Acoustic Signal-Based Cutting Tool Condition Monitoring in Woodworking Using Deep Learning

Acoustic Signal-Based Cutting Tool Condition Monitoring in Woodworking Using Deep Learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Acoustic Signal-Based Cutting Tool Condition Monitoring in Woodworking Using Deep Learning Srđan Svrzić, Mlađan Popović, Mladen Furtula, Zoran Nikolić, Marija Đurković, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8969198/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Effective monitoring of wood cutting processes is crucial for maintaining cutting quality, extending tool life, and reducing energy and material waste. While conventional methods rely on cutting forces, power consumption, or temperature measurements, this study explores acoustic signal analysis as an alternative, leveraging its sensitivity to tool condition, cutting dynamics, and transient high-frequency events. The primary contribution of this study lies in the development of frequency–intensity–time representations derived from acoustic signals generated during particleboard cutting and in the subsequent classification of these patterns using deep learning algorithms. In parallel, cutting power consumption and tool temperature were analyzed as reference monitoring variables, with the aim to prove the hypothesis that deep learning-based classification of acoustic data could achieve comparable to or higher accuracy and reliability. Experimental investigations were conducted using two circular saw blades with different dullness conditions saw blades operating at feed rates of 10 m/min and 15 m/min. The workpiece material was melamine-faced particleboard. Statistical analysis revealed highly significant differences in power consumption between cutting conditions, while temperature measurements did not consistently distinguish between feed rates or tool wear states. Acoustic signals were acquired, pre-processed, and classified using several deep learning architectures. The highest classification accuracy of 92.00% was achieved using ConvNetXt-XL, while other architectures, including ResNet, EfficientNet, VGG19, DenseNet, MobileNet V2, Inception V3, and SqueezeNet achieved accuracies between 73.00% and 91.00%. The obtained results demonstrate the effectiveness of acoustic signals as a reliable monitoring variable for intelligent diagnostics and condition monitoring in wood machining processes. Wood industry Particleboard Cutting tools Maintenance Sound analysis Deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 1.0 Introduction The global wood industry presents a dynamic and ever-growing field valued at approximately 1 trillion USD, while consuming more than 4 billion cubic meters of wood according to the F.A.O. report [1]. It is estimated that its gross growth will continue by 25% until 2030. The major bottleneck of the wood industry sector is the low total utilization, which is estimated to be approximately 30–55% for solid wood products [2] and 71–98% for wood-based panels (MDF and particleboards) applications. At the same time, the global woodworking machinery market size is valued at $ 5.03 billion in 2024 and is projected to grow to $ 7.88 billion by 2032 [3]. Proper maintenance and appropriate operation of wood processing machine tools are critical for cost and material loss reduction [4,5]. Poor organization and management of tooling can lead to an increase in hidden costs, resulting in productivity losses of approximately 20–30% [6,7] of working time and additional tool-related cost increases of 10–30% [8]. Circular saw blades used in wood panel processing are exposed to abrasive constituents, adhesives, and coatings, resulting in progressive loss of sharpness that is difficult to detect without interrupting production. Consequently, there is a clear need for non-invasive, robust, and information-rich monitoring approaches capable of operating in real production environments and providing reliable indicators of tool condition. Particleboard was selected as the focus of this study due to its extensive use in furniture manufacturing and panel processing, as well as its relatively uniform internal structure when produced under standardized industrial conditions. Under controlled manufacturing parameters, such as adhesive formulation, wood species composition, and pressing regime, particleboard exhibits stable mechanical and physical properties [9]. While density gradients across the board thickness may influence local stiffness and modulus of elasticity [10], these variations are systematic and reproducible, making particleboard a suitable material for controlled machining studies. In industrial practice, particleboard is predominantly processed by circular sawing, where tool wear has an immediate and measurable impact on cutting power, surface integrity, and process stability. Previous research on particleboard machining has largely focused on tool geometry and cutting parameters, including rake angle, clearance angle, bevel angle, cutting speed, feed rate, and tooth configuration [11,12]. These studies have demonstrated strong links between tool geometry, cutting forces, airborne dust emission, and surface quality. However, they typically rely on conventional monitoring variables, such as cutting forces, vibration signals, or power consumption. While cutting power has proven to be a robust and industrially accessible indicator of process load [13,14,15,16,17], it represents an aggregated response of the system and may fail to capture localized or transient phenomena associated with early-stage tool wear. Tool temperature, although investigated as a monitoring variable [13,18], often lacks sufficient sensitivity under moderate cutting conditions and stable regimes. Acoustic signals generated during wood machining constitute a non-contact and information-rich process output that has received comparatively limited attention. Cutting sound originates from tool–material interaction, chip formation, and frictional effects, and is inherently sensitive to high-frequency and transient events. Early work on acoustic emission in machining demonstrated its potential for process monitoring [19], yet its application in woodworking has remained largely exploratory. Recent studies have shown that, with appropriate signal processing, sound can be used to identify cutting regimes, tool types, and even wood species [17,20,21]. However, these investigations did not directly address tool condition monitoring nor systematically evaluate the robustness of acoustic features under closely related operating conditions. The increasing adoption of artificial intelligence in manufacturing offers new opportunities to exploit such complex signals. Deep learning methods, particularly convolutional neural networks (CNNs), have demonstrated exceptional performance in extracting discriminative patterns from high-dimensional data. In wood science, CNNs have been successfully applied to wood species identification [22,23], fiber segmentation from X-ray CT images [24], and indirect assessment of tool wear through image-based analysis of drilled holes [25]. Comprehensive reviews by [26] further emphasize the potential of deep learning architectures—including CNNs, autoencoders, and recurrent networks—for intelligent machining and tool monitoring. Nevertheless, the application of deep learning to airborne acoustic signals for direct tool condition monitoring in circular sawing of wood-based panels has not yet been systematically investigated. Building on these findings, the present study explores the use of acoustic signals as the primary input for a deep learning-based tool condition monitoring during particleboard cutting. Acoustic data are transformed into time–frequency representations and classified using a representative set of modern CNN architectures. Cutting power consumption and tool temperature are simultaneously analyzed as established reference variables to benchmark the performance and reliability of the acoustic approach. The engineering motivation lies in demonstrating that non-contact acoustic monitoring can provide sensitivity comparable to, or exceeding, that of conventional monitoring proxies, while offering superior responsiveness to transient and high-frequency phenomena. The novelty and contribution of this work lie in: (i) systematically converting cutting sound into spectrogram-based representations suitable for CNN analysis, (ii) benchmarking multiple state-of-the-art CNN architectures under consistent experimental conditions, and (iii) evaluating the practical implications of acoustic-based monitoring for tool-condition assessment, maintenance planning, and process reliability in industrial woodworking environments. 2.0 Materials and Methods 2.1 Materials The experiments were conducted using melamine-faced particleboard, designated as type P2 in accordance with the European standard EN 312:2010. typically employed in the manufacture of furniture. The particleboards, with a nominal thickness of 18 mm, were sampled from the same stock at a domestic wholesaler company (Belgrade, Serbia). These boards were manufactured by Egger Group GmbH in a factory located in Răducăeni, Romania. The particleboard samples are divided into four groups according to specific machine cutting variables in the main experiment. These variables are as follows: the state of the saw (new, used) and the feed rate (10 and 15 m/min). Each sample group, consisting of three panels, is designated as follows: Used U10 - used saw; feed rate: 10 m/min, Used U15 - used saw; feed rate: 15 m/min, New N10 - new saw; feed rate: 10 m/min, New N15 - new saw; feed rate: 15 m/min. Basic physical and mechanical tests were performed to determine both the conformity of particleboard samples with the requirements of the EN 312 standard, as well as to examine the variations between the sample groups. In that aspect, the properties of particleboards, such as: density (EN 323, 1993), moisture content (EN 322, 1993), bending strength, modulus of elasticity (EN 310, 1993) and internal bond (EN 319, 1993) are determined (Fig. 2 (a) to (c)) and presented in Table 1 . Table 1 Statistical comparison of basic physical and mechanical properties of sample groups at the significance level of α = 0.05 Property Unit Sample group (mean value ± st. dev.) ANOVA statistics Used u10 Used u15 New n10 New n15 F P-value F crit. Thickness mm 18.19 ± 0.02 18.18 ± 0.007 18.18 ± 0.026 18.18 ± 0.015 0.30759 0.81968 2.94669 Density kg/m³ 658.8 ± 26.31 647.9 ± 20.2 667.7 ± 9.89 658.3 ± 7.98 1.66113 0.19796 2.94669 Moisture content % 7.13 ± 0.093 7.11 ± 0.127 7.07 ± 0.078 7.10 ± 0.082 0.29887 0.82559 3.49030 Bending strength N/mm² 17.20 ± 2.364 18.64 ± 1.237 17.94 ± 0.495 17.19 ± 1.434 1.21845 0.32887 3.09839 MOE N/mm² 2623 ± 79.1 2910 ± 26.2 2666 ± 177.4 2743 ± 169.7 5.69749 0.00548 * 3.09839 Internal bond N/mm² 0.431 ± 0.065 0.417 ± 0.045 0.449 ± 0.045 0.428 ± 0.056 0.52531 0.66849 2.94669 * Significant difference 2.2 Cutting Tools and Conditions The particleboard panels were cut using two FREUD LU2C 1200 circular saw blades, with characteristics presented at Figs. 1 and 2 h and Table 2 . The first piece was evaluated in its original condition, whereas the second piece had been utilized for approximately 700 meters with similar cutting conditions. Table 2 Tool and cutting conditions Cutting tool LU2C 1200 RPM 4000 min − 1 Feed per tooth 0.03125 mm Tooth shape ATB Number of teeth 80 Diameter (mm) 250 Body thickness b (mm) 2.2 Cutting width B (mm) 3.2 Rake angle (°) 18 Clear angle (°) 5 Inclination angle (°) 10 Tool override (mm) 10 The tool condition in terms of saw tooth dullness was estimated according to photos made by microscope (Fig. 2 g). The photos taken under the microscope should illustrate a brand new, unpackaged circular saw blade and the one of the same type that has already been used. It was expected that the tooth edges and surfaces for the new saw are in perfect condition, while the teeth image of a used tool should have traces of use in the form of different deviations of cutting edges, cracks, and material deposits on cutting surfaces. Throughout the course of each experimental stage, the thermal imaging camera was maintained in a state of recording (Fig. 2 j). The maximum temperature values for each cut were duly noted and meticulously documented. In accordance with the findings of the research, the tool temperature results were subjected to analysis using the ANOVA statistic. 2.3 Experimental Setup The study was conducted in the Laboratory of Machines and Apparatus at the Faculty of Forestry, University of Belgrade (Beograd, Serbia). The machining system used for this study was a Minimax CU 410K combined machine (SCM, Rimini, Italy) equipped with a 3 kW three-phase asynchronous motor (Fig. 2 d). The speed of the motor was set by a customized frequency controller (Fig. 2 f) to 4000 rpm with a corresponding frequency of 50.5 Hz. The noise occurring in the cutting process was recorded, in WAV format (Fig. 2 n) using a dbx RTA-M measurement microphone with an electret condenser on the back (Fig. 2 e). The RTA-M is an omnidirectional, low-profile frequency measurement microphone specifically designed to record all frequencies from 20 Hz to 20 kHz, ensuring accurate "real-time "pinging" analysis of the audio signal. It is operated with phantom power. To reduce the effects of vibration, the microphone is housed in a vibration-damping rack. The Focusrite Scarlet SOLO USB audio interface was connected to a PC (Fig. 2 o). The cutting power measuring device (Fig. 2 i) has a Circutor CW-TAN active power transformer for unbalanced three-phase systems with the following characteristics: alternating current 5 A, alternating voltage 230 V, frequency 50 Hz, accuracy 0.5% and analog voltage output 0–10. The possible measuring ranges are 5, 10 and 15 kW. The measured cutting power data are given in Watts. The operator selects the expected range for a better resolution of the results. The Dino-Lite Edge USB microscope with 470x magnification was used for visual observation of the tool condition (Fig. 2 g). The temperature of the saw teeth was measured using a FLIR E50 thermal imaging camera (Fig. 2 j). The recordings were in the form of movie clips which were later processed and statistically analyzed (Fig. 2 m). 2.4. Signal Acquisition and Processing Audacity, a cross-platform open-source audio software, was used to record and process the audio signals. The measurements were carried out at a sampling rate of 44100 Hz. The sounds generated in the cutting process were captured with the microphone and recorded on the PC as wave files. Spectral analysis was performed on these recordings using the fast Fourier transform (FFT) and the short-time Fourier transform (STFT) provided by MATLAB R2023 Edition (Fig. 2 o). On these recordings, spectral analysis was performed using FFT and STFT (Fig. 2 s). Spectrograms are 3D (frequency-time-power) charts obtained by STFT of original sound signals (Fig. 2 p). The entire system for cutting power data acquisition is based on the Power Expert software platform, and the sampling rate has been set to 1000 Hz (Fig. 2 k). The spectrograms further served as training data for the deep learning networks (Fig. 2 r) considered in this study. 2.5 Deep Learning Architectures and Training Protocol The eight convolutional neural network (CNN) architectures were considered to form a representative and engineering-relevant benchmark for spectrogram image classification. ResNet introduces residual or “skip” connections that allow gradients to propagate more effectively through deep networks, mitigating the vanishing gradient problem. This design enables the training of very deep models, which is beneficial for capturing subtle spectral variations in acoustic data that may indicate tool wear or cutting conditions [27]. MobileNetV2 relies on depthwise separable convolutions and inverted residuals, significantly reducing computational complexity and memory requirements [28]. While lighter than most CNNs, it still maintains strong accuracy. This makes it particularly attractive for deployment in edge AI scenarios, where low-power, real-time monitoring systems are required. Inception V3 employs inception modules that process multiple filter sizes in parallel, capturing both fine and coarse features within the same layer [29]. This ability to extract multi-scale representations is advantageous for spectrograms, where patterns may occur at different frequencies and time windows simultaneously. SqueezeNet was designed with parameter reduction in mind, using fire modules that “squeeze” with 1×1 convolutions before “expanding” into a mix of 1×1 and 3×3 filters [30]. This results in a model under 5 MB in size, yet capable of competitive accuracy. For industrial monitoring, this compactness allows easier integration into embedded systems and sensor nodes. DenseNet establishes direct connections between each layer and all subsequent layers, promoting feature reuse and improving gradient flow [31]. This architecture can capture both local and global acoustic structures efficiently, which is useful when subtle harmonic patterns must be linked with broader spectral trends. VGG19 represents a classical deep CNN, characterized by its simple yet powerful architecture of stacked 3×3 convolutions followed by fully connected layers [32]. Although computationally heavier, VGG19 provides a strong reference baseline and facilitates comparison with newer, more efficient architectures. EfficientNet-B0 was added as a modern CNN with a principled balance of accuracy and computational cost via compound scaling [33]. Finally, ConvNeXt represents a modernized CNN design, assessing whether recent convolutional improvements translate into measurable gains for the considered challenge [34]. Overall, the benchmark set was intentionally constructed to cover a spectrum from ultra-compact (SqueezeNet) through mobile-efficient (MobileNetV2), canonical high-accuracy (ResNet/DenseNet), multi-scale feature extraction (Inception V3), classic deep baseline (VGG19), and modern accuracy/efficiency (EfficientNet-B0, ConvNeXt), enabling a balanced evaluation in terms of accuracy, efficiency, and practical industrial applicability. All classification models were pre-trained on the ImageNet [35] dataset and fine-tuned using the PyTorch framework. The training was performed using the Adam optimization algorithm, [36] with the cross-entropy loss function and the initial learning rate of 1e-4 (which was decreased by a factor of 0.1 every 7 epochs). All the computations were performed on the Lambda workstation with the AMD Threadripper 3970X (32 cores, 3.79 GHz processor), 128 GB RAM, and two Titan RTX (24 GB) + NVLink GPUs. From the four sets of 250 measurements for each class (one set per class), 200 samples per class were randomly selected for artificial neural network (CNN) training, while the remaining were randomly split for the purpose of validation and testing (25 per class). To improve robustness and reduce overfitting, we applied online augmentation for the training data set tailored to time–frequency representations. Specifically, we applied: time shift/random temporal cropping, SpecAugment masking (time and frequency masking), and additive Gaussian noise (waveform or spectrogram). When augmenting at the waveform level, we additionally used a small time-stretch and pitch shift. Augmentations were applied only during training with a fixed probability (not on validation/test). To quantify performance variability, we repeated training five times with different random seeds while keeping the test set fixed (25 samples per class). We computed mean ± standard deviation of metrics considered for the algorithms evaluation were: Accuracy quantifies the proportion of correct predictions across all classes, expressed as \(\frac{TP+TN}{TP+TN+FP+FN}\) . F1 – Score (Macro Average) computes the unweighted arithmetic mean of class-specific F1-scores, defined as \(\frac{1}{C}{\sum}_{i=1}^{C}=2\frac{Precision\bullet Recall}{Precision+Recall}\) , where C is the number of classes. This metric equally weights all classes, making it suitable for scenarios where minority class performance is critical; Cohen’s Kappa measures inter-rater agreement corrected for random chance: \(\kappa=\left({p}_{0}-{p}_{\epsilon}\right)/\left(1-{p}_{\epsilon}\right)\) , where p 0 is observed accuracy and p ε is expected chance agreement. Values are interpreted as: ≤0 (no agreement), 0.01–0.20 (slight), 0.21–0.60 (moderate), 0.61–0.80 (substantial), 0.81-1.00 (near-perfect). MCC (Matthews Correlation Coefficient) — a unique measure of the quality of predictions (correlation true vs. predicted), robust to class imbalance. MAE (Mean Absolute Error) — the average "distance" between the real and predicted class; useful when the classes have an order (ordinal). 3.0 Results and Discussion The results of this study are presented in the form of visual aids, including photographs, screen captures, tables and diagrams. As illustrated in Figs. 3 (a) and (b), the photographic evidence provides a rudimentary understanding of the sawtooth condition. It is possible to observe the deterioration of the tooth edges and related surfaces of a used saw. As demonstrated in Fig. 4, an investigation into cutting power measurements revealed the potential to generate a diagram illustrating the average cutting power for varying feed rates and saw conditions. 3.1 Cutting power The cutting power curves are presented as a mean value for each measurement, calculated at the precise moment of each cut. It was observed that the highest mean cutting power was exhibited by U15, followed by N15, N10 and finally U10. It was hypothesized that the mean cutting power would increase both with machining length and feed rate [37,38]. As posited by [39], intensive tool wear, as indicated by flank wear width, manifests at a distance of approximately 550 meters. However, as asserted by [40], this wear may occur much earlier, at around 200 meters of machining distance. Thereafter, there is a consistent deterioration until 2750 meters. One potential explanation for this phenomenon is that, during the initial phases of tool wear, the process is purely cutting, whereas in later phases, it is more akin to the breaking off of processed particleboard. In the context of a pure cutting process, an increase in friction between tooth surfaces and the material is observed. This rise in friction results in an enhancement of the cutting power. In addition, it is conceivable that the initial stages may witness a decline in kerf width due to pronounced tool wear, particularly along the lateral edges. This phenomenon is attributed to the diminished friction, consequently leading to a reduction in power consumption during the latter stages of operation. This is attributable to the more stable and uniform tool wear characteristics that characterize these latter stages. Nevertheless, an apparent discrepancy in the consumed power is evident for both the new and used saws when an increased feed rate is employed. Statistical analysis of the cutting power achieved during the experiment showed significant influence of both the saw condition and the feed rate (Tables 3 and 4 ). The maximal cutting power was calculated by subtracting the idle machine power from the total cutting power. Cutting power increased with the feed rate, regardless of the tool condition. However, the tool condition performed differently in regard to the feed rate. The new saw achieved lower maximal cutting power during higher feed rate, than the used one. The results were opposite for the lower feed rate. Table 3 ANOVA statistics of maximal cutting power values, at the significance level of α = 0.05 Sample group Count Mean (W) Standard deviation ANOVA statistics F P-value F crit N10 25 650.41 30.988 124.389 5.93·10 − 33 * 2.69939 N15 25 710.27 38.432 U10 25 587.40 28.358 U15 25 747.99 27.151 * Significant difference Table 4 The statistical analysis of maximal cutting power values between the single pairs of sample groups (Bonferroni test) Sample group pair p-value (T-test) Significant difference Bonferroni correction (adjusted α) N10 vs N15 2.0099 · 10 − 07 Yes 0.00833 N10 vs U10 1.2643 · 10 − 09 Yes N10 vs U15 7.5126 · 10 − 16 Yes N15 vs U10 3.6171 · 10 − 17 Yes N15 vs U15 0.000212346 Yes U10 vs U15 2.4164 · 10 − 25 Yes 3.2 Temperature The temperature of the saw blade was the focus of continuous monitoring during each cutting run, and the peak temperature values were subjected to statistical comparison across the sample groups. The results of the analysis of variance (ANOVA) indicated a statistically significant difference among the groups, thereby confirming that one or both of the experimental variables (saw condition and/or feed rate) affected the temperature buildup (see Table 5 ). In order to evaluate the impact of specific experimental variables, a post-hoc Bonferroni test was used to compare the individual pairs of sample groups (see Table 6 ). The findings indicated that the saw condition (new vs. used) exerted a substantial influence on the saw blade temperature. Statistical analysis revealed that both the N10 and N15 sample groups exhibited mean values of saw temperature that were significantly higher than those observed in the U10 and U15 samples. However, the saw temperature remained unaffected by the feed rate, as evidenced by the lack of statistical significance in temperature variations between 10 and 15 meters per minute for both new and used saws. Table 5 Analysis of variance (ANOVA) of maximum saw temperature values between the sample groups, at the significance level of α = 0.05 Sample group Count Mean (°C) Standard deviation ANOVA statistics F P-value F crit. N10 80 61.28 10.307 97.09043 2.499· 10 − 44 * 2.63373 N15 76 57.93 5.008 U10 79 47.08 6.302 U15 79 45.39 5.362 * Significant difference Table 6 The statistical analysis of maximum saw temperature values between the single pairs of sample groups (Bonferroni test) Sample group pair p-value (T-test) Significant difference Bonferroni correction (adjusted α) N10 vs N15 0.011631 No 0.00833 N10 vs U10 8.9690 · 10 − 20 Yes N10 vs U15 2.0668 · 10 − 24 Yes N15 vs U10 2.1804 · 10 − 23 Yes N15 vs U15 5.9649 · 10 − 32 Yes U10 vs U15 0.072467 No 3.3 Sound signal As demonstrated in Fig. 5, a comparison of average sound intensity levels was facilitated through the analysis of noise recordings in the context of the cutting time domain. The mean integral of all intensities at all frequencies is the average sound intensity for a single cutting run. As was the case in the preceding instance, the maximum values for sound intensity were observed for U15, with N15 following in second place. However, these values were only recorded during the final two seconds of the run. The sound intensity curves for N10 and U10 demonstrate a high degree of similarity. As was the case previously, it is evident that there is a certain discrepancy in the sound intensity produced by the tool in relation to the varying feed rates. However, no such variation is observed in the tool's dullness. It is evident that the sound signal is significantly impacted by its frequency domain; therefore, the calculation of the mean FFT values was deemed appropriate. The FFT was conducted at a computing frequency of 100 Hz. Subsequent to this, the mean value for all frequencies was calculated. The results pertaining to the mean FFT values for all stages of measurement are illustrated in Figs. 7 (a) and (b). As illustrated in Fig. 7 , it is evident that the FFT graphs for all four experimental stages demonstrate a similar pattern of behavior. The characteristic peaks manifest at equivalent frequencies, exhibiting solely disparate levels of intensity. As the graph of the FFT for N10 is not clearly visible in Fig. 7 a, Fig. 7 b provides a more comprehensive overview of its performance. It was also evident that the trend previously identified in the context of cutting power and average sound intensity measurements was also apparent in the case of average FFT graphs. The highest values exhibited were observed for U15, followed by N10 and U10. It has been established that the dominant peak for all average FFT graphs occurs precisely at 5200 Hz. This is consistent with the natural frequency of the saw. It is evident that other peaks manifest as higher harmonics of the fundamental frequency. The graphs are distinguished solely by the intensity of their peaks. As demonstrated in Figs. 6 and 7 , the graph was unable to achieve the requisite level of accuracy for process recognition. Consequently, the sound recordings for all 1,000 measurements in the WAV format were subjected to STFT with a Hann window with 50% overlapping, smoothing the spectral line considerably and also giving frequency-time-power domain of the sound signal. As demonstrated in Figs. 8 (a) and (b), the STFT provided a graphical representation of sound signals in the form of spectrograms. As demostrated in Figs. 9 (a)–(d), spectrogram representations of the acoustic signals were employed as the primary input data for the convolutional neural network (CNN). From each of the four datasets comprising 250 measurements, 200 samples were randomly selected for network training, while the remaining 50 samples were reserved for testing purposes. Several state-of-the-art CNN architectures were evaluated in this study, including MobileNetV2, ResNet, Inception V3, SqueezeNet, DenseNet, VGG19, EfficientNet-B7—selected for its balance between accuracy and computational efficiency—and ConvNeXt-XL, which represents a modern CNN design paradigm [34]. The classification performance achieved by these architectures is summarized in Table 7 and forms the basis for the following discussion, which focuses on the comparative effectiveness of the applied models and their suitability for sound-based monitoring of the wood machining process. Table 7 CNN metrics (performance of the developed deep learning models for machining sound classification Model Accuracy F1 (Macro) MCC MAE Cohen's Kappa MobileNetV2 81.00 ± 3.92% 0.810 ± 0.04 0.747 ± 0.12 0.270 ± 0.06 0.747 ± 0.05 ResNet 91.00%±2.86% 0.911 ± 0.02 0.881 ± 0.10 0.120 ± 0.05 0.880 ± 0.03 Inception_v3 80.00%±4.00% 0.801 ± 0.04 0.734 ± 0.15 0.270 ± 0.06 0.733 ± 0.05 SqueezeNet 73.00%±4.44% 0.730 ± 0.04 0.641 ± 0.19 0.380 ± 0.07 0.640 ± 0.05 DenseNet 82.00%±3.84% 0.821 ± 0.03 0.761 ± 0.12 0.220 ± 0.05 0.760 ± 0.05 VGG19 84.00%±3.67% 0.840 ± 0.03 0.787 ± 0.09 0.260 ± 0.06 0.787 ± 0.04 EfficientNet-B7 89.00%±3.13% 0.890 ± 0.03 0.854 ± 0.08 0.140 ± 0.04 0.853 ± 0.04 ConvNeXt-XL 92.00%±2.71% 0.920 ± 0.02 0.893 ± 0.08 0.110 ± 0.03 0.893 ± 0.03 The results shown in Table 7 indicate a clear hierarchy of performance among the analyzed CNN architectures in the task of multi-class audio processing classification. ConvNeXt-XL achieved the best overall performance, with the highest accuracy (92.00%), the highest Macro F1 (0.920), the highest MCC (0.893), and the lowest MAE (0.110), indicating high discriminative power and stability in the presence of class imbalance. The high value of Cohen's κ (0.893) confirms the strong agreement of the model with the benchmarks, significantly above the level of random guess. ResNet shows very competitive results (Accuracy = 91.00%, Macro F1 = 0.911, MCC = 0.881, κ = 0.880), confirming that residual connections effectively alleviate the gradient degradation problem and enable stable learning of deep representations. Although slightly inferior to ConvNeXt-XL, ResNet maintains an excellent balance between accuracy and robustness, with a low MAE (0.120). EfficientNet-B7 achieves solid performance (Accuracy = 89.00%, κ = 0.853), demonstrating the good efficiency of the architecture based on scaling depth, width and resolution. However, compared to ResNet and ConvNeXt-XL, a slight decrease in robustness is observed (lower MCC and higher MAE), suggesting a greater sensitivity to confusion between classes. DenseNet (Accuracy = 82.00%, κ = 0.760) shows correct but moderate results. Although the feature reuse mechanism contributes to learning stability, the performance indicates a limited ability to separate classes compared to more modern architectures. MobileNetV2 achieves 81.00% accuracy with κ = 0.747, which confirms good generalization considering the low complexity of the model. However, in this experimental environment its efficiency advantage does not come with performance at the level of the leading models, indicating a trade-off between parametric economy and discriminative power. Models VGG19, Inception_v3 and SqueezeNet make up the bottom rank in terms of performance. SqueezeNet is the weakest (Accuracy = 73.00%, Macro F1 = 0.730, κ = 0.640), which clearly indicates the limited representative power of the extremely lightweight architecture. VGG19 (Accuracy = 84.00%, κ = 0.787) and Inception_v3 (Accuracy = 80.00%, κ = 0.733) show a pronounced sensitivity to confusion between classes, which is reflected in lower MCC values and higher MAE, especially for borderline samples. Overall, the results confirm that the more modern architectures (ConvNeXt-XL and ResNet) provide the best compromise between classification accuracy, robustness and stability. Differences between models arise not only from overall accuracy, but also from the ability to reliably separate classes, which is clearly quantified through MCC, Macro F1 and Cohen's κ. These findings indicate that architectural innovations play a crucial role in the quality of latent representations in the processing sound classification tasks. Discussion of the obtained results with respect to previous studies Tool condition monitoring in machining processes has been extensively studied using a variety of sensing modalities. Traditional approaches have relied primarily on measurements of cutting force, vibration, power consumption, and tool temperature. For instance, [41] and [42] investigated cutting force as a reliable indicator of tool wear, while [43] expanded this line of research to engineered wood products. However, force measurements often require invasive instrumentation that is difficult to integrate into industrial environments. Vibration-based methods have also been widely explored [12,44], but their sensitivity to machine dynamics and environmental noise may limit their robustness in factory settings. Another established line of research concerns acoustic emission monitoring, which has demonstrated potential in capturing high-frequency signals directly linked to cutting phenomena [19,13]. Studies by [14] and [15] confirmed that sound analysis can support adaptive control in wood machining, while [16,20] and [21] showed that acoustic features correlate with tool wear and sawing conditions. Parallel studies investigated power consumption [45] and tool temperature [18,40,13] as indirect indicators of machining states. While these variables provide valuable information, they often fail to distinguish subtle differences in tool condition, especially at early wear stages. Compared with these approaches, the present study highlights several novel contributions. First, it focuses on acoustic signal classification using deep neural networks, rather than conventional statistical or shallow learning techniques. Previous works in wood machining often relied on handcrafted features or limited classifiers, whereas this research employs advanced CNN architectures (ResNet, MobileNetV2, DenseNet, etc.) that automatically extract discriminative features from spectrograms. Second, the results demonstrate superior classification accuracy, with ResNet achieving 96.2%, which is higher than most accuracies reported in prior woodworking signal studies that typically ranged between 80–90% [25,26]. A third contribution lies in the multi-sensor framework, where acoustic monitoring was complemented with power and temperature measurements. Although these auxiliary variables did not yield sufficient accuracy independently, their combined analysis provided valuable context for validating acoustic-based predictions. This integration reflects a step toward holistic monitoring systems suitable for Industry 4.0 manufacturing environments. Finally, the study emphasizes practical applicability. By evaluating lightweight architectures such as MobileNetV2 and SqueezeNet, the work demonstrates the feasibility of deploying acoustic-based monitoring on edge devices for real-time process control. This is an important distinction from previous studies that primarily validated methods in controlled laboratory environments without considering industrial integration. In the context of Industry 4.0 and smart manufacturing, the proposed approach provides distinct advantages over previous monitoring strategies. Earlier studies based on cutting force [41,42], vibration [12,44], or power and temperature signals [13,18,45] demonstrated useful insights but generally required invasive sensors, complex instrumentation, or lacked the robustness needed for integration in industrial environments. In contrast, acoustic sensing is low-cost, non-invasive, and easily deployable, which makes it well-suited for large-scale industrial applications. By leveraging deep learning models such as ResNet and MobileNetV2, this study further ensures high classification accuracy with architectures that are also computationally efficient, thus opening the possibility of real-time edge deployment in production lines. This combination of acoustic monitoring and lightweight CNNs reflects a shift from purely experimental validation toward practical, scalable solutions for predictive maintenance, process optimization, and adaptive control in woodworking and beyond. Hence, the study not only advances academic understanding but also aligns closely with the industrial demands of next-generation manufacturing systems. Study limitations and future research directions While the present study has demonstrated the potential of deep learning applied to acoustic signals for tool condition monitoring in woodworking, several avenues remain open for future research. First, the experiments were conducted under controlled laboratory conditions with a limited number of feed rates and two saw blade states. Expanding the study to include a wider range of cutting speeds, tool geometries, and machining conditions would enable the development of more generalizable models capable of handling diverse industrial scenarios. Second, although this research employed multiple CNN architectures with excellent classification accuracy, further exploration of hybrid and ensemble models could improve robustness, particularly in borderline cases where signal features overlap between classes. Incorporating attention mechanisms or transformer-based architectures may also enhance the ability to capture long-range dependencies in acoustic data. Third, integration of acoustic signals with other sensor modalities, such as vibration, image-based inspection, or current monitoring, could lead to multi-modal fusion frameworks that provide more comprehensive assessments of tool wear and process quality. Fourth, future research should investigate real-time deployment strategies, including the use of edge AI devices and embedded platforms. This would allow the development of compact, low-power monitoring systems directly applicable to Industry 4.0 smart factories. Finally, long-term industrial case studies are needed to validate system performance in dynamic production environments, where noise, variability of workpieces, and operator influences may affect signal quality. Such validation would bridge the gap between experimental research and large-scale industrial adoption. 4.0 Conclusions Based on the information presented, it can be concluded that the findings from the CNN classification offer a satisfactory tool for monitoring the cutting process. Any result with an accuracy level above 95% can be considered excellent, providing a powerful tool for process control and observation. The results obtained from cutting power measurements were also expected and satisfactory, making this metric a valuable, robust, and reliable indicator of overall process load. However, statistical analysis of temperature measurement results did not demonstrate the expected causality. Although some anticipated outcomes were observed, such as significant temperature variations for specific feed rates, there were no clear indicators to determine the tool's condition regarding dullness. Cutting power provided excellent classification results among the different sample groups, but there are certain limitations to this approach, including material heterogeneity, the provision of only a single scalar value, slower response, and the effects of motor efficiency and transmission losses. Temperature measurement did not provide satisfactory classification among the defined sample groups. Statistical analysis showed that it was not possible to distinguish the feed rate for the same tool condition. Sound, as a multi-dimensional signal, was preprocessed to produce inputs for CNN classification datasets in the form of spectrogram images. The accuracy of CNN classification among different groups exceeded 80%, which is considered excellent. Accuracies were 92.00%, 91.00%, 89.00%., 84.00%, 82.00%, 81.00%, 80.00% and 73.00% for ConvNet Xt-XL, ResNet, EfficientNet-B7, VGG19, DenseNet, MobileNetV2, Inception V3, and SqueezeNet, respectively. Declarations Funding This work was supported by the Ministry of Science, Technological Development and Innovation of the Republic of Serbia, under registration number 451-03-65/2024-03/200169, dated 05.02.2024. Author contributions 1. SS: conceptualization, data curation, formal analysis, investigation, resources, methodology, supervision, writing. 2. MP: formal analysis, investigation, validation. 3. MF: conceptualization, formal analysis, investigation. 4. ZN: software, supervision. 5. MĐ: funding acquisition, resources. 6. AD: project administration. 7. PT: resources. 8. AV: data curation, methodology, resources, software, supervision, validation, visualization, writing. Acknowledgements The authors would like to thank Marko Stojičić, Faculty of Forestry employee, who greatly assisted throughout entire investigation process, Miloš Marinković from EGGER Serbia and Darex Home Company, Belgrade, Serbia for obtaining necessary laminated particle boards and Faculty of Forestry, Belgrade, Serbia for participating in publishing costs. Data available statement Data available on request from the authors. The data that support the findings of this study are available from the corresponding author upon reasonable request. Ethics statement Ethics approval was waived for this study because no patients’ data were reported. During the preparation of this work, the authors used Grammarly, InstaText and ChatGPT to improve readability. The authors reviewed and edited the content and take full responsibility for the publication. References Food and Agriculture Organization of the United Nations, Global Forest Products Facts and Figs. 2024, FAO, Rome, Italy, 2024, doi: 10.4060/cd8005en. K. Hassan, A. Villa, S. Kuittinen, J. Jänis, and A. Pappinen, “An assessment of side-stream generation from the Finnish forest industry,” Journal of Material Cycles and Waste Management, vol. 21, pp. 265–280, 2018, doi: 10.1007/s10163-018-0787-5. Fortune Business Insights, Woodworking Machinery Market Size, Share & Industry Analysis by Type and Application, and Regional Forecast , 2025. Available online: https://www.fortunebusinessinsights.com/machinery-industry J.-W. Sun, L.-F. Xi, S.-C. Du, and E.-S. Pan, “Tool maintenance optimization for multi-station machining systems with economic consideration of quality loss and obsolescence,” Robotics and Computer-Integrated Manufacturing, vol. 26, no. 2, pp. 145–155, 2010, doi: 10.1016/j.rcim.2009.07.005. J. He, H. Gao, S. Li, L. Guo, Y. Lei, and A. Cao, “An intelligent maintenance decision-making based on cutters’ economic life,” International Journal of Production Economics, vol. 267, Art. no. 109075, 2024, doi: 10.1016/j.ijpe.2023.109075. G. Byrne, D. Dornfeld, and B. Denkena, “Advancing cutting technology,” CIRP Annals – Manufacturing Technology, vol. 52, no. 2, pp. 483–507, 2003, doi: 10.1016/S0007-8506(07)60200-5. F. Klocke, Manufacturing Processes 1: Cutting. Berlin, Germany: Springer, 2011, doi: 10.1007/978-3-642-11979-8. Woodworking Network, 2025. Available online: https://www.woodworkingnetwork.com/ G. Nemli, I. Aydin, and E. Zekovic, “Evaluation of some properties of particleboard as a function of manufacturing parameters,” Materials & Design, vol. 28, no. 4, pp. 1169–1176, 2007, doi: 10.1016/j.matdes.2006.01.015. E.-D. Wong, P. Yang, M. Zhang, Q. Wang, T. Nakao, K.-F. Li, and S. Kawai, “Analysis of the effects of density profile on bending properties of particleboard using FEM,” European Journal of Wood and Wood Products, vol. 61, no. 1, pp. 66–72, 2003, doi: 10.1007/s00107-002-0350-9. J. P. Davim (Ed.), Wood Machining. Hoboken, NJ, USA: Wiley, 2013, pp. 176–177. V. Nasir and J. Cool, “Characterization, optimization, and acoustic emission monitoring of airborne dust emission during wood sawing,” International Journal of Advanced Manufacturing Technology, vol. 109, pp. 2365–2375, 2020, doi: 10.1007/s00170-020-05842-5. V. Nasir, J. Cool, and F. Sassani, “Acoustic emission monitoring of sawing process using artificial intelligence,” International Journal of Advanced Manufacturing Technology, vol. 102, pp. 4179–4197, 2019, doi: 10.1007/s00170-019-03526-3. R. Licow, D. Chuchala, M. Deja, K. A. Orlowski, and P. Taube, “Effect of pine impregnation and feed speed on sound level and cutting power in wood sawing,” Journal of Cleaner Production, 2020, doi: 10.1016/j.jclepro.2020.122833. M. Derbas, S. Frömel-Frybort, C. Laaber, and M. Riegler, “Sound analysis of mechanical wood cutting processes as a basis for adaptive process control,” in Proc. 9th Hardwood Conf., Sopron, Hungary, 2021. S. Svrzić, M. Đurković, G. Danon, M. Furtula, and D. Stanojević, “Acoustic emission analysis in circular saw cutting of beech wood,” BioResources, vol. 16, no. 4, pp. 8239–8257, 2021, doi: 10.15376/biores.16.4.8239-8257. S. Svrzić, M. Đurković, A. Vukićević, Z. Nikolić, V. Mihailović, and A. Dedić, “Sound classification and power consumption relation for wood machining monitoring,” European Journal of Wood and Wood Products, vol. 82, no. 6, pp. 1953–1962, 2024, doi: 10.1007/s00107-024-02139-2. R. Igaz, R. Kminiak, K. Lubos, M. Nemec, and T. Gregel, “Temperature monitoring in CNC machining of solid wood,” Sustainability, vol. 11, no. 1, Art. no. 95, 2018, doi: 10.3390/su11010095. Y. Murase, K. Matsumoto, and T. Ohuchi, “Acoustic emission and cutting resistance of solid wood, MDF and particleboard,” Journal of the Faculty of Agriculture, Kyushu University, vol. 53, no. 2, pp. 485–490, 2008. S. Svrzić et al., “Sound signal processing and deep learning for determining circular saw blade speed,” in Proc. 6th Int. Sci. Conf. Wood Technology and Product Design, Ohrid, North Macedonia, 2023. M. Mirić-Milosavljević et al., “Signal processing and machine learning for identifying idling noises of circular saw blades,” BioResources, vol. 19, no. 1, pp. 1744–1756, 2024. Y. Sun et al., “Wood species recognition with small data using deep learning,” International Journal of Computational Intelligence Systems, vol. 14, no. 1, pp. 1451–1460, 2021, doi: 10.2991/ijcis.d.210423.001. A. R. de Geus et al., “Timber section analysis using deep learning,” Multimedia Tools and Applications, vol. 79, pp. 34513–34529, 2020, doi: 10.1007/s11042-020-09212-x. P. Kibleur et al., “Deep learning segmentation of wood fiber bundles,” Composites Science and Technology, 2022, doi: 10.1016/j.compscitech.2022.109287. A. Jegorowa et al., “Deep learning for drill wear classification in chipboard,” Wood Science and Technology, vol. 55, no. 1, pp. 1–23, 2021, doi: 10.1007/s00226-020-01245-7. V. Nasir and F. Sassani, “A review on deep learning in machining and tool monitoring,” International Journal of Advanced Manufacturing Technology, vol. 115, pp. 2683–2709, 2021. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE CVPR, 2016, pp. 770–778. M. Sandler et al., “MobileNetV2: Inverted residuals and linear bottlenecks,” in Proc. IEEE CVPR, 2018, pp. 4510–4520. C. Szegedy et al., “Rethinking the inception architecture,” in Proc. IEEE CVPR, 2016, pp. 2818–2826. F. N. Iandola et al., “SqueezeNet: AlexNet-level accuracy with fewer parameters,” arXiv:1602.07360, 2016. G. Huang et al., “Densely connected convolutional networks,” in Proc. IEEE CVPR, 2017, pp. 4700–4708. K. Simonyan and A. Zisserman, “Very deep convolutional networks,” arXiv:1409.1556, 2014. M. Tan and Q. V. Le, “EfficientNet: Rethinking model scaling,” in Proc. ICML, 2020. Z. Liu et al., “A ConvNet for the 2020s,” in Proc. IEEE CVPR, 2022. O. Russakovsky et al., “ImageNet Large Scale Visual Recognition Challenge,” IJCV, 2015. D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. ICLR, 2015. L. Cristóvão, Machining Properties of Wood, PhD thesis, Luleå University of Technology, Sweden, 2013. M. Ekevad, L. Cristóvão, and B. Marklund, “Wear of teeth of circular saw blade,” Wood Material Science & Engineering, vol. 7, pp. 150–153, 2012. V. P. Astakhov and J. P. Davim, Machining and Tool Wear. London, UK: Springer, 2008. R. Bendikiene and G. Keturakis, “Influence of tool characteristics on wear,” Journal of Wood Science, 2017. A. Naylor et al., “Mechanical cutting force model,” BioResources, vol. 7, no. 3, pp. 2883–2894, 2012. B. Porankiewicz et al., “Cutting forces in machining pine wood,” BioResources, vol. 6, no. 4, pp. 3687–3713, 2011. G. Goli et al., “Specific cutting forces in engineered wood,” Materials, vol. 11, no. 12, Art. no. 2575, 2018. G. Kovatchev and V. Atanasov, “Vibration during longitudinal milling,” Acta Facultatis Xylologiae Zvolen, vol. 63, no. 1, pp. 85–92, 2021. J. Kováč et al., “Analysis of cutting conditions in cross-cutting wood,” BioResources, vol. 16, no. 1, pp. 1029–1041, 2021. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 24 Mar, 2026 Reviewers agreed at journal 19 Mar, 2026 Reviews received at journal 18 Mar, 2026 Reviewers agreed at journal 17 Mar, 2026 Reviewers invited by journal 27 Feb, 2026 Editor assigned by journal 27 Feb, 2026 Submission checks completed at journal 26 Feb, 2026 First submitted to journal 25 Feb, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8969198","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":598098966,"identity":"781a52cf-338c-4832-b95d-a67e1c9fdd3f","order_by":0,"name":"Srđan Svrzić","email":"data:image/png;base64,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","orcid":"","institution":"University of Belgrade","correspondingAuthor":true,"prefix":"","firstName":"Srđan","middleName":"","lastName":"Svrzić","suffix":""},{"id":598098967,"identity":"21d5ca93-e9a0-47ba-8bf8-9e626978bed5","order_by":1,"name":"Mlađan Popović","email":"","orcid":"","institution":"University of Belgrade","correspondingAuthor":false,"prefix":"","firstName":"Mlađan","middleName":"","lastName":"Popović","suffix":""},{"id":598098968,"identity":"db2f2fef-f3c8-430a-ae9b-594a13ea0d54","order_by":2,"name":"Mladen Furtula","email":"","orcid":"","institution":"University of Belgrade","correspondingAuthor":false,"prefix":"","firstName":"Mladen","middleName":"","lastName":"Furtula","suffix":""},{"id":598098969,"identity":"5e282d92-1e69-4669-8eea-5cae8fa9d4cd","order_by":3,"name":"Zoran Nikolić","email":"","orcid":"","institution":"University of Belgrade","correspondingAuthor":false,"prefix":"","firstName":"Zoran","middleName":"","lastName":"Nikolić","suffix":""},{"id":598098973,"identity":"41e2d72d-6a70-465c-97db-ba5629154c66","order_by":4,"name":"Marija Đurković","email":"","orcid":"","institution":"University of Belgrade","correspondingAuthor":false,"prefix":"","firstName":"Marija","middleName":"","lastName":"Đurković","suffix":""},{"id":598098975,"identity":"c203a168-749b-49bb-be16-e65022935e75","order_by":5,"name":"Aleksandar Dedić","email":"","orcid":"","institution":"University of Belgrade","correspondingAuthor":false,"prefix":"","firstName":"Aleksandar","middleName":"","lastName":"Dedić","suffix":""},{"id":598098976,"identity":"9dd596f7-0b5b-4b50-af15-3c90db6b6a27","order_by":6,"name":"Petar Todorović","email":"","orcid":"","institution":"University of Kragujevac","correspondingAuthor":false,"prefix":"","firstName":"Petar","middleName":"","lastName":"Todorović","suffix":""},{"id":598098977,"identity":"7500ad85-4ca2-4a5f-8aee-d63fb34eebd0","order_by":7,"name":"Arso Vukićević","email":"","orcid":"","institution":"University of Kragujevac","correspondingAuthor":false,"prefix":"","firstName":"Arso","middleName":"","lastName":"Vukićević","suffix":""}],"badges":[],"createdAt":"2026-02-25 15:10:32","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8969198/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8969198/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103890568,"identity":"0a539d9a-d3d3-48ea-b2d4-7e0bc50914f6","added_by":"auto","created_at":"2026-03-04 07:58:19","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":163519,"visible":true,"origin":"","legend":"\u003cp\u003eCutting tool geometry: (B) cutting width; (b) body thickness; rake(18°) and clear angle (5°)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/bb59b6979da5ec8d1b101eba.png"},{"id":103890571,"identity":"c8156cb8-5384-4624-81be-36165a67e1ae","added_by":"auto","created_at":"2026-03-04 07:58:20","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":471223,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental flowchart illustrating: investigation of the physical and mechanical properties of particleboard (a, b, c); cutting process and machine subsystems (d, e, f); examination of saw blade dullness (g, h); data acquisition devices and hardware (i, j, k, o); acquired datasets (l, m, n); and classification of tool condition and feed rate (p, r, s)\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/60f50a4b46aa1dc2a8701cdc.png"},{"id":103890577,"identity":"6ae5a8c2-bc48-4380-bc48-3ca9fedf3614","added_by":"auto","created_at":"2026-03-04 07:58:22","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":108160,"visible":true,"origin":"","legend":"\u003cp\u003eSaw tooth condition for (a) new and (b) used saw\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/80b12a707877f285b1aa640d.png"},{"id":103890554,"identity":"f1d64935-a2a8-4143-bee4-d2b2a960075b","added_by":"auto","created_at":"2026-03-04 07:58:15","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":14725,"visible":true,"origin":"","legend":"\u003cp\u003eAverage cutting power for different feed rates and saw dullness\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/438524e30d9f419998dd3ee9.png"},{"id":103890588,"identity":"32729785-0d84-4eb8-96b5-6578acbf65aa","added_by":"auto","created_at":"2026-03-04 07:58:30","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":24977,"visible":true,"origin":"","legend":"\u003cp\u003eAverage sound intensities during cutting process\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/6ed9f0a0b0b8b47853592e53.png"},{"id":103890569,"identity":"bf4b7840-81a0-4bba-a276-ad20a2931d3c","added_by":"auto","created_at":"2026-03-04 07:58:20","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":165163,"visible":true,"origin":"","legend":"\u003cp\u003eScreenshot of thermal imaging camera FLIR E50\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/6120f7d741c26d6a41ae2c02.png"},{"id":103890570,"identity":"7468810f-594d-4ca0-983e-0ab96500287f","added_by":"auto","created_at":"2026-03-04 07:58:20","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":115508,"visible":true,"origin":"","legend":"\u003cp\u003eAverage FFT values (a) evenly highlighted for all measurements; (b) only highlighted for new saw at feed rate u = 10 m/min\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/9d4001de9250cdef3f5e50e4.png"},{"id":103890566,"identity":"7a955693-0ea4-4307-8db0-4f28a1d7cec9","added_by":"auto","created_at":"2026-03-04 07:58:19","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":851765,"visible":true,"origin":"","legend":"\u003cp\u003eSpectrogram example for new U10 (a) 2D and (b) 3D presentation\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/1255e778a8af9a4cc8396972.png"},{"id":103890573,"identity":"3728185b-75de-4b6f-81c2-876758ced635","added_by":"auto","created_at":"2026-03-04 07:58:20","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":1043372,"visible":true,"origin":"","legend":"\u003cp\u003ePrepared spectrograms as input data for CNN (a) new U10; (b) new U15; (c) used U10 and (d) used U15\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/6b43aa6db66dd621ab750070.png"},{"id":103890579,"identity":"1b1eaf83-0093-45ec-9c12-ef09a520988c","added_by":"auto","created_at":"2026-03-04 07:58:23","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":109257,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrices for the considered predictive models.\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/17c038a6b4cb06fb4a0592ba.png"},{"id":103890607,"identity":"9be3546b-518c-439c-9140-68ecba0ea95e","added_by":"auto","created_at":"2026-03-04 07:58:37","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4471313,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8969198/v1/8fcb45bf-c01a-4dde-aa22-fed26a6cc0ec.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Acoustic Signal-Based Cutting Tool Condition Monitoring in Woodworking Using Deep Learning","fulltext":[{"header":"1.0 Introduction","content":"\u003cp\u003eThe global wood industry presents a dynamic and ever-growing field valued at approximately 1 trillion USD, while consuming more than 4\u0026nbsp;billion cubic meters of wood according to the F.A.O. report [1]. It is estimated that its gross growth will continue by 25% until 2030. The major bottleneck of the wood industry sector is the low total utilization, which is estimated to be approximately 30\u0026ndash;55% for solid wood products [2] and 71\u0026ndash;98% for wood-based panels (MDF and particleboards) applications. At the same time, the global woodworking machinery market size is valued at \u003cspan\u003e$\u003c/span\u003e5.03\u0026nbsp;billion in 2024 and is projected to grow to \u003cspan\u003e$\u003c/span\u003e7.88\u0026nbsp;billion by 2032 [3].\u003c/p\u003e \u003cp\u003eProper maintenance and appropriate operation of wood processing machine tools are critical for cost and material loss reduction [4,5]. Poor organization and management of tooling can lead to an increase in hidden costs, resulting in productivity losses of approximately 20\u0026ndash;30% [6,7] of working time and additional tool-related cost increases of 10\u0026ndash;30% [8]. Circular saw blades used in wood panel processing are exposed to abrasive constituents, adhesives, and coatings, resulting in progressive loss of sharpness that is difficult to detect without interrupting production. Consequently, there is a clear need for non-invasive, robust, and information-rich monitoring approaches capable of operating in real production environments and providing reliable indicators of tool condition.\u003c/p\u003e \u003cp\u003eParticleboard was selected as the focus of this study due to its extensive use in furniture manufacturing and panel processing, as well as its relatively uniform internal structure when produced under standardized industrial conditions. Under controlled manufacturing parameters, such as adhesive formulation, wood species composition, and pressing regime, particleboard exhibits stable mechanical and physical properties [9]. While density gradients across the board thickness may influence local stiffness and modulus of elasticity [10], these variations are systematic and reproducible, making particleboard a suitable material for controlled machining studies. In industrial practice, particleboard is predominantly processed by circular sawing, where tool wear has an immediate and measurable impact on cutting power, surface integrity, and process stability.\u003c/p\u003e \u003cp\u003ePrevious research on particleboard machining has largely focused on tool geometry and cutting parameters, including rake angle, clearance angle, bevel angle, cutting speed, feed rate, and tooth configuration [11,12]. These studies have demonstrated strong links between tool geometry, cutting forces, airborne dust emission, and surface quality. However, they typically rely on conventional monitoring variables, such as cutting forces, vibration signals, or power consumption. While cutting power has proven to be a robust and industrially accessible indicator of process load [13,14,15,16,17], it represents an aggregated response of the system and may fail to capture localized or transient phenomena associated with early-stage tool wear. Tool temperature, although investigated as a monitoring variable [13,18], often lacks sufficient sensitivity under moderate cutting conditions and stable regimes.\u003c/p\u003e \u003cp\u003eAcoustic signals generated during wood machining constitute a non-contact and information-rich process output that has received comparatively limited attention. Cutting sound originates from tool\u0026ndash;material interaction, chip formation, and frictional effects, and is inherently sensitive to high-frequency and transient events. Early work on acoustic emission in machining demonstrated its potential for process monitoring [19], yet its application in woodworking has remained largely exploratory. Recent studies have shown that, with appropriate signal processing, sound can be used to identify cutting regimes, tool types, and even wood species [17,20,21]. However, these investigations did not directly address tool condition monitoring nor systematically evaluate the robustness of acoustic features under closely related operating conditions.\u003c/p\u003e \u003cp\u003eThe increasing adoption of artificial intelligence in manufacturing offers new opportunities to exploit such complex signals. Deep learning methods, particularly convolutional neural networks (CNNs), have demonstrated exceptional performance in extracting discriminative patterns from high-dimensional data. In wood science, CNNs have been successfully applied to wood species identification [22,23], fiber segmentation from X-ray CT images [24], and indirect assessment of tool wear through image-based analysis of drilled holes [25]. Comprehensive reviews by [26] further emphasize the potential of deep learning architectures\u0026mdash;including CNNs, autoencoders, and recurrent networks\u0026mdash;for intelligent machining and tool monitoring. Nevertheless, the application of deep learning to airborne acoustic signals for direct tool condition monitoring in circular sawing of wood-based panels has not yet been systematically investigated.\u003c/p\u003e \u003cp\u003eBuilding on these findings, the present study explores the use of acoustic signals as the primary input for a deep learning-based tool condition monitoring during particleboard cutting. Acoustic data are transformed into time\u0026ndash;frequency representations and classified using a representative set of modern CNN architectures. Cutting power consumption and tool temperature are simultaneously analyzed as established reference variables to benchmark the performance and reliability of the acoustic approach. The engineering motivation lies in demonstrating that non-contact acoustic monitoring can provide sensitivity comparable to, or exceeding, that of conventional monitoring proxies, while offering superior responsiveness to transient and high-frequency phenomena.\u003c/p\u003e \u003cp\u003eThe novelty and contribution of this work lie in: (i) systematically converting cutting sound into spectrogram-based representations suitable for CNN analysis, (ii) benchmarking multiple state-of-the-art CNN architectures under consistent experimental conditions, and (iii) evaluating the practical implications of acoustic-based monitoring for tool-condition assessment, maintenance planning, and process reliability in industrial woodworking environments.\u003c/p\u003e"},{"header":"2.0 Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n\u003ch2\u003e2.1 Materials\u003c/h2\u003e\n\u003cp\u003eThe experiments were conducted using melamine-faced particleboard, designated as type P2 in accordance with the European standard EN 312:2010. typically employed in the manufacture of furniture. The particleboards, with a nominal thickness of 18 mm, were sampled from the same stock at a domestic wholesaler company (Belgrade, Serbia). These boards were manufactured by Egger Group GmbH in a factory located in Răducăeni, Romania.\u003c/p\u003e\n\u003cp\u003eThe particleboard samples are divided into four groups according to specific machine cutting variables in the main experiment. These variables are as follows: the state of the saw (new, used) and the feed rate (10 and 15 m/min). Each sample group, consisting of three panels, is designated as follows:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eUsed U10 - used saw; feed rate: 10 m/min,\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eUsed U15 - used saw; feed rate: 15 m/min,\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eNew N10 - new saw; feed rate: 10 m/min,\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eNew N15 - new saw; feed rate: 15 m/min.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eBasic physical and mechanical tests were performed to determine both the conformity of particleboard samples with the requirements of the EN 312 standard, as well as to examine the variations between the sample groups. In that aspect, the properties of particleboards, such as: density (EN 323, 1993), moisture content (EN 322, 1993), bending strength, modulus of elasticity (EN 310, 1993) and internal bond (EN 319, 1993) are determined (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e(a) to (c)) and presented in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab1\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eStatistical comparison of basic physical and mechanical properties of sample groups at the significance level of \u0026alpha;\u0026thinsp;=\u0026thinsp;0.05\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth rowspan=\"2\" align=\"left\"\u003e\n\u003cp\u003eProperty\u003c/p\u003e\n\u003c/th\u003e\n\u003cth rowspan=\"2\" align=\"left\"\u003e\n\u003cp\u003eUnit\u003c/p\u003e\n\u003c/th\u003e\n\u003cth colspan=\"4\" align=\"left\"\u003e\n\u003cp\u003eSample group (mean value\u0026thinsp;\u0026plusmn;\u0026thinsp;st. dev.)\u003c/p\u003e\n\u003c/th\u003e\n\u003cth colspan=\"3\" align=\"left\"\u003e\n\u003cp\u003eANOVA statistics\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eUsed u10\u003c/strong\u003e\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eUsed u15\u003c/strong\u003e\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eNew n10\u003c/strong\u003e\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003e\u003cstrong\u003eNew n15\u003c/strong\u003e\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eF\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eP-value\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eF crit.\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eThickness\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003emm\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e18.19\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e18.18\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.007\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e18.18\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.026\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e18.18\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.015\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.30759\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.81968\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2.94669\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDensity\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ekg/m\u0026sup3;\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e658.8\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;26.31\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e647.9\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;20.2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e667.7\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;9.89\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e658.3\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;7.98\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1.66113\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.19796\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2.94669\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMoisture content\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e%\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e7.13\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.093\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e7.11\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.127\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e7.07\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.078\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e7.10\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.082\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.29887\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.82559\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e3.49030\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eBending strength\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eN/mm\u0026sup2;\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e17.20\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;2.364\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e18.64\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;1.237\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e17.94\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.495\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e17.19\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;1.434\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e1.21845\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.32887\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e3.09839\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMOE\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eN/mm\u0026sup2;\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e2623\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;79.1\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e2910\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;26.2\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e2666\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;177.4\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e2743\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;169.7\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e5.69749\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.00548 *\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e3.09839\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eInternal bond\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eN/mm\u0026sup2;\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.431\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.065\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.417\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.045\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.449\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.045\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.428\u003c/p\u003e\n\u003cp\u003e\u0026plusmn;\u0026thinsp;0.056\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.52531\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e0.66849\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"char\" char=\".\"\u003e\n\u003cp\u003e2.94669\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003ctfoot\u003e\n\u003ctr\u003e\n\u003ctd colspan=\"9\"\u003e* Significant difference\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tfoot\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n\u003ch2\u003e2.2 Cutting Tools and Conditions\u003c/h2\u003e\n\u003cp\u003eThe particleboard panels were cut using two FREUD LU2C 1200 circular saw blades, with characteristics presented at Figs.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eh and Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. The first piece was evaluated in its original condition, whereas the second piece had been utilized for approximately 700 meters with similar cutting conditions.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003ctable id=\"Tab2\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eTool and cutting conditions\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eCutting tool\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eLU2C 1200\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eRPM\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e4000 min\u003csup\u003e\u0026minus;\u0026thinsp;1\u003c/sup\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFeed per tooth\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0.03125 mm\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eTooth shape\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eATB\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eNumber of teeth\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e80\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDiameter (mm)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e250\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eBody thickness b (mm)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e2.2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eCutting width B (mm)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e3.2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eRake angle (\u0026deg;)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e18\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eClear angle (\u0026deg;)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e5\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eInclination angle (\u0026deg;)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e10\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eTool override (mm)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e10\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eThe tool condition in terms of saw tooth dullness was estimated according to photos made by microscope (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eg). The photos taken under the microscope should illustrate a brand new, unpackaged circular saw blade and the one of the same type that has already been used. It was expected that the tooth edges and surfaces for the new saw are in perfect condition, while the teeth image of a used tool should have traces of use in the form of different deviations of cutting edges, cracks, and material deposits on cutting surfaces.\u003c/p\u003e\n\u003cp\u003eThroughout the course of each experimental stage, the thermal imaging camera was maintained in a state of recording (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ej). The maximum temperature values for each cut were duly noted and meticulously documented. In accordance with the findings of the research, the tool temperature results were subjected to analysis using the ANOVA statistic.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n\u003ch2\u003e2.3 Experimental Setup\u003c/h2\u003e\n\u003cp\u003eThe study was conducted in the Laboratory of Machines and Apparatus at the Faculty of Forestry, University of Belgrade (Beograd, Serbia). The machining system used for this study was a Minimax CU 410K combined machine (SCM, Rimini, Italy) equipped with a 3 kW three-phase asynchronous motor (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ed). The speed of the motor was set by a customized frequency controller (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ef) to 4000 rpm with a corresponding frequency of 50.5 Hz. The noise occurring in the cutting process was recorded, in WAV format (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003en) using a dbx RTA-M measurement microphone with an electret condenser on the back (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ee). The RTA-M is an omnidirectional, low-profile frequency measurement microphone specifically designed to record all frequencies from 20 Hz to 20 kHz, ensuring accurate \"real-time \"pinging\" analysis of the audio signal. It is operated with phantom power. To reduce the effects of vibration, the microphone is housed in a vibration-damping rack. The Focusrite Scarlet SOLO USB audio interface was connected to a PC (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eo).\u003c/p\u003e\n\u003cp\u003eThe cutting power measuring device (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ei) has a Circutor CW-TAN active power transformer for unbalanced three-phase systems with the following characteristics: alternating current 5 A, alternating voltage 230 V, frequency 50 Hz, accuracy 0.5% and analog voltage output 0\u0026ndash;10. The possible measuring ranges are 5, 10 and 15 kW. The measured cutting power data are given in Watts. The operator selects the expected range for a better resolution of the results.\u003c/p\u003e\n\u003cp\u003eThe Dino-Lite Edge USB microscope with 470x magnification was used for visual observation of the tool condition (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eg).\u003c/p\u003e\n\u003cp\u003eThe temperature of the saw teeth was measured using a FLIR E50 thermal imaging camera (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ej). The recordings were in the form of movie clips which were later processed and statistically analyzed (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003em).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n\u003ch2\u003e2.4. Signal Acquisition and Processing\u003c/h2\u003e\n\u003cp\u003eAudacity, a cross-platform open-source audio software, was used to record and process the audio signals. The measurements were carried out at a sampling rate of 44100 Hz. The sounds generated in the cutting process were captured with the microphone and recorded on the PC as wave files. Spectral analysis was performed on these recordings using the fast Fourier transform (FFT) and the short-time Fourier transform (STFT) provided by MATLAB R2023 Edition (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eo). On these recordings, spectral analysis was performed using FFT and STFT (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003es). Spectrograms are 3D (frequency-time-power) charts obtained by STFT of original sound signals (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ep).\u003c/p\u003e\n\u003cp\u003eThe entire system for cutting power data acquisition is based on the Power Expert software platform, and the sampling rate has been set to 1000 Hz (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ek). The spectrograms further served as training data for the deep learning networks (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003er) considered in this study.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n\u003ch2\u003e2.5 Deep Learning Architectures and Training Protocol\u003c/h2\u003e\n\u003cp\u003eThe eight convolutional neural network (CNN) architectures were considered to form a representative and engineering-relevant benchmark for spectrogram image classification. ResNet introduces residual or \u0026ldquo;skip\u0026rdquo; connections that allow gradients to propagate more effectively through deep networks, mitigating the vanishing gradient problem. This design enables the training of very deep models, which is beneficial for capturing subtle spectral variations in acoustic data that may indicate tool wear or cutting conditions [27]. MobileNetV2 relies on depthwise separable convolutions and inverted residuals, significantly reducing computational complexity and memory requirements [28]. While lighter than most CNNs, it still maintains strong accuracy. This makes it particularly attractive for deployment in edge AI scenarios, where low-power, real-time monitoring systems are required. Inception V3 employs inception modules that process multiple filter sizes in parallel, capturing both fine and coarse features within the same layer [29]. This ability to extract multi-scale representations is advantageous for spectrograms, where patterns may occur at different frequencies and time windows simultaneously. SqueezeNet was designed with parameter reduction in mind, using fire modules that \u0026ldquo;squeeze\u0026rdquo; with 1\u0026times;1 convolutions before \u0026ldquo;expanding\u0026rdquo; into a mix of 1\u0026times;1 and 3\u0026times;3 filters [30]. This results in a model under 5 MB in size, yet capable of competitive accuracy. For industrial monitoring, this compactness allows easier integration into embedded systems and sensor nodes. DenseNet establishes direct connections between each layer and all subsequent layers, promoting feature reuse and improving gradient flow [31]. This architecture can capture both local and global acoustic structures efficiently, which is useful when subtle harmonic patterns must be linked with broader spectral trends. VGG19 represents a classical deep CNN, characterized by its simple yet powerful architecture of stacked 3\u0026times;3 convolutions followed by fully connected layers [32]. Although computationally heavier, VGG19 provides a strong reference baseline and facilitates comparison with newer, more efficient architectures. EfficientNet-B0 was added as a modern CNN with a principled balance of accuracy and computational cost via compound scaling [33]. Finally, ConvNeXt represents a modernized CNN design, assessing whether recent convolutional improvements translate into measurable gains for the considered challenge [34]. Overall, the benchmark set was intentionally constructed to cover a spectrum from ultra-compact (SqueezeNet) through mobile-efficient (MobileNetV2), canonical high-accuracy (ResNet/DenseNet), multi-scale feature extraction (Inception V3), classic deep baseline (VGG19), and modern accuracy/efficiency (EfficientNet-B0, ConvNeXt), enabling a balanced evaluation in terms of accuracy, efficiency, and practical industrial applicability.\u003c/p\u003e\n\u003cp\u003eAll classification models were pre-trained on the ImageNet [35] dataset and fine-tuned using the PyTorch framework. The training was performed using the Adam optimization algorithm, [36] with the cross-entropy loss function and the initial learning rate of 1e-4 (which was decreased by a factor of 0.1 every 7 epochs). All the computations were performed on the Lambda workstation with the AMD Threadripper 3970X (32 cores, 3.79 GHz processor), 128 GB RAM, and two Titan RTX (24 GB) + NVLink GPUs.\u003c/p\u003e\n\u003cp\u003eFrom the four sets of 250 measurements for each class (one set per class), 200 samples per class were randomly selected for artificial neural network (CNN) training, while the remaining were randomly split for the purpose of validation and testing (25 per class). To improve robustness and reduce overfitting, we applied online augmentation for the training data set tailored to time\u0026ndash;frequency representations. Specifically, we applied: time shift/random temporal cropping, SpecAugment masking (time and frequency masking), and additive Gaussian noise (waveform or spectrogram). When augmenting at the waveform level, we additionally used a small time-stretch and pitch shift. Augmentations were applied only during training with a fixed probability (not on validation/test).\u003c/p\u003e\n\u003cp\u003eTo quantify performance variability, we repeated training five times with different random seeds while keeping the test set fixed (25 samples per class). We computed mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation of metrics considered for the algorithms evaluation were: Accuracy quantifies the proportion of correct predictions across all classes, expressed as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\frac{TP+TN}{TP+TN+FP+FN}\\)\u003c/span\u003e\u003c/span\u003e. F1 \u0026ndash; Score (Macro Average) computes the unweighted arithmetic mean of class-specific F1-scores, defined as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\frac{1}{C}{\\sum}_{i=1}^{C}=2\\frac{Precision\\bullet Recall}{Precision+Recall}\\)\u003c/span\u003e\u003c/span\u003e, where \u003cem\u003eC\u003c/em\u003e is the number of classes. This metric equally weights all classes, making it suitable for scenarios where minority class performance is critical; Cohen\u0026rsquo;s Kappa measures inter-rater agreement corrected for random chance: \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\kappa=\\left({p}_{0}-{p}_{\\epsilon}\\right)/\\left(1-{p}_{\\epsilon}\\right)\\)\u003c/span\u003e\u003c/span\u003e, where \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003e0\u003c/em\u003e\u003c/sub\u003e is observed accuracy and \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003e\u0026epsilon;\u003c/em\u003e\u003c/sub\u003e is expected chance agreement. Values are interpreted as: \u0026le;0 (no agreement), 0.01\u0026ndash;0.20 (slight), 0.21\u0026ndash;0.60 (moderate), 0.61\u0026ndash;0.80 (substantial), 0.81-1.00 (near-perfect). MCC (Matthews Correlation Coefficient) \u0026mdash; a unique measure of the quality of predictions (correlation true vs. predicted), robust to class imbalance. MAE (Mean Absolute Error) \u0026mdash; the average \"distance\" between the real and predicted class; useful when the classes have an order (ordinal).\u003c/p\u003e\n\u003c/div\u003e"},{"header":"3.0 Results and Discussion","content":"\u003cp\u003eThe results of this study are presented in the form of visual aids, including photographs, screen captures, tables and diagrams. As illustrated in Figs. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e (a) and (b), the photographic evidence provides a rudimentary understanding of the sawtooth condition. It is possible to observe the deterioration of the tooth edges and related surfaces of a used saw. As demonstrated in Fig. 4, an investigation into cutting power measurements revealed the potential to generate a diagram illustrating the average cutting power for varying feed rates and saw conditions.\u003c/p\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 Cutting power\u003c/h2\u003e\n \u003cp\u003eThe cutting power curves are presented as a mean value for each measurement, calculated at the precise moment of each cut. It was observed that the highest mean cutting power was exhibited by U15, followed by N15, N10 and finally U10. It was hypothesized that the mean cutting power would increase both with machining length and feed rate [37,38].\u003c/p\u003e\n \u003cp\u003eAs posited by [39], intensive tool wear, as indicated by flank wear width, manifests at a distance of approximately 550 meters. However, as asserted by [40], this wear may occur much earlier, at around 200 meters of machining distance. Thereafter, there is a consistent deterioration until 2750 meters. One potential explanation for this phenomenon is that, during the initial phases of tool wear, the process is purely cutting, whereas in later phases, it is more akin to the breaking off of processed particleboard. In the context of a pure cutting process, an increase in friction between tooth surfaces and the material is observed. This rise in friction results in an enhancement of the cutting power. In addition, it is conceivable that the initial stages may witness a decline in kerf width due to pronounced tool wear, particularly along the lateral edges. This phenomenon is attributed to the diminished friction, consequently leading to a reduction in power consumption during the latter stages of operation. This is attributable to the more stable and uniform tool wear characteristics that characterize these latter stages. Nevertheless, an apparent discrepancy in the consumed power is evident for both the new and used saws when an increased feed rate is employed.\u003c/p\u003e\n \u003cp\u003eStatistical analysis of the cutting power achieved during the experiment showed significant influence of both the saw condition and the feed rate (Tables \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e). The maximal cutting power was calculated by subtracting the idle machine power from the total cutting power. Cutting power increased with the feed rate, regardless of the tool condition. However, the tool condition performed differently in regard to the feed rate. The new saw achieved lower maximal cutting power during higher feed rate, than the used one. The results were opposite for the lower feed rate.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eANOVA statistics of maximal cutting power values, at the significance level of \u0026alpha;\u0026thinsp;=\u0026thinsp;0.05\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSample group\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eCount\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003cp\u003e(W)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eStandard deviation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" align=\"left\"\u003e\n \u003cp\u003eANOVA statistics\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eF crit\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e650.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e30.988\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" align=\"left\"\u003e\n \u003cp\u003e124.389\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" align=\"left\"\u003e\n \u003cp\u003e5.93\u0026middot;10\u003csup\u003e\u0026minus;\u0026thinsp;33\u003c/sup\u003e *\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" align=\"left\"\u003e\n \u003cp\u003e2.69939\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e710.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e38.432\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eU10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e587.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e28.358\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eU15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e747.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e27.151\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\"\u003e* Significant difference\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eThe statistical analysis of maximal cutting power values between the single pairs of sample groups (Bonferroni test)\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSample group pair\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ep-value (T-test)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSignificant difference\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eBonferroni correction (adjusted \u0026alpha;)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN10 vs N15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2.0099 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;07\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"6\" align=\"left\"\u003e\n \u003cp\u003e0.00833\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN10 vs U10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.2643 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;09\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN10 vs U15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7.5126 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;16\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN15 vs U10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.6171 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;17\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN15 vs U15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.000212346\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eU10 vs U15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2.4164 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;25\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Temperature\u003c/h2\u003e\n \u003cp\u003eThe temperature of the saw blade was the focus of continuous monitoring during each cutting run, and the peak temperature values were subjected to statistical comparison across the sample groups. The results of the analysis of variance (ANOVA) indicated a statistically significant difference among the groups, thereby confirming that one or both of the experimental variables (saw condition and/or feed rate) affected the temperature buildup (see Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e). In order to evaluate the impact of specific experimental variables, a post-hoc Bonferroni test was used to compare the individual pairs of sample groups (see Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e). The findings indicated that the saw condition (new vs. used) exerted a substantial influence on the saw blade temperature. Statistical analysis revealed that both the N10 and N15 sample groups exhibited mean values of saw temperature that were significantly higher than those observed in the U10 and U15 samples. However, the saw temperature remained unaffected by the feed rate, as evidenced by the lack of statistical significance in temperature variations between 10 and 15 meters per minute for both new and used saws.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab5\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAnalysis of variance (ANOVA) of maximum saw temperature values between the sample groups, at the significance level of \u0026alpha;\u0026thinsp;=\u0026thinsp;0.05\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSample group\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eCount\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003cp\u003e(\u0026deg;C)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" align=\"left\"\u003e\n \u003cp\u003eStandard deviation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" align=\"left\"\u003e\n \u003cp\u003eANOVA statistics\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eF crit.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e61.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10.307\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" align=\"left\"\u003e\n \u003cp\u003e97.09043\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" align=\"left\"\u003e\n \u003cp\u003e2.499\u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;44\u003c/sup\u003e *\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" align=\"left\"\u003e\n \u003cp\u003e2.63373\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e57.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5.008\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eU10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e47.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6.302\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eU15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e45.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5.362\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\"\u003e* Significant difference\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tab6\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eThe statistical analysis of maximum saw temperature values between the single pairs of sample groups (Bonferroni test)\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSample group pair\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ep-value (T-test)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSignificant difference\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eBonferroni correction (adjusted \u0026alpha;)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN10 vs N15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.011631\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"6\" align=\"char\"\u003e\n \u003cp\u003e0.00833\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN10 vs U10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e8.9690 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;20\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN10 vs U15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2.0668 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;24\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN15 vs U10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2.1804 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;23\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN15 vs U15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e5.9649 \u0026middot; 10\u003csup\u003e\u0026minus;\u0026thinsp;32\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eU10 vs U15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.072467\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3 Sound signal\u003c/h2\u003e\n \u003cp\u003eAs demonstrated in Fig.\u0026nbsp;5, a comparison of average sound intensity levels was facilitated through the analysis of noise recordings in the context of the cutting time domain. The mean integral of all intensities at all frequencies is the average sound intensity for a single cutting run. As was the case in the preceding instance, the maximum values for sound intensity were observed for U15, with N15 following in second place. However, these values were only recorded during the final two seconds of the run. The sound intensity curves for N10 and U10 demonstrate a high degree of similarity. As was the case previously, it is evident that there is a certain discrepancy in the sound intensity produced by the tool in relation to the varying feed rates. However, no such variation is observed in the tool\u0026apos;s dullness.\u003c/p\u003e\n \u003cp\u003eIt is evident that the sound signal is significantly impacted by its frequency domain; therefore, the calculation of the mean FFT values was deemed appropriate. The FFT was conducted at a computing frequency of 100 Hz. Subsequent to this, the mean value for all frequencies was calculated. The results pertaining to the mean FFT values for all stages of measurement are illustrated in Figs.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e(a) and (b).\u003c/p\u003e\n \u003cp\u003eAs illustrated in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e, it is evident that the FFT graphs for all four experimental stages demonstrate a similar pattern of behavior. The characteristic peaks manifest at equivalent frequencies, exhibiting solely disparate levels of intensity. As the graph of the FFT for N10 is not clearly visible in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003ea, Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003eb provides a more comprehensive overview of its performance. It was also evident that the trend previously identified in the context of cutting power and average sound intensity measurements was also apparent in the case of average FFT graphs. The highest values exhibited were observed for U15, followed by N10 and U10. It has been established that the dominant peak for all average FFT graphs occurs precisely at 5200 Hz. This is consistent with the natural frequency of the saw. It is evident that other peaks manifest as higher harmonics of the fundamental frequency. The graphs are distinguished solely by the intensity of their peaks.\u003c/p\u003e\n \u003cp\u003eAs demonstrated in Figs.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e and \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e, the graph was unable to achieve the requisite level of accuracy for process recognition. Consequently, the sound recordings for all 1,000 measurements in the WAV format were subjected to STFT with a Hann window with 50% overlapping, smoothing the spectral line considerably and also giving frequency-time-power domain of the sound signal. As demonstrated in Figs.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e (a) and (b), the STFT provided a graphical representation of sound signals in the form of spectrograms.\u003c/p\u003e\n \u003cp\u003eAs demostrated in Figs.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e9\u003c/span\u003e (a)\u0026ndash;(d), spectrogram representations of the acoustic signals were employed as the primary input data for the convolutional neural network (CNN). From each of the four datasets comprising 250 measurements, 200 samples were randomly selected for network training, while the remaining 50 samples were reserved for testing purposes. Several state-of-the-art CNN architectures were evaluated in this study, including MobileNetV2, ResNet, Inception V3, SqueezeNet, DenseNet, VGG19, EfficientNet-B7\u0026mdash;selected for its balance between accuracy and computational efficiency\u0026mdash;and ConvNeXt-XL, which represents a modern CNN design paradigm [34]. The classification performance achieved by these architectures is summarized in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e and forms the basis for the following discussion, which focuses on the comparative effectiveness of the applied models and their suitability for sound-based monitoring of the wood machining process.\u0026nbsp;\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab7\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eCNN metrics (performance of the developed deep learning models for machining sound classification\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF1 (Macro)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMCC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMAE\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCohen\u0026apos;s Kappa\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMobileNetV2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e81.00\u0026thinsp;\u0026plusmn;\u0026thinsp;3.92%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.810\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.747\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.270\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.747\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eResNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.00%\u0026plusmn;2.86%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.911\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.881\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.120\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.880\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInception_v3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e80.00%\u0026plusmn;4.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.801\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.734\u0026thinsp;\u0026plusmn;\u0026thinsp;0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.270\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.733\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSqueezeNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e73.00%\u0026plusmn;4.44%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.730\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.641\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.380\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.640\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDenseNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e82.00%\u0026plusmn;3.84%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.821\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.761\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.220\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.760\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVGG19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e84.00%\u0026plusmn;3.67%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.840\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.787\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.260\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.787\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEfficientNet-B7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.00%\u0026plusmn;3.13%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.890\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.854\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.140\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.853\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eConvNeXt-XL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.00%\u0026plusmn;2.71%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.920\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.893\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.110\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.893\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eThe results shown in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e indicate a clear hierarchy of performance among the analyzed CNN architectures in the task of multi-class audio processing classification. ConvNeXt-XL achieved the best overall performance, with the highest accuracy (92.00%), the highest Macro F1 (0.920), the highest MCC (0.893), and the lowest MAE (0.110), indicating high discriminative power and stability in the presence of class imbalance. The high value of Cohen\u0026apos;s \u0026kappa; (0.893) confirms the strong agreement of the model with the benchmarks, significantly above the level of random guess.\u003c/p\u003e\n \u003cp\u003eResNet shows very competitive results (Accuracy\u0026thinsp;=\u0026thinsp;91.00%, Macro F1\u0026thinsp;=\u0026thinsp;0.911, MCC\u0026thinsp;=\u0026thinsp;0.881, \u0026kappa;\u0026thinsp;=\u0026thinsp;0.880), confirming that residual connections effectively alleviate the gradient degradation problem and enable stable learning of deep representations. Although slightly inferior to ConvNeXt-XL, ResNet maintains an excellent balance between accuracy and robustness, with a low MAE (0.120).\u003c/p\u003e\n \u003cp\u003eEfficientNet-B7 achieves solid performance (Accuracy\u0026thinsp;=\u0026thinsp;89.00%, \u0026kappa;\u0026thinsp;=\u0026thinsp;0.853), demonstrating the good efficiency of the architecture based on scaling depth, width and resolution. However, compared to ResNet and ConvNeXt-XL, a slight decrease in robustness is observed (lower MCC and higher MAE), suggesting a greater sensitivity to confusion between classes.\u003c/p\u003e\n \u003cp\u003eDenseNet (Accuracy\u0026thinsp;=\u0026thinsp;82.00%, \u0026kappa;\u0026thinsp;=\u0026thinsp;0.760) shows correct but moderate results. Although the feature reuse mechanism contributes to learning stability, the performance indicates a limited ability to separate classes compared to more modern architectures.\u003c/p\u003e\n \u003cp\u003eMobileNetV2 achieves 81.00% accuracy with \u0026kappa;\u0026thinsp;=\u0026thinsp;0.747, which confirms good generalization considering the low complexity of the model. However, in this experimental environment its efficiency advantage does not come with performance at the level of the leading models, indicating a trade-off between parametric economy and discriminative power.\u003c/p\u003e\n \u003cp\u003eModels VGG19, Inception_v3 and SqueezeNet make up the bottom rank in terms of performance. SqueezeNet is the weakest (Accuracy\u0026thinsp;=\u0026thinsp;73.00%, Macro F1\u0026thinsp;=\u0026thinsp;0.730, \u0026kappa;\u0026thinsp;=\u0026thinsp;0.640), which clearly indicates the limited representative power of the extremely lightweight architecture. VGG19 (Accuracy\u0026thinsp;=\u0026thinsp;84.00%, \u0026kappa;\u0026thinsp;=\u0026thinsp;0.787) and Inception_v3 (Accuracy\u0026thinsp;=\u0026thinsp;80.00%, \u0026kappa;\u0026thinsp;=\u0026thinsp;0.733) show a pronounced sensitivity to confusion between classes, which is reflected in lower MCC values and higher MAE, especially for borderline samples.\u003c/p\u003e\n \u003cp\u003eOverall, the results confirm that the more modern architectures (ConvNeXt-XL and ResNet) provide the best compromise between classification accuracy, robustness and stability. Differences between models arise not only from overall accuracy, but also from the ability to reliably separate classes, which is clearly quantified through MCC, Macro F1 and Cohen\u0026apos;s \u0026kappa;. These findings indicate that architectural innovations play a crucial role in the quality of latent representations in the processing sound classification tasks.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eDiscussion of the obtained results with respect to previous studies\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eTool condition monitoring in machining processes has been extensively studied using a variety of sensing modalities. Traditional approaches have relied primarily on measurements of cutting force, vibration, power consumption, and tool temperature. For instance, [41] and [42] investigated cutting force as a reliable indicator of tool wear, while [43] expanded this line of research to engineered wood products. However, force measurements often require invasive instrumentation that is difficult to integrate into industrial environments. Vibration-based methods have also been widely explored [12,44], but their sensitivity to machine dynamics and environmental noise may limit their robustness in factory settings.\u003c/p\u003e\n \u003cp\u003eAnother established line of research concerns acoustic emission monitoring, which has demonstrated potential in capturing high-frequency signals directly linked to cutting phenomena [19,13]. Studies by [14] and [15] confirmed that sound analysis can support adaptive control in wood machining, while [16,20] and [21] showed that acoustic features correlate with tool wear and sawing conditions. Parallel studies investigated power consumption [45] and tool temperature [18,40,13] as indirect indicators of machining states. While these variables provide valuable information, they often fail to distinguish subtle differences in tool condition, especially at early wear stages.\u003c/p\u003e\n \u003cp\u003eCompared with these approaches, the present study highlights several novel contributions. First, it focuses on acoustic signal classification using deep neural networks, rather than conventional statistical or shallow learning techniques. Previous works in wood machining often relied on handcrafted features or limited classifiers, whereas this research employs advanced CNN architectures (ResNet, MobileNetV2, DenseNet, etc.) that automatically extract discriminative features from spectrograms. Second, the results demonstrate superior classification accuracy, with ResNet achieving 96.2%, which is higher than most accuracies reported in prior woodworking signal studies that typically ranged between 80\u0026ndash;90% [25,26].\u003c/p\u003e\n \u003cp\u003eA third contribution lies in the multi-sensor framework, where acoustic monitoring was complemented with power and temperature measurements. Although these auxiliary variables did not yield sufficient accuracy independently, their combined analysis provided valuable context for validating acoustic-based predictions. This integration reflects a step toward holistic monitoring systems suitable for Industry 4.0 manufacturing environments.\u003c/p\u003e\n \u003cp\u003eFinally, the study emphasizes practical applicability. By evaluating lightweight architectures such as MobileNetV2 and SqueezeNet, the work demonstrates the feasibility of deploying acoustic-based monitoring on edge devices for real-time process control. This is an important distinction from previous studies that primarily validated methods in controlled laboratory environments without considering industrial integration.\u003c/p\u003e\n \u003cp\u003eIn the context of Industry 4.0 and smart manufacturing, the proposed approach provides distinct advantages over previous monitoring strategies. Earlier studies based on cutting force [41,42], vibration [12,44], or power and temperature signals [13,18,45] demonstrated useful insights but generally required invasive sensors, complex instrumentation, or lacked the robustness needed for integration in industrial environments. In contrast, acoustic sensing is low-cost, non-invasive, and easily deployable, which makes it well-suited for large-scale industrial applications. By leveraging deep learning models such as ResNet and MobileNetV2, this study further ensures high classification accuracy with architectures that are also computationally efficient, thus opening the possibility of real-time edge deployment in production lines. This combination of acoustic monitoring and lightweight CNNs reflects a shift from purely experimental validation toward practical, scalable solutions for predictive maintenance, process optimization, and adaptive control in woodworking and beyond. Hence, the study not only advances academic understanding but also aligns closely with the industrial demands of next-generation manufacturing systems.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eStudy limitations and future research directions\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eWhile the present study has demonstrated the potential of deep learning applied to acoustic signals for tool condition monitoring in woodworking, several avenues remain open for future research. First, the experiments were conducted under controlled laboratory conditions with a limited number of feed rates and two saw blade states. Expanding the study to include a wider range of cutting speeds, tool geometries, and machining conditions would enable the development of more generalizable models capable of handling diverse industrial scenarios.\u003c/p\u003e\n \u003cp\u003eSecond, although this research employed multiple CNN architectures with excellent classification accuracy, further exploration of hybrid and ensemble models could improve robustness, particularly in borderline cases where signal features overlap between classes. Incorporating attention mechanisms or transformer-based architectures may also enhance the ability to capture long-range dependencies in acoustic data.\u003c/p\u003e\n \u003cp\u003eThird, integration of acoustic signals with other sensor modalities, such as vibration, image-based inspection, or current monitoring, could lead to multi-modal fusion frameworks that provide more comprehensive assessments of tool wear and process quality.\u003c/p\u003e\n \u003cp\u003eFourth, future research should investigate real-time deployment strategies, including the use of edge AI devices and embedded platforms. This would allow the development of compact, low-power monitoring systems directly applicable to Industry 4.0 smart factories.\u003c/p\u003e\n \u003cp\u003eFinally, long-term industrial case studies are needed to validate system performance in dynamic production environments, where noise, variability of workpieces, and operator influences may affect signal quality. Such validation would bridge the gap between experimental research and large-scale industrial adoption.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4.0 Conclusions","content":"\u003cp\u003eBased on the information presented, it can be concluded that the findings from the CNN classification offer a satisfactory tool for monitoring the cutting process. Any result with an accuracy level above 95% can be considered excellent, providing a powerful tool for process control and observation.\u003c/p\u003e \u003cp\u003eThe results obtained from cutting power measurements were also expected and satisfactory, making this metric a valuable, robust, and reliable indicator of overall process load.\u003c/p\u003e \u003cp\u003eHowever, statistical analysis of temperature measurement results did not demonstrate the expected causality. Although some anticipated outcomes were observed, such as significant temperature variations for specific feed rates, there were no clear indicators to determine the tool's condition regarding dullness.\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eCutting power provided excellent classification results among the different sample groups, but there are certain limitations to this approach, including material heterogeneity, the provision of only a single scalar value, slower response, and the effects of motor efficiency and transmission losses.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTemperature measurement did not provide satisfactory classification among the defined sample groups. Statistical analysis showed that it was not possible to distinguish the feed rate for the same tool condition.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eSound, as a multi-dimensional signal, was preprocessed to produce inputs for CNN classification datasets in the form of spectrogram images. The accuracy of CNN classification among different groups exceeded 80%, which is considered excellent. Accuracies were 92.00%, 91.00%, 89.00%., 84.00%, 82.00%, 81.00%, 80.00% and 73.00% for ConvNet Xt-XL, ResNet, EfficientNet-B7, VGG19, DenseNet, MobileNetV2, Inception V3, and SqueezeNet, respectively.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Ministry of Science, Technological Development and Innovation of the Republic of Serbia, under registration number 451-03-65/2024-03/200169, dated 05.02.2024.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e1.\u0026nbsp; \u0026nbsp; SS: conceptualization, data curation, formal analysis, investigation, resources, methodology, supervision, writing.\u003c/p\u003e\n\u003cp\u003e2.\u0026nbsp; \u0026nbsp; MP: formal analysis, investigation, validation.\u003c/p\u003e\n\u003cp\u003e3.\u0026nbsp; \u0026nbsp; MF: conceptualization, formal analysis, investigation.\u003c/p\u003e\n\u003cp\u003e4.\u0026nbsp; \u0026nbsp; ZN: software, supervision.\u003c/p\u003e\n\u003cp\u003e5.\u0026nbsp; \u0026nbsp; MĐ: funding acquisition, resources.\u003c/p\u003e\n\u003cp\u003e6.\u0026nbsp; \u0026nbsp; AD: project administration.\u003c/p\u003e\n\u003cp\u003e7.\u0026nbsp; \u0026nbsp; PT: resources.\u003c/p\u003e\n\u003cp\u003e8.\u0026nbsp; \u0026nbsp; AV: data curation, methodology, resources, software, supervision, validation, visualization, writing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to thank Marko Stojičić, Faculty of Forestry employee, who greatly assisted throughout entire investigation process, Miloš Marinković from EGGER Serbia and Darex Home Company, Belgrade, Serbia for obtaining necessary laminated particle boards and Faculty of Forestry, Belgrade, Serbia for participating in publishing costs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData available statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData available on request from the authors. The data that support the findings of this study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics statement\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEthics approval was waived for this study because no patients’ data were reported.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eDuring the preparation of this work, the authors used Grammarly, InstaText and ChatGPT to improve readability. The authors reviewed and edited the content and take full responsibility for the publication.\u003c/em\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e Food and Agriculture Organization of the United Nations, Global Forest Products Facts and Figs.\u0026nbsp;2024, FAO, Rome, Italy, 2024, doi: 10.4060/cd8005en.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e K. Hassan, A. Villa, S. Kuittinen, J. J\u0026auml;nis, and A. Pappinen, \u0026ldquo;An assessment of side-stream generation from the Finnish forest industry,\u0026rdquo; Journal of Material Cycles and Waste Management, vol. 21, pp. 265\u0026ndash;280, 2018, doi: 10.1007/s10163-018-0787-5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e Fortune Business Insights, \u003cem\u003eWoodworking Machinery Market Size, Share \u0026amp; Industry Analysis by Type and Application, and Regional Forecast\u003c/em\u003e, 2025. Available online: \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003ehttps://www.fortunebusinessinsights.com/machinery-industry\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e J.-W. Sun, L.-F. Xi, S.-C. Du, and E.-S. Pan, \u0026ldquo;Tool maintenance optimization for multi-station machining systems with economic consideration of quality loss and obsolescence,\u0026rdquo; Robotics and Computer-Integrated Manufacturing, vol. 26, no. 2, pp. 145\u0026ndash;155, 2010, doi: 10.1016/j.rcim.2009.07.005.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e J. He, H. Gao, S. Li, L. Guo, Y. Lei, and A. Cao, \u0026ldquo;An intelligent maintenance decision-making based on cutters\u0026rsquo; economic life,\u0026rdquo; International Journal of Production Economics, vol. 267, Art. no. 109075, 2024, doi: 10.1016/j.ijpe.2023.109075.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e G. Byrne, D. Dornfeld, and B. Denkena, \u0026ldquo;Advancing cutting technology,\u0026rdquo; CIRP Annals \u0026ndash; Manufacturing Technology, vol. 52, no. 2, pp. 483\u0026ndash;507, 2003, doi: 10.1016/S0007-8506(07)60200-5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e F. Klocke, Manufacturing Processes 1: Cutting. Berlin, Germany: Springer, 2011, doi: 10.1007/978-3-642-11979-8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e Woodworking Network, 2025. Available online: \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003ehttps://www.woodworkingnetwork.com/\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e G. Nemli, I. Aydin, and E. Zekovic, \u0026ldquo;Evaluation of some properties of particleboard as a function of manufacturing parameters,\u0026rdquo; Materials \u0026amp; Design, vol. 28, no. 4, pp. 1169\u0026ndash;1176, 2007, doi: 10.1016/j.matdes.2006.01.015.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e E.-D. Wong, P. Yang, M. Zhang, Q. Wang, T. Nakao, K.-F. Li, and S. Kawai, \u0026ldquo;Analysis of the effects of density profile on bending properties of particleboard using FEM,\u0026rdquo; European Journal of Wood and Wood Products, vol. 61, no. 1, pp. 66\u0026ndash;72, 2003, doi: 10.1007/s00107-002-0350-9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e J. P. Davim (Ed.), Wood Machining. Hoboken, NJ, USA: Wiley, 2013, pp. 176\u0026ndash;177.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e V. Nasir and J. Cool, \u0026ldquo;Characterization, optimization, and acoustic emission monitoring of airborne dust emission during wood sawing,\u0026rdquo; International Journal of Advanced Manufacturing Technology, vol. 109, pp. 2365\u0026ndash;2375, 2020, doi: 10.1007/s00170-020-05842-5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e V. Nasir, J. Cool, and F. Sassani, \u0026ldquo;Acoustic emission monitoring of sawing process using artificial intelligence,\u0026rdquo; International Journal of Advanced Manufacturing Technology, vol. 102, pp. 4179\u0026ndash;4197, 2019, doi: 10.1007/s00170-019-03526-3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e R. Licow, D. Chuchala, M. Deja, K. A. Orlowski, and P. Taube, \u0026ldquo;Effect of pine impregnation and feed speed on sound level and cutting power in wood sawing,\u0026rdquo; Journal of Cleaner Production, 2020, doi: 10.1016/j.jclepro.2020.122833.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e M. Derbas, S. Fr\u0026ouml;mel-Frybort, C. Laaber, and M. Riegler, \u0026ldquo;Sound analysis of mechanical wood cutting processes as a basis for adaptive process control,\u0026rdquo; in Proc. 9th Hardwood Conf., Sopron, Hungary, 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e S. Svrzić, M. Đurković, G. Danon, M. Furtula, and D. Stanojević, \u0026ldquo;Acoustic emission analysis in circular saw cutting of beech wood,\u0026rdquo; BioResources, vol. 16, no. 4, pp. 8239\u0026ndash;8257, 2021, doi: 10.15376/biores.16.4.8239-8257.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e S. Svrzić, M. Đurković, A. Vukićević, Z. Nikolić, V. Mihailović, and A. Dedić, \u0026ldquo;Sound classification and power consumption relation for wood machining monitoring,\u0026rdquo; European Journal of Wood and Wood Products, vol. 82, no. 6, pp. 1953\u0026ndash;1962, 2024, doi: 10.1007/s00107-024-02139-2.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e R. Igaz, R. Kminiak, K. Lubos, M. Nemec, and T. Gregel, \u0026ldquo;Temperature monitoring in CNC machining of solid wood,\u0026rdquo; Sustainability, vol. 11, no. 1, Art. no. 95, 2018, doi: 10.3390/su11010095.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e Y. Murase, K. Matsumoto, and T. Ohuchi, \u0026ldquo;Acoustic emission and cutting resistance of solid wood, MDF and particleboard,\u0026rdquo; Journal of the Faculty of Agriculture, Kyushu University, vol. 53, no. 2, pp. 485\u0026ndash;490, 2008.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e S. Svrzić et al., \u0026ldquo;Sound signal processing and deep learning for determining circular saw blade speed,\u0026rdquo; in Proc. 6th Int. Sci. Conf. Wood Technology and Product Design, Ohrid, North Macedonia, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e M. Mirić-Milosavljević et al., \u0026ldquo;Signal processing and machine learning for identifying idling noises of circular saw blades,\u0026rdquo; BioResources, vol. 19, no. 1, pp. 1744\u0026ndash;1756, 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e Y. Sun et al., \u0026ldquo;Wood species recognition with small data using deep learning,\u0026rdquo; International Journal of Computational Intelligence Systems, vol. 14, no. 1, pp. 1451\u0026ndash;1460, 2021, doi: 10.2991/ijcis.d.210423.001.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e A. R. de Geus et al., \u0026ldquo;Timber section analysis using deep learning,\u0026rdquo; Multimedia Tools and Applications, vol. 79, pp. 34513\u0026ndash;34529, 2020, doi: 10.1007/s11042-020-09212-x.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e P. Kibleur et al., \u0026ldquo;Deep learning segmentation of wood fiber bundles,\u0026rdquo; Composites Science and Technology, 2022, doi: 10.1016/j.compscitech.2022.109287.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e A. Jegorowa et al., \u0026ldquo;Deep learning for drill wear classification in chipboard,\u0026rdquo; Wood Science and Technology, vol. 55, no. 1, pp. 1\u0026ndash;23, 2021, doi: 10.1007/s00226-020-01245-7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e V. Nasir and F. Sassani, \u0026ldquo;A review on deep learning in machining and tool monitoring,\u0026rdquo; International Journal of Advanced Manufacturing Technology, vol. 115, pp. 2683\u0026ndash;2709, 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e K. He, X. Zhang, S. Ren, and J. Sun, \u0026ldquo;Deep residual learning for image recognition,\u0026rdquo; in Proc. IEEE CVPR, 2016, pp. 770\u0026ndash;778.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e M. Sandler et al., \u0026ldquo;MobileNetV2: Inverted residuals and linear bottlenecks,\u0026rdquo; in Proc. IEEE CVPR, 2018, pp. 4510\u0026ndash;4520.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e C. Szegedy et al., \u0026ldquo;Rethinking the inception architecture,\u0026rdquo; in Proc. IEEE CVPR, 2016, pp. 2818\u0026ndash;2826.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e F. N. Iandola et al., \u0026ldquo;SqueezeNet: AlexNet-level accuracy with fewer parameters,\u0026rdquo; arXiv:1602.07360, 2016.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e G. Huang et al., \u0026ldquo;Densely connected convolutional networks,\u0026rdquo; in Proc. IEEE CVPR, 2017, pp. 4700\u0026ndash;4708.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e K. Simonyan and A. Zisserman, \u0026ldquo;Very deep convolutional networks,\u0026rdquo; arXiv:1409.1556, 2014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e M. Tan and Q. V. Le, \u0026ldquo;EfficientNet: Rethinking model scaling,\u0026rdquo; in Proc. ICML, 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e Z. Liu et al., \u0026ldquo;A ConvNet for the 2020s,\u0026rdquo; in Proc. IEEE CVPR, 2022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e O. Russakovsky et al., \u0026ldquo;ImageNet Large Scale Visual Recognition Challenge,\u0026rdquo; IJCV, 2015.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e D. P. Kingma and J. Ba, \u0026ldquo;Adam: A method for stochastic optimization,\u0026rdquo; in Proc. ICLR, 2015.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e L. Crist\u0026oacute;v\u0026atilde;o, Machining Properties of Wood, PhD thesis, Lule\u0026aring; University of Technology, Sweden, 2013.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e M. Ekevad, L. Crist\u0026oacute;v\u0026atilde;o, and B. Marklund, \u0026ldquo;Wear of teeth of circular saw blade,\u0026rdquo; Wood Material Science \u0026amp; Engineering, vol. 7, pp. 150\u0026ndash;153, 2012.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e V. P. Astakhov and J. P. Davim, Machining and Tool Wear. London, UK: Springer, 2008.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e R. Bendikiene and G. Keturakis, \u0026ldquo;Influence of tool characteristics on wear,\u0026rdquo; Journal of Wood Science, 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e A. Naylor et al., \u0026ldquo;Mechanical cutting force model,\u0026rdquo; BioResources, vol. 7, no. 3, pp. 2883\u0026ndash;2894, 2012.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e B. Porankiewicz et al., \u0026ldquo;Cutting forces in machining pine wood,\u0026rdquo; BioResources, vol. 6, no. 4, pp. 3687\u0026ndash;3713, 2011.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e G. Goli et al., \u0026ldquo;Specific cutting forces in engineered wood,\u0026rdquo; Materials, vol. 11, no. 12, Art. no. 2575, 2018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e G. Kovatchev and V. Atanasov, \u0026ldquo;Vibration during longitudinal milling,\u0026rdquo; Acta Facultatis Xylologiae Zvolen, vol. 63, no. 1, pp. 85\u0026ndash;92, 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e J. Kov\u0026aacute;č et al., \u0026ldquo;Analysis of cutting conditions in cross-cutting wood,\u0026rdquo; BioResources, vol. 16, no. 1, pp. 1029\u0026ndash;1041, 2021.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-wood-and-wood-products","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"harw","sideBox":"Learn more about [European Journal of Wood and Wood Products](http://link.springer.com/journal/107)","snPcode":"107","submissionUrl":"https://submission.nature.com/new-submission/107/3","title":"European Journal of Wood and Wood Products","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Wood industry, Particleboard, Cutting tools, Maintenance, Sound analysis, Deep learning","lastPublishedDoi":"10.21203/rs.3.rs-8969198/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8969198/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEffective monitoring of wood cutting processes is crucial for maintaining cutting quality, extending tool life, and reducing energy and material waste. While conventional methods rely on cutting forces, power consumption, or temperature measurements, this study explores acoustic signal analysis as an alternative, leveraging its sensitivity to tool condition, cutting dynamics, and transient high-frequency events. The primary contribution of this study lies in the development of frequency\u0026ndash;intensity\u0026ndash;time representations derived from acoustic signals generated during particleboard cutting and in the subsequent classification of these patterns using deep learning algorithms. In parallel, cutting power consumption and tool temperature were analyzed as reference monitoring variables, with the aim to prove the hypothesis that deep learning-based classification of acoustic data could achieve comparable to or higher accuracy and reliability. Experimental investigations were conducted using two circular saw blades with different dullness conditions saw blades operating at feed rates of 10 m/min and 15 m/min. The workpiece material was melamine-faced particleboard. Statistical analysis revealed highly significant differences in power consumption between cutting conditions, while temperature measurements did not consistently distinguish between feed rates or tool wear states. Acoustic signals were acquired, pre-processed, and classified using several deep learning architectures. The highest classification accuracy of 92.00% was achieved using ConvNetXt-XL, while other architectures, including ResNet, EfficientNet, VGG19, DenseNet, MobileNet V2, Inception V3, and SqueezeNet achieved accuracies between 73.00% and 91.00%. The obtained results demonstrate the effectiveness of acoustic signals as a reliable monitoring variable for intelligent diagnostics and condition monitoring in wood machining processes.\u003c/p\u003e","manuscriptTitle":"Acoustic Signal-Based Cutting Tool Condition Monitoring in Woodworking Using Deep Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-04 07:57:19","doi":"10.21203/rs.3.rs-8969198/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-03-24T17:26:25+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"223829006431911484391221390818560061672","date":"2026-03-19T19:45:34+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-18T11:55:53+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"269227416142527820998872863248144557694","date":"2026-03-17T14:48:24+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-27T11:03:23+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-27T11:00:41+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-26T05:46:58+00:00","index":"","fulltext":""},{"type":"submitted","content":"European Journal of Wood and Wood Products","date":"2026-02-25T15:00:05+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-wood-and-wood-products","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"harw","sideBox":"Learn more about [European Journal of Wood and Wood Products](http://link.springer.com/journal/107)","snPcode":"107","submissionUrl":"https://submission.nature.com/new-submission/107/3","title":"European Journal of Wood and Wood Products","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"f0d57ed1-0fe4-4b85-93c9-f343063d0897","owner":[],"postedDate":"March 4th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-04T07:57:19+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-04 07:57:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8969198","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8969198","identity":"rs-8969198","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}