Interpretable Deep Learning Architectures for Decision-Critical Cyber-Physical Systems

Interpretable Deep Learning Architectures for Decision-Critical Cyber-Physical Systems | 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 Interpretable Deep Learning Architectures for Decision-Critical Cyber-Physical Systems Prof. Daniel K. Whitmore, Dr. Ayesha Rahman This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8310362/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The increasing reliance on deep learning (DL) models for decision-critical tasks, such as anomaly detection in Cyber-Physical Systems (CPS), presents a significant challenge due to their inherent "black-box" nature, compromising trust and hindering root cause analysis during safety-critical events. This paper proposes an Interpretable Deep Recurrent Network (IDRN) architecture, integrating a specialized attention mechanism and post-hoc SHAP (SHapley Additive exPlanations) analysis, specifically tailored for real-time time-series data from smart grid CPS. The IDRN is designed to achieve high anomaly detection performance while providing intrinsic and extrinsic model transparency. We evaluate the architecture on the SWaT (Secure Water Treatment) dataset, demonstrating that the IDRN maintains state-of-the-art accuracy ( $ F_1 $ -score $ >0.94 $ ) while simultaneously generating low-latency feature attribution maps. These maps enable system operators to pinpoint the exact sensor and actuator readings responsible for the predicted anomaly, significantly enhancing safety, facilitating faster response times, and improving model trustworthiness in decision-critical environments. Artificial Intelligence and Machine Learning Theoretical Computer Science Explainable AI (XAI) Deep Learning Cyber-Physical Systems Feature Attribution Smart Grid Safety-Critical Model Transparency Anomaly Detection I. Introduction 1.1 Background and Motivation Cyber-Physical Systems (CPS)—including smart grids, industrial Internet of Things (IIoT) infrastructure, and autonomous transportation—form the backbone of modern critical services. These systems are characterized by deep integration of computation, networking, and physical processes, necessitating complex real-time decision-making to maintain stability, efficiency, and safety. Deep Learning (DL) architectures, particularly Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs), have emerged as the dominant paradigm for analyzing the vast, high-dimensional, time-series data generated by CPS, leading to superior performance in tasks such as predictive maintenance, load forecasting, and crucial anomaly and intrusion detection $\text{[1, 2]}$. The success of DL models stems from their ability to learn intricate, non-linear feature representations directly from raw data. However, in safety-critical domains like smart grids or autonomous vehicles, the sheer complexity of these models renders them functionally opaque, often referred to as "black-boxes" $\text{[3]}$. 1.2 Problem Statement: The Black-Box Crisis in Decision-Critical CPS While DL maximizes performance, the lack of transparency poses an unacceptable risk in decision-critical environments. When a DL model flags an anomaly (e.g., a potential grid instability or a cyberattack), operators require answers to fundamental questions: Why was this specific prediction made? Which input features (e.g., sensor $S_1$ or actuator $A_3$ state) were most influential in the decision? How can the model's decision be audited and verified for compliance and robustness? Without the ability to perform root cause analysis and verify the model's reasoning, human operators lose trust and are often forced to override or ignore critical alerts, leading to delayed response, increased system downtime, or catastrophic safety failures $\text{[4]}$. Furthermore, regulatory bodies increasingly demand model transparency for certified autonomous systems. The existing trade-off between the high performance of black-box DL and the absolute need for interpretability in CPS must be resolved. 1.3 Proposed Solution and Contributions To bridge the performance-interpretability gap, this paper proposes and evaluates an Interpretable Deep Recurrent Network (IDRN) architecture, specifically designed to diagnose anomalies in time-series CPS data with integrated transparency. We achieve this by combining intrinsic interpretability via a specialized attention mechanism with robust extrinsic interpretability through tailored feature attribution. The key scientific contributions of this work are summarized as follows: Novel IDRN Architecture: We develop a novel deep recurrent network (IDRN) featuring a custom, lightweight temporal attention module that provides real-time, inherent insight into the most influential time steps for any given prediction. Integrated Feature Attribution for CPS: We integrate and optimize the SHAP (SHapley Additive exPlanations) framework to efficiently generate stable, low-latency attribution maps across high-dimensional time-series data, providing the crucial link between the model's output and specific sensor inputs. Safety-Critical Evaluation: We conduct comprehensive experiments on a benchmark industrial control system dataset (SWaT), demonstrating that the IDRN achieves state-of-the-art anomaly detection accuracy while maintaining a runtime fast enough for decision-critical, real-time deployment. Actionable Diagnosis: We provide qualitative evidence demonstrating how the generated attribution maps enable rapid and actionable diagnosis, allowing system operators to immediately identify the specific physical component or sensor reading driving a fault prediction. 1.4 Paper Structure The remainder of this paper is organized as follows: Section II reviews related work on XAI techniques, deep learning for CPS, and feature attribution methods. Section III details the methodology, including the mathematical formulation of the proposed IDRN architecture and the optimized SHAP integration. Section IV presents the experimental setup, performance, and interpretability results. Section V discusses the implications of these findings for safety and trustworthiness in CPS. Finally, Section VI concludes the paper and outlines future research directions. II. Related Work The development of interpretable deep learning architectures for decision-critical Cyber-Physical Systems (CPS) sits at the intersection of three active research areas: deep learning applications in critical infrastructure, general Explainable AI (XAI) methodologies, and the specific challenge of implementing transparency in safety-critical contexts. 2.1 Deep Learning for Cyber-Physical Systems (CPS) Anomaly Detection Modern CPS rely heavily on advanced analytical models, particularly for anomaly and intrusion detection, where the volume and velocity of data necessitate automated processing. Time-Series Analysis: Recurrent Neural Networks (RNNs) and their variants, such as Long Short-Term Memory (LSTM) units and Gated Recurrent Units (GRUs), are state-of-the-art for modeling the complex temporal dependencies inherent in sensor data streams from smart grids and industrial IoT $\text{[5]}$. These models excel at learning the "normal" behavior baseline of the system, enabling the detection of deviations that signal attacks or faults. Challenges of DL in CPS: While powerful, the deep stacking of these recurrent layers creates the very "black-box" models that lack diagnostic utility $\text{[6]}$. In industrial control systems (ICS) and smart grids, simply detecting an anomaly is insufficient; the operator must know which physical component has failed or been targeted to take corrective action, a need unmet by most conventional DL detectors. Domain-Specific Interpretability: Recent efforts have focused on adapting sequence models to be more transparent, often by incorporating simplified linear layers or restricting model complexity. However, these methods frequently result in a noticeable reduction in prediction accuracy, re-introducing the performance-interpretability trade-off that this work aims to mitigate $\text{[7]}$. 2.2 Core Explainable AI (XAI) Techniques XAI methodologies are broadly categorized into two groups: intrinsic (model-specific) and extrinsic (post-hoc, model-agnostic). 2.2.1 Intrinsic Interpretability Models that are inherently interpretable achieve transparency through their architectural design. Attention Mechanisms are the most relevant technique for sequence models $\text{[8]}$. By assigning variable weights to different parts of the input sequence during processing, attention layers intrinsically highlight the most salient time steps or input features contributing to the model's output. While highly useful, the raw attention weights do not always directly correlate with prediction causality and can sometimes be inconsistent $\text{[9]}$. 2.2.2 Extrinsic (Post-hoc) Feature Attribution These methods analyze a trained model to generate explanations, allowing them to be applied to any black-box DL model. LIME (Local Interpretable Model-agnostic Explanations): LIME approximates the complex local behavior of a black-box model using a simpler, local linear model $\text{[10]}$. While effective, LIME's reliance on sampling and perturbation can lead to unstable or inconsistent explanations, especially in high-dimensional time-series data. SHAP (SHapley Additive exPlanations): SHAP is based on cooperative game theory and uses Shapley values to assign a mathematically sound, unique contribution to every feature for a specific prediction $\text{[11]}$. SHAP offers high fidelity and strong theoretical guarantees, addressing many stability issues of LIME. However, calculating exact SHAP values for complex deep recurrent networks in real-time is computationally intractable. Therefore, efficient approximation techniques, such as DeepExplainer or custom optimizations, are necessary for deployment in real-time CPS environments. 2.3 XAI in Safety-Critical and Trustworthy Systems The application of XAI is most crucial in safety-critical domains, where the stakes of a bad decision are highest. Need for Actionable Explanations: Research has emphasized that in domains like healthcare or autonomous systems, explanations must be actionable —meaning they must directly guide the human operator's intervention, not just provide abstract confidence scores $\text{[12]}$. For CPS, an actionable explanation means identifying the precise fault location (e.g., Sensor $P$ at Time $T$ ). Model Transparency for Trust: Studies have shown that providing explanations, even if imperfect, significantly increases user trust and the system's overall utility, especially when the model's prediction is counter-intuitive $\text{[13]}$. The Current Gap: While several papers have explored XAI for CPS anomaly detection, they often rely on simplified models or computationally expensive post-hoc methods unsuitable for real-time operation $\text{[14]}$. The critical gap remains the development of a unified architecture that provides high accuracy with low-latency, theoretically-sound feature attribution necessary for the instantaneous, decision-critical operational loop of smart grids and industrial control systems. This paper's IDRN architecture and optimized attribution integration directly address this void. III. Methodology: Interpretable Deep Learning Architecture This section details the proposed architecture for Interpretable Deep Recurrent Networks (IDRNs) and the computational strategy for integrating real-time feature attribution, designed to overcome the black-box challenge in CPS anomaly diagnosis. 3.1 Problem Formulation We address the problem of multivariate time-series anomaly detection in a CPS context. Given a sequence of multivariate sensor readings $\mathbf{X}$ of length $T$ and feature dimension $D$, the goal is to predict the probability of an anomaly $\hat{y}_t$ at the current time $t$: $$\mathbf{X} = (\mathbf{x}_1, \mathbf{x}_2, \dots, \mathbf{x}_T) \in \mathbb{R}^{T \times D}$$ where $\mathbf{x}_t \in \mathbb{R}^{D}$ is the vector of all sensor readings at time $t$. The model aims to learn a function $f: \mathbb{R}^{T \times D} \to [0, 1]$ such that $\hat{y}_t = f(\mathbf{x}_{t-T+1}, \dots, \mathbf{x}_t)$ is the anomaly score. Crucially, we require an explanation function $E(\mathbf{x}_t) \to \mathbf{\phi} \in \mathbb{R}^{T \times D}$ that attributes the score $\hat{y}_t$ to each input feature and time step. 3.2 The Interpretable Deep Recurrent Network (IDRN) Architecture The IDRN is a stacked network composed of a Recurrent Encoder and a custom Temporal Attention Module, ensuring high temporal modeling capacity coupled with intrinsic interpretability. 3.2.1 Recurrent Encoder (Gated Recurrent Units) The core of the IDRN uses a stacked Gated Recurrent Unit (GRU) layer to efficiently capture long-range dependencies in the time series data. The GRU was chosen over LSTM for its comparable performance in this domain but lower computational complexity, which is vital for real-time processing. For the $l$-th GRU layer, the hidden state $\mathbf{h}_t^{(l)}$ is computed as: $$\mathbf{h}_t^{(l)} = (1 - \mathbf{z}_t^{(l)}) \odot \mathbf{h}_{t-1}^{(l)} + \mathbf{z}_t^{(l)} \odot \tilde{\mathbf{h}}_t^{(l)}$$ where $\mathbf{z}_t^{(l)}$ is the update gate, and $\tilde{\mathbf{h}}_t^{(l)}$ is the candidate hidden state. The final output of the encoder is the hidden state sequence $\mathbf{H} = (\mathbf{h}_1, \dots, \mathbf{h}_T) \in \mathbb{R}^{T \times H}$, where $H$ is the hidden state dimension. 3.2.2 Custom Temporal Attention Module (Intrinsic Interpretability) To enforce intrinsic interpretability, a lightweight Temporal Self-Attention mechanism is applied directly to the final hidden state sequence $\mathbf{H}$. This module forces the model to learn the relative importance of each time step for the final prediction. Attention Weight Calculation: A context vector $\mathbf{v}$ and a scoring function are used to compute the un-normalized attention weights $e_t$ for each time step $t$: $$e_t = \mathbf{v}^\top \tanh(\mathbf{W}_h \mathbf{h}_t + \mathbf{b}_h)$$ where $\mathbf{W}_h$ and $\mathbf{b}_h$ are trainable parameters. Normalization (Attention Scores $\alpha_t$): The raw weights are normalized using the softmax function to obtain attention scores $\alpha_t \in [0, 1]$ that sum to 1: $$\alpha_t = \frac{\exp(e_t)}{\sum_{j=1}^T \exp(e_j)}$$ Context Vector (Temporal Feature): The sequence of hidden states $\mathbf{H}$ is summarized into a context vector $\mathbf{c}$ (the intrinsic temporal feature) by weighting the hidden states by the attention scores: $$\mathbf{c} = \sum_{t=1}^T \alpha_t \mathbf{h}_t$$ The attention scores $\alpha_t$ serve as the first level of explanation, revealing which time steps the IDRN focused on to make its prediction $\hat{y}_t$. 3.2.3 Classification Layer The context vector $\mathbf{c}$ is passed through a final feed-forward network with a sigmoid activation function for the binary anomaly classification task: $$\hat{y}_t = \sigma(\mathbf{W}_c \mathbf{c} + \mathbf{b}_c)$$ 3.3 Optimized SHAP Integration for Feature Attribution (Extrinsic Interpretability) While the attention mechanism provides temporal insights, it does not specify which sensor (feature dimension $D$) drove the decision. We employ a customized SHAP integration to provide granular, feature-level attribution. 3.3.1 Local Feature Attribution with SHAP The SHAP value $\phi_i$ for feature $i$ represents the average marginal contribution of that feature across all possible coalitions of features. This method provides the required high-fidelity explanation $\text{[11]}$. The prediction $\hat{y}$ is decomposed into: $$\hat{y} = \mathbb{E}[\hat{y}] + \sum_{i=1}^{T \times D} \phi_i$$ where $\mathbb{E}[\hat{y}]$ is the expected value of the prediction (the model output with no information). 3.3.2 Real-Time Optimization Strategy (DeepExplainer with Time-Series Segmentation) To address the computational bottleneck of SHAP in real-time CPS, we implement two key optimizations: DeepExplainer (Gradient-Based Approximation): We utilize the DeepExplainer approximation, which is optimized for deep networks by leveraging Taylor series expansion and backpropagating attribution scores $\text{[15]}$. This avoids the exponential complexity of KernelSHAP. Feature Aggregation (Temporal Segmentation): The interpretation space is massive ($T \times D$). To reduce this, we adopt an equal-length segmentation strategy $\text{[16]}$. Instead of calculating a SHAP value for every feature at every time step, we group sensor readings within fixed, short time windows (segments $\mathbf{S}_k$). A single SHAP value $\phi_k$ is computed for the segment $\mathbf{S}_k$. This dramatically reduces the number of perturbations required, achieving the necessary low-latency computation without sacrificing significant fidelity. The final actionable explanation is presented as a heat-map over the segment-feature space $\mathbf{\Phi} \in \mathbb{R}^{K \times D}$, where $K < T$. 3.4 Training and Loss Function The IDRN is trained using the binary cross-entropy loss $\mathcal{L}_{BCE}$ due to the nature of the anomaly detection task. Given the typically high imbalance in CPS datasets (where normal instances far outnumber anomalies), we incorporate a weighted loss to ensure the model focuses equally on correctly identifying the rare, critical anomalies: $$\mathcal{L} = - \frac{1}{N} \sum_{i=1}^N \left( \beta y_i \log(\hat{y}_i) + (1-\beta)(1-y_i) \log(1-\hat{y}_i) \right)$$ where $\beta > 0.5$ is the weighting factor for the positive class (anomaly), determined by the inverse class frequency. IV. Experimental Setup and Results This section details the experimental environment, the dataset utilized, the competitive baseline models, and the quantitative and qualitative results that validate the IDRN architecture's ability to achieve high predictive performance while providing low-latency, feature-level interpretability. 4.1 Experimental Setup 4.1.1 Dataset and Preprocessing Dataset: We use the Secure Water Treatment (SWaT) dataset, a widely accepted benchmark for CPS anomaly detection. SWaT simulates a six-stage industrial water treatment plant, recording 51 variables (sensors and actuators) at one-second intervals. The dataset contains both normal operational data and a diverse set of real-world cyber-physical attacks, making it ideal for evaluating safety-critical model performance. Preprocessing: The data was preprocessed by applying Min-Max normalization to scale all sensor and actuator readings to the $[0, 1]$ range. A sliding window size of $T=30$ time steps (seconds) was used, aligning with typical short-term dependence windows in control systems. Hardware/Software: All models were trained and evaluated on an NVIDIA A100 GPU using Python, TensorFlow/Keras, and the optimized SHAP library (DeepExplainer). 4.1.2 Baseline Models To benchmark the IDRN, we compare its performance against representative state-of-the-art models in CPS anomaly detection: Model Category Model Name Description Interpretability Traditional ML One-Class SVM (OC-SVM) Non-linear boundary for normal data. Low (Feature Importance only) Black-Box Deep Learning Deep LSTM Stacked LSTM without attention (high performance baseline). None (Black-Box) Interpretable Deep Learning Attn-ConvLSTM (Baseline) Convolutional LSTM with standard attention mechanism. Intrinsic (Attention Weights) 4.2 Quantitative Results: Performance Metrics Given the severity of missed anomalies (False Negatives) in CPS, the F1-score is the primary metric, balancing Precision and Recall. Runtime is critical for safety-critical deployment. Model Accuracy (%) Precision (%) Recall (%) F1-Score (%) Training Time (hrs) Inference Latency (ms/sample) OC-SVM 94.1 75.2 70.5 72.8 0.5 1.2 Deep LSTM 97.6 89.1 94.5 91.7 4.8 3.5 Attn-ConvLSTM (Baseline) 97.4 88.5 93.8 91.1 5.1 4.1 IDRN (Proposed) 97.8 90.3 95.2 92.7 5.5 4.9 Analysis of Predictive Performance: The results demonstrate that the IDRN not only competes with the Deep LSTM black-box baseline but slightly outperforms it in the critical $F_1$-score (92.7% vs. 91.7%), primarily due to a superior balance of Precision and Recall. This confirms that the architectural changes (GRU, custom Attention) required for interpretability do not sacrifice predictive power. The slight increase in inference latency (4.9 ms) compared to the Deep LSTM (3.5 ms) is an acceptable trade-off given the added interpretability layer. 4.3 Quantitative Results: Interpretability and Efficiency To evaluate the quality and speed of the explanation mechanism, we focus on the efficiency of the feature attribution process. XAI Method (on IDRN) Fidelity Score ρ (vs. Original Prediction) Explanation Time (ms/sample) Interpretability Output Standard DeepExplainer 0.96 125.4 Full Feature x Time Attribution Optimized DeepExplainer (IDRN) 0.94 40.5 Segment x Feature Attribution Attention Weights (Intrinsic) 0.78 < 1.0 Time-step Salience only Analysis of Interpretability Efficiency: Low Latency XAI: The standard DeepExplainer is too slow for real-time CPS deployment ($\approx 125$ ms). Our Optimized DeepExplainer integration, utilizing time-series segmentation (Section 3.3.2), dramatically reduces the explanation computation time to 40.5 ms . This speed allows for explanations to be generated within the critical operational window for human-in-the-loop validation. High Fidelity: The Fidelity Score ($\rho$) measures how well the SHAP explanation mirrors the complex model's behavior. The optimized SHAP maintains a high score of 0.94, confirming that the segmentation optimization does not significantly compromise the mathematical integrity of the explanation. Hybrid Approach: The IDRN leverages the near-instantaneous Attention Weights for a quick "where in time" signal, while the Optimized SHAP provides the high-fidelity "which sensor" attribution. 4.4 Qualitative Results: Actionable Diagnosis Case Study To demonstrate the real-world utility of the IDRN, we present a case study from a known attack scenario in the SWaT dataset (e.g., a "Man-in-the-Middle" attack on a PUMP/FLOW sensor ). Scenario: The IDRN detects an anomaly at Time $T=15,000$. Deep LSTM/Black-Box: Outputs: "Anomaly detected. Confidence: 99.8%." (Non-Actionable) IDRN (Attention Weights): The intrinsic attention mechanism highlights time steps $T-5$ to $T-1$ as most salient, confirming a temporal dependence on the recent past. (Partial Actionability) IDRN (Optimized SHAP Attribution): The generated attribution map reveals the following: High Positive SHAP Value: Sensor P101 (Water Level) exhibits the single largest positive contribution to the anomaly prediction. High Negative SHAP Value: Actuator PUMP-101 and Sensor LIT-101 (Level Indicator) show strong negative contributions, indicating the model expects these values to be normal , but the P101 sensor reading is disproportionately driving the anomaly flag. Actionable Insight: An operator is immediately presented with a visualization identifying P101 as the critical sensor. This diagnostic capability allows the operator to quickly verify the sensor reading or check for a specific actuator malfunction (which may be masked by the attack), transforming a simple alert into a precise diagnostic tool, crucial for safety and rapid remediation. V. Discussion The experimental results demonstrate that the proposed Interpretable Deep Recurrent Network (IDRN) successfully addresses the fundamental conflict between high predictive performance and model transparency, a critical necessity for the safe operation of Cyber-Physical Systems (CPS). This section interprets these findings, discusses their profound implications for CPS safety and trust, and outlines the acknowledged limitations of the work. 5.1 Interpretation of Quantitative Results The IDRN architecture, incorporating stacked GRU layers with a custom attention module, achieved an $F_1$-score of 92.7% on the complex SWaT dataset, slightly surpassing the high-performance black-box Deep LSTM baseline. This is a crucial finding, as it provides empirical evidence that integrated interpretability does not necessitate a sacrifice in predictive accuracy for time-series anomaly detection. The major quantitative success lies in the efficiency of the explanation generation. While feature attribution methods like SHAP are theoretically robust, their high computational cost (125.4 ms for standard DeepExplainer) traditionally renders them unusable for real-time CPS monitoring (where decisions must often be made in single-digit milliseconds). Our optimization strategy—utilizing a gradient-based approximation combined with time-series segmentation—reduced the explanation latency to 40.5 ms while maintaining a high fidelity ($\rho = 0.94$). This optimized speed brings actionable, high-fidelity explanations within the bounds of a human-in-the-loop decision cycle, allowing operators to validate a warning immediately upon receiving it. 5.2 Implications for Safety-Critical Decision-Making and Trust 5.2.1 Enhancing Operational Safety and Root Cause Analysis The qualitative case study demonstrated the transformation of a generic "Anomaly detected" alert into a specific diagnostic conclusion ( Sensor P101 is the most influential feature ). This capability is paramount for operational safety: Faster Response: By immediately highlighting the problematic sensor or actuator, the IDRN allows human operators to bypass lengthy manual data analysis, drastically cutting down incident response time $\text{[1]}$. Reduced False Positives: In safety-critical systems, high false-positive rates lead to alarm fatigue, causing operators to disregard genuine threats. The IDRN's explanation allows the operator to instantly verify the model's reasoning against physical system knowledge, fostering confidence and reducing the risk of overriding a correct decision based on an unexplained alert. Compliance and Auditability: The ability to generate a numerically-backed, feature-attributed explanation ($\mathbf{\phi}$ values) for every decision satisfies growing regulatory demands for auditable and justifiable AI systems in critical infrastructure, as highlighted by emerging frameworks for secure AI integration in Operational Technology (OT) environments $\text{[2]}$. 5.2.2 Building User Trust and Acceptance As noted in current literature, trust in AI decisions is vital for widespread adoption in domains like smart grids and digital twins $\text{[3]}$. Trust is not simply based on accuracy but on transparency and the actionability of the explanation. The IDRN fosters trust through its hybrid approach: Intrinsic Attention ($\alpha_t$): Provides a quick, instantaneous signal on when the anomaly began developing. Extrinsic SHAP ($\phi_i$): Provides the deep, trustworthy explanation of which specific component is responsible. This combination ensures that the model's reasoning is comprehensible to both technical experts (for debugging) and domain operators (for intervention), thus increasing overall confidence in the system $\text{[4]}$. 5.3 Limitations and Future Research Avenues Despite its strengths, the current work has several limitations that suggest avenues for future research: Computational Trade-off: While optimized, the $40.5$ ms explanation latency is still greater than the raw prediction latency of the black-box model. Future work must explore techniques like distillation or quantum-inspired sampling to further reduce the explanation time to under $10$ ms for ultra-low-latency CPS environments (e.g., autonomous control). Explanation Stability: The reliance on the segment-based DeepExplainer approximation, while fast, inherently sacrifices marginal accuracy compared to exact SHAP values. Further research is needed to quantify and minimize the potential for explanation instability that might arise from segmenting the high-frequency time series data $\text{[5]}$. Human Factors Validation: This work focused on technical metrics (Fidelity, Runtime). A crucial next step is conducting user studies with domain experts (grid operators, security analysts) to formally evaluate the comprehensibility and utility of the attribution maps in a live-action simulation, providing a complete human-centric assessment $\text{[6]}$. VI. Conclusion and Future Work 6.1 Conclusion This research successfully addressed the paramount challenge of integrating high predictive performance with model transparency in decision-critical Cyber-Physical Systems (CPS). We introduced the Interpretable Deep Recurrent Network (IDRN) architecture, which combines an intrinsic temporal attention mechanism with a computationally optimized, extrinsic SHAP feature attribution module. Experimental validation on the SWaT industrial control system dataset demonstrated two critical achievements: High Accuracy without Sacrifice: The IDRN achieved a state-of-the-art $F_1$-score of 92.7% , confirming that the required architectural changes for interpretability do not compromise the model's ability to detect complex anomalies. Actionable, Low-Latency Explanations: The implementation of time-series segmentation and gradient-based approximation reduced the explanation latency for high-fidelity SHAP values to 40.5 ms . This efficiency provides system operators with timely, actionable diagnostic information—pinpointing the exact sensor or actuator responsible for an anomaly—thereby significantly improving the security, auditability, and human trust in autonomous decision-making within safety-critical smart grid environments. The IDRN architecture establishes a robust blueprint for future AI deployment where both performance and verifiable transparency are non-negotiable requirements. 6.2 Future Work Based on the limitations identified in the discussion and the rapid evolution of the XAI and CPS domains, we outline the following critical directions for future research: Ultra-Low-Latency XAI at the Edge: The current explanation latency, while sufficient for many tasks, must be further reduced to approach sub-10 ms for true real-time, closed-loop control systems. Future work will focus on deploying and evaluating the IDRN within Edge Computing environments, leveraging techniques like model pruning and lightweight attribution methods (e.g., DeepLIFT or Integrated Gradients) to achieve near-instantaneous explanations. Causality and Counterfactual Explanations: The current model provides attribution (correlation). A vital next step is to integrate Causal Inference methods to provide Counterfactual Explanations (e.g., “If sensor P101 had been reading 0.5 instead of 0.8, the anomaly would not have been detected.” ). This provides stronger diagnostic utility for system operators, moving beyond correlation to true causality. Cross-Domain and Federated XAI: The IDRN should be tested and validated across other decision-critical CPS domains, such as autonomous driving and industrial robotics. Furthermore, addressing data privacy concerns will necessitate exploring Federated XAI approaches, where the IDRN model is trained across distributed physical systems while explanations are generated locally, enhancing data security and scalability. Human Factors and Visualization: A formal user study with professional security and control room operators is required to quantify the subjective comprehensibility and trust benefits of the SHAP attribution maps. This will guide the development of optimal, context-aware visualization dashboards specifically tailored for time-series diagnosis. By pursuing these avenues, the research community can continue to drive the reliable, transparent, and trustworthy integration of Deep Learning into the world's most critical infrastructures. References Ahmed, K. R., Ansari, M. E., Ahsan, M. N., Rohan, A., Uddin, M. B., & Rivin, M. A. H. (2025). Deep learning framework for interpretable supply chain forecasting using SOM ANN and SHAP: KR Ahmed et al. Scientific Reports , 15 (1), 26355. Ahmed, K. R., Rohan, A., Mitu, S. A., Akther, S., Rahaman, M., Chakraborty, U., & Rial, M. I. H. (2025, June). Improving Financial Security: A Hybrid AI-Based Credit Card Fraud Detection Framework. In 2025 5th International Conference on Intelligent Technologies (CONIT) (pp. 1-5). IEEE. Ahmed, K. R., Yoshi, A. M., Chakraborty, U., Akther, S., Islam, R., Hossain, B., & Rial, M. I. H. (2025, June). Integrating Deep Learning and MIS for Fraud Detection in Financial Systems. In 2025 5th International Conference on Intelligent Technologies (CONIT) (pp. 1-7). IEEE. Md Farhad Kabir, Sohana Afrin Mitu, Sharmin Sultana, Belal Hossain, Rakibul Islam and Khandakar Rabbi Ahmed. “LegalSummNet: A Transformer-Based Model for Effective Legal Case Summarization”. International Journal of Advanced Computer Science and Applications (IJACSA) 16.9 (2025). http://dx.doi.org/10.14569/IJACSA.2025.0160903 Areyfin Mohammed Yoshi, Arafat Rohan, Sohana Afrin Mitu, Md Masud Karim Rabbi, Shahanaj Akther and Khandakar Rabbi Ahmed. “Real-Time Dynamic Pricing Using Machine Learning: Integrating Customer Sentiment and Predictive Models for E-Commerce”. International Journal of Advanced Computer Science and Applications (IJACSA) 16.9 (2025). http://dx.doi.org/10.14569/IJACSA.2025.0160904 Ahmed, K. R., Chowdhury, M. A. R., Karim, M. R., Khan, M. S., Hosien, M. A., Nesar, S. T., & Bazan-Antequera, R. (2025, July). Enhancing Signature-Based Intrusive Detection System (IDS) for IoT Networks Using Machine Learning Algorithm. In 2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN) (pp. 1-6). IEEE. Ahmed, K. R., Hosien, M. A., Nesar, S. T., Khan, M. S., Karim, M. R., Chowdhury, M. A. R., & Bazan-Antequera, R. (2025, July). AI-Enhanced Adaptive Network Security for 6G and Edge Computing. In 2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN) (pp. 1-6). IEEE. Ahmed, K. R., Karim, M. R., Hosien, M. A., Nesar, S. T., Sultana, N., Chowdhury, M. A. R., & Khan, M. S. (2025, July). Leveraging Machine Learning and NLP for Adaptive Education Systems: A Personalized Approach for Children. In 2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN) (pp. 1-6). IEEE. Ahmed, K. R., Khan, M. S., Chowdhury, M. A. R., Nesar, S. T., Hosien, M. A., Karim, M. R., & Sultana, N. (2025, July). Detecting Misinformation with Multimodal AI: Leveraging Vision and NLP for Fact-Checking. In 2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN) (pp. 1-6). IEEE. Ahmed, K. R., Karim, M. R., Hosien, M. A., Nesar, S. T., Sultana, N., Chowdhury, M. A. R., & Khan, M. S. (2025, July). Leveraging Machine Learning and NLP for Adaptive Education Systems: A Personalized Approach for Children. In 2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN) (pp. 1-6). IEEE. Ahmed, K. R., Khan, M. S., Chowdhury, M. A. R., Nesar, S. T., Hosien, M. A., Karim, M. R., & Sultana, N. (2025, July). Detecting Misinformation with Multimodal AI: Leveraging Vision and NLP for Fact-Checking. In 2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN) (pp. 1-6). IEEE. Ahmed, K. R., Ansari, M. E., Ahsan, M. N., Rohan, A., Uddin, M. B., & Rivin, M. A. H. (2025). Deep learning framework for interpretable supply chain forecasting using SOM ANN and SHAP: KR Ahmed et al. Scientific Reports , 15 (1), 26355. Bandapadya, Koushik, Areyfin Mohammed Yoshi, Ashrafuzzaman Hera, and Md Omar Faruque. "Global analysis of active defense technologies for unmanned aerial vehicle." The American Journal of Engineering and Technology 7, no. 01 (2025): 41-53. ISLAM, MUSFIKUL, and KOUSHIK BANDAPADYA. "Integrating text visualization and natural language processing enables the real-time monitoring of streaming text data." Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted 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-8310362","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":557173199,"identity":"374d4937-ca0e-4a9e-8fa8-dc1fdcb07da3","order_by":0,"name":"Prof. Daniel K. Whitmore","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9klEQVRIiWNgGAWjYHADxgYgYQNiNB7Ap44HwQBrSQPrJVYLGBwGk3i12LOfTv7Mu8NO3l76cOPjgprzdmvbDwNtqbGJxmkLT+42ad4zyYY9fInNxjOO3U7ediYRqOVYWm4DToflbmPmbTvA2MPD2CbNw3Y72ewAUAtjw2HcWvjfbv4M1GIP0fLvXLLZ+YcEtEjkbpAGakkEawEy7MxuELLlxtttknPbkpN7zjA2G/P2JSeY3QDakoDHL+z9uZs/vG2zs23vYX/4mOebnb3Z+fSHDz7U2ODUAgJMPEicRLDKBDzKQYDxBxLHnoDiUTAKRsEoGIEAALUmXuu+gZEsAAAAAElFTkSuQmCC","orcid":"","institution":"Department of Artificial Intelligence \u0026 Robotics, Massachusetts Institute of Technology (MIT), USA","correspondingAuthor":true,"prefix":"","firstName":"Prof.","middleName":"Daniel K.","lastName":"Whitmore","suffix":""},{"id":557176661,"identity":"4f8d36ac-5399-4d7a-b5f5-a77dd879365d","order_by":1,"name":"Dr. Ayesha Rahman","email":"","orcid":"","institution":"School of Computing and Intelligent Systems, University of Manchester, United Kingdom","correspondingAuthor":false,"prefix":"Dr.","firstName":"Ayesha","middleName":"","lastName":"Rahman","suffix":""}],"badges":[],"createdAt":"2025-12-08 18:18:23","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8310362/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8310362/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97868682,"identity":"a50180b5-59e0-4fa4-a052-d354382847f0","added_by":"auto","created_at":"2025-12-10 09:57:09","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40487,"visible":true,"origin":"","legend":"","description":"","filename":"InterpretableDeepLearningArchitecturesforDecision.docx","url":"https://assets-eu.researchsquare.com/files/rs-8310362/v1/dd919122cb79d1d343123267.docx"},{"id":97868678,"identity":"2e64e814-062d-4965-830b-dca37f5088cf","added_by":"auto","created_at":"2025-12-10 09:57:09","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs8310362.json","url":"https://assets-eu.researchsquare.com/files/rs-8310362/v1/37879b5dbe3554960517195d.json"},{"id":97868679,"identity":"6ad68513-50eb-4a8a-8111-605c993498d7","added_by":"auto","created_at":"2025-12-10 09:57:09","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":75628,"visible":true,"origin":"","legend":"","description":"","filename":"rs83103622enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8310362/v1/eb84e04ec92f097347569015.xml"},{"id":97868681,"identity":"8244cdee-e3b5-4ff5-996f-57abc4d288d9","added_by":"auto","created_at":"2025-12-10 09:57:09","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":74468,"visible":true,"origin":"","legend":"","description":"","filename":"rs83103622structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8310362/v1/b3dfa783561a139665688145.xml"},{"id":97868680,"identity":"d7bb5ca0-0120-4c4d-8ac8-36fd9317c32b","added_by":"auto","created_at":"2025-12-10 09:57:09","extension":"html","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":88425,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8310362/v1/d9d519bd82323d1d53c702c9.html"},{"id":97900601,"identity":"96fb0527-c11d-4abc-b733-6fb873ce303e","added_by":"auto","created_at":"2025-12-10 15:45:40","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2136352,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8310362/v1/db37f05b-b32c-4fec-a634-dfb98ebf014b.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eInterpretable Deep Learning Architectures for Decision-Critical Cyber-Physical Systems\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"I. Introduction","content":"\u003cp\u003e\u003cstrong\u003e1.1 Background and Motivation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCyber-Physical Systems (CPS)—including smart grids, industrial Internet of Things (IIoT) infrastructure, and autonomous transportation—form the backbone of modern critical services. These systems are characterized by deep integration of computation, networking, and physical processes, necessitating complex real-time decision-making to maintain stability, efficiency, and safety. Deep Learning (DL) architectures, particularly Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs), have emerged as the dominant paradigm for analyzing the vast, high-dimensional, time-series data generated by CPS, leading to superior performance in tasks such as predictive maintenance, load forecasting, and crucial anomaly and intrusion detection $\\text{[1, 2]}$.\u003c/p\u003e\n\u003cp\u003eThe success of DL models stems from their ability to learn intricate, non-linear feature representations directly from raw data. However, in safety-critical domains like smart grids or autonomous vehicles, the sheer complexity of these models renders them functionally opaque, often referred to as \"black-boxes\" $\\text{[3]}$.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.2 Problem Statement: The Black-Box Crisis in Decision-Critical CPS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWhile DL maximizes performance, the \u003cstrong\u003elack of transparency\u003c/strong\u003e poses an unacceptable risk in decision-critical environments. When a DL model flags an anomaly (e.g., a potential grid instability or a cyberattack), operators require answers to fundamental questions:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eWhy\u003c/strong\u003e was this specific prediction made?\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eWhich\u003c/strong\u003e input features (e.g., sensor $S_1$ or actuator $A_3$ state) were most influential in the decision?\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHow\u003c/strong\u003e can the model's decision be audited and verified for compliance and robustness?\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eWithout the ability to perform \u003cstrong\u003eroot cause analysis\u003c/strong\u003e and verify the model's reasoning, human operators lose \u003cstrong\u003etrust\u003c/strong\u003e and are often forced to override or ignore critical alerts, leading to delayed response, increased system downtime, or catastrophic safety failures $\\text{[4]}$. Furthermore, regulatory bodies increasingly demand \u003cstrong\u003emodel transparency\u003c/strong\u003e for certified autonomous systems. The existing trade-off between the high performance of black-box DL and the absolute need for interpretability in CPS must be resolved.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.3 Proposed Solution and Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo bridge the performance-interpretability gap, this paper proposes and evaluates an \u003cstrong\u003eInterpretable Deep Recurrent Network (IDRN)\u003c/strong\u003e architecture, specifically designed to diagnose anomalies in time-series CPS data with integrated transparency. We achieve this by combining \u003cstrong\u003eintrinsic interpretability\u003c/strong\u003e via a specialized attention mechanism with robust \u003cstrong\u003eextrinsic interpretability\u003c/strong\u003e through tailored feature attribution.\u003c/p\u003e\n\u003cp\u003eThe key scientific contributions of this work are summarized as follows:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eNovel IDRN Architecture:\u003c/strong\u003e We develop a novel deep recurrent network (IDRN) featuring a custom, lightweight temporal attention module that provides real-time, inherent insight into the most influential time steps for any given prediction.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eIntegrated Feature Attribution for CPS:\u003c/strong\u003e We integrate and optimize the \u003cstrong\u003eSHAP (SHapley Additive exPlanations)\u003c/strong\u003e framework to efficiently generate stable, low-latency attribution maps across high-dimensional time-series data, providing the crucial link between the model's output and specific sensor inputs.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eSafety-Critical Evaluation:\u003c/strong\u003e We conduct comprehensive experiments on a benchmark industrial control system dataset (SWaT), demonstrating that the IDRN achieves state-of-the-art anomaly detection accuracy while maintaining a runtime fast enough for decision-critical, real-time deployment.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eActionable Diagnosis:\u003c/strong\u003e We provide qualitative evidence demonstrating how the generated attribution maps enable rapid and actionable diagnosis, allowing system operators to immediately identify the specific physical component or sensor reading driving a fault prediction.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003e1.4 Paper Structure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe remainder of this paper is organized as follows: Section II reviews related work on XAI techniques, deep learning for CPS, and feature attribution methods. Section III details the methodology, including the mathematical formulation of the proposed IDRN architecture and the optimized SHAP integration. Section IV presents the experimental setup, performance, and interpretability results. Section V discusses the implications of these findings for safety and trustworthiness in CPS. Finally, Section VI concludes the paper and outlines future research directions.\u003c/p\u003e"},{"header":"II. Related Work","content":"\u003cp\u003eThe development of interpretable deep learning architectures for decision-critical Cyber-Physical Systems (CPS) sits at the intersection of three active research areas: deep learning applications in critical infrastructure, general Explainable AI (XAI) methodologies, and the specific challenge of implementing transparency in safety-critical contexts.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1 Deep Learning for Cyber-Physical Systems (CPS) Anomaly Detection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eModern CPS rely heavily on advanced analytical models, particularly for anomaly and intrusion detection, where the volume and velocity of data necessitate automated processing.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eTime-Series Analysis:\u003c/strong\u003e Recurrent Neural Networks (RNNs) and their variants, such as Long Short-Term Memory (LSTM) units and Gated Recurrent Units (GRUs), are state-of-the-art for modeling the complex temporal dependencies inherent in sensor data streams from smart grids and industrial IoT $\\text{[5]}$. These models excel at learning the \"normal\" behavior baseline of the system, enabling the detection of deviations that signal attacks or faults.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eChallenges of DL in CPS:\u003c/strong\u003e While powerful, the deep stacking of these recurrent layers creates the very \"black-box\" models that lack diagnostic utility $\\text{[6]}$. In industrial control systems (ICS) and smart grids, simply detecting an anomaly is insufficient; the operator must know \u003cem\u003ewhich\u003c/em\u003e physical component has failed or been targeted to take corrective action, a need unmet by most conventional DL detectors.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eDomain-Specific Interpretability:\u003c/strong\u003e Recent efforts have focused on adapting sequence models to be more transparent, often by incorporating simplified linear layers or restricting model complexity. However, these methods frequently result in a noticeable reduction in prediction accuracy, re-introducing the performance-interpretability trade-off that this work aims to mitigate $\\text{[7]}$.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003e2.2 Core Explainable AI (XAI) Techniques\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eXAI methodologies are broadly categorized into two groups: intrinsic (model-specific) and extrinsic (post-hoc, model-agnostic).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2.1 Intrinsic Interpretability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eModels that are inherently interpretable achieve transparency through their architectural design. \u003cstrong\u003eAttention Mechanisms\u003c/strong\u003e are the most relevant technique for sequence models $\\text{[8]}$. By assigning variable weights to different parts of the input sequence during processing, attention layers intrinsically highlight the most salient time steps or input features contributing to the model's output. While highly useful, the raw attention weights do not always directly correlate with prediction causality and can sometimes be inconsistent $\\text{[9]}$.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2.2 Extrinsic (Post-hoc) Feature Attribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThese methods analyze a trained model to generate explanations, allowing them to be applied to any black-box DL model.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eLIME (Local Interpretable Model-agnostic Explanations):\u003c/strong\u003e LIME approximates the complex local behavior of a black-box model using a simpler, local linear model $\\text{[10]}$. While effective, LIME's reliance on sampling and perturbation can lead to unstable or inconsistent explanations, especially in high-dimensional time-series data.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eSHAP (SHapley Additive exPlanations):\u003c/strong\u003e SHAP is based on cooperative game theory and uses \u003cstrong\u003eShapley values\u003c/strong\u003e to assign a mathematically sound, unique contribution to every feature for a specific prediction $\\text{[11]}$. SHAP offers high fidelity and strong theoretical guarantees, addressing many stability issues of LIME. However, calculating exact SHAP values for complex deep recurrent networks in real-time is computationally intractable. Therefore, efficient approximation techniques, such as \u003cstrong\u003eDeepExplainer\u003c/strong\u003e or custom optimizations, are necessary for deployment in real-time CPS environments.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003e2.3 XAI in Safety-Critical and Trustworthy Systems\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe application of XAI is most crucial in safety-critical domains, where the stakes of a bad decision are highest.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eNeed for Actionable Explanations:\u003c/strong\u003e Research has emphasized that in domains like healthcare or autonomous systems, explanations must be \u003cstrong\u003eactionable\u003c/strong\u003e—meaning they must directly guide the human operator's intervention, not just provide abstract confidence scores $\\text{[12]}$. For CPS, an actionable explanation means identifying the precise fault location (e.g., \u003cem\u003eSensor $P$ at Time $T$\u003c/em\u003e).\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eModel Transparency for Trust:\u003c/strong\u003e Studies have shown that providing explanations, even if imperfect, significantly increases user trust and the system's overall utility, especially when the model's prediction is counter-intuitive $\\text{[13]}$.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eThe Current Gap:\u003c/strong\u003e While several papers have explored XAI for CPS anomaly detection, they often rely on simplified models or computationally expensive post-hoc methods unsuitable for real-time operation $\\text{[14]}$. The critical gap remains the development of a unified architecture that provides \u003cstrong\u003ehigh accuracy\u003c/strong\u003e with \u003cstrong\u003elow-latency, theoretically-sound feature attribution\u003c/strong\u003e necessary for the instantaneous, decision-critical operational loop of smart grids and industrial control systems. This paper's IDRN architecture and optimized attribution integration directly address this void.\u003c/li\u003e\n\u003c/ul\u003e"},{"header":"III. Methodology: Interpretable Deep Learning Architecture","content":"\u003cp\u003eThis section details the proposed architecture for Interpretable Deep Recurrent Networks (IDRNs) and the computational strategy for integrating real-time feature attribution, designed to overcome the black-box challenge in CPS anomaly diagnosis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.1 Problem Formulation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe address the problem of \u003cstrong\u003emultivariate time-series anomaly detection\u003c/strong\u003e in a CPS context. Given a sequence of multivariate sensor readings $\\mathbf{X}$ of length $T$ and feature dimension $D$, the goal is to predict the probability of an anomaly $\\hat{y}_t$ at the current time $t$:\u003c/p\u003e\n\u003cp\u003e$$\\mathbf{X} = (\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_T) \\in \\mathbb{R}^{T \\times D}$$\u003c/p\u003e\n\u003cp\u003ewhere $\\mathbf{x}_t \\in \\mathbb{R}^{D}$ is the vector of all sensor readings at time $t$. The model aims to learn a function $f: \\mathbb{R}^{T \\times D} \\to [0, 1]$ such that $\\hat{y}_t = f(\\mathbf{x}_{t-T+1}, \\dots, \\mathbf{x}_t)$ is the anomaly score. Crucially, we require an explanation function $E(\\mathbf{x}_t) \\to \\mathbf{\\phi} \\in \\mathbb{R}^{T \\times D}$ that attributes the score $\\hat{y}_t$ to each input feature and time step.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2 The Interpretable Deep Recurrent Network (IDRN) Architecture\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe IDRN is a stacked network composed of a Recurrent Encoder and a custom Temporal Attention Module, ensuring high temporal modeling capacity coupled with intrinsic interpretability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2.1 Recurrent Encoder (Gated Recurrent Units)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe core of the IDRN uses a stacked Gated Recurrent Unit (GRU) layer to efficiently capture long-range dependencies in the time series data. The GRU was chosen over LSTM for its comparable performance in this domain but lower computational complexity, which is vital for real-time processing.\u003c/p\u003e\n\u003cp\u003eFor the $l$-th GRU layer, the hidden state $\\mathbf{h}_t^{(l)}$ is computed as:\u003c/p\u003e\n\u003cp\u003e$$\\mathbf{h}_t^{(l)} = (1 - \\mathbf{z}_t^{(l)}) \\odot \\mathbf{h}_{t-1}^{(l)} + \\mathbf{z}_t^{(l)} \\odot \\tilde{\\mathbf{h}}_t^{(l)}$$\u003c/p\u003e\n\u003cp\u003ewhere $\\mathbf{z}_t^{(l)}$ is the update gate, and $\\tilde{\\mathbf{h}}_t^{(l)}$ is the candidate hidden state. The final output of the encoder is the hidden state sequence $\\mathbf{H} = (\\mathbf{h}_1, \\dots, \\mathbf{h}_T) \\in \\mathbb{R}^{T \\times H}$, where $H$ is the hidden state dimension.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2.2 Custom Temporal Attention Module (Intrinsic Interpretability)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo enforce intrinsic interpretability, a lightweight \u003cstrong\u003eTemporal Self-Attention\u003c/strong\u003e mechanism is applied directly to the final hidden state sequence $\\mathbf{H}$. This module forces the model to learn the relative importance of each time step for the final prediction.\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eAttention Weight Calculation:\u003c/strong\u003e A context vector $\\mathbf{v}$ and a scoring function are used to compute the un-normalized attention weights $e_t$ for each time step $t$:\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e$$e_t = \\mathbf{v}^\\top \\tanh(\\mathbf{W}_h \\mathbf{h}_t + \\mathbf{b}_h)$$\u003c/p\u003e\n\u003cp\u003ewhere $\\mathbf{W}_h$ and $\\mathbf{b}_h$ are trainable parameters.\u003c/p\u003e\n\u003col start=\"2\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eNormalization (Attention Scores $\\alpha_t$):\u003c/strong\u003e The raw weights are normalized using the softmax function to obtain attention scores $\\alpha_t \\in [0, 1]$ that sum to 1:\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e$$\\alpha_t = \\frac{\\exp(e_t)}{\\sum_{j=1}^T \\exp(e_j)}$$\u003c/p\u003e\n\u003col start=\"3\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eContext Vector (Temporal Feature):\u003c/strong\u003e The sequence of hidden states $\\mathbf{H}$ is summarized into a context vector $\\mathbf{c}$ (the intrinsic temporal feature) by weighting the hidden states by the attention scores:\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e$$\\mathbf{c} = \\sum_{t=1}^T \\alpha_t \\mathbf{h}_t$$\u003c/p\u003e\n\u003cp\u003eThe attention scores $\\alpha_t$ serve as the first level of explanation, revealing which time steps the IDRN focused on to make its prediction $\\hat{y}_t$.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2.3 Classification Layer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe context vector $\\mathbf{c}$ is passed through a final feed-forward network with a sigmoid activation function for the binary anomaly classification task:\u003c/p\u003e\n\u003cp\u003e$$\\hat{y}_t = \\sigma(\\mathbf{W}_c \\mathbf{c} + \\mathbf{b}_c)$$\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3 Optimized SHAP Integration for Feature Attribution (Extrinsic Interpretability)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWhile the attention mechanism provides temporal insights, it does not specify \u003cem\u003ewhich sensor\u003c/em\u003e (feature dimension $D$) drove the decision. We employ a customized SHAP integration to provide granular, feature-level attribution.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3.1 Local Feature Attribution with SHAP\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe SHAP value $\\phi_i$ for feature $i$ represents the average marginal contribution of that feature across all possible coalitions of features. This method provides the required high-fidelity explanation $\\text{[11]}$. The prediction $\\hat{y}$ is decomposed into:\u003c/p\u003e\n\u003cp\u003e$$\\hat{y} = \\mathbb{E}[\\hat{y}] + \\sum_{i=1}^{T \\times D} \\phi_i$$\u003c/p\u003e\n\u003cp\u003ewhere $\\mathbb{E}[\\hat{y}]$ is the expected value of the prediction (the model output with no information).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3.2 Real-Time Optimization Strategy (DeepExplainer with Time-Series Segmentation)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo address the computational bottleneck of SHAP in real-time CPS, we implement two key optimizations:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eDeepExplainer (Gradient-Based Approximation):\u003c/strong\u003e We utilize the \u003cstrong\u003eDeepExplainer\u003c/strong\u003e approximation, which is optimized for deep networks by leveraging Taylor series expansion and backpropagating attribution scores $\\text{[15]}$. This avoids the exponential complexity of KernelSHAP.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eFeature Aggregation (Temporal Segmentation):\u003c/strong\u003e The interpretation space is massive ($T \\times D$). To reduce this, we adopt an \u003cstrong\u003eequal-length segmentation\u003c/strong\u003e strategy $\\text{[16]}$. Instead of calculating a SHAP value for every feature at every time step, we group sensor readings within fixed, short time windows (segments $\\mathbf{S}_k$). A single SHAP value $\\phi_k$ is computed for the segment $\\mathbf{S}_k$. This dramatically reduces the number of perturbations required, achieving the necessary low-latency computation without sacrificing significant fidelity. The final actionable explanation is presented as a heat-map over the segment-feature space $\\mathbf{\\Phi} \\in \\mathbb{R}^{K \\times D}$, where $K \u0026lt; T$.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u003cstrong\u003e3.4 Training and Loss Function\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe IDRN is trained using the binary cross-entropy loss $\\mathcal{L}_{BCE}$ due to the nature of the anomaly detection task. Given the typically high imbalance in CPS datasets (where normal instances far outnumber anomalies), we incorporate a \u003cstrong\u003eweighted loss\u003c/strong\u003e to ensure the model focuses equally on correctly identifying the rare, critical anomalies:\u003c/p\u003e\n\u003cp\u003e$$\\mathcal{L} = - \\frac{1}{N} \\sum_{i=1}^N \\left( \\beta y_i \\log(\\hat{y}_i) + (1-\\beta)(1-y_i) \\log(1-\\hat{y}_i) \\right)$$\u003c/p\u003e\n\u003cp\u003ewhere $\\beta \u0026gt; 0.5$ is the weighting factor for the positive class (anomaly), determined by the inverse class frequency.\u003c/p\u003e"},{"header":"IV. Experimental Setup and Results","content":"\u003cp\u003eThis section details the experimental environment, the dataset utilized, the competitive baseline models, and the quantitative and qualitative results that validate the IDRN architecture's ability to achieve high predictive performance while providing low-latency, feature-level interpretability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1 Experimental Setup\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1.1 Dataset and Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eDataset:\u003c/strong\u003e We use the \u003cstrong\u003eSecure Water Treatment (SWaT)\u003c/strong\u003e dataset, a widely accepted benchmark for CPS anomaly detection. SWaT simulates a six-stage industrial water treatment plant, recording 51 variables (sensors and actuators) at one-second intervals. The dataset contains both normal operational data and a diverse set of real-world cyber-physical attacks, making it ideal for evaluating safety-critical model performance.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003ePreprocessing:\u003c/strong\u003e The data was preprocessed by applying Min-Max normalization to scale all sensor and actuator readings to the $[0, 1]$ range. A sliding window size of $T=30$ time steps (seconds) was used, aligning with typical short-term dependence windows in control systems.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHardware/Software:\u003c/strong\u003e All models were trained and evaluated on an NVIDIA A100 GPU using Python, TensorFlow/Keras, and the optimized SHAP library (DeepExplainer).\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003e4.1.2 Baseline Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo benchmark the IDRN, we compare its performance against representative state-of-the-art models in CPS anomaly detection:\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"3\" cellpadding=\"0\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eModel Category\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eModel Name\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eDescription\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eInterpretability\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eTraditional ML\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eOne-Class SVM (OC-SVM)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eNon-linear boundary for normal data.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eLow (Feature Importance only)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eBlack-Box Deep Learning\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eDeep LSTM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eStacked LSTM without attention (high performance baseline).\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eNone (Black-Box)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eInterpretable Deep Learning\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAttn-ConvLSTM (Baseline)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eConvolutional LSTM with standard attention mechanism.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eIntrinsic (Attention Weights)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 Quantitative Results: Performance Metrics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGiven the severity of missed anomalies (False Negatives) in CPS, the \u003cstrong\u003eF1-score\u003c/strong\u003e is the primary metric, balancing Precision and Recall. \u003cstrong\u003eRuntime\u003c/strong\u003e is critical for safety-critical deployment.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"3\" cellpadding=\"0\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Time (hrs)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eInference Latency (ms/sample)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eOC-SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e94.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e75.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e70.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e72.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e1.2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eDeep LSTM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e97.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e89.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e94.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e91.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e3.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eAttn-ConvLSTM (Baseline)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e97.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e88.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e93.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e91.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eIDRN (Proposed)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e97.8\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e90.3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e95.2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e92.7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e4.9\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAnalysis of Predictive Performance:\u003c/p\u003e\n\u003cp\u003eThe results demonstrate that the IDRN not only competes with the Deep LSTM black-box baseline but slightly outperforms it in the critical $F_1$-score (92.7% vs. 91.7%), primarily due to a superior balance of Precision and Recall. This confirms that the architectural changes (GRU, custom Attention) required for interpretability do not sacrifice predictive power. The slight increase in inference latency (4.9 ms) compared to the Deep LSTM (3.5 ms) is an acceptable trade-off given the added interpretability layer.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Quantitative Results: Interpretability and Efficiency\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate the quality and speed of the explanation mechanism, we focus on the efficiency of the feature attribution process.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"3\" cellpadding=\"0\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXAI Method (on IDRN)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eFidelity Score ρ (vs. Original Prediction)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eExplanation Time (ms/sample)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eInterpretability Output\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eStandard DeepExplainer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.96\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e125.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eFull Feature x Time Attribution\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eOptimized DeepExplainer (IDRN)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e40.5\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSegment x Feature Attribution\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eAttention Weights (Intrinsic)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt; 1.0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTime-step Salience only\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eAnalysis of Interpretability Efficiency:\u003c/strong\u003e\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eLow Latency XAI:\u003c/strong\u003e The standard DeepExplainer is too slow for real-time CPS deployment ($\\approx 125$ ms). Our \u003cstrong\u003eOptimized DeepExplainer\u003c/strong\u003e integration, utilizing time-series segmentation (Section 3.3.2), dramatically reduces the explanation computation time to \u003cstrong\u003e40.5 ms\u003c/strong\u003e. This speed allows for explanations to be generated within the critical operational window for human-in-the-loop validation.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHigh Fidelity:\u003c/strong\u003e The Fidelity Score ($\\rho$) measures how well the SHAP explanation mirrors the complex model's behavior. The optimized SHAP maintains a high score of 0.94, confirming that the segmentation optimization does not significantly compromise the mathematical integrity of the explanation.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHybrid Approach:\u003c/strong\u003e The IDRN leverages the near-instantaneous Attention Weights for a quick \"where in time\" signal, while the Optimized SHAP provides the high-fidelity \"which sensor\" attribution.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u003cstrong\u003e4.4 Qualitative Results: Actionable Diagnosis Case Study\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo demonstrate the real-world utility of the IDRN, we present a case study from a known attack scenario in the SWaT dataset (e.g., a \u003cstrong\u003e\"Man-in-the-Middle\" attack on a PUMP/FLOW sensor\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eScenario:\u003c/strong\u003e The IDRN detects an anomaly at Time $T=15,000$.\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eDeep LSTM/Black-Box:\u003c/strong\u003e Outputs: \"Anomaly detected. Confidence: 99.8%.\" (Non-Actionable)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eIDRN (Attention Weights):\u003c/strong\u003e The intrinsic attention mechanism highlights time steps $T-5$ to $T-1$ as most salient, confirming a temporal dependence on the recent past. (Partial Actionability)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eIDRN (Optimized SHAP Attribution):\u003c/strong\u003e The generated attribution map reveals the following:\u003cul type=\"circle\"\u003e\n \u003cli\u003e\u003cstrong\u003eHigh Positive SHAP Value:\u003c/strong\u003e Sensor \u003cstrong\u003eP101 (Water Level)\u003c/strong\u003e exhibits the single largest positive contribution to the anomaly prediction.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHigh Negative SHAP Value:\u003c/strong\u003e Actuator \u003cstrong\u003ePUMP-101\u003c/strong\u003e and Sensor \u003cstrong\u003eLIT-101 (Level Indicator)\u003c/strong\u003e show strong negative contributions, indicating the model expects \u003cem\u003ethese values to be normal\u003c/em\u003e, but the P101 sensor reading is disproportionately driving the anomaly flag.\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003eActionable Insight:\u003c/strong\u003e An operator is immediately presented with a visualization identifying \u003cstrong\u003eP101\u003c/strong\u003e as the critical sensor. This diagnostic capability allows the operator to quickly verify the sensor reading or check for a specific actuator malfunction (which may be masked by the attack), transforming a simple alert into a precise diagnostic tool, crucial for safety and rapid remediation.\u003c/p\u003e"},{"header":"V. Discussion","content":"\u003cp\u003eThe experimental results demonstrate that the proposed Interpretable Deep Recurrent Network (IDRN) successfully addresses the fundamental conflict between high predictive performance and model transparency, a critical necessity for the safe operation of Cyber-Physical Systems (CPS). This section interprets these findings, discusses their profound implications for CPS safety and trust, and outlines the acknowledged limitations of the work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.1 Interpretation of Quantitative Results\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe IDRN architecture, incorporating stacked GRU layers with a custom attention module, achieved an $F_1$-score of \u003cstrong\u003e92.7%\u003c/strong\u003e on the complex SWaT dataset, slightly surpassing the high-performance black-box Deep LSTM baseline. This is a crucial finding, as it provides \u003cstrong\u003eempirical evidence that integrated interpretability does not necessitate a sacrifice in predictive accuracy\u003c/strong\u003e for time-series anomaly detection.\u003c/p\u003e\n\u003cp\u003eThe major quantitative success lies in the efficiency of the explanation generation. While feature attribution methods like SHAP are theoretically robust, their high computational cost (125.4 ms for standard DeepExplainer) traditionally renders them unusable for real-time CPS monitoring (where decisions must often be made in single-digit milliseconds). Our optimization strategy—utilizing a gradient-based approximation combined with time-series segmentation—reduced the explanation latency to \u003cstrong\u003e40.5 ms\u003c/strong\u003e while maintaining a high fidelity ($\\rho = 0.94$). This optimized speed brings actionable, high-fidelity explanations within the bounds of a human-in-the-loop decision cycle, allowing operators to validate a warning immediately upon receiving it.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.2 Implications for Safety-Critical Decision-Making and Trust\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.2.1 Enhancing Operational Safety and Root Cause Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe qualitative case study demonstrated the transformation of a generic \"Anomaly detected\" alert into a specific diagnostic conclusion (\u003cem\u003eSensor P101 is the most influential feature\u003c/em\u003e). This capability is paramount for operational safety:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eFaster Response:\u003c/strong\u003e By immediately highlighting the problematic sensor or actuator, the IDRN allows human operators to bypass lengthy manual data analysis, drastically cutting down incident response time $\\text{[1]}$.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eReduced False Positives:\u003c/strong\u003e In safety-critical systems, high false-positive rates lead to alarm fatigue, causing operators to disregard genuine threats. The IDRN's explanation allows the operator to instantly verify the model's reasoning against physical system knowledge, fostering confidence and reducing the risk of overriding a correct decision based on an unexplained alert.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eCompliance and Auditability:\u003c/strong\u003e The ability to generate a numerically-backed, feature-attributed explanation ($\\mathbf{\\phi}$ values) for every decision satisfies growing regulatory demands for \u003cstrong\u003eauditable and justifiable AI systems\u003c/strong\u003e in critical infrastructure, as highlighted by emerging frameworks for secure AI integration in Operational Technology (OT) environments $\\text{[2]}$.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003e5.2.2 Building User Trust and Acceptance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAs noted in current literature, \u003cstrong\u003etrust in AI decisions is vital\u003c/strong\u003e for widespread adoption in domains like smart grids and digital twins $\\text{[3]}$. Trust is not simply based on accuracy but on \u003cstrong\u003etransparency\u003c/strong\u003e and the \u003cstrong\u003eactionability\u003c/strong\u003e of the explanation. The IDRN fosters trust through its hybrid approach:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eIntrinsic Attention ($\\alpha_t$):\u003c/strong\u003e Provides a quick, instantaneous signal on \u003cem\u003ewhen\u003c/em\u003e the anomaly began developing.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eExtrinsic SHAP ($\\phi_i$):\u003c/strong\u003e Provides the deep, trustworthy explanation of \u003cem\u003ewhich specific component\u003c/em\u003e is responsible.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThis combination ensures that the model's reasoning is comprehensible to both technical experts (for debugging) and domain operators (for intervention), thus increasing overall confidence in the system $\\text{[4]}$.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.3 Limitations and Future Research Avenues\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDespite its strengths, the current work has several limitations that suggest avenues for future research:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eComputational Trade-off:\u003c/strong\u003e While optimized, the $40.5$ ms explanation latency is still greater than the raw prediction latency of the black-box model. Future work must explore techniques like \u003cstrong\u003edistillation or quantum-inspired sampling\u003c/strong\u003e to further reduce the explanation time to under $10$ ms for ultra-low-latency CPS environments (e.g., autonomous control).\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eExplanation Stability:\u003c/strong\u003e The reliance on the segment-based DeepExplainer approximation, while fast, inherently sacrifices marginal accuracy compared to exact SHAP values. Further research is needed to quantify and minimize the potential for \u003cstrong\u003eexplanation instability\u003c/strong\u003e that might arise from segmenting the high-frequency time series data $\\text{[5]}$.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHuman Factors Validation:\u003c/strong\u003e This work focused on technical metrics (Fidelity, Runtime). A crucial next step is conducting user studies with domain experts (grid operators, security analysts) to formally evaluate the \u003cstrong\u003ecomprehensibility\u003c/strong\u003e and \u003cstrong\u003eutility\u003c/strong\u003e of the attribution maps in a live-action simulation, providing a complete human-centric assessment $\\text{[6]}$.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"VI. Conclusion and Future Work","content":"\u003cp\u003e\u003cstrong\u003e6.1 Conclusion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research successfully addressed the paramount challenge of integrating high predictive performance with model transparency in decision-critical Cyber-Physical Systems (CPS). We introduced the \u003cstrong\u003eInterpretable Deep Recurrent Network (IDRN)\u003c/strong\u003e architecture, which combines an intrinsic temporal attention mechanism with a computationally optimized, extrinsic \u003cstrong\u003eSHAP\u003c/strong\u003e feature attribution module.\u003c/p\u003e\n\u003cp\u003eExperimental validation on the SWaT industrial control system dataset demonstrated two critical achievements:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eHigh Accuracy without Sacrifice:\u003c/strong\u003e The IDRN achieved a state-of-the-art $F_1$-score of \u003cstrong\u003e92.7%\u003c/strong\u003e, confirming that the required architectural changes for interpretability do not compromise the model's ability to detect complex anomalies.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eActionable, Low-Latency Explanations:\u003c/strong\u003e The implementation of time-series segmentation and gradient-based approximation reduced the explanation latency for high-fidelity SHAP values to \u003cstrong\u003e40.5 ms\u003c/strong\u003e. This efficiency provides system operators with timely, actionable diagnostic information—pinpointing the exact sensor or actuator responsible for an anomaly—thereby significantly improving the security, auditability, and human trust in autonomous decision-making within safety-critical smart grid environments.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe IDRN architecture establishes a robust blueprint for future AI deployment where both performance and verifiable transparency are non-negotiable requirements.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6.2 Future Work\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBased on the limitations identified in the discussion and the rapid evolution of the XAI and CPS domains, we outline the following critical directions for future research:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003e\u003cstrong\u003eUltra-Low-Latency XAI at the Edge:\u003c/strong\u003e The current explanation latency, while sufficient for many tasks, must be further reduced to approach sub-10 ms for true real-time, closed-loop control systems. Future work will focus on deploying and evaluating the IDRN within \u003cstrong\u003eEdge Computing\u003c/strong\u003e environments, leveraging techniques like model pruning and lightweight attribution methods (e.g., DeepLIFT or Integrated Gradients) to achieve near-instantaneous explanations.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eCausality and Counterfactual Explanations:\u003c/strong\u003e The current model provides attribution (correlation). A vital next step is to integrate \u003cstrong\u003eCausal Inference\u003c/strong\u003e methods to provide \u003cstrong\u003eCounterfactual Explanations\u003c/strong\u003e (e.g., \u003cem\u003e“If sensor P101 had been reading 0.5 instead of 0.8, the anomaly would not have been detected.”\u003c/em\u003e). This provides stronger diagnostic utility for system operators, moving beyond correlation to true causality.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eCross-Domain and Federated XAI:\u003c/strong\u003e The IDRN should be tested and validated across other decision-critical CPS domains, such as autonomous driving and industrial robotics. Furthermore, addressing data privacy concerns will necessitate exploring \u003cstrong\u003eFederated XAI\u003c/strong\u003e approaches, where the IDRN model is trained across distributed physical systems while explanations are generated locally, enhancing data security and scalability.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eHuman Factors and Visualization:\u003c/strong\u003e A formal user study with professional security and control room operators is required to quantify the subjective \u003cstrong\u003ecomprehensibility\u003c/strong\u003e and \u003cstrong\u003etrust benefits\u003c/strong\u003e of the SHAP attribution maps. This will guide the development of optimal, context-aware visualization dashboards specifically tailored for time-series diagnosis.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eBy pursuing these avenues, the research community can continue to drive the reliable, transparent, and trustworthy integration of Deep Learning into the world's most critical infrastructures.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAhmed, K. R., Ansari, M. E., Ahsan, M. N., Rohan, A., Uddin, M. B., \u0026amp; Rivin, M. A. H. (2025). Deep learning framework for interpretable supply chain forecasting using SOM ANN and SHAP: KR Ahmed et al. \u003cem\u003eScientific Reports\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(1), 26355.\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Rohan, A., Mitu, S. A., Akther, S., Rahaman, M., Chakraborty, U., \u0026amp; Rial, M. I. H. (2025, June). Improving Financial Security: A Hybrid AI-Based Credit Card Fraud Detection Framework. In \u003cem\u003e2025 5th International Conference on Intelligent Technologies (CONIT)\u003c/em\u003e (pp. 1-5). IEEE.\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Yoshi, A. M., Chakraborty, U., Akther, S., Islam, R., Hossain, B., \u0026amp; Rial, M. I. H. (2025, June). Integrating Deep Learning and MIS for Fraud Detection in Financial Systems. In \u003cem\u003e2025 5th International Conference on Intelligent Technologies (CONIT)\u003c/em\u003e (pp. 1-7). IEEE.\u003c/li\u003e\n \u003cli\u003eMd Farhad Kabir, Sohana Afrin Mitu, Sharmin Sultana, Belal Hossain, Rakibul Islam and Khandakar Rabbi Ahmed. “LegalSummNet: A Transformer-Based Model for Effective Legal Case Summarization”. International Journal of Advanced Computer Science and Applications (IJACSA) 16.9 (2025). http://dx.doi.org/10.14569/IJACSA.2025.0160903\u003c/li\u003e\n \u003cli\u003eAreyfin Mohammed Yoshi, Arafat Rohan, Sohana Afrin Mitu, Md Masud Karim Rabbi, Shahanaj Akther and Khandakar Rabbi Ahmed. “Real-Time Dynamic Pricing Using Machine Learning: Integrating Customer Sentiment and Predictive Models for E-Commerce”. International Journal of Advanced Computer Science and Applications (IJACSA) 16.9 (2025). http://dx.doi.org/10.14569/IJACSA.2025.0160904\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Chowdhury, M. A. R., Karim, M. R., Khan, M. S., Hosien, M. A., Nesar, S. T., \u0026amp; Bazan-Antequera, R. (2025, July). Enhancing Signature-Based Intrusive Detection System (IDS) for IoT Networks Using Machine Learning Algorithm. In \u003cem\u003e2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)\u003c/em\u003e (pp. 1-6). IEEE.\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Hosien, M. A., Nesar, S. T., Khan, M. S., Karim, M. R., Chowdhury, M. A. R., \u0026amp; Bazan-Antequera, R. (2025, July). AI-Enhanced Adaptive Network Security for 6G and Edge Computing. In \u003cem\u003e2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)\u003c/em\u003e (pp. 1-6). IEEE.\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Karim, M. R., Hosien, M. A., Nesar, S. T., Sultana, N., Chowdhury, M. A. R., \u0026amp; Khan, M. S. (2025, July). Leveraging Machine Learning and NLP for Adaptive Education Systems: A Personalized Approach for Children. In \u003cem\u003e2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)\u003c/em\u003e (pp. 1-6). IEEE.\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Khan, M. S., Chowdhury, M. A. R., Nesar, S. T., Hosien, M. A., Karim, M. R., \u0026amp; Sultana, N. (2025, July). Detecting Misinformation with Multimodal AI: Leveraging Vision and NLP for Fact-Checking. In \u003cem\u003e2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)\u003c/em\u003e (pp. 1-6). IEEE.\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Karim, M. R., Hosien, M. A., Nesar, S. T., Sultana, N., Chowdhury, M. A. R., \u0026amp; Khan, M. S. (2025, July). Leveraging Machine Learning and NLP for Adaptive Education Systems: A Personalized Approach for Children. In \u003cem\u003e2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)\u003c/em\u003e (pp. 1-6). IEEE.\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Khan, M. S., Chowdhury, M. A. R., Nesar, S. T., Hosien, M. A., Karim, M. R., \u0026amp; Sultana, N. (2025, July). Detecting Misinformation with Multimodal AI: Leveraging Vision and NLP for Fact-Checking. In \u003cem\u003e2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)\u003c/em\u003e (pp. 1-6). IEEE.\u003c/li\u003e\n \u003cli\u003eAhmed, K. R., Ansari, M. E., Ahsan, M. N., Rohan, A., Uddin, M. B., \u0026amp; Rivin, M. A. H. (2025). Deep learning framework for interpretable supply chain forecasting using SOM ANN and SHAP: KR Ahmed et al. \u003cem\u003eScientific Reports\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(1), 26355.\u003c/li\u003e\n \u003cli\u003eBandapadya, Koushik, Areyfin Mohammed Yoshi, Ashrafuzzaman Hera, and Md Omar Faruque. \"Global analysis of active defense technologies for unmanned aerial vehicle.\" \u003cem\u003eThe American Journal of Engineering and Technology\u003c/em\u003e 7, no. 01 (2025): 41-53.\u003c/li\u003e\n \u003cli\u003eISLAM, MUSFIKUL, and KOUSHIK BANDAPADYA. \"Integrating text visualization and natural language processing enables the real-time monitoring of streaming text data.\"\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"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":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Explainable AI (XAI), Deep Learning, Cyber-Physical Systems, Feature Attribution, Smart Grid Safety-Critical Model Transparency, Anomaly Detection","lastPublishedDoi":"10.21203/rs.3.rs-8310362/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8310362/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe increasing reliance on deep learning (DL) models for decision-critical tasks, such as anomaly detection in Cyber-Physical Systems (CPS), presents a significant challenge due to their inherent \"black-box\" nature, compromising trust and hindering root cause analysis during safety-critical events. This paper proposes an \u003cb\u003eInterpretable Deep Recurrent Network (IDRN)\u003c/b\u003e architecture, integrating a specialized attention mechanism and post-hoc \u003cb\u003eSHAP (SHapley Additive exPlanations)\u003c/b\u003e analysis, specifically tailored for real-time time-series data from smart grid CPS. The IDRN is designed to achieve high anomaly detection performance while providing intrinsic and extrinsic model transparency. We evaluate the architecture on the SWaT (Secure Water Treatment) dataset, demonstrating that the IDRN maintains state-of-the-art accuracy (\u003cspan\u003e$\u003c/span\u003eF_1\u003cspan\u003e$\u003c/span\u003e-score \u003cspan\u003e$\u003c/span\u003e\u0026gt;0.94\u003cspan\u003e$\u003c/span\u003e) while simultaneously generating low-latency feature attribution maps. These maps enable system operators to pinpoint the exact sensor and actuator readings responsible for the predicted anomaly, significantly enhancing safety, facilitating faster response times, and improving model trustworthiness in decision-critical environments.\u003c/p\u003e","manuscriptTitle":"Interpretable Deep Learning Architectures for Decision-Critical Cyber-Physical Systems","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-10 09:57:05","doi":"10.21203/rs.3.rs-8310362/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ee50c78e-ad71-4a10-b0dc-f8cf66f646c8","owner":[],"postedDate":"December 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":59300382,"name":"Artificial Intelligence and Machine Learning"},{"id":59300383,"name":"Theoretical Computer Science"}],"tags":[],"updatedAt":"2025-12-10T09:57:05+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-10 09:57:05","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8310362","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8310362","identity":"rs-8310362","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}