An Investigation into Mechanomyography for Signal Extraction and Classification of Human Lower Limb Activity

To mitigate the difficulties associated with the extraction of Mechanomyography (MMG) signals from raw Acceleration (ACC) data and the subsequent classification of human lower limb activities based on MMG signals, the Feature Mode Decomposition (FMD) algorithm has been utilized for the isolation of the MMG signal. Simultaneously, surface Electromyography (sEMG) signals were recorded to perform correlation analyses, thereby validating the effectiveness of the extracted Mechanomyography (MMG) signals. The results demonstrate that the envelope entropy derived from the FMD was the lowest among the observed values, and the composite signal obtained via FMD displayed the highest correlation with the sEMG signal. This indicates that FMD is capable of efficiently isolating the MMG signal while maintaining the maximal quantity of muscle contraction data. To address the challenge of classifying human lower limb activities, a comprehensive feature extraction procedure was implemented, resulting in the derivation of 448 unique features from multi-channel mechanomyography (MMG) signals. Subsequently, Kernel Principal Component Analysis (KPCA) was employed to diminish the feature set's dimensionality. This was succeeded by the deployment of a Temporal Convolutional Network integrated with an Attention mechanism (TCN-Attention) to train the classification model. Additionally, the Sine-Cosine Northern Goshawk Optimization (SCNGO) algorithm was leveraged for optimization purposes. The findings indicate that FMD exhibited the minimum envelope entropy value of 8.13, concurrently attaining the maximum correlation coefficient of 0.87 between MMG and sEMG signals. Significantly, the SCNGO-TCN-Attention model demonstrated superior classification accuracy, attaining an exceptional accuracy rate of 98.44%.