Abstract: Aiming at the problem of extracting mechanomyography (MMG) signals from raw acceleration data and classifying the six common lower limb activities based on MMG signals, a MMG signal filtering and denoising method based on the feature mode decomposition (FMD) algorithm, a time-frequency domain feature extraction method for MMG signals, a feature set dimensionality reduction method based on kernel principal component analysis (KPCA), and a lower limb activity classification method based on a temporal convolutional network (TCN) incorporating an attention mechanism were proposed. Analysis shows that the envelope entropy value of the MMG signal obtained through the FMD algorithm was the smallest, only 8.13, indicating that FMD can efficiently extract MMG signals and remove random noise. Additionally, compared to the original signal, the power spectral density of the MMG signal obtained via the FMD algorithm showed a significant reduction in the low-frequency band where motion artifacts resided, while the reduction was minimal in the frequency band where MMG signals were located. This demonstrates that FMD effectively removed artifact interference while retaining the majority of MMG signal data. For the classification of the six common lower limb activities based on MMG signals, a comprehensive feature extraction strategy was adopted, extracting 448 features from 16-channel MMG signals and reducing the feature set dimensionality through KPCA. A TCN incorporating an attention mechanism was also constructed to train the classification model. Furthermore, by applying the sine-cosine northern goshawk optimization (SCNGO) algorithm to optimize the network’s hyperparameters, analysis reveals that the proposed model achieves outstanding classification accuracy, reaching 98.4%.