Abstract: To accurately identify the muscle fatigue levels of miners engaged in long-term operations in coal mines, reduce human-factor risks, and precisely prevent safety accidents caused by miner fatigue, a muscle fatigue level recognition method for miners based on multi-scale nonlinear coupling features of surface electromyography (sEMG) signals is proposed. First, subjects wearing EMG sensors performed bicep curl movements with dumbbells of different masses (4, 6, 8, 10 kg) to simulate miners’ daily manual lifting tasks, thereby completing the collection of EMG signals corresponding to different muscle fatigue levels. The acquired signals were preprocessed for noise removal using a 4th-order Butterworth bandpass filter with a cutoff frequency of 50 Hz. Next, an optimized multivariate variational mode decomposition (MVMD) algorithm was employed to achieve mode-separation decomposition of sEMG signals across different time–frequency scales. Principal component analysis (PCA) was further applied to select the principal intrinsic mode functions (IMFs) to extract the multiscale sample entropy (MSE) of the main IMFs as a nonlinear coupled feature vector. Finally, a classification model based on a weighted long short-term memory network (WLSTM) was constructed. The multi-scale nonlinear coupled feature vectors were input into the WLSTM model for training and recognition, ultimately achieving classification and identification of muscle fatigue levels during miners’ manual lifting operations. Experimental results on a self-built dataset showed that the WLSTM model achieved 96.8% in accuracy, 99.5% in recall, 96.7% in F₁-score, and 98.1% in average recognition precision. Compared with the standard LSTM model, WLSTM improved these metrics by 4.2%, 4.6%, 3.0%, and 3.9%, respectively. Additionally, ablation experiments conducted by separately removing the MVMD and PCA modules resulted in average reductions of 6.4%, 8.2%, 6.4%, and 7.6% across the four performance metrics, confirming the effectiveness of incorporating MVMD and PCA modules in enhancing the performance of the WLSTM classification model.