In order to investigate the dynamic coupling mechanism between central regulation and peripheral responses, a three-dimensional electromyography (EMG) fatigue evaluation model for upper limb multi-joint movements was developed using a brain-computer interface (BCI) training approach. Surface EMG signals from four channels, including the brachioradialis muscle, were recorded from 30 participants during four types of upper limb movements (fist clenching, wrist flexion, wrist elevation, and elbow flexion). A comprehensive evaluation system integrating subjective perception, peripheral physiology, and central performance was constructed by combining the fatigue scale-14 (FS-14) subjective fatigue scale with common spatial pattern classification accuracy.The results are as follows. Significant progressive decreases are observed in time-domain features across three stages (p<0.05), although compensatory enhancement is found in specific channels. Dynamic changes in high-frequency energy distribution are demonstrated by time-frequency analysis, which confirms the preferential recruitment mechanism of type IIb fast-twitch muscle fibers. The optimal contributions are determined through entropy-based weighting for subjective fatigue, system performance, and peripheral physiology. The composite fatigue index exhibits significant progressive increases across stages (p<0.01). It is indicated that the model effectively integrates multi-dimensional fatigue signals, providing new insights for dynamic monitoring and rehabilitative interventions in neuromuscular fatigue.Experimental data indicate that greater classification accuracy fluctuations (maximum variation 0.13) are demonstrated in movement execution tasks compared to motor imagery tasks. It is suggested that BCI system designs require task-specific fatigue intervention strategies.