Accurate and reliable muscle fatigue detection plays an important role in rehabilitation robot and human-machine cooperation. Considering frequency change of surface EMG signal after muscle fatigue, this paper proposes a muscle fatigue detection method based on improved band energy entropy, which uses fractional weighted S-transform energy entropy and fractional weighted wavelet packet decomposition energy entropy to detect muscle fatigue, and compares them with sample entropy, dispersion entropy and fractional fuzzy dispersion entropy. A total of 32 subjects participated in the muscle fatigue experiment. The results show that the improved frequency band energy entropy performs optimally in terms of noise robustness, data stability, and anti-interference performance. The improved S-transform energy entropy achieved the highest sensitivity to static muscle fatigue among all the compared complexity algorithms. The absolute slope of complexity variation with muscle fatigue is the largest, with an average value of -15.946 5×10-3. The improved wavelet packet energy entropy (-3.100 3×10-3) detection algorithm can also detect muscle fatigue well, and achieves comparable performance to fractional fuzzy dispersion entropy (-2.602 6×10-3). Among all algorithms, this algorithm exhibits the lowest time consumption, second only to the dispersion entropy algorithm. In the dynamic muscle fatigue test, the improved S-transform energy entropy also performs optimally. The improved frequency band energy entropy provided an effective new tool for analysis of sEMG signal complexity.