The invention discloses a fault energy region boundary recognition and feature extraction method based on an instantaneous spectral entropy and a signal noise energy difference. A time domain boundaryand a frequency domain boundary of an energy concentration region are determined effectively in the time-frequency power density spectrum of an underwater robot dynamic signal, wherein the total energy in the boundary is used as a fault feature and the mapping relationship between the extracted time-frequency energy fault feature and the fault degree is unique. And when the fault feature is usedfor constructing a fault sample to carry out thruster fault degree classification, the fault degree classification accuracy of the test sample is 100%.