Abstract: Aimed at the nonstationary and nonlinear vibration signals of hydropower units, a fault identification method is constructed combining the IMF energy moment and the bi-direction long short-term memory neural network (BiLSTMNN). First, we use the complementary ensemble empirical mode decomposition (CEEMD) method to process the normal and fault vibration signals from a hydropower unit, and obtain the intrinsic mode functions (IMF) and the residual components with different frequencies. Then, the IMF energy moment is calculated and used as the fault feature. And we use the fault features as inputs and the fault categories as outputs, and train BiLSTMNN to obtain a fault identifier for the unit. The operation state of the unit can be identified as a normal or specific fault type by combining this identifier with the IMF energy moment characteristics of the real-time signals. Finally, two sets of comparative experiments are designed based on the sample data collected on a rotor test stand and from the on-site observation of a hydropower unit. The results show our new method is effective in mining signal features and can achieve a high accuracy of fault diagnosis.

Key words: intrinsic mode function, energy moment, bi-directional long short-term memory neural network, fault diagnosis, vibration signal of hydropower unit