Abstract: Feature extraction and intelligent recognition of the vibration signals of pump turbines are significant to reliable and safe operation of a pumped storage power station. Due to its complicated operational conditions, a pump turbine in operation can create a large number of physical sources that excite its vibrations, and the frequency components of the vibration signals are quite complicated. The traditional methods suffer a poor accuracy of feature extraction from a complicated vibration signal. To improve the accuracy, this paper describes a new model of feature extraction and intelligent recognition of the vibration signals, based on the variational mode decomposition (VMD), bubble entropy (BE), and long short-term memory (LSTM) neural network. First, this method analyzes the vibration signal using VMD and obtains several modes. Then for each mode, its BE value is calculated and a BE eigenvector is constructed. Finally, the eigenvectors of the vibration signal are trained and recognized using a LSTM neural network. We have verified the method against the complicated vibration signals measured at the top cover of a pump turbine at the Pushihe pumped storage station, and achieved a signal recognition accuracy of 97.87%, indicating its important engineering application value.

Key words: pump turbine, vibration signal, variational mode decomposition, bubble entropy, long short-term memory