A fault identification method of hydraulic pump fusing long short-term memory and synchronous compression wavelet transform

Abstract

The hydraulic pump (HP) is the “power heart” of the hydraulic transmission system. It is widely used in high-pressure, high flow, and flow regulation situations, such as hydraulic presses and engineering machinery. Due to the harsh working environment and complexity of the structure, faults in HPs are often not detected and treated in time, resulting in some economic losses and even serious accidents. This research introduces a long short-term memory (LSTM) and visual geometry group (VGG) fused fault diagnosis approach for HPs. First, the one-dimensional non-stationary acoustic signal is transformed into the high-frequency two-dimensional time–frequency image through synchronous compressed wavelet transform, which serves as the input for the convolutional neural network. Second, VGG is introduced for the fusion of extracted features to increase the attention of important features and thus enhance the representation ability of the network. Third, LSTM is incorporated for the mining of time series. Finally, the fault diagnosis results are verified through the Softmax layer. To verify the performance of the proposed model, five types of faults are identified through experiments. As compared with other traditional diagnostic methods, the proposed model has superior accuracy and stability.