Intelligent fault diagnosis of hydraulic piston pump combining improved LeNet-5 and PSO hyperparameter optimization

• Manually adjusting the hyperparameters of LeNet-5 model is uncertain. • The Adam, Adagrad and Adadelta optimizers are utilized in Improve-LeNet-5 model. • Improve-LeNet-5 model can classify and identify the fault state of piston pump. • PSO is employed to optimize the hyperparameters of Improve-LeNet-5 model. • PSO-Improve-CNN model possesses high accuracy and strong robustness. The hydraulic axial piston pump is the power heart of the hydraulic transmission system in aerospace equipment and industrial filed. Its stable operation will directly affect the safety and reliability of the whole equipment. It is very significant to realize its health status monitoring and intelligent fault diagnosis. In view of the restrictions of traditional mechanical fault diagnosis in the dependence on a large number of signal processing technologies and expert diagnosis experience, as well as the time-consuming of data preprocessing, it is very meaningful to explore new ideas and methods to realize intelligent fault diagnosis of hydraulic piston pump. Based on the standard LeNet-5 model, the kernel size and kernel number are improved, and the batch normalization layers are added to the network architecture. Based on the Improve-LeNet-5 model, the recognition accuracy is chosen as the target value of the fitness function, the hyperparameters of the Improve-LeNet-5 model are automatically optimized via particle swarm optimization (PSO), including the learning rate, the number of convolution kernels, batch size, and the number of neurons in the fully connected layer. Finally, the PSO-Improve-CNN diagnostic model is constructed. And it is employed to classify and identify five signals data of hydraulic piston pump: normal state, swash plate wear, sliding slipper wear, loose slipper and center spring failure. Research result indicates that the recognition accuracy of PSO-Improve-CNN model can reach 98.71%, and the highest recognition accuracy can reach 99.06%, which are respectively higher than the standard LeNet-5 and Improve-LeNet-5 about 5.23% and 2.25%. By comparing with AlexNet, VGG11, VGG13, VGG16, and GoogleNet, the PSO-Improve-CNN model presents the highest diagnostic accuracy, less time in training and testing, and greater robustness. The comprehensive performance of the proposed model is demonstrated to be much stronger.