An Improved Conv-LSTM Method for Gear Fault Detection.

Reducing the occurrence of gear failures and extending their service life is a vital issue in industrial production. To solve the problem that the method of gear fault detection with Convolution Neural Network (CNN) is difficult to extract the temporal features of the vibration data, an improved Convolutional-LSTM (Conv-LSTM) gear fault detection method was proposed. First, the raw data was fed into the convolutional layer, followed by the pooling and LSTM layers. A batch normalisation layer (BN) was added after the convolutional layer to speed up convergence. Second, to reduce the complexity of the model, a Global Maximum Pooling layer (GMP) was used to replace the flattened layer, and the Hinge functions are used as loss functions. Finally, classification is carried out by the Softmax classifier. The overall accuracy of model architecture could reach 99.64% on the University of Connecticut gear fault dataset. The results show that the proposed method is effective and can meet gear fault diagnosis's accuracy and timeliness requirements.