Deep Convolutional and LSTM Recurrent Neural Networks for Rolling Bearing Fault Diagnosis Under Strong Noises and Variable Loads

To research the problems of the rolling bearing fault diagnosis under different noises and loads, a dual-input model based on a convolutional neural network (CNN) and long-short term memory (LSTM) neural network is proposed. The model uses both time domain and frequency domain features to achieve end-to-end fault diagnosis. One-dimensional convolutional and pooling layers are utilized to extract the spatial features and retain the sequence features of the data. In addition, an LSTM layer is employed to extract the sequence features. Finally, a dense layer is applied for fault classification. To enhance recognition accuracy under different noises and loads, three techniques are applied to the proposed model, including taking time-frequency domain signals as input, using the CNN-LSTM model, and adopting the mini-batch and batch normalization methods. The Case Western Reserve University and Drivetrain Diagnostics Simulator data sets are used to construct experiments under different conditions, including varying loads and different noises. The proposed model can achieve a high fault recognition rate under variable load and noise conditions as well as satisfactory anti-noise and load adaptability.