Bayesian optimization and channel-fusion-based convolutional autoencoder network for fault diagnosis of rotating machinery

Deep learning methods are essential for the application of data driven technologies on fault diagnosis of rotating machinery. However, the generalization and performance of deep learning methods for fault diagnosis are highly dependent on the selection of hyper parameters and the design of network structure. To solve aforementioned challenge in fault diagnosis and obtain a powerful model, a convolutional network based on Bayesian optimization and channel fusion mechanism is newly developed. In the proposed network, a convolutional autoencoder network is firstly applied to extract the compressed features and reconstruct the input data. Then, the channel fusion mechanism is introduced to deduce the error of reconstructed input data. Next, a hybrid loss function is defined through summing mean square error and cross-entropy. Finally, a Bayesian optimization scheme is designed to optimize the hyper parameters of the designed network. The effectiveness of the proposed method is verified by a laboratory bearing dataset and an industrial bearing dataset, respectively. Five classical fault diagnosis methods are also tested as comparison. The experimental results indicate the proposed method can achieve the outstanding performance in fault diagnosis both in accuracy and efficiency.