A New Deep Fusion Network for Automatic Mechanical Fault Feature Learning

Mechanical fault diagnosis is essential in ensuring the safety of production and economic development. In the field of fault diagnosis, deep learning has been extensively used due to its excellent feature learning ability. However, it still suffers from several issues; for example, 1) simultaneous requirements of features from multiple aspects, including sparsity and robustness, are hardly met due to the limited feature learning ability of a single model, and 2) most methods deal with preprocessed signals instead of original time domain signals because of the noise interference and deficiency of a single model. To solve these problems, this study proposes a new deep fusion network for fault feature learning, which combines two types of deep learning models, namely, sparse autoencoder and contractive autoencoder, which are respectively applied to enhance features’ sparsity and robustness and thereby guarantee the representativeness of extracted features and gain strong anti-interference ability. Consequently, fault diagnosis with original time domain signals can be realized. Bearing and gearbox fault diagnosis experiments are conducted to verify the performance of the presented network. Results show that the diagnosis accuracies for two cases are higher than those of networks based on single contractive autoencoder and sparse autoencoder. These results demonstrate that the proposed fusion network has superior feature learning ability relative to single model networks and can deal with original time domain signals by simultaneously enhancing features’ sparsity and robustness.