Fault feature extraction and diagnosis method based on deep belief network

As the equipment is increasingly complicated, it is much more challenging to artificially extract and select fault features using expertise and signal processing technology. In addition, traditional shallow architectures, i.e., BP neural network, SVM, are not capable enough at learning the complex nonlinear relationships between equipment health status and its represented signals. Deep belief network (DBN) has unique advantages in feature extraction, high-dimensional and nonlinear data processing. Hence, a novel fault feature extraction and diagnosis method are proposed based on deep belief network. Deep neural network can be directly trained using original time domain signal and utilized for smart fault diagnosis. This proposed method can adaptively extract the fault features and automatically identify machinery health conditions, overcoming the dependence on massive signal processing technologies and expertise. Moreover, periodic time domain signals are not required, which makes this method with high applicability and generality. The effectiveness of the proposed method is validated using datasets from simulations and bearings. The diagnostic results prove that the proposed method is able to conduct fault feature extraction and diagnosis effectively under various operating conditions, fault locations and levels from the raw signals to obtain superior diagnosis accuracy.