A new deep auto-encoder method with fusing discriminant information for bearing fault diagnosis

•We propose a new deep auto-encoder method with fusing discriminant information.•We presents a gradient descent method to optimize the new model.•This method can improve features’ representative ability with no fine-tuning operation.•This method can improve model’s numerical stability on insufficient amount of training data.•We find a symmetric structure among multiple fault conditions in form of relatedness matrix. In recent years, deep learning techniques have been proved a promising tool for bearing fault diagnosis. However, to extract deep features with better representative ability, how to introduce discriminant information about different fault types into the deep learning model is still challenging. Moreover, as deep learning techniques heavily rely on mass of measuring data, relatively small amounts of data may cause over-fitting and reduce model stability as well. To solve such problems, a new deep auto-encoder method with fusing discriminant information about multiple fault types is proposed for bearing fault diagnosis. First, a new loss function is designed by introducing structural discriminant information. Specifically, to improve the feature’s representative ability, a new discriminant regularizer is designed in the loss function by using maximum correlation entropy. And to represent the structural information among multiple fault types, a relation matrix for fault types is introduced, then a new regularizer with a symmetric constraint on this matrix is constructed. Second, a gradient descent method is provided to optimise this loss function, and the optimal deep features, as well as fault relatedness, are learned simultaneously. Experimental results on CWRU and IMS bearing data sets show that, compared to several state-of-the-art diagnosis methods, the proposed method can effectively improve the diagnostic accuracy with acceptable time efficiency. And the results on the Kruskal–Wallis Test indicate the proposed method has better numerical stability.