A new multi-layer adaptation cross-domain model for bearing fault diagnosis under different operating conditions

Abstract Bearing faults under different operating conditions often cannot be diagnosed by models trained under a single operational condition. Additionally, the extraction of domain-invariant features in domain adaptation (DA) algorithms is also a challenge. To address the aforementioned issues, a multi-layer adaptation model based on an improved sparse autoencoders (SAEs) and dual-domain distance mechanism (ISAE-DDM) is proposed. First, the feature extraction capability of traditional SAEs is enhanced by a strategy that combines mean squared error with mean absolute error. Subsequently, the features of data under multiple hidden layers are extracted by the ISAE. Then, the distribution discrepancy between the source domain and target domain are measured by a dual-domain distance approach that combines Wasserstein distance with multi-kernel maximum mean discrepancy. Then, the domain distance loss under each hidden layer is assigned different weighting parameters. Finally, a joint metric DA mechanism across multiple hidden layers is constructed to achieve a better domain alignment. The performance of the proposed method is demonstrated by two different bearing experiments. Moreover, this model exhibits higher stability, and generalization capabilities compared to other methods.