The Application of a Lightweight Domain-Adversarial Neural Network in Bearing Fault Diagnosis

Smart manufacturing is a rising research hotspot in Industry 4.0 era, in which deep learning has been getting more and more involved with the theoretical research of fault diagnosis in recent years. However, the deficiency of well labelled data in practical scenarios and unexpected massive weights in a deep learning model have seriously hindered the implementation of deep learning. Aiming at solving these problems, transfer learning strategies and lightweight structures are adopted, a novel lightweight deep learning model with transferring strategy call Lightweight Domain-Adversarial Neural Network (LDANN) is proposed in this paper. Efficient bottleneck residual blocks are adapted and embedded into the model to construct a lightweight feature extractor, and adversarial mechanism is implemented between the label predictor and the domain classifier to complete the domain adaptation. The model has been verified with a bearing dataset from CWRU and it is proven outperforms the comparison models. We further discuss the performance of LDANN and raise a future research orientation.