GBLN: a new generative broad learning network for rolling bearing fault diagnosis

Broad learning system (BLS) is an effective machine learning framework that expands the breadth of the network by increasing the number of neurons in each layer. However, traditional BLS uses the squared loss function, susceptible to adverse effects of noise. Therefore, a new generative broad learning network (GBLN) is proposed. In GBLN, a wave constraint term is designed to directly simulate the waveform features of the data, achieve accurate mining of state features, and maintain good stability even in the presence of noisy mutations. Meanwhile, the wave constraint term can lead the model to the optimal solution faster, because it provides a clearer optimisation direction for the model, which is directly optimised for the waveform features. In addition, GBLN exhibits a robustness by providing a quantified metric for the excessive penalty for preventing bias, thus maintaining a balanced approach to managing noise. The rolling bearing fault diagnosis experimental results show that compared with the comparative method, the proposed GBLN method improves the recognition accuracy by 2%−8%, especially under strong noise interference conditions.