Semi-supervised high-precision fault diagnosis in rotating machinery with limited labeled data

Abstract Accurate and fast fault diagnosis is the core to ensure the normal operation of rotating machinery. The high cost of manual fault sample labeling and the large variation of mechanical characteristics of different rotating machinery bring challenges to fault diagnosis of small-sample and zero-sample problems. It is difficult to solve these challenges via existing supervised methods. To this end, in this article, a fault diagnosis model based on Semi-supervised Data Enhancement Broad Learning System (SSDEBLS) is proposed. Specifically, a low-pass filter is applied into the noise reduce of the vibration signal, and the frequency domain features are extracted using the fast Fourier transform algorithm. With the extracted features, a semi-supervised learning method based on hypergraph and label propagation algorithm are used to obtain the pseudo-label of the unlabeled samples. Finally, the Data Enhancement Broad Learning System method is used to build a fault classifier by continuously feeding data into the model to iteratively update the parameters. The experimental results show that in the actual application of the fault diagnosis of rotating machinery, the 100% precision can be obtained from the same rotating machinery with only 5% of the labeled data as training data, which can effectively solve the small-sample problem, while 94% precision can be obtained from different rotating machinery, indicating that SSDEBLS can solve the zero sample problem on a new equipment. In addition, the proposed SSDEBLS demonstrates a rapid fault detection response owing to the hierarchical framework of the Broad Learning System.