Diagnosis of composite faults for complex industrial machinery: A label-assisted self-supervised clustering approach

Early detection of faults is crucial to ensure the effective operation of industrial machinery. Incipient faults are hard to detect because the fault signals are weak, lack sufficient labels, and there is interference caused by multiple sensor signals and a large amount of noise background. Once a fault occurs, it will usually lead to composite faults that are coupled together, seriously affecting the accuracy of fault diagnosis. To overcome these challenges, this paper proposes a Label-assisted Self-supervised Clustering (LASSC) approach, which consists of two stages. The first stage is to mine the coupling characteristics of multi-sensor signals, and the second stage is to judge the composite fault states. Due to significant differences in magnitude, some fault features are ignored when learning from multi-sensor signals. To cope with this, constraint boundaries-based normalization is employed. Unlabeled data are projected onto a weight matrix through feature clustering, and the composite fault features are mined through unsupervised clustering. With the constraint of the assistance labels, self-supervised clustering integrates information from all channels to diagnose composite faults. Experiments are conducted using the proposed LASSC approach on the CWRU dataset to validate the accuracy and computational efficiency of the model. In addition, a series of experiments are designed on the excavator dataset, such as sample imbalance and limited labeled data, to verify the adaptability of the model in practical industrial applications. In addition, the application of this method in excavator cases revealed the causes of composite fault through interpretability analysis, enabling more targeted maintenance recommendations for industrial machinery faults.