Learning to compare: A pseudo-fault sample-guided relational network for machinery fault detection and localization

Deep learning models have demonstrated remarkable potential in machinery fault detection. Nonetheless, most approaches rely on the availability of faulty samples, which are often unattainable in actual engineering due to the expensive data acquisition costs and rare fault occurrences. To address this challenge, a new pseudo-fault sample-guided relational network is proposed for fault detection and localization of machinery. First, a multiscale residual learning module is designed to extract robust features from input vibration signals while suppressing irrelevant interference. Then, the positive and negative feature pairs are created to characterize the relations between normal patterns and fault patterns. Finally, a capsule relation metric module is built to capture the fine-grained relations between the feature pairs and output relation scores, thereby identifying the health state of machinery. Specially, a pseudo-fault sample repository is constructed to guide the network in learning to compare the differences between normal state samples and other state samples. Extensive experimental results on three datasets demonstrate that our method can achieve compelling detection performance even in the absence of real faulty samples.