A few-shot fine-grained fault diagnosis framework based on pseudo-supervised task representation

Traditional fault diagnosis methods are typically coarse-grained, with clear distinctions between faults. These methods necessitate extensive labeled data for effective training and lack the capability to diagnose at more fine-grained granularity with limited data. Furthermore, labeling data is labor-intensive and, in fact, the data is usually unlabeled in many complex engineering systems. To tackle these challenges, we propose a novel few-shot fine-grained fault diagnosis framework based on pseudo-supervised task representation, PSTR, which aims to generate pseudo-labels and assign them to the original samples, then samples with high information entropy are selected to construct meta-tasks. Firstly, feature representations are crafted by a network utilizing contrastive learning, thereby enhancing class separation. To construct pseudo-supervised tasks, clustering is applied to the feature representations, leading to the generation of pseudo-labels for the original samples. Subsequently, a perturbation strategy is introduced to effectively cope with the weak diversity of the task, obtaining meta-representations enriched with the fine-grained features provided by the large clusters and with the variability given by the small clusters to complete fault diagnosis. Furthermore, during the meta-training process, clusters containing high information are selected by entropy and low-scoring samples are filtered out using an evaluation model. It achieves an average accuracy of 83.37 % on both public datasets in diagnosing unseen fine-grained faults, outperforming popular methods.