Unsupervised contrastive learning state evaluation framework for early abnormal bearings

Abstract

Rotational machinery is indispensable in various equipment, making the early detection of its potential malfunctions a crucial area for research. Although data-driven methods show promise in anomaly detection, their practicality is hindered by the scarcity of abnormal data, since machinery usually operates normally and failures are rare and unpredictable. Unsupervised contrastive learning enables early anomaly detection without relying on abnormal samples, offering a promising solution. Typically, it generates positive pairs via data augmentation and considers all other samples as negative pairs. However, these augmentation strategies may be inappropriate for early anomaly detection, as mechanical faults are predominantly reflected in frequency-domain characteristics. To address these challenges, this article proposes an unsupervised anomaly detection contrastive learning (ADCL) state evaluation framework for bearing early anomaly detection, which employs two proxies for the clustering of representation vectors. Specifically, we propose a lightweight separable self-attention for the construction of the autoencoder. Then, pseudo-abnormal data are generated by using the fault feature injection methods. The soft contrastive learning loss takes the anchor as the clustering target for normal samples and simultaneously optimizes the hypersphere center and the decision boundary, the latter of which is learned based on the contrast between normal and pseudo-abnormal samples. Finally, the cosine distance of the latent representation from the hypersphere centroid is utilized to assess the state of the target object, thereby enhancing sensitivity towards abnormal changes in feature distribution. Experimental results on two bearing datasets show the effectiveness of the proposed ADCL framework.