Normality Aggregation and Abnormality Separation Contrastive Learning for Mechanical Anomaly Detection

Accurate anomaly detection (AD) is essential for the safe operation of high-end equipment. Intelligent detection approaches have the capacity for automatic abnormality discovery from big data. However, uncontrollable unsupervised training is prone to cause model collapse, at which inliers and outliers become indistinguishable. Contrastive learning (CL) provides a practicable solution via instance contrast. However, the objective of classical contrastive loss contradicts the ideal sample distribution for AD. In light of this, a normality aggregation and abnormality separation CL framework (NA2SCL) is constructed for mechanical AD, which specifies the contrastive objective for pseudo inliers and pseudo outliers, respectively, thereby learning an inlier-clustering and outlier-separate latent space for AD. Specifically, identity-preserving and identity-shift transformations are designed to generate virtual normal and abnormal samples of vibration signals. Then, the contrastive objective is integrated via attracting virtual inliers, repelling virtual outliers, and discriminating transformation identity. A soft weighted mechanism is further performed on pseudo positive samples to eliminate the negative aggregating effect. Moreover, the minimum cosine distance is adopted to determine the detection threshold and compute anomaly scores. Experimental results on single-condition, multicondition, and varying-condition cases demonstrate that the framework can capture the ideal distribution suitable for AD.