NorCLR: A Normality-Aggregated Contrastive Learning Framework for Mechanical Anomaly Detection

Anomaly detection is critical for the operating safety of mechanical equipment. Existing unsupervised training paradigm of anomaly detection may suffer from model collapse. The emergence of contrastive learning provides a practicable solution, however, the goal of traditional contrastive loss contradicts with the ideal distribution of samples in anomaly detection. To this end, a normality-aggregated contrastive learning framework is proposed for mechanical anomaly detection. First, we design two forms of transformation, i.e., identity-preserving and distribution-shift transformation, to generate virtual positive and negative samples of vibration signals. Then, the vanilla contrastive loss is modified with the ideal prior distribution of normal and abnormal samples, which aims to attract inliers and repel outliers. Besides, a soft weighted mechanism is applied on normal samples to avoid negative aggregation of false positive samples. Finally, the minimum cosine distance between the tested sample and all training samples are adopted as the anomaly scores. Experiments on single-condition and multi-condition scenarios validate the superiority of the proposed framework for anomaly detection.