Orthogonal Latent Compression for Streaming Anomaly Detection in Industrial Vision

Han Gao, Huiyuan Luo, Fei Shen, Zhengtao Zhang

Abstract: Although existing industrial anomaly detection methods perform well, they are trained on offline datasets collected in advance and remain unchanged once the training is complete. Simultaneously, they assume that the data is static without any drift. However, data in industrial scenarios, especially in sequential assembly lines, usually arrives dynamically in streams and suffers from data drift over time, such as lighting variations and digital noise. The offline training paradigm and inability to dynamically update of existing methods are inconsistent with the data characteristics of streaming dynamics, and it is also difficult to quickly adapt to streaming data drift. To this end, we propose a streaming anomaly detection method that can not only learn dynamically based on the production line data stream, but also adapt to data drift as quickly as possible by effectively utilizing a small amount of drifted training data. The core idea of the proposed method is to compress the features into an orthogonal latent space and constrain the features with nearest reconstruction and maximum separability to maximally capture the normal patterns of the data. Extensive experiments on three real industrial datasets demonstrate our method’s excellent performance in stream anomaly detection tasks and rapid adaptability to data drifts. Additionally, our method has lower modeling complexity and higher computational efficiency. It also achieves state-of-the-art performance in offline industrial image anomaly detection and localization tasks. Source code will be released upon paper acceptance.