FAB: Few-Shot Adversarial Learning Formed Decision Boundary for Industrial Anomaly Detection

Abstract

Few-shot anomaly detection (FSAD) plays a pivotal role in the condition monitoring of industrial equipment. Adapting anomaly detection to few-shot settings is essential for improving the generalization capability of models across diverse industrial scenarios. However, this task remains highly challenging, since discriminative representations are difficult to learn from only a limited number of normal samples. To address this issue, this study proposes an end-to-end adversarial training framework for industrial FSAD, aiming to establish robust decision boundaries under data-scarce conditions. The primary contributions include three parts: dual-purpose label utilization, a diversity enhancement mechanism, and discriminative feature learning. In the first part, the generator exploits the dual semantics of labels to simultaneously encode sample authenticity and machine operating status. By synthesizing pseudosamples associated with adaptive soft labels, the proposed framework is able to construct robust decision margins around the sparse distribution of normal data. In the second part, a diversity enhancement mechanism is introduced by integrating a discrepancy regularizer and a mutual information estimation network. This design promotes the generation of diverse pseudosamples while preserving semantic consistency with the underlying patterns of genuine normal signals. In the third part, the discriminator incorporates circle loss within a metric learning framework to optimize the topology of the latent space, thereby learning more discriminative feature representations and establishing well-defined decision boundaries. Extensive experiments on two real-world industrial datasets demonstrate the effectiveness of the proposed method. Comparative evaluations against state-of-the-art approaches further confirm its superior performance.