Practical Test-Time Domain Adaptation for Industrial Condition Monitoring by Leveraging Normal-Class Data

Machine learning has driven significant advancements across diverse domains. However, models often experience performance degradation when applied to data distributions that differ from those encountered during training, a challenge known as domain shift. This issue is particularly relevant in industrial condition monitoring, where data originate from heterogeneous sensors operating under varying conditions, hardware configurations, or environments. Domain adaptation is a well-known method to address this problem; however, the proposed methods are not directly applicable in real-world condition monitoring scenarios. This study addresses such challenges by introducing a Normal-Class Test-Time Domain Adaptation (NC-TTDA) framework tailored for condition monitoring applications. The proposed framework detects distributional shifts in sensor data and adapts pretrained models to new operating conditions by exploiting readily available normal-class samples, without requiring labeled target data. Furthermore, it integrates seamlessly with automated machine learning (AutoML) workflows to support hyperparameter optimization, model selection, and test-time adaptation within an end-to-end pipeline. Experiments conducted on six publicly available condition monitoring datasets demonstrate that the proposed approach achieves robust generalization under domain shift, yielding average AUROC scores above 99% and low false positive rates across all target domains. This work emphasizes the need for practical solutions to address domain adaptation in condition monitoring and highlights the effectiveness of NC-TTDA for real-world industrial monitoring applications.