Anomalous Sound Detection for Slider Machine Using Contrastive Learning Under Domain-Shifted Conditions

In this paper, we use contrastive learning to address anomalous sound detection under domain shifts for the slider industrial machine. Contrastive learning is an innovative approach that uses augmentation techniques to learn the invariant attributes of samples. Our work involves assessing the effect of each of several augmentation methods on the slider machine and fine-tuning the parameters of the most effective method. We consider domain-shifted data to better simulate real-world cases, which tend to have characteristics that differ from those seen in the training data. Based on our experiments, we propose the preferred augmentation technique and parameters for slider domain-shifted data. The results show that using slider-specific techniques and parameters improves performance on slider domain-shifted data by 3.0% when compared to using more generic techniques and parameters.