Improving Anomalous Sound Detection by Distance Matrix-Based Visualization of Measurement Flaws

Abstract

Although recent DNN-based methods have improved the performance of anomalous sound detection systems, it is still difficult to deploy a system in a real environment without performance degradation. This is often due to measurement flaws such as sensor variability, poor setup, or environmental noise. Since such adverse effects are difficult to model by machine learning, a practical approach to this issue is for humans to identify such flaws and correct them. To this end, we propose a method to visualize measurement flaws as a heatmap based on the distance matrix of the samples in the dataset. This method is designed to find unexpected flaws in the measurement process. Using this method, we were able to identify measurement flaws of anomalous sound detection systems in real production lines. The robustness of anomalous sound detection can be improved by correcting the flaws found by our method.