[5K3-OS-38b] OS-38

2:45 PM - 3:00 PM JST(5:45 AM - 6:00 AM UTC)

[5K3-OS-38b-04] A Visualization Approach for Dataset Quality Adaptable to Continuously Varying Measurement Error Factors

〇Kosei Ozeki¹, Nobuaki Tanaka¹, Toshiyuki Kuriyama¹ (1. Mitsubishi Electric Corporation)

Keywords:

Quality Control,Anomalous Detection,Product Inspection,Predictive Maintenance,Data Visualization

Recent advances in machine learning have accelerated the development of anomaly detection techniques utilizing acoustic and vibration signals. However, in practical operation, detection accuracy often degrades compared to simulation environments due to domain shifts, such as variations in measurement systems. As a practical solution to this issue, identifying the causes of domain shifts and compensating for measurement conditions is an effective approach. In this study, we propose a linear-regression-based analysis that combines observed signals with continuous sub-labels—metadata unrelated to device malfunctions, such as background noise levels or recording dates—to visualize and quantify measurement deficiencies. This approach yields actionable insights that can help improve measurement quality in real-world settings.