Comparing Different Feature Extraction Methods in Condition Monitoring Applications

Artificial intelligence (AI) has been widely adopted in industrial applications, making machine learning (ML) on edge devices essential. However, the power consumption of edge devices often limits their computational capabilities. In this study, we aimed to address this challenge by utilizing feature extraction (FE) techniques, explicitly considering domain shift scenarios where the data characteristics may change in real-world applications. We introduced a simple but effective FE method, the time-frequency feature extractor (TFEx). By using AutoML and performing over 6000 cross-validations on multiple time series datasets, we compared TFEx with six other FE methods. Our results showed that TFEx was the most effective method in reducing the dimensions of raw data while maintaining accuracy. Additionally, the structure of TFEx makes it suitable for implementation on edge devices, and even a simplified version can be used in many cases without significant loss of accuracy.