Comparative Analysis of Time, Frequency, and Wavelet Features for Tool Condition Monitoring Using Machine Learning under Cross-Tool Validation

Abstract

Tool condition monitoring (TCM) is critical for ensuring high machining quality and preventing catastrophic tool failures in smart manufacturing. While data-driven approaches increasingly rely on complex machine learning architectures and hybrid feature extraction (combining time, frequency, and time-frequency domains), the specific computational and accurate contributions of each domain remain poorly understood—a critical bottleneck for real-time edge-computing deployment. This study presents a systematic benchmarking analysis comparing time-domain, frequency-domain (FFT), wavelet-based, and hybrid feature sets for tool wear classification using the standard PHM 2010 Data Challenge dataset. To ensure industrial relevance, model generalization is rigorously assessed through cross-tool validation using a robust leave-one-out approach. Operating under strict computational efficiency constraints, a Decision Tree and Random Forest framework is employed. The results demonstrate that simpler time-domain features perform exceptionally robustly, with the Random Forest model achieving 92.3% classification accuracy. Crucially, hybrid features offer only marginal improvements at the expense of high computational overhead, while standalone frequency and wavelet domains prove less effective under cross-tool constraints. Furthermore, our analysis reveals two critical pitfalls for practical TCM design: feature selection techniques fail to improve accuracy, and standard segment normalization severely degrades performance by inadvertently removing vital signal magnitude information linked to tool wear. These insights provide concrete, computationally efficient guidelines for deploying lean, high-accuracy machine learning models on real-time industrial edge devices.