High-frequency time series data in manufacturing creates computational challenges for machine learning applications. This study compares four feature extraction methods (PAA, PCA, catch22, tsfresh) for classifying surface-based defects in screw connections. Using 12,500 tightening operations across eight defect classes, we evaluate these methods based on classification performance, computational efficiency, and memory usage. Results show tsfresh variants achieve the highest accuracy (up to 11% improvement over benchmark), while PAA offers the best balance between performance and computational efficiency. For minimal resource usage, catch22 maintains performance within ± 6% of benchmark using only 24 features. These findings provide concrete guidelines for selecting feature extraction techniques based on application requirements.