Random wavelet kernels for interpretable fault diagnosis in industrial systems

Deep learning is a powerful method for fault diagnosis, but its "black-box" nature raises concerns in critical applications. This paper presents an interpretable, lightweight method combining random convolution kernel transformation (ROCKET) with wavelet kernels, which offer systematic time-frequency analysis and intuitive insights. Principal component analysis (PCA) is used to extract relevant patterns, forming a health indicator that guides maintenance decisions. A case study on linear actuator fault diagnosis demonstrates the method's balance of interpretability and computational efficiency, making it a valuable tool for reliable asset health monitoring in resource-limited settings.