Beyond deep features: Fast random wavelet kernel convolution for weak-fault feature extraction of rotating machinery

Extracting meaningful and significant signal features is the eternal theme of mechanical fault diagnosis. Recently deep learning has promoted the success of automatic feature extraction in various engineering fields. However, a network must be deep enough to extract hidden fault information, limiting its interpretability, efficiency, and generalization. This paper argues that even shallow features can surpass deep features. To this end, one-layer Random waVELet (RaVEL) kernel convolution is proposed to extract shallow features for weak-fault diagnosis (WFD) of rotating machinery. First, a large number of diverse convolution kernels are fitted by randomizing the scaling and shifting parameters of multiple wavelet bases. Then, by randomly convolving the input signal with kernels, a huge number of convolved outputs are obtained. Finally, global pooling and secondary activation are to evaluate and sift out useful features. Due to the diversity of extracted features, this structure requires no training or tuning of extraction model, resulting in reliable and efficient inference. Additionally, this paper theoretically reveals the correlation of inner products in random feature space and signal space. Experiments show that RaVEL outperforms the state-of-the-art (SOTA) methods in WFD tasks, and it works 28 times faster on CPU than the SOTA ResNet-based method on GPU. Codes are available at https://github.com/fyancy/RaVEL.