Sinc-Based Multiplication-Convolution Network for Small-Sample Fault Diagnosis and Edge Application

Data-driven intelligent diagnosis models need massive monitoring data to train themselves for satisfactory recognition performance. Nevertheless, in many industrial practices, collecting fault data is often expensive and time consuming, which leads to small-sample fault diagnosis becoming a valuable research hotspot. Inspired by multiscale mode characteristics and feature enhancement learning, this study proposes a Sinc-based multiplication-convolution network (SincMCN) for intelligent fault diagnosis under small-sample conditions. It works in the frequency domain and consists of only three layers, including a feature multiplication separator, a feature convolution extractor, and a classifier. In the feature separator, a series of Sinc-based multiplication filtering kernels (SincMFKs) is directionally designed for improving the utilization of fault-sensitive features of spectrum samples. The dot products between SincMFKs and spectrum samples are stacked into activated mode spectrum images (AMSIs) with rich fault-related features retained. Subsequently, a 2-D convolutional layer is employed as the feature convolution extractor to abstract high-level feature representations from the AMSIs. Finally, a fully connected layer is used as the classifier for achieving a fast and precise condition identification. Especially, to better train the proposed SincMCN, an orthogonal loss term is presented to help the feature multiplication separator extract the discriminative fault knowledge as much as possible. Compared to five end-to-end models, experimental results show that SincMCN has better diagnosis accuracy and stronger competitiveness for few-shot diagnosis. In particular, analytic SincMFKs not only cut down the model parameters for edge diagnosis, but also shows powerful application potentials for online monitoring of rotating machinery.