Lightweight rotating machinery fault diagnosis based on quadratic convolutional neural network and evidence fusion of multi-source sensor information

Abstract Rotating machinery is the core component of various modern equipment. Timely and accurate fault detection is of great significance to ensure the safe operation of mechanical equipment and the stability of production efficiency. Deep learning-based methods have triggered a boom in the field of machinery fault diagnosis. However, due to factors such as noise interference in industrial applications, an increase in accuracy is often accompanied by an increase in computational cost, leading to difficulties in deploying the model in engineering applications. In order to find a balance between robustness and complexity, this paper proposes a lightweight intelligent fault diagnosis framework called (QSMC-EFN). Firstly, a new nonlinear quadratic neural network is used to extract features initially and then combined with a multiscale separable convolutional block to extract multiscale features of the vibration signal. Different features from multiple scales are subsequently fused using a statistical Dempster-Shafer (DS) theory of evidence, fed into a lightweight feature extraction layer consisting of multi-scale separated convolution and broadcast attention mechanisms to gain further the ability to extract more critical features. Through experiments on the planetary gearbox and variable speed bearing datasets, it is verified that the proposed fault diagnosis system ensures effectiveness while also achieving a lightweight model.