SKND-TSACNN: A novel time-scale adaptive CNN framework for fault diagnosis of rotating machinery

Abstract

Recently, deep learning models represented by convolutional neural networks (CNN) have played an increasingly important role in rotating machinery fault diagnosis (FD). However, complex working conditions and uncertain noise interference pose severe challenges to the accuracy and anti-noise abilities of CNN-based FD models. To address these, based on a selective kernel block (SKB), a novel time-scale adaptive CNN (TSACNN) and neural network denoiser (NND) are constructed in this study, and an FD framework with strong generalization and anti-noise abilities is established through their deep fusion. Based on the end-to-end idea, the kernel adaptive capability of the SKB enables the traditional CNN to dynamically extract the time-domain information of different-scales hidden in signals, thus enhancing the FD accuracy under complex working conditions. The NND constructed based on the SKB (named SKND) overcomes the defect of information loss in the process of encoding and decoding and effectively improves the anti-noise ability of the TSACNN by improving the signal-to-noise ratio of the input signals. Experimental results show that TSACNN can achieve better performance than four state-of-the-art CNN-based FD models under both noise interference and complex working conditions and that SKND can achieve better denoising ability than a state-of-the-art NND model. By deep fusion of SKND and TSACNN, the proposed SKND-TSACNN can achieve satisfactory FD accuracy and robustness under complex working conditions and strong noise interference.