Fault Diagnosis With Robustness and Lightweight Synergy Under Noisy Environment

Abstract

Over the years, deep learning-based fault diagnosis approaches have made remarkable progress. However, the diagnosis accuracy has been on a steady decline because of the noise interference in industrial applications. Additionally, the accuracy improvement is often accompanied by an increase in computational cost, which makes it difficult to achieve engineering-embedded deployment. To find the balance between robustness and complexity, a new intelligent fault diagnosis framework named lightweight multi-scale dilated causal convolutional attention network (Li-MDCCAN) is proposed. Based on dual fusion mechanisms, the proposed network realizes high robustness by adopting multi-scale strategy and attention modules, which explore multi-scale and multi-dimensional features along three parallel paths. Furthermore, lightweight convolutional blocks are introduced to reduce the network’s complexity. The performance of Li-MDCCAN is verified with six different datasets and compared with those of four advanced fault diagnosis networks. The experimental results indicate the effectiveness and superiority of Li-MDCCAN in increasing robustness and decreasing complexity. Moreover, prior sensor placement is discussed to guide applications under limited engineering conditions.