A novel multi-scale quadratic convolutional network for bearing fault diagnosis: Handling noisy conditions

Abstract

Despite recent advances in deep learning-based fault diagnosis, it is challenging to build a reliable fault diagnosis model for rolling bearings under noisy conditions. Rolling bearing vibration signals are frequently corrupted by various types of noises, which poses significant challenges for accurate fault feature extraction and degrades diagnostic performance. This paper proposes a multi-scale hybrid information fusion method based on a quadratic embedded-attention convolutional network (MSHQAN) for bearing fault diagnosis. Specifically, a multi-scale feature extraction (MFE) module is proposed to extract representative features from the vibration signals. In the MFE module, a quadratic mixed layer is introduced to model the complex nonlinear relationships and multi-scale features in the vibration signal. Furthermore, a spatial-refined feature unit is put forward to refine the features extracted by the quadratic mixed unit to obtain discriminative spatial features. Additionally, a hybrid feature weight fusion module is proposed to achieve optimal fusion of cross-scale hybrid features via an adaptive weight allocation strategy. Moreover, a quadratic mixed-convolutional pooling module is developed to further strengthen the discriminative representation of fault-related features through in-depth refinement. To enhance model interpretability, a Grad-CAM++ method is employed to identify fault-related regions under noisy conditions. Experimental results validate the effectiveness of the proposed MSHQAN on two public bearing datasets.