A novel sparse-aware contrastive learning network with adaptive gating neurons for extreme class imbalance diagnosis scenarios

Abstract

The intelligent diagnosis model can exhibit excellent diagnostic performance with sufficient training samples and under ideal conditions. However, in engineering environments, it is difficult to obtain balanced fault samples from equipment, and the interference from noise components in signals makes it challenging for deep learning models to extract fault features, posing significant challenges for intelligent diagnosis. To address these issues, this paper proposes a novel sparse perception contrastive learning network with adaptive gating neurons for extreme class imbalance diagnosis scenarios. Specifically, we first propose an adaptive gating neurons residual network, derive and establish a mathematical relationship between the adaptive gating neuron and learnable weighted autocorrelation functions, demonstrating the model’s ability to extract relevant features from vibration signals and perform adaptive noise reduction. Building upon this, we propose a Sparse Perception Cross-entropy Loss (SPCL) function, which focuses on tail-class fault samples that are difficult to cluster based on class sparsity in the feature space. Furthermore, to further enhance the diagnostic performance of the contrastive learning model in class-imbalanced bearing faults, we propose a Center Contrastive Loss (CCL) function. CCL calculates the centers of each class using classifier weights optimized by SPCL function, ensuring that all fault classes are represented in each mini-batch during learning, thereby enabling effective contrastive learning. The efficacy of the introduced methodologies is confirmed through experimental outcomes obtained from both publicly accessible and proprietary datasets. Experimental results demonstrate that our proposed method significantly outperforms other methods in the intelligent diagnosis of bearing faults in scenarios with extreme class imbalance.