Contrastive Dynamic Residual Shrinkage Network for Industrial Process Fault Diagnosis

Convolutional Neural Networks (CNNs) have become a popular choice for industrial process fault diagnosis owing to the outstanding feature extraction ability. However, the susceptibility of process data to noise often cause CNNs to capture irrelevant or redundant information during feature learning, thereby affecting feature discriminability and reducing diagnostic accuracy. To solve this problem, a contrastive dynamic residual shrinkage (CDRS) network is designed to improve feature expression in industrial process fault diagnosis. First, we propose a dynamic residual shrinkage module, which dynamically aggregates multiple convolutional kernels to enrich feature diversity without significantly expanding model complexity. Besides, a shrinkage function is employed to adaptively determine soft threshold, suppressing irrelevant information and refining feature quality. Second, inspired by contrast learning, we design a new loss function to optimize the feature space by reducing the intra-class distances and enlarging inter-class separability, thus improving feature discriminability for downstream fault classification. Finally, experimental validation on the Tennessee Eastman (TE) process demonstrates the effectiveness of the proposed CDRS network, achieving an average accuracy improvement of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4.80 \%}$</tex> over existing approaches.