A Fault Diagnosis Method Based on MCAG-ResNet for Industrial Processes

Industrial process fault diagnosis remains challenging because one-dimensional time-series data often involve complex dynamics, noise disturbances, and temporal dependencies, which hinder reliable fault representation and robust diagnostic decisions under complex operating conditions. To address these challenges, a fault diagnosis method for industrial processes based on the Multiscale Convolution-Attention-GRU Residual Network (MCAG-ResNet) is proposed. MCAG-ResNet integrates multiscale feature learning, attention-based feature recalibration, temporal dependency modeling, and residual learning in a unified architecture to enhance discriminative fault representation and diagnostic robustness. In addition, normalization and lightweight data augmentation are incorporated to improve training stability and generalization performance. Validation on the Tennessee Eastman (TE) and Continuous Stirred Tank Reactor (CSTR) datasets demonstrates the effectiveness, generalization capability, and diagnostic stability of the MCAG-ResNet in complex industrial process fault diagnosis. Further analyses, including variable contribution, feature importance, noise robustness, hyperparameter sensitivity, performance–complexity, and statistical stability analyses, verify its interpretability, robustness, parameter rationality, practical applicability, and stability.