Fault Diagnosis of Dynamic Chemical Processes Based on Improved Residual Network Combined with a Gated Recurrent Unit

Aiming at the challenge of distinguishing the contributions of variables in dynamic chemical process data, this paper proposes a novel fault diagnosis method based on the IResNet-GRU model. First, we utilize principal component analysis to compute a correlation matrix, which serves as the input for an attention module. This approach enables the evaluation of feature contributions to predictions, thereby identifying the root-cause variables responsible for faults. Concurrently, we enhance the residual network (ResNet) with the attention module to assign weights to the extracted features. The improved ResNet (IResNet) can differentiate the significance of the monitored variables. Second, we augment the raw data into two-dimensional data using sliding window technology, capturing spatial and temporal data features. Finally, a gated recurrent unit is integrated to extract dynamic features from the augmented two-dimensional data effectively. The proposed method is validated using the Tennessee–Eastman chemical process. The diagnosis results demonstrate that the proposed method outperforms conventional methods.