Fault Diagnosis for Limited Similar Batch Process using Two-dimensional Adversarial Domain Adaptation-based Graph Convolutional Network

Batch process fault diagnosis is critical for ensuring the quality and safety of batch processes production. However, recognizing fault types in low-volume and similar batch processes poses a significant challenge. This paper proposes a novel domain-adaptive transfer learning model named two-dimensional adversarial domain adaptation-based graph convolutional network (2D-ADGCN) for limited similar batch process fault diagnosis. Firstly, the method creates enhanced samples by integrating statistical information within a two-dimensional sliding window, significantly expanding the scale of the analyzed data by sliding the window in both time and batch dimensions. Secondly, by constructing an adversarial domain adaptation graph convolutional network, the data from the source domain (original batches) and the target domain (similar batches) are mapped into a shared feature space, reducing the distribution discrepancy between the two domains. Then, leveraging the labeled samples from the source domain, the model learns discriminative representations for the samples and performs classification on the target domain samples, enabling fault diagnosis for similar batches. Finally, the effectiveness of 2D-ADGCN is validated through industrial case studies of two Lactobacillus plantarum fed-batch fermentation processes, demonstrating its significant potential for fault diagnosis in low-volume and similar batch processes.