Fault diagnosis of batch processes for small samples based on contrastive learning

Abstract

Fault diagnosis plays a critical role in process engineering. Existing methods often depend on large datasets for continuous process. However, for batch processes, some key variables are transient and often measured offline. Hence, the size of available datasets is small, making it difficult to effectively extract useful features for diagnosis. To overcome this limitation, this paper proposes a gated transformer network based on supervised contrastive learning (SCGTN), specifically designed for fault diagnosis in small-sample batch processes. SCGTN incorporates a dual-channel gated transformer network to independently extract features from the temporal dimension and multi-variable statistics of batch process data. In this proposed framework, a supervised contrastive cost function has been incorporated as one of the loss terms into the total loss function to enhance the discriminative power of the learned representations in the feature space. The model parameters are then optimized by considering both the supervised contrastive loss and the cross-entropy loss. Experimental results demonstrate that this method can effectively capture deep feature representations and perform reliable fault diagnosis in small-sample scenarios. When compared to four other methods, SCGTN exhibits superior prediction accuracy and stronger generalization capabilities.