A few-shot class-incremental fault diagnosis model for non-analytical indicator diagrams based on serialization and contrastive learning

Abstract In pumping unit well fault diagnosis, non-analytical indicator diagrams frequently occur due to the privacy of field data and historical accumulation, resulting in invalid information and pixel redundancy that reduce diagnostic efficiency. To address this, a serialization method is proposed to convert two-dimensional redundant diagrams into one-dimensional sequences, enhancing information density and constructing a serialized fault dataset. To further tackle low diagnostic accuracy and catastrophic forgetting in Few-Shot Class-Incremental Learning, a Few-Shot Class-Incremental fault diagnosis model for non-analytical indicator diagrams based on serialization and contrastive learning (CLIFS-Net) is proposed, which is specifically designed for the proposed special serialized data format. First, the self-supervised contrastive learning module is pretrained on base classes and then frozen, learning stable and transferable representations from serialized data by capturing local contour variations and global morphology. Then, a lightweight adapter and sample replay strategy, combined with a Nearest-Mean-of-Exemplars classifier, enable flexible adaptation to new classes and alleviate forgetting. Experimental results on the serialized non-analytical indicator diagrams dataset show that the proposed backbone and classifier respectively enhance baseline methods by 2.39% and 4.23%. The integrated CLIFS-Net achieves an average diagnostic accuracy of 92.72%, higher than the overall average of 87.94% for baseline, backbone-modified, and classifier-modified methods.