Robust contrastive learning based on evidential uncertainty for open-set semi-supervised industrial fault diagnosis

Semi-supervised learning is increasingly applied in industrial fault diagnosis, presuming that the label spaces of labeled samples and unlabeled samples are identical. However, unknown faults in real-world scenarios present a challenge for traditional closed-set approaches. To this end, we propose a novel framework for open-set semi-supervised industrial fault diagnosis, named Evidential Robust Contrastive Learning (ERCL). The theory of evidence is introduced to explicitly assess sample uncertainty, guiding the evidential robust contrastive representation module to implement instance-level training strategies for each unlabeled sample. Additionally, an adaptive out-of-distribution detection module is developed to detect unknown faults by evaluating the difference in mutual information distributions between in-distribution (ID) and out-of-distribution (OOD) samples, thereby avoiding over-reliance on prior knowledge. The proposed framework is validated with the Tennessee Eastman process and polyester esterification process. As proved in the experiments, ERCL exhibits superior diagnosis accuracy in open scenarios with limited labeled data.