A Two-Stage Feature-Extraction-Based Fault Detection Method for Industrial Processes With Imperfect Data

Data distributions of normal and abnormal operational conditions may overlap to some extent, making it challenging for traditional fault detection models to extract the intrinsic features of normal data, which brings serious challenge to fault detection. Against this background, we propose a new two-stage feature-extraction (TSFE)-based fault detection method with variable missing repair mechanism for industrial processes with imperfect data. The proposed method can effectively compress the data feature-space to extract intrinsic features. First, an out-of-distribution (OOD) data generation method is designed. Second, a new feature compression auto-encoder (FCAE) is constructed for stage-one feature extraction. Positive and negative sample feature pairs are constructed based on normal data and OOD data. Then, a new feature compression supervised contrastive learning (FCSCL) network is constructed for stage-two feature extraction. In addition, the proposed method includes a variable missing repair mechanism to handle imperfect data. Finally, the effectiveness of the proposed method is verified using a real hot strip mill process dataset. The experimental results indicate that the proposed method achieves an average correct monitoring rate of up to 95.5%, with a false alarm rate of only 1%. This demonstrates that high fault detection performance can still be maintained even in the presence of variable signals missing.