Novel convex model-based approach for data-driven fault diagnosis considering uncertainty

Fault diagnosis plays a critical role in various engineering practices by ensuring system reliability and safety, where timely detection mitigates potential hazards and minimizes operational downtime. However, the presence of uncertainties inevitably induces fluctuation characteristics in measured signals, significantly compromising the accuracy of traditional data-driven fault diagnosis approaches. To address this challenge, this paper proposes a novel convex model-based fault diagnosis (CMFD) framework to improve the accuracy of fault classification. By reviewing some fundamental concepts of evidence theory, a universal data-driven fault diagnosis framework is presented first. To measure the fluctuation phenomenon in statistical features, an uncertain feature extraction method is proposed by means of two convex models. Based on the geometric characteristic of convex models, a novel volume-based strategy for basic probability assignment (BPA) determination is subsequently proposed to deliver quantitative pattern matching. Considering the potential conflicts in multi-source diagnostic results, an evidence-based information fusion procedure is introduced to obtain consistent outcomes. Eventually, two case studies are investigated to validate the effectiveness of the proposed models and methods. Compared to existing probabilistic solutions for uncertainty scenarios, CMFD adapts better to limited samples and operates without distributional assumptions. By pioneering geometric exploitation of convex models for the similarity measure of uncertain fault features, CMFD bypasses the reliance on auxiliary classifiers and converts probabilistic discrepancies into more intuitive geometric relationships, significantly enhancing the interpretability of uncertainty-aware fault diagnosis.