An Unsupervised Knowledge and Data Dual-Driven Based Fault Diagnosis for Industrial Process

Fault diagnosis (FD) is essential for ensuring the safety and reliability of industrial processes. While data-driven approaches have demonstrated remarkable diagnostic performance, their heavy reliance on large-scale labeled datasets poses significant challenges due to the high cost and effort required for manual annotation. Unsupervised FD methods offer a promising alternative; however, their effectiveness is often hindered by the lack of principled integration of fault-related knowledge, resulting in suboptimal interpretability and diagnostic reliability. To address this limitation, this paper proposes an unsupervised knowledge and data dual-driven (KDDD) fault diagnosis framework that embeds fault characteristics into the feature extraction process by leveraging variable associations. Specifically, we hypothesize that fault patterns manifest in the structural relationships among monitoring variables. Based on this assumption, an association graph is constructed to capture these dependencies and subsequently integrated into the feature extraction model, enabling a more informed and interpretable fault representation. Extensive experiments on a widely adopted chemical process benchmark demonstrate the effectiveness of the proposed method in enhancing fault feature extraction and improving diagnostic performance.