A novel quality-related generative data-driven fault diagnosis method for complex industrial processes with incomplete data

Abstract Spatiotemporal incomplete data pose significant challenges to the robust diagnosis of quality-related faults in complex industrial environments. Missing data, often attributed to sensor failures, communication issues, or storage limitations, can compromise the reliability of diagnostic systems and diminish the effectiveness of conventional fault detection methods. To address these challenges, this paper proposes a novel integrated framework combining a generative quality-aware (QA) generative adversarial imputation network (GAIN)-based imputation with a graph-based spatiotemporal transformer encoder, a broad learning system (BLS) classifier and an enhanced uncertainty estimation mechanism. Compared to the standard GAIN, our proposed QA-GAIN incorporates a specific quality indicator into adversarial training, reducing the root mean square error (RMSE) by 23%–53% across missing rates of 10%–80% compared to traditional imputation methods. For downstream monitoring, the spatiotemporal transformer with graph convolution and uncertainty weighting reduces Hotelling’s T 2 false alarm rate (FAR) by up to 27% and improves the fault detection rate (FDR) by up to 15% compared to transformer-based baseline models under severe missingness. The BLS-inspired classifier uses latent features efficiently to enhance decision boundaries, while uncertainty calibration shows up to 64% lower expected calibration error (ECE) than standard Monte Carlo (MC) dropout, providing more reliable probability estimates and higher confidence in correct predictions. Evaluations on the steel industry hot strip mill process (HSMP) and the chemical industry Tennessee Eastman process (TEP) datasets confirm the proposed framework’s ability to recover missing data, understand spatiotemporal patterns and deliver strong and trustworthy predictions for quality-related fault diagnosis.