Trustworthy process monitoring based on canonical correlation analysis considering epistemic uncertainty

Due to scarce fault data and frequent operational condition changes of actual industrial processes, data-based models suffer from epistemic uncertainty, which significantly impacts the reliability of monitoring results. To address this issue, this paper proposes an uncertainty canonical correlation analysis (UCCA) method based on normal uncertainty distribution. Firstly, uncertainty theory is adopted to transform point data into range data for better characterization of epistemic uncertainty. The mean and variance of variables are redefined. Subsequently, a novel uncertainty feature extraction model called UCCA is proposed by integrating canonical correlation analysis (CCA) approach, which fully considers the issue of incomplete cognition caused by limited sample size and varying operating conditions. Finally, a trustworthy monitoring method based on UCCA is proposed to suppress overconfident judgments caused by epistemic uncertainty, which reduces the risk of false alarms. Moreover, the monitoring results are provided while the reliability of assessment outcomes is evaluated through belief measure. Experiments demonstrate that compared to other algorithms, UCCA achieves an average FDR increase of 14.01% and FAR decrease of 1.48% in continuous stirred tank reactor. In wastewater treatment process, it achieves an average FDR increase of 1.96% and FAR decrease of 0.40%.