Abstract Ensuring safe and reliable operation in modern industrial systems requires advanced monitoring frameworks capable of distinguishing between quality‐related and process‐related faults. Conventional multivariate statistical methods, such as partial least squares (PLS) and canonical correlation analysis (CCA), and their dynamic extensions, often fail to capture nonlinear relationships and exhibit limited robustness in noisy or large‐scale scenarios. To overcome these limitations, this study develops a deep variational canonical correlation analysis (DVCCA) framework for intelligent process monitoring. By embedding canonical correlation into a probabilistic latent variable model, DVCCA enables simultaneous modelling of nonlinear dependencies between process and quality‐related variables while retaining interpretability through its probabilistic structure. A set of monitoring indices, including , , and , are constructed to capture latent subspace deviations and reconstruction errors, with control limits estimated via kernel density estimation. In the three‐phase flow facility, DVCCA achieves a fault detection rate (FDR) of 71.06% with a false alarm rate (FAR) as low as 0.06% for quality‐related faults, and an FDR of 98.69% for process‐related faults. In the nuclear power plant case, the proposed approach attains nearly 100% detection accuracy while maintaining FARs close to zero. These results confirm the effectiveness and robustness of DVCCA for online monitoring in safety‐critical industrial applications.
Zhang et al. (Sun,) studied this question.