Abstract
Industrial process data typical exhibit nonlinear and time-varying characteristics, which limit the effectiveness of traditional process monitoring methods. Kernel-based techniques can address nonlinearity by projecting data into high-dimensional feature spaces, but their application to time-varying systems remains limited due to two key challenges: the near-singularity of the kernel matrix during updates and the high computational cost of updating implementation. This study proposes a kernel-based time-varying vector autoregressive model to characterize system dynamics. An adaptive regularization strategy is introduced to balance model bias and numerical stability by dynamically tuning the regularization parameter. To reduce computational burden in updating implementation, the eigenvalue decomposition of the kernel matrix is reformulated as that of diagonal plus rank-one matrix and arrow matrix, thereby reducing at least one-third of the computational cost. To facilitate a practical model, a favorable sample incorporation strategy is designed to suppress the influence of faulty samples and enhance model robustness. The experimental results show that the proposed method achieves superior average fault detection rates and false alarm rates on five industrial datasets compared with existing advanced time-varying process monitoring methods, demonstrating good generalization performance.
Funding Information
  • National Key R&D Program of China (2024YFB3311405)
  • National Natural Science Foundation of China (62303039)
  • National Natural Science Foundation of China (62433004)

This publication has 46 references indexed in Scilit: