Robust Latent Dynamics Modeling for Highly Instrumented Processes

Due to the massive sensors and industrial feedback control loops in highly instrumented industries, collected process data are usually high dimensional with latent dynamics. Dynamic latent variables (DLVs) methods aim to reduce dimensionality by extracting dynamic low-dimensional features, but their effectiveness is reduced by overfitting and numerical instability problems as the number of sensors increases. To tackle this challenge, we introduce the dynamic shrinkage latent autoregressive (DS-LAR) modeling algorithm for robust latent dynamic modeling in highly instrumented processes. Specifically, we propose a novel objective function with a capacity control term that can prevent overfitting while extracting the most predictable features. Meanwhile, we ensure nonsingularity during the solving procedure through the newly designed term, effectively alleviating numerical instability. Moreover, we develop two initialization approaches with analytical solutions and a hyperparameter selection strategy to enhance the algorithm’s efficiency and robustness. The effectiveness of the DS-LAR algorithm is demonstrated through case studies, including a simulated data set and a real oscillation data set from Eastman Chemical Company.