A multirate gated variational triple-latent-variable model for monitoring industrial data with heterogeneous sampling rates

Abstract

Background

Industrial process monitoring frequently involves the analysis of multivariate time-series data collected at heterogeneous sampling rates, presenting substantial challenges for accurate fault detection and feature extraction.

Methods

To address these issues, we propose a multirate gated variational triple-latent-variable (MG-VTLV) model that effectively captures nonlinear dynamic dependencies among variables sampled at various rates. The MG-VTLV model is built upon the variational autoencoder framework and incorporates three mutually independent latent variables—quality-relevant, quality-irrelevant, and process-irrelevant—to characterize and disentangle multi-level correlations. A multirate gating unit (MRGU) is embedded in both the encoder and decoder, allowing the model to adaptively adjust its parameters based on real-time data availability and enabling robust dynamic feature extraction under asynchronous sampling conditions.

Results

Experimental evaluations using both the simulated Tennessee Eastman platform and a real-world coal-fired power plant demonstrate that MG-VTLV outperforms existing methods in terms of fault detection accuracy and robustness, particularly under conditions of limited or imbalanced sampling.