Process monitoring and dynamic fusion of complex industrial systems: A reconstruction-based Bayesian framework

In the last decade, the automation complexity and sophistication of multiple sensors in modern industrial systems have grown significantly with the fast transformation from Industry 4.0 to 5.0. This transformation of Industry 4.0 to 5.0 has still not been carefully investigated for dynamic monitoring and fusion information. Traditional monitoring techniques are not advanced to address these significant challenges when evaluating the intricate information collected from sophisticated sensors and computing systems. This paper presents an innovative, intelligent dynamic fusion framework that utilizes machine learning (ML) and deep learning (DL) to combine dynamic Bayesian global-local preserving projection (DBGLPP), mutual information entropy (MIE), stacked autoencoders (SAE), kernel density estimation (KDE), and reconstruction-based contributions (RBC). The novel dynamic fusion framework addresses the issues of real-time dynamic monitoring in physio-chemical systems. The approach seeks to investigate, categorize, isolate, identify, and diagnose anomalies and faults. The framework's feasibility and functionality are evaluated using recently established models such as Wavelet-PCA, CWT-3D-CNN, DALSTM-AE, and the newly proposed dynamic, intelligent fusion monitoring framework (DBGLPP-SAE) as model validation baselines. The proposed innovative methodology has been tested by assessment of the ethanol-water distillation column (DC) and the Tennessee Eastman Process (TEP) as baseline benchmarks. The results and findings showed that the novel dynamic intelligent fusion framework can address the issues and challenges of real-world complex industrial systems. It can robustly detect abnormalities, classify and isolate faults, identify root channels, and diagnose problematic variables.