摘要
•Innovative machine learning-based multi-model fusion for process monitoring.•Ensures robust detection and diagnosis of abnormal operational conditions.•Integrates dynamic correlation and reconstruction contribution for accuracy.•Validated using ethanol-water distillation and a three-phase flow facility.•Performance compared with traditional baseline approaches.
Over the last twenty years, industrial chemical processes have grown increasingly complex and dynamic due to rapid developments in manufacturing automation and digital sensor technology. Interrelated systems and intricate control processes are safety challenges in industrial processes. Conventional methods are limited to a single scale and presume that the data is static or minimally dynamic. However, these methods cannot handle sophisticated automated industrial processes and dynamically correlated information. Process monitoring is a vital area of study in real-time processes to enhance performance and process safety. This paper presents a novel intelligent machine learning (ML) based multi-model fusion monitoring framework to deal with the safety challenges associated with physio-chemical processes. This framework combines data-driven machine learning methods using distributed canonical correlation analysis (DCCA), autoencoder (AE), and reconstruction-based contribution (RBC). The efficacy of the ML frameworks is assessed and distinguished using an ethanol-water mixture distillation column (DC) system and Three-Phase flow facility benchmark as the case study scenarios. The proposed novel frameworks are validated using advanced methodologies such as stacked-AE (SAE) and long short-term memory-AE (LSTM-AE). The findings demonstrate that the suggested (DCCA-AE) framework is more effective and resilient in detecting anomalies and variable density diagnosis than the current SAE and LSTM-AE methods. This technique allows for robust detection, reliable identification, and diagnosis of abnormal safety situations.
[Display omitted]