Grant: Discovery Grants Program - Individual

Chemical processes can experience faults leading to potential failures during their operations; process fault prediction is widely used to predict failures so that operators can correct them in advance. Failures can cause serious and sometimes catastrophic incidents in chemical plants. Therefore, safe and environmentally diligent process operation requires monitoring tools with accurate fault identification and prediction capability. Conventional unsupervised multivariate statistical learning strategies based on only normal operating data are becoming inadequate for training predictive monitors for increasingly complicated industrial processes. Supervised learning employing both normal and fault data is necessary, but insufficient fault-labelled data limits its application. Transfer learning can train a predictive monitor using similar incidents from other facilities. However, critical issues must be resolved before this emerging technology can be applied in the process industry. Although transfer learning has been widely studied in the literature and appears encouraging, little has been done for process monitoring applications, where multivariate statistical monitoring technologies predominate. Moreover, process knowledge has been largely missed in developing transfer learning approaches. This proposal will address these problems and close the gaps between promise and reality. Hence our short-term objectives are to investigate predictive process monitoring problems with scarce fault-labelled data and develop solution strategies through integrated process-knowledge and data-feature transfer. Our long-term objectives are to create an integrated framework for predictive monitoring, diagnosis, and decision support for safe operation. This decision support system will include monitoring to predict incidents, diagnosis to locate the source of events, and decision support to guide operators to reestablish safe operations. This proposed research will solve the long-standing problem of scarce fault-labelled data in process monitoring, thereby establishing a new framework for process fault prediction from three perspectives: 1) generalization of classical multivariate statistical models for fault prediction; 2) transfer learning using the proposed generalized model; 3) integrated transfer of process knowledge and data features. The proposed research program will establish a theoretically sound and methodologically solid foundation for process fault detection and prediction. Our solutions will apply to a wide range of industries that employ automation and autonomous systems. Our research program will train highly qualified personnel in process system engineering, machine learning, and data analytics. They will be the next generation of academic and technical leaders who will integrate these technologies into process plants to boost the competitiveness and safety of the Canadian industry.