(375h) Online Fault Detection and Diagnosis of Industrial Processes Via Data Augmentation and Integrative Learning of Process Knowledge and Fault Propagation Map

Effective operation of industrial processes requires successful fault detection and diagnosis. Robust and fast online process monitoring is still challenging for industrial practitioners even though large volumes of historical and online sensor data are available. These challenges stem from: 1) fault scenarios in chemical processes are naturally complex and cannot be exhaustively enumerated or predicted, 2) sensor measurements continuously produce massive arrays of high-dimensional big data streams that are often nonparametric and heterogeneous, and 3) the strict environmental, health, and safety requirements established in the facilities demand high reliability and accuracy of any process monitoring and fault detection tool. To address these challenges, in this talk, we present a robust and reliable process monitoring framework that performs the fault detection and diagnosis tasks simultaneously. The fault detection is based on multivariate statistical process control (SPC) that can monitor nonparametric and heterogeneous high-dimensional data streams and detect process anomalies as early as possible while maintaining a pre-specified in-control average run length. After a fault is detected by the detection module, the faulty data are sent to the diagnosis module for fault classification. The classification is performed by a machine learning model, e.g. a random forest model, trained by historical faulty data and process knowledge. First, a univariate SPC algorithm is applied on each faulty data stream to find the chronological sequence of variables that showed any mean shift up to the time that the alarm is raised. This sequence is further accompanied by the process knowledge and fed to the machine learning model to obtain the corresponding faulty scenario.

Based on the chronological sequence obtained from univariate SPC algorithm, a fault propagation map is constructed for each fault scenario to facilitate fault classification. The fault propagation maps can be used for fault classification and diagnosis by graph neural network (GNN). Alternatively, we incorporate the local statistics obtained from the multivariate SPC algorithm in a Bayesian network framework for causality analysis. Both methods are compared using the classic Tennessee Eastman Process (TEP) benchmark dataset. It’s worth highlighting that our proposed process monitoring framework only uses a fraction of the TEP timeseries dataset, as fault diagnosis or classification must be done by the time a fault is detected (an alarm is raised). This practical constraint limits fault classification accuracy, as the model does not have access to the full timeseries data in the TEP dataset for training and testing. To address this challenge, we introduce data augmentation strategies to increase the volume of training data subject to this practical limitation. In this talk, we will demonstrate the effectiveness of data augmentation on improving fault classification accuracy.