A data-driven approach to fault diagnostics for industrial process plants based on feature extraction and inferential statistics

Accurate detection and diagnostics of faults in complex industrial plants are important for preventing unplanned downtime, optimizing operations and maintenance decisions, minimizing repair time, and optimizing spare part logistics. It is often infeasible to generate accurate physics-based models of complex equipment; therefore, and due to lower computational complexity, data-driven methods are frequently employed. We propose a novel method for data-driven fault diagnostics and validate it using the Tennessee Eastman process (TEP) benchmark. It is assumed that the time of the onset of the fault is known, such that time-series data from the process both before and after occurrence of the fault can be extracted. For each of the measured time-series, several statistical features are extracted. A statistical significance level is computed for each feature using inferential statistics measures. The matrix of significance levels serves as a ``fingerprint'' of each fault category and is used as input to a feedforward neural network. We show that the network can be trained to achieve high classification accuracy on data from the TEP benchmark model