Early Malfunction Diagnosis of Industrial Process Units Utilizing Online Linear Trend Profiles and Real-Time Classification

The early detection of potential malfunctions at process systems can significantly reduce downtime and improve their overall operability. In that context, this paper demonstrates the behavior and response, through a comparative analysis, of novel data-driven diagnosis methods for interdependent time series. The proposed real-time slope statistic profile method utilizes a self-adaptive sliding window based on a real-time classification technique of linear trend profiles of both interdependent time series and internal condition so as to avoid misdetections. The calculation of the linear trend profile is based on a standard parametric linear trend test, and the selection of possible incidents is based on its two-level cross-checking. All possible combinations for the calculation of the trend test and cross-checking are created to explore their efficiency. The proposed methods are tested against real data sets from a chemical process system of the Centre for Research and Technology Hellas/Chemical Process Energy and Resources Institute derived from specific scenarios during nominal operating conditions.