Fault Detection and Isolation in a Spiral-Wound Reverse Osmosis (RO) Desalination Plant

Abstract

Sensor fault detection and isolation (SFDI) approaches, based on support vector regression (SVR) plant sensor models and self-organizing-map (SOM) analysis, were investigated for application to reverse osmosis (RO) desalination plant operation. SFDI-SVR and SFDI-SOM were assessed using operational data from a small spiral-wound RO pilot plant and synthetic faulty data generated as perturbations relative to normal plant operational data. SFDI-SVR was achieved without false negative (FN) detections for sensor deviations of ≳|10%| and FN detections of, at the most, ≲|5%|, and for sensor deviations of ≳|4%| at sensor fault detection (FD) thresholds of up to ∼|4%|. False positive (FP) detections were almost invariant, with respect to sensor FD, being ≲|5%| for sensor deviations of ≳|5%|. Corrections of faulty sensor readings were within SVR model accuracy (AARE < 1%) for SFDI-SVR and ≲|5%| for SFDI-SOM. Although SFDI-SOM has lower detection accuracy, it requires a single overall plant model (or SOM), while providing pictorial representation of plant operation and depiction of faulty operational trajectories.