A strategy for detection and isolation of sensor failures and process upsets

A novel approach is proposed to isolate sensors that are affected by the root cause of nonconforming operation and to distinguish between failed sensors and process upsets. Systems having multivariate nature can be monitored by building a principal component analysis (PCA) model using historical data. T2 and sum-of-squared-prediction error (SPE) of the calibration model facilitate fault detection and isolation on-line. These two measures are complementary in explaining the events captured and not captured by the model. In this paper, we put more emphasis on the importance of using the T2 and the SPE together for fault detection and identification. Correlation coefficient criterion was utilized to infer about the state of the correlation structure between one sensor and its closest neighbor for distinguishing between sensor failures and process upsets. Faulty measurements were reconstructed from available sensors using the calibration model and an optimization algorithm which in turn unveiled more process upsets. The strategy is illustrated on a benchmark industrial liquid-fed ceramic melter.