Sensor fault detection via multiscale analysis of prediction model residuals

In this paper, a new approach to sensor validation in real time is described that is based on (1) representation of the sensor signal by wavelets, (2) decomposition of the signal for different frequency ranges, (3) formation of a prediction model using information at an appropriate frequency (level of decomposition), (4) calculation of residuals, namely, the difference between the predictions of the prediction model and the values of the signal decomposed at the same frequency, and (5) diagnosis of the existence of faults via tests on the residuals. The proposed strategy is able to eliminate the effect of noise and process changes from the effects of physical changes in the sensor itself. To demonstrate the effectiveness of the proposed strategy, a simulated noisy signal from a thermocouple in a continuous stirred tank reactor and a temperature signal from an operating pilot distillation column were analyzed. The results of the diagnosis indicated that the proposed strategy had a low type I (false alarm) rate and a satisfactory type II (failure to detect faults) rate.