Diagnostics based on the Statistical Correlation of Sensors

The paper describes a new strategy for real-time sensor diagnostics that is based on the statistical correlation of various sensor signal pairs. During normal fault-free operation there is a certain correlation between the sensor signals which is lost in the event of a fault. The proposed algorithm quantifies the correlation between sensor signal pairs using real-time scalar metrics based on the Mahalanobis-distance concept. During normal operation all metrics follow a similar pattern, however in the event of a fault; metrics involving the faulty sensor would increase in proportion to the magnitude of the fault. Thus, by monitoring this pattern and using a suitable fault-signature table it is possible to isolate the faulty sensor in real-time. Preliminary simulation results suggest that the strategy can mitigate the false-alarms experienced by most model-based diagnostic algorithms due to an intrinsic ability to distinguish nonlinear vehicle behavior from actual sensor faults.