Winner take all experts network for sensor validation.

Abstract

The validation of sensor measurements has become an integral part of the operation and control of modern industrial equipment. The sensor under harsh environment must be shown to consistently provide the correct measurements. Analysis of the validation hardware or software should trigger an alarm when the sensor signals deviate appreciably from the correct values. Neural network based models can be used to on-line estimate critical sensor values when neighboring sensor measurements are used as inputs. The underlying assumption is that the neighboring sensors share an analytical relationship. The discrepancy between the measured and predicted sensor values may then be used as an indicator for sensor health. The proposed Winner Take All Experts (WTAE) network based on a 'divide and conquer' strategy significantly reduces the computational time required to train the neural network. It employs a growing fuzzy clustering algorithm to divide a complicated problem into a series of simpler sub-problems and assigns an expert to each of them locally. After the sensor approximation, the outputs from the estimator and the real sensor readings are compared both in the time domain and the frequency domain. Three fault indicators are used to provide analytical redundancy to detect the sensor failure. In the decision stage, the intersection of three fuzzy sets accomplishes a decision level fusion, which indicates the confidence level of the sensor health. Two data sets, the Spectra Quest Machinery Fault Simulator data set and the Westland vibration data set, were used in simulations to demonstrate the performance of the proposed WTAE network. The simulation results show the proposed WTAE is competitive with or even superior to the existing approaches.