Failure Diagnostic System for Air-Operated Control Valves Using Neural Network

A prototype failure diagnosis system has been developed using neural network technology for the actuators of air-operated valves. Because actual failure data were not easily available, the data of 30 failure patterns were experimentally obtained using more than 10 sensors. The time series data of sensor signals are Fourier transformed. The data of magnitude spectrum, phase difference and other quantities are used as the characteristic parameters in our failure diagnosis. From the data, appropriate information for use in failure diagnosis was extracted. Furthermore, similarities among failure characteristics were found by fuzzy clustering and statistical analysis. The new system that we developed consists of many sub-networks and one main network. Each sub-network is related to one specific sensor signal, and deals with the magnitude spectra from the sensor signal. The main network makes the final decision according to the outputs from the sub-networks and other data. In our system, the number of network connections can be reduced by approximately 40% without degradation of the recognition capability in comparison with the conventional system that uses only one neural network.