Fault prediction in electrical valves using temporal Kohonen maps

This paper presents a proactive maintenance scheme for the prediction of faults in electrical valves. In our case study, these valves are used for controlling the oil flow in a distribution network. A system implements temporal self-organizing maps for the prediction of faults. These faults lead to deviations either on torque, on the valve end position or on opening/closing time. For fault prediction, one map is trained using data from a mathematical model devised for the electrical valve. The training is performed by fault injection based on three parameter deviations over this same mathematical model. The map learns the energies of the torque and the position that are computed using the wavelet packet transform. Once the map is trained, the system is ready for on-line monitoring of the valve. During the on-line testing phase, the system computes the Euclidean distance and the activation of data series. The biggest activation determines which is the winner neuron of the map for one data series. The obtained results demonstrate a new solution for prediction behavior of these valves.