Enhancement and laboratory implementation of neural network detection of short circuit faults in DC transit system

Abstract

The continuing development of neural networks for real-time detection of DC short circuit faults in DC transit systems is described. The Discrete Wavelet Transform has been previously applied to detect any surges in DC third rail current waveform. In the event of a surge, the wavelet transform extracts a feature vector from the current waveform and feeds it to a self-organising neural network. The neural network then determines whether the feature vector belongs to a fault current surge due to short circuit occurring across the DC-link capacitor within a chopper train, or to a normal surge due to the starting of the chopper train. Perfect classification has been achieved between different cases of train starting and short circuits. The robustness of the neural network is further proven using extended data sets for training and testing and laboratory implementation. A new simulation model for the inverter trains is developed. A new type of short circuit fault occurring between third rail and track is modelled. For testing the fault detection scheme within the laboratory environment, a hardware model of the DC transit system is carefully built using an induction motor, laboratory power supplies and electronic models of railway components. Two neural networks are trained using the simulation data, and tested using the simulation and laboratory-measured data. Perfect classification is again achieved from both neural networks. The laboratory environment provides a valuable platform for fine-tuning the neural networks and for in-depth studies into the effects of practical constraints such as measurement noises and nonlinearities of the hardware models.