Recurrent Neural Network Based On-line Fault Diagnosis Approach for Power Electronic Devices

294 fault patterns of 12-pulse waveform controlled rectifier circuit (CRC) are studied firstly; a special fault classification method according to the rectifier aberrant voltage waveforms is put forward then. 12-dimension fault patterns and the corresponding fault codes are obtained through logic pre-processing of the fault rectifier voltage waveforms. A recurrent neural network (RNN) with deviation error units is used to construct a fault mapping. The fault patterns are used as the input of the network, after being calculated by the neural network the 12-dimension fault codes are obtained to indicate the fault state of the system. Simulation and experiment study demonstrate that the proposed technique is low time consuming with high fault identification rate, it improves the performance of the existing neural network based on-line fault diagnosis methods effectively. The proposed method is suitable for the fault diagnosis of complex power electronic devices or systems.