Fault Diagnosis for Power Electronic System with Fault Gradation Using Assembly Neural Network Group

The existing methods of fault diagnosis almost treat all the faults equally. This always leads to large expense of the whole diagnosis system and serious faults are not controlled efficiently. This paper proposes a new fault diagnosis approach to power electronic system with fault gradation using assembly BP(Back-Propagation) neural network group consisting of 3 sub BP neural networks. According to the hazard extents and the occurrence frequencies of different faults, the faults are divided into different grades. The higher the fault grade, the larger the number of the used sub neural networks is. Experimental results show that the diagnosis correctness rate of the faults with the highest grade is 100% , and the diagnosis correctness rates of the other faults with lower grade, which are of less hazard and lower occurrence frequency, are also about 95%.Because the final output is the sum of the products of the sub neural networks´ outputs and their confidence degrees, the wrong output of one sub neural network has little influence to the final output of the assembly neural network group. Therefore, our approach makes the correctness rate of the fault diagnosis rise greatly. The approach proposed in this paper also can be expanded to other fault diagnoses, such as mechanical systems.