System Level Fault Diagnosis Algorithm Research Based On BP Neural Network

With higher error precision and strong nonlinear mapping ability, BP neural network is currently the mostly used feed forward neural networks. And it is easy to program, which make it to be a hot modern intelligent diagnostics. Genetic algorithm is a random selection algorithm, and it is able to search in the global scope. To improve BP algorithm easy to fall into local minimum problem and slow convergence, we combine Genetic algorithm with BP neural network algorithm, which can reduce the diagnosis time as well as improve the efficiency of the algorithm.In this thesis, Genetic algorithm to optimize BP neural network diagnostic algorithm (GA-BPFD) is first proposed for the system-level diagnosis based on Chwa & Hakimi model. The thesis contains of three improvements:BP neural network construction, activation function optimization, preprocessing of the weights and bias.Firstly, research the selection rules of network layers and nodes in the design phase of the network structure. And do optimization analysis for BP Activation function, thereby improving the network generalization performance. Secondly, we use the genetic algorithm to preprocess the test report in order to obtain a set of minimum error BP neural network weights and bias values; then we put the weights and bias values obtained through previous step as the initial BP neural network weights and bias values, diagnose the faults according to the test report. We also do some proof and analysis for the GA convergence as well as the GA-BPFD correctness. Finally, the evaluation model of BP network and GA-BP network are respectively trained, learned and tested by using the network toolbox of Matlab and writing GA-BP program. Experimental results show that, compared with BPFD algorithm, the improved algorithm GA-BPFD has a high diagnostic accuracy and lower time complexity, and the constructed network also has good generalization ability.