Abstract: Conventional artificial neural networks （ANNs） still have problems of easy plunging into local minimum,slow convergence rate and bad generalization capacity.For solving these problems,a quantum genetic algorithm was used to obtain initial weight value and bias values of ANNs.The quantum state and quantum gate operation of the quantum algorithm could improve the optimization performance as well as the quality of the initial values effectively,which laid a foundation for global optimum approximation in the later algorithm.After then,Levenberg-Marquardt algorithm,a fast algorithm,was adopted to get the optimal solution by optimizing the weight value and the threshold of the multilayer feedforward ANNs.By doing that,the problem of local minimum convergence in the networks training was solved.Before evaluation or judging,gas concentration and gas generation speed were used to judge whether transformer fault,and the dissolved gas-in-oil data of normal transformers and fault transformer was used by individual ANNs,the former given transformer health level and the latter given transformer fault type.In this way,the complexity of the net was decreased prominently and the effect of training and applying were improved.The proposal algorithm is used to evaluate and diagnose on-site transformer based on dissolved gas-in-oil data,and the accuracy is higher than 95%.