Fault Diagnosis Model Based on Optimal Neural Network Structure

Abstract

Depth of neural network in the choice of network structure, the activation function and the optimization function aspect, mainly rely on the experience of the operator to determine, led to its optimization in swarm intelligence algorithm (SI), only in the specific network structure to find the optimal weights and bias, limiting the depth of the further development of neural network in fault diagnosis field. Therefore, this paper proposes a fault diagnosis model which combines adaptive genetic algorithm and deep feed-forward network. Through the verification of MNIST, CIFAR10 standard data set and CSTV simulation experiment data, the fault diagnosis accuracy of the proposed fault diagnosis model is significantly improved compared with the original model, which is 12.95% higher on the CIFAR10 data set and 99.03% higher on the MNIST data set. On the CSTV simulation experimental data, the fault diagnosis model proposed in this paper achieves 99.994% accuracy in fault diagnosis accuracy, convergence speed and stability due to the original model, and it does not require experimentals to debug the neural network parameters, which improves the efficiency of training neural network.