基于双神经网络结构的新型人工神经网络训练算法

A Novel ANN Training Method Based on Dual Neural Network Architecture

基于视觉神经系统生理学特征,提出了一种可加速神经网络训练过程的双神经网络结构,它同时提供了一种初始化神经网络的新方法。首先,对原始数据进行小波分解,将获得的近似系数在辅助神经网络中进行训练;然后将训练得到的权值和阈值传递给主神经网络;最后,利用主神经网络对全部的输入输出信号进行训练。通过非线性函数逼近、非线性动态系统辨识和井底压力预测仿真实验,并和常规的神经网络结构进行比较,结果表明:在达到相同精度的前提下,双神经网络结构极大地缩短了训练时间。

Based on physiological characteristics of human visual neural system,a new DNNA(dual neural networks architecture)was proposed to accelerate ANN training process,including a new approach to initiate weights and biases of the neural network.Firstly,having the original data decomposed with discrete wavelet transform;and then having the obtained approximation coefficients trained in an assistant neural network and the weights and thresholds obtained in training passed to the main neural network;and finally,having the main neural network employed to train all input/output signals.Through the nonlinear function approximation,nonlinear dynamic system identification,the simulation of bottom hole pressure prediction as well as the comparison with the conventional neural network structure,the results show that,the proposed DNNA can dramatically reduce the whole training time while preserving the same accuracy.