神经网络诊断模型的隐层压缩算法

Hidden Layers Compression Algorithm for Neural Network Based Diagnosis Models

基于多层前向网络的诊断模型在设备故障诊断领域应用比较广泛。但在多层前向网络的设计和训练问题上，单隐层的隐层单元数选取一直非常困难，一般采用试凑法，既费时又不一定保证收敛。本文针对这一难题，从设备故障诊断的工程实际出发，采用隐层压缩算法，获得了性能较好、网络规模又较小的神经网络诊断模型。给出了隐层压缩算法中原始网络的构造方法，并且证明只要原始网络的隐层单元数不小于训练样本数减１，该原始网络就可以学会全部训练样本。由于所构造的原始网络往往存在隐层冗余，隐层压缩算法可以迫使冗余隐层单元在训练中死亡，从而达到精简、压缩隐层的目的。文中给出的实例研究验证了该隐层压缩算法的有效性。

Diagnosis models based on multilayer feedforward neural networks are widely used in the field of machinery fault diagnosis.But how to determine the number of hidden units in the hidden layer is very difficult.In order to solve this problem,a hidden layers compression algorithm is proposed.With help of this algorithm,the neural network base diagnosis models which have better network performances and little network complexity are obtained conveniently.The method for constructing an initial network is given,it is proved that the initial network can learn all training examples if its hidden units are not fewer than all training examples minus one.The initital network,however,usually has redundancy in its hidden layer.The comperssion algorithm can force its redundant hidden units to die so that there are no useless hidden units after removal of dead hidden units.The example study in this paper indicates that this algorithm is effective.