选择用于平台罗经故障检测的神经网络结构的方法

Model Selection in Neural Networks for Stabilized Gyrocompass Failure Detection

在基于神经网络的平台罗经故障检测中,为了提高故障检测的灵敏度,要求神经网络拟合平台罗经动态系统时均方预测误差的均值及其标准差均小。影响神经网络对系统拟合精度的因素既有隐层节点数也有输入延迟数。本文以均方预测误差的均值和标准差共同作为评价神经网络逼近平台罗经动态系统性能的指标,并借用系统化交叉证实法的结构,且增加一个外循环用以同时选择输入延迟数,构建选择用于平台罗经故障检测的神经网络结构的方法。

In some applications of neural networks, such as stabilized gyrocompass failure detection based on neural networks, in order to promote the failure detection sensitivity, both the average and the standard deviation of mean squared prediction errors are needed to be small. The precision approximation to the stabilized gyrocompass by the neural networks is influenced by the number of both hidden nodes and input delays. So that, instead of only the average of prediction errors, both the average and the standard deviation of mean squared prediction errors should be used as criteria to select not only the number of hidden nodes but also the number of input delays as well simultaneously. Making use of the framework of cross validation, such a strategy of model selection in neural networks was designed in this article.