一种深层过程神经网络及其在信号分类中的应用

A Deep Layer Process Neural Network and Its Application in Signal Classification

针对时变信号小样本集建模分类问题,提出一种深层多尺度径向基过程神经网络(DLMS-RBFPNN)。该模型由时变信号输入层、多尺度径向基核变换层、全连接层和感知机分类器构成。兼顾时变信号的频谱特征和分布形态的多样性,基于径向基过程神经网络,通过将不同宽度参数的Gauss核函数进行线性叠加,构成多尺度核,完成不同尺度上对过程信号形态特征的提取、辨识和相似性度量。通过在径向基核函数层之上叠加全连接层和分类器,实现时变信号不同尺度特征的融合和分类。DLMS-RBFPNN具有较少的模型参数,适用于小样本集建模,在机制上可提高对时变信号过程细节特征和趋势特征的辨识及记忆能力。在分析DLMS-RBFPNN性质的基础上,建立一种基于动态聚类算法的核中心函数确定方法以及基于PSO的模型参数优化求解算法。以旋转机械基于示功图信号的故障诊断为例进行实验,结果验证了模型和算法的有效性。

Aiming at the modeling classification problem of small sample set of time-varying signals,this paper proposes a model based on deep layer multi-scale radial basis function process neural network(DLMS-RBFPNN).The model consists of a time-varying signal input layer,a multi-scale radial basis function kernel transform layer,a fully connected layer and a perceptron classifier.Taking into account the diversity of the spectral characteristics and distribution patterns of time-varying signals,the radial basis process neural network and the Gauss kernel function with different width parameters are linearly superposed to form a multi-scale kernel and complete the identification and similarity measure of process signal morphological features at different scales.The model realizes the fusion and classification of different scale features of time-varying signals,by superposing fully connected layers and classifiers on the radial basis kernel function layer.DLMS-RBFPNN has fewer model parameters and is suitable for small sample set modeling.It can improve the recognition and memory ability of time-varying signal process features.This paper analyzes the nature of the DLMS-RBFPNN model,proposes the center determination method of kernel function based on dynamic clustering,and constructs the model parameter optimization algorithm based on PSO.The fault diagnosis of rotating machinery based on dynamometer diagram is taken as an example.The actual data processing results show the effectiveness of the model and algorithm.