基于RSPNN的制粉系统故障诊断  

Fault Diagnosis Based on RSPNN for Pulverizing Systems

作　　者：费树岷[1] 李延红[1] 柴琳[1] 

机构地区：[1]东南大学自动化学院,江苏南京210096

出　　处：《控制工程》2012年第3期412-415,共4页Control Engineering of China

基　　金：国家自然科学基金(No.60804017;60835001;60904020;60974120);教育部博士点基金(No.20070286039;20070286001)

摘　　要：针对发电厂制粉系统故障与征兆对应关系复杂及过程信息的不确定性及传统BP神经网络故障诊断的缺点,提出了基于粗糙集概率神经网络(RSPNN)的制粉系统故障诊断方法,以改善传统BP神经网络初始值敏感、易使学习过程陷入局部极小值以及样本数据过大时训练速度慢等问题。首先采用自组织映射神经网络(SOMNN)对连续样本数据进行离散化;再利用基于区分矩阵的HORAFA算法对离散化样本数据进行RS属性约简,并将约简结果作为概率神经网络(PNN)的输入;最后利用PNN作为诊断决策分类器,输出故障模式,并进行了仿真研究。仿真结果表明,该方法不仅优化神经网络的拓扑结构,降低神经网络的训练时间,而且能准确、快速地诊断制粉系统故障类型,同时对发电厂制粉系统及其相关设备的在线故障诊断问题有一定启发性。Due to the complicated relationship between the faults and corresponding symptoms of pulverizing systems, uncertainty of information, and the shortcomings of the general BP learning algorithm is training neural networks, a fault diagnosis system based on rough sets probabilistic neural networks （RSPNN） is proposed to deal with the traditional problems appearing in fault diagnosis techniques such as the sensitive initial value, the learning process into a local minimum and the slow training. Firstly, continuous attributes are quantized by SOM. Secondly, a HORAFA method based on distinguish matrix is used in the heuristic reduction of RS to reduce the samples as the input of the probabilistic neural networks （PNN）. Then, the PNN is used as a classifier to predict fault. The simulation results show that the method optimizes the structure of neural network, decreases the computation complexity, improves the diagnosis correctness, and provides inspiration about on-line fault diagnosis for pulverizing system and related equipment.

关 键 词：制粉系统 故障诊断 粗糙集 概率神经网络 

分 类 号：TP27[自动化与计算机技术—检测技术与自动化装置]