节点文献
基于小波神经网络的设备故障诊断方法研究
Research on Fault Diagnosis Method of Equipment Based onWavelet Neural Network
【作者】 孙士慧;
【导师】 赵仕俊;
【作者基本信息】 中国石油大学 , 检测技术与自动化装置, 2008, 硕士
【摘要】 神经网络以其固有的记忆能力、自学习能力以及强容错性为故障诊断问题提供了一个新方法。本文针对科学实验中广泛使用的平流泵的故障特点,深入研究了BP神经网络的故障诊断方法。首先用小波包分析技术做信号处理。选取db3小波函数,用硬阈值小波包降噪的方法将信号降噪,然后进行小波包分解与重构,以提取信号的能量特征向量,并将得到的特征向量作为神经网络的输入。本文采用具有一个隐含层的三层BP神经网络进行故障诊断,深入分析故障诊断的结果后发现:第一,网络容易陷入极小值而导致诊断失败;第二,网络的隐含层节点数难以确定。为了解决上述问题,本文研究设计了GA+BP算法。该方法是将遗传算法与神经网络相结合。首先,GA对BP神经网络做前期优化,确定出最佳网络结构及该结构对应的初始权值、阈值和网络的学习速率;然后,构造具有最佳结构和参数的神经网络来进行故障诊断。GA+BP算法的设计中,把每个染色体分解为连接基因和参数基因,对这两部分采取不同的遗传操作。连接基因采用二进制编码方法,参数基因采用实数编码方法;连接基因采用一点交叉方式和基本变异方式,参数基因中的权阈基因和速率基因各自采用算术交叉方式和非均匀变异方式。另外,交叉算子和变异算子都采用自适应的方法。GA+BP神经网络与BP神经网络故障诊断的结果对比后可以看到:第一,GA+BP神经网络比BP神经网络的工作量少,且克服了陷入局部极小的缺点,有更好的训练性能;第二,GA+BP神经网络的故障诊断准确率高于BP神经网络。由此可见,GA+BP神经网络能够更好的进行平流泵的故障诊断工作。
【Abstract】 Neural network offers a new method for fault diagnosis owing to its memory ability,self-learning ability and strongly fault tolerance. This paper makes research on the faultdiagnosis method of neural network deeply based on the fault characteristics of pump whichis widely used in experiment.Wavelet packet analysis is used to do the signal processing. Wavelet db3 is chosen, andall signals are de-noised by hard threshold de-noising method. Then wavelet packetdecomposes and constructs the energy eigenvectors which are regarded as the inputeigenvectors of the neural network.A three-layer BPNN is applied to do the fault diagnosis. The results of simulation showthat the network traps in local minimum easily, and both the number of hidden neurons andthe learning rate are difficult to decide either.In order to solve these questions above, this paper designs GA+BP algorithm. In thisalgorithm, genetic algorithm is used to optimize the number of hidden neurons, the initialweights and thresholds, and the learning rate of BPNN first, and then fault diagnosis is doneby this neural network which has the optimum structure and parameters. In GA+BP neuralnetwork, each chromosome is divided into the connection genes and the parameter genes, anddifferent genetic operations are carried on two parts. Connection genes are binary type andparameter genes are real-valued. Mixed crossover and mutation operations are operated on theconnection genes and parameter genes separately. It means the connection genes adoptsingle-point crossover and simple mutation, and the parameter genes adopt arithmeticcrossover and non-uniform mutation. Both the crossover and mutation operators adoptself-adaptive method. Comparing the simulation results of GA+BP neural network with BPNN, we know thatGA+BP neural network has less work but high training performance, and the local minimumis inexistent. In addition, the GA+BP neural network can diagnose the failure more correctlythan BPNN. In conclusion, GA+BP neural network can accomplish the pump fault diagnosismuch better.
【Key words】 fault diagnosis; wavelet packet; neural network; genetic algorithm;