Font Size: a A A

Based On Multi-nuclear Co-space Model Of Extreme Learning Machine Clustering Diagnosis Methods Research

Posted on:2016-01-12Degree:DoctorType:Dissertation
Country:ChinaCandidate:K GaoFull Text:PDF
GTID:1108330473467176Subject:Electrical engineering
Abstract/Summary:PDF Full Text Request
The digital circuit develops faster than the analog circuit because of its high degree of integration, wide application and stable reliability. People tend to study the fault diagnosis of digital circuit for the difficulty of analog circuits in testing and fault diagnosis technology, which invest more human, material and financial resources. The analog circuit takes in a little part of the whole circuit system study area. However, the reliability of the whole circuit system mostly depends on the reliability of the analog circuit, it is important to study the testing and fault diagnosis technology of the analog circuit.The current feature extraction methods have certain limitations, such as wavelet analysis, the main extraction, curve wave characteristics of principal component analysis, independent component analysis and other methods. In the process of feature data extraction in high dimensional data, it often needs nodes to increase the data extraction and the complexity of the network. The curve wave analysis affects the accuracy in processing of redundant data of the reconstruction. PCA method can reduce in a sample effectively, but it decreases the accuracy of the data for the minimum variance of projection. Independent component analysis to higher order statistics reveals hidden in the data independent of statistical nature, but whose objective function of optimization and the classification still exists problems. Classical neural network algorithms have some questions in the training time and classification accuracy, outstanding performance for low efficiency and recognition rate like BP neural network, RBF neural network, Hopfield neural network, ART neural network, LVQ learning vector quantization and support vector machine, etc.This paper presents a extraction method of the common spatial patterns (CSP),which is making classification diagnosis, combined with a fast neural network of the way to the overloading learning machine.The main content and innovation are:1. The CSP is introduced into the fault diagnosis of analog circuits proposed to extract features, and the method of analog circuit fault diagnosis based on the common spatial patterns:analog circuit fault types of sample data for PCA feature extraction; reduced the sample parameters and dimension of multiple channels; from the source data dimensionality reduction; feature data to obtain a small amount of data; by using CSP (MKCSP) method for multi class feature data in the same sub space on the map in order to find out the orthogonal projection matrix, whitening and transformed; reduced the overlap interval between fault modes so as to the maximum variance projection of multi class fault modes; improved the accuracy of fault diagnosis.2. The paper puts forward out of the fault diagnosis and classification method of machine learning:the study of model learning algorithm research based on a variety of artificial neural network; studied the algorithm and model simulation, neural network learning machine training performance and the effect of analog circuit fault diagnosis classification recognition rate; it doesn’t need to iterate until the numerical approximation of a function,which is difficult to grasp the training process, and the training time is relatively so long,that the fault diagnosis and recognition rate is relatively low.The paper is based on single-hidden layer feedforward networks, SLFNs by Huang etc,set the appropriate hidden layer nodes so as to input weights and hide layer deviation of random assignment, input layer and the hidden layer randomly generated and the weights of the connections between the neurons in the hidden layer threshold, minimized norm least squares solution as the output of the network weights by using Moore-Penrose generalized inverse.The process of learning does not require any adjustment, which only need to set the number of hidden layer neurons, so we can quickly obtain the global optimum that is only good generalization performance solutions.3. It combined with the analog circuit fault diagnosis method based on the classification of common spatial pattern feature extraction method and limit of machine learning, with the Sallen-Key band passed filter circuit simulation of neural network, through the common spatial patterns and overrun of machine learning research in the application of analog circuit fault diagnosis and classification. It is extrcated by the way to PCA with the low-pass filter circuit fault data of Sallen-Key, using CSP (MKCSP) method for multi class feature data in the same sub space mapping, and the kernel function operation in order to reduce the amount of computation of nonlinear systems by means of orthogonal and whitening operation on two sampling data matrix and matrix diagonalization, so as to obtain the optimal projection direction, achieved one of these categories the feature information of maximizing deviations.It combined with the analog circuit fault diagnosis method based on the classification of common spatial pattern feature extraction method and limit of machine learning, with the Sallen-Key band passed filter circuit simulation of neural network, through the common spatial patterns and overrun of machine learning research in the application of analog circuit fault diagnosis and classification.It is extrcated by the way to PCA with the low-pass filter circuit fault data of Sallen-Key, using CSP (MKCSP) method for multi class feature data in the same sub space mapping, and the kernel function operation in order to reduce the amount of computation of nonlinear systems by means of orthogonal and whitening operation on two sampling data matrix and matrix diagonalization, so as to obtain the optimal projection direction, achieved one of these categories the feature information of maximizing deviations.The paper is studied the actual number of analog circuits wireless data transmission terminal to realize fault diagnosis and classification.4. It’s used to oscilloscope measurements, circuit system in normal and fault state of the node voltage under different gain in the experiment,the application of common spatial pattern algorithm, the PCA extraction of feature information of albino diversification to construct spatial filter and independent, projection mapping feature with discreted, the establishment of multi category of spatial filter, and then used machine learning classifier training over learning and classification.Although the measurement results is not only by the change of the precision of measuring instruments, the instability of measurement error, time, different manufacturers PCB plate material (copper), the influence factors of electronic components parameters and the stability of the input signal, but also the better the theoretical simulation results. However, this method is much better than other neural network methods, to improve the effect of classification and the diagnostic ability.
Keywords/Search Tags:Analog circuits, Fault diagnosis, Extreme learning machine (ELM), Kernel principal component analysis(PCA), Common spatial patterns (CSP)
PDF Full Text Request
Related items