Intelligent Detection Technologies Of Disturbing Signals For Tokamak Magnetic Fluids

Magnetic fluid cracking is one of the main threats of experimental equipment and personnel safety in Tokamak experiments. The disturbance signal diagnosis and classification is one of the most important methods of judging whether the magnetic fluid will be broken or not. The automatic and accurate recognition of the magnetic fluid disturbance signals caused by the magnetic fluid unstability using an intelligent detection technology is a hot topic for the researchers.In this thesis, combined with the characteristics of magnetic fluid disturbance signals caused by the magnetic fluid unstability, the improved S-Transform algorithm is used to extract the feature vector based on frequency feature points. The classification stability and accuracy rate are analyzed about the radial basis function neural network classifier and the hyper sphere support vector machine classifier that both are based on the PCACL(Principal Component Analysis and Cross Learning) K-means clustering algorithm. Finally, the article discusses the factors that affect the stability of magnetic fluid and proposes the framework design, the overall structure and the function model. The main jobs of this dissertation are as follows:1. In the link of signal feature extraction, the collected magnetic fluid disturbance signals are transformed from the time domain to the frequency-time domain by using S-Transform algorithm. The results of signal segment pretreatment are a group of feature vectors about frequency. Because the S-Transform algorithm can not obtain the optimal time resolution and frequency resolution at the same time, this dissertation adds adjustment parameters to balance both resolution. In order to select the best parameters, this thesis proposes the parameter evaluation criteria which are standards for energy loss degree. The experimental results demonstrate the effectiveness of the improvement.2. This thesis constructs a signal classifier of radial basis function neural network using PCACL K-means clustering algorithm which has the ability to separate the disturbance signals caused the interference from the ones caused by the magnetic fluid unstability. In this dissertation, the classifier activation function centers, the weight vectors and the variance are optimized asynchronously. The classifier activation function is obtained through the PCACL K-means clustering algorithm. The PCACL K-means clustering algorithm is an improved algorithm based on the enhanced K-means clustering algorithm, it has been provided to train the activation function of hidden layer centers. This improved algorithm improves the possibility that it converges to an optimal result which is irrelevant to the initial specified centers position. Experiments show that the radial basis function neural network classifier using PCACL K-means clustering algorithm has better generalization ability and classification accuracy rate than the one using enhanced K-means clustering algorithm, meeting the design targets of the control system.3. This thesis discusses a variety of popular support vector machine algorithms which forms a solid theoretical basis for the new algorithm by comparing their advantages and disadvantages. Most support vector machine algorithms are in pursuit of class interval as large as possible, which also embodies the structural risk minimization principle. In the case of multi-class classification, according to some support vector machine algorithms being not overcome the influence brought by the difference of the number of training samples, a new hypersphere support vector machine algorithm using PCACL K-means clustering algorithm is proposed. This support vector machine defines weighted coefficient to adjust the maximum isolation edge between the positive with the hypersphere and the negative with the hypersphere, which can effectively reduce the risk of miscarriage of justice when the number of training samples is sparse. The original hypersphere support vector machine centers are obtained by PCACL K-means clustering algorithm, which needs less training time. Also this thesis proposes an improved incremental learning method to train the hypersphere support vector machine. Experiments show that this new hypersphere support vector machine which is trained under circumstances the sample data are sparse reduces the max isolation edge value, but improves the generalization ability, satisfying with the design concept of the control system.4. According to the results of theoretical research, the framework design scheme, overall structure and function model of the system of remote intelligent magnetic fluid disturbance signal diagnosis system is discussed, which can provide a reference method for the detection and control of Tokamak magnetic fluid disturbance unstability. The successful operation of the supporting system verifies the correctness and feasibility of the theoretical research results.This dissertation is on the basis of “Active Control of Tokamak Resistive Wall Membrane Steady Operating Conditions”(NO. 2008CB717807)(Ministry of Science and Technology 973 Program) and the sub item “The Research on Some Basic Techniques of Tokamak Device”(NO.10875027) under subject of National Science Foundation “Ministry of Science and Technology ITER Plan(Domestic Supporting Research Plan)”. The research background is based on the Chinese academy of science nuclear fusion device HT-7 and the research results lay a solid foundation for deepening studies on nuclear fusion experiment and longtime discharge.