A New Method Of Analog Circuit Diagnosis Based On Independent Component Analysis

A new analog circuit fault diagnosis method is studied in this thesis. Analog circuit diagnosis is belong to the circuit test domain, which needs more efficient method with the circuit scale increasing and the data processing accelerating. Nowadays, most analog circuit diagnosis technology can’t meet the real requirement for the slow development in this area and it has been an abstraction in circuit test domain.At first, two test node selected methods have been introduced in this thesis. One is the node excluded method, which selected the expectations and variances of the fault fuzzy sets as the judging parameters, and the node that had the least expectation was selected first, when there were more than one nodes that had the same expectation, selecting the node with less variance. Fault modes were excluded after every iterative if they have been able to be identified, which can simplify the intersection computation by decreasing the number of the fault modes to be recognized. The simulation results showed that this method consumed less time and can find out the optimized nodes set. The other test nodes selected method was based on faults code using the fuzzy fault nodes set, which let the node with larger number of sets have the more priority. It stopped when all the fault modes had different codes with the least nodes number. Intersection computation of the fuzzy sets were not needed in this method, which made it possible to realize the automatism and real time of the system through programming to complete simple sorting work.Two methods were presented to solve the unaccountable fault modes in analog circuit when there is only one fault element of the testable circuit. In order to make the fault modes accountable, the voltage sensitivity sequences were abstracted to set up the fault dictionary for its invariability in linear circuit. The second method was adopted for most situation, any one of the elements’values varied from zero to infinity can be view as one fault by parameter analysis, and the fault feature trochoids can be illustrated to make the fault accountable which took the soft faults and hardware faults of the same element to be one fault mode. Hardware fault was translated into one of the special case of the soft faults.Based on independent analysis technology, a new feature extraction method was introduced in the thesis for its importance in analog circuit diagnosis. At first, the point of penetration between independent component analysis and analog circuit diagnosis was found. The application foundation of independent component analysis was based on the independent assumption of the data, and the observed data were the linear mixture or non-linear mapping function of these independent components. In the analog circuit, the variety of the circuit elements parameters are independent, and the circuit variable can be expressed as the function of the circuit elements, which is the basis of the feature extraction method of this paper. This method assumed that the signals had non-Gaussian distribution and were independent between each other. The method presented in this paper was different from the Gaussian distribution and correlativity assumption. Two feature extraction algorithms were described in the thesis by taking entropy and kurtosis as two non-Gaussian measurements. Projection pursuit were adopted to lower the data dimension. Projection pursuit was used to find the "most interesting" distribution that included the most valuable element of the data. In fact, the Gaussian distribution is the most random and has the least value for data processing, and non-Gaussian distribution is the most ordinal and has the best architecture feature. In this thesis, success fault classifier was achieved by making the multi-mode variables away from the Gaussian distribution, aiming to make the entropy and kurtosis maximized.In order to make the classifier applicable for non-linearity circuit, back propagation neural network was adopted to design the fault modes classifier which had three layers:feature vector was as the input, nonlinearity sign function was taken as the transfer function for the hidden layer and binary codes were applicable to the output layer. The classifier had intelligence after training by supervisory learning.Simulation results showed that the method presented in the thesis can be successfully used to identify the fault modes in our test circuits, including the multi-testable nodes circuits, the single-testable node circuits, the linearity circuit and the nonlinearity circuit.