Analog Circuit Fault Diagnosis Approaches Using Neural Networks Based On Pattern Recognition Theory

The fault diagnosis of the analog circuits is an advanced synthesis intercrossed subj ect,which is an application technology absorbing the new theory,technology and method in other subjects and fields.In the past 40 years,it has developed a great deal of theory and approaches.But the approaches are limited to deal with fault diagnosis because of the variety and complexity of the analog circuits,especially the large-scale analog Circuits with the tolerance or the SOft fault.ArtifiCial Neural Networks (ANN)which have been one of the most active research areas recently can be contributed to solve problems in various practical fields.In this paper, analog circuit fault diagnosis approach based on pattern recognition,and ANN is expatiated.In the paper,based on the existing literature research foundation an analog circuit catastrophic fault location approach by using feedforward networks with back—propagation learning is realized.By this approach,the simulation require ments before test are reduced because fewer training samples are needed,and the fault location process is fast.This method is very efficient in location of single hard fault wit component tolerances.The measureme nt space feature and the general characterization concept of single and double soft fault in linear circuits are presented.According to this concept,a linear circuits soft fault location approach using subhidden layer BPNN is established with element tolerance,and it is shown that this approach is successful in fault location.A double fault feature extraction., method for linear circuits with element toleranceS are alSO presented.By this method,the double fault general characterization can be calculated by single fault general characterization which can be calculated by single fault feature.This method makes simulation before test more simple.According to the actual project requirements,practical BP algorithm is presented and realized on personal computer.After further optimization and improvements, a subhidden layer BPNN algorithm which support unlimited units is realized. The algorithm is powerful,simple and user—friendly,highly compatibility and expansion . According to the complexity of actual project , this algorithm allow users designing the structure of the neural network.