Development of parametric fault diagnosis algorithms for analog circuits using soft computing techniques

Abstract

Electronic systems for the applications of telecommunications, multimedia, biomedical equipments, networking, and real-time control systems are driving today's consumer electronics market. These systems require both digital and analog functionality. Integrated Circuits (IC) with digital and analog circuits are called as mixed-signal circuits. In a mixed signal circuit the majority parts are digital, but about 80 % of faults occur in the analog parts. The fault diagnosis of analog circuits is more complicated than the digital circuits. One of the challenges in analog circuit testing is the variability arising from the manufacturing process. The parametric faults or soft fault refers to parameter deviation in circuits that do not affect its connectivity but affect its characteristics. The circuit is still functioning, but some of its characteristic properties lie outside their allowable intervals. Soft faults are undesirable and they are difficult to detect. The presence of parametric faults in an analog circuit reduces the overall performance of the circuit and it is very difficult to detect. Therefore, parametric fault diagnosis problem provides an ideal platform to carryout research on related methods and that is the motivating factor of this research work. Fault diagnosis approaches for analog circuits are classified into two types, Simulation After Test (SAT) and Simulation Before Test (SBT). The SBT approach builds data dictionary through simulation and the data dictionary used to identify and locate faults. The SBT techniques are classified as fault dictionary technique and probabilistic technique. Fault dictionary technique is one of the popular techniques and can be divided into traditional fault dictionary method and intelligent fault dictionary method. The traditional fault dictionary approach is the simplest form in which it compares the measured response with responses recorded in the dictionary and when a matching response is detected, likely faults are identified. Since this method is based on direct comparison of the measurement results with a fault dictionary to diagnose fault, this method is time consuming, especially for large circuits. In Intelligent fault dictionary method, the fault dictionary is constructed based on machine learning techniques, such as Artificial Neural Network (ANN), Support Vector Machine (SVM), Wavelet Neural Network (WNN) etc., are used for model the networks and hence used to diagnose faults. This research attempts to develop a fault diagnosis models based on ANN, SVM and WNN classifiers combined with Bio inspired optimization algorithms to determine the optimum parameters for the classifiers. The developed fault diagnosis models are applied to two benchmark circuits Sallen key band pass filter and Tow Thomas low pass filter circuits. This research investigates and compares the performance of Artificial Neural network with back propagation training and weight optimized ANN through Genetic Algorithm (GA-ANN), weight optimized ANN through Crow Search Algorithm (CSA-ANN), and weight optimized ANN through Particle Swarm optimization (PSO -ANN) applied to two benchmark circuits such as Sallen key band pass filter and Tow Thomas low pass filter circuits. Simulation results indicate that the proposed PSO-ANN provides significant improvement in classification accuracy compared with ANN, GA-ANN and CSA-ANN methods. The research is further extended through implementing SVM classifier with Radial Basis Function (RBF) kernel for faults diagnosis in analog circuits. Also SVM combined with bio inspired optimization techniques Genetic Algorithm (GA), Crow Search Algorithm (CSA) and Particle Swarm Optimization (PSO) in order to find the optimal SVM parameter such as regularization parameter (C) and the width of the kernel (γ) of radial basis function. Experimental results for SVM, (GA-SVM),(CSA-SVM) and (PSOSVM) are detailed. Simulation results indicate that PSO based SVM model performs better than SVM, GA-SVM and CSA-SVM fault diagnosis models. A hybrid machine learning methodology, Wavelet Neural Network (WNN) in which Wavelet Transform is combined with Neural Networks is constructed and implemented to diagnose faults in analog circuits. In addition with WNN, the weights of WNN are optimized through the evolutionary technique Genetic Algorithm (GA). The simulation results obtained for WNN, GA based WNN are compared to assess the effectiveness of the algorithms with bench mark circuits.. Among the proposed methods, GA based WNN fault diagnosis model outperforms compared with other methods such as WNN, ANN, GA-ANN, CSA-ANN, PSO -ANN, SVM, GA -SVM, CSA-SVM and PSO - SVM models for fault diagnosis in analog circuits.