Comparison of Euclidean distance based neural networks for analog Integrated Circuits fault recognition- LVQS &SOM

The advent of integrated circuits (ICs) and hence the subsequent miniaturization of electronic circuitry has brought out considerable difficulties encountered during identification of faults in integrated circuits during the testing phase of manufacturing and subsequent mass production. Artificial neural network (ANN) augurs well in handling such complex tasks in such systems as it generalizes well without the need to explicitly define the relationship between variables. There has been resurgence in interest among researchers in utilizing ANN for recognizing faults in analog circuits. This work aims at analyzing the role played by the various training parts of both the Euclidean distance based ANNs namely, the self organizing feature maps(SOM) and various versions of learning vector quantization neural network (LVQNN)i.e., LVQ1, LVQ 2, LVQ 2.1 and LVQ. Extensive studies have been conducted to ascertain the role played by learning rate and other unique parameters such as the role played by normalization as a part of preprocessing technique and the number of iterations for convergence. Moreover the results have been compared with the generalized multilayer feedforward network with back propagation algorithm. The best combination of network parameters was also determined. For this purpose an analog filter circuit with 1 fault free and 10 single hard faults was simulated using SPICE simulation software. Experimental results demonstrate the high classification accuracy and the adaptability of both the Euclidean classifiers and its suitability for fault recognition in analog circuits.