On the Construction of Neuromorphic Fault Dictionaries for Analog Integrated Circuits

Abstract

Methods of machine learning are actively used to construct neuromorphic fault dictionaries that provide the fault diagnostics of analog and mixed-signal integrated circuits in an associative mode. Many problems of the neural network (NN) training associated with the large amount of input data can be solved by reducing the size of the training data sets and using only their significant characteristics. In this paper, a route for the formation of a neuromorphic fault dictionary (NFD) is presented, a method based on the calculation of the entropy for choosing the significant characteristics of the training set is proposed, and the corresponding algorithm is developed. The results of the experimental studies for analog filter are shown demonstrating high efficiency of the proposed method: reduction by a factor of 192 in the NN training time, and coverage up to 95.0% of catastrophic faults and up to 84.81% of parametric faults by the resulting NFD in the course of diagnostics.