Using neural networks to solve testing problems

Abstract

This paper discusses using neural networks for diagnosing circuit faults. As a circuit is tested, the output signals from a Unit Under Test can vary as different functions are invoked by the test. When plotted against time, these signals create a characteristic trace for the test performed. Sensors in the ATS can be used to monitor the output signals during test execution. Using such an approach, defective components can be classified using a neural network according to the pattern of variation from that exhibited by a known good card. This provides a means to develop testing strategies for circuits based upon observed performance rather than domain expertise. Such capability is particularly important with systems whose performance, especially under faulty conditions, is not well documented or where suitable domain knowledge and experience does not exist. Thus, neural network solutions may in some application areas exhibit better performance than either conventional algorithms or knowledge-based systems. They may also be retrained periodically as a background function, resulting with the network gaining accuracy over time.