A modular fault-diagnostic system for analog electronic circuits using neural networks with wavelet transform as a preprocessor

We have developed a modular analog circuit fault-diagnostic system based on neural networks using wavelet decomposition, principal component analysis, and data normalization as preprocessors. Our proposed system has the ability to identify faulty components or modules in an analog circuit by analyzing its impulse response. In this approach, the circuit is divided into modules, which, in turn, are divided into smaller submodules successively. At each level, where a module is divided into submodules, a neural network is trained to identify the submodule that inherits the fault of interest from the parent module. This procedure finds the faulty component or module of any desirable size in an analog circuit by consecutive divisions of modules as many times as necessary. Our proposed approach has three advantages over the traditional neural-network-based diagnostic systems, which directly look for faulty components in the entire circuit. First, the performance of the modular systems is reliable and robust independent of the circuit size and can successfully classify similar fault classes with a significant overlap in the feature space where the traditional approach completely fails. Second, the modular approach requires significantly smaller neural network architectures, leading to much more efficient training. Third, for large real circuit boards, our diagnostic system proceeds to systematically reduce the size of the faulty modules until it is feasible to replace it.