Study On Fault Diagnosis In Analog Circuits Based On Constructive Neural Network

Analog circuit fault diagnosis has been studied for several decades, it always been the difficult and hot effected by the element tolerance, nonlinearity, temperature drift and other factors. Analog circuit has the most problems arise in electronic devices, though it has the small proportion. Electronic device reliability relies heavily on the reliability of the analog circuits. Traditional method has significant limitations, such as weak ability to process data, long-time diagnosis, complex and so on, under the background that the circuit scale increases. Artificial intelligence methods especially neural network method provide a new direction for analog circuit fault diagnosis, it adapt to the diagnose nonlinear circuits well, make diagnosis easy for it doesn’t rely on physical circuits. It is shortly in diagnostic time, accuracy, error correction and fault tolerant, and it modeling complex. Constructive neural network based on cover theory is a new neural network model proposed in recent years. It modeling simple, robust, and computing well when compared with other neural network, which suitable for industrial in complex environments with massive data, besides it can reduce computing time greatly.This paper is based on neural network theory, conquering the problems in exist fault-diagnosis systems, such as pre-processing data, extracting fault features, modeling complex, difficult to find and learn new fault-patterns and so on, using constructive neural network method based on cover theory to diagnosis analog circuit faults, the new system achieve good diagnostic results. First, we build the constructive neural network based on sphere neighborhood theory rely on the M-P neuron spherical model, which can diagnose analog circuit fault with 100% accuracy when it disturbed by ±4% interference. Then, for the problem fault can not be diagnosed with ±15% interference, we set a reject pattern and increase neurons to learn samples which can not be diagnosed, retrain neural network, it can diagnose the new fault and improve overall accuracy. In view of the problem that the practical industrial applications need to deal with the huge amounts of data and diagnosis system need to reduce optimization, we establish new neural networks based on domain cover and fuzzy-cover theory, diagnosis range from the biggest soft fault expand to all soft fault mode, accuracy can reach 89.3% and 89.3% respectively, and also can reduce the number of neurons, saves the overhead, reduce diagnose time, the calculation difficulty and the amount of calculation. We use fuzzy covering algorithm to diagnose biggest soft failure mode, improve the diagnostic accuracy reach 85.71% by single choice 100% by three choices. Experiments show that this method has good fault-tolerant and generalization ability, especially suitable for circuit fault diagnosis under complicated environment, which has good development prospects.