A Neural-Network-Based Fault Diagnosis Method Of Analog Circuits And Its Application

A neural-network-based fault diagnosis approach for analog circuits can be considered as pattern recognition problem. Fault diagnosis is actually the mapping process from the neural network input to its output after fault features are extracted from the units under test (UUT). So, neural-network-based analog fault diagnosis method includes the processes of fault feature extraction from the signals of UUT and the construction of neural network architectures. At present, many literatures about the two processes of neural-network-based analog hard or single-value-soft fault diagnosis technique focus on the combination of wavelet analysis, evolutionary algorithm etc. and neural network, including front-processing and the neural network optimization techniques. However, few literatures compare the advantages and disadvantages of these optimization algorithms and discuss their suitable applications. Besides, the practicality of single-value-soft fault diagnosis is questioned, as the aim of analog fault diagnosis is to differentiate between the situation of the elements' parameters falling into their tolerance ranges and hard/soft faults. And so far, it is difficult to diagnose the overlapped faults. Therefore, after the difficulties of neural-network-based analog fault diagnosis approach are discussed in this dissertation, some new analog fault diagnosis methods and the technical solutions of relevant automatic test and diagnosis system (ATDS) are developed, which forms a systematic neural-network-based analog fault diagnosis method. The main contents and achievements of the paper are as follows.1. Wavelet method of analog fault diagnosis is discussed. Based on the wavelet coefficients obtained from the wavelet de-noising and decomposition of test node voltages of analog circuits, the fault features extraction methods based on the maximum absolute value of the components of wavelet coefficients, the squares of the components of wavelet coefficients and wavelet-fractal are proposed in this paper. The first two methods use wavelet transformation to de-noise and decompose the test node voltages of analog circuits, and then calculate the maximum absolute value or the squares of the components of wavelet coefficients, which are then preprocessed by principal component analysis (PCA) and normalization. These preprocessed data form fault features, which are used as the inputs of neural network to identify fault classes. The difference between wavelet-fractal method and the first two methods is that fault features of analog circuits are obtained by calculating box dimension of the de-noised signals for wavelet-fractal method. The advantages and disadvantages of the three methods and their suitable applications are discussed in detail in this paper and the correctness of the methods is further verified through diagnosis examples.2. A unified neural-network-based fault diagnosis of analog circuits is proposed. Based on the idea of unified description of hard faults and soft faults of analog circuits, the vectors consisted of the skewness and standard deviations (STDs) of the test node voltages of analog circuits are proposed to form fault features. At the process of simulation before test (SBT), conduct the parameter scan and alternating current (AC) analyses on the circuits under test (CUTs) to obtain test node voltages for the calculation of their skewness and STDs, which constitutes the fault feature vectors. And then, choose the training and testing samples of the neural network according to the trochoid trends in the two dimensional coordinate system, whose x-coordinate and y-coordinate are STD and skewness respectively. The trained neural network can correctly distinguish between the no-fault state and hard/soft fault state at the process of simulation after test (SAT). The diagnosis principles and its suitable application are discussed in detail in this paper. The correctness of the proposed method is further verified by diagnosis examples and its superiority is proved by the comparison of the proposed method to the relevant literatures.3. An optimized neural network based fault diagnosis method of analog circuits is discussed. In allusion to back propagation (BP) neural network (NN) easily falling into local optimum, genetic algorithm (GA)-BPNN-, Immune GA-BPNN (IGA-BPNN)-, particle swarm algorithm-BPNN-, grouped particle swarm algorithm-BPNN-based analog fault diagnosis methods are proposed. The four methods substitute the gradient descent algorithm and optimize the architectures of BPNN by GA, IGA, particle swarm algorithm and grouped particle swarm algorithm, respectively. The principle of "survival of the fittest" is introduced to BP neural network to improve its convergence in GA-BP neural network. IGA-BP neural network combines GA with the mechanism of immune system to optimize BPNN, which makes BPNN have global search capability. The velocity-position model of particle swarm algorithm is introduced to BPNN to improve its immature convergence, of which model is simple and easily realized. Grouped particle swarm algorithm is developed on the basis of particle swarm algorithm and has the characteristics of reorganisation and mutation, which ensure BPNN converge to its global optimal solution. The specific application and effects of the four methods are discussed in a comparing way in detail in this paper. 4. From the form and the causes of overlapped faults of analog circuits, the re-classification method of analog overlapped fault diagnosis is discussed and the fusion neural network based analog overlapped fault diagnosis method is proposed, respectively. For the re-classification method of analog overlapped fault diagnosis, the overlapped areas of fault features are divided out into new fault classes, firstly. Next, use the conventional neural network based analog fault diagnosis method to identify the re-classified fault classes. The fusion neural network based analog overlapped fault diagnosis method includes the steps of SBT and real-time diagnosis. At the step of SBT, the overlapped fault classes are distributed into different sub-neural-networks and different kinds of fault features are used to train the sub-neural-networks. The target outputs of the sub-neural-networks are used as the training samples of the fusion neural network to train the neural network. At the step of real-time diagnosis, the preprocessed signals of the sampled outputs of the CUTs are sent into the diagnosis system. And then, the fault classes can be determined by the outputs of the fusion neural network. The basic principles and applied conditions of the two methods are discussed in this paper, and the effectiveness of those methods is further verified by diagnosis examples.5. The technical solution of digital signal processor (DSP) based ATDS and ATDS's typical application solutions are given in detail. Under guidance of theory and methods of analog fault diagnosis, the experimental device of DSP-based ATDS is built. The basic principles and the software solutions of the ATDS's composition, the DSP-based mainboard, the excitation module and the decision-making module are listed by doing experiments on analog chips and analog integrated circuits in this paper. The effectiveness and correctness of the proposed method in this paper are further validated by the experimental results.