The Research On Theory And Approaches Of Neural Networks For Digital Filters Design And Harmonics Measurement

Digital filter, which is as the main cell of digital signal processing, has been widely applied to electronics systems, communication systems, image processing and other signal processing fields. At the same time, to guarantee secure stable operation of power systems, it has been an important requirement for measuring precisely the power systems harmonics. In this dissertation, the research is focused on seeking after the theory and approaches for designing FIR digital filters and measuring power harmonics based on neural networks technology.For one-dimensional (1-D) FIR linear-phase filters designing problem, a non-equiripple neural networks design approach is firstly presented which adopts cosine basis functions as the hidden layer neuron activation functions. For controlling effectively the errors of filters'passband and stopband, a novel equiripple neural networks design approach is put forward. The approach can implement equiripple filter design by adjusting the weighted error function at iterations. The convergence theorems proves that the neural networks algorithms can converge to its global minimum, and the design results show that they are more efficient and superior than some traditional methods.Two-dimensional (2-D) digital filters are of the use in digital image processing and other 2-D digital signal processing fields. The 1-D filters designing approaches are extended to 2-D FIR linear-phase filters design by according to the magnitude response of 2-D filters. Firstly, a neural networks approach for designing quadrant symmetric 2-D filters is presented, and the approach is extended to designing 2-D FIR filters with arbitrary magnitude responses. On this foundation, an equiripple 2-D FIR linear-phase filters design theory and approach is studied using neural networks. The approach can control effectively the passband ripple and the stopband attenuation. The simulation results show that the approach can design better 2-D FIR filters than other optimal methods, such as least-square method, close least-square method and semidifinite programming (SDP) method, et al.In some signal processing applications, the need arises for the design of complex FIR filters to meet some specifications that cannot be achieved by real filters. In this dissertation, complex FIR filters designing theory and approaches are studied. According to the frequency-response characteristics of 1-D and 2-D FIR digital filter, a novel strategy, which is used to design complex filters by minimizing simultaneously the magnitude and phase errors of filters, is proposed. According the strategy, the neural networks approaches for designing 1-D and 2-D complex FIR filters are studied, respectively. Design results show that they can obtain better design effects in passband ripple, stopband attenuation, and group delay error than SDP method, weighted least-square (WLS) method and minimax design method.With the extensive application of broadband digital systems, the design of narrow transition-width linear-phase FIR digital filters has been an important studying objective for broadband digital signal processing. In this dissertation, two narrow transition-width FIR filters design approaches with neural networks algorithm are studied using frequency-response masking (FRM) technology. According to the frequency-responses characteristics of FRM filters, a novel joint non-linear neural networks optimization approach is proposed, in which the set of FRM filter coefficients of all sub-filters is treated as a single design vector and the coefficients of overall sub-filters are optimized simultaneously. The design method is readily extended to multistage FRM filters. In addition, considering the highly non-linear optimal problem in designing FRM filter, a linear neural networks design approach is put forward in which the highly nonlinear problem with respect to the coefficients of the sub-filters is decomposed into several linear neural network optimization problems. Some simulation results show that the two design approaches can design better FRM filters than WLS method, SDP method and genetic algorithm design method.On power systems harmonics problems, a windowed FFT and neural networks method is firstly presented for power harmonic measurement. The harmonic analysis precision of this approach relies more on the precision of frequency correction using windowed FFT algorithm, so a new neural networks analysis method is proposed. In order to improve harmonic analysis effectiveness, a novel high precision and wide range power systems harmonic analysis approach is proposed base on neural networks algorithm. This method can obtain very high harmonic analysis precision by updating simultaneity the power frequency, magnitude and phase. The results show that in asynchronous sampling conditions, and for the considered frequency varying from 40 to 60 Hz, the fundamental frequency, and harmonic magnitudes and phases can be estimated at high accuracy. Considering the inter-harmonic problem in power systems, a windowed FFT and neural network based approach is proposed for inter-harmonic measurement. The simulation results show that the method can estimate accurately the harmonic and inter-harmonic parameters from the signals in power systems.