Reasearch On The Neural Networks Optimization Technique For Communication Channel Modeling

It has been an important research activity aimed at developing the theoretical foundations of interconnected communication techniques and control theory. The paper implements artificial neural network technique of intelligent control theory into channel modeling for communications. On the basis of obtaining the training data denoting channel characteristics through simulation, optimization techniques for neural based modeling are studied for several typical communication channels.Channels are the transmission media for communication system. In order to transmit information with high quality and great capacity on limited frequency bands, people need to grasp the radio propagation characteristics and its rules. Appropriate channel models should be established. Artificial neural networks are nonlinear dynamic system with the property of parallel information processing, adaptable processing and excellent learning and imitation. It has the characteristics of self-organizing and self-study, approaching any nonlinear system, getting the optimization solution satisfying kinds of restriction conditions rapidly under complicated environments, high robustness and capacity of fault tolerance. Compared with traditional channel modeling methods-theoretically computation and actual measurement, it has the virtues with decreasing the quantity of computation, adaptitude and high precision.Firstly, basic theory of artificial neural network and the key problems for neural based models development are studied. The approaches and process for modeling are analyzed. The training algorithms, struture forms and deficiency are compared for neural based modeling. According to the difficulty of getting training data, the sensitivity information and high complexity and nonlinearity for models, improvements of struture forms for channel modeling are proposed.Secondly, power line noise channel and multi-path channel optimization modeling techniques based on artificial neural network are proposed. In order to get the training and test data for neural networks modeling, power line multi-path channel model and noise channel (including colored background noise, narrow band noise and pulse noise) model are developed in Matlab and related data are analyzed. Clustered Markov chains are studied for pulse noise characteristics. Based on the training and test data, suitable network structures and training algorithms are chosed, and neural based background noise channel, narrow band noise channel, multi-path channel and pulse noise channel models are developed. The influences on the performance of neural models with different network structures and number of data are compared. Neural models are analyzed and tested and prove to be accurate and run fast.Thirdly, in order to reduce the design complexity of Gauss power density spectrum filter for DRM short wave channel, Gauss power density spectrum filter is designed through the neural network optimization technique. In order to get the training data for modeling, DRM receiver system is studied and simulated based on the DRM standard. Two-dimensional Wiener filtering estimation is selected for getting the training data for neural modeling. On the basis of these high quality data and channel estimation, neural based time domain and frequency domain models for DRM channel are developed. Due to the less computation and accuracy, neural models are good methods for speeding up the system simulation.At last, neural based wireless mobile channel modeling techniques are proposed. Neural network structure and frequency model for modeling multi-path channel are proposed. The neural model based CDF and PDF curve for Rayleigh and Rice fading channel are agreed with that in Matlab. According to the diagram of Doppler fading generator in 3GPP2, tap time-delayed neural network model is developed for generating time domain waves for the output of Doppler shading channel. Cosine based neural network is developed for the frequency domain characteristic of Doppler shading shape filter. The results are satisfactory. The neural model runs fast and accurately, and it has the advantage with easily controlled transition and stop band. The paper also developed the neural model for AWGN, path loss model suitable for 3GPP2. The curves of relationship between channel correlation coefficients and antenna interval, relationship between correlation of polarized antennas and different slant angles are approximated with neural model. From the modeling results, neural network is a fast and effective modeling method.