Blind Signals Detection Directly Using Hopfield Neural Network Based On Continuous Multi-valued Neurons

The development of the wireless communication technology will lead the time-varying feature of wireless channel be sharpening, the cases of fast time-varying feature of wireless channels become increase, the cases of short data communication become increase, and diversifies the modulation schemes and makes signal constellation dense to be an inevitable trend. All these factors require the blind signal processing(BSP) algorithms own the ability to meet some exacting requirements. Most of those BSP algorithms based on Artificial Neural Networks(ANN) are bound up with those cost functions of traditional BSP algorithms. Its essence is still the traditional cost functions of BSP algorithms playing active roles. Although part of such algorithms can reduce the degree of dependence on the data length, their computation complexity is very high and can only be used in those cases of simple modulation. This dissertation is focused on finding a method of blind signal detection directly using ANN, which has the ability to be fit for these cases of"do not rely on statistical information, short data communication and dense constellation". The research works can be summarized the following points.(1) A BSP algorithm whose empiric risk is composed by Constant Modulus Algorithm(CMA) and Constellation Match Error(CME) cost functions is improved based on the references. The improved cost function is formulated using linear support vector regression based on structural risk minimization principle whose empiric risk is composed by CMA and CME cost functions. Then equalizer can be obtained through Iterative Reweighted Least Squares(IRWLS) method for solving this optimization problem.(2) A new optimization problem of BSP is formulated. This optimization problem does not require any statistical information of the source signal sequence, and the special optimization problem is suitable to be solved using Hopfield Neural Network based on Continuous Multi-valued Neurons(HNNCMVN). Then a specific weight matrix is devised. By considering the constellation characteristics of the source signal, the different constraints of the optimization problem are given from the views of the polar coordinate and rectangular coordinate system. Meantimes, the change law of the HNNCMVN initiation step is shown.(3) A novel blind signals detection directly algorithm using HNNCMVN with Amplitude and Phase Continuous Activation(APCA) is proposed. Considering the characteristics of MPSK signals, two types of continuous multi-valued activation functions are designed and the methods of select their parameters is illustrated briefly. The new energy functions of synchronous and asynchronous mode of the HNNCMVN are derived proved, respectively. By considering the characteristics of QAM signals, the new continuous amplitude and phase multi-valued activation functions are designed, analyzed and discussed, respectively. The blind signal detection ability of HNNCMVN is conformed in the case of the signals with missing statistics information.(4) To solve the dense QAM signals blind detection issue from another point of view. An new algorithm using Hopfield Neural Network based on HNNCMVN with In-phase/Quadrature Amplitude Component Continuous Activation(IQACCA) is proposed. Firstly, the method of selecting the scope of amplification factor from the point of activation functions is shown in detail. Secondly, the method and proof of the special energy functions of synchronous and asynchronous mode of the HNNCMVN are shown, respectively. Furthermore, some theorems of the new energy functions are given out and proved. Thirdly, a new initial state vector approach of this specific HNNCMVN is designed to speed up the algorithm convergence rate. The applicability of detecting low-order QAM signals using higher order QAM activation function is verified, and analyzed these special phenomena and give out the solution finally.The research results show the blind signals detection directly algorithms using HNNCMVN can confront those cases of"do not rely on statistical information, short data communication and dense constellation"effectively.