Blind Detection Algorithm Based On Complex Hopfield Neural Network

After nearly 30 years, many substantial contributions to blind signals detection have been made in various scientific fields, such as speech source separation and recognition, biological signal processing, and wireless communication, etc. Unfortunately, some important theoretical and practical issues are still unsolved and they will be focused in this dissertation.The main contributions of this dissertation are as follows:(1) The blind equalization algorithms based on High Order Statistics (HOS) and Second Order Statistics (SOS) are introduced in chapter2. The disadvantages and unsolved problems of these algorithms are investigated. The HOS algorithms are based on adaptive equalization, so that the equalizers should be designed. The SOS algorithms used to blindly identify the channel firstly. A well appreciated shortcoming of these algorithms based on HOS is the large residual error often exhibited by higher-order statistical estimates, especially in the cases of those complex constellations of QAM. They often suffer from slow convergence and excessive data length requirement. Furthermore, the HOS algorithms are limited in Guassian noise case. On the other hand, those algorithms based on SOS require the conditions that the transmitted signal being independent and white with zero mean must be satisfied. While colored signals or the ill-conditioned channels such as those with zeros on the unit circle or with common zeros among subchannels, those algorithms tend to be very unreliable and may not be appropriate for sensor networks.(2) Considering HOS algorithms need large amount of data, and SOS algorithms showing poor performance in the case of the SIMO channels with common zeros, new algorithms are proposed based on optimization performance function of finite alphabet constraint. Based on the subspace relationship between received signals and transmitted signals, the performance function can detect signals blindly which directly use constellations in themselves but do not rely on second or high order statistics of constellation signals and received signals. Due to the new performance functions are not relied on Bezout Identity that is suitable to any signals, it shows a promise that the method will be applied better to any kind of channels.(3) Varieties of optimization algorithms are explored, including ? approximation algorithm, Particle Swarm Optimization (PSO), Immune algorithm (IA) and Genetic algorithm (GA). Simulations demonstrate that the optimization algorithms designed by this performance function are able to detect the two-state signals (such as BPSK, QPSK) satisfactorily, which get a better performance using a short receiving sequence. With the increased complexity of signal constellation, different levels of the signals produce different effect for the weights of the cost function. The algorithms will be easily trapped in local minima. And the algorithm complexities show that those methods may not be appropriate for MQAM and MPSK signals.(4) Three novel Multi-valued complex Hopfield neural network (MCHNN) constructed by the performance function based on alphabet constraint is proposed in this dissertation: "Complex Hopfield Neural Network with Real-Imaginary-type Soft-Multistate- activation-function, CHNN_RISM" and "Complex Hopfield Neural Network with Real-Imaginary-type Hard-Multistate- activation-function, CHNN_RIHM" for MQAM signals blind detection; "Complex Hopfield Neural Network with Amplitude-Phase-type Hard-Multistate-activation-function, CHNN_APHM" for MPSK signals blind detection. The Amplitude-Phase-type multistate activation functions and new energy functions for MCHNN are constructed. The special energy functions own the ability to describe the dynamics characteristics of the new MCHNN, which the energy functions in the existing references cannot explain. The stabilities for MCHNN with asynchronous and synchronous operating mode are also analyzed separately. Furthermore, to verify the effectiveness of MCHNN, the weighted matrix of MCHNN is constructed by the specific cost function for blind detection of signals. Because of the characteristics of the MCHNN, different levels of the signals produce equal effect for the weight of these energy functions,and the energy functions at the true solution point are fixed value relevant to the length of signals. The algorithms only use the information of the alphabet that the transmitted signals belonged. In addition to required the transmitted signals are "iid" and "belong to alphabet", the algorithms neither need more assumption of noise statistics nor other statistics information of transmitted signals. Simulation results show that the proposed MCHNN can be used to blindly detect dense MQAM constellation signals with shorter received signals (similar to SOS algorithms) and the global minimal value of the MCHNN energy function is verified, the algorithms are also proved to be appropriate for channel with common zeros.