Design Of A Blind Equalizer Based On Long Short-term Memory Neural Network

In the modern era where the development of science and technology is increasingly important,channel equalization as an important technology for improving communication quality,has always been the focus of scholars.In order to solve the problem of inter-symbol interference,equalization techniques are usually used.In many cases,there is no condition for adaptive equalization.Blind equalization does not require training sequence and can be applied in many scenarios.Therefore,improving the performance of the blind equalizer is of great significance.In this thesis,blind equalization and machine learning related technologies are combined to design an improved neural network blind equalizer,which is mainly improved from the training algorithm and the structure.In terms of training algorithms,the advantages and disadvantages and performance of several gradient-based optimization algorithms are compared.According to these algorithms,this thesis simulates these optimization algorithms.The simulation results show that the algorithm for reducing the variance of the learning rate in the initial stage of training is superior to other optimization algorithms in terms of convergence speed and mean square error performance.According to the structure of the blind equalizer,the recurrent neural network is selected as the basic structure of the blind equalizer.Aiming at the disadvantage that the gradient of the simple recurrent neural network disappears and it is easy to fail to converge,the long-short term memory network in the simple recurrent neural network is analyzed,and its structure is improved.The blind equalizer of improved nested long short-term memory neural network can process information more reasonably.Then the simulation was also implemented,and the simulation results show that blind equalizer of the improved structure has better performance.The design and verification of the hardware circuit of the improved blind equalizer were carried out,the obtained steady-state mean square error is-33.97 d B;Python's simulation steady-state mean square error reached-34.63 d B and converged after 1850 iterations.In summary,the expected results have been achieved.