Research On Deep Learning Based Channel Estimation Algorithms

Massive Multi-input Multi-output(Massive MIMO)technology can greatly improve the utilization of spatial dimensions,thereby improving spectrum efficiency and reducing energy consumption.It is considered to be one of the core technologies in the fifth generation mobile communication system(5G).Orthogonal frequency division multiplexing(OFDM)technology can effectively combat narrowband interference and frequency selective fading,and can achieve low complexity through Inverse Fast Fourier Transform(IFFT),which is widely used in practical systems.Either of these two technologies or the combined MIMO-OFDM technology will play a key role in 5G,B5G and next-generation mobile communication systems,and have broad application prospects.Channel estimation is the process of estimating certain parameters of the channel model based on the received data information,and is a key step in Massive MIMO/OFDM communication.Through channel estimation,the receiver can obtain the impulse response of the channel,thereby providing necessary Channel State Information(CSI)for subsequent precoding and beamforming in Massive MIMO,or coherent demodulation in OFDM.Therefore,the effect and accuracy of estimation will directly affect the overall performance of the communication system.There are some problems of traditional channel estimation algorithms,such as requirements for priori channel statistics,large pilot overhead,high complexity,and poor robustness.Moreover,in Massive MIMO scenarios,the parameters to be estimated increase sharply with the increase of the antenna number.At the same time,the effects of pilot contamination and low-precision quantization also need to be considered in practical systems.These factors above cause great challenges to the design of channel estimation algorithms.In recent years,artificial intelligence,especially Deep Learning(DL),has achieved remarkable results in the fields of computer vision,natural language processing,and speech recognition.Researchers in the field of wireless communication also expect to apply DL to intelligent communication systems.As a key step in communication systems,channel estimation has attracted much attention in the research of intelligent communication.Because the algorithm based on DL does not require priori channel statistical characteristics,and the iterative training can make the deep neural networks infinitely close to the actual channel scenario,it shows amazing performance advantages in reducing algorithm complexity,processing heterogeneous data,and robustness.However,research on the combination of DL and channel estimation is still in the early stage of exploration,and there are some problems.1)Most of the existing channel estimation algorithms based on DL only use the deep learning network as a black box,and are not optimized for the channel estimation scenario,which means that there is a problem of excessive training overhead.Since the channel is time-varying,the offline training network cannot be used for continuous data transmission,which need to be adjusted or retrained.2)Most of the algorithms based on DL are designed for a simple communication system,without considering the influence of practical factors.Especially in the face of largescale MIMO scenarios,the high parameter dimension will lead to an increase in the input dimension of the neural network and a complex network structure,which will affect the channel estimation performance.Therefore,we adopt the idea of deep learning and propose new channel estimation algorithms from these two perspectives.The first one is based on meta-learning,which is verified in an OFDM baseband system.Then an algorithm for a Massive MIMO communication system with pilot contamination and low-precision Analog to Digital Converters(ADCs)is proposed.For the first problem,a channel estimation algorithm based on Meta Learning is proposed.Meta Learning is a new DL strategy.The algorithm strategy of meta-learning is different from ordinary supervised learning.The way of defining task set and pre-training makes the algorithm have the ability of rapid convergence with few data shots.The proposed OFDM channel estimation algorithm based on meta-learning defines the channel model with different parameters as the task set.Through the meta-training of different sub-tasks,the DL network can not only learn the nonlinear mapping of the channel function,but also have good adaptability for each subtask.This makes the model converge quickly according to a small amount of pilot data in the current channel state,which improves the robustness and adaptability of the algorithm.For the case where the actual scene is inconsistent with the training data,an online fine-tuning scheme is proposed,which can solve the shortcomings of large offline training overhead.The simulation results show that the proposed algorithm can improve the transmission error rate of the OFDM system,and has robustness to different channel states and the number of pilots.For the second problem,channel estimation in a Massive MIMO system with pilot contamination and low-precision ADCs is considered.On the basis of the first research point,the channel estimation problem is decomposed into regression and interpolation.Among them,regression can be solved by the method of the first research point.For the interpolation problem,a method in the field of computer vision,i.e.,Super Resolution(SR)is introduced.The interpolation optimization problem is modeled as a SR reconstruction problem,where the result of the least squares channel estimation is uesd as the initial input.Then the Channel Transfer Function matrix calculated by the re-ceived pilot is regarded as a low-resolution "picture",which will be fed into the SR neural network to remove interference and noise.After that,the high-resolution "pic-ture" will be obtained.In this way,the proposed algorithm can be deployed in Massive MIMO systems with low precision ADCS,whcih means it can reduce power consumption,while ensuring the accuracy of CSI estimation.Simulation experiments show that the proposed DL channel estimation scheme can reduce the performance degradation caused by pilot contamination and low-precision ADC,thereby effectively improving the accuracy of channel estimation.