1. Massive MIMO technical background
Massive MIMO technology has spatial multiplexing gain, diversity gain and beamforming capabilities. By configuring multiple antennas at the transmitting end and receiving end, multiple reception and multiple transmission can be realized, and space resources can be fully utilized without increasing spectrum resources and transmission power. , doubles the channel capacity and reduces multi-user interference, which has significant performance advantages. However, these gains are based on the premise that the base station can accurately obtain uplink and downlink channel state information (CSI, channel state information).
2. Massive MIMO channel state information feedback faces challenges
The uplink CSI acquisition is relatively easy, the user end sends training pilots, and the base station estimates the channel response of each user based on the received signal; while the downlink CSI acquisition is more difficult, and it is also a difficult problem that needs to be solved.
In the time division duplex (TDD, time division duplex) system, the base station can perform channel estimation through the training pilots sent on the uplink, and then use the channel reciprocity to obtain the CSI of the downlink.
In the frequency division duplex (FDD, frequency division duplex) system, the uplink and downlink work on different frequency points, and the channel reciprocity is very weak. Therefore, the CSI of the downlink needs to be passed by the user end first. The downlink pilot is estimated and sent back to the base station through the feedback link.
Complete CSI backhaul needs to consume a lot of air interface resource overhead, so the protocol usually uses vector quantization (VQ, vector quantization) or codebook-based methods to reduce overhead, but these methods lose channel information to a certain extent, and their generation The amount of feedback will multiply as the number of transmitting antennas increases. In the massive MIMO system, the base station uses a large-scale antenna array, which makes the codebook design complexity and the corresponding feedback amount increase significantly. Therefore, the traditional feedback scheme is not advisable in the massive MIMO system.
3. Traditional technical means and problems
It is found that as the number of base station antennas increases, the channel matrix of the user terminal in the space-frequency domain can be expressed in a sparse form due to the limited local scatterers of the base station. Therefore, a CSI feedback scheme based on compressive sensing (CS, compressive sensing) is proposed.
Theoretically, the correlated CSI can be transformed into an uncorrelated sparse vector on some basis, and then use the CS method to randomly project it to obtain a dimensionality reduction measurement value; the measurement value occupies a small amount of resource overhead. It is transmitted back to the base station through the feedback link, and the base station recovers the original sparse channel vector from the low-dimensional compressed measurement value according to the CS algorithm.
By exploiting the space-time correlation of CSI, the CS method does not rely on statistics, which simplifies the compression process and reduces the feedback overhead to a certain extent. However, traditional CS-based methods still have the following three problems:
1) The CS method relies heavily on the prior assumption of the channel structure, that is, the channel satisfies sparsity on some transform basis, while the actual channel is not completely sparse on any basis, and may even have no interpretable structure;
2) The CS method uses random projections to obtain low-dimensional compressed signals, and does not make full use of the structural characteristics of the channel;
3) Most of the existing CS algorithms for CSI recovery are iterative algorithms, which have large computational overhead and slow running speed, and do not meet the real-time requirements of the actual system.
Therefore, a massive MIMO feedback mechanism that can highly compress CSI information to reduce transmission overhead and recover CSI quickly and accurately from the highly compressed feedback information needs to be proposed urgently, and the introduction of breakthrough new technologies is imminent.
4. Existing AI technical means
4.1 Deep learning-based CSI feedback architecture CsiNet
The structure of CsiNet is similar to an automatic encoder, including two parts: an encoder and a decoder.
- The encoder belongs to the user end and is used for CSI compression, that is, to compress the original N-dimensional channel matrix into M-dimensional codewords by using the sparse characteristics of the channel matrix.
- The decoder belongs to the base station and is used for CSI reconstruction, that is, to restore the received codeword s to the original channel matrix
The working mechanism of CsiNet can be summarized as follows: After receiving the channel matrix in the space-frequency domain, the user terminal obtains the truncated matrix through two-dimensional DFT, and then uses the encoder to generate a compressed codeword s; then the codeword s is returned through the feedback link After the base station receives the codeword s, the base station uses a decoder to reconstruct the channel matrix in the angular delay domain; finally, the restored channel matrix in the space-frequency domain is obtained by inverse DFT.
The network structure of CsiNet is shown in Fig.1.
4.2 CSI Feedback Based on Long Short-Term Memory Network (CsiNet-LSTM)
In many typical massive MIMO application scenarios, the channel changes slowly, and a frame of channel data collected has time correlation, which can be used to compress the channel matrix more efficiently.
The long short-term memory network (LSTM, long short-term memory) extends the CsiNet architecture to improve the network's compromise between compression ratio and restoration quality.
In the CsiNet-LSTM network structure, the two modules of CsiNet encoder and CsiNet decoder follow the network structure in CsiNet. CsiNet-LSTM uses two different compression rates when performing angle-delay domain feature extraction and restoration reconstruction on the channel matrix. The first CsiNet module employs a high compression rate, which enables to preserve sufficient structural information of the first channel matrix for subsequent high-resolution restoration. Since there is a correlation between the remaining channels and the first channel, which contain less effective information, the subsequent T-1 channel matrices can be coded with a low compression rate. Before restoration and reconstruction, the first high-compression-rate encoded codeword is concatenated in front of all low-compression-rate codewords, and the channel correlation information is fully utilized for decoding. The decoded output forms a sequence of length T and sends it to the 3-layer LSTM. The LSTM can implicitly learn the time correlation through the input of the previous moment, and then merge it with the input of the current moment, thereby improving the reconstruction of low compression rate. quality.
4.3 CSI feedback based on two-way channel reciprocity
In the FDD system, since the uplink and downlink work in different frequency bands, the channel reciprocity is not obvious. Therefore, the user end needs to feed back the downlink CSI to the base station. Since both the uplink and downlink channels can be expressed as a function of the physical environment composed of multipath and scatterers, existing studies have shown that there is a certain correlation between the two-way channels of the FDD system. Therefore, literature [25] focuses on It aims at studying the correlation between uplink and downlink CSI in FDD system, and using uplink CSI to improve the recovery accuracy of downlink CSI.
Using the correlation between the magnitude and absolute value of uplink and downlink CSI, two CSI feedback architectures, DualNet-MAG and DualNet-ABS, are proposed to reduce the feedback overhead of CSI and improve the reconstruction accuracy.
The user end first separates the amplitude and phase of the downlink CSI that needs to be fed back, and then feeds back the amplitude to the base station after being compressed and encoded by the encoder, while the phase is directly quantized based on the amplitude distribution and then fed back to the base station, so the total feedback overhead is A larger proportion of is from phase feedback. When performing CSI reconstruction, the base station jointly decodes the downlink amplitude obtained by feedback and the estimated uplink amplitude, and makes full use of the two-way reciprocity to improve the accuracy of amplitude reconstruction, and the downlink CSI can be recovered by adding the quantized phase obtained by feedback .
The network structure of DualNet-MAG is shown in Figure 3, and the structure of DualNet-ABS is similar to it.
4.4 CSI Differential Feedback Based on Time Correlation
In a massive MIMO system, the collected channel data sequence has time correlation, because the CSI feedback cycle is shorter than the channel coherence time [24], and time correlation can efficiently compress the channel matrix.
At the initial time t1, the differential feedback network needs to perform CSI high-precision feedback under the premise of a large compression rate, which is to provide more accurate prior information for subsequent CSI feedback. From time t2, the differential feedback network will feed back the prediction error. Compared with the process at time t1, the differential feedback network can achieve high feedback bandwidth efficiency with the support of prior information, and only uses a smaller compression rate.
5. National smart network open innovation platform - AI-based CSI compression feedback open task
5.1 Using AI to study the problems faced
The introduction of artificial intelligence represented by deep learning (DL, deep learning) technology into the massive MIMO channel state information feedback scheme provides a new design idea to solve the problem of CSI feedback, but the use of AI technology still faces the following problems:
- Researchers need to build simulation capabilities to build datasets
In order to meet research needs, researchers need to build simulations to output a large number of relevant data sets for research according to research needs, which consumes a lot of time on simulation capabilities.
2. The generalization of the model is not good, and abundant data is needed for generalization research
The generalization of the model has always been a hot research direction in the industry. For those who study the generalization of the model, they need data sets with richer scenarios for model training. It takes a lot of energy to build simulation capabilities, and it takes a long time to generate data. set.
3. Lack of systematic evaluation indicators for the network
Existing algorithm model evaluation is mostly based on reconstruction accuracy (such as NMSE or cosine similarity), and the model has not been applied to the system to verify network performance indicators. At present, there is a lack of a simulator that can embed the model into the communication system.
5.2 Introduction of AI-based CSI Compression Feedback Task
Based on the above research background, the "Smart Network" national new generation artificial intelligence open innovation platform has constructed an AI-based CSI compression feedback open task. Users can use the AI algorithm development and training environment provided by the platform, and use rich data sets online and flexibly The simulation capability of self-configuring parameters to build data sets for AI model training can meet various research needs such as model accuracy and model generalization. In addition, this task also provides a system-level simulation environment that can verify the effect of the AI model in the entire communication process, and provides a variety of system-level indicators for multi-dimensional verification and evaluation of the algorithm.
5.3 Advantages of Platform Open Tasks
Based on the pain points of AI research, the research task is opened. The advantages of the platform are as follows:
- The platform provides AI algorithm development and training environment
- The platform provides rich data sets
- The platform provides simulation capabilities for generating data sets, and parameters can be configured to generate data sets according to research needs
- The platform provides system simulation capabilities and various communication indicators to conduct online reasoning verification and evaluation of AI models
- The platform provides data interface and message interface to call data sets and simulation capabilities
5.4 Platform Open Task Research Objectives
This task hopes to attract AI algorithm researchers to participate in the study of communication AI algorithms. For this task, the algorithm is mainly inspected from the following aspects:
- Channel Feature Information Feedback Accuracy (NSME)
- Influence of Channel Feature Information Feedback on Network Performance
- The overhead of channel feature information feedback is the compression rate
- Model generalization under different channel environments
6. Summary
Deep learning (Deep Learning, DL) provides a new idea for solving the CSI feedback problem in massive MIMO systems. Although the CSI feedback technology based on deep learning makes up for the shortcomings of traditional methods to a certain extent, and has great potential in the application of future mobile communication systems, the challenges of wireless communication systems in terms of model generalization and ultra-low delay , requires deeper, more efficient, more accurate, and more universal AI algorithm output, and also looks forward to data and environmental support that meet the needs of such AI algorithm incubation.
Based on this, the AI-based CSI compression feedback task is opened on the open innovation platform to provide researchers with open simulation capabilities and data sets to meet various research needs.
Pay attention to the public account of Wangzhiquan on WeChat , and get the information of the open innovation platform as soon as possible