A deep learning-based massive MIMO signal detection method

Abstract

The application discloses a large-scale MIMO signal detection method based on deep learning and belongs to the technical field of communication.The GROAMP-Net intelligent detection network model is designed and built, the OAMP iteration method is combined with the deep learning thought, and the detection precision is obviously improved; due to the setting of the trainable parameters, the adaptability of the model is greatly enhanced and is more flexible, and the model can be applied to time-varying channels; meanwhile, the introduction of the residual structure and the gate unit effectively solves the problem that the network performance is degraded due to the error accumulation in the iteration process with the deepening of the network, the value of the trainable parameter is kept in a reasonable range, and the accuracy of the model detection process is further ensured.The algorithm has obvious advantages compared with the existing algorithm, and can solve the problem of insufficient detection precision of the traditional non-intelligent signal detection method and the existing intelligent signal detection algorithm.

Description

TECHNICAL FIELD

[0001] The application is a signal detection method of a wireless communication system, in particular, a large-scale multiple-in multiple-out (MIMO) communication system signal detection method based on deep learning, belonging to the field of communication technology. BACKGROUND

[0002] Large-scale MIMO technology has become one of the core technologies supporting 5G and is also considered a key enabling technology for 6G because it has many advantages such as improving system capacity, spectral efficiency, user experience rate, reducing inter-user interference, enhancing full-dimensional coverage, and saving energy consumption.

[0003] However, the development and application of large-scale MIMO technology also faces many problems. Because the actual wireless communication transmission environment is very complex, when the electromagnetic wave carrying useful information propagates in it, it will not only be interfered by external noise sources and other electromagnetic waves, but also be blocked by various obstacles in space and affected by the geographical environment, such as inter-symbol interference (ISI) caused by multipath effect. How to solve a series of problems represented by ISI is a problem that large-scale MIMO technology must overcome. Nowadays, the mainstream solution is to perform signal reconstruction at the receiving end, which can reduce the impact of various interferences on the performance of the communication system. This process is called signal detection. For large-scale MIMO communication systems, signal detection technology can improve the overall performance, so signal detection technology occupies an important position in the research of large-scale MIMO.

[0004] The signal detection technology has been researched for decades since it was proposed. The classical detection methods mainly include linear and nonlinear methods. The typical representative of the nonlinear detection method is the maximum likelihood (ML) algorithm, and the linear detection method mainly includes the matching filter (MF), the zero force (ZF) algorithm, and the minimum mean square error (MMSE) algorithm. In the large-scale MIMO system, these classical methods have fatal problems. For example, the ML algorithm is difficult to be practically applied due to its extremely high complexity, the MF method has poor performance in the environment with a large number of antennas, the ZF algorithm does not consider the channel noise, and the MMSE has relatively high computational complexity. Researchers introduce the iterative approximate solution idea into the signal detection field on the basis of the classical algorithm, so as to reduce the computational complexity of the algorithm as much as possible without losing too much detection accuracy. However, in practical application, the overall performance of the iterative detection method will be significantly degraded with the increase of the number of users. Obviously, in order to meet the communication requirements of the large-scale MIMO system, the signal detection method needs to be further improved and optimized.

[0005] With the rapid development of artificial intelligence technology, especially deep learning, researchers introduce deep learning into the signal detection field to construct a detection network, which greatly improves the detection performance compared with the traditional method and the iterative detection algorithm. However, the signal detection network based on deep learning still has problems. For example, DetNet introduces gradient descent into the traditional ML algorithm and combines the deep learning idea to achieve a detection accuracy much higher than the traditional method, but the number of parameters to be trained is related to the product of the number of transmit and receive antennas, which will cause a large actual time cost in the large-scale MIMO system. OAMP-Net is a deep detection network based on the OAMP iterative method, which increases the detection accuracy while keeping the algorithm complexity unchanged. Moreover, the model is more flexible through the trainable parameters, which can better adapt to the time-varying channel. However, in OAMP-Net, the estimation results of each network layer still have deviations, and the comprehensive performance of the model is sensitive to the key parameters, and unreasonable parameters will obviously affect the detection accuracy. SUMMARY

[0006] The technical problem to be solved by the present application is to overcome the deficiencies and defects of the prior art, and to provide a large-scale MIMO signal detection method based on deep learning assistance, which solves the problem that the detection accuracy of the existing signal detection algorithm cannot meet the requirements of the large-scale MIMO communication system.

[0007] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] A large-scale MIMO signal detection method based on deep learning, comprising the following steps:

[0009] Step 1, generating a training data set and a test data set;

[0010] Step 2, constructing a GROAMP-Net intelligent signal detection network model, which is based on an improved OAMP iterative method to recover the original transmitted signal;

[0011] Step 3, inputting the training data set generated in step 1 into the GROAMP-Net intelligent signal detection network model constructed in step 2 for offline training, observing the loss function curve to ensure that the model converges well; using the test data set to test the converged model, and obtaining the trained GROAMP-Net intelligent signal detection network model after the test is passed;

[0012] Step 4, using the trained GROAMP-Net intelligent signal detection network model for large-scale MIMO signal detection.

[0013] Further, the specific way of step 1 is:

[0014] In the Python environment, a large-scale MIMO channel model of Rayleigh fading is built, the number of transmit and receive antennas is set to 32, a bit stream is generated by calling a random function, a 4QAM modulation method is selected to modulate the bit stream into a symbol sequence and introduce additive channel noise, and the corresponding transmitted signal, channel matrix, received signal and noise power are generated in a 10dB signal-to-noise ratio environment for training data set; the relevant data generated in the signal-to-noise ratio environment of 0-9dB and 15-20dB is used as the test data set.

[0015] Further, the GROAMP-Net intelligent signal detection network model is a model-driven deep learning network, which includes ten serial layers with the same structure; each serial layer completes the same linear estimation and nonlinear estimation steps, and the entire network structure is equivalent to the optimized OAMP iterative process; wherein the input of the first layer is the original received signal, the channel matrix, the noise power, and the input of each subsequent layer is the original received signal, the channel matrix, the noise power, and the transmitted signal estimation result output in the previous layer.

[0016] Further, four trainable parameters are introduced in each layer of the GROAMP-Net intelligent signal detection network model, two of which are used to adjust the change step of the linear estimation and nonlinear estimation steps in the OAMP iterative method, and the other two are used as residual coefficients to construct a residual structure; the residual structure is used to correct the results of linear estimation and nonlinear estimation, thereby reducing the influence of the error caused by iterative approximation on the overall performance of the network.

[0017] In addition, a gating unit is arranged in each layer of the GROAMP-Net intelligent signal detection network model, which is used for controlling and adjusting trainable parameters of the two steps of linear estimation and nonlinear estimation, so as to prevent unreasonable values of the trainable parameters in the training process.

[0018] Further, the gating unit is composed of two control gates, and the system functions of the two control gates are respectively a sigmoid function and a Leaky-ReLu function; the control mode of the gating unit on the two trainable parameters is that the two trainable parameters are multiplied after passing through the two control gates, and then a new value is obtained and used for subsequent processing.

[0019] Further, in step 3, the optimization mode of offline training is Adam, and the selected loss function is the sum of the mean square errors of the output estimated value and the true value of the transmitted signal.

[0020] Further, the specific mode of step 4 is:

[0021] The trained GROAMP-Net model is deployed at the receiving end of the large-scale MIMO communication system, and the received signal, the channel matrix and the noise power are taken as inputs, and after processing by the trained model, the output value is the recovered transmitted signal estimated value.

[0022] The GROAMP-Net of the present application has the following advantages:

[0023] (1) Compared with the traditional signal detection method for large-scale MIMO system, the GROAMP-Net of the present application is an intelligent detection network constructed according to the deep learning idea, and not only the operation ability is improved, but also the introduction of trainable parameters enhances the flexibility of the network model.

[0024] (2) Compared with the existing large-scale MIMO system signal detection algorithm based on deep learning, the GROAMP-Net of the present application avoids the "gradient disappearance" and "gradient explosion" phenomenon caused by deepening the network layers through the residual structure, and at the same time, the gating unit controls the value change of the training parameters within a reasonable range, so that a higher detection accuracy can be achieved. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is the network structure diagram of the GROAMP-Net of the present application.

[0026] Figure 2 is a schematic diagram of the gating unit structure in the GROAMP-Net of the present application.

[0027] Figure 3is a broken line graph of the symbol error rate detected by the GROAMP-Net of the application and several reference algorithms with the change of the signal-to-noise ratio under the condition of 0-9dB signal-to-noise ratio.

[0028] Figure 4 is a broken line graph of the symbol error rate detected by the GROAMP-Net of the application and several reference algorithms with the change of the signal-to-noise ratio under the condition of 15-20dB signal-to-noise ratio. DETAILED DESCRIPTION

[0029] The application will be described in detail below with reference to the accompanying drawings.

[0030] A large-scale MIMO system signal detection method based on deep learning, the steps are as follows:

[0031] Step 1, build a large-scale MIMO channel model, generate and obtain the required subsequent transmission signal x, receiving signal y, channel matrix H and other related data, and construct training data set and test data set;

[0032] Step 2, expand the OAMP iterative detection process, and build a deep detection network model GROAMP-Net by combining deep learning related methods, in which, on the one hand, trainable parameters are added to realize the intelligentization of the network, and on the other hand, the trainable parameters and the intermediate values and output values in the iteration are corrected and optimized;

[0033] Step 3, input the training data set generated in step 1 into the GROAMP-Net intelligent signal detection network model constructed in step 2 for offline training, observe the loss function curve, and ensure that the model converges well; test the converged model using the test data set, and obtain the trained GROAMP-Net intelligent signal detection network model after the test is passed;

[0034] Step 4, use the trained GROAMP-Net intelligent signal detection network model to detect large-scale MIMO signals.

[0035] The specific implementation mode of GROAMP-Net in step 2 is as follows:

[0036] Introduce trainable parameters γ t and θ t , which are used to adjust the transformation step of the linear estimation and the nonlinear estimation in the OAMP iterative process, respectively.

[0037] Introduce residual structure to correct the intermediate value obtained by the linear estimation and the output value obtained by the nonlinear estimation which need to be iterated for multiple times, and prevent the network performance from degrading due to error accumulation.

[0038] The gating unit structure is constructed to control and optimize the trainable parameters by controlling the gate idea to prevent unreasonable values and protect the network performance. The gating unit is composed of control gates with sigmoid and Leaky-ReLu system functions.

[0039] In step 3, the training process is an end-to-end offline training process, a single training is 200 epochs, the training round is 20, the batch size of reading data is 1024 each time, and the Adam optimization algorithm is used to optimize the training process to minimize the loss function.

[0040] The method is based on the OAMP iterative method, combined with deep learning theory, and each iteration process is equivalent to a layer of neural network, thereby constructing an intelligent detection network model. By adding trainable parameters to enhance the adaptability and flexibility of the model, and introducing residual structure to correct the estimated value and output value of each network layer, setting the gating unit to optimize the value of trainable parameters, the method realizes higher detection accuracy compared with OAMP iterative detection and OAMP-Net without increasing additional computational complexity.

[0041] The following is a more specific example:

[0042] A deep learning-based large-scale MIMO system signal detection method, comprising the following steps:

[0043] Step 1: Construct a large-scale MIMO communication system, set the number of transmit and receive antennas M and N to be 32, use 4QAM modulation, generate channel noise through a random function, and build a system model. Generate the transmit signal x, channel matrix H and receive signal y through the built large-scale MIMO system model, which is used as the data set required for subsequent detection network model training, verification and testing. The signal-to-noise ratio of the training environment is set to 10dB, and the signal-to-noise ratio interval of the test environment is 0-9dB and 15-20dB.

[0044] Step 2: Construct the GROAMP-Net network model as shown in Figure 1 The model is composed of multiple identical series of layers, which means that the operation process is the same but the input parameters are different. The tth series of layers takes y, H, noise variance σ 2 and the output value of the previous layer as input to complete the calculation process contained in the OAMP algorithm.

[0045] Specifically, the calculation process of the OAMP algorithm mainly includes four steps. First, calculate the error variance v t 2 of the linear estimate, then get the intermediate value r t through the linear estimate, and then correct the estimated value and output value of each network layer according to vt 2 Update the error variance t of the nonlinear estimate t 2 Finally, based on r t and t t 2 The output value of this layer is obtained by performing nonlinear estimation. GROAMP-Net is based on the OAMP algorithm, combining deep learning concepts to transform the iterative process of OAMP into a deep network structure, by setting the trainable parameter γ. t and θ t To influence precision r t and t t 2 Providing appropriate transformation step sizes enhances the adaptability and resilience of the network structure.

[0046] Still see Figure 1 The GROAMP-Net network model introduces a residual structure to handle the r that needs to be updated multiple times during the iteration process. t and On the one hand, it performs error correction; on the other hand, it sets up a gating unit to control the trainable parameter γ. t and θ t Optimize it to keep its value within a reasonable range.

[0047] The residual structure is achieved by introducing two trainable parameters α, each initially set to 1. t and β t Achieve, α t and β t Used for processing and r t The specific processing method is shown in the following formula.

[0048]

[0049] r t =r middle +(1-β t )r middle (2)

[0050]

[0051]

[0052] in, here For linear MMSE matrices

[0053]

[0054] r middle and respectively are intermediate values in the linear estimation and nonlinear estimation processes under the residual structure, η t denotes the nonlinear estimation process of the t-th layer of the network model.

[0055] The gating unit structure as shown in Figure 2 is used to realize the optimization of γ t and θ t , which consists of two system functions controlled by the gating unit, sigmoid and Leaky-ReLu, and is converted into the formula

[0056] γ t = sigmoid(γ t )·Leaky-ReLu(γ t ) (5)

[0057] θ t = sigmoid(θ t )·Leaky-ReLu(θ t ) (6)

[0058] Step 3: The GROAMP-Net network model is trained on the relevant channel data generated by the large-scale MIMO system model in step 1, and the optimization method is Adam, and the loss function is the sum of the mean square error of the output estimation value and the true value of the transmitted signal x, and the specific expression is as follows:

[0059]

[0060] Step 4: Select the OAMP algorithm similar to GROAMP-Net, the OAMP-Net algorithm, the representative deep learning-based intelligent detection network DetNet, and the classic linear signal detection algorithms ZF and MMSE as reference algorithms, and test them on the relevant channel data generated by the large-scale MIMO system model in step 1, and observe the detection accuracy of the algorithms with the symbol error rate as the index. The results are as follows:

[0061] As shown in Figure 3 , the symbol error rate detection process of several reference algorithms changes with the signal-to-noise ratio in the 0-9dB signal-to-noise ratio environment, and it can be seen that in the low signal-to-noise ratio environment, various detection algorithms will be severely affected by noise, and the detection accuracy is low, and there is no obvious difference between different algorithms. But with the increase of signal-to-noise ratio, GROAMP-Net is more sensitive to channel quality improvement compared with reference algorithms, and its symbol error rate decreases more, and the detection accuracy gradually shows an advantage compared with other reference algorithms.

[0062] As shown in Figure 4The detection error symbol rate of GROAMP-Net can be quickly stabilized at a lower level in a high signal-to-noise ratio environment, the detection accuracy is slightly better than that of OAMP-Net and OAMP algorithm, compared with OAMP-Net, about 1dB performance is achieved, compared with the typical deep learning network-based DetNet algorithm, more obviously better than the two classical algorithms ZF and MMSE, and in general, the higher the signal-to-noise ratio, the more obvious the advantage.

[0063] In summary, in view of the problem of insufficient detection accuracy of traditional non-intelligent signal detection methods and existing intelligent signal detection algorithms, the GROAMP-Net intelligent detection network model is designed and built, the OAMP iterative method is combined with the deep learning idea, the detection accuracy is significantly improved, the adaptability of the model is greatly enhanced and more flexible due to the setting of the trainable parameters γ and θ, and the model can be applied to time-varying channels, at the same time, the introduction of the residual structure and the gating unit effectively solves the problem of network performance degradation caused by error accumulation in the iteration process with the deepening of the network, the value of the trainable parameter is kept within a reasonable range, and the accuracy of the model detection process is further ensured. The simulation results show that the present application has obvious superiority compared with the existing algorithms.

Claims

1. A method for detecting large-scale MIMO signals based on deep learning, characterized in that, Includes the following steps: Step 1: Generate the training dataset and the test dataset; the specific method is as follows: A large-scale MIMO channel model with Rayleigh fading was built in a Python environment. The number of receive antennas M and transmit antennas N were both set to 32. A bit stream was generated by calling a random function. The bit stream was modulated into a symbol sequence using 4QAM modulation and additive channel noise was introduced. The corresponding transmit signal x, channel matrix H, receive signal y, and noise power were generated in a 10dB signal-to-noise ratio environment and used as the training dataset. Relevant data were generated in signal-to-noise ratio environments of 0~9dB and 15~20dB and used as the test dataset. Step 2: Construct the GROAMP-Net intelligent signal detection network model, which is based on the improved OAMP iterative method to recover the original transmitted signal; Step 3: Input the training dataset generated in Step 1 into the GROAMP-Net intelligent signal detection network model constructed in Step 2 for offline training, observe the loss function curve, and ensure that the model converges well; use the test dataset to test the converged model, and after passing the test, the trained GROAMP-Net intelligent signal detection network model is obtained. Step 4: Use the trained GROAMP-Net intelligent signal detection network model to perform large-scale MIMO signal detection; The GROAMP-Net intelligent signal detection network model is a model-driven deep learning network containing ten identical cascaded layers. Each cascaded layer performs the same linear and nonlinear estimation steps, and the entire network structure is equivalent to the optimized OAMP iterative process. The input of the first layer is the original received signal, the channel matrix, and the noise power. The input of each subsequent layer is the original received signal, the channel matrix, the noise power, and the estimated transmitted signal output from the previous layer. Each layer of the GROAMP-Net intelligent signal detection network model introduces four trainable parameters. Two of these parameters are used to adjust the step size of the linear and nonlinear estimation steps in the OAMP iterative method, respectively. The other two parameters serve as residual coefficients to construct the residual structure. The residual structure is used to correct the results of the linear and nonlinear estimation, thereby reducing the impact of the error caused by iterative approximation on the overall network performance. In addition, each layer of the GROAMP-Net intelligent signal detection network model is equipped with a gating unit to control the trainable parameters that adjust the step size of the two steps of linear estimation and nonlinear estimation, so as to prevent them from taking unreasonable values ​​during the training process. The gating unit consists of two types of control gates, whose system functions are the sigmoid function and the Leaky-ReLU function, respectively. The gating unit controls the two trainable parameters by passing each trainable parameter through the two control gates and then multiplying them to obtain a new value, which is then used for subsequent processing. The GROAMP-Net network model introduces a residual structure to handle the multiple iterations required during the iterative process. and On the one hand, it performs error correction; on the other hand, it sets up a gating unit to control the trainable parameters. and Optimize it to keep its value within a reasonable range; The residual structure is achieved by introducing two trainable parameters with an initial value of 1. and accomplish, and Used for processing respectively and The specific processing method is shown in the following formula. (1) (2) (3) (4) in, Let y be the output of the (t-1)th concatenated layer, y be the received signal, and H be the channel matrix. This is the intermediate value obtained from the linear estimation in the t-th cascaded layer. Let be the error variance of the nonlinear estimation in the t-th cascaded layer. and These are intermediate values ​​in the linear and nonlinear estimation processes under the residual structure, respectively. This refers to the nonlinear estimation process of the t-th layer of the network model. ,here It is a linear MMSE matrix, and N is the number of transmit antennas.

2. The method for detecting large-scale MIMO signals based on deep learning according to claim 1, characterized in that, In step 3, the offline training optimization method is Adam, and the loss function selected is the sum of the mean square error of the output estimate and the true value of the transmitted signal.

3. The method for detecting large-scale MIMO signals based on deep learning according to claim 1, characterized in that, The specific method for step 4 is as follows: The trained GROAMP-Net model is deployed at the receiver of a large-scale MIMO communication system. The received signal, channel matrix, and noise power are used as inputs. After processing by the trained model, the output value is the estimated value of the recovered transmitted signal.

Images