An artificial intelligence-assisted MIMO detection method

Abstract

The present invention discloses an artificial intelligence-assisted MIMO detection method. First, the steps with large computational complexity in MIMO detection are determined, and the required input and output training data are collected; then, a multi-layer perceptron (MLP) model in deep learning is used to learn and map the steps to be optimized to obtain trained model parameters; then, the computational process with high complexity in the MIMO detection algorithm is converted into a neural network forward propagation process, that is, several matrix multiplications and matrix addition operations, so as to achieve the purpose of reducing the complexity of the detection algorithm. The effectiveness of the proposed detection algorithm is verified in different scenarios, and the detection performance can approach the effect of the original MIMO detection algorithm, and compared with the original MIMO detection algorithm, the computational complexity can be greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile communications, and in particular relates to an artificial intelligence-assisted MIMO detection method. Background Art

[0002] Multiple-Input Multiple-Output (MIMO) technology is the main technical approach to improve the spectrum efficiency of wireless communications. The main advantage of the MIMO system is that it can use spatial multiplexing technology without adding additional bandwidth and power, greatly improving the spectrum efficiency of the communication system. Therefore, modern communication systems often use MIMO and high-order quadrature amplitude modulation (QAM) to maximize spectrum efficiency.

[0003] However, the widespread deployment of high-dimensional MIMO systems is hindered by some limitations. In practical scenarios, the communication system is equipped with a large number of radio frequency units. When performing signal processing, the complexity of the transmitter and receiver of the receiver will be very high, and high energy consumption will be generated. In addition, the transmitted signal is sent from multiple antennas at the transmitter end of the MIMO system, and is received by multiple antennas at the receiver end. It is the result of the superposition of multiple signals, which makes it relatively difficult to detect the signal at the receiver end. As the number of antennas and the modulation order increase, the design of low-complexity and high-efficiency MIMO receivers faces huge technical challenges. For example, symbol detection can no longer rely on maximum likelihood detection or spherical decoding, because their complexity grows exponentially with the number of transmitting antennas and modulation order. This is the main bottleneck for the implementation of high-dimensional MIMO. Therefore, the research on high-performance MIMO detection technology is of great significance.

[0004] For MIMO communication systems, under the premise that the receiver obtains channel state information and the transmitted symbols obey uniform distribution, the maximum likelihood detector (ML) needs to traverse all possible vectors, as shown in the following formula, where u represents the vector taken from the n-dimensional symbol set The transmission symbol vector, H is the m×n MIMO complex channel matrix, and y is the channel reception symbol vector:

[0005]

[0006] The maximum likelihood detector is an NP-hard problem, which is too complex to be applied in practice. Therefore, how to simplify the signal detection algorithm and approach the performance of maximum likelihood detection has become a hot issue in the communications industry.

[0007] The Expectation Propagation (EP) algorithm is an algorithm with low complexity and high performance. The EP algorithm is an algorithm in Bayesian machine learning that can effectively process continuous distributions through moment matching. Therefore, it has the ability to process more complex and general approximate functions.

[0008] In the EP detection algorithm, a Gaussian approximation of the posterior distribution of the transmitted symbol vector is first constructed, i.e., q EP (u)≈p(u|y). The EP algorithm iteratively constructs q that matches the first two moments of each MIMO dimension. EP (u). The reason for this is that for a larger number of transmitting antennas n, it will become difficult to calculate directly from the posterior probability p(u|y). The final effect of EP detection is also very good. Even for very large MIMO orders, such as 250 transmitting antennas, only 7 to 8 iterations are needed to approximate the original distribution, and the complexity of each iteration depends on the inversion of the n×n matrix. Simulation results show that the number of iterations required for the EP algorithm is independent of the scale of the system and the modulation order, and the complexity of the algorithm is maintained at O(n 3 ). In addition, EP detection is a soft-output algorithm that provides an a posteriori probability estimate for each received symbol, which can naturally be used for further channel decoding.

[0009] With the improvement of science and technology, new technologies such as artificial intelligence have emerged in people's vision like mushrooms after rain. Integrating artificial intelligence technology into communication will become a mainstream direction of the development of the 6th Generation Mobile Communication Technology (6G). In addition, with the improvement of hardware level and computer computing power, machine learning technology has been widely used.

[0010] Machine learning is a type of artificial intelligence that requires a large amount of data to train an ideal model that meets the needs. Neural networks are an algorithm in machine learning and belong to supervised learning. Supervised learning means that the output of the training data is real and certain, and can accurately give the training error to guide the model to train and learn. Neural networks are composed of a large number of interconnected nodes. During the iterative data transmission process, the parameter values ​​of the nodes will be gradually optimized to continuously reduce the loss function, so as to achieve the purpose of accurately mapping the input data to the output data.

[0011] Deep learning is a branch of machine learning technology. It has received great attention in recent years and has been studied in depth by numerous experts and scholars. Deep learning can be understood as a vertical extension of neural network algorithms. Based on the basic neural network, it conducts detailed research on the number of network layers and network reasoning and operation processes to enable the network model to have a more powerful learning ability and enable the network to learn extremely complex operation processes after a large number of iterative training. In recent years, deep learning has been widely used in various fields, such as natural language processing, computer vision, and recommendation algorithms.

[0012] In the past, most mainstream ideas relied on the powerful learning ability of neural networks to improve the performance of algorithms. However, the essence of network reasoning, that is, the prediction of new data, is a number of matrix multiplication operations. Therefore, the existing problem is how to reduce the complexity of the MIMO detection algorithm while maintaining the original detection performance as much as possible and avoiding excessive performance loss. Summary of the invention

[0013] The purpose of the present invention is to provide an artificial intelligence-assisted MIMO detection method to solve the technical problem of reducing the complexity of the MIMO detection algorithm while ensuring a certain MIMO detection performance.

[0014] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0015] An artificial intelligence-assisted MIMO detection method comprises the following steps:

[0016] Step 1: Identify the more complex steps in the MIMO detection process and collect the required input and output training data;

[0017] Step 2: Use the multi-layer perceptron MLP model in deep learning to learn and map the steps to be optimized to obtain the trained model parameters;

[0018] Step 3: Convert the computational process with high complexity in the MIMO detection algorithm into a neural network forward propagation process, i.e., several matrix multiplication and matrix addition operations, so as to reduce the complexity of the detection algorithm.

[0019] Furthermore, in step 1, the required input and output training data are collected, and the model structure is designed based on comprehensive consideration of computational complexity and learning ability, which specifically includes the following steps:

[0020] Step 1.1, before the MIMO detection begins, initialize several arrays to store the input and output data to be collected. During the MIMO detection process, save the input and output data that need to be mapped each time. Under different signal-to-noise ratio conditions, since the data distribution of different signal-to-noise ratios is quite different, in order to obtain better training results, the signal-to-noise ratio is divided into different intervals, and different signal-to-noise ratio intervals are learned separately using different neural network models. When collecting data, it is also collected separately according to the intervals divided by the signal-to-noise ratio.

[0021] Step 1.2, comprehensively consider the computational complexity and learning ability of the neural network used in the detection process, and determine the structure of the model, including the number of hidden layers and the number of nodes in each hidden layer. The number of nodes in the input layer corresponds to the number of types of input data, and the number of nodes in the output layer corresponds to the number of types of output data.

[0022] Furthermore, the multi-layer perceptron MLP model in deep learning is used to learn and map the steps to be optimized to obtain the trained model parameters, which specifically includes the following steps:

[0023] Step 2.1, import the collected training data, divide the input and output data as well as the training and test data, and make sure that the divided test set accounts for a certain proportion of the total data set to ensure the generalization ability of the model;

[0024] Step 2.2: Use the z-score method to standardize the data to alleviate the impact of data dimension. The z-score standardization will standardize the data to a data distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula:

[0025]

[0026] μ and σ are the mean and standard deviation of the input data of the training set, respectively, and the input data of the training set and the test set are standardized.

[0027] Step 2.3, then perform data format processing; after the standardization is completed, convert the data set into a tensor type, and create a data loader to iteratively pass in batches of data for training;

[0028] Step 2.4: Build a neural network model and obtain the model after training.

[0029] Furthermore, the more complex calculation process in the MIMO detection algorithm is converted into a neural network forward propagation process, which specifically includes the following steps:

[0030] Step 3.1, respectively extract the weight matrix and bias parameters of the trained model, and save the node parameters of each model into a separate data file, so as to select the corresponding model according to the signal-to-noise ratio; and save the mean and standard deviation of the corresponding model of each signal-to-noise ratio in the standardized state into the data file;

[0031] Step 3.2: When iterating the signal-to-noise ratio, the corresponding model parameters and the corresponding standardized mean and standard deviation are imported according to the signal-to-noise ratio; then, the neural network forward propagation process is embedded in the MIMO detection algorithm.

[0032] Furthermore, in step 3.2, the neural network forward propagation process is embedded in the MIMO detection algorithm. The specific process is:

[0033] In the MIMO detection process, after obtaining the input variables, their format is integrated into the input data of the neural network; then, the input data is standardized; after that, the core forward propagation process of the neural network is reproduced to obtain the output result.

[0034] The MIMO detection method based on artificial intelligence assistance of the present invention has the following advantages:

[0035] The present invention adopts the multi-layer perceptron model in deep learning to learn and map some of the more complex steps in the MIMO detection algorithm, and converts the complex calculation process into a number of matrix multiplication operations of the neural network forward propagation, so as to achieve the purpose of reducing the complexity of the detection algorithm. After selecting different MIMO scenarios for experiments, the effectiveness and strong adaptability of this solution have been verified. On the one hand, it reduces the complexity of the MIMO detection algorithm, and on the other hand, it maintains the original detection performance as much as possible to avoid excessive performance loss, and has good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic block diagram of the process of the artificial intelligence-assisted MIMO detection method provided by the present invention.

[0037] Figure 2 A schematic block diagram of the MIMO system simulation process provided by the present invention.

[0038] Figure 3 This is the simulation result diagram of EP detection hard decision symbol error rate SER.

[0039] Figure 4 This is the simulation result diagram of EP detection soft decision bit error rate BER.

[0040] Figure 5 A schematic block diagram of the MLP module training process provided by the present invention.

[0041] Figure 6 A schematic block diagram of the MLP-assisted EP detection algorithm process provided by the present invention.

[0042] Figure 7 This is a comparison chart of the hard decision symbol error rate (SER) of the detection results of the present invention and the original EP detection algorithm after four iterations in an 8-input 8-output MIMO channel scenario.

[0043] Figure 8 This is a comparison chart of the hard decision symbol error rate (SER) of the detection results of the present invention and the original EP detection algorithm after four iterations in a fading channel scenario with 4 inputs and 8 outputs. DETAILED DESCRIPTION

[0044] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of an artificial intelligence-assisted MIMO detection method of the present invention in conjunction with the accompanying drawings.

[0045] Combination Figure 1 A MIMO detection method based on artificial intelligence assistance of the present invention is specifically described, comprising the following steps:

[0046] The expectation propagation (EP) detection algorithm with outstanding performance and low complexity is simulated, the time-consuming steps in the detection process are analyzed, and the optimization ideas of the neural network in the EP detection algorithm are clarified; then the required input and output training data are collected during the simulation iteration process, and the model structure is designed based on comprehensive consideration of computational complexity and learning ability; then the multi-layer perceptron (MLP) model in deep learning is used to learn and map the steps that need to be optimized to obtain the trained model parameters; finally, the more complex calculation process in the EP detection algorithm is converted into a neural network forward propagation process, that is, several matrix multiplication and matrix addition operations, in order to achieve the purpose of reducing the complexity of the detection algorithm. The specific process is as follows:

[0047] First, a relatively complete MIMO detection scenario is simulated. The process is as follows: Figure 2 As shown:

[0048] (1) Define some important parameters in the simulation scenario, such as the number of transmitting antennas, the number of receiving antennas, the modulation method, the number of transmitted symbols, the target channel coding rate, the transmission block length, the signal-to-noise ratio, etc. Some of these parameters, such as the modulation method, the number of transmitting and receiving antennas, etc., can be adjusted as needed to verify the applicability of the detection algorithm to different scenarios.

[0049] (2) Generate constellation points according to the modulation method, calculate the power of each constellation point symbol, and further calculate the average symbol power of the constellation point.

[0050] (3) Iterate different bit signal-to-noise ratios. Iterate each packet under a certain signal-to-noise ratio. In each packet, a binary information bit sequence is randomly generated, and the length of the sequence is the transmission block length.

[0051] (4) Channel coding is performed on the information bit sequence, and the 5G ToolBox provided by MATLAB is directly used to add cyclic redundancy check code, code block segmentation, error correction coding and rate matching in sequence. Among them, error correction coding is more critical, and its purpose is to add some redundant codewords to enable the receiving end to automatically correct errors occurring in data transmission. LDPC coding is used in the present invention, which has a simple mathematical description and is suitable for parallel processing and hardware implementation, and has good performance and flexibility.

[0052] (5) The channel coded sequence is modulated to obtain the transmitted signal u, and the MIMO channel matrix H is randomly generated according to the number of transmitting and receiving antennas. The dimension is RxAntNum×TxAntNum. Each element in the channel matrix is ​​normally distributed and the modulus is 1. Then the transmitted signal passes through the channel and Gaussian noise ω is added to obtain the received signal y. This process is y=Hu+ω.

[0053] (6) In order to better understand the reasoning process of the EP algorithm, the complex variables of transmitted symbols, received symbols, channel matrix, and noise variance are expressed as real variables, that is, the complex-valued MIMO system is reconstructed into a real-valued MIMO system. The complex-valued MIMO system model y = Hu + ω can be equivalent to the real part and the imaginary part sequence Two sequences, sending signals Receiving Signals Gaussian noise Channel Matrix At this time, the real-valued channel model and the posterior probability distribution are as follows:

[0054]

[0055]

[0056] represents the posterior probability of sending a real-valued symbol vector, Indicates the mean The covariance matrix is The normal distribution of The meaning is if The function value is 1, otherwise it is 0. is the variance of the complex-valued noise. The variances of the real and imaginary parts of the complex-valued noise are is half the variance of the complex channel noise, is the symbol set of the real and imaginary parts of the symmetric M-QAM signal, and the symbol average energy It is half of the average energy of the complex signal symbol.

[0057] (7) Perform EP detection algorithm.

[0058] (8) Directly based on the results of EP detection, hard-determine the number of erroneous symbols. At the same time, EP detection will return the symbol probability of the detection result, so the expectation can be calculated based on the output posterior probability estimate, and soft demodulation can be performed. The specific steps of demodulation are: rate matching, LDPC decoding, decoding block segmentation, cyclic redundancy check, which is the inverse process of channel coding; then soft-determine the number of erroneous bits.

[0059] (9) Based on the detection results, the simulation results of the symbol error rate (SER) of the hard decision and the bit error rate (BER) of the soft decision are obtained. The simulation results are shown in Figure 3 and Figure 4 shown.

[0060] Then, the more time-consuming steps in the detection process are analyzed to clarify the optimization ideas of neural networks in the EP detection algorithm.

[0061] In the EP detection algorithm, the most complex and time-consuming step is to calculate the Gaussian approximation q EP The mean and variance of (u) are calculated because the inversion operation in this process consumes a lot of time. Secondly, during the EP inner loop iteration, the probability density function q of the current detection signal is calculated. (l)\i (u i ) and variance and updates The process is also relatively time-consuming.

[0062] Based on the above analysis of the more time-consuming steps in the detection process, we try to calculate the Gaussian approximation q EP (u) and calculate the probability density function q of the current detection signal (l)\i (u i ) and variance and updates These three steps are simplified by considering the mean of the original posterior probability distribution p(u|y) and variance And the probability density function q of the detection signal (l)\i (u i ) and variance These four variables are used as the input of the neural network, and the probability density function q of the detection signal after iteration is (l+1)\i (ui ) and variance As the output of the neural network, the relatively complex steps in the EP detection process can be skipped and the mean value of the output result of the neural network can be directly used. and variance Generate a probability density function of normal distribution and get the detection result.

[0063] Since the essence of a neural network is that input data is output after being calculated at the network nodes, the intermediate calculation process can be equivalent to a matrix multiplication operation. Therefore, the calculation process can actually be equivalent to several matrix multiplication operations, and the calculation complexity can be greatly reduced compared to the complex formula calculation performed directly. Another advantage of a neural network is that no matter how large the data scale of the input data is, the calculation process is still a few matrix operation steps, which can be effectively parallelized. In the case of batch data, the calculation efficiency can be greatly improved.

[0064] After that, the required input and output training data are collected during the simulation iteration process, and the model structure is designed based on comprehensive consideration of computational complexity and learning ability. The process is as follows:

[0065] (1) Before the simulation starts, six arrays are initialized to store the six types of variable data that need to be collected. During the simulation, the data that needs to be mapped in each EP iteration is saved. Under different signal-to-noise ratios, the distribution of data varies greatly. For example, when the signal-to-noise ratio is 10, Roughly 10 -3 order of magnitude, and when the signal-to-noise ratio is 30, Roughly 10 -5 Since the data distribution of different signal-to-noise ratios is quite different, in order to obtain better training results, the signal-to-noise ratio is divided into different intervals, and different neural network models are used for learning in different signal-to-noise ratio intervals. Therefore, a total of 5 models are trained, and the bit signal-to-noise ratio Make the following divisions: Train the first model, Train the second model, Train the third model, Train the fourth model, Train the fifth model. Therefore, when collecting data, it is also divided into these 5 cases.

[0066] Finally, the amount of data corresponding to each signal-to-noise ratio interval is 307,200. Each piece of data has 6 features, the first 4 features are input, and the last 2 features are output, that is, the data scale is 307,200×6.

[0067] (2) The essence of a neural network model is matrix multiplication, and the matrix parameters are what the model needs to train. The number of hidden layers determines the number of matrix multiplications required for prediction, and to a large extent determines the algorithm complexity of a model prediction. The number of nodes in each layer of the neural network will affect the learning ability of the model. If the number of nodes is greater, the learning and fitting abilities of the model will be stronger, and the amount of data required will be greater, otherwise underfitting will occur. If the number of nodes is smaller, the learning ability of the model is also limited, and it is not easy to learn complex distributions, and it is easy to overfit.

[0068] After many attempts, considering the computational complexity and learning ability, the final model structure is 4 nodes in the input layer, corresponding to the four input variables and There are two hidden layers with 32 and 16 nodes respectively, and an output layer with 2 nodes, corresponding to the two output variables and

[0069] Next, the multi-layer perceptron (MLP) model in deep learning is used to learn and map the steps to be optimized to obtain the trained model parameters. The process is as follows: Figure 5 As shown:

[0070] (1) Import the collected training data and divide the input and output data as well as the training and test data. The divided test set accounts for 20% of the total data set to ensure the generalization ability of the model.

[0071] (2) Then, the z-score method is used to standardize the data to alleviate the impact of data dimension. Z-score standardization will standardize the data to a data distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula:

[0072]

[0073] μ and σ are the mean and standard deviation of the input data of the training set, respectively, and the input data of the training set and the test set are standardized.

[0074] (3) Then the data format is processed. After the standardization process is completed, the data format is numpy format, and the Pytorch process of implementing neural networks requires data in tensor format. Tensor type data is faster in neural networks and can create higher-dimensional matrices and vectors. Therefore, it is necessary to convert the data set to a tensor type, and also convert the data format to float32. The default double data type cannot be trained in the Pytorch framework. Next, create a data loader and iteratively pass in data in batches for training.

[0075] (4) Build a neural network model. In the present invention, three fully connected layers of 4×32, 32×16, and 16×2 are used to achieve data learning while ensuring the simplicity of reasoning; the activation function uses the ReLU() function to maximize the calculation speed and the speed of training convergence; the optimizer uses the Adam optimizer, and the initial learning rate is set to 0.001; the error loss function uses the L1 loss function, that is, the minimum absolute value deviation. The model is obtained after training.

[0076] Finally, the more complex calculation process in the EP detection algorithm is transformed into the neural network forward propagation process, as shown in the following figure. Figure 6 As shown:

[0077] (1) First, extract the weight matrix and bias parameters of the trained model. Since the extracted parameters are tensor-like data, in order to better reproduce, it is also necessary to convert the data type to numpy type and save the parameters as .mat files for easy calling in MATLAB. Each model saves a mat file to store the model parameters so that the corresponding model can be selected according to the signal-to-noise ratio.

[0078] In addition, the mean and standard deviation of each signal-to-noise ratio corresponding model during standardization need to be saved in a mat file. Before making a prediction, the data needs to be standardized with the mean and variance of the training data set before the next prediction can be made.

[0079] (2) First, when iterating the signal-to-noise ratio, the corresponding model parameters and the corresponding standardized mean and standard deviation are imported according to the signal-to-noise ratio. Subsequently, further modifications are made in the EP detection algorithm.

[0080] In the EP iterative inner loop, the mean of the posterior probability distribution p(u|y) is obtained and variance And the probability density function q of the previous round of detection signals (l)\i (u i ) and variance After these four variables are obtained, their formats are integrated into the input data of the neural network; subsequently, the input data is standardized.

[0081] After that, it is the core reproduction process of the neural network. The process of the neural network prediction in the present invention is as follows, where n is the number of transmitting antennas:

[0082]

[0083]

[0084]

[0085] Finally, the output results are the probability density function q of the detection signal in this round of iteration. (l)\i (u i ) And the probability density function q of the detection signal in this round of iteration (l)\i (u i )

[0086] The iterative process of the MLP-assisted EP algorithm only requires three matrix multiplication operations and matrix additions. From the above formula, the main computational complexity of this process is: O(2n×4×32)+O(2n×32×16)+O(2n×16×2). After simplification, the final computational complexity is O(n), while the computational complexity of the original EP detection algorithm is O(n 3 ), the complexity is greatly reduced.

[0087] like Figure 7 As shown, it is a comparison chart of the hard decision symbol error rate (SER) of the detection results of the invention and the original EP detection algorithm after 4 iterations in the MIMO channel scenario with 8 inputs and 8 outputs. Although the MLP-based EP detection algorithm is not able to completely match the performance of the original EP algorithm, it can approximate the EP algorithm to a large extent. Especially in the case of low signal-to-noise ratio (SNR≤16dB), it can be relatively close to the performance of the EP algorithm; and in the case of high signal-to-noise ratio (SNR>16dB), the accuracy requirements for the neural network fitting mapping data will be higher, and the performance difference is relatively large, but the error is also within an acceptable range, and the performance is still much higher than that of the traditional MMSE detection algorithm.

[0088] Since the training data uses the simulation data of 8-input 8-output MIMO channel, in order to verify the universality of the model, a fading channel different from the training data is used, such as Figure 8 As shown, it is a comparison chart of the hard decision symbol error rate SER of the detection results of the present invention and the original EP detection algorithm after 4 iterations in a fading channel scenario with 4 inputs and 8 outputs. The performance of the present invention can still approach that of the original EP algorithm in most cases.

[0089] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. An artificial intelligence-assisted MIMO detection method, characterized in that: The following steps are involved: Step 1: Determine the steps with high complexity in the MIMO detection process, including the steps with complexity of O(n 3 ), collect the required input and output training data, and design the model structure based on comprehensive consideration of computational complexity and learning ability; The specific steps include: Step 1.1, before the MIMO detection begins, initialize several arrays to store the input and output data to be collected. During the MIMO detection process, save the input and output data that need to be mapped each time; divide the signal-to-noise ratio into different intervals, and use different neural network models to learn different signal-to-noise ratio intervals separately; when collecting data, also collect separately according to the intervals divided by the signal-to-noise ratio; Step 1.2, comprehensively consider the computational complexity and learning ability of the neural network used in the detection process, determine the structure of the model, including the number of hidden layers and the number of nodes in each hidden layer, the number of nodes in the input layer corresponds to the number of types of input data, and the number of nodes in the output layer corresponds to the number of types of output data; Step 2: Using the multi-layer perceptron MLP model in deep learning, learn and map the high-complexity steps that need to be optimized in step 1 to obtain trained model parameters; specifically, the following steps are included: Step 2.1, import the collected training data, divide the input and output data as well as the training and test data, and the divided test set accounts for 20% of the total data set to ensure the generalization ability of the model; Step 2.2: Use the standard score z-score method to standardize the data to alleviate the impact of data dimension. The z-score standardization will standardize the data to a data distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula: z＝ x-μ σ μ is the mean of the input data of the training set, σ is the standard deviation of the input data of the training set, and z is the output result with a mean of 0 and a standard deviation of 1 after standardization. The input data of the training set and the test set are standardized; Step 2.3, then perform data format processing; after the standardization is completed, convert the data set into a tensor type, and create a data loader to iteratively pass in batches of data for training; Step 2.4: Build a neural network model and obtain the model after training; Step 3: Convert the computational process with high complexity in the MIMO detection algorithm into a neural network forward propagation process, i.e., several matrix multiplication and matrix addition operations, to reduce the complexity of the detection algorithm; specifically, the following steps are included: Step 3.1, respectively extract the weight matrix and bias parameters of the trained model, and save the node parameters of each model into a separate data file, so as to select the corresponding model according to the signal-to-noise ratio; and save the mean and standard deviation of the corresponding model of each signal-to-noise ratio in the standardized state into the data file; Step 3.2, when iterating the signal-to-noise ratio, import the corresponding model parameters and the corresponding standardized mean and standard deviation according to the signal-to-noise ratio; then, embed the neural network forward propagation process into the MIMO detection algorithm; In step 3.2, the forward propagation process of the neural network is embedded in the MIMO detection algorithm. The specific process is as follows: in the MIMO detection process, after obtaining the input variables, their format is integrated into the input data of the neural network; then, the input data is standardized; after that, the core forward propagation process of the neural network is reproduced to obtain the output result.

Images