A deep learning based joint device activity detection and channel estimation method

Abstract

The application relates to a deep learning-based joint device activity detection and channel estimation method, which comprises the following steps: introducing a blind source separation module composed of an SVD module and a CICA module in combination, so that a pilot signal matrix Y can be separated from a pilot matrix S through the blind source separation module act , and then inputting the pilot signal matrix Y and the pilot matrix S act into a neural network module for data feature extraction and iterative calculation, thereby improving the prediction accuracy of the neural network model and generating more accurate prediction data. On the basis of improving the accuracy of joint activity detection and channel estimation, the transmission overhead is reduced, and the compromise between pilot overhead and accuracy is realized.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a wireless communication technology based on deep learning. Background Technology

[0002] The "Internet of Everything" has become a crucial development direction for the communications industry. In the future, most terminal devices will connect to networks, gradually converging into a massive, highly intelligent global Internet of Things. In traditional mobile communication systems, random access is a necessary process for establishing a wireless link between a terminal and user equipment. Only after random access is completed can normal data transmission occur between the terminal and the user equipment network. Unlicensed random access, as a new technology, allows users to access the network randomly without prior authorization, thereby improving user capacity and access efficiency. The breakthrough of this method lies in its improvement of the traditional licensed random access mechanism, making it possible to increase the number of users.

[0003] However, a key problem remains to be solved in unlicensed random access methods: device activity detection and channel estimation (JADCE). When multiple users access the network simultaneously, the access point typically needs to perform user activity detection to identify active users and channel estimation to obtain the channel state information of these active users. By estimating channel state information, such as time-domain or frequency-domain response, the received data can be corrected and recovered to improve the accuracy of the received signal, thereby allowing users to access and communicate effectively.

[0004] Researchers have proposed solutions to the JADCE problem using deep learning (DL). For example, patent number CN 112910806 A discloses a joint channel estimation and user activation detection method based on deep neural networks. This method utilizes the data feature extraction capabilities of neural network modules to detect pilot information and user-transmitted data signals, thereby achieving active user identification and channel estimation scheme design. Pilot signals, as special signals used for channel estimation and equalization in communication systems, are typically inserted into the data stream transmitted by user equipment so that terminals can use these known pilot signals to estimate the channel characteristics of the corresponding user equipment, such as fading and delay.

[0005] Increasing the length of pilot data to provide richer information is a common method to improve the accuracy of joint activity detection and channel estimation. However, increasing the length of pilot data often requires more valuable spectrum resources and transmission time, negatively impacting transmission overhead. Therefore, there is an urgent need for a solution that can reduce transmission overhead, thereby achieving a trade-off between pilot overhead and the accuracy of joint activity detection and channel estimation. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based joint device activity detection and channel estimation method that can reduce transmission overhead while improving the accuracy of device activity detection and channel estimation, thus achieving a trade-off between pilot overhead and accuracy.

[0007] This invention is achieved through the following technical solution:

[0008] A deep learning-based joint device activity detection and channel estimation method includes the following steps:

[0009] S1, Blind Source Separation

[0010] The preamble signal matrix Y received by the terminal is transmitted into the blind source separation module, which includes an SVD module and a CICA module. The SVD module decomposes the preamble signal matrix Y to obtain an estimate of its rank. The CICA module performs signal separation on the preamble signal matrix Y based on the estimate, separating out the pilot matrix S. act And output it to the neural network module; the pilot matrix S act The set of pilot data emitted by the active user equipment;

[0011] S2, Data Prediction

[0012] The neural network module receives the preamble signal matrix Y and the pilot matrix S. act It utilizes data feature extraction and iterative generation to output predicted data. The predicted data Includes user equipment activity data and channel status data.

[0013] The SVD module, short for Singular Value Decomposition, is a linear algebraic decomposition method widely used in data analysis, signal processing, and image compression. The SVD module decomposes a matrix into the product of three matrices and uses singular values ​​to estimate the rank of the matrix. Using the existing SVD module, the user first decomposes the preamble signal matrix Y into a noise signal subspace and a pure noise space, thereby estimating the rank of the preamble signal matrix Y. The rank of the preamble signal matrix Y is the number of active user devices.

[0014] The CICA module, short for Complex Independent Component Analysis, is a method for blind source separation (BSS). In CICA, it is assumed that the observed signal is a linear mixture of multiple independent source signals, and the goal of CICA is to separate the original independent source signals from these mixed signals.

[0015] First, the rank of the preamble signal matrix Y is estimated using the SVD method. Then, the estimated rank and the preamble signal matrix Y are input into the CICA module. The user can then separate the pilot matrix S based on the CICA module in the prior art. act .

[0016] Deep learning models are machine learning models based on multi-layered neural network structures. They can model and predict complex data patterns through multi-level feature extraction and representation learning. Examples include Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). In the fields of joint device activity detection and channel estimation, deep learning models can be used to learn complex activity patterns and features from signals from multiple users, achieving efficient detection of multi-user activities. They can also learn channel characteristics and data patterns from pilot signals in the data, achieving more accurate channel estimation. Especially in complex multipath propagation environments, deep learning models can learn nonlinear channel characteristics, thereby improving the accuracy of channel estimation.

[0017] Compared to directly inputting the preamble signal matrix Y into the deep learning module for data prediction, this invention first introduces a blind source separation module composed of the SVD module and the CICA module. This module can first separate the pilot matrix S from the preamble signal matrix Y. act Then, the preamble signal matrix Y and the pilot matrix S act The data is input into the neural network module for feature extraction and iterative calculation, which improves the prediction accuracy of the neural network model and enables the generation of more accurate prediction data. The predicted data The system includes user equipment activity data and channel state data. The user equipment activity data contains information about whether a user is active or inactive, which can be used to distinguish between active and inactive users. The channel state data includes the time-domain or frequency-domain response of the channel. The purpose of this data is to estimate the time-domain or frequency-domain response of the channel to correct and recover received data, thereby improving the accuracy of the received signal.

[0018] This paper introduces an innovative blind source separation module that effectively separates the pilot matrix, enabling a one-to-one correspondence between pilot signals and active user equipment (UFOs). Using this technique, the separated pilot matrix, which corresponds one-to-one with active UFOs, is input into a deep learning module, significantly improving the accuracy of the deep learning model. Compared to traditional methods that increase pilot data length, this method effectively reduces transmission overhead while improving the accuracy of joint activity detection and channel estimation. Notably, this method not only improves accuracy but also achieves a trade-off between pilot overhead and accuracy. In wireless communication, pilot overhead is a critical performance indicator that directly affects the system's transmission efficiency and reliability. By combining blind source separation and deep learning, we successfully improved the system's accuracy and transmission efficiency without increasing pilot overhead. In conclusion, the proposed method combining blind source separation and deep learning improves the accuracy of joint activity detection and channel estimation in wireless communication while reducing pilot overhead, opening up new possibilities for the development of wireless communication technology.

[0019] As a preferred embodiment of the present invention, the neural network module is a UGAN network model, and the training method of the UGAN network model is as follows:

[0020] SA, generating training data

[0021] Based on MATLAB, a simulated preamble signal matrix Y' is generated. The simulated preamble signal matrix Y' contains a known simulated user data signal matrix x' and other simulated interference signal data. After performing the S1 and blind source separation on the simulated preamble signal matrix Y', a training dataset is generated. The training dataset contains a simulated active signal matrix Yact' and a simulated pilot matrix Sact'.

[0022] SB, training to generate prediction data

[0023] The training dataset is fed into the UGAN network model for training, and the predicted data is output.

[0024] In this case, we chose the UGAN network model as the neural network module. The UGAN network model is a mature deep learning model in the existing technology, possessing powerful feature learning and classification capabilities, and is suitable for signal processing and communication fields. During the training process of the UGAN network model, Generative Adversarial Networks (GANs) are a well-known training method. Through competition and cooperation between the generator and discriminator, the generated fake data becomes increasingly closer to the real data. In our case, during the training process of the UGAN network model, we continuously optimize the network parameters to make the predicted data increasingly closer to the real data.

[0025] Before training the UGAN network model, we need to generate the simulated preamble matrix Y' using the Matlab platform. This matrix contains the known simulated user data signal matrix x' and other interference signal data, such as noise and time delay. Using the simulated preamble matrix Y', we can perform S1 and blind source separation, extracting the simulated pilot matrix Sact' from Y', ensuring a one-to-one correspondence between active user equipment and the pilot matrix. The simulated preamble matrix Y' and the simulated pilot matrix Sact' are then used as the training dataset and fed into the UGAN network model for training, thereby generating the predicted data.

[0026] Through the above training process, we have successfully trained a UGAN network model with high accuracy and stability. This model can generate the predicted data. This allows for a more accurate estimation of the activity of the user equipment and the corresponding channel quality.

[0027] As a preferred embodiment of the present invention, the expression for the analog preamble signal matrix Y' is specifically as follows:

[0028] Y' = SX' + Z,

[0029] in, This represents the set of pilot matrices emitted by N devices, where L represents the pilot length, and N and M represent the number of user equipment and the number of antennas, respectively, which can be customized by the trainer. The noise matrix is ​​represented by a standard Gaussian normal distribution.

[0030] In communication systems, noise is an unavoidable factor that interferes with and affects signal transmission. The noise matrix Z represents an L×M matrix composed of randomly generated noise, whose elements follow a standard Gaussian normal distribution. Here, L represents the pilot length, and M represents the number of antennas. The elements in this matrix are randomly generated and follow a standard Gaussian normal distribution. Analyzing and processing the noise matrix Z can improve the precision and accuracy of signal processing and data analysis.

[0031] S represents the set of pilot matrices emitted by N devices. The pilot length L is a key parameter that can be defined by the user to ensure the stability and accuracy of signal transmission. In practical applications, we can optimize the accuracy and stability of signal transmission by adjusting the value of the pilot length L.

[0032] As a preferred embodiment of the present invention, the expression for the analog user data signal matrix x' is:

[0033] X'=AH

[0034]

[0035]

[0036] Where A represents the user equipment state matrix. It is the channel matrix of AWGM, where N and M represent the number of user equipment and the number of antennas, respectively, and can be customized by the trainer;

[0037] In a communication system, the analog user data signal matrix X' = AH describes the data signals transmitted by the user equipment and the state of the channel. The user equipment state matrix A not only describes the activity level of the user equipment but also reflects its communication status. The channel matrix H is used to evaluate channel quality, reflecting changes and losses in signal transmission.

[0038] In our case, the channel matrix H was chosen as an AWGN channel. AWGN channels, also known as additive white Gaussian noise channels, are a common noise model in communication systems. This noise can affect signal transmission; therefore, when constructing the channel matrix H, we used independently sampled random numbers from a Gaussian distribution with zero mean and N0 / 2 variance to construct an N×M matrix.

[0039] As a preferred embodiment of the present invention, the performance evaluation method of the UGAN network model is as follows: [The method involves] evaluating the predicted data... Normalized error analysis (NMSE) is performed on the simulated user data signal matrix x' to obtain the NMSE value. The NMSE value is used to evaluate the prediction data generated by the UGAN network model. The accuracy.

[0040] NMSE is an abbreviation for "Normalized Mean Squared Error," a metric used to measure the difference between model predictions and true values. It is commonly used to evaluate the performance of generative models, including the UGAN network model. In the UGAN network model, NMSE assesses the degree of difference between the samples generated by the generator and the true samples. This is achieved by calculating the predicted data... The mean square error (MSE) between the simulated user data signal matrix x' and the target data matrix, after normalization, yields a metric for evaluating the generator's performance, namely the NMSE value. This value helps assess the performance of the UGAN network model in generating the predicted data. The accuracy.

[0041] As a preferred embodiment of the present invention, the UGAN network model includes a generator and a discriminator, and the activation functions of both the generator and the discriminator are ReLU activation functions.

[0042] The ReLU activation function, short for Rectified Linear Unit, is a commonly used activation function. It converts negative input values ​​to 0 while keeping positive values ​​unchanged, thus increasing the non-linear expressive power of the network. Using ReLU activation in the generator and discriminator can enhance the model's representational power, enabling it to better learn and simulate the distribution of data. By employing ReLU activation in the UGAN network model, we can construct a generative adversarial network with strong representational power and stability, providing powerful support for various machine learning tasks.

[0043] As a preferred embodiment of the present invention, the generator is a recurrent neural network consisting of multiple cascaded basic blocks, wherein the basic blocks include Unet blocks and projection blocks;

[0044] The generator described in this case is the recurrent neural network (RNN). The RNN is a relatively mature neural network in the prior art, composed of multiple cascaded basic blocks, each of which includes a Unet block and a projection block. This structure is designed based on a deep understanding of data distribution and the need for fine-grained processing. The Unet block enables fine-grained feature extraction, and the projection block normalizes the features, thereby better adapting to various data distributions.

[0045] In the recurrent neural network (RNN), the Unet block and the projection block are typically used for modeling and generating sequential data. The Unet block uses skip connections to concatenate the feature maps of the encoder and decoder, thus preserving more spatial information during decoding. This design allows the generator to better understand the contextual information of the input data and generate richer and more accurate output results.

[0046] The projection block normalizes the features, making their scale, range, and other attributes more stable, thus enabling it to better handle input data of different scales. Furthermore, the projection block adjusts the feature dimensions, allowing the generator's output to better match the dimensions of the target data.

[0047] By cascading multiple basic blocks, recurrent neural networks can achieve progressive abstraction and refined processing of input data. This network structure has significant advantages in processing complex data, effectively extracting key information from the data and generating high-quality output results, namely the predicted data.

[0048] As a preferred embodiment of the present invention, the UGAN network model employs a block training strategy to train the generator. Specifically, the block training strategy involves setting an initial learning rate η0, a learning rate η1, and a learning rate η2. First, the i-th basic block is trained with the initial learning rate η0 until convergence, and then all the basic blocks are trained with the learning rate η1 and the learning rate η2 until convergence. The initial learning rate η0 > the learning rate η1 > the learning rate η2.

[0049] In the training process of the UGAN network model, in order to achieve better performance and stability, we adopted an innovative block-based training strategy. The core idea of ​​this strategy is to decompose the entire generator into multiple basic blocks and train each basic block step by step.

[0050] First, we set an initial learning rate η0, which determines the step size for updating the model parameters. During training, we first train the i-th basic block using the initial learning rate η0 until that basic block converges. This approach allows the model to better adapt to the data distribution and ensures the generator has a good starting point for subsequent training.

[0051] Once the i-th basic block converges, we begin training the remaining basic blocks using two different learning rates, η1 and η2. These two learning rates are typically lower than the initial learning rate α, which helps the model fine-tune parameters more effectively later in training, further improving the generator's performance.

[0052] The advantage of this block-based training strategy is that it allows us to optimize each basic block while maintaining the overall performance of the model. Furthermore, by gradually reducing the learning rate, we can better control the update speed of the model parameters, thus avoiding overfitting and underfitting. In summary, by adopting a block-based training strategy, we can train the UGAN network model more effectively, improving the performance and stability of its generator.

[0053] In summary, the present invention has the following beneficial effects:

[0054] 1. A blind source separation module was introduced, which can effectively separate the pilot matrix and achieve a one-to-one correspondence between the pilot matrix and active user devices. Using this technique, the separated pilot matrix was input into the deep learning module, thereby significantly improving the accuracy of the deep learning model.

[0055] 2. By calculating the predicted data The mean square error (MSE) between the simulated user data signal matrix x' and the target data matrix, after normalization, yields a metric for evaluating the generator's performance, namely the NMSE value. This value helps assess the performance of the UGAN network model in generating the predicted data. The accuracy.

[0056] 3. By cascading multiple basic blocks, recurrent neural networks can achieve progressive abstraction and refined processing of input data. This network structure has significant advantages in processing complex data, effectively extracting key information from the data and generating high-quality output results, namely the predicted data.

[0057] 4. By adopting a block training strategy, we can train the UGAN network model more effectively and improve the performance and stability of its generator. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the UGAN network model training process of the present invention;

[0059] Figure 2 This is a schematic diagram of the Unet module structure of the present invention;

[0060] Figure 3 This is a comparison chart of the NMSE results for the two methods. Detailed Implementation

[0061] The present invention will be further described in detail below with reference to the accompanying drawings.

[0062] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

[0063] The training steps for the UGAN network model are as follows: Figure 1 As shown,

[0064] S1. Generate training dataset

[0065] S11. Each user generates an information bit stream s of length L\2. After encoding and modulation, a pilot signal S to be transmitted is obtained. The modulation method is Binary Phase Shift Keying (BPSK). The modulated signal matrix S reaches the terminal through the AWGN channel.

[0066] S12, such as Figure 1 As shown, the analog preamble signal matrix received by the terminal is:

[0067] Y′=SX′+Z (1)

[0068] in Represents the pilot matrix, Represents the AWGN channel matrix. Represents the diagonal matrix of device activity Let L represent the noise matrix, which follows a standard Gaussian normal distribution. L, N, and M represent the pilot length, number of users, and number of antennas, respectively.

[0069] S13. The reconfigurable analog preamble matrix based on the sparsity of the device activity matrix is ​​as follows:

[0070] Y′=[H] :,N [S] N,: +Z=H act S act +Z (2)

[0071] Where N represents the user set, S act 'H' represents the analog pilot matrix transmitted by the active device. act This represents the channel matrix corresponding to the active device.

[0072] Since the number of active user devices is H act S act The rank of the matrix. Therefore, by calculating H... act S act The rank of K′ a This allows us to calculate the number of active user devices. From Equation 2, we can deduce that the rank of the analog preamble matrix Y' is H. act S act The rank of K′ a Therefore, the number of active user devices can be obtained by calculating the rank of the analog preamble matrix Y'.

[0073] The rank of the analog preamble matrix Y' is determined using the Singular Value Decomposition (SVD) algorithm: the received analog preamble matrix Y' is decomposed.

[0074] Y′=U∑V H (3)

[0075] in Kmax = min(NL) > K a Kmax is Y act The singular value of ′.

[0076] Further, under high signal-to-noise ratio conditions, the received analog preamble signal matrix Y act It is divided into a noise signal subspace and a pure noise space.

[0077] The noise signal subspace is represented as follows:

[0078]

[0079] Pure noise space can be represented as:

[0080]

[0081] Therefore, the number of active devices can be estimated as follows:

[0082]

[0083] S14, based on The analog preamble signal matrix received by the device is shown in formula (2). Complex Independent Component Analysis (CICA) is used to analyze the Y signal. act Blind source separation yields the analog pilot matrix S corresponding to the active device. act ′.

[0084] S15, S separated based on CICA algorithm act Before feeding the data into the UGAN network, the complex data is first converted into real numbers. The specific operation is as follows:

[0085]

[0086] Get the dataset n represents the number of user devices.

[0087] S2, Generate Prediction Data

[0088] The UGAN network begins training on the dataset. The UGAN network model is composed of a generator network and a discriminator network cascaded together. The generator consists of multiple basic blocks, each containing two components: a Unet block and a projection block. Many Unet blocks and projection blocks form a recurrent neural network (RNN). The activation functions of both the generator network and the discriminator network are ReLU activation functions.

[0089] Unet module, such as Figure 2 As shown, the network consists of three layers: the first layer is a convolutional layer with ReLU activation function; the second layer contains V convolutional blocks with identical structures, each containing a convolutional layer (Conv layer), a batch normalization (BN) layer, and a ReLU activation layer; the third layer contains one convolutional layer.

[0090] During the feedforward phase of the input data, dimensionality reduction is performed first, followed by dimensionality augmentation to extract data features. In the dimensionality reduction phase, a 1D convolution with a kernel size of 3, a stride of 2, and padding of 1 is first used for sampling. Then, a 1D convolution with a kernel size of 3 and a stride of 2, along with a ReLU activation function, is used to extract features. Tensors are then superimposed with tensors of the same size, and this process is repeated twice. In the dimensionality augmentation phase, a 1D transformation with a kernel size of 2 and a stride of 2 is used. Next, a 1D convolution with a kernel size of 3, a stride of 1, and a ReLU activation function is used to extract features, and this process is repeated twice. In the discriminator network design, a 1D convolution with a kernel size of 3, a stride of 1, and a ReLU activation function is used to extract features. Then, a 1D convolution with a kernel size of 3 and a stride of 2 is used for downsampling. Finally, a 1D convolution with a kernel size of 3 is used to obtain the final output.

[0091] During training As input to the UGAN network generator, where (·) + This represents the pseudo-inverse of a matrix. The input to a basic block is... The projection block projects the output value onto... The null space, corresponding to the output of the Unet block is The loss function of the generator is:

[0092]

[0093] The loss function of the discriminator is:

[0094]

[0095] The generator is trained using a block-based training strategy with an initial learning rate of η0. The i-th basic block is trained with a learning rate of η0 until convergence, and then all blocks are trained with learning rates of η1 and η2 until convergence.

[0096] The output of the Unet block after K iterations:

[0097]

[0098] S3, Performance Evaluation

[0099] In step S2, generating prediction data, the UGAN network model generates prediction data. Then the predicted data With real data The NMSE was calculated by comparison, and the expression is as follows:

[0100]

[0101] in It is an estimated value. It is the actual value.

[0102] UGAN network model testing steps

[0103] The terminal receives a preamble signal matrix Y, but the user activity state and corresponding channel quality in the preamble signal matrix Y are unknown. A blind source separation module, composed of an SVD module and a CICA module, is introduced to separate the pilot matrix S from the preamble signal matrix Y received by the terminal. act Then, the preamble signal matrix Y and the pilot matrix S are... act By inputting data features into the trained UGAN network model for feature extraction and iterative calculation, the prediction accuracy of the neural network model is improved, enabling the generation of more accurate prediction data. The predicted data The system includes user equipment activity data and channel state data. The user equipment activity data contains information about whether the user is active or inactive, which can be used to distinguish between active and inactive users. The channel state data is used to estimate the channel quality corresponding to the user equipment, so that the user equipment can effectively access and communicate without prior authorization.

[0104] In this example, the above implementation uses MATLAB R2020a, PyCharm 2020, and PyTorch 3.8 (Python 3.8) to simulate the above steps. The simulation data settings are as follows: pilot length L = 32, 40, 48, 58, 64, 72, 80; number of devices N = 64; preset number of active devices Ka = 8; number of base station antennas M = 8; learning rate η0 = 5-10. -4 , eta1=0.2eta0, eta2=0.02eta0, standard deviation σ 2 =0.1. Calculate and plot a comparison of the loss error under different schemes as the pilot length increases, as shown in the figure. Figure 3 As shown.

[0105] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based joint device activity detection and channel estimation method, characterized in that, Specifically, the steps include the following: S1, Blind Source Separation The terminal receives a preamble signal matrix Y, which is then transmitted to a blind source separation module. This module includes an SVD module and a CICA module. The SVD module decomposes the preamble signal matrix Y to obtain an estimate of its rank. The CICA module then performs signal separation on the preamble signal matrix Y based on this estimate, separating out the pilot matrix. And output it to the neural network module; the pilot matrix A collection of pilot data emitted by active user equipment; S2, Data Prediction The neural network module receives the preamble signal matrix Y and the pilot matrix. It utilizes data feature extraction and iterative generation to output predicted data. The predicted data Includes user equipment activity data and channel status data; The neural network module is a UGAN network model, and the training method of the UGAN network model is as follows: SA, generating training data Based on MATLAB, a simulated preamble signal matrix Y' is generated. The simulated preamble signal matrix Y' contains a known simulated user data signal matrix x' and other simulated interference signal data. After performing the S1 and blind source separation on the simulated preamble signal matrix Y', a training dataset is generated. The training dataset contains the simulated preamble signal matrix Y' and the simulated pilot matrix Sact'. SB, training to generate prediction data The training dataset is fed into the UGAN network model for training, and the predicted data is output. ; The expression for the simulated preamble signal matrix Y' is as follows: Y'=SX'+Z, in, This represents the set of pilot matrices emitted by N devices. The matrix represents an L×N array, where L, N, and M represent the pilot length, the number of user equipment, and the number of antennas, respectively, and are defined by the trainer. The noise matrix is ​​represented by a standard Gaussian normal distribution. Represents an L×M matrix; The expression for the simulated user data signal matrix x' is: X'=AH, ， ， Where A represents the user equipment state matrix. H is the channel matrix of the AWGM. Represents an N×M matrix.

2. The method for joint device activity detection and channel estimation based on deep learning according to claim 1, characterized in that, The training method for the UGAN network model also includes training and generating prediction data located at the SB. Subsequent SC and performance evaluation: The predicted data... Normalized error analysis (NMSE) is performed on the simulated user data signal matrix x' to obtain the NMSE value. The NMSE value is used to evaluate the prediction data generated by the UGAN network model. The accuracy.

3. A method for joint device activity detection and channel estimation based on deep learning according to claim 1 or 2, characterized in that, The UGAN network model contains a generator and a discriminator, and the activation functions of both the generator and the discriminator are ReLU activation functions.

4. The method for joint device activity detection and channel estimation based on deep learning according to claim 3, characterized in that, The generator is a recurrent neural network consisting of multiple cascaded basic blocks, including Unet blocks and projection blocks.

5. The method for joint device activity detection and channel estimation based on deep learning according to claim 4, characterized in that, The UGAN network model employs a block-based training strategy to train the generator. Specifically, the block-based training strategy involves setting an initial learning rate. Learning rate and learning rate 2 First, using the initial learning rate... Train the i-th basic block until convergence, then use the learning rate... and the learning rate two Training all the aforementioned basic blocks converges; the initial learning rate >The learning rate is one >The learning rate two .

Images