A new channel image acquisition method and system

Abstract

The present invention relates to the technical field of the intersection and integration of artificial intelligence (AI) neural network technology and wireless communication systems, and discloses a novel method and system for acquiring channel images. Based on data-driven AI neural network channel estimation, the present invention acquires sufficient channel images for neural network training. This invention overcomes the limitation of previous channel image acquisition methods that depend on the type and geometry of deployed antennas, enabling AI neural networks to be effectively used for channel estimation in wireless communication systems.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross-integration of artificial intelligence neural network technology and wireless communication system, and specifically relates to a novel method for acquiring channel images. Background Art

[0002] In the field of image processing, image restoration is divided into two methods: image noise reduction and image reconstruction. In recent years, AI neural network technology has been used to achieve performance that surpasses traditional techniques for both noise reduction and reconstruction. In the field of wireless communications, large-scale antenna array technology, by equipping more antennas, can significantly improve data throughput and spectral efficiency, and is therefore widely used in various wireless communication systems. As more antennas are installed at the base station or terminal side, the equivalent wireless channel matrix becomes sparser, and the variation between adjacent elements decreases, giving the wireless channel the characteristics of an image. Inspired by this, researchers are working to apply AI to the traditional field of channel estimation.

[0003] AI-based channel estimation can be categorized into model-driven and data-driven approaches, depending on how the AI ​​neural network is trained. Model-driven approaches assume that random wireless channels follow a certain mathematical distribution. Some studies model wireless channels as random distributions following a sparse Gaussian mixture (GM) and use expectation maximization (EM) to estimate the GM parameters of the wireless channel. In contrast, data-driven approaches do not rely on prior expert knowledge of the wireless channel distribution. Instead, they use collected channel data to train a neural network to approximate the wireless channel distribution.

[0004] In practice, wireless channels vary greatly depending on the scenario, such as urban, rural, indoor shopping malls, and indoor factories, making them difficult to describe using a single mathematical distribution. Data-driven channel estimation techniques, on the other hand, often assume a 3D lens antenna array (MxN) at the base station. By rearranging the MxN channel matrix, a 2D channel map (MxNx1) can be obtained. This channel map acquisition method relies on antenna geometry, including shape and size, and therefore has significant limitations. To obtain a 64x64 channel map, a wireless communication system equipped with lens antennas requires 4096 antenna elements, while a linear array antenna requires 64 antenna elements at both the base station and the user end. This channel map acquisition method consumes expensive hardware resources, resulting in high implementation costs. Furthermore, it is often impractical in practice due to device size and power constraints.

[0005] For a multiple-input multiple-output (MIMO) wireless communication system equipped with multiple antenna units, assuming that the number of antenna units equipped on the base station side is M, the number of antenna units equipped on the user equipment side is N, and the number of transmission streams is S (S≤N), the uplink channel matrix can be expressed as H∈C MXS , where C represents the complex domain. In the training phase, the pre-designed reference signal x t ∈C SX1 , sent by the user end at time t and received by the base station side t Expressed as

[0006] y t =H t x t +n (1)

[0007] where n∈C MX1 Represents noise.

[0008] Since the mobile communication device at the user end is powered by a battery, its power consumption and computing power are relatively limited, so in practice N = S = 1. t It is known that the least squares (LS) channel estimation can be expressed as

[0009]

[0010] Among them H LS is the original channel signal H t The least squares estimate of the reference signal x t Generally, orthogonality is satisfied. Also often used to replace.

[0011] Since the least squares channel estimation is simple to implement, it has been widely used in the industry. However, its disadvantages are also obvious. Since it does not take the noise term into account, when the channel conditions are bad, the noise will be greatly amplified, resulting in a significant decrease in channel estimation performance. Therefore, the linear minimum mean square error (LMMSE) channel estimation is often used to improve the least squares LS channel estimation.

[0012]

[0013] where R HH =E{HH H} is the autocorrelation matrix of the channel signal, β is a constant determined by the constellation point of the transmitted signal (for example, for 16-QAM, β=17 / 9), and SNR represents the signal-to-noise ratio.

[0014] The LMMSE-based method can improve the channel estimation performance, but its implementation complexity is very high, especially the channel statistical information R HH The acquisition of is very difficult and inefficient in practical applications.

[0015] Data-driven AI channel estimation features offline training and online inference. This allows for complex and time-consuming offline training, while online inference allows for rapid response and eliminates the need for prior channel statistics, providing a new technical direction for channel estimation.

[0016] As mentioned earlier, model-driven AI channel estimation technology is not universally applicable. In practical applications, it is difficult to obtain the mathematical distribution of each wireless channel in advance. Data-driven AI channel estimation has strong generalization capabilities and great practical application prospects. The key to its successful deployment is to obtain sufficient channel images for training.

[0017] Through the above analysis, the problems and defects of the existing technology are as follows:

[0018] Traditional LS channel estimation has poor performance. LMMSE can significantly improve performance, but it is complex and expensive to run online. AI-assisted channel estimation can achieve both high performance and fast online operation. However, existing technologies rely on the type and geometry of deployed antennas to obtain channel images, which is highly limited. It also consumes expensive hardware resources and is costly to implement. In practice, this is constrained by limited device size and power consumption. Summary of the Invention

[0019] In response to the above-mentioned technical problems in existing channel image acquisition methods, the present invention proposes a new channel image acquisition method, which enables data-driven artificial intelligence channel estimation to be effectively deployed in wireless communication systems.

[0020] The present invention is implemented as follows: a novel channel image acquisition method comprises the following steps:

[0021] Step 1: Image collection. During the offline training phase, collect signals at time T (T≥1) to obtain the original channel signals of the three-dimensional tensor H = H1, H2, ..., H T ]∈C MXSXT and the received signal H after passing through the channel LS =H LS, , H LS, ,…,H LS, ]∈C MXSXT , respectively as the real signal used for neural network training and the input signal to be estimated, where M is the number of antennas of the receiving device, S is the number of transmitted data streams, and T is the time observation length;

[0022] Step 2: Dimensional reorganization, three-dimensional tensor channel signals H and H LS Perform dimension reorganization operation, fix the time dimension T, transform each MXS matrix into a vector of MSX1, and then arrange it according to the time dimension to obtain matrices H′ and H′ of size MSXT LS ;

[0023] Step 3: Sliding window selection, use a sliding window of size PXK with a step size p to select the real signal H' and the received signal H' LS Slide along the time dimension and select a channel image of size PXK, where P≤MS, K is the time dimension and K≤T;

[0024] Step 4: Perform real and imaginary part separation on the obtained PXK channel image to obtain a PXKX2 channel image.

[0025] Step 5: Neural network training: Get the PXKX2 channel image for neural network training, get the weight coefficient of each layer, and deploy the trained neural network to the device.

[0026] Step 6: Online speculation. In the online speculation stage, the system waits for K moments to obtain the initial received signals from the 1st to the Kth moment.

[0027] Step 7: The observation signal of the system at time K+1 is composed of the signal from the 2nd to the K+1th, and so on:

[0028] For the received signal, a sliding window of size LXQ is used with a step size q to select and generate an LXQ channel image;

[0029] Perform real and imaginary part separation on the obtained LXQ channel image to obtain the LXQX2 channel image;

[0030] The obtained LXQX2 channel image is input into the trained neural network model to obtain the corresponding true channel estimation value.

[0031] Furthermore, the time dimension is used to generate the channel tensors and channel images required by the neural network.

[0032] Furthermore, in the offline training stage, the original channel signal at time T (T≥1) and the signal after passing through the channel are used as the real signal and input signal for neural network training, respectively.

[0033] Furthermore, the input signal for training can be the signal y received directly after passing through the channel t, or it can be a signal that has been preliminarily processed. The preliminarily processed methods include but are not limited to: least squares (LS) and least linear mean square (LMMSE).

[0034] Furthermore, a sliding window in the time dimension acts on the real signal and the input signal respectively to select one or more channel images.

[0035] Furthermore, good channel images are selected to train the neural network and obtain the weight coefficients of each layer.

[0036] Furthermore, in the online inference phase, the system waits for a certain initial time K (K≥1) to obtain the initial received signal from the 1st to the Kth time. The K time here can be consistent with the sliding window time dimension K used in the previous offline training phase, or it can be inconsistent.

[0037] Furthermore, the received signal at the K+1th time is composed of the received signals from the 2nd to the K+1th time, and so on.

[0038] Furthermore, the online inference stage also uses a sliding window in the time dimension to act on the received signal to select one or more channel images.

[0039] Furthermore, the sliding window size and step size in the online inference phase may be the same as or different from those used in the offline training phase.

[0040] Furthermore, the initial received signals or channel images from time 1 to time K can be fed into the trained neural network to obtain an estimated value. The received signal or channel image obtained at time K+1 is fed into the trained neural network to obtain the channel estimate at time K+1, and so on.

[0041] Another object of the present invention is to provide a novel channel image acquisition system, comprising:

[0042] The offline training module is used to collect the signal at time T (T≥1) during the offline training phase and obtain the original channel signal H = [H1, H2, ..., H T ]∈C MXSXT and the received signal H after passing through the channel LS =[H LS,1 , H LS,2 ,…,H LS,T ]∈C MXSXT , respectively as the real signal used for neural network training and the input signal to be estimated, where M is the number of antennas of the receiving device, S is the number of transmitted data streams, and T is the time observation length;

[0043] Dimensional reorganization module for the three-dimensional tensor channel signals H and H LSPerform dimension reorganization operation, fix the time dimension T, transform each MXS matrix into a vector of MSX1, and then arrange it according to the time dimension to obtain matrices H′ and H′ of size MSXT LS ;

[0044] Sliding window selection module, used for a sliding window of size PXK with a step size p between the obtained real signal H′ and the received signal H′ LS Slide along the time dimension and select a channel image of size PXK, where P≤MS, K is the time dimension and K≤T;

[0045] A real and imaginary part separation operation module is used to perform a real and imaginary part separation operation on the obtained PXK channel image to obtain a PXKX2 channel image;

[0046] The neural network training module is used to obtain the PXKX2 channel image for neural network training, obtain the weight coefficient of each layer, and deploy the trained neural network to the device end;

[0047] The online inference module is used in the online inference stage. The system waits for K moments to obtain the initial received signals from the 1st to the Kth moments.

[0048] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs any one of the steps of the above-mentioned improved novel channel image acquisition method.

[0049] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform any one of the steps of the above-mentioned improved novel channel image acquisition method.

[0050] Another object of the present invention is to provide an information data processing terminal, which is used to implement the above-mentioned improved system for acquiring new channel images of urban scene remote sensing images based on the SwinUnet model.

[0051] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0052] First, in view of the technical problems existing in the above-mentioned prior art and the difficulty of solving these problems, this paper closely combines the technical solutions to be protected by the present invention and the results and data during the research and development process, and analyzes in detail and in depth how the technical solutions of the present invention solve the technical problems and some creative technical effects brought about by solving the problems. The specific description is as follows:

[0053] The present invention proposes a new channel image acquisition method based on data-driven artificial intelligence neural network channel estimation. The first priority is to obtain enough channel images for training the neural network. The present invention solves the previous limitation that channel image acquisition depends on the type of deployed antennas and their geometric dimensions, and proposes a new and effective channel image acquisition method, so that artificial intelligence neural networks can be effectively used for channel estimation in wireless communication systems in practice.

[0054] Second, considering the technical solution as a whole or from the perspective of the product, the technical effects and advantages of the technical solution to be protected by the present invention are described in detail as follows:

[0055] From a product perspective, the new channel image acquisition method proposed in the present invention is more universal and can enable AI-assisted channel estimation to be more widely used in wireless communication systems. In practice, not all wireless communication systems are equipped with large-scale antenna units on both the transmitting and receiving ends. Taking WiFi as an example, its latest standard currently only supports a maximum of 8 antennas on the base station side and a maximum of 4 antennas on the user side. The more common configuration is 2 or 4 antennas on the base station side and 1 antenna on the user side. According to the existing channel image acquisition method, the channel image obtained is of a dimension of 2X1 or 4X1. After dimensional reorganization, the maximum channel image that can be obtained is 2X2. The amount of information is insufficient and it is difficult to be effectively used for training and inference of artificial intelligence neural networks. The method proposed in the present invention, by introducing the time dimension and sliding window selection, can easily obtain a sufficiently effective channel image for neural network training, and the online inference stage can also be better performed.

[0056] Third, as auxiliary evidence for the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0057] The present invention proposes a novel channel image acquisition method, which fills the technical gap in the industry at home and abroad by introducing the time dimension and the idea of ​​sliding window selection.

[0058] In the field of wireless communications, there is a widespread technical prejudice against using artificial intelligence for physical layer algorithms such as channel estimation. Many people believe that it is not feasible. There are two fundamental reasons for this: one is the high complexity and high cost brought about by artificial intelligence assistance, and the other is the universality problem of artificial intelligence.

[0059] The first reason is that traditional thinking leads us to believe that effective AI-assisted channel estimation requires massive antennas at both the transmitter and receiver ends, which is prohibitively expensive. The second reason is that model-driven AI-assisted channel estimation often assumes that the channel follows a certain mathematical distribution, resulting in limited generalization in practice.

[0060] This invention, based on data-driven, artificial intelligence-assisted channel estimation, possesses excellent generalization capabilities. Furthermore, as previously analyzed, this invention overcomes the limitation of previous channel image acquisition methods, which relied on the type and geometry of deployed antennas. It proposes a novel and effective channel image acquisition method, enabling the effective use of artificial intelligence neural networks for channel estimation in wireless communication systems, effectively overcoming the technical biases in this field. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of a novel channel image acquisition method provided by an embodiment of the present invention;

[0062] Figure 2 This is a module diagram of a new channel image acquisition system provided by an embodiment of the present invention;

[0063] Figure 3 This is a channel estimation curve diagram for a user moving at a speed of 3 kilometers per hour in a 3GPP UMa environment provided by an embodiment of the present invention;

[0064] Figure 4 This is a channel estimation curve diagram provided by an embodiment of the present invention for a user moving at a speed of 60 kilometers per hour in a 3GPP UMa environment. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] In order to enable those skilled in the art to fully understand how to implement the present invention, this section provides an explanatory embodiment that expands on the technical solutions of the claims.

[0067] As mentioned above, the model-driven AI channel estimation technology is not very versatile. In practical applications, it is difficult to obtain the mathematical distribution of each wireless channel in advance. Data-driven AI channel estimation has strong generalization capabilities and has good practical application prospects. The key to its successful deployment is to obtain enough channel images for training. Figure 1 、 Figure 2 As shown, an embodiment of the present invention proposes a new channel image acquisition method, which enables data-driven artificial intelligence channel estimation to be effectively deployed in wireless communication systems. The specific implementation method is described as follows.

[0068] In the offline training phase, we collect the observation signals at time T, and the corresponding original channel signals are H1, H2, ..., H T, this original channel is used for offline training. Its acquisition method can be obtained by modeling the wireless channel according to the parameters of various types of wireless communication environments. Accordingly, the channel after least squares estimation is H LS,1 , H LS,2 ,…,H LS,T Combining these channel matrices at time T, we get the three-dimensional tensors of the channel signals H = [H1, H2, ..., H T ]∈C MXSXT , H LS =[H LS,1 , H LS,2 ,…,H LS,T ]∈C MXSXT We take the obtained three-dimensional tensor channel signals H and H LS They are used as the original channel dataset and the contaminated channel dataset respectively for offline training of artificial intelligence neural network.

[0069] For the three-dimensional tensor channel signals H and H LS Perform dimension reorganization operation, fix the time dimension T, transform each MXS matrix into a vector of MSX1, and then arrange it according to the time dimension to obtain matrices H′ and H′ of size MSXT LS .

[0070] Use a PXK sliding window with a sliding step size p to act on H′ and H′ LS The time dimension T is used to select matrix elements to obtain the channel picture of PXK, where P≤MS, K≤T. In particular, if P=MS, the total number of channel pictures obtained is The Floor(x) function represents the rounding down of the variable x. In the present invention, the channel matrix selected by the sliding window is collectively referred to as a channel image.

[0071] The real and imaginary parts of the obtained PXK channel images are separated to obtain the three-dimensional channel tensor of PXKX2. These PXK channel images or three-dimensional channel tensors of PXKX2 are input into the neural network for training to obtain the weight coefficient of each layer. The trained neural network can then be deployed to the device end.

[0072] Accordingly, in the online inference phase, it is necessary to wait for K time points during the system initialization phase to obtain the initial noise observation tensor H LS =[H LS,1 , H LS,2 ,…,H LS,K ], follow the above steps to obtain the channel image of PXKX2, input it into the trained neural network, and after the trained operation of each layer, the estimated value of the original channel can be obtained.

[0073] From the K+1th moment on, take the noise observation tensor HLS =[H LS,2 , H LS,3 ,…,H LS,K+1 ] as the input tensor, and follow the above steps to get the channel image of PXKX2, input the trained neural network, and the target channel H K+1 It can be estimated by the deployed neural network, which means that from the K+1th moment onwards, the system can perform real-time estimation of the channel.

[0074] For the target channel H at time K+2 K+2 The estimate of the noise observation tensor H can be taken LS =[H LS,3 , H LS,4 ,…,H LS,K+2 ] as the input tensor, which is estimated by the deployed neural network, and so on.

[0075] It should be noted that the noise observation tensor H used for offline training and online inference is LS , or you can use y in formula (1) directly without the least squares channel estimation t replace.

[0076] In order to prove the creativity and technical value of the technical solution of the present invention, this section provides application examples of the claimed technical solution on specific products or related technologies.

[0077] To reduce user-side costs, 3GPP-standard wireless communication systems typically use only a single antenna. To ensure that the acquired channel image contains sufficient information, the base station must be equipped with a large-scale antenna, which increases the cost of acquiring the channel image. The novel channel image acquisition method proposed in this paper can be applied to 3GPP-standard wireless communication systems, breaking the limitations of traditional technical solutions on the size of antennas at both the transmitting and receiving ends.

[0078] Wireless communication systems based on the WiFi standard are primarily designed for indoor coverage. A common configuration is two or four base station antennas and one user antenna. Existing methods for acquiring channel images have dimensions of 2x1 or 4x1. After dimensional reorganization, the maximum channel image size can be 2x2. This lacks sufficient information, making it difficult to effectively train and infer artificial intelligence neural networks. Existing channel estimation algorithms are also very limited in accuracy. The method proposed in this paper, by introducing the time dimension and sliding window selection, can easily generate a sufficiently effective channel image for neural network training.

[0079] Similarly, the new channel image acquisition method proposed in the present invention can be applied to other wireless communication systems such as industrial Internet standards.

[0080] The following content describes the test process with data, charts, etc.

[0081] like Figure 3 and Figure 4 As shown in the figure, we model the wireless environment according to the urban macro station UMa (Urban Macro) defined by 3GPP, set the base station to be equipped with 64 antenna units and the user end to be equipped with 1 antenna unit. When the user moves at speeds of 3 kilometers per hour and 60 kilometers per hour respectively, the deep learning compressed sensing algorithm DL-CS based on the channel image acquisition method of the present invention achieves optimal performance at low signal-to-noise ratios and near-optimal LMMSE channel estimation performance at high signal-to-noise ratios, but its implementation complexity is much lower than LMMSE.

[0082] Figure 3 The channel estimation for a user moving at 3 km / h in a 3GPP UMa environment is shown; Figure 4 The figure shows the channel estimation for a user moving at 60 kilometers per hour in a 3GPP UMa environment.

[0083] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0084] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A new channel image acquisition method, characterized in that: The following steps are involved: Step 1: Offline training. During the offline training phase, the signal at time T is collected, where T , get the original channel signal of the three-dimensional tensor and the received signal after passing through the channel , respectively as the real signal used for neural network training and the input signal to be estimated, where M is the number of antennas of the receiving device, S is the number of transmitted data streams, and T is the time observation length; Step 2: Dimensional reorganization, three-dimensional tensor channel signal and Perform dimension reorganization operation, fix the time dimension T, transform each M×S matrix into a MS×1 vector, and then arrange it according to the time dimension to obtain a matrix of size MS×T and ; Step 3: Sliding window selection, use a sliding window of size P×K with a step size of p to select the real signal and receive signals Slide on the time dimension and select a channel image of size P×K, where P , K is the time dimension and K ; Step 4: Real and imaginary part separation operation is performed on the obtained P×K channel image to obtain a P×K×2 channel image; Step 5: Neural network training: Get the P×K×2 channel image for neural network training, get the weight coefficient of each layer, and deploy the trained neural network to the device. Step 6: Online speculation. In the online speculation stage, the system waits for K moments to obtain the initial received signals from the 1st to the Kth moment. Step 7: The observation signal of the system at the K+1th time is composed of the signals from the 2nd to the K+1th, and so on.

2. The novel channel image acquisition method according to claim 1, characterized in that: The time dimension is used to generate the channel tensors and channel images required by the neural network; In the offline training phase, the original channel signal at time T and the signal after passing through the channel, where T , respectively, as the real signal and input signal used for neural network training.

3. The novel channel image acquisition method according to claim 1, characterized in that: The input signal used for training is the signal received directly after passing through the channel , or a signal that has been preliminarily processed, where the preliminarily processed methods include but are not limited to: least squares (LS), least linear mean square (LMMSE); The sliding window in the time dimension acts on the real signal and the input signal respectively to select one or more channel images.

4. The novel channel image acquisition method according to claim 1, wherein: Select good channel images to train the neural network and obtain the weight coefficients of each layer; In the online inference phase, the system waits for a certain initial time K to obtain the initial received signal from the 1st to the Kth time, where K , where the K moment is consistent with or inconsistent with the sliding window time dimension K used in the previous offline training phase; The received signal at the K+1th time is composed of the received signals from the 2nd to the K+1th time, and so on.

5. The novel channel image acquisition method according to claim 1, characterized in that: The online inference stage also uses a sliding window in the time dimension to act on the received signal to select one or more channel images; The sliding window size and its step size in the online inference phase are the same as or different from those used in the offline training phase.

6. The novel channel image acquisition method according to claim 1, characterized in that: The initial received signals or channel images from the 1st to the Kth time are input into the trained neural network to obtain an estimated value; the received signal or channel image obtained at the K+1th time is input into the trained neural network to obtain the channel estimate at the K+1th time, and so on.

7. The novel channel image acquisition method according to claim 1, characterized in that: Step 7: The observation signal of the system at time K+1 is composed of the signal from the 2nd to the K+1th, and so on: For the received signal, a sliding window of size L×Q is used with a step size q to generate an L×Q channel image; Perform real and imaginary part separation on the obtained L×Q channel image to obtain an L×Q×2 channel image; The obtained L×Q×2 channel image is input into the trained neural network model to obtain the corresponding true channel estimation value.

8. A channel image acquisition system, characterized in that: include: The offline training module is used to collect signals at time T during the offline training phase, where T , get the original channel signal of the three-dimensional tensor and the received signal after passing through the channel , respectively as the real signal used for neural network training and the input signal to be estimated, where M is the number of antennas of the receiving device, S is the number of transmitted data streams, and T is the time observation length; Dimensional reorganization module for three-dimensional tensor channel signals and Perform dimension reorganization operation, fix the time dimension T, transform each M×S matrix into a MS×1 vector, and then arrange it according to the time dimension to obtain a matrix of size MS×T and ; Sliding window selection module, used for a sliding window of size P×K with a step size p in the obtained real signal and receive signals Slide on the time dimension and select a channel image of size P×K, where P , K is the time dimension and K ; A real and imaginary part separation operation module is used to perform a real and imaginary part separation operation on the obtained P×K channel image to obtain a P×K×2 channel image; The neural network training module is used to obtain the P×K×2 channel image for neural network training, obtain the weight coefficient of each layer, and deploy the trained neural network to the device end; The online inference module is used in the online inference stage. The system waits for K moments to obtain the initial received signals from the 1st to the Kth moments.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the new channel image acquisition method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the novel channel image acquisition method according to any one of claims 1 to 7.

Images