An AI-based noise reduction channel estimation method, device and computer equipment

Abstract

The application relates to the technical field of wireless communication, and discloses a channel estimation method and device based on AI noise reduction and a computer device, the method comprising the following steps: extracting a demodulation reference signal from a received signal to perform least square channel estimation, obtaining an LS channel estimation matrix to calculate a phase offset between subcarriers and a common time offset index; generating a first phase compensation factor corresponding to each subcarrier based on the phase offset between the subcarriers, and generating a common second phase compensation factor based on the common time offset index; performing phase rotation compensation on the LS channel estimation matrix by using the first phase compensation factor and the second phase compensation factor, obtaining an input feature matrix from which the influence of time offset is removed, inputting the input feature matrix into a pre-trained AI noise reduction network model, and outputting a noise-reduced channel estimation matrix. The application can effectively separate noise and channel information in the received signal, significantly improve the anti-noise capability and precision of channel estimation, and ensure low calculation complexity and easy deployment.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a channel estimation method, apparatus, and computer device based on AI noise reduction. Background Technology

[0002] In modern wireless communication systems (such as 5G / 6G, IoT, and satellite communications), accurately obtaining channel state information (CSI) is a crucial prerequisite for equalization, demodulation, and decoding at the receiver. Traditional channel estimation methods mainly include least squares (LS) and minimum mean square error (MMSE). However, while the LS method is computationally simple, it is highly sensitive to noise interference in low signal-to-noise ratio environments, leading to a significant decrease in estimation accuracy. Although the MMSE method achieves better performance, it involves large-scale matrix inversion operations. With the increase in antenna size in massive MIMO and millimeter-wave communications, its computational complexity increases exponentially, making it difficult to meet the low latency requirements of real-time communication systems.

[0003] To overcome the bottlenecks of traditional methods, channel estimation and denoising using artificial intelligence technologies such as deep learning has become a research hotspot in recent years. Although existing AI channel estimation methods have shown certain advantages in extracting complex channel features and suppressing noise, they still face serious challenges in practical engineering deployments. Specifically, existing AI models often have overly complex network structures, resulting in high computational overhead during inference and making it difficult to achieve real-time operation on resource-constrained terminal devices or base stations. At the same time, existing solutions lack effective preprocessing mechanisms for physical layer characteristics such as timing offset, resulting in insufficient generalization ability of the models to different channel scenarios and difficulty in balancing high-precision denoising effects with low-cost computational resource consumption. Summary of the Invention

[0004] This invention provides a channel estimation method, apparatus, and computer device based on AI noise reduction, which can effectively separate noise and channel information in the received signal, significantly improve the noise resistance and accuracy of channel estimation, while ensuring low computational complexity and ease of deployment.

[0005] This invention provides a channel estimation method based on AI noise reduction, comprising:

[0006] S1. Extract the demodulation reference signal DMRS from the received signal, and perform least squares channel estimation on the demodulation reference signal to obtain the LS channel estimation matrix;

[0007] S2. Calculate the phase offset between subcarriers and the common time offset index based on the LS channel estimation matrix. ;

[0008] S3. Generate a first phase compensation factor corresponding to each subcarrier based on the phase offset between the subcarriers, and generate a common second phase compensation factor based on the common time offset index;

[0009] S4. Perform phase rotation compensation on the LS channel estimation matrix using the first phase compensation factor and the second phase compensation factor to obtain the input feature matrix with the time offset effect removed.

[0010] S5. Input the input feature matrix into the pre-trained AI denoising network model and output the denoised channel estimation matrix; wherein, the AI ​​denoising network model adopts a dual-branch structure, including a first branch for feature extraction and downsampling and a second branch for feature mapping and upsampling; both the first branch and the second branch contain a denoising module, which is composed of a self-attention unit, a layer normalization unit and a one-dimensional convolution unit.

[0011] Further, in S2, the phase offset between the subcarriers The calculation formula is:

[0012] ;

[0013] Where k is the DMRS subcarrier index, M is the number of subcarriers in the interval, and N is the number of subcarriers occupied by the DMRS. The LS channel estimation matrix is... This indicates the conjugate transpose. This indicates the calculation of the phase angle;

[0014] The common time offset index The calculation formula is:

[0015] ;

[0016] in, The number of FFT points in the system. This is the index of the starting position of resource block RB.

[0017] Further, in S3, the first phase compensation factor and the second phase compensation factor The calculation formulas are as follows:

[0018] ; ;

[0019] Where k is the subcarrier index of the DMRS. M is the number of subcarriers spaced between the two DMRS subcarriers.

[0020] Furthermore, in S4, the input feature matrix after removing the time bias effect... The calculation method is as follows:

[0021] ;

[0022] in, This represents the element-wise multiplication operation of a matrix.

[0023] Furthermore, the one-dimensional convolutional unit in the noise reduction module specifically includes cascaded point convolution and depth convolution; the computation path of the noise reduction module also includes a Gaussian error linear unit activation function cascaded in series.

[0024] Furthermore, the pre-trained AI denoising network model is trained in the following way:

[0025] Construct an initial training dataset, which contains the LS channel estimation matrix. and the corresponding ideal channel estimation matrix ;

[0026] Regarding the and Phase rotation compensation is performed using the calculation methods described in steps S2 to S4 respectively, resulting in an input feature matrix free from time offset effects. and label matrix ; wherein, the tag matrix The calculation formula is: ;by As model input, As model labels, the mean absolute error is used as the loss function to iteratively train the AI ​​noise reduction network model.

[0027] Furthermore, the iterative training process also includes a data augmentation step:

[0028] In each iteration, a random pair of inputs is used. and as a label The data is truncated in the frequency domain and padded with zeros to maintain data dimension consistency; at the same time, a preset percentage a% of the data is forcibly retained without the truncation and zero-padding operations.

[0029] Furthermore, in S5, the process of outputting the denoised channel estimation matrix using the AI ​​denoising network model adopts a two-level parallel deployment architecture:

[0030] The first level is time-slot-level parallelism: an independent thread is allocated to each processing time slot through a thread pool, and noise reduction calculations are performed in parallel according to the time slot; the second level is antenna-level parallelism: within the processing thread of a single time slot, inference calculations are performed concurrently for different receiving antennas; and the process also introduces a CPU core binding mechanism, which binds the thread performing the calculation to a specified CPU core and improves its scheduling priority.

[0031] The present invention also provides a channel estimation device based on AI noise reduction. Based on the channel estimation method based on AI noise reduction described above, the device includes:

[0032] The extraction unit is used to extract the demodulation reference signal DMRS from the received signal and perform least squares channel estimation on the demodulation reference signal to obtain the LS channel estimation matrix.

[0033] The calculation module is used to calculate the phase offset between subcarriers and the common time offset index based on the LS channel estimation matrix. ;

[0034] The generation module is used to generate a first phase compensation factor corresponding to each subcarrier based on the phase offset between the subcarriers, and to generate a common second phase compensation factor based on the common time offset index.

[0035] The compensation module is used to perform phase rotation compensation on the LS channel estimation matrix using the first phase compensation factor and the second phase compensation factor to obtain an input feature matrix that removes the time offset effect.

[0036] The output module is used to input the input feature matrix into a pre-trained AI denoising network model and output the denoised channel estimation matrix; wherein, the AI ​​denoising network model adopts a dual-branch structure, including a first branch for feature extraction and downsampling and a second branch for feature mapping and upsampling; both the first branch and the second branch contain a denoising module, which is composed of a self-attention unit, a layer normalization unit and a one-dimensional convolution unit.

[0037] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0038] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0039] The beneficial effects of this invention are as follows:

[0040] This invention establishes an AI-based denoising channel estimation method that enables adaptive channel denoising. Simultaneously, it establishes a data preprocessing method, reducing the model's data size requirements and lowering computational overhead during training. Furthermore, this invention constructs a model deployment framework that efficiently deploys the model to the communication system while conserving computational resources. By comprehensively considering neural network model structure, data preprocessing methods, and the model deployment framework, this invention significantly improves system throughput while meeting the latency requirements of the communication system. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the channel estimation method based on AI noise reduction of the present invention.

[0042] Figure 2 This is a schematic diagram of the time offset compensation process in this invention.

[0043] Figure 3 This is an architecture diagram of the AI ​​noise reduction network model in this invention.

[0044] Figure 4 This is a schematic diagram of the noise reduction module in this invention.

[0045] Figure 5 This is a model deployment framework diagram of the AI ​​noise reduction network model in this invention.

[0046] Figure 6 This is a schematic diagram of the channel estimation device for AI noise reduction according to the present invention.

[0047] Figure 7 This is a schematic diagram of the internal structure of the computer device in this invention.

[0048] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0049] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0050] This invention provides a channel estimation method and deployment system based on artificial intelligence (AI) denoising technology. It studies denoising in channel estimation based on neural networks, aiming to adaptively reduce channel noise while minimizing computational overhead during training and inference, thereby meeting the system's real-time requirements. This method can be applied to 5G / 6G mobile communication systems, the Internet of Things (IoT), satellite communication, and other scenarios requiring high-precision channel estimation.

[0051] like Figure 1 As shown, the present invention provides a channel estimation method based on AI noise reduction, comprising:

[0052] S1. Extract the demodulation reference signal (DMRS) from the received signal, and perform least-squares channel estimation on the demodulation reference signal to obtain the LS (Least Squares) channel estimation matrix; S2. Calculate the phase offset between subcarriers and the common time offset index based on the LS channel estimation matrix. .

[0053] like Figure 2 As shown, this involves performing TO (Timing Offset) compensation to eliminate the phase offset caused by TO. TO is divided into time offset between subcarriers and common time offset:

[0054] Phase offset between subcarriers The calculation formula is:

[0055] ;

[0056] Where k is the DMRS subcarrier index; M is the number of subcarriers in the interval, which is set to 2 in this embodiment; N is the number of subcarriers occupied by DMRS, which is set to 1638 in this embodiment; The LS channel estimation matrix is... This indicates the conjugate transpose. This indicates the calculation of the phase angle;

[0057] The common time offset index The calculation formula is:

[0058] ;

[0059] in, The system's FFT points are set to 4096 in this embodiment; This is the index of the starting position of resource block RB.

[0060] S3. Generate a first phase compensation factor corresponding to each subcarrier based on the phase offset between the subcarriers, and generate a common second phase compensation factor based on the common time offset index.

[0061] First phase compensation factor and the second phase compensation factor The calculation formulas are as follows:

[0062] ; ;

[0063] Where k is the subcarrier index of the DMRS. M is the number of subcarriers spaced between the two DMRS subcarriers.

[0064] S4. The LS channel estimation matrix is ​​phase rotated and compensated using the first phase compensation factor and the second phase compensation factor to obtain the input feature matrix with the time offset effect removed.

[0065] Input feature matrix after removing time bias The calculation method is as follows:

[0066] ;

[0067] in, This represents the element-wise multiplication operation of a matrix.

[0068] S5. Input the input feature matrix into the pre-trained AI denoising network model and output the denoised channel estimation matrix.

[0069] like Figure 3 As shown, the AI ​​denoising network model is a hybrid neural network structure that integrates attention mechanisms and convolutional feature extraction. It employs a dual-branch structure, including a first branch for feature extraction and downsampling (performing dimensionality compression and main feature extraction on the input) and a second branch for feature mapping and upsampling (performing dimensionality enhancement and feature mapping on the results of the first branch). Both the first and second branches contain denoising modules, which consist of self-attention units, layer normalization units, and one-dimensional convolutional units. The attention units and layer normalization units are used to construct global dependencies, and by concatenating 1D convolutions and the GELU activation function (Gaussian Error Linear Unit), they are used to enhance the extraction of local features. Figure 4 As shown, the noise reduction module adopts a series structure of "point convolution - depthwise convolution - point convolution", and the specific calculation path is as follows:

[0070] Pointwise Conv1d: First, the input features are passed through the first point convolutional layer (KernelSize=1), which expands the number of feature channels (for example, to twice the original number of channels) to enrich the expression dimension of the features.

[0071] Nonlinear activation (GELU): The upgraded features are activated by Gaussian error linear units (GELU), introducing nonlinear factors;

[0072] Feature extraction (Depthwise Conv1d): The activated features are fed into a deep convolutional layer (Kernel Size=k, such as k=3 or 5), and spatial (frequency domain) feature extraction is performed independently on each channel. This step can effectively reduce the number of parameters.

[0073] Pointwise Conv1d: Finally, after a second pointwise convolutional layer (Kernel Size=1), the number of feature channels is compressed back to the original dimension (or the dimension required for the next layer), and a residual connection is made with the input. This structure significantly reduces computational complexity while maintaining feature extraction capabilities.

[0074] In one embodiment, the one-dimensional convolutional unit in the noise reduction module specifically includes cascaded point convolution and depth convolution; the computation path of the noise reduction module also includes a Gaussian error linear unit activation function cascaded in series.

[0075] In one embodiment, the pre-trained AI noise reduction network model is trained in the following manner:

[0076] (1) In a computer simulation environment (e.g., a link-level simulation platform based on the 3GPP protocol), the DMRS signal of the transmitting end and the received signal after transmission through the channel are generated using a preset channel model (e.g., CDL or TDL channel model).

[0077] Get input data At the simulated receiver, least squares (LS) channel estimation is performed on the received signal to obtain the LS channel estimation matrix including noise and interference. .

[0078] Get tag data The current channel frequency response (CFR) true value is directly read from the channel generator of the simulation platform and used as the ideal channel estimation matrix. Pairs Save the dataset to create the initial training dataset. In this embodiment, a total of 50,000 training data points were collected by traversing different signal-to-noise ratio (SNR) and Doppler shift scenarios.

[0079] in, This represents the channel estimation matrix obtained by LS. This represents the channel matrix obtained from the ideal channel estimation. Specifically, and These two matrices have the same dimensions. ,in It is the number of subcarriers occupied by the DMRS signal. It is the number of symbols occupied by the DMRS signal. This refers to the number of receiving antennas.

[0080] In this embodiment, Set to 1638, Set to 2, Set it to 4.

[0081] (2) Regarding the above and Phase rotation compensation is performed using the calculation methods described in steps S2 to S4, ultimately yielding an input feature matrix free from time-off effects. and label matrix , as the training set; wherein, the label matrix The calculation formula is: ;

[0082] (3) with As model input, As model labels, the AI ​​denoising network model is iteratively trained using MAE (Mean Absolute Error) as the loss function, with a training learning rate of 0.0001 and an iteration count of 100.

[0083] The iterative training process also includes a data augmentation step:

[0084] In each iteration, a random pair of inputs is used. and as a label Truncation is performed in the frequency domain dimension, and zero-padding is performed to maintain data dimension consistency; at the same time, a preset percentage a% (e.g., a value of 5) of the data is forcibly retained without the truncation and zero-padding operations.

[0085] In one embodiment, such as Figure 5 As shown, the process of outputting the denoised channel estimation matrix using the AI ​​denoising network model adopts a two-level parallel deployment architecture:

[0086] The first level is slot-level parallelism: an independent thread is allocated to each processing slot through a thread pool, and noise reduction calculations are performed in parallel on a slot-by-slot basis. The second level is antenna-level parallelism: within a processing thread of a single slot, inference calculations are performed concurrently for different receiving antennas. Furthermore, the process introduces CPU core binding and thread priority control mechanisms. In computing environments that support hardware affinity settings (such as physical machine-based base station BBUs or edge computing devices), specific threads performing inference are bound to designated physical CPU cores through operating system interfaces (such as `sched_setaffinity` in Linux), and the scheduling policies of these threads are set to real-time priorities (such as `SCHED_FIFO`). In this way, threads are forced to run on dedicated cores, minimizing time jitter caused by operating system context switching and cache misses, thereby ensuring the determinism of channel estimation processing latency.

[0087] like Figure 6 As shown, the present invention also provides a channel estimation device based on AI noise reduction. Based on the channel estimation method based on AI noise reduction described above, the device includes:

[0088] Extraction unit 1 is used to extract the demodulation reference signal DMRS from the received signal and perform least squares channel estimation on the demodulation reference signal to obtain the LS channel estimation matrix.

[0089] Calculation module 2 is used to calculate the phase offset between subcarriers and the common time offset index based on the LS channel estimation matrix. ;

[0090] The generation module 3 is used to generate a first phase compensation factor corresponding to each subcarrier based on the phase offset between the subcarriers, and to generate a common second phase compensation factor based on the common time offset index.

[0091] Compensation module 4 is used to perform phase rotation compensation on the LS channel estimation matrix using the first phase compensation factor and the second phase compensation factor to obtain an input feature matrix that removes the time offset effect;

[0092] Output module 5 is used to input the input feature matrix into a pre-trained AI denoising network model and output the denoised channel estimation matrix; wherein, the AI ​​denoising network model adopts a dual-branch structure, including a first branch for feature extraction and downsampling and a second branch for feature mapping and upsampling; both the first branch and the second branch contain a denoising module, which is composed of a self-attention unit, a layer normalization unit and a one-dimensional convolution unit.

[0093] Each of the above modules is used to perform the corresponding steps in the AI-based noise reduction channel estimation method. The specific implementation method is as described in the above method embodiment, and will not be repeated here.

[0094] like Figure 7 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 7As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores all the data required for the AI-based noise reduction channel estimation method. The network interface allows communication with external terminals via a network connection. The computer program is executed by the processor to implement the AI-based noise reduction channel estimation method.

[0095] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0096] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described AI-based noise reduction channel estimation methods.

[0097] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0098] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0099] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A channel estimation method based on AI noise reduction, characterized in that, include: S1. Extract the demodulation reference signal DMRS from the received signal, and perform least squares channel estimation on the demodulation reference signal to obtain the LS channel estimation matrix; S2. Calculate the phase offset between subcarriers and the common time offset index based on the LS channel estimation matrix. Wherein, the phase offset between the subcarriers The calculation formula is: Where k is the DMRS subcarrier index, M is the number of subcarriers in the interval, and N is the number of subcarriers occupied by the DMRS. The LS channel estimation matrix is... This indicates the conjugate transpose. This indicates the calculation of the phase angle; the common time offset index The calculation formula is: ； in, The number of FFT points in the system. This is the index of the starting position of resource block RB; S3. Generate a first phase compensation factor corresponding to each subcarrier based on the phase offset between the subcarriers, and generate a common second phase compensation factor based on the common time offset index; S4. Perform phase rotation compensation on the LS channel estimation matrix using the first phase compensation factor and the second phase compensation factor to obtain the input feature matrix with the time offset effect removed. S5. Input the input feature matrix into the pre-trained AI denoising network model and output the denoised channel estimation matrix; wherein, the AI ​​denoising network model adopts a dual-branch structure, including a first branch for feature extraction and downsampling and a second branch for feature mapping and upsampling; both the first branch and the second branch contain a denoising module, which is composed of a self-attention unit, a layer normalization unit and a one-dimensional convolution unit.

2. The channel estimation method based on AI noise reduction according to claim 1, characterized in that, In S3, the first phase compensation factor and the second phase compensation factor The calculation formulas are as follows: ; ; Where k is the subcarrier index of the DMRS. M is the number of subcarriers spaced between the two DMRS subcarriers.

3. The channel estimation method based on AI noise reduction according to claim 2, characterized in that, In S4, the input feature matrix after removing the time offset effect. The calculation method is as follows: ； in, This represents the element-wise multiplication operation of a matrix.

4. The channel estimation method based on AI noise reduction according to claim 1, characterized in that, The one-dimensional convolutional unit in the noise reduction module specifically includes cascaded point convolution and depth convolution; the computation path of the noise reduction module also includes a Gaussian error linear unit activation function cascaded in series.

5. The channel estimation method based on AI noise reduction according to claim 1, characterized in that, The pre-trained AI noise reduction network model was trained in the following way: Construct an initial training dataset, which contains the LS channel estimation matrix. and the corresponding ideal channel estimation matrix ; Regarding the and Phase rotation compensation is performed using the calculation methods described in steps S2 to S4 respectively, resulting in an input feature matrix free from time offset effects. and label matrix ; wherein, the tag matrix The calculation formula is: ; by As model input, As model labels, the mean absolute error is used as the loss function to iteratively train the AI ​​noise reduction network model.

6. The channel estimation method based on AI noise reduction according to claim 5, characterized in that, The iterative training process also includes a data augmentation step: In each iteration, a random pair of inputs is used. and as a label The data is truncated in the frequency domain and padded with zeros to maintain data dimension consistency; at the same time, a preset percentage a% of the data is forcibly retained without the truncation and zero-padding operations.

7. The channel estimation method based on AI noise reduction according to claim 1, characterized in that, In S5, the process of outputting the denoised channel estimation matrix using the AI ​​denoising network model adopts a two-level parallel deployment architecture: The first level is time-slot-level parallelism: an independent thread is allocated to each processing time slot through a thread pool, and noise reduction calculations are performed in parallel according to the time slot; the second level is antenna-level parallelism: within the processing thread of a single time slot, inference calculations are performed concurrently for different receiving antennas; and the process also introduces a CPU core binding mechanism, which binds the thread performing the calculation to a specified CPU core and improves its scheduling priority.

8. A channel estimation device based on AI denoising, based on the channel estimation method based on AI denoising according to any one of claims 1-7, characterized in that, The device includes: The extraction unit is used to extract the demodulation reference signal (DMRS) from the received signal and perform least-squares channel estimation on the demodulation reference signal to obtain the LS channel estimation matrix; the calculation module is used to calculate the phase offset between subcarriers and the common time offset index based on the LS channel estimation matrix. The generation module is used to generate a first phase compensation factor corresponding to each subcarrier based on the phase offset between the subcarriers, and to generate a common second phase compensation factor based on the common time offset index. The compensation module is used to perform phase rotation compensation on the LS channel estimation matrix using the first phase compensation factor and the second phase compensation factor to obtain an input feature matrix that removes the time offset effect. The output module is used to input the input feature matrix into a pre-trained AI denoising network model and output the denoised channel estimation matrix; wherein, the AI ​​denoising network model adopts a dual-branch structure, including a first branch for feature extraction and downsampling and a second branch for feature mapping and upsampling; both the first branch and the second branch contain a denoising module, which is composed of a self-attention unit, a layer normalization unit and a one-dimensional convolution unit.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images