Channel estimation method of deep convolutional generative adversarial network based on Swin Transform
By combining Swin Transformer and DCGAN channel estimation methods, deep features of complex channels are extracted and channel estimation is optimized through the generation adversarial mechanism, the problem of inaccurate channel estimation in complex channel environments is solved, and high-resolution channel response recovery and communication are achieved Performance improvement.
Patent Information
- Application Number
- CN202411308846.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-19
- Publication Date
- 2024-12-27
AI Technical Summary
The existing MIMO-OTFS channel estimation methods are difficult to provide sufficient estimation accuracy in complex channel environments with high dynamic, nonlinearity, and deep learning-based methods fail to fully combine the advantages of different models to improve estimation effects.
Deep convolution based on Swin Transformer is used to generate adversarial networks, deep features of complex channel environments are extracted through Swin Transformer, and the adversarial training of DCGAN generator and discriminator generates preliminary channel estimation results that match the actual channel response, combining with the decoder and The upsampling module is further optimized to restore high-resolution channel response.
It significantly improves the accuracy and robustness of channel estimation, especially in high dynamic and strong multipath fading environments, it can show excellent anti-interference ability and improves communication performance and bit error rate performance.
Smart Images
Figure CN119210945A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a channel estimation method of a deep convolutional generative adversarial network based on Swin Transformer, and belongs to the technical field of wireless communications. Background Art
[0002] In recent years, with the rapid development of wireless communication technology, MIMO technology and OTFS technology have become important components of the next generation of wireless communication systems due to their superior performance in complex channel environments. MIMO technology can significantly improve the system's spectrum efficiency and anti-interference ability through the configuration of multiple antennas, while OTFS technology further enhances the robustness of the signal under time and frequency selective fading conditions through modulation in the delay-Doppler domain. However, in the face of complex wireless channel environments, how to accurately and effectively perform channel estimation is still a difficult problem that needs to be solved urgently.
[0003] In recent years, remarkable results have been achieved in the fields of image processing and natural language processing. These models can extract high-dimensional features from large amounts of data and generate high-quality output results through deep neural network structures. Therefore, using deep learning technology to improve the accuracy of channel estimation has become a research hotspot in the field of wireless communications. As a new type of visual Transformer, SwinTransformer achieves efficient extraction of complex features by introducing a sliding window mechanism and a hierarchical structure; while the Deep Convolutional Generative Adversarial Network (DCGAN) can generate channel estimation results close to the actual channel response through adversarial training between the generator and the discriminator.
[0004] However, most existing MIMO-OTFS channel estimation methods rely on traditional linear estimation and statistical assumptions, which makes it difficult to provide sufficient estimation accuracy in highly dynamic, nonlinear and complex channel environments. In addition, current deep learning-based channel estimation methods mostly focus on the application of a single technology and fail to fully combine the advantages of different deep learning models to improve the estimation effect. Summary of the invention
[0005] The technical problem to be solved by the present invention is to provide a channel estimation method of a deep convolutional generative adversarial network based on Swin Transformer, which solves the problem of inaccurate channel estimation in a highly dynamic, nonlinear and complex wireless channel environment.
[0006] The technical solution of the present invention is: a channel estimation method based on a deep convolutional generative adversarial network of Swin Transformer, the method combines Swin Transformer and DCGAN, and is used in a MIMO-OTFS system in a complex wireless channel environment. The method comprises the following steps:
[0007] S1: receiving a signal modulated by a MIMO antenna array from a wireless channel and preprocessing the signal, and then transforming the signal into a time-frequency domain through OTFS modulation;
[0008] S2: The preprocessed and time-frequency transformed OTFS signal is input into the encoder composed of Swin Transformer. The encoder uses the multi-layer attention mechanism in Swin Transformer to extract deep features that can reflect the changes in complex channel environment.
[0009] S3: The deep features extracted by the encoder are input into the DCGAN module together with the reference signal. The DCGAN module generates a preliminary channel estimation result that matches the actual channel response through the confrontation between the generator and the discriminator.
[0010] S4: The preliminary channel estimation results generated by the DCGAN module are input into the decoder, which further processes and optimizes the preliminary channel estimation results to restore accurate channel characteristics;
[0011] S5: Input the channel features processed by the decoder into the upsampling module for optimization processing, and restore the high-resolution channel response through upsampling;
[0012] S6: Using the restored high-resolution channel response, the MIMO signal modulated by OTFS is demodulated to finally restore the original data symbols and realize channel estimation of data in a complex channel environment.
[0013] Optionally, the preprocessing includes the following steps: symbol synchronization, frequency offset correction and removal of guard intervals, dividing the preprocessed signal samples into a training set and a test set, the training set is used for model training and optimization, and the test set is used to evaluate the performance of the model, the preprocessing step ensures that the signal is in an ideal state before entering the OTFS modulator, and ensures the accuracy and reliability of the subsequent channel estimation process, and the OTFS modulation performs time-frequency domain transformation through the following formula:
[0014]
[0015] Optional, where X T，V [k,l] is the symbol in the delay-Doppler domain, k is the delay index, l is the Doppler shift index, M is the number of samples of the Doppler shift, N is the number of samples of the delay, x[n,m] is the symbol in the time-frequency domain, n is the time index, and m is the frequency index.
[0016] Optionally, the encoder uses the multi-layer attention mechanism in Swin Transformer to extract deep features that can reflect changes in complex channel environments. The specific operation is: Swin Transformer includes a two-layer processing structure: the first layer adopts a window-based multi-head self-attention mechanism to calculate the relationship between channel features in a local non-overlapping delay-Doppler domain window, effectively extract local channel characteristics, while maintaining low computational complexity; the second layer adopts an offset window-based multi-head self-attention mechanism to establish connections between adjacent windows by offsetting windows between different levels to capture a wider range of spatial, temporal and frequency correlations; each layer also includes a multi-layer perceptron module, a GeLU nonlinear activation function and a layer normalization module to optimize the feature extraction process, and enhance the gradient propagation and training stability of the network through residual connections; the output of the window-based multi-head self-attention mechanism in the first layer is:
[0017]
[0018] Optional, where represents the output of the window-based multi-head self-attention mechanism in the first layer, h l-1 The normalized output of the representation layer is processed by the window-based multi-head self-attention mechanism module and h l-1 Add together to get
[0019] Optionally, the output of the window-based multi-head self-attention mechanism module is then input into the multi-layer perceptron module after layer normalization. The output of the multi-layer perceptron module is compared with The final output of the first layer is added together:
[0020]
[0021] Optional, where represents the output of the window-based multi-head self-attention mechanism in the first layer, h l Represents the final output of the first layer.
[0022] Optionally, the final output h of the first layer l As the input of the second layer, it is first normalized after reaching the second layer, and then input into the multi-head self-attention mechanism module based on the offset window. The output of the multi-head self-attention mechanism based on the offset window is the same as h l After the residual connection and layer normalization, it is input into the multi-layer perceptron module, and the final output is:
[0023]
[0024] Optional, where represents the output of the multi-head self-attention mechanism based on the offset window, h l+1 is the final output of the second layer.
[0025] Optionally, the specific operation of generating a preliminary channel estimation result that matches the actual channel response is: in the DCGAN module, the generator and the discriminator interact with each other through adversarial training. The generator is responsible for generating a channel estimation result similar to the actual channel response based on the input data, while the discriminator continuously distinguishes the difference between the generated channel estimation and the actual channel response. In each training, the generator tries to deceive the discriminator to generate an estimation result that is closer to the real channel, and the discriminator helps the generator improve the output by feeding back the generated error. As the training progresses, the discriminator's ability to distinguish gradually increases, and the generator also continuously optimizes its own performance during the adversarial process, and the output channel estimation result is getting closer and closer to the actual channel response. In the end, the generator can generate a preliminary channel estimation result that matches the real channel, providing a reliable foundation for subsequent channel estimation.
[0026] Optionally, the specific operation of step S4 is: the preliminary channel estimation result generated by the DCGAN module is input into the decoder for further processing, the decoder is composed of multiple convolutional layers, batch normalization layers and ReLU activation function components, the components work together to optimize the input preliminary channel estimation results layer by layer, the convolutional layer is responsible for extracting and enhancing the spatial features in the channel estimation results, the batch normalization layer stabilizes the training process and accelerates convergence through standardization, and the ReLU activation function introduces nonlinear transformation to enhance the expressive power of the model. Through the operation, the decoder can refine and optimize the preliminary channel estimation results, and gradually restore more accurate channel features, thereby improving the accuracy and reliability of the overall channel estimation.
[0027] Optionally, the specific operation of step S5 is: the channel features processed by the decoder will be input into the upsampling module for further optimization processing. The upsampling module first globally integrates and dimensionally expands the input channel features through a fully connected layer to capture more global context information and obtain the expanded channel features. Subsequently, the expanded channel features enter a convolutional layer, which performs convolution operations on feature maps to capture local spatial information, thereby improving the resolution and detail of the features. By means of a sliding window, the convolutional layer can aggregate features within the neighborhood, highlight important details, and reduce noise interference, thereby restoring a higher resolution channel response. Ultimately, the channel response after upsampling can more accurately reflect the actual channel conditions.
[0028] Optionally, the specific operation of step S6 is: first, the demodulator adjusts the received signal according to the channel response using the restored high-resolution channel response to reduce the interference and distortion caused by the time-varying characteristics and multipath effects. Subsequently, the compensated received signal is mapped back to the original data symbol to eliminate the noise and distortion introduced by the channel, and finally the accurate original data symbol is restored to achieve channel estimation and data recovery in a complex channel environment. The OTFS demodulation process is expressed by the following formula:
[0029]
[0030] Optional, where x[n,m] is the symbol in the time-frequency domain, n is the time index, m is the frequency index, M is the number of samples of Doppler shift, N is the number of samples of delay, and X T，V [k,l] is a symbol in the delay-Doppler domain, k is the delay index, and l is the Doppler shift index.
[0031] The beneficial effects of the present invention are:
[0032] 1. A novel MIMO-OTFS channel estimation scheme combining Swin Transformer and DCGAN is proposed to solve the channel estimation problem in complex wireless channel environments.
[0033] 2. By utilizing the multi-layer attention mechanism and sliding window strategy of Swin Transformer, the deep features of the channel can be effectively extracted. At the same time, the generative adversarial mechanism of the DCGAN module can further optimize the channel estimation results to make them closer to the actual channel response. Preliminary assumptions show that this method is expected to significantly improve the accuracy and robustness of channel estimation, especially in high-dynamic and strong multipath fading environments, and is expected to show excellent anti-interference capabilities.
[0034] 3. Through the joint optimization processing of the decoder and upsampling module, high-resolution channel response can be restored, further improving the communication performance and bit error rate performance of the system. This provides a possible reference direction for the design of future wireless communication systems. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 It is a schematic diagram of the process of the present invention;
[0036] Figure 2 A diagram showing the structure of a channel estimation network according to the present invention;
[0037] Figure 3 This is the DCGAN structure diagram of the present invention. DETAILED DESCRIPTION
[0038] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods. The following examples are only used to more clearly illustrate the technical solutions of the present invention, but the protection scope of the present invention is not limited to the examples described below.
[0039] At present, deep learning methods are gradually applied to the research on MIMO-OTFS channel estimation. However, in complex channel environments, especially when multipath effects are significant and time-frequency selectivity is strong, the accuracy of many channel estimation methods is still limited. Although deep learning technology has been introduced into the research on channel estimation, the MIMO-OTFS channel estimation method combining Swin Transformer and deep convolutional generative adversarial network (DCGAN) has not been studied in depth. Swin Transformer, with its excellent multi-scale feature extraction capability, can extract deep features in complex channel environments, while DCGAN has unique advantages in generating realistic channel models. The present invention proposes a MIMO-OTFS channel estimation method combining Swin Transformer and DCGAN. By using Swin Transformer as an encoder to extract the time-frequency two-dimensional features of the channel, and using DCGAN to perform fine restoration of the channel, the method aims to cope with the highly nonlinear and dynamic changes of the channel environment, thereby improving the accuracy and robustness of channel estimation. Especially under the OTFS modulation framework, the method can more effectively capture the complex characteristics of the channel and adapt to the changing channel environment. This method is expected to provide an innovative solution for efficient and reliable MIMO-OTFS channel estimation in future complex wireless communication systems.
[0040] The method of the present invention comprises the steps of designing a channel estimation model, extracting channel features using Swin Transformer, performing channel recovery in combination with DCGAN, and analyzing the accuracy and performance of channel estimation.
[0041] Example 1: A specific implementation example of the present invention provides a channel estimation method based on a deep convolutional generative adversarial network of Swin Transformer, the method is as follows Figure 1-3 As shown, the following steps are included:
[0042] like Figure 1 As shown, step S1 is performed to receive a signal modulated by a MIMO antenna array from a wireless channel and pre-process the signal. Subsequently, the signal is transformed in the time-frequency domain through OTFS modulation.
[0043] like Figure 1As shown, step S2 is performed to input the preprocessed and time-frequency transformed OTFS signal into the encoder composed of Swin Transformer. The encoder uses the multi-layer attention mechanism in Swin Transformer to extract deep features that can reflect the changes in complex channel environment. The structure of Swin Transformer is shown in Figure 2 As shown. The output of the window-based multi-head self-attention mechanism in the first layer is:
[0044]
[0045] in, represents the output of the window-based multi-head self-attention mechanism in the first layer, h l-1 The normalized output of the representation layer is processed by the window-based multi-head self-attention mechanism module and h l-1 Add together to get
[0046] Then, the output of the window-based multi-head self-attention mechanism module is input into the multi-layer perceptron module after layer normalization. The final output of the first layer is added together:
[0047]
[0048] in, represents the output of the window-based multi-head self-attention mechanism in the first layer, h l Represents the final output of the first layer.
[0049] The final output of the first layer is h l As the input of the second layer, it is first normalized after reaching the second layer, and then input into the multi-head self-attention mechanism module based on the offset window. The output of the multi-head self-attention mechanism based on the offset window is the same as h l After the residual connection and layer normalization, it is input into the multi-layer perceptron module, and the final output is:
[0050]
[0051] in, represents the output of the multi-head self-attention mechanism based on the offset window, h l+1 is the final output of the second layer.
[0052] like Figure 1As shown, step S3 is performed to input the deep features extracted by the encoder together with the reference signal into the DCGAN module. The DCGAN module generates a preliminary channel estimation result that matches the actual channel response through the confrontation between the generator and the discriminator. The structure of the generator and the discriminator is shown in Figure 3 shown.
[0053] like Figure 1 As shown, step S4 is performed to input the preliminary channel estimation result generated by the DCGAN module into the decoder, and the decoder further processes and optimizes the preliminary channel estimation result to restore the accurate channel characteristics; the structure of the decoder is as follows Figure 2 shown.
[0054] like Figure 1 As shown, step S5 is performed to input the channel features processed by the decoder into the upsampling module for optimization processing, and the high-resolution channel response is restored by upsampling; the structure of the upsampling module is as follows Figure 2 shown.
[0055] like Figure 1 As shown, step S6 is performed to demodulate the MIMO signal modulated by OTFS by using the restored high-resolution channel response, and finally restore the original data symbols to achieve channel estimation of data in a complex channel environment.
[0056] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.
Claims
1. A channel estimation method based on Swin Transformer and deep convolutional generative adversarial network, characterized in that: The method comprises the following steps: S1: receiving a signal modulated by a MIMO antenna array from a wireless channel and preprocessing the signal, and then transforming the signal into a time-frequency domain through OTFS modulation; S2: The preprocessed and time-frequency transformed OTFS signal is input into the encoder composed of Swin Transformer. The encoder uses the multi-layer attention mechanism in Swin Transformer to extract deep features that can reflect the changes in complex channel environment. S3: The deep features extracted by the encoder are input into the DCGAN module together with the reference signal. The DCGAN module generates a preliminary channel estimation result that matches the actual channel response through the confrontation between the generator and the discriminator. S4: The preliminary channel estimation results generated by the DCGAN module are input into the decoder, which further processes and optimizes the preliminary channel estimation results to restore accurate channel characteristics; S5: Input the channel features processed by the decoder into the upsampling module for optimization processing, and restore the high-resolution channel response through upsampling; S6: Using the restored high-resolution channel response, the MIMO signal modulated by OTFS is demodulated to finally restore the original data symbols and realize channel estimation of data in a complex channel environment.
2. According to claim 1, a channel estimation method based on Swin Transformer and deep convolutional generative adversarial network is characterized in that: In step S1, the preprocessing includes the following steps: symbol synchronization, frequency offset correction and removal of guard interval, dividing the preprocessed signal samples into a training set and a test set, the training set is used for model training and optimization, and the test set is used to evaluate the performance of the model. The preprocessing step ensures that the signal is in an ideal state before entering the OTFS modulator, ensuring the accuracy and reliability of the subsequent channel estimation process. The OTFS modulation performs time-frequency domain transformation, which is expressed by the following formula: Among them, X T，V [k,l] is the symbol in the delay-Doppler domain, k is the delay index, l is the Doppler shift index, M is the number of samples of the Doppler shift, N is the number of samples of the delay, x[n,m] is the symbol in the time-frequency domain, n is the time index, and m is the frequency index.
3. According to claim 1, a channel estimation method based on Swin Transformer and deep convolutional generative adversarial network is characterized in that: The encoder uses the multi-layer attention mechanism in Swin Transformer to extract deep features that can reflect changes in complex channel environments. The specific operation is as follows: Swin Transformer includes a two-layer processing structure: the first layer adopts a window-based multi-head self-attention mechanism to calculate the relationship between channel features in a local non-overlapping delay-Doppler domain window, effectively extract local channel characteristics, while maintaining low computational complexity; the second layer adopts an offset window-based multi-head self-attention mechanism to establish connections between adjacent windows by offsetting windows between different levels, capturing a wider range of spatial, temporal and frequency correlations; each layer also includes a multi-layer perceptron module, a GeLU nonlinear activation function and a layer normalization module to optimize the feature extraction process, and enhance the gradient propagation and training stability of the network through residual connections; the output of the window-based multi-head self-attention mechanism in the first layer is: in, represents the output of the window-based multi-head self-attention mechanism in the first layer, h l-1 The normalized output of the representation layer is processed by the window-based multi-head self-attention mechanism module and h l-1 Add together to get Then, the output of the window-based multi-head self-attention mechanism module is input into the multi-layer perceptron module after layer normalization. The final output of the first layer is added together: in, represents the output of the window-based multi-head self-attention mechanism in the first layer, h l Represents the final output of the first layer; The final output of the first layer is h l As the input of the second layer, it is first normalized after reaching the second layer, and then input into the multi-head self-attention mechanism module based on the offset window. The output of the multi-head self-attention mechanism based on the offset window is the same as h l After the residual connection and layer normalization, it is input into the multi-layer perceptron module, and the final output is: in, represents the output of the multi-head self-attention mechanism based on the offset window, h l+1 is the final output of the second layer.
4. The channel estimation method based on Swin Transformer and deep convolutional generative adversarial network according to claim 1, characterized in that: The specific operation of generating the preliminary channel estimation result that matches the actual channel response is as follows: in the DCGAN module, the generator and the discriminator interact with each other through adversarial training. The generator is responsible for generating a channel estimation result similar to the actual channel response based on the input data, and the discriminator continuously distinguishes the difference between the generated channel estimation and the actual channel response. In each training, the generator attempts to deceive the discriminator to generate an estimation result that is closer to the actual channel, and the discriminator helps the generator improve the output by feeding back the generated error. As the training progresses, the discriminator's ability to distinguish is gradually enhanced, and the generator also continuously optimizes its own performance in the adversarial process. The output channel estimation result is getting closer and closer to the actual channel response. Finally, the generator generates a preliminary channel estimation result that matches the real channel, providing a reliable basis for subsequent channel estimation.
5. The channel estimation method based on Swin Transformer and deep convolutional generative adversarial network according to claim 1, characterized in that: The specific operation of step S4 is: the preliminary channel estimation result generated by the DCGAN module is input into the decoder for further processing. The decoder is composed of multiple convolutional layers, batch normalization layers and ReLU activation function components. The components work together to optimize the input preliminary channel estimation results layer by layer. The convolutional layer is responsible for extracting and enhancing the spatial features in the channel estimation results. The batch normalization layer stabilizes the training process and accelerates convergence through standardization. The ReLU activation function introduces nonlinear transformation to enhance the expression ability of the model. Through the above operation, the decoder can refine and optimize the preliminary channel estimation results.
6. The channel estimation method based on Swin Transformer and deep convolutional generative adversarial network according to claim 1, characterized in that: The specific operation of step S5 is: the channel features processed by the decoder will be input into the upsampling module for further optimization processing. The upsampling module first globally integrates and dimensionally expands the input channel features through a fully connected layer to capture more global context information and obtain the expanded channel features. Subsequently, the expanded channel features enter a convolutional layer. The convolutional layer captures local spatial information by performing convolution operations on the feature map. By means of a sliding window, the convolutional layer can aggregate features in the neighborhood, highlight important details, and reduce noise interference, thereby restoring high-resolution channel responses.
7. The channel estimation method based on Swin Transformer and deep convolutional generative adversarial network according to claim 1, characterized in that: The specific operation of step S6 is: first, the demodulator adjusts the received signal according to the channel response using the restored high-resolution channel response to reduce the interference and distortion caused by the time-varying characteristics and multipath effect. Then, the compensated received signal is mapped back to the original data symbol to eliminate the noise and distortion introduced by the channel, and finally the accurate original data symbol is restored to achieve channel estimation and data recovery in a complex channel environment. The OTFS demodulation process is expressed by the following formula: Where x[n,m] is the symbol in the time-frequency domain, n is the time index, m is the frequency index, M is the number of samples of Doppler shift, N is the number of samples of delay, and X T，V [k,l] is a symbol in the delay-Doppler domain, k is the delay index, and l is the Doppler shift index.