Large-scale MIMO-OTFS system channel estimation method based on DML technology

Abstract

The invention provides a method for carrying out channel estimation on a large-scale MIMO-OTFS (Multiple Input Multiple Output-OTFS) system by utilizing a DML (Discrete Multiplexing Language) combined training layered global DLCE (Digital Light Control Element) network, which comprises the following steps of: 1, inputting a received pilot signal into a scene classifier to obtain a channel scene index; and inputting the pilot signal to a corresponding feature extractor to extract channel features based on the predicted channel scene index. And step 3, inputting the output of the feature extractor into a feature mapper to reconstruct the whole downlink cascade channel. The HGDLCE network extracts different channel characteristics for different channel scenes and finally reconstructs the channel, so that the channel estimation precision under different channel scenes is improved. Simulation results show that the scheme can obtain better channel estimation precision than a traditional scheme.

Description

Technical Field

[0001] This invention relates to the field of wireless communication, and more specifically to a channel estimation method for a large-scale MIMO-OTFS system based on DML technology. Background Technology

[0002] Massive MIMO-OTFS technology supports parallel data transmission and simultaneous communication across multiple user frequency bands, improving spectrum utilization and the robustness of communication systems in high-speed mobile communication. Meanwhile, accurate channel estimation for massive MIMO-OTFS systems is crucial for effective information reception. However, traditional algorithms cannot adapt to changes in channel degrees of freedom and sparsity caused by switching between different scenarios and variations in the number of antennas. This invention addresses the downlink communication between a reconfigurable intelligent surface (RIS) in a massive MIMO-OTFS system and a user, using distributed machine learning (DML) technology. It trains a downlink channel estimation neural network with different training datasets from different channel scenarios. To address the robustness issue of the distributed learning scheme, a hierarchical neural network architecture based on DML is proposed, capable of constructing the entire cascaded channel through a feature mapper. Summary of the Invention

[0003] This invention addresses the problem that traditional algorithms cannot adapt to changes in channel degrees of freedom and sparsity caused by switching between different scenarios and changes in the number of antennas. It proposes a channel estimation method based on DML technology for RIS-assisted large-scale MIMO-OTFS downlink systems.

[0004] The channel estimation method for large-scale MIMO-OTFS systems based on DML technology includes the following steps:

[0005] First, let's explain the data model of this algorithm. The entire cell can be divided into R regions. Users located in the same region have similar channel characteristics, while users in different regions have different channel characteristics. The number of users in the r-th region is C. r (r = 1, 2, ..., R). Let... This indicates the channel from BS to RIS. Let R represent the channel from RIS to the k-th user in the r-th region, where k = 1, 2, ..., R.

[0006] Step 1, assuming the BS sends a signal to the user via a reflection link with the RIS, then the downlink transmission of the user-received signal is as follows:

[0007]

[0008] in, The reflection matrix at RIS is represented by Φ = diag(φ1, φ2, ..., φ). N ) is a diagonal matrix, where φ n This represents the reflection coefficient of the nth RIS unit. x represents the precoding vector at the BS end. r,k It is the signal transmitted from the BS end, n r,k This represents additive noise.

[0009] Step 2, the H represented by the traditional Saleh-Valenzuela (SV) channel model BS As shown below

[0010]

[0011] in, The complex gain of the l1-th path is given by the number of principal paths between the base station and the RIS. These represent the azimuth and elevation angles of the l1-th path to the RIS point, respectively. These represent the azimuth and elevation angles at points B and C, respectively. It can be observed that H... BS It is universal for all users, meaning that RIS is deployed after H in the cell. BS Almost unchanged.

[0012] Step 3, similarly, It can be represented as

[0013]

[0014] Among them, L r,k This represents the number of main channels between RIS and the k-th user in the r-th region. Represents the complex gain of the l2-th path. These represent the azimuth and elevation angles of the l2nd path at RIS, respectively. and These represent the normalized array steering vectors associated with BS and RIS, respectively. For a typical uniform planar array, i.e., N1×N2 (N=N1×N2), It can be represented as

[0015]

[0016] Where n1 = [0, 1, ..., N1-1] T n² = [0, 1, ..., N²-1] T d = λ / 2 represents the antenna spacing, and λ represents the carrier wavelength.

[0017] Step 4, set φ = [φ1, φ2, ..., φN Substituting into formula (1), we can obtain

[0018]

[0019] in, This represents the downlink concatenated channel corresponding to the k-th user located in the r-th region. Since the RIS lacks the ability to process signals independently, the concatenated channel H can only be estimated at the user end. r,k Instead of estimating H separately BS and

[0020] Step 5: In the above-mentioned cascaded channel estimation scheme for traditional SV channels, the RIS is deployed after the cell H. BS It remains almost unchanged, but when the user is on the move, The SV statistical model used will change, therefore the channel model of the OTFS system can be added to the SV channel model to update it. Updated as follows

[0021]

[0022] Under the condition of satisfying the biorthogonal waveform, formula (6) can be simplified to

[0023]

[0024] At this time H r,k The SV model was converted to the SV-OTFS channel model.

[0025] Step 6, in order to estimate the downlink concatenated channel H r,k The BS needs to transmit known pilot signals to the users via the RIS in Q time slots. According to formula (5), the pilot signal received by the k-th user in the r-th region in the q-th time slot (q=1,2,…,Q) is... It can be represented as

[0026]

[0027] Where, p r,k,q This represents the pilot signal transmitted by the BS, φ q Let n represent the reflection vector of the q-th time slot at RIS. r,k,q The received noise in the q-th time slot follows a complex Gaussian distribution with a mean of 0 and a variance of 0.

[0028] Step 7, assume p r,k,q =1, after transmission through Q time slots, we can obtain a pilot vector, expressed as: but It can also be written as

[0029]

[0030] in, according to Formula (9) can be expressed as

[0031]

[0032] in, h r,k =vec(H r,k ), w r,k And Θ are pre-designed as fixed values ​​for channel estimation, and like the pilot signal, Ψ r,k The information is known to both the BS and the user. Downlink cascaded channel estimation is based on this information. and Ψ r,k To estimate h r,k .

[0033] Step 8: From the expression of the cascaded channel in formula (10), it can be seen that the size of the cascaded channel is N times the size of the traditional massive MIMO channel. The DNN can be deployed on the user side for downlink cascaded channel estimation to establish a nonlinear mapping relationship from the received pilot signal to the cascaded channel, as shown below.

[0034]

[0035] in, This represents a nonlinear mapping function with weights θ.

[0036] Step 9: To train the DNN, the user needs to collect enough training data in advance. Let the training dataset be... D r,k Let represent the size of the training dataset for the k-th user located in the r-th region. The loss function can be expressed as:

[0037]

[0038] in, Representation and Input The corresponding output of the neural network, and the label. It can be obtained through traditional LS-based channel estimation schemes.

[0039] Step 10, the purpose of training the DNN is to minimize the above loss function by optimizing the weights θ, that is...

[0040]

[0041] Step 11: After determining the objective function, train the DNN on the training dataset through an iterative process. In each iteration t, the base station updates the weights θ by considering the gradient vectors uploaded by all users, as shown below.

[0042]

[0043] Where, θ t and θ t+1 These are the weights for the t-th iteration and the (t+1)-th iteration, respectively. g(θ) t ) is θ t gradient vector, η t It is the learning rate.

[0044] Step 12: The base station broadcasts the updated weights θ to all users. t+1 This is used for the next iteration. Subsequently, all users are in their own datasets. Calculate the gradient vector g r,k (θ t Simultaneously, the channel scene index is predicted by inputting the received pilot signals into a scene classifier. The user then reassembles the set of gradient vectors from all local sources. Uploaded to the base station to update the weights.

[0045] Predictive Channel Scenario Index The pilot signal is input to the corresponding feature extractor to extract channel features. The output of the feature extractor is then input to the feature mapper to reconstruct the entire downlink cascaded channel. Attached Figure Description

[0046] Figure 1 This is a model of an intelligent reflective surface-assisted wireless communication system in an example of the present invention.

[0047] Figure 2 This is a flowchart of the distributed machine learning process based on the GDLCE network in an example of the present invention.

[0048] Figure 3 This is the downlink channel estimation process based on a hierarchical GDLCE network in an example of the present invention.

[0049] Figure 4 This invention provides a comparison of the NMSE performance of different algorithms under the SV-OTFS channel full-cell model in an example of the present invention.

[0050] Figure 5 This invention presents a comparison of the NMSE performance of different algorithms in a full-cell scenario using the DeepMIMO dataset. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, it should be noted that these embodiments are not intended to limit the present invention. Equivalent transformations or substitutions in function, method, or structure made by those skilled in the art based on these embodiments are all within the protection scope of the present invention.

[0052] This invention provides a channel estimation method for large-scale MIMO-OTFS systems based on DML technology, which utilizes DML technology for collaborative training among all users in the base station and cell. Specifically, firstly, a downlink channel estimation (DLCE) network shared by the base station and all users is constructed. Then, based on the local training datasets available to all users, a global DLCE network (Global DLCE, GDLCE) is jointly trained using DML. Finally, a hierarchical GDLCE (HGDLCE) neural architecture is proposed to improve the channel estimation accuracy under different channel scenarios.

[0053] DLCE networks include the following steps:

[0054] Step 1: Deploy the DNN on the user side for downlink cascaded channel estimation to establish a nonlinear mapping relationship between the received pilot signal and the cascaded channel.

[0055] Step 2, let the training dataset be... D r,k Let represent the size of the training dataset for the k-th user located in the r-th region. Then the loss function can be expressed as: The goal of training a DNN is to minimize the aforementioned loss function by optimizing the weights, i.e.

[0056] Step 3, update the weights θ using gradient descent. In each iteration t: θ t+1 =θ t -η t g(θ t ).

[0057] Where, θ t and θ t+1 These are the weights for the t-th iteration and the (t+1)-th iteration, respectively. g(θ) t ) is θ t gradient vector, η t It's the learning rate. Users can directly estimate the channel based on the trained DNN.

[0058] The steps for GDLCE networking are as follows:

[0059] Step 1, all users in their own dataset Calculate the gradient vector g r,k (θ t ).

[0060] Step 2, collect the gradient vectors from all local users. Uploaded to the base station.

[0061] Step 3: The base station updates the weights by considering all uploaded gradient vectors, as shown below.

[0062]

[0063] Step 4: Finally, the base station broadcasts the updated weights θ to all users. t+1 This is used for the next iteration.

[0064] Based on DML training, all users will share the GDLCE network. Compared to neural networks trained on local training datasets of individual users, the GDLCE network is able to learn the overall channel characteristics of the cell.

[0065] The steps for HGDLCE networking are as follows:

[0066] Step 1: Input the received pilot signal into the scene classifier to obtain the channel scene index.

[0067] Step 2, based on the predicted channel scenario index The pilot signal is input into the corresponding feature extractor to extract channel features.

[0068] Step 3: The output of the feature extractor is fed into the feature mapper to reconstruct the entire downlink cascaded channel.

[0069] The proposed HGDLCE network is trained on different datasets for different channel scenarios based on DML technology. The scene classifier is trained separately, while the feature extractor and feature mapper are trained jointly. The channel scene index is obtained using a traditional angle estimation algorithm, and the cross-entropy between the scene classifier's output and the channel scene index label is used as the loss function to train the scene classifier.

[0070] Figure 2 This demonstrates the process by which the weights of the GDLCE network are trained in four steps during each iteration.

[0071] Figure 3The structure of the HGDLCE network is shown, and the scene classifier consists of five layers. The first and third layers are convolutional layers with 32 filters of 3×3 kernels each, using ReLU as the activation function. The second and fourth layers are max-pooling layers with 2×2 kernels. The fifth layer is a linear layer. Each user's feature extractor has the same architecture, consisting of three identical convolutional layers. Each convolutional layer consists of 32 filters of 3×3 kernels, a batch normalization operation, and a ReLU for activation. The feature mapper is a linear layer that maps the output of the feature extractor to a downlink concatenated channel vector with 2NM elements.

[0072] Figure 4 The paper presents a performance comparison of different NMSE algorithms under a full-cell model. As shown in the figure, the DLCE scheme, trained separately for the three scenarios, cannot achieve reliable channel estimation across the entire cell. The LS-based and MMSE-based schemes have better estimation accuracy in the high SNR range, but require significant pilot overhead. The two DML-based schemes achieve even better channel estimation performance.

[0073] Figure 5 This paper presents a comparison of the NMSE performance of different algorithms on the DeepMIMO dataset across the entire cell scenario. The test samples were randomly generated from three channel scenarios. Since channel scenario 2 shares some features with channel scenarios 1 and 3 of the DeepMIMO dataset, the DLCE network trained on channel scenario 2 outperforms the DLCE network trained on channel scenario 3. Overall, the HGDLCE network achieves the best channel estimation accuracy on the DeepMIMO dataset compared to other schemes.

[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A channel estimation method for a large-scale MIMO-OTFS system based on DML technology, characterized by the following steps: S1. Construct a downlink system model for RIS-assisted base station-user communication, and further derive the SV-OTFS model for downlink concatenated channels. S2. Derive the expression for the known pilot signal that the base station sends to the user via RIS in Q time slots. S3. Deploy the DNN on the user side to establish a nonlinear mapping relationship between the received pilot signal and the cascaded channel. S4. Channel estimation is performed for different types of channel scenarios for multiple users using the proposed HGDLCE network-based method.

2. A channel estimation method for a large-scale MIMO-OTFS system based on DML technology, characterized by the following steps: In S1, the specific steps are as follows: S1.1 The system model is as follows: This invention addresses the downlink communication between a Relay-Assisted Base Station (RIS) and users in a large-scale MIMO-OTFS system. It employs a base station with a uniform planar array of M antennas to communicate with multiple single-antenna users. An RIS with N reflective elements is deployed between the base station and users to enhance communication. The RIS can be controlled by the base station via a separate radio link. The entire cell can be divided into R regions. Users within the same region have similar channel characteristics, while users in different regions have different channel characteristics. The number of users in the r-th region is C. r (r = 1, 2, ..., R). Let... This indicates the channel from BS to RIS. Let R represent the channel from RIS to the k-th user in the r-th region, where k = 1, 2, ..., R. S1.

2. Assuming the BS sends a signal to the user via a reflection link with the RIS, the downlink transmission of the user-received signal is as follows: S1.3, H as represented by the traditional Saleh-Valenzuela (SV) channel model BS As shown below H BS It is universal for all users, meaning that RIS is deployed after H in the cell. BS Almost unchanged. Similarly, It can be represented as S1.

4. Set φ=[φ1,φ2,…,φ N Substituting into the formula described in S1.2, we can obtain... in, This represents the downlink concatenation channel corresponding to the k-th user located in the r-th region. S1.5 When the user is moving, The SV statistical model used will change, therefore the channel model of the OTFS system is added to the SV channel model to update it. Updated as follows Under the condition of satisfying the biorthogonal waveform, the above equation can be simplified to: At this time H r,k The SV model was converted to the SV-OTFS channel model.

3. A channel estimation method for a large-scale MIMO-OTFS system based on DML technology, characterized in that, In step S2, the specific steps for deriving the pilot signal are as follows: S2.1 Pilot signal received at the k-th user in the r-th region during the q-th time slot (q = 1, 2, ..., Q). It can be represented as S2.2, Assume p r,k,q =1, after transmission through Q time slots, we can obtain a pilot vector, expressed as: The formula described in S1.6 can be expressed as follows: in, h r,k =vec(H r,k ), w r,k And Θ are pre-designed as fixed values ​​for channel estimation, and like the pilot signal, Ψ r,k The information is known to both the BS and the user. Downlink concatenated channel estimation is based on this information. and Ψ r,k To estimate h r,k .

4. A channel estimation method for a large-scale MIMO-OTFS system based on DML technology, characterized in that, In step S3, the specific steps for deploying the DNN are as follows: S3.1 Deploying the DNN on the user side for downlink cascaded channel estimation establishes a nonlinear mapping relationship between the received pilot signal and the cascaded channel, as shown below. Among them, f θ : This represents a nonlinear mapping function with weights θ. S3.2, Let the training dataset be... D r,k Let represent the size of the training dataset for the k-th user located in the r-th region. The loss function can be expressed as: in, Representation and Input The corresponding output of the neural network, and the label. It can be obtained through traditional LS-based channel estimation schemes. S3.3 The purpose of training the DNN is to minimize the above loss function by optimizing the weights θ, that is... S3.4 After determining the objective function, the DNN is trained on the training dataset through an iterative process. In each iteration t, the base station updates the weights θ by considering the gradient vectors uploaded by all users, as shown in the following equation. Where, θ t and θ t+1 These are the weights for the t-th iteration and the (t+1)-th iteration, respectively. g(θ) t ) is θ t gradient vector, η t It is the learning rate.

5. A channel estimation method for a large-scale MIMO-OTFS system based on DML technology, characterized in that, In step S4, the specific steps of the channel estimation method based on the HGDLCE network are as follows: S4.1 Input the received pilot signal into the scene classifier to obtain the channel scene index. S4.2, Prediction-based Channel Scenario Index The pilot signal is input into the corresponding feature extractor to extract channel features. S4.3 The output of the feature extractor will be input to the feature mapper to reconstruct the entire downlink cascaded channel.

6. A wireless communication receiver simulation program product, comprising a computer program that, when executed, implements the system model of claim 2.

7. A wireless communication information acquisition algorithm program product, comprising a computer program that, when executed, implements the information acquisition algorithm of claim 5.

8. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that... When the processor executes the program, it implements the channel estimation method for a large-scale MIMO-OTFS system based on DML technology as described in any one of claims 1 to 5.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that... When the computer program is executed by the processor, it implements the channel estimation method for a large-scale MIMO-OTFS system based on DML technology as described in any one of claims 1 to 5.

Images