A method for predicting a multi-intelligent reflecting surface channel based on transfer learning
By constructing a physical prior convolutional neural network architecture and combining transfer learning and layer-by-layer unfreezing strategies, the problems of insufficient generalization ability and large data requirements of deep learning models in multi-intelligent reflective surface environments are solved, achieving efficient channel prediction and improved robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHENGDU TECH UNIV
- Filing Date
- 2025-09-28
- Publication Date
- 2026-08-04
AI Technical Summary
Existing technologies lack the generalization ability of deep learning models in multi-intelligent reflective surface environments, have high target domain data requirements, low prediction accuracy, and fail to fully utilize the physical characteristics of the channel.
A physical prior convolutional neural network architecture is constructed using a transfer learning-based approach. Channel features are extracted through frequency-domain selective convolutional filters and phase alignment layers. Combined with weight regularization and layer-by-layer unfreezing strategies, channel prediction across environments is achieved.
It significantly reduces the data requirements of the target domain, improves prediction accuracy and robustness, adapts to the differences in multiple IRS environments, and maintains efficient operation in complex communication scenarios.
Smart Images
Figure CN121308883B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication technology, specifically, it relates to a multi-intelligent reflective surface channel prediction method based on transfer learning. Background Technology
[0002] With the rapid development of 5G and future 6G communication systems, intelligent reflective surfaces (IRS), as an emerging technology capable of programmable control of electromagnetic wave propagation, have attracted widespread attention. By deploying intelligent reflective surfaces in the environment and flexibly adjusting their reflection coefficients, the coverage, spectral efficiency, and energy efficiency of the system can be significantly improved, making it particularly suitable for scenarios with dense urban areas, congested links, and users at the cell edge.
[0003] Currently, scholars both domestically and internationally have conducted extensive research on channel prediction and estimation for single intelligent reflective surface systems, proposing various deep learning-based methods, such as super-resolution neural networks, compressed sensing combined with residual networks, and LSTM time-correlation modeling. These methods can effectively capture channel characteristics and improve prediction accuracy in single intelligent reflective surface scenarios. However, in multi-intelligent reflective surface systems, the propagation environment differs significantly between each intelligent reflective surface and the user, leading to strong environmental dependence and imbalance in the channel. When traditional single intelligent reflective surface learning models are directly transferred to new intelligent reflective surface scenarios, they often require a large amount of new training data, exhibiting insufficient generalization ability, and significantly increasing training and pilot overhead.
[0004] On the other hand, existing deep learning models are mostly "black box" structures, lacking utilization of the physical laws of communication. For example, in frequency division duplex systems, the phase difference caused by frequency offset in the uplink and downlink channels is difficult to be fully captured by traditional pure data-driven neural networks, thus affecting the model's prediction performance across frequency bands and environments. Existing technologies mainly suffer from the following problems: First, the model's generalization ability is insufficient. In multi-intelligent reflective surface scenarios, deep learning models usually need to be retrained in each new environment, which is costly. Second, when the data of the target intelligent reflective surface subsystem is limited, the prediction accuracy of models trained by traditional methods drops significantly. Finally, existing convolutional neural networks or fully connected neural networks fail to fully integrate the physical characteristics of intelligent reflective surface channels, resulting in a disconnect between the model structure and the channel propagation mechanism, leading to insufficient learning efficiency and robustness.
[0005] Therefore, there is an urgent need for a multi-intelligent reflective surface channel prediction method that can achieve efficient migration and combine physical prior knowledge under the condition of a small amount of target domain data, so as to solve the problems of high data requirements, low migration efficiency and insufficient prediction accuracy in the existing technology. Summary of the Invention
[0006] The purpose of this invention is to provide a multi-intelligent reflective surface (MIS) channel prediction method based on transfer learning, which mainly addresses the shortcomings of existing deep learning models in multi-IRS environments, such as poor generalization, large target domain data requirements, and low prediction accuracy.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A multi-intelligent reflector surface channel prediction method based on transfer learning includes the following steps:
[0009] S1, Construct a multi-IRS system based on frequency division duplex;
[0010] S2, establish channel feature representations as the input and output of the neural network, and construct the training dataset for the neural network;
[0011] S3, establish a physical prior convolutional neural network architecture that embeds physical characteristics;
[0012] S4 utilizes the source domain dataset to pre-train the physical prior convolutional neural network architecture to learn common feature representations across environments;
[0013] S5, target domain adaptation by direct transfer learning of physical prior convolutional neural network architecture on the target domain subsystem;
[0014] S6 performs channel prediction and performance evaluation on the physical prior convolutional neural network architecture after transfer learning.
[0015] Further, in step S1, the multi-IRS system includes K distributed access point-smart reflector pairs, each pair covering an independent service area, which is further divided into 5 non-overlapping sub-areas, thus forming 5K IRS auxiliary subsystems; each IRS auxiliary subsystem has APs equipped with M antennas, each IRS auxiliary subsystem contains N reflector elements, serves I single-antenna users, and defines an uplink cascaded channel. Downlink concatenated channel Where H represents the cascaded channel matrix, U represents the uplink channel, D represents the downlink channel, i represents the user index, and k represents the subsystem index.
[0016] Furthermore, in step S2, the channel feature representation is as follows: the uplink and downlink channels in the complex domain are expanded into real vectors with concatenated real and imaginary parts, respectively.
[0017] Furthermore, in step S2, the training dataset is processed by applying the source domain K... S The data from each IRS auxiliary subsystem is aggregated to obtain the target domain K.T The data from each IRS auxiliary subsystem are divided to obtain the fitness set and the test set.
[0018] Furthermore, the physical prior convolutional neural network architecture includes:
[0019] Frequency-domain selective convolution filter: It adopts a grouped convolution method to independently convolve the channel of each antenna in the frequency domain, which approximately realizes the time delay domain filtering, thereby extracting the multipath structure features in the IRS channel;
[0020] Convolutional feature extraction layer: includes multiple two-dimensional convolutional layers, batch normalization layers and ReLU activation layers, used to learn the spatial-frequency correlation features of the uplink channel layer by layer;
[0021] Phase alignment layer: For the deterministic phase rotation caused by the uplink and downlink carrier frequency difference in the frequency division duplex system, the FFT → phase rotation → IFFT processing method is used to compensate for it, so that the output characteristics are consistent with the physical channel characteristics.
[0022] Output layer: The features are projected onto the two output channels of the convolution kernel using 1×1 convolution, which represent the real and imaginary parts of the downlink channel, respectively.
[0023] Furthermore, in step S4, during the pre-training of the convolutional neural network architecture, the optimization objective adopts the mean squared error loss function and is combined with a weight regularization term to ensure network convergence stability and obtain the pre-trained parameter set.
[0024] Furthermore, the specific process of step S5 is as follows:
[0025] Initial stage: Freeze all convolutional feature extraction layers and only unfreeze the output layer to prevent the model from forgetting source domain knowledge in the early adaptation stage;
[0026] Intermediate stage: After completing 1 / 3 of the adaptation steps, the ReLU activation layer is unfrozen; after completing 2 / 3 of the adaptation steps, the batch normalization layer is unfrozen, thus achieving hierarchical unfreezing.
[0027] Termination phase: After completing all adaptation steps, the target domain-specific parameters are obtained and validated on the test set.
[0028] Furthermore, in this invention, the channel prediction method is as follows: input the uplink channel of the target domain into the model after transfer learning, and output the downlink concatenated channel prediction value.
[0029] Furthermore, in this invention, the performance evaluation method is as follows: calculate the normalized mean square error between the prediction result and the real channel, and evaluate the final result by the average NMSE over all target tasks.
[0030] Compared with the prior art, the present invention has the following beneficial effects:
[0031] (1) This invention significantly reduces the target domain data requirement. Traditional No-transfer models require almost 100% target domain data to achieve high prediction accuracy, while this invention only requires 30% target domain data to achieve comparable or even better performance, effectively reducing pilot overhead and training costs. Secondly, the channel prediction accuracy is greatly improved. Under the same experimental conditions, the NMSE of downlink channel prediction is reduced by an average of 2–4 dB, and the prediction accuracy advantage is significant compared to models such as Direct-test (no adaptation) and ordinary CNN (no physical prior). At the same time, this method can quickly adapt to the differences in multiple IRS environments. Through a layer-by-layer unfreezing transfer strategy, the model can be transferred from the source domain to the target domain within a limited number of iterations, effectively avoiding catastrophic forgetting and ensuring stable operation on 20 target IRS subsystems.
[0032] (2) This invention exhibits excellent robustness: as the number of AP antennas (MM) or IRS reflector units (NN) increases, it still maintains a low prediction error. Compared with baseline models such as No-transfer and CNN, it performs better in large-scale high-dimensional channel scenarios and has strong scalability. Furthermore, it demonstrates outstanding adaptability to frequency offsets, maintaining a low prediction error even with uplink / downlink frequency spacing δ. f In experiments where the frequency was increased from 50MHz to 800MHz, the performance degradation was much smaller than that of the control model, demonstrating that the phase alignment layer can effectively compensate for the non-reciprocity of the FDD system.
[0033] (3) The contribution of the physical prior structure of this invention is also significant. Ablation experiments show that removing the frequency domain filter or phase alignment layer leads to a decrease in prediction performance. When the physical prior is completely removed, the model performance drops to -6.5dB, which is much lower than the -16.2dB of the complete model, fully verifying the necessity of the structural design. Finally, this method has high engineering application value, can work efficiently in 6G multi-IRS deployment scenarios, and is suitable for complex environments such as urban blocks and large-scale millimeter-wave communication. It balances low overhead, strong generalization and high accuracy, and has significant prospects for industrial application. Attached Figure Description
[0034] Figure 1 This is a flowchart of the method of the present invention.
[0035] Figure 2 This is a schematic diagram of a simulation scenario according to an embodiment of the present invention. Detailed Implementation
[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.
[0037] Example
[0038] like Figure 1 , 2 As shown, this invention discloses a multi-intelligent reflector surface channel prediction method based on transfer learning.
[0039] First, a multi-IRS system model is constructed, defining the system as a frequency division duplex multiple-input multiple-output orthogonal frequency division multiplexing architecture. This system consists of multiple distributed access points and intelligent reflective surfaces. Each pair of AP-IRS covers an independent service area, and each service area is divided into five non-overlapping sub-areas, forming multiple IRS auxiliary subsystems. Here, the IRS auxiliary subsystem refers to each IRS and its link's signal transmitter (access point) and signal receiver (user terminal), meaning each pair of AP-IRS covers an independent service area. Figure 2 In this context, "subsystem" refers to a subsystem consisting of each IRS and its signal transmitters (access points) and receivers (user terminals) on its links; "AP" represents signal access points; "Users" represents user terminals; "scatters" represents scatterers in the communication environment; and "IRS" represents intelligent reflective surfaces. Figure 2 The communication scenario depicted is that multiple IRSs are randomly distributed on both sides of two intersecting roads in a city. The access point (AP) receives signals from the base station and then transmits the signals to the user terminal via IRS reflection. The IRS bypasses obstacles to increase the signal coverage.
[0040] Specifically, the experimental environment in this embodiment deploys 18 IRSs, each paired with an access point, forming 18 Access Point-Intelligent Reflecting Surface (AP-IRS) pairs, further divided into 90 IRS-user subsystems. Each subsystem's AP is equipped with 4 transmit antennas, and the IRS contains 144 tunable reflective elements, serving 10 single-antenna users. Uplink and downlink concatenated channels are defined as the concatenated forms of AP-IRS channels and IRS-user channels, respectively. In practical applications, it is assumed that the direct link between the user and the AP is blocked, and the user communicates with the AP only through the IRS, simulating complex urban blocks or large-scale millimeter-wave communication scenarios.
[0041] Then, channel feature representation and dataset construction are performed. The uplink and downlink channels in the complex domain are expanded into real vectors with concatenated real and imaginary parts, respectively, which serve as the input and output of the neural network. The uplink channel... By AP–IRS channel With IRS – User Channel Cascaded to obtain; downlink channel By AP–IRS channel With IRS – User Channel The cascaded structure achieves consistent propagation path delay and gain between the uplink and downlink channels, but exhibits a frequency offset δ. f .
[0042] For the data from 70 subsystems in the source domain, a training dataset was constructed by aggregating the data to learn common features across environments. For the data from 20 subsystems in the target domain, the data was divided into an adaptation set and a test set, used for model fine-tuning and performance validation, respectively. In practice, the dataset was obtained based on the DeepMIMO open database O1 scenario, simulating a 60GHz millimeter-wave communication system, with users and scatterers distributed on both sides of two intersecting streets to ensure data diversity.
[0043] The core of this invention is the design of a Physics-aware CNN. The Physics-aware CNN comprises multiple structural modules. The frequency-selective convolutional filter employs a grouped convolution approach, independently convolving the channel of each antenna in the frequency domain to approximate delay-domain filtering, thereby extracting multipath structural features from the IRS channel. The convolutional feature extraction layer contains multiple layers of two-dimensional convolution operations, with each convolution layer followed by a batch normalization layer and a ReLU activation function, learning the spatial-frequency correlation features of the uplink channel layer by layer. The phase alignment layer addresses the deterministic phase rotation problem caused by the uplink and downlink carrier frequency difference in FDD systems, using Fast Fourier Transform, phase rotation compensation, and Inverse Fourier Transform for phase correction, ensuring that the output features align with the physical channel characteristics. The output layer uses a 1×1 convolution to project the preceding features onto two channels, representing the real and imaginary parts of the downlink channel, respectively. The design of these modules, by introducing physical priors, significantly reduces the learning difficulty of the model and improves prediction accuracy.
[0044] In the source domain pre-training phase, the Physics-aware CNN is trained on the source domain dataset to learn common feature representations across environments. The optimization objective employs the mean squared error loss function, combined with a weight regularization term to ensure network convergence stability, ultimately yielding the pre-trained parameter set. Specific training parameter settings are as follows: the learning rate is 1×10⁻⁶. -4 The batch size is 128, the training epochs are 100, the optimizer is Adam, the momentum parameters are (0.9, 0.999), and the regularization uses a weight decay coefficient λ1 = 10. -3 The goal of this stage is to make the model more generalizable so that it can be quickly transferred to the target domain later. This is achieved using a dataset D containing 70 source domain subsystems. Tr (S) Train a Physics-aware CNN on the CNN. After training, the parameter set is obtained.
[0045] In the target domain adaptation phase, a layer-by-layer unfreezing fine-tuning strategy is adopted to prevent the model from forgetting source domain knowledge in the early stages of adaptation. In the initial stage, all convolutional feature extraction layers are frozen, and only the output layer parameters are updated. In the intermediate stage, the ReLU activation layers are unfrozen after completing 1 / 3 of the adaptation steps. In the final stage, the batch normalization layers are unfrozen after completing 2 / 3 of the adaptation steps. This strategy achieves a smooth transition from the source domain to the target domain by introducing target domain data information in stages, while preventing catastrophic forgetting. Specific training parameters are set as follows: learning rate of 1×10⁻³, batch size of 64, adaptation steps of 15, regularization using a weight decay coefficient λ₁ = 10⁻³, and loss function is mean squared error.
[0046] In the channel prediction and performance evaluation stage, the uplink channel of the target domain is input into the finely tuned model, and the predicted value of the downlink concatenated channel is output. The normalized mean square error (NMSE) between the prediction result and the actual channel is calculated. The final result is quantified and evaluated by the average NMSE across all target tasks. Experimental results show that when only 30% of the data is available in the target domain, the method of this invention can achieve prediction accuracy comparable to that of the No-transfer method using nearly 100% of the data. Compared with the Direct-test model, the performance improvement is significant, verifying the necessity of target domain adaptation. The Physics-aware CNN outperforms the ordinary CNN throughout the process, proving that physical prior design can improve generalization ability. In addition, when the number of AP antennas $M$ increases from 4 to 32, the prediction accuracy of the method of this invention is consistently better than the baseline model. When the number of IRS reflection units N$ increases from 40 to 144, the method of this invention still maintains a low NMSE in complex large-scale IRS scenarios, while the performance of the CNN baseline model decreases significantly. As the number of AP antennas or IRS reflective units gradually increases, the present invention can still maintain a low NMSE, while the performance of ordinary CNN baseline models drops significantly, demonstrating the robustness and scalability of the present invention in complex communication environments.
[0047] When the uplink / downlink frequency interval δ f When the frequency range increases from 50MHz to 800MHz, the method of this invention is almost unaffected at small frequency differences (100–400MHz); at large frequency differences (>400MHz), although the accuracy decreases, the performance is still significantly better than the control method; the phase alignment layer effectively compensates for the phase rotation caused by frequency offset, so that the model remains robust at different frequency intervals.
[0048] Ablation experiments further validated the importance of the physical prior structure. The complete model's NMSE was approximately -16.2 dB; removing the phase alignment layer resulted in a performance degradation of approximately 0.8 dB; removing the frequency domain filter resulted in a performance degradation of approximately 1.6 dB; and completely removing the physical prior structure reduced the model's performance to -6.5 dB, significantly lower than the complete model's -16.2 dB. This fully demonstrates the necessity and effectiveness of the physical prior design. In terms of engineering applications, this invention can operate efficiently in 6G multi-IRS deployment scenarios, is suitable for complex environments such as urban blocks and large-scale millimeter-wave communications, and has significant industrial application prospects.
[0049] In summary, this invention, by combining source domain pre-training with target domain stepwise fine-tuning and embedding physical characteristic constraints, achieves cross-environment knowledge transfer, significantly reduces the amount of target domain data required, and improves the accuracy and robustness of channel prediction. In practical applications, this invention is suitable for complex communication scenarios and has good scalability and engineering application value.
[0050] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.
Claims
1. A multi-intelligent reflective surface channel prediction method based on transfer learning, characterized in that, Includes the following steps: S1, Construct a multi-IRS system based on frequency division duplex; S2, establish channel feature representations as the input and output of the neural network, and construct the training dataset for the neural network; S3, establish a physical prior convolutional neural network architecture embedding physical characteristics; wherein, the physical prior convolutional neural network architecture includes: Frequency-domain selective convolutional filter: It adopts a grouped convolution method to independently convolve the channel of each antenna in the frequency domain to achieve time delay domain filtering, thereby extracting the multipath structure features in the IRS channel; Convolutional feature extraction layer: includes multiple two-dimensional convolutional layers, batch normalization layers and ReLU activation layers, used to learn the spatial-frequency correlation features of the uplink channel layer by layer; Phase alignment layer: For the deterministic phase rotation caused by the uplink and downlink carrier frequency difference in the frequency division duplex system, the FFT → phase rotation → IFFT processing method is used to compensate for it, so that the output characteristics are consistent with the physical channel characteristics. Output layer: The features are projected onto the two output channels of the convolution kernel using 1×1 convolution, which represent the real and imaginary parts of the downlink channel, respectively; S4 utilizes the source domain dataset to pre-train the physical prior convolutional neural network architecture to learn common feature representations across environments; S5, target domain adaptation by direct transfer learning of physical prior convolutional neural network architecture on the target domain subsystem; S6 performs channel prediction and performance evaluation on the physical prior convolutional neural network architecture after transfer learning.
2. The multi-intelligent reflector surface channel prediction method based on transfer learning according to claim 1, characterized in that, In step S1, the multi-IRS system includes K distributed access point-smart reflector pairs, each pair covering an independent service area, which is further divided into 5 non-overlapping sub-areas, thus forming 5K IRS auxiliary subsystems; each IRS auxiliary subsystem has an AP equipped with M antennas, each IRS auxiliary subsystem contains N reflector elements, serves I single-antenna users, and defines an uplink cascaded channel. Downlink concatenated channel Where H represents the cascaded channel matrix, U represents the uplink channel, D represents the downlink channel, i represents the user index, and k represents the subsystem index.
3. The multi-intelligent reflector surface channel prediction method based on transfer learning according to claim 2, characterized in that, In step S2, the channel feature representation is as follows: the uplink and downlink channels in the complex domain are expanded into real vectors with concatenated real and imaginary parts, respectively.
4. The multi-intelligent reflector surface channel prediction method based on transfer learning according to claim 3, characterized in that, In step S2, the training dataset is trained by applying the source domain K. S The data from each IRS auxiliary subsystem is aggregated to obtain the target domain K. T The data from each IRS auxiliary subsystem are divided to obtain the fitness set and the test set.
5. The multi-intelligent reflector surface channel prediction method based on transfer learning according to claim 4, characterized in that, In step S4, during the pre-training of the physical prior convolutional neural network architecture, the optimization objective adopts the mean squared error loss function and is combined with the weight regularization term to ensure network convergence stability and obtain the pre-trained parameter set.
6. The multi-intelligent reflector surface channel prediction method based on transfer learning according to claim 5, characterized in that, The specific process of step S5 is as follows: Initial stage: Freeze all convolutional feature extraction layers and only unfreeze the output layer to prevent the model from forgetting source domain knowledge in the early adaptation stage; Intermediate stage: After completing 1 / 3 of the adaptation steps, the ReLU activation layer is unfrozen; after completing 2 / 3 of the adaptation steps, the batch normalization layer is unfrozen, thus achieving hierarchical unfreezing. Termination phase: After completing all adaptation steps, the target domain-specific parameters are obtained and validated on the test set.
7. The multi-intelligent reflector surface channel prediction method based on transfer learning according to claim 6, characterized in that, The channel prediction method is as follows: input the uplink channel of the target domain into the model after transfer learning, and output the downlink concatenated channel prediction value.
8. The multi-intelligent reflector surface channel prediction method based on transfer learning according to claim 7, characterized in that, The performance evaluation method is as follows: calculate the normalized mean square error between the prediction result and the actual channel, and evaluate the final result by the average NMSE over all target tasks.