Intelligent air interface optimization method based on cooperation of large model and small model

Abstract

The invention discloses an intelligent air interface optimization method based on cooperation of a large model and a small model. The method comprises the following steps: acquiring general environment knowledge by constructing an air interface large model obtained by training on a diversified air interface data set; obtaining special environment knowledge through an air interface small model trained by an air interface data set in a specific scene; and fusing the output of the air interface large model and the output of the air interface small model to obtain a high-precision target result, thereby realizing collaborative utilization of general environment knowledge and special environment knowledge. According to the invention, the adaptation of the air interface large model to a specific environment is realized without accessing the parameters of the air interface large model, and the improvement of the intelligent air interface task performance based on the air interface large model is realized with low complexity.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to an intelligent air interface optimization method based on the collaboration of large and small models. Background Technology

[0002] Large-scale models are a potential key technology in next-generation mobile communication systems. By training deep and complex neural networks on massive amounts of data, large-scale models have achieved significant breakthroughs in fields such as natural language processing and computer vision. Large-scale models demonstrate the potential to address long-standing complex challenges in the design and optimization of communication systems, bringing about a paradigm shift in communication systems.

[0003] In intelligent air interface (IAI) tasks, although large-model-based IAI methods have achieved good performance across various tasks, their high complexity makes it difficult to quickly adapt to specific environments for further performance improvements. First, training and fine-tuning large models requires extremely high computational and energy costs. Second, in multi-user scenarios, customizing large models for each terminal device leads to inefficient inference due to small batch sizes and excessive memory overhead. Third, fine-tuning on limited datasets can easily result in catastrophic forgetting of existing knowledge. Finally, the parameters and architecture of proprietary large models are often unavailable, making fine-tuning impractical. These issues collectively limit the applicability of direct adaptive methods in real-world systems. Achieving rapid and low-cost adaptation of large models to specific environments with low overhead is a key bottleneck encountered in the application of large models in intelligent air interface. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide an intelligent air interface optimization method based on the collaboration of large and small models. Without accessing the parameters of the large air interface model, the method can adapt the large air interface model to a specific environment with low computational overhead, thereby improving the performance of the intelligent air interface.

[0005] Technical Solution: A smart air interface optimization method based on the collaboration of a large model and a small model. This method constructs a large air interface model and a small air interface model to achieve complementary collaboration between the two models on the air interface side. The large air interface model serves as a general channel knowledge base, responsible for learning wireless propagation patterns across scenarios. The small air interface model acts as an environment plugin, responsible for quickly capturing environmental features under specific conditions and dynamically correcting the output of the large air interface model through differential modeling. This allows the large air interface model to adapt to specific environments without accessing its parameters. The large and small air interface models share the same input. The output label during training is The steps include: S1, for a specific air interface task, uses a hybrid air interface dataset consisting of data from multiple environments and configurations. When the cost function and At the minimum, large air interface models are obtained through separate training. With air interface reference small model ; S2, with large air interface model Using the output as a reference, the air interface dataset is used. Air interface reference small model Perform output distribution alignment training, when the cost function Training stops when the minimum value is reached. S3 will be a small model of air interface proxy. The model parameters are set to match the air interface reference small model. The model parameters are the same; then the air interface dataset is used. Air interface proxy small model Training is performed when the cost function Training stops when the minimum value is reached. S4, large air interface model Air interface reference small model Small model of air interface proxy The outputs are fused to obtain the result of the air interface task.

[0006] Furthermore, the input Not limited to channels, received signals, pilot signals, and statistical information; the output tag It is not limited to channel reconstruction, receiving symbols, location information, and beamforming.

[0007] Furthermore, the air interface tasks are not limited to channel estimation, signal detection, and channel feedback.

[0008] Furthermore, the air interface reference mini-model and the air interface proxy mini-model have the same model structure.

[0009] Furthermore, the cost function when training a large air interface model. for: ,

[0010] Cost function when training an air reference small model for: , in, This represents the cost function used for a specific air interface task; Depending on the different air interface task requirements, the cost function is not limited to the mean squared error cost function or the cross-entropy cost function.

[0011] Furthermore, the cost function for training the air-to-ground reference mini-model. for: , in, , These are fixed hyperparameters. Let be the cross-entropy cost function.

[0012] Furthermore, the cost function for training the small air interface agent model. for: , in, These are trainable hyperparameters.

[0013] Furthermore, in step S4, the fusion method is described as follows: , in, This indicates the final output result of the air interface task.

[0014] Compared with the prior art, the significant advantages of this invention are as follows: 1. This invention achieves complementary collaboration between the large and small models on the air interface side by constructing a two-layer model structure of "basic knowledge – environment adaptation". The large air interface model, as a general channel knowledge base, is responsible for learning the wireless propagation rules across scenarios; the small air interface model, as an environment plugin, is responsible for quickly capturing environmental features under specific environments and dynamically correcting the output of the large model through differential modeling. 2. When the parameters of the large air interface model are inaccessible, this invention first aligns the large air interface model with the small air interface reference model. On this basis, the small air interface reference model and the small air interface proxy model simulate the output distribution changes brought about by the large model in the process of adapting to a specific environment, so as to achieve efficient adaptation of the large model to a specific environment and further improve the performance of intelligent air interface tasks. Attached Figure Description

[0015] Figure 1 This is an overall framework diagram of the present invention; Figure 2 This is a schematic diagram illustrating the construction of general environmental knowledge, taking a channel reconstruction task as an example. Figure 3 This is a schematic diagram illustrating the output alignment of an air interface task, taking the channel reconstruction task as an example. Figure 4 This is a schematic diagram illustrating adaptation to a specific environment, taking a channel reconstruction task as an example. Figure 5 This is a schematic diagram of the operation of an air interface task, taking the channel reconstruction task as an example. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0017] This invention designs an intelligent air interface optimization method based on the collaboration of a large model and a small model. By constructing a two-layer model structure of "basic knowledge – environment adaptation," it achieves complementary collaboration between the large and small models on the air interface side. The large air interface model, as a general channel knowledge base, is responsible for learning the wireless propagation rules across scenarios. The small air interface model, as an environment plugin, is responsible for quickly capturing environmental characteristics under specific conditions and dynamically correcting the output of the large model through differential modeling. This achieves low-complexity adaptation of the large air interface model to specific environments without accessing the parameters of the large air interface model, thus improving the performance of air interface tasks. The specific environment refers to a specific wireless propagation environment between transmit and receive, including but not limited to a specific area. In the intelligent air interface design based on the large model, when the parameters of the large air interface model are inaccessible, the method of this invention first aligns the large air interface model with the reference small air interface model. Based on this, the reference small air interface model and the proxy small air interface model simulate the output distribution changes brought about by the large model during the adaptation to specific environments, thereby achieving efficient adaptation of the large model to specific environments. This method is a general architecture for intelligent air interface tasks, such as... Figure 1 As shown, by adjusting the model structures of the large air interface model and the small air interface model (including the air interface reference small model and the air interface proxy small model), this method is applicable to any intelligent air interface task, such as channel estimation, signal detection, and channel feedback. In this method, the large air interface model, the air interface reference small model, and the air interface proxy small model can adopt any neural network architecture. The number of parameters in the large air interface model must be greater than the number of parameters in the air interface reference small model (or the air interface proxy small model); the air interface reference small model and the air interface proxy small model have the same model structure. Furthermore, the large air interface model, the air interface reference small model, and the air interface proxy small model have the same input. It is not limited to channels, received signals, pilot signals, statistical information, etc.; the output label during training is... It is not limited to channel reconstruction, receiving symbols, location information, beamforming, etc.

[0018] A smart air interface optimization method based on the collaboration of large and small models includes the following steps: Step 1, Building general environmental knowledge; In a hybrid air interface dataset composed of multiple environments and configuration data Above, a large air interface model is obtained through training. With air interface reference small model ; use The cost function is used to train the large air interface base model and the small air interface reference model, minimizing the cost function. When training the large air interface model, the cost function... The description is as follows: (1) When training an air-to-the-ground reference small model, the cost function The description is as follows: (2) in, This represents the cost function used for a specific air interface task. It can be freely chosen according to the different requirements of the air interface task, and is not limited to the mean squared error cost function, cross-entropy cost function, etc.

[0019] Step 2, align air interface task output; Large model with air interface Using the output as a reference, the air interface dataset is used. Air interface reference small model Perform output distribution alignment training; use The cost function is used to train an air-to-reference small model, minimizing the cost function. The description is as follows: (3) in, and To fix the hyperparameters, Let be the cross-entropy cost function.

[0020] Step 3, adapting to specific environments; Air interface proxy small model The model parameters are set to the air reference small model after step 2. The model parameters are then used, followed by the air interface dataset of the target environment. Air interface proxy small model Conduct training.

[0021] use The cost function is used to train the small air interface agent model, minimizing the cost function. The description is as follows: (4) in, The trainable hyperparameters are trained together with the model parameters of the air interface agent mini-model.

[0022] Step 4: Run the air interface task; Large model for air interface Air interface reference small model Small model of air interface proxy The outputs are fused to obtain the result of the air interface task.

[0023] The outputs of the large air interface model, the small air interface reference model, and the small air interface proxy model are fused. The fusion method is described below: (5) in, This indicates the final output result of a specific air interface task when the method described is used to complete that task.

[0024] Taking a channel reconstruction task as an example, consider a large-scale multiple-input multiple-output system. The base station is configured with 32 transmit antennas, and each user is configured with a single receive antenna. Orthogonal frequency division multiplexing is used, and the number of subcarriers is 32. The channel matrix to be reconstructed is... After conversion to a real matrix, the dimension is 32×32×2. The user equipment will then use the channel matrix. Through random projection matrix Compressed into low-dimensional codewords The process is described as follows: (6) in, This is represented as a vector form. When the base station receives... Afterwards, through the analysis of Performing a pseudo-inverse operation and restoring the result to matrix form yields a rough estimate of the original channel: (7) in, This represents the transformation from vector form to the original shape. This represents the pseudo-inverse operation of the matrix. Subsequently, this initial reconstruction result is further processed by a neural network to improve accuracy: (8) in, This represents a neural network model for channel reconstruction. For the finally recovered channel. When When using the method described in this invention, a large air interface model is reconstructed from the channel. Channel Reconstruction Reference Small Model Small model of channel reconstruction agent The system consists of three channel reconstruction models. The large air interface channel reconstruction model must have more parameters than the small reference channel reconstruction model (or the small proxy channel reconstruction model). The small reference and small proxy models have the same model structure. The implementation will proceed as follows: Step F1: Pre-training of the large channel reconstruction base model and the small channel reconstruction reference model; like Figure 2 As shown, in a hybrid channel dataset consisting of multiple environments and configuration data... Above, a large air interface model for channel reconstruction was obtained through training. With channel reconstruction reference small model ; use and The cost function is used to train both the large-scale air interface model and the small-scale reference model for channel reconstruction, respectively, to minimize the cost function. The cost function for training the large-scale air interface model for channel reconstruction is described as follows: , in, Let be the mean squared error cost function.

[0025] When training the channel reconstruction proxy mini-model, the cost function is described as follows: .

[0026] Step F2, reconstruct channel alignment; like Figure 3 As shown, a large-scale air interface model is reconstructed using the channel. The reconstructed channel output is used as a reference, employing the channel dataset. Small reference model for channel reconstruction The output reconstructed channel is then subjected to distribution alignment training. use The cost function is used to train a small reference model for channel reconstruction, minimizing the cost function. The cost function is described as follows: , in, and To fix the hyperparameters, Let be the cross-entropy cost function.

[0027] Step F3, Channel Environment Adaptation; like Figure 4 As shown, channel data in the target environment Above, the channel reconstruction reference small model from step (2) is used. Parameter initialization channel reconstruction proxy small model And fine-tuned the small model of the channel reconstruction agent.

[0028] use The channel reconstruction proxy mini-model is trained using the cost function to minimize the cost function, which is described as follows: , in, The trainable hyperparameters are trained together with the model parameters of the channel reconstruction agent mini-model.

[0029] Step F4: The channel reconstruction model runs; like Figure 5 As shown, the large air interface model for channel reconstruction... Channel Reconstruction Reference Small Model Small model of channel reconstruction agent The reconstructed channels are fused to obtain the optimized result. The fusion method is described as follows: .

[0030] This embodiment is only for illustrating the technical concept of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. An intelligent air interface optimization method based on the collaboration of large and small models, which achieves complementary collaboration between the large and small models on the air interface side by constructing a large air interface model and a small air interface model; wherein, The large air interface model serves as a general channel knowledge foundation, responsible for learning wireless propagation patterns across scenarios. The small air interface model acts as an environment plugin, responsible for quickly capturing environmental features under specific conditions and dynamically correcting the output of the large model through differential modeling. This allows the large air interface model to adapt to specific environments without accessing the parameters of the large air interface model. The large air interface model and the small air interface model have the same input. The output label during training is Its characteristic is that it includes the following steps: S1, for a specific air interface task, uses a hybrid air interface dataset consisting of data from multiple environments and configurations. Large air interface models were obtained through separate training. With air interface reference small model When the cost function and Stop training when the minimum value is reached; S2, with large air interface model Using the output as a reference, the air interface dataset is used. Air interface reference small model Perform output distribution alignment training, when the cost function Stop training when the minimum value is reached; S3 will be a small model of air interface proxy. The model parameters are set to match the air interface reference small model. The model parameters are the same; then the air interface dataset is used. Air interface proxy small model Training is performed when the cost function Stop training when the minimum value is reached; S4, large air interface model Air interface reference small model Small model of air interface proxy The outputs are fused to obtain the result of the air interface task.

2. The intelligent air interface optimization method based on the collaboration of large and small models according to claim 1, characterized in that, The input Includes channel, received signal, pilot signal, and statistical information; the output tag This includes channel reconstruction, receiving symbols, location information, and beamforming.

3. The intelligent air interface optimization method based on the collaboration of large and small models according to claim 1, characterized in that, The tasks include channel estimation, signal detection, and channel feedback.

4. The intelligent air interface optimization method based on the collaboration of large and small models according to claim 1, characterized in that, The air interface reference mini-model and the air interface proxy mini-model have the same model structure.

5. The intelligent air interface optimization method based on the collaboration of large and small models according to any one of claims 1-4, characterized in that, Cost function when training large air interface models for: Cost function when training an air reference small model for: in, This represents the cost function used for a specific air interface task; Depending on the different air interface task requirements, the cost function can be either the mean squared error cost function or the cross-entropy cost function.

6. The intelligent air interface optimization method based on the collaboration of large and small models according to any one of claims 1-4, characterized in that, Cost function for training an air-to-the-ground reference small model for: in, , These are fixed hyperparameters. Let be the cross-entropy cost function.

7. The intelligent air interface optimization method based on the collaboration of large and small models according to any one of claims 1-4, characterized in that, Cost function for training small air interface agent models for: in, These are trainable hyperparameters.

8. The intelligent air interface optimization method based on the collaboration of large and small models according to any one of claims 1-4, characterized in that, In step S4, the fusion method is described as follows: in, This indicates the final output of the air interface task. These are trainable hyperparameters.

Images