A method for constructing an intelligent wireless propagation environment with adaptive capabilities

Abstract

The present invention discloses a method for constructing an intelligent wireless propagation environment with adaptive capability, including: according to the current scene, a perception module performs multimodal data perception, and performs data enhancement on a single modal data sample, respectively inputs it into a ResNet neural network, and inputs S 3D tensors of different modalities into a transformer-XL fusion model to obtain a feature sequence after multimodal data fusion; and obtains control information of the wireless propagation environment after edge calculation of the feature sequence, and transmits the control information to RIS, IAB, and UAV-RIS relay devices respectively; when the scene changes, the feature representation of the new multimodal data is input into a transformer-XL2 fusion model for feature fusion, and a new feature sequence is obtained after fusion, and the control information of the wireless propagation environment is obtained after edge calculation of the new feature sequence, and is transmitted to RIS, IAB, and UAV-RIS relay devices respectively. The present invention can quickly perceive the wireless propagation environment and realize the real-time reconfiguration of the wireless propagation environment.

Description

Technical Field

[0001] The invention belongs to the field of communication technology, and in particular relates to a method for constructing an intelligent wireless propagation environment with self-adaptation capability. Background Art

[0002] As global digitalization grows at an exponential rate, the current mobile communication network is insufficient to meet the application requirements of multi-dimensional perception, ubiquitous intelligent connection of people, machines and objects, ultra-high speed, ultra-low latency, ultra-high density wireless connection and other application requirements brought about by the development of technologies such as metaverse, web3.0 and adaptive AI. As the next generation of mobile communication technology, 6G wireless network needs to fundamentally change the way of network construction and deployment, and promote mankind into a new era of deep integration of virtual and reality. In recent years, user-centric intelligent network is considered to be one of the disruptive technologies in the 6G era. This network is a flexible and reconfigurable three-dimensional network architecture. From the three dimensions of "network simplicity", "network autonomy" and "intelligent endogenous", it creates "intelligent ubiquitous connectivity" through technologies such as control base station and data base station decoupling, elastic RAN, minimalist protocol stack, semantic communication, endogenous AI network architecture, and digital twin network architecture. However, the highly dynamic wireless environment (including spatiotemporal attributes such as user location and mobility) and dynamic multi-service requirements pose great challenges to the intelligent wireless access network to achieve the vision of control user separation, multi-air interface fusion, and plug-and-play wireless connection.

[0003] The impact of uncontrollable wireless environment includes signal attenuation limiting the transmission distance, multipath propagation causing signal fading, reflection / refraction of a large number of objects causing uncontrollable interference, etc. Therefore, the inability to actively control according to actual needs to eliminate the negative impact of the ever-changing wireless environment on the signal is the main factor hindering the development of current wireless networks.

[0004] In recent years, the concept of Smart Radio Environments (SRE) with the capabilities of "programmable wireless channels and reconfigurable wireless environments" has gradually become a research hotspot for wireless communications, especially high-frequency communications such as millimeter wave / THz. It uses reconfigurable intelligent surfaces (RIS) to turn the wireless environment itself into an intelligent and reconfigurable entity. By dynamically controlling the propagation of radio waves, it is expected to achieve a paradigm shift in wireless connections from "changing oneself to adapt to the environment" to "changing the environment to adapt oneself", so as to meet the growing demand for ubiquitous connectivity, extreme throughput and high reliability in the next generation of communication systems.

[0005] The idea first involved an internal integrated controller group that can interact locally and communicate globally to achieve a given electromagnetic property. A large number of low-cost metasurfaces coated in a wireless environment and controlled by software can effectively reshape the propagation of electromagnetic waves, including wave absorption, abnormal reflection, polarization deflection and focusing, etc. Therefore, objects that are considered uncontrollable by the traditional communication paradigm become programmable objects that facilitate communication and information processing, and the wireless environment itself becomes an entity that can be programmed, configured and optimized.

[0006] At present, the research on intelligent wireless environment mainly focuses on the realization of hardware demonstration platform and point-to-point experimental testing. Even in the more in-depth research of RIS, the current research mainly focuses on the design of RIS-assisted communication system, channel estimation and system optimization, focusing on the fields of wireless resource management and allocation.

[0007] Disadvantages of existing technology:

[0008] Although the traditional method has partially improved the wireless propagation environment, it still faces the problem of difficult channel estimation. The larger the size of RIS, the more difficult it is to achieve channel estimation. It is difficult to apply to complex and changeable scenarios, ultra-low latency and ultra-high reliability scenarios such as the Industrial Internet, V2X vehicle networking and 6G deterministic networks. Summary of the invention

[0009] The technical problem to be solved by the present invention is that in the face of the future development needs of low-power, high-throughput, and low-latency communications, the traditional wireless communication mode is gradually no longer adaptable. The fundamental reason is that the wireless signal used for transmission experiences a random wireless channel brought by an uncontrollable wireless environment. However, the wireless environment used to transmit information is not controlled and optimized as a communication subject. In order to solve the above problem, a method for constructing an intelligent wireless propagation environment with adaptive capabilities is provided.

[0010] The object of the present invention is achieved in the following manner:

[0011] A method for constructing an intelligent wireless propagation environment with self-adaptation capability comprises the following steps:

[0012] S1: According to the current scenario, the perception module performs multimodal data perception and performs data enhancement on the data samples of a single modality, and inputs them into the ResNet neural network to obtain the 3D tensor H*W*C of a single modality after convolution; where H represents the image length, W represents the image width, and C represents the number of channels;

[0013] S2: Input S 3D tensors of different modalities into the transformer-XL fusion model, and use the self-attention mechanism to perform multi-scale fusion on the single-modal tensors obtained in step S1 at different feature layers to obtain the feature sequence (S*H*W)*C after multi-modal data fusion; where S represents the number of different modalities, H represents the image length, W represents the image width, and C represents the number of channels;

[0014] The multimodal data is trained by the transformer-XL fusion model to generate corresponding feature experience, which is stored in the feature experience library. The experience feature refers to the weight set generated after the multimodal data is trained by the fusion model. The weight will be fine-tuned during the training process as the scene changes, and the weights of different scenes are integrated to form feature experience. The experience library will initially have some fusion features and their relationship vectors with the control end.

[0015] S3: edge computing is performed on the feature sequence after fusion of multimodal data to obtain control information of the wireless propagation environment, and the control information is transmitted to RIS, IAB, and UAV-RIS relay devices respectively, and the RIS, IAB, and UAV-RIS relay devices make corresponding changes according to the control information; the edge computing includes inputting the feature sequence after fusion of multimodal data into the LSTM neural network through the central processing unit, and inputting the relationship vector in the experience library into the LSTM neural network at the same time. The LSTM neural network embodies physical space information, which can establish a connection with the control information, and then obtains the relationship between the features and the relationship between the features and the control information, and outputs the control information of the wireless propagation environment, the information includes the position of the drone carrying RIS at the next moment, the phase of each unit module of RIS, and the node information of IAB;

[0016] S4: When the scene changes, the perception module of the scene before the change is used to perceive the new multimodal information, and the transformer-XL model is fine-tuned using transfer learning based on the feature experience library to obtain a new fusion model transformer-XL2;

[0017] S5: The feature representation of the new multimodal data is input into the transformer-XL2 fusion model for feature fusion. After fusion, a new feature sequence (S 1 *H*W)*C, and sent to step S3 for edge computing again. After edge computing, the control information of the new wireless propagation environment is transmitted to RIS, IAB, and UAV-RIS relay devices respectively.

[0018] The perception module includes millimeter wave radar, visual sensor, laser radar and GPS sensor.

[0019] The data enhancement in S1 includes brightness, gamma correction and hue.

[0020] Beneficial effects of the present invention: Compared with the prior art, the present invention can quickly perceive the wireless propagation environment, realize real-time reconfiguration of the wireless propagation environment, and realize migration evolution on this basis, so as to realize the adaptation of the wireless propagation environment in a low-cost, low-power, flexible and adaptive manner, thereby providing ultra-low latency, ultra-high reliability and high-throughput wireless communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the present invention.

[0022] Figure 2 It is a structural diagram of the LSTM neural network of the present invention. DETAILED DESCRIPTION

[0023] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0024] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same technical meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0025] like Figure 1 As shown, a method for constructing an intelligent wireless propagation environment with self-adaptation capability comprises the following steps:

[0026] S1: According to the current scenario, the perception module performs multimodal data perception and performs data enhancement on the data samples of a single modality, and inputs them into the ResNet neural network to obtain the 3D tensor H*W*C of a single modality after convolution; where H represents the image length, W represents the image width, and C represents the number of channels;

[0027] S2: Input S 3D tensors of different modalities into the transformer-XL fusion model, and use the self-attention mechanism to perform multi-scale fusion on the single-modal tensors obtained in step S1 at different feature layers to obtain the feature sequence (S*H*W)*C after multi-modal data fusion; where S represents the number of different modalities, H represents the image length, W represents the image width, and C represents the number of channels;

[0028] The multimodal data is trained by the transformer-XL fusion model to generate corresponding feature experience, which is stored in the feature experience library. The experience library will initially have fusion features and their relationship vectors with RIS, IAB, and UAV-RIS relay devices. The experience feature refers to the weight set generated after the multimodal data is trained by the fusion model. The weight will be fine-tuned during the training process as the scene changes, and the weights of different scenes will be integrated to form feature experience.

[0029] S3: edge computing is performed on the feature sequence after fusion of multimodal data to obtain control information of the wireless propagation environment, and the control information is transmitted to RIS, IAB, and UAV-RIS relay devices respectively, and the RIS, IAB, and UAV-RIS relay devices make corresponding changes according to the control information; the edge computing includes inputting the feature sequence after fusion of multimodal data into the LSTM neural network through the central processor, and inputting the relationship vector in the experience library into the LSTM neural network. The LSTM neural network embodies physical space information and can establish a connection with the control information (for example, knowing the location and movement trajectory of the terminal, the base station beam can be controlled for tracking), thereby obtaining the relationship between the features and the relationship between the features and the control information, and outputting the control information of the wireless propagation environment, the information includes the next moment position of the drone carrying RIS, the phase of each unit module of RIS, and the node information of IAB;

[0030] S4: When the scene changes, the perception module of the scene before the change is used to perceive the new multimodal information, and the transformer-XL model is fine-tuned using transfer learning based on the feature experience library to obtain a new fusion model transformer-XL2;

[0031] S5: The feature representation of the new multimodal data is input into the transformer-XL2 fusion model for feature fusion. After fusion, a new feature sequence (S 1 *H*W)*C, where S 1 The modal number of the new feature sequence is sent to step S3 for edge calculation again. After the edge calculation, the control information of the new wireless propagation environment is transmitted to the RIS, IAB, and UAV-RIS relay devices respectively. The RIS, IAB, and UAV-RIS relay devices make corresponding changes according to the control information.

[0032] The feature experience is applied to the transformer-XL fusion model to achieve fusion feature evolution. At the same time, the new feature experience generated is used to update and expand the feature experience to the feature experience library;

[0033] The perception module includes millimeter wave radar, visual sensor, laser radar and GPS sensor.

[0034] The data enhancement in S1 includes brightness, gamma correction and hue.

[0035] Furthermore, the structure diagram of the LSTM neural network is as follows: Figure 2 As shown, where x t is the input, h t The specific working steps of LSTM neural network: LSTM is composed of three gate units: forget gate, input gate and output gate. Since LSTM is a time-series neural network (simply understood as the current moment, the previous moment and the next moment), it is impossible for it to retain all information. It stores the required information and forgets the unnecessary information.

[0036] The input gate determines how much of the network's input data needs to be saved to the unit state at the current moment.

[0037] The forget gate determines how much of the cell state at the previous moment needs to be retained to the current moment.

[0038] The output gate controls how much of the current cell state needs to be output to the current output value (i.e. the next moment);

[0039] The fused features are input into the LSTM neural network, and this process is used for continuous iteration and updating; the fused feature vector and its relationship vector with the control end device are input into the LSTM neural network to obtain the hidden state vector of each data feature, and its corresponding matrix is ​​obtained according to the fused data feature vector and its relationship with the control end device, and its weight is continuously adjusted through the regression loss function until the loss is minimized, and the optimal solution of the relationship between the fused features and the control end (the control end includes RIS, IAB, and UAV-RIS relay device) is obtained, and the control information of the wireless propagation environment is output.

[0040] The present invention can quickly perceive the wireless propagation environment and realize real-time reconfiguration of the wireless propagation environment, and on this basis realize migration evolution, so as to realize the adaptation of the wireless propagation environment in a low-cost, low-power, flexible and adaptive manner, thereby providing wireless communication with ultra-low latency, ultra-high reliability and high throughput.

[0041] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several changes and improvements can be made without departing from the overall concept of the present invention, which should also be regarded as the scope of protection of the present invention.

Claims

1. A method for constructing an intelligent wireless propagation environment with adaptive capabilities, characterized in that: The steps include: S1: According to the current scenario, the perception module performs multimodal data perception and performs data enhancement on the data samples of a single modality, and inputs them into the ResNet neural network to obtain the 3D tensor H*W*C of a single modality after convolution; where H represents the image length, W represents the image width, and C represents the number of channels; S2: Input S 3D tensors of different modalities into the transformer-XL fusion model, and use the self-attention mechanism to perform multi-scale fusion on the single-modal tensors obtained in step S1 at different feature layers to obtain the feature sequence (S*H*W)*C after multi-modal data fusion; where S represents the number of different modalities, H represents the image length, W represents the image width, and C represents the number of channels; The multimodal data is trained by the transformer-XL fusion model to generate corresponding feature experience, which is stored in the feature experience library. The feature experience refers to the weight set generated after the multimodal data is trained by the fusion model. The weight will be fine-tuned during the training process as the scene changes, and the weights of different scenes are integrated to form feature experience. The experience database initially has some fusion features and their relationship vectors with the control end; S3: edge computing is performed on the feature sequence after fusion of multimodal data to obtain control information of the wireless propagation environment, and the control information is transmitted to the RIS, IAB, and UAV-RIS relay devices respectively, and the RIS, IAB, and UAV-RIS relay devices make corresponding changes according to the control information; The edge computing includes inputting the feature sequence after multimodal data fusion into the LSTM neural network through the central processor, and inputting the relationship vector in the experience library into the LSTM neural network. The LSTM neural network reflects the physical space information, which can be connected with the control information, and then obtain the relationship between the features and the relationship between the features and the control information, and output the control information of the wireless propagation environment, which includes the position of the drone carrying RIS at the next moment, the phase of each unit module of RIS, and the node information of IAB; S4: When the scene changes, the perception module of the scene before the change is used to perceive the new multimodal information, and the transformer-XL model is fine-tuned using transfer learning based on the feature experience library to obtain a new fusion model transformer-XL2; S5: The feature representation of the new multimodal data is input into the transformer-XL2 fusion model for feature fusion. After fusion, a new feature sequence (S 1 *H*W)*C, and sent to step S3 for edge computing again. After edge computing, the control information of the new wireless propagation environment is transmitted to RIS, IAB, and UAV-RIS relay devices respectively.

2. The method for constructing an intelligent wireless communication environment with self-adaptive capability according to claim 1, characterized in that: The perception module includes millimeter wave radar, visual sensor, laser radar and GPS sensor.

3. The method for constructing an intelligent wireless propagation environment with self-adaptation capability according to claim 1, characterized in that: The data enhancement in S1 includes brightness, gamma correction and hue.

Images