Environment backscattering-based integrated vehicle-road cloud communication system and method

Abstract

The present application belongs to the intelligent traffic technology field, and specifically relates to a vehicle-road-cloud integrated communication system and method based on environmental backscattering. The method comprises the following steps: step one, constructing a vehicle-road-cloud integrated simulation environment; step two, defining reinforcement learning decision elements; step three, cloud strategy training and model issuing; step four, vehicle online decision and communication simulation; and step five, strategy closed-loop updating and performance evaluation. The present application realizes a closed-loop optimization mechanism of "cloud training-road distribution-vehicle execution" by integrating an AmBC module on the vehicle side and deploying a lightweight deep Q network intelligent agent on the cloud platform. The present application allows vehicles to communicate without interference through the reflection of RSU broadcast signals when cellular users occupy the spectrum, and at the same time, the DQN strategy dynamically selects three modes of "idle / active V2X / AmBC", thereby maximizing the long-term average throughput.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent transportation, and in particular to a vehicle-road-cloud integrated communication system and method based on environmental backscattering. BACKGROUND

[0002] With the rapid development of intelligent and connected vehicles and smart transportation infrastructure, vehicle-road-cloud integrated collaboration has become a key architecture to support advanced autonomous driving and real-time traffic management. In this architecture, vehicles, roadside units, and cloud platforms exchange environmental perception, driving intent, and control instructions through V2X communication. However, current mainstream V2X technologies generally use active radio frequency transmission, which faces serious challenges in urban high-density areas. On the one hand, the number of cellular users in 5G and future mobile communication systems has increased dramatically, causing the shared spectrum to be in a long-term congestion state. If vehicles actively communicate, they will not only interfere with cellular users but also be refused access by the base station for violating spectrum coexistence rules. On the other hand, roadside units need to continuously broadcast traffic information, relying on mains power supply and dedicated spectrum, which has high deployment costs and limited coverage, making it difficult to be widely promoted in rural areas or temporary scenarios.

[0003] To alleviate the spectrum pressure, environmental backscattering communication technology has attracted attention due to its characteristics of no need for active transmission and extremely low power consumption. This technology allows devices to transmit information by reflecting existing radio frequency signals in the environment (such as base station or roadside unit broadcasts), thereby avoiding adding interference. However, existing researches mostly use it in static or low-rate Internet of Things scenarios and have not effectively integrated it into the dynamic and changing vehicle networking environment. At the same time, although some works have tried to use deep reinforcement learning to optimize communication strategies, the models used are usually complex in structure and have high computational overhead, making it difficult to run in real time on vehicle-mounted or roadside edge devices. More importantly, these methods mostly assume a stable environment and do not consider the state distribution shift problem caused by the continuous evolution of real traffic flow and wireless channels, and directly updating the model online can easily lead to performance degradation. Therefore, there is an urgent need for a vehicle-road-cloud integrated communication solution that can not only use environmental backscattering to achieve interference-free communication but also adapt to dynamic environments through lightweight intelligent decision-making and support long-term stable evolution. SUMMARY

[0004] To solve the above problems, the application provides a vehicle-road-cloud integrated communication system and method based on environmental backscattering, which integrates an environmental backscattering communication module at the vehicle end, deploys a lightweight deep Q network strategy on the cloud end, and constructs a closed-loop cooperative mechanism of "cloud training-road distribution-vehicle execution-data backflow-strategy update", so that the vehicle can intelligently select three communication modes of idle, active V2X or backscattering in a congested scenario of spectrum occupied by cellular users, and realize reliable communication with no interference, low power consumption and high throughput. The method not only uses the broadcast signal of the road side unit as the carrier source of backscattering to avoid adding radio frequency interference, but also ensures efficient operation of the system on resource-limited edge devices and adaptation to long-term evolution of the real environment through compact state representation, distillation-guided lightweight strategy network and stability-ensuring continuous learning mechanism.

[0005] The technical scheme of the application in combination with the drawings is described as follows:

[0006] In a first aspect, the application provides a vehicle-road-cloud integrated communication method based on environmental backscattering, comprising:

[0007] Step one, build a vehicle-road-cloud integrated simulation environment;

[0008] S11, build a city road scene in PreScan, deploy virtual RSUs at an interval of 200 meters, set vehicle flow, cellular user distribution and RSU safety zone radius to 200 meters;

[0009] S12, integrate a 5.9GHz V2X channel model, including path loss, shadow fading and Doppler effect, set RSU transmission power, backscattering reflection coefficient and modulation rate;

[0010] S13, realize three virtual modules: virtual RSU module, virtual OBU module and cloud training server module;

[0011] Step two, define reinforcement learning decision elements;

[0012] S21, the state is a 5-dimensional vector: spectrum occupation flag, distance to the nearest RSU, distance to the neighboring vehicle, whether the CU is in the safety zone, vehicle speed, all normalized to [0, 1];

[0013] S22, the action space contains three items: idle, active V2X and backscattering;

[0014] S23, reward = communication capacity gain - interference penalty; if active transmission when the spectrum is busy and the CU is not in the safety zone, give a strong negative reward of -100;

[0015] Step three, cloud strategy training and model distribution;

[0016] S31, first train the teacher DQN, then train the student DQN, the loss contains TD error and KL divergence;

[0017] S32, convert the student model to FP16 format, compress to about 50KB, and only when the throughput improvement is greater than or equal to 5% is marked as a deliverable version;

[0018] S33, the policy model is delivered to the local cache of all virtual RSUs for vehicles to download on demand;

[0019] Step four, vehicle online decision and communication simulation;

[0020] S41, the vehicle automatically loads the latest policy model after entering the RSU coverage area;

[0021] S42, collect state, infer action and execute communication behavior every 10ms, and record log;

[0022] S43, the RSU receives mixed signals, recovers backscattering data through serial interference cancellation and differential demodulation;

[0023] Step five, policy closed loop update and performance evaluation;

[0024] S51, every 5 minutes, the RSU aggregates vehicle logs in the jurisdiction and uploads them to the cloud;

[0025] S52, the cloud uses double buffer hybrid sampling to incrementally train the model and only updates the output layer;

[0026] S53, after the new strategy is verified to meet the requirements through multi-scenario A / B testing, a new version is generated to enter the next round of distribution, forming a closed loop optimization.

[0027] Further, the specific method of step S11 is as follows:

[0028] A typical urban road scene is built in PreScan, including crossroads, straight lanes, and elevated ramps, and virtual road side units (RSUs) are set at intervals of 200 meters; high-precision maps are imported, vehicle flow, cellular user distribution and mobile models are set, and the RSU safety zone radius is defined as 200 meters;

[0029] The specific method of step S12 is as follows:

[0030] In the PreScan combined with MATLAB / Simulink environment, a 5.9GHz V2X channel model is integrated, including path loss, shadow fading and Doppler effect; set the RSU broadcast signal parameters: center frequency 5.9GHz, bandwidth 10MHz, OFDM modulation, transmit power 30dBm; at the same time, model the backscattering link, set the vehicle reflection coefficient α=0.5, and the backscattering modulation rate is 10kbps;

[0031] Specifically, the following models are included:

[0032] 1. Path loss model;

[0033] Path loss Varies with distance The standard free-space propagation model is mixed with the urban microcell model:

[0034] ;

[0035] wherein, is the path loss at the reference distance ; is the path loss exponent; is the Euclidean distance between the vehicle and the RSU; is the shadowing random variable, which follows a zero-mean normal distribution;

[0036] 2. Received signal power model;

[0037] The received power of the RSU transmitted signal at the vehicle is:

[0038] ;

[0039] wherein, is the RSU transmit power; is the calculated path loss; is the received power of the RSU broadcast signal at the vehicle;

[0040] 3. Environmental backscatter link model;

[0041] The vehicle changes the reflection coefficient a of the RSU signal by modulating the antenna impedance to realize information coding; assuming that the reflection coefficient varies in the range a ∈ [0, 1], the equivalent transmit power of the backscatter signal is:

[0042] ;

[0043] wherein, is the set average reflection coefficient; is the backscatter power of the vehicle “virtual transmission”;

[0044] 4. Backscatter communication rate model;

[0045] The backscatter modulation rate is determined by the channel bandwidth and the modulation method, and in this simulation, it is set as:

[0046] ;

[0047] wherein, available bandwidth; backscattering link signal-to-noise ratio;

[0048] 5, Doppler shift model;

[0049] Doppler shift is:

[0050] ;

[0051] wherein, vehicle speed; angle between vehicle moving direction and signal arrival direction; speed of light, center frequency;

[0052] The specific method of the step S13 is as follows:

[0053] Virtual RSU module: periodically broadcast traffic status, used for receiving and decoding backscattering signals, and built-in 8GB cache for storing strategy model;

[0054] Virtual OBU module: mounted on each simulation vehicle, used for integrating perception, decision-making, and communication sub-modules;

[0055] Cloud training server module: running on local workstation or cluster, used for being responsible for strategy training, version management, and secure distribution.

[0056] Further, the specific method of the step S21 is as follows:

[0057] Each OBU collects the following 5-dimensional states in each 10ms simulation step: 1, spectrum occupation flag: determined by local energy detection; 2, Euclidean distance to the nearest virtual RSU; 3, distance to the nearest other simulation vehicle; 4, whether there is a CU located in the RSU safety zone; and 5, vehicle instantaneous speed;

[0058] All dimensions are normalized to [0, 1] and used as strategy network input;

[0059] The specific method of the step S22 is as follows:

[0060] The virtual OBU module selects three communication behaviors: 1, action 0, i.e., idle: no communication; 2, action 1, i.e., active V2X: sending BSM message at 20dBm power; and 3, action 2, i.e., backscattering: modulating RSU broadcast signal to send vehicle status information at a modulation rate of 10kbps;

[0061] The specific method of the step S23 is as follows:

[0062] Reward = Communication Capacity Gain - Interference Penalty; wherein, if action 1 is selected when the spectrum is occupied and the CU is not in the safe zone, a strong negative reward of -100 is imposed; the backscatter mode always gets a positive reward, encouraging energy-efficient communication.

[0063] Further, the specific method of the step S31 is as follows:

[0064] In the cloud simulation environment, a teacher DQN model is first trained, which is a deep Q network with 3 fully connected layers and 128 neurons in each layer; then a student DQN model is trained based on the knowledge distillation method, which is a simplified student network containing 2 fully connected layers and 32 neurons in each layer; the student DQN model is jointly optimized by minimizing the KL divergence between the TD error and the teacher model output, aiming to obtain a lightweight strategy network with performance close to the teacher model; double experience replay pools are used for training, i.e., each pool has a capacity of 10,000, the optimizer is Adam, the initial exploration rate ε = 1.0, and the exponential decay is to 0.01;

[0065] The specific method of the step S32 is as follows:

[0066] After training, the weight parameters of the student DQN model are converted from the original 32-bit floating-point format (FP32) to the 16-bit half-precision floating-point format (FP16), which is realized by the PyTorch standard quantization tool chain, and the model file size is compressed from about 100KB to about 50KB; then, the student DQN model is assigned a semantic version number, where the major version number represents architectural changes, the minor version number represents performance improvements, and the revision number represents training rounds; next, a closed-loop simulation is run in a set of validation scenarios independent of the training set, and key indicators are counted: average communication throughput, interference event rate, and backscatter decoding success rate; only when the average throughput of the new model in all validation scenarios is improved by no less than 5% compared to the current online version, and the number of interference events does not increase significantly (i.e., the increase is ≤2%), the state is marked as “ready to be issued”, otherwise the training result is discarded and the failure reason is recorded for subsequent analysis;

[0067] The specific method of the step S33 is as follows:

[0068] The FP16 compressed policy model marked as "issuable" is pushed to all virtual RSU nodes in the simulation environment through a secure transmission channel; each virtual RSU maintains a policy cache area in the local file system, storing multiple historical models by version number; when the simulation vehicle enters the communication coverage range of a certain RSU, that is, the received signal strength RSRP is greater than -85dBm, the virtual OBU module initiates a policy synchronization request, and the virtual RSU module responds by issuing the latest "issuable" version of the model file to the vehicle through the direct connection interface; the vehicle receives it and checks it, and loads it into the lightweight reasoning engine after confirming that it is correct, for subsequent real-time decision-making.

[0069] Further, the specific method of step S41 is as follows:

[0070] When the simulation vehicle enters the coverage range of a certain virtual RSU, that is, the received signal strength is greater than -85dBm, the virtual OBU module automatically loads the latest policy model from the virtual RSU module cache;

[0071] The specific method of step S42 is as follows:

[0072] In each 10ms simulation step, the OBU performs the following procedures:

[0073] a. Collect the current 5-dimensional state;

[0074] b. Input the lightweight DQN model and get the Q value through forward reasoning;

[0075] c. Select the action corresponding to argmax Q;

[0076] d. Update the communication behavior according to the action: if it is active V2X, generate BSM and count the channel interference; if it is backscattering, superimpose the modulation component in the RSU received signal;

[0077] e. Record the state, action, and reward, and temporarily store them locally;

[0078] The specific method of step S43 is as follows:

[0079] The virtual RSU module receives the mixed signal in each time slot, first demodulates the broadcast main signal itself, then extracts the backscattering sub-signal through serial interference cancellation, recovers the vehicle data through differential demodulation, and counts the decoding success rate.

[0080] Further, the specific method of step S51 is as follows:

[0081] Every 5 minutes of simulation time, each virtual RSU aggregates the logs of all vehicles in the jurisdiction, i.e., action sequence, throughput, interference event, and decoding result, and uploads them to the cloud server module through the simulation backhaul link;

[0082] The specific method of the step S52 is as follows:

[0083] The cloud stores the new data into a short-term buffer, mixes the new data with long-term historical data, adopts a double-buffer sampling strategy to perform incremental training on an existing DQN model, freezes the bottom-layer parameters, only updates the output layer, sets the learning rate to 0.001, and the like.

[0084] The specific method of the step S53 is as follows:

[0085] The trained new strategy is subjected to A / B testing in multiple PreScan scenarios, indexes such as average throughput, interference times, and decision delay are evaluated, a new version is generated after reaching the standard, and the new version enters a next round of distribution cycle to form a closed loop of "simulation execution-data feedback-strategy optimization".

[0086] In a second aspect, the present application further provides a vehicle-road-cloud integrated communication system based on environmental backscattering, comprising:

[0087] A system modeling and parameter initialization module is configured to construct a vehicle-road-cloud integrated simulation environment, deploy virtual RSUs, OBUs and cloud servers, and complete initial parameter configuration;

[0088] A reinforcement learning decision element definition module is configured to define a 5-dimensional state space, a 3-item action space and a composite reward function containing strong negative rewards;

[0089] A cloud platform DQN strategy training module is configured to execute a "teacher-student" knowledge distillation framework, train and compress a lightweight DQN strategy model, and perform strict performance verification;

[0090] A vehicle online decision and communication execution module is configured to realize strategy loading, real-time state perception, action decision and execution of three communication modes on the vehicle side;

[0091] A strategy closed loop updating module is configured to aggregate roadside data, perform cloud incremental training, and drive continuous iteration and optimization of the strategy through A / B testing.

[0092] The present application has the following beneficial effects:

[0093] 1. The application proposes a vehicle-road cloud cooperation method combining environmental backscatter communication and lightweight reinforcement learning, which has the primary value of realizing reliable communication without interference in a highly congested spectrum. Traditional vehicle networking relies on active transmission, which is prone to conflict with cellular users in shared frequency bands. However, the present scheme makes full use of the signals broadcast by roadside units as environmental radio sources, enabling vehicles to achieve information feedback through passive modulation, thereby fundamentally avoiding the addition of radio frequency interference. This mechanism not only aligns with the green communication philosophy of "zero net additional power" in future high-density wireless environments, but also significantly reduces the power consumption requirements of the vehicle-mounted end, providing a physical layer foundation for long-term operation on resource-constrained edge nodes.

[0094] 2. The application breaks through the inherent contradiction between model complexity and performance at the intelligent decision-making level. By introducing a knowledge distillation guided lightweight strategy training framework and combining communication field priors to structurally constrain the action space, a minimalist neural network can still effectively approximate the complex state-action mapping relationship. This method discards the inertia of relying on large model stacking and instead improves the decision-making ability of small models from the perspectives of representation efficiency and learning guidance, not only meeting the stringent latency requirements of vehicle-mounted real-time reasoning, but also providing a reproducible technical path for efficient reinforcement learning in edge intelligent systems.

[0095] 3. The application constructs a closed-loop optimization system with long-term evolution capability, effectively addressing the challenges posed by dynamic non-stationarity in real traffic environments. By designing a double-buffer experience management and strategy stability guard mechanism, the system can absorb new environmental knowledge while retaining historical effective experiences and preventing harmful updates, thereby avoiding catastrophic forgetting and strategy shock. This mechanism expands the traditional offline training paradigm into a continuous learning process of "simulation training-online execution-data feedback-safe evolution", enabling the vehicle-road cooperation system to self-improve over time, laying a methodological foundation for the robustness and sustainability of intelligent transportation systems. BRIEF DESCRIPTION OF DRAWINGS

[0096] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0097] Figure 1 The flowchart of the application;

[0098] Figure 2 The throughput evolution diagram of the lightweight student model;

[0099] Figure 3 The interference suppression capability evolution diagram;

[0100] Figure 4 Evolution graph for communication mode selection;

[0101] Figure 5 Graph for teacher DQN training curve;

[0102] Figure 6 Evolution graph for distillation training loss. DETAILED DESCRIPTION

[0103] The application will be described in further detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the application, and not to limit the application. In addition, it should be noted that, for the sake of description, only the parts related to the application are shown in the drawings, not all structures.

[0104] Embodiment one:

[0105] Referring to Figure 1 The application provides a vehicle-road-cloud integrated communication method based on environmental backscattering, and the core is to construct a closed-loop cooperative mechanism of "cloud training-road distribution-vehicle execution-data backflow-strategy update". The method comprises the following steps:

[0106] Step one, build a vehicle-road-cloud integrated simulation environment;

[0107] S11, build a city road scene in PreScan, deploy virtual RSUs at an interval of 200 meters, set vehicle flow, cellular user distribution and RSU safety zone radius to 200 meters, and the specific steps are as follows:

[0108] A typical city road scene is built in PreScan, including key nodes such as crossroads, straight lanes and elevated ramps. Virtual road side unit (RSU) positions are set at an interval of 200 meters. A high-precision map is imported, vehicle flow (such as 1200 vehicles per hour), cellular user distribution and mobile model are set, and the RSU safety zone radius is defined as 200 meters.

[0109] S12, integrate a 5.9GHz V2X channel model, including path loss, shadow fading and Doppler effect, set RSU transmission power, backscattering reflection coefficient and modulation rate, and the specific steps are as follows:

[0110] A 5.9 GHz V2X channel model was integrated into the PreScan and MATLAB / Simulink environment, incorporating path loss, shadowing fading, and Doppler effects. RSU broadcast signal parameters were set as follows: center frequency 5.9 GHz, bandwidth 10 MHz, OFDM modulation, and transmit power 30 dBm. Simultaneously, an environmental backscatter link was modeled, with vehicle reflection coefficient α = 0.5 and backscatter modulation rate set to 10 kbps.

[0111] Specifically, it includes the following models:

[0112] 1. Path loss model;

[0113] Path loss With distance The changes are modeled using a hybrid approach combining the standard free-space propagation model and the urban micro-cell model:

[0114] ;

[0115] in, For reference distance Path loss at the location, This is the path loss index. This represents the Euclidean distance between the vehicle and the RSU. Let be a shadowed fading random variable that follows a zero-mean normal distribution.

[0116] 2. Received signal power model;

[0117] The received power of the RSU transmitted signal at the vehicle is:

[0118] ;

[0119] in This refers to the RSU transmit power. The path loss is calculated using the above formula. The power of the RSU broadcast signal received by the vehicle is the basis for subsequent backscatter modulation and determines the available carrier strength.

[0120] 3. Environmental backscattering link model;

[0121] The vehicle encodes information by modulating its antenna impedance to change the reflection coefficient α of the RSU signal. Assuming the reflection coefficient varies within the range of α∈[0,1], the equivalent transmit power of the backscattered signal is:

[0122] ;

[0123] in, The set average reflectance coefficient; Backscatter power "virtually transmitted" to vehicles.

[0124] 4. Backscatter communication rate model

[0125] The backscatter modulation rate is determined by the channel bandwidth and the modulation method. In this simulation, we set:

[0126]

[0127] where is the available bandwidth; is the backscatter link signal-to-noise ratio.

[0128] 5. Doppler shift model

[0129] Due to the movement of vehicles, Doppler shift will be generated, which will affect the communication performance. The Doppler shift is:

[0130]

[0131] where is the vehicle speed, is the angle between the vehicle movement direction and the signal arrival direction; is the speed of light, is the center frequency.

[0132] S13, virtual RSU module, i.e. broadcast + decoding + cache strategy, virtual OBU module, i.e. perception + decision + communication, and cloud training server module, i.e. training + distribution + version management, as follows:

[0133] Virtual RSU module: periodically broadcast traffic status (SPAT / MAP), receive and decode backscatter signals, and built-in 8GB cache for storing strategy model;

[0134] Virtual OBU module: mounted on each simulation vehicle, integrating perception, decision, and communication sub-modules;

[0135] Cloud training server module: running on local workstation or cluster, responsible for strategy training, version management, and secure distribution.

[0136] Step two, define reinforcement learning decision elements

[0137] S21, state is a 5-dimensional vector: spectrum occupation flag, distance to the nearest RSU, distance to the neighboring vehicle, whether the CU is in the safety zone, and vehicle speed, all normalized to [0, 1], as follows:

[0138] ​​Each OBU collects the following 5-dimensional states in each 10ms simulation step: 1. Spectrum occupancy flag: determined by local energy detection; 2. Euclidean distance to the nearest virtual RSU; 3. Distance to the nearest other simulated vehicle; 4. Whether there is a CU located in the RSU safety zone; 5. Vehicle instantaneous speed.

[0139] All dimensions are normalized to [0, 1] and used as inputs to the policy network.

[0140] The action space contains three items: idle, active V2X, and backscattering, as follows:

[0141] The virtual OBU can choose three communication behaviors: 1. Action 0 (idle): no communication; 2. Action 1 (active V2X): transmit BSM messages at 20 dBm power; 3. Action 2 (backscattering): modulate the RSU broadcast signal to transmit low-rate state information.

[0142] The reward is the communication capacity gain minus the interference penalty. If active transmission is performed when the spectrum is busy and the CU is not in the safety zone, a strong negative reward of -100 is given, as follows:

[0143] The reward is the communication capacity gain minus the interference penalty. If action 1 is selected when the spectrum is occupied and the CU is not in the safety zone, a strong negative reward of -100 is applied; the backscattering mode always receives a positive reward, encouraging energy-efficient communication.

[0144] Step three, cloud policy training and model distribution;

[0145] S31, first train the teacher DQN, then train the student DQN, the loss contains TD error and KL divergence, as follows:

[0146] In the cloud simulation environment, first train the teacher DQN (3 layers x 128 neurons), then train the student DQN (2 layers x 32 neurons, a total of 1347 parameters), and the loss function contains TD error and KL divergence. Training uses a double experience replay pool (capacity of 10,000 each), the optimizer is Adam, the initial exploration rate ε = 1.0, and it is exponentially decayed to 0.01.

[0147] S32, convert the student model to FP16 format, compress it to about 50KB, and only when the throughput improvement is ≥5% is it marked as a deliverable version, as follows:

[0148] After training, the weight parameters of the student DQN model are converted from the original 32-bit floating-point format (FP32) to the 16-bit half-precision floating-point format (FP16), which is achieved by the PyTorch standard quantization tool chain. Without significantly losing inference accuracy, the model file size is compressed from about 100KB to about 50KB. Then, a semantic version number is assigned to the model, where the major version number represents architectural changes, the minor version number represents performance improvements, and the revision number represents training rounds. Next, a closed-loop simulation is run in a set of validation scenarios independent of the training set, and key indicators are calculated: average communication throughput, interference event rate, and backscatter decoding success rate. Only when the average throughput of the new model in all validation scenarios is improved by no less than 5% compared to the current online version, and the number of interference events does not increase significantly (increase ≤2%), the state is marked as "deliverable", otherwise the training result is discarded and the failure reason is recorded for subsequent analysis.

[0149] S33, the policy model is delivered to the local cache of all virtual RSUs for vehicles to download on demand, as follows:

[0150] The FP16 compressed policy model marked as "deliverable" is pushed to all virtual RSU nodes in the simulation environment through a secure transmission channel. Each virtual RSU maintains a policy cache area in the local file system, storing multiple historical models by version number. When a simulation vehicle enters the communication coverage of an RSU (defined as received signal strength RSRP>-85dBm), its OBU module initiates a policy synchronization request, and the RSU responds by delivering the latest "deliverable" version of the model file to the vehicle through a direct interface. After receiving the file, the vehicle checks it and loads it into the lightweight inference engine for subsequent real-time decision-making if it is correct.

[0151] Step four, online decision-making and communication simulation of vehicles;

[0152] S41, the vehicle automatically loads the latest policy model after entering the RSU coverage area (RSRP>-85dBm), as follows:

[0153] When a simulation vehicle enters the coverage area of a virtual RSU (received signal strength>-85dBm), its OBU module automatically loads the latest policy model from the RSU cache.

[0154] S42, collect state, inference action, and execute communication behavior every 10ms, and record logs, as follows:

[0155] In each 10ms simulation step, the OBU performs the following processes:

[0156] a. Collect the current 5-dimensional state;

[0157] b. Input the light DQN model, and forward inference to get the Q value;

[0158] c. Select the action corresponding to argmax Q;

[0159] d. Update the communication behavior according to the action: if it is proactive V2X, generate BSM and count the channel interference; if it is backscattering, superimpose the modulation component in the RSU received signal;

[0160] e. Record the state, action, reward, etc. Log, temporarily stored in the local.

[0161] S43, the RSU receives the mixed signal, recovers the backscattering data through serial interference cancellation and differential demodulation, specifically as follows:

[0162] The virtual RSU receives the mixed signal in each time slot, demodulates the broadcast main signal itself first, then extracts the backscattering sub-signal through serial interference cancellation, recovers the vehicle data through differential demodulation, and counts the decoding success rate.

[0163] Step five, policy closed loop update and performance evaluation;

[0164] S51, every 5 minutes, the RSU aggregates the vehicle logs in the jurisdiction and uploads to the cloud, specifically as follows:

[0165] Every 5 minutes of simulation time, each virtual RSU aggregates the logs (action sequence, throughput, interference event, decoding result) of all vehicles in the jurisdiction, and uploads them to the cloud server module through the simulation backhaul link.

[0166] S52, the cloud uses a double-buffer hybrid sampling strategy to incrementally train the model, only updating the output layer, specifically as follows:

[0167] The cloud stores new data into the short-term buffer, mixes it with long-term historical data, and uses a double-buffer sampling strategy to incrementally train the existing DQN model, freezes the bottom layer parameters, only updates the output layer, and sets the learning rate to 0.001.

[0168] S53, after the new strategy is verified to meet the standard through multi-scenario A / B testing, a new version is generated to enter the next round of distribution, forming a closed loop optimization, specifically as follows:

[0169] The trained new strategy is subjected to A / B testing in multiple PreScan scenarios (peak, night, rainy day, etc.), and the average throughput, interference frequency, decision delay and other indicators are evaluated. After meeting the standard, a new version is generated to enter the next round of distribution cycle, forming a closed loop of "simulation execution-data feedback-strategy optimization".

[0170] Embodiment two:

[0171] The embodiment provides a kind of based on environment backscattering car-road cloud integration communication system implementation, for the based on environment backscattering car-road cloud integration communication method described in embodiment one, specifically includes:

[0172] System modeling and parameter initialization module: for building car-road-cloud integrated simulation environment, deployment virtual RSU, OBU and cloud server, and complete initial parameter configuration.

[0173] Reinforcement learning decision element definition module: for defining 5-dimensional state space, 3-item action space and complex reward function containing strong negative reward.

[0174] Cloud platform DQN strategy training module: for executing "teacher-student" knowledge distillation framework, training and compression is obtained Light DQN strategy model, and strict performance verification is carried out.

[0175] Vehicle online decision and communication execution module: for realizing the strategy loading of vehicle end, real-time state perception, action decision and the execution of three communication modes.

[0176] Strategy closed loop update module: for aggregating roadside data, executing cloud incremental training, and driving continuous iteration and optimization of strategy by A / B test.

[0177] Embodiment three:

[0178] Referring to Figures 2 to 6 , the effect of the application is visualized.

[0179] Figure 2 (Light student model throughput evolution schematic diagram): it shows that the average communication throughput of student model increases steadily and finally converges with the increase of training rounds, which proves the effectiveness of its learning.

[0180] Figure 3 (Interference suppression capability evolution graph): it shows that the average number of interference events per step decreases significantly during training process, which verifies the effectiveness of strong negative reward mechanism for avoiding spectrum conflict.

[0181] Figure 4 (Communication mode selection evolution graph): it clearly reflects the selection ratio change of three actions (Idle, V2X, AmBC) during training process. With the deepening of training, the agent learns to use AmBC more in suitable scenarios, achieving the balance between energy saving and reliability.

[0182] Figure 5 (Teacher DQN training curve graph): it shows that the teacher model can successfully learn the complex strategy with high yield, and provides a reliable knowledge source for guiding student model.

[0183] Figure 6 (Figure of distillation training loss evolution): The common downward trend of KL divergence loss and TD loss shows that the student model has achieved a good balance between imitating the teacher's behavior and optimizing the strategy autonomously.

[0184] In summary, the application successfully combines the environmental backscattering technology with lightweight reinforcement learning through the specific implementation manner described above, and constructs an efficient, reliable and long-term evolution capable vehicle-road cloud integrated communication system.

[0185] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A vehicle-road-cloud integrated communication method based on environmental backscattering, characterized in that, include: Step 1: Construct an integrated vehicle-road-cloud simulation environment; S11. Build an urban road scene in PreScan and deploy virtual roadside units (RSUs). The specific method is as follows: In PreScan, a typical urban road scenario is built, including intersections, straight lanes, and elevated ramps. Virtual roadside units (RSUs) are set at 200-meter intervals. A high-precision map is imported, and vehicle flow, cellular user distribution, and movement models are set. The radius of the RSU safe zone is defined as 200 meters. S12. An integrated 5.9 GHz V2X channel model, including path loss, shadowing fading, and Doppler effect, is used to set RSU broadcast signal parameters, vehicle backscatter reflection coefficient, and modulation rate. The specific method is as follows: In the PreScan and MATLAB / Simulink environment, a 5.9 GHz V2X channel model was integrated, including path loss, shadowing fading, and Doppler effect. The RSU broadcast signal parameters were set as follows: center frequency 5.9 GHz, bandwidth 10 MHz, OFDM modulation, and transmit power 30 dBm. At the same time, the environmental backscatter link was modeled, and the vehicle reflection coefficient α = 0.5 and the backscatter modulation rate was set to 10 kbps. The V2X channel model specifically includes the following: 1: Path loss model; Path loss With distance The changes are modeled using a hybrid approach combining the standard free-space propagation model and the urban micro-cell model: ； In the formula, For reference distance Path loss at the location; The path loss index; This is the Euclidean distance between the vehicle and the RSU; Let be a shadow fading random variable, following a zero-mean normal distribution; 2: Received signal power model; The received power of the RSU transmitted signal at the vehicle is: ； In the formula, RSU transmit power; The calculated path loss; The power of the RSU broadcast signal received by the vehicle; 3: Environmental backscattering link model; The vehicle modulates the antenna impedance to change the reflection coefficient α of the RSU signal, thereby achieving information encoding; assuming the reflection coefficient varies within the range of α∈[0,1], the equivalent transmit power of the backscattered signal is: ； In the formula, The set average reflectance coefficient; The incident signal power from the RSU; The backscatter power of the vehicle's "virtual emission"; 4: Backscatter communication rate model; The backscatter modulation rate is determined by the channel bandwidth and the modulation scheme. In this simulation, we set: ； In the formula, Available bandwidth; The signal-to-noise ratio of the backscatter link; 5: Doppler frequency shift model; The Doppler frequency shift is: ； In the formula, For vehicle speed; The angle between the vehicle's direction of movement and the direction the signal arrives; At the speed of light, The center frequency; S13. Implement the three main virtual modules: virtual RSU module, virtual OBU module, and cloud training server module. The specific methods are as follows: Virtual RSU module: periodically broadcasts traffic status, receives and decodes backscatter signals, and has an 8 GB cache for storing policy models; Virtual OBU module: Mounted on each simulated vehicle, used to integrate perception, decision-making, and communication sub-modules; Cloud training server module: runs on a local workstation or cluster and is responsible for policy training, version management and secure distribution; Step 2: Define the decision-making elements of reinforcement learning; S21. Each OBU collects the following 5-dimensional status in every 10 ms simulation step:

1. Spectrum occupancy flag: determined by local energy detection; 2. Euclidean distance to the nearest virtual RSU; 3. Distance to the nearest other simulated vehicle; 4. Whether there is a CU located within the RSU safe zone; 5. Instantaneous vehicle speed; All dimensions are normalized to [0, 1] and then used as input to the policy network; S22. The virtual OBU module selects three communication behaviors:

1. Action 0, i.e., idle: no communication is performed; 2. Action 1, i.e., active V2X: BSM messages are sent at a power of 20 dBm; 3. Action 2, i.e., backscatter: the RSU broadcast signal is modulated and vehicle status information is sent at a modulation rate of 10 kbps. S23, Reward = Communication Capacity Gain - Interference Penalty; Where, if Action 1 is selected when the spectrum is occupied and the CU is not in the safe zone, a strong negative reward of -100 is applied; Backscatter mode always receives a positive reward, encouraging energy-saving communication; Step 3: Cloud-based strategy training and model deployment; S31. First train the teacher's DQN model, then train the student's DQN model; S32. Convert the student DNQ model to FP16 format; S33. Distribute the compressed and verified student DQN model to the local cache of all virtual RSUs; Step 4: Simulation of online vehicle decision-making and communication; S41. After a vehicle enters the RSU coverage area, the latest student DQN model is automatically loaded. The specific method is as follows: When a simulated vehicle enters the coverage area of ​​a virtual RSU (i.e., the received signal strength is > -85 dBm), the virtual OBU module automatically loads the latest policy model from the virtual RSU module cache; S42. Collect status, inference actions, execute communication behaviors, and record logs. The specific method is as follows: Within each 10 ms simulation step, the OBU executes the following procedure: a. Collect the current 5D status; b. Input the lightweight DQN model and obtain the Q value through forward inference; c. Select the action corresponding to argmax Q; d. Update communication behavior based on actions: If it is active V2X, generate BSM and take channel interference into account; If it is backscattering, then the modulation component is superimposed on the RSU received signal; e. Record status, actions, and rewards, and temporarily store them locally; S43 and RSU receive mixed signals; the specific method is as follows: The virtual RSU module receives mixed signals in each time slot, first demodulates its own broadcast main signal, then extracts the backscattered secondary signal through serial interference cancellation, recovers the vehicle data through differential demodulation, and counts the decoding success rate. Step 5: Policy closed-loop update and performance evaluation; S51 and RSU aggregate vehicle logs within their jurisdiction and upload them to the cloud; S52. The cloud uses double-buffered hybrid sampling to incrementally train the currently deployed student DQN model, updating only the output layer. S53. After the new strategy has been verified and met the standards through multi-scenario A / B testing, a new version will be generated and enter the next round of distribution, forming a closed-loop optimization.

2. The vehicle-road-cloud integrated communication method based on environmental backscattering according to claim 1, characterized in that, The specific method for step S31 is as follows: In a cloud simulation environment, a teacher DQN model is first trained, which is a deep Q-network with three fully connected layers and 128 neurons per layer. Then, a student DQN model is trained based on the knowledge distillation method. This is a simplified student network with two fully connected layers and 32 neurons per layer. The student DQN model is jointly optimized by minimizing the KL divergence between the TD error and the teacher model output, aiming to obtain a lightweight policy network with performance close to that of the teacher model. Training uses a dual experience replay pool with a capacity of 10,000 each, the optimizer is Adam, the initial exploration rate ε=1.0, and it decays exponentially to 0.

01. The specific method for step S32 is as follows: After training, the weight parameters of the student DQN model were converted from the original 32-bit floating-point format (FP32) to 16-bit half-precision floating-point format (FP16). This operation was performed using the PyTorch standard quantization toolchain, compressing the model file size from approximately 100 KB to about 50 KB. Subsequently, a semantic version number was assigned to the student DQN model, where the major version number indicates an architecture change, the minor version number indicates a performance improvement, and the revision number indicates the training epoch. Next, closed-loop simulations were run in a set of validation scenarios independent of the training set, and key indicators were statistically analyzed: average communication throughput, interference event occurrence rate, and backscatter decoding success rate. The new model was marked as "ready to be deployed" only when the average throughput of the new model in all validation scenarios was at least 5% higher than the current online version, and the number of interference events did not increase significantly (i.e., the increase was ≤2%). Otherwise, the training results were discarded and the reasons for failure were recorded for subsequent analysis. The specific method for step S33 is as follows: The FP16 compressed policy model marked as "deployable" is pushed to all virtual RSU nodes in the simulation environment through a secure transmission channel. Each virtual RSU maintains a policy cache in its local file system, storing multiple historical models by version number. When a simulated vehicle enters the communication coverage area of ​​a certain RSU, defined as when the received signal strength RSRP > -85 dBm, the virtual OBU module will actively initiate a policy synchronization request. After the virtual RSU module responds, it will distribute the latest "deployable" version of the model file to the vehicle through the direct connection interface. After receiving the file, the vehicle will verify it and, if it is correct, load it into the lightweight inference engine for subsequent real-time decision-making.

3. The vehicle-road-cloud integrated communication method based on environmental backscattering according to claim 1, characterized in that, The specific method for step S51 is as follows: Every 5 minutes of simulation time, each virtual RSU aggregates the logs of all vehicles in its jurisdiction, namely action sequences, throughput, interference events, and decoding results, and uploads them to the cloud server module through the simulation backhaul link; The specific method for step S52 is as follows: The cloud stores new data in a short-term buffer and mixes it with long-term historical data. A double-buffer sampling strategy is used to incrementally train the existing DQN model, freezing the underlying parameters and updating only the output layer. The learning rate is set to 0.

001. The specific method for step S53 is as follows: The trained new strategy is A / B tested in multiple PreScan scenarios to evaluate metrics such as average throughput, number of disturbances, and decision latency. Once the targets are met, a new version is generated and enters the next round of distribution, forming a closed loop of "simulation execution – data feedback – strategy optimization".

4. A vehicle-road-cloud integrated communication system based on environmental backscattering, comprising the vehicle-road-cloud integrated communication method based on environmental backscattering as described in any one of claims 1-3, characterized in that, System modeling and parameter initialization module: used to build an integrated vehicle-road-cloud simulation environment, deploy virtual RSUs, OBUs and cloud servers, and complete the initial parameter configuration; The reinforcement learning decision element definition module is used to define a 5-dimensional state space, a 3-action space, and a composite reward function that includes strong negative rewards. The cloud platform DQN policy training module is used to execute the "teacher-student" knowledge distillation framework, train and compress a lightweight DQN policy model, and perform rigorous performance verification. Vehicle Online Decision and Communication Execution Module: Used to implement vehicle-side policy loading, real-time status perception, action decision-making, and execution of three communication modes; The strategy closed-loop update module is used to aggregate roadside data, perform incremental training in the cloud, and drive continuous iteration and optimization of the strategy through A / B testing.

Images