RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning

Abstract

The invention discloses an RIS-ISCC functional slicing method oriented to vehicle-edge collaborative reasoning, which is used for uniformly virtualizing spectrum resources of space-air-ground nodes, an RIS reflection unit and reasoning capability into functional slices for providing perception, communication and calculation energy as required. In a sensing and communication layer, spectrum resources of space-air-ground nodes are divided into sensing detection and communication function slices, and RIS reflection units are synchronously divided to enhance track detection and task unloading capability; in the calculation level, an isomorphic reasoning model is deployed at vehicle and edge nodes, and flexible reconstruction of a vehicle-edge reasoning function is realized through dynamic division of network entrances and exits of the isomorphic reasoning model, so that calculation resource allocation is indirectly adjusted. A system utility maximization problem is modeled as a constrained long-time accumulation optimization problem, and is solved through a dynamic slicing algorithm of DRL heterogeneous resources. The algorithm combines traffic flow prediction, a network state and a task load, adaptively divides spectrum resources, configures an RIS function and determines an inference model segmentation strategy.

Description

RIS-ISCC Functional Slicing Method for Vehicle-Edge Collaborative Reasoning Technical Field

[0001] This invention belongs to the field of vehicle networking technology, specifically a RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning. Background Technology

[0002] With the continuous evolution of intelligent transportation and autonomous driving, Space-Air-Ground Integrated Vehicular Networks (SAGVNs) are emerging. [1] It is considered a crucial supporting architecture for next-generation intelligent mobility. Its goal is to build an open, secure, and agile network system, breaking down barriers between different networks and integrating ground-based, airborne, and space-based communication resources to provide seamless information interaction and full-area coverage for in-vehicle services. This system is based on terrestrial cellular networks, supplemented by drone small base stations and low-Earth orbit satellite networks, forming a three-dimensional, interconnected, multi-layered network structure that provides vehicles with all-time, all-area access and data services. The ground-based network provides high-capacity communication and edge computing support for densely populated urban areas; the airborne network can be flexibly deployed according to traffic and mission loads to achieve rapid coverage and road network monitoring; and the space-based network achieves wide-area interconnection through a constellation of low-Earth orbit satellites, providing robust protection for remote or emergency scenarios.

[0003] In this architecture, the functions of vehicle-to-everything (V2X) have expanded from traditional communication services to a deep integration of perception and intelligent reasoning. Autonomous vehicles can generate over 1GB of multimodal data per second, encompassing video streams, radar point clouds, and high-precision map information. [2] Video inference is particularly crucial for autonomous driving tasks such as object detection, scene understanding, and decision control. Video inference is extremely sensitive to communication and computing capabilities: on the one hand, real-time detection and segmentation tasks rely on high-throughput video data transmission and low-latency result return; on the other hand, the computing power of onboard terminals is limited, making it difficult to complete complex neural network inference calculations under energy-constrained conditions. Therefore, edge intelligence-based collaborative inference has become the mainstream solution. This involves offloading some or all video inference tasks to airborne or ground-based edge nodes through vehicle-edge-air-space computing collaboration to compensate for local computing bottlenecks and shorten end-to-end response time. However, due to differences in task complexity, node load, and network status, a single communication or computing mechanism cannot simultaneously meet the comprehensive requirements of video inference tasks in terms of reliability, latency, and accuracy. Therefore, integrated sensing and computing technologies are needed. [3] The proposed approach integrates sensing, communication, and computing (inference) in a unified architecture, enabling wireless signals to perform environmental perception (vehicle trajectory detection) tasks while transmitting information, and to collaborate with multi-layer computing resources to achieve vehicle-edge joint inference, thereby balancing communication latency, sensing accuracy, and inference efficiency.

[0004] In the integrated air-space-ground vehicle network, the Intelligent Reflecting Surface (RIS) is crucial. [4] The introduction of RIS provides new physical support for unleashing the potential of Integrated Sensing–Communication–Computation (ISCC). As a programmable electromagnetic reconfiguration device, RIS can reshape the channel propagation environment without increasing additional power consumption by flexibly modulating the amplitude and phase of the incident wave, thereby synergistically enhancing communication (e.g., data or feature offloading), sensing (e.g., vehicle trajectory detection), and computing (e.g., vehicle-side joint inference).

[0005] At the communication level, RIS can dynamically construct vehicle-side-air-sky multi-hop reflection links, effectively mitigating non-line-of-sight blocking and channel fading, and improving signal-to-noise ratio and link stability. [5] This provides a reliable guarantee for high-quality video data transmission between vehicles and edge nodes. At the perception level, the high-dimensional reflective array of the RIS can achieve beam focusing and direction reconstruction in some spectrum resources, thereby enhancing the system's spatial detection and scene perception capabilities. When airborne and ground-based nodes work together, the RIS can form a virtual synthetic aperture, achieving high-precision reconstruction of vehicle trajectory and motion state based on reflected echo signals, realizing spectrum-level fusion of communication and perception. [6] .

[0006] At the computing level, the programmability of RIS can be extended to task scheduling and inference collaboration. By dynamically configuring reflection units and beam paths, it provides multi-channel task migration paths for video inference models, enabling tasks to be flexibly unloaded between vehicle-based, air-based, or space-based nodes according to channel status, node load, and computing requirements.

[0007] Therefore, RIS becomes the key physical enabling technology driving ISCC implementation in SAGVN.

[0008] While RIS-enabled ISCC offers new system potential for integrated air-space-ground vehicle-to-everything (V2X) connectivity, several challenging problems still need to be addressed in the dynamic environment of V2X:

[0009] One of the challenges is the correlation modeling and adaptive partitioning of heterogeneous resources.

[0010] Space-air-ground multi-layer networks exhibit significant heterogeneity in bandwidth, computation, and reflection resources, and the electromagnetic reconfiguration behavior of RIS (Resonance Injection System) simultaneously affects communication channel gain and sensing resolution. This multi-source, multi-domain resource coupling makes it difficult for traditional independent allocation mechanisms to maintain global consistency and real-time performance. For integrated space-air-ground networks, researchers have proposed multi-dimensional slicing and resource allocation schemes from different perspectives. Chen et al. [7] A hierarchical, multi-dimensional resource slicing framework assisted by low-Earth orbit satellites is proposed. By combining a Long Short-Term Memory (LSTM) model and a Soft Actor-Critic (SAC) algorithm, the transmission rate is improved while meeting the Service Level Agreement (SLA). (Slice-Soft-SAGIN) [8] It is an integrated space-air-ground network resource allocation framework that integrates resource detection and allocation modules for ground, air, and satellite resources, providing customized resource orchestration schemes for each slice. (Lyu et al.) [9] A slice-based online management framework for an integrated air-space-ground vehicle network is proposed, which periodically schedules workflows, deploys drones, and updates resources to provide customized services with performance isolation. (Zhou et al.)

[10] Centralized and distributed multi-agent deep deterministic policy gradient algorithms are proposed to manage resources of slices with high throughput, low latency, and wide coverage, aiming to obtain Pareto optimal solutions. Considering the network topology and coverage changes caused by satellite orbiting the Earth, Kak et al.

[11] This method abstracts dynamic network topologies into multiple static network topologies, and performs slice deployment and resource management on each static network topology. It is suitable for scenarios with a small number of nodes. (Rodrigues et al.)

[12] Centralized and distributed learning analysis is employed to analyze user request patterns and routing information, reserving slice resources for various tasks offloaded to satellite or ground base stations.

[0011] Another challenge lies in the uncertainty of reasoning under the fusion of synesthesia and computation.

[0012] The deep integration of ISCC (Intelligent Edge Computing) enables communication, sensing, and computing to mutually reinforce each other, but it also introduces significant uncertainties: fluctuations in communication link status can alter the sensing signal-to-noise ratio, sensing errors can affect task offloading decisions, and computing results, in turn, influence resource allocation. This multi-factor, bidirectional dependency makes inference performance difficult to predict. In vehicular networks empowered by mobile edge computing, the allocation of sensing, communication, and computing resources affects base station sensing performance. To address this, Li et al.

[13] A cooperative power, bandwidth, and computational resource allocation scheme is proposed to maximize the overall utility, including sensing accuracy, sensing information processing latency, and communication and computational latency for computational tasks. Wen et al. addressed the signal design problem of integrating sensing, communication, and computation.

[14] A dual-function signal design method is proposed, dividing the antenna array into sensing and computation signals to improve computational accuracy while ensuring sensing performance. Considering the resource contention and performance conflicts in integrated sensing, communication, and computing networks, Zhao et al.

[15] A wireless resource scheduling architecture based on bilateral matching theory is proposed, modeling the device association and subchannel allocation problem as a bilateral one-to-one matching problem with externalities. This method is applicable to multi-cell, multi-user, and multi-functional smart IoT scenarios. Considering the coupling between sensing, computation, and communication processes in edge intelligent inference tasks, Xing et al.

[16] A task-oriented multi-device integrated sensing, communication, and computing scheme is proposed. By splitting the model into device-side feature extraction and server-side inference, a non-convex optimization problem is constructed with sensing power, communication time, and quantization bit depth as decision variables. Considering the limited sensing range of a single vehicle and the randomness of the arrival of collaborative tasks in the vehicle-to-everything (V2X) scenario, Zhao et al.

[17] An integrated perception, computing, and communication system for vehicle-to-everything (V2X) networks is proposed. Through a collaborative perception and data fusion architecture between vehicles and roadside units, the joint optimization problem of communication, perception, and computing resources is modeled as a mixed-integer nonlinear programming problem with task completion rate as the objective. (Li et al.)

[18] By integrating sensing, communication, and computing technologies, a dual-function signal transmission framework is proposed, including two schemes: "shared" and "separated". The joint beamforming optimization problem in both schemes is modeled as a non-convex optimization problem and solved using semi-definite relaxation and Gaussian randomization methods.

[0013] In summary, existing slicing mechanisms mostly focus on the division of communication or computing dimensions, without establishing a unified modeling and joint optimization mechanism for spectrum, RIS reflection units and inference computing resources. Furthermore, ISCC research mainly focuses on static local resource optimization, lacking unified modeling of uncertainties in multiple factors of perception, communication and computing, as well as support for dynamic real-time offloading of multiple exit nodes in highly dynamic scenarios. Summary of the Invention

[0014] This invention proposes a functional slicing method based on RIS-enhanced ISCC to support video inference applications related to autonomous driving in an integrated air-space-ground vehicle network. Its goal is to maximize the completion rate of latency-sensitive inference tasks under a given offloading strategy.

[0015] A RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning is proposed. A resource slicing framework is constructed in the air-space-ground integrated vehicle network SAGVN, and resource slicing is implemented under the resource slicing framework. Base stations, UAVs and satellites in SAGVN are used as edge nodes. Base stations refer to ground base stations, UAVs refer to UAV-borne base stations, and satellites refer to satellite-borne base stations.

[0016] The SAGVN also includes multiple intelligent RIS units;

[0017] First, build the resource slicing framework:

[0018] The resource slicing framework is a RIS-ISCC functional resource slicing system framework based on "functional provision". Under this architecture, spectrum resources, intelligent reflector resources and computing resources are uniformly virtualized into functional slices; RIS stands for intelligent reflector and ISCC stands for integrated sensing and computing.

[0019] At the perception and communication level of the slicing framework: the spectrum resources of the air-space-ground nodes are divided into two functional slices: perception and detection, and communication. Correspondingly, the RIS reflection unit is synchronously divided into two parts, which are used to enhance vehicle trajectory detection and task offloading, respectively, to enhance the vehicle trajectory detection capability of the perception function and the task offloading capability of the communication function.

[0020] At the computational level of the slicing framework: isomorphic inference models are deployed on vehicles and edge nodes, and the network structure of the inference model is synchronously divided into entry and exit points, realizing the reconstruction of vehicle-edge inference functions, thereby indirectly adjusting the allocation of computing resources;

[0021] Then, implement resource slicing:

[0022] The system utility maximization problem is modeled as a constrained long-term cumulative optimization problem and solved by a reinforcement learning algorithm for RIS-ISCC functional slicing. This algorithm, as a slicing algorithm, combines traffic flow prediction, network state and task load to adaptively allocate spectrum resources and configure RIS functions, and determine the inference model segmentation strategy.

[0023] The slicing algorithm design process is as follows:

[0024] S1. Construct a resource-aware vehicle trajectory detection model and a vehicle-side communication model to quantify the impact of slicing on detection error and communication latency, respectively.

[0025] S2. The long-term optimization problem with the goal of maximizing the completion rate of latency-sensitive inference tasks is modeled as a Markov decision process, and a reinforcement learning algorithm based on a dual deep Q-network DDQN is designed. The algorithm dynamically and collaboratively optimizes the spectrum and the entry / exit selection of RIS resource allocation and inference model based on ISCC functional slice states, vehicle trajectories and real-time load.

[0026] The main contributions of this invention include:

[0027] RIS–ISCC Functional Slicing Framework: This invention innovatively constructs a functional resource slicing framework from the perspective of "providing synesthetic computing functions." Spectrum resources, reflector resources, and computing resources are virtualized into functional slices, enabling perception (trajectory detection), communication (task offloading), and computing (vehicle-edge collaborative inference). At the perception and communication function level, spectrum resources on the space-air-ground nodes are divided into perception detection and communication; correspondingly, the unit blocks on each reflector RIS are also divided into two parts to enhance vehicle trajectory detection and task offloading. At the computing function level, isomorphic inference models are deployed on vehicles and edge nodes, with network entry / exit points synchronously divided to indirectly regulate vehicle-edge computing resources.

[0028] A reinforcement learning algorithm for RIS–ISCC functional slicing is proposed: A resource-aware vehicle trajectory detection model and a vehicle-side communication model are constructed to quantify the impact of resource slicing on detection error and communication latency, respectively. Based on this, a long-term optimization problem aiming to maximize the completion rate of latency-sensitive inference tasks is modeled as a Markov decision process, and a reinforcement learning algorithm based on a dual deep Q-network (DDQN) is designed. This algorithm can dynamically and collaboratively optimize spectrum / RIS resource allocation and model entry / exit selection based on ISCC functional slice status, vehicle trajectory, and real-time load, thereby continuously improving system utility.

[0029] This invention presents a RIS-enhanced ISCC functional slicing method for the Internet of Vehicles (IoV), constructing a dynamic resource slicing framework based on "functional provision," which unifies communication, sensing, and computing resources into programmable functional units. Experiments demonstrate that this method significantly improves task success rate in dynamic traffic environments and maintains a high success rate even with increased vehicle density, exhibiting strong robustness. The framework's resource abstraction mechanism and DRL optimization algorithm can be flexibly adapted to various edge nodes and business scenarios, and are easily migrated to other integrated sensing, communication, and computing systems such as UAV swarms and industrial IoT. Attached Figure Description

[0030] Figure 1 illustrates a RIS-ISCC-enabled vehicle-edge collaborative inference scenario;

[0031] Figure 2 shows the adaptive vehicle-edge model segmentation at the entrance / exit;

[0032] Figures 3(a) and 3(b) illustrate the impact of resource slicing on vehicle trajectory detection and communication transmission rate, respectively.

[0033] Figure 3(a) shows and Impact on 3D positioning error and communication rate

[0034] Figure 3(b) shows Impact on 3D positioning error and communication rate;

[0035] Figure 4 shows the implementation framework of the slicing algorithm;

[0036] Figures 5(a) and 5(b) show the effects of task unloading success rate and reward on training rounds, respectively.

[0037] Figure 5(a) shows the task unloading success rate.

[0038] Figure 5(b) illustrates reward acquisition;

[0039] Figures 6(a) and 6(b) illustrate the impact of the number of available resource blocks on the task completion rate, respectively.

[0040] Figure 6(a) illustrates the impact of increasing spectrum resource blocks on task completion rate.

[0041] Figure 6(b) illustrates the impact of increased computing resources on task completion rate;

[0042] Figure 7 shows the cumulative distribution of delay time for successfully completed tasks;

[0043] Figure 8 illustrates the impact of vehicle density on the mission unloading success rate.

[0044] Figure 9 illustrates the impact of delay time on the success rate of task unloading. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0046] 1. Overview of the Solution

[0047] Reconfigurable Intelligent Surface (RIS) enhanced Integrated Sensing–Communication–Computation (ISCC) is considered an important development direction for the future of connected vehicles. However, most existing connected vehicle resource slicing methods focus on differentiated services or resource quota allocation, and few works can support the collaborative reasoning requirements in integrated sensing–computation scenarios.

[0048] To address the aforementioned issues, this invention proposes a RIS-ISCC functional slicing framework for vehicle-edge collaborative inference. Unlike traditional service- or resource-providing network slicing mechanisms, this framework unifies the spectrum resources of air-space-ground nodes, RIS reflection units, and inference capabilities into functional slices, providing on-demand sensing, communication, and computing functions. At the sensing and communication levels, the spectrum resources of air-space-ground nodes are divided into two functional slices: sensing and communication. Simultaneously, RIS reflection units are partitioned to enhance trajectory detection and task offloading capabilities. At the computing level, homogeneous inference models are deployed on vehicles and edge nodes. The dynamic partitioning of their network entry and exit points enables flexible reconfiguration of vehicle-edge inference functions, thereby indirectly adjusting the allocation of computing resources. The system utility maximization problem is modeled as a constrained long-term cumulative optimization problem and solved using a dynamic slicing algorithm based on deep reinforcement learning (DRL) for heterogeneous resources. This algorithm adaptively partitions spectrum resources, configures RIS functions, and determines the inference model partitioning strategy by combining traffic flow prediction, network state, and task load.

[0049] Simulation results show that the proposed framework has good adaptability and stability in dynamic road network environments, and can significantly improve the success rate of task delivery and the efficiency of system resource utilization.

[0050] In the following text, "base station" specifically refers to a ground base station, "drone" refers to an airborne base station of a drone, and "satellite" refers to a low-orbit satellite base station.

[0051] Table 1: Definitions of key symbols in the following text

[0052]

[0053] 2 RIS-ISCC Functional Resource Slicing Architecture

[0054] As shown in Figure 1, this invention considers a RIS-ISCC-enabled vehicle-edge collaborative inference scenario. A dual-timescale, dual-layer edge control system, consisting of two edge controllers, determines resource slicing and task scheduling. The edge controllers are deployed near the base station, responsible for centralized management of the base station and UAVs; the ground workstation is dedicated to satellite task scheduling. The system status information received by both is uniformly provided by the edge controllers. It is assumed that time is divided into a series of slice windows, where each window... The duration is Time is divided into multiple equal-length slice windows, and each slice window is further divided into equal-length scheduling time slots. Slice Window The set of included scheduling slots is defined as follows: The sets of edge nodes (ground base stations, drones, low-Earth orbit satellites) and RIS are represented as follows: , At the start of the slicing window, the edge controller formulates a RIS-ISCC slicing scheme, i.e., slicing (redistribution), based on collected historical task information. and The sensory resources of the mid-node and the inference model (computational resources) for vehicle-edge sharing. In small-scale time slots... Within the control area, the edge controller integrates slice resource usage and network situation changes, and uses reflective surfaces to batch schedule inference tasks generated by vehicles. Edge nodes, acting as collaborators, receive, process, and return results to the target vehicles.

[0055] 2.1 RIS-ISCC Functional Resource Slicing Framework

[0056] Unlike traditional RAN slicing mechanisms based on communication services or resource allocation, this invention constructs a RIS-ISCC functional resource slicing framework from the perspective of providing synergistic computing functions to support autonomous driving applications in integrated air-space-ground vehicle networks. In this framework, slicing no longer merely represents a static allocation of spectrum or computing resources, but is abstracted as a set of synergistic computing functions that can be dynamically enabled, combined, and reconfigured to support communication, perception, and collaborative reasoning tasks on demand.

[0057] To achieve the above objectives, the communication spectrum, RIS reflection units, and inference execution capabilities of heterogeneous nodes in the air, space, and ground are first pooled uniformly and mapped to functional-level resource entities through network function virtualization. Under the centralized control of the software-defined network, the edge controller continuously senses the traffic density, link status, and task queue load within the slice window, and dynamically adjusts the functional forms provided by the synergistic computing at different stages accordingly. At the same time, the reflection configuration of the RIS and the execution position of the inference model are incorporated into a unified control loop, enabling communication enhancement, perception assurance, and inference collaboration to evolve collaboratively within the same functional slice framework, rather than being optimized independently.

[0058] At the communication and sensing level, let the set of edge nodes be... .node In the slice window The total number of available spectrum resource blocks is For the set of foundation and air foundation nodes Its spectrum resources are mapped into two types of functional slices, used for communication functions and sensing functions respectively, with a corresponding number of resource blocks. and (satisfy Through this modeling method, RIS in The reflective units are divided and Two parts (satisfying) RIS reflection capability is explicitly characterized as a programmable system capability that supports the sensing function, rather than just a passive channel enhancement means, so that reflection configuration can be functionally optimized in conjunction with spectrum allocation.

[0059] At the inference level, the proposed slicing method does not target traditional computing resources, but rather reflects the dynamic segmentation and flexible reconstruction of the inference model structure between the vehicle and edge nodes. The vehicle and edge nodes share a single... A pre-trained inference model with residual blocks, where the vehicle set is represented as... Within each slice window, the edge controller dynamically determines the inference output of the vehicle-side model based on the synesthetic computing function requirements and network status. And the entry point for the corresponding collaborator model. ,in( ).

[0060] This model-structure-based partitioning approach essentially depicts the redistribution of inference functionality between the vehicle and edge nodes, rather than a static division of computing power or processing capabilities. Different exit options correspond to different inference functionality provision modes:

[0061] 1) Vehicle-side full reasoning ( ): Task reasoning relies entirely on onboard resources;

[0062] 2) Vehicle-side complete unloading ( The original task data is offloaded to edge processing;

[0063] Car-side collaborative reasoning ( The task is partially inferenced on the vehicle side, and intermediate features are output from the vehicle-side model. Unload to the entry point of the edge node The edge nodes complete the remaining inference and return the calculation results. Take the red (black) line in Figure 2 as an example. The intermediate features of the task originate from the exit of the vehicle-end model. ( The entry point for uploading to the collaborator model. ( ).

[0064] 2.2 RIS-assisted vehicle-side communication model

[0065] This section constructs a model that comprehensively considers the impact of the number of reflective elements and phase configuration on channel gain and data transmission rate.

[0066] Assuming RIS Include One reflective unit, wherein the reflective unit The phase is represented as The phase matrix of the reflecting surface is represented as follows:

[0067] (1)

[0068] RIS To edge nodes (vehicle The vector representation of the channel is as follows Therefore, the vehicle via RIS To edge nodes The gain of the cascaded channel (CC) is calculated as follows:

[0069] (2)

[0070] Set edge nodes To the vehicle The distance is Then the edge The calculation results are sent back to the vehicle. At that time, the channel gain was calculated as

[0071] (3)

[0072] in, It is path loss. It is the path loss index. This is a small-scale Rayleigh fading channel. It operates at the maximum-ratio-reflecting (MRR) level.

[20] Below, RIS Upper reflector unit The phase is determined as

[0073] (4)

[0074] The gain of the equivalent channel can be calculated as follows:

[0075] (5)

[0076] It should be noted that in the considered scenario, the edge nodes, acting as collaborators, must opportunistically cover moving vehicles after processing their tasks to complete result delivery. Since downlink delivery of inference results with small datasets has almost no performance bottleneck, RIS is only used to enhance uplink transmission. Accordingly, a binary variable regarding (uplink) task scheduling is defined. A value of 1 indicates that in the time slot Below, vehicle RIS The task Unload to edge node Otherwise, it is 0. Given ,vehicle Upload task To edge nodes (base stations or drones) The uplink rate at that time was quantized as

[0077] (6)

[0078] exist In the case of RIS assistance, equation (6) degenerates into one without RIS assistance (relying only on RIS assistance). The pattern of (middle edge nodes).

[0079] Given ,vehicle Upload task To satellite The uplink speed at that time is

[0080] (7)

[0081] Edge nodes (base stations or drones) Return mission To the vehicle The downlink rate at that time is

[0082] (8)

[0083] satellite Return mission To the vehicle The downlink rate at that time is

[0084] (9)

[0085] Unlike Equation (6), Equations (8) and (9) do not reference any reflector resources when characterizing the downlink.

[0086] 2.3 RIS-enhanced vehicle trajectory detection model

[0087] Unlike traditional ground-based detection methods, this section constructs a RIS-enhanced vehicle trajectory detection model, in which the addition of UAVs and RIS enhances the detection performance of the base station from the dimensions of line-of-sight (quality, accuracy) and perception range.

[0088] Set up vehicles Located in Base station lie in .vehicle Relative to base station The elevation angle was detected as

[0089] (10)

[0090] base station Observed vehicles Azimuth angle is

[0091] (11)

[0092] Let the time difference between the radar transmitted wave and the echo be... Due to the two-way propagation delay of the signal, the base station With vehicles The detection range was estimated to be

[0093] (12)

[0094] in, It is the speed of light. Let's assume it is located at... RIS to base station (vehicle The straight-line distance is ( ).vehicle Signal via RIS Reaching the edge node The multipath distance was determined as

[0095] (13)

[0096] Based on equation (4) and the inertia rule, the vehicle exist The position of the moment is

[0097] (14)

[0098] Set edge nodes The received power is The resource block size is (kHz). In non-cooperative (no RIS and drone) mode, for the probe Arrival vehicle The distance measurement error was calculated as

[0099] (15)

[0100] In a cooperative model (with RIS and drones), a collaborative sensing resource detection model was constructed to characterize the impact of the quantity of sensing resources on vehicle trajectory detection error. We assume base stations... A group of drones can be dispatched Drones With base station The distance between them is Then drone Due to distance and base station The attenuation of the generated sensing signal was calculated as

[0101] (16)

[0102] in Reference distance. Efficiency factor. and This is introduced to reflect the enhanced vehicle trajectory detection performance of RIS. As an extension of Equation (15), the detection error in cooperative mode is calculated as follows:

[0103] (17)

[0104] in, Noise power spectral density. Given , and The value in equation (7) is the same as They are inversely correlated. Assume the vehicle... With base station Angle measurement error meets The 3D positioning error of the vehicle sensed by the base station can be calculated as follows:

[0105] (18)

[0106] Visualization using equations (6), (17) and (18) , and The impact on vehicle trajectory detection and communication transmission rate. Figure 3(a) shows... time and Increased impact on the communication transmission rate of 3D positioning errors. Between 0 and 20 The downward trend is most pronounced when the value is between 0 and 8. Greater than 20, When the number of resource blocks exceeds 8, the rate of error reduction gradually slows down. It is evident that the optimal number of base station sensing resource blocks should be between 15 and 25, while the optimal number for UAV sensing resources should be between 3 and 8. As shown in Figure 3(b), the settings are... ,along with With the increase of [variable], the 3D positioning error has been on a downward trend. [Selection] , and It can balance communication awareness capabilities as much as possible.

[0107] 3. Problem Modeling

[0108] To evaluate the impact of resource slicing on task offloading performance, this section constructs an end-to-end latency model oriented towards offloading. This model comprehensively considers factors such as RIS configuration, vehicle trajectory, and node load to quantify the impact of resource slicing schemes on task offloading latency. The system dynamically selects the offloading path based on the resource slice status (such as communication / sensing resource block partitioning, RIS phase configuration, and model exit location) to achieve successful task delivery under latency constraints.

[0109] Within the proposed framework, the goal of resource slicing is to achieve the following within each slice window: Within the system, optimal communication, sensing, and computing resources are allocated for downstream task offloading. The edge controller is responsible for optimizing the slicing scheme, including communication / sensing resource allocation, RIS unit partitioning, RIS phase configuration, and model exit location. Task scheduling is performed by the DDPG strategy deployed on the edge controller under a given slicing scheme. Therefore, the resource slicing optimization problem is independent of the specific scheduling process, and its goal is to provide optimal resource allocation for the DDPG strategy to maximize the long-term task success rate of the system.

[0110] Set up vehicles The latency for performing full inference is Edge nodes The full inference latency of the model is Then the task Reasoning to the model number The processing delay of each export is approximated as

[0111] (19)

[0112] Similar task From the model The processing delay for each entry point to continue inference until the end is calculated as follows:

[0113] (20)

[0114] vehicle The generated task set and the number of elements are represented as and In the vehicle In the processing queue, the task is in line Previously and required in the vehicle The set of task indices processed above is represented as These tasks cover three types: vehicle-side local inference, vehicle-edge collaborative inference (intermediate feature offloading), and full offloading. Reasoning task The processing delay is calculated as .

[0115] Define binary variables The representative vehicle is in In one time slot, the vehicle The task needs to be completed. Unload to edge nodes Otherwise, it is 0. Define a binary variable. A value of 1 indicates that in the window Below, the model's exit Activated, otherwise 0. Set task. From the vehicle model's exit The amount of data uninstalled is Given ,Task From vehicles Model export Uploaded to base station or drone The delay is calculated as

[0116] (twenty one)

[0117] According to formula (15), the latency already covers the possible use of RIS, demonstrating the support of resource slicing for communication capabilities. (Task) From vehicles Model export Uploaded to satellite The delay is

[0118] (twenty two)

[0119] edge nodes The collected task set and the number of elements are represented as and At edge nodes In the processing queue, the task is in line The previous set of task indexes was represented as Let the edge nodes be defined. The full inference latency of the model is Edge nodes The processing latency of tasks in the processing queue is calculated as follows: .

[0120] The return latency is determined by the downlink rate and is affected by the communication resources reserved in the resource slice. (Base station or drone) The return task in The reasoning results to the vehicle The return delay is expressed as

[0121] (twenty three)

[0122] satellite Return mission The reasoning results to the vehicle The return delay is expressed as

[0123] (twenty four)

[0124] Given ,Task From generation to calculation result via nodes Delivery to vehicles The total delay was calculated as

[0125] (25)

[0126] Set a task The maximum tolerable delay is We define a binary variable with respect to the delay constraint.

[0127] (26)

[0128] Representative vehicle With edge nodes After the collaborative task unloading is completed, the task processing latency meets the requirements; otherwise, it is 0. Assume a base station... Detected vehicles The position error threshold is We define a binary variable regarding the trajectory prediction accuracy constraint.

[0129] (27)

[0130] Combining (26) and (27), define

[0131] (28)

[0132] This indicates that the vehicle's trajectory is basically consistent with the predicted results, and the task... The result is successfully delivered within the specified time; otherwise, it is 0. (In the slice window...) Below, the total reward for completing the task is defined as...

[0133] (29)

[0134] in, Indicates vehicle and edge nodes By dividing the model's entrance and exit Collaborate to complete the task The profit factor. The total loss function resulting from an incomplete task is defined as...

[0135] (30)

[0136] in, Indicates vehicle and edge nodes By dividing the model's entrance and exit Collaboration unfinished task The loss factor.

[0137] The long-term functional resource slice utility maximization problem centered on windows is modeled as follows:

[0138]

[0139] (31a) (31b)

[0140] (31c)

[0141] (31d)

[0142] (31e)

[0143] (31f)

[0144] question To maximize the completion rate of latency-sensitive inference tasks, decision-making elements such as collaborative inference, RIS phase configuration, and resource slice allocation are integrated. Constraint (31a) ensures that the total amount of resources allocated to communication and radar slices on each edge node does not exceed its available spectrum resources. Under the constraint (31b), the total number of all RIS units used must not exceed the total number of units held by the system. Furthermore, (31c) explicitly defines the RIS reflection path and its phase adjustment mechanism. Constraints (31d) and (31e) are sufficiently small constants. make sure

[0145]

[0146] The unloading decision is determined according to this rule. Under the constraint (31f), the uniqueness of the model's exit selection is ensured.

[0147] 4 Design of DRL Algorithm for RIS-ISCC Functional Slices

[0148] 4.1 Transformation of Markov Decision Processes

[0149] In order to solve the problem Resource allocation for functional slices is transformed into a Markov Decision Process (MDP). Within the MDP framework, the edge controller is abstracted as an agent that interacts with the dynamic environment comprised of vehicles, edge nodes, RIS (Resource Allocation System), and task queues. The system operates on a slice window basis, where the agent observes the environmental state, performs actions (i.e., resource allocation and task scheduling decisions), and receives corresponding reward feedback from the environment. The goal is to learn an optimal policy that maximizes long-term cumulative rewards. The specific elements of the MDP are described below:

[0150] State space: states Captured in the time window All key system information affecting decision-making, including the drone swarms that each base station can cooperate with, and historical performance (i.e., the previous time window). Task success rate, average latency, and average energy consumption), available resources (i.e. , as well as The environmental state during training is represented as...

[0151] (31)

[0152] Action space: Action Corresponding to the edge controller in each slice window The joint decision made internally includes: the number of spectrum resource blocks allocated to each base station and drone for communication and sensing. ), and allocate the number of reflection units on each RIS for communication and sensing ( Determine the unloading exit location of the vehicle-side inference model. And to determine the real-time phase matrix configuration for the relevant RIS. The resource allocation actions performed by the system during the training phase are...

[0153] (32)

[0154] Reward function: Reward Used to evaluate within a time window The immediate performance of actions taken within a window. It is calculated based on the processing results of all tasks within that window, with the goal of encouraging successful task completion and meeting latency requirements, while penalizing failed or timed-out tasks. Within each time window... At the end, a summary calculation is performed based on the results of all tasks completed within that time window. The reward function can be defined as:

[0155] (33)

[0156] Through the above MDP modeling, the complex coupled optimization problem of resource slicing and task scheduling is transformed into a sequence decision problem.

[0157] 4.2 Algorithm Design Based on Dual-Depth Q-Network

[0158] To address the aforementioned high-dimensional, continuous state-action space MDP problem, this section presents a deep reinforcement learning algorithm based on Double Deep Q-Network (DDQN). DDQN effectively mitigates the overestimation and training instability problems commonly encountered in Q-Learning under function approximation by introducing a target network and an experience replay mechanism.

[0159] As shown in Figure 4, the algorithm framework is deployed on the edge controller. It outputs a resource slicing policy based on the system input state. This policy network consists of multiple fully connected layers, using ReLU as the hidden layer activation function. Each slice window... Internally, the edge controller assigns parameters to each task based on the real-time accuracy requirements of the vehicle task, traffic density, and vehicle task queue status. Intelligently determine the optimal unloading strategy and dynamically adjust the unloading exit of the vehicle-side model. and RIS Reflection phase matrix The system dynamically adjusts the allocation of communication and sensing resource blocks between the base station and the UAV, as well as the operating mode of the RIS (such as the number of unit blocks used for communication and sensing, and the unit phase), using time windows as a cycle. Edge controllers are responsible for making decisions regarding these slices and task scheduling, and are deployed on edge nodes. Based on changes in the environment, the intelligent agent makes the optimal resource slicing plan.

[0160]

[0161] 4. Simulation Experiments and Result Analysis

[0162] This invention establishes a small-scale testbed to verify the proposed method. In the testbed, RK3588, Atlas 300I, and RTX 4090 were used to simulate vehicles and drones, base stations, and low-Earth orbit satellites, respectively. ResNet50 was deployed on each platform.

[25] The model was used to simulate inference for autonomous driving tasks, which were derived from the BDD100K Driving Videos dataset. Experiments were conducted in Python 3.6 and implemented using the PyTorch 1.7.1 deep learning framework. This paper uses PyTorch to build a convolutional neural network, and the original and target neural networks use the same architecture during model training. Fully connected layers are used between all hidden layers, with ReLU activation functions, and the final layer uses the softmax activation function.

[0163] The traffic flow data selected for the experiment came from the OpenITS open data platform. The arrival rate of vehicle-generated tasks followed a Poisson distribution. To ensure the diversity of task information in the simulation experiment, the delay constraint for each task was given randomly with probability within a limited range. Other parameters are shown in Table 1.

[0164] Table 1 Simulation Parameters

[0165]

[0166] To objectively evaluate performance, this experiment selected four representative ISCC functional slicing strategies for comparison, as shown in Table II. Since different resource slicing methods affect the latency of the inference task, this experiment selected three representative resource slicing methods for comparison, as shown in Table 3.

[0167] Table 2 lists the ablation classifications of the proposed methods.

[0168]

[0169] Table 3 Classification of benchmark methods

[0170]

[0171] 4.1 Impact of Training Rounds on Performance

[0172] Figures 5(a) and 5(b) show the success rates of the proposed method compared to three baseline methods. These graphs visually reflect the performance evolution of different models during training. The success rate curve of the proposed method (Proposed-1) shows a rapid upward trend followed by eventual convergence. After approximately 400-500 training epochs, its success rate reaches and stabilizes at a peak of 0.62-0.64. This indicates that the synergistic effect of RIS communication enhancement, inference model partitioning, and radar detection can significantly and efficiently improve the task completion success rate. To explore the contribution of each component, we conducted ablation experiments:

[0173] The Proposed-2 model converged to a success rate of 0.51–0.53. This indicates that the lack of UAV-assisted radar detection prevents the system from predicting network topology changes, making it difficult for the model to make optimal task offloading decisions in high-speed mobile environments, thus limiting performance improvement. The Proposed-3 model converged to a success rate of 0.49–0.54. This shows that a fixed inference model partitioning cannot adapt to dynamically changing network environments and task requirements. The Proposed-4 model performed the worst, achieving a success rate of only 0.46–0.49. This verifies that the absence of RIS communication enhancement has a significant impact on system performance, indicating that RIS resource partitioning is a key component for improving system utility.

[0174] Comprehensive comparison shows that the Proposed-1 method performs best among all baseline methods, fully demonstrating the importance of the synergy between RIS communication enhancement, inference model partitioning, and radar detection.

[0175] 4.2 The impact of resource quantity and delay constraints on task completion rate

[0176] The second set of experiments evaluated the impact of increasing the available spectrum resource blocks on the task success rate. Figure 6(a) shows the impact of increasing spectrum resources on the task success rate when the number of computing resources is fixed at 16. As the number of spectrum resource blocks increases from 10 to 20, the task success rate of each scheme shows a continuous upward trend. After the number of spectrum resource blocks increases to 16 or 18, the increase in task success rate of each scheme begins to slow down and tends to stabilize. Sufficient spectrum resources are a necessary condition for improving system performance, but not the only condition. Next, we examine the impact of increasing computing resources on performance when the number of sub-channels is fixed at 16. As shown in Figure 6(b), the task success rate rises rapidly in the initial stage. When the computing resources increase from 10 to 16, the task success rate shows a significant rapid increase. This indicates that when computing resources are scarce, it is the main bottleneck limiting the system's task processing capacity. However, when the number of computing resource blocks increases to 16, especially from R=16 to R=20, the system performance no longer shows a significant improvement. This is because the upper limit of the system's total task processing capacity is determined by both types of resources. When the number of tasks and communication bandwidth are saturated, simply increasing computing resources is unlikely to continue to improve system performance. Given fixed spectrum and computing resources, the resource allocation method proposed in this paper outperforms other methods, demonstrating that the resource allocation of the proposed method is more reasonable.

[0177] Figure 7 illustrates the probability distribution of task completion latency for this scheme and other schemes, with a fixed 16 blocks of spectrum and computing resources each. As shown in Figure 7, the task success rate continuously increases and eventually approaches saturation as the maximum latency requirement increases. When the task latency requirement is between 100ms and 300ms, the system must complete a large number of perception, communication, and inference processes within a very short time window, resulting in a low task success rate for all schemes, approximately 30% to 36%. When the latency requirement increases to 400ms to 600ms, the success rate of the proposed scheme rapidly increases to approximately 45% to 65%, showing a significant advantage over other schemes. With the latency requirement further relaxed to 700ms to 800ms, the task success rate of the proposed scheme reaches over 75%. When the maximum latency requirement increases to 9s or even higher, the task completion rate tends to stabilize at 75% to 80%. Under the same maximum task latency constraint, the proposed scheme is generally superior to other schemes.

[0178] 4.3 The impact of vehicle density on task completion rate

[0179] Figure 8 evaluates the impact of changes in traffic flow on task unloading success rate when the average vehicle task arrival rate is 0.5 tasks / s and the total resource volume is fixed. The vehicle density is adjusted from 0.1 vehicles / m². 2 Increased to 0.8 vehicles / m 2During the process, the success rate of all four schemes showed a downward trend. When the vehicle density was 0.1 vehicles / m²... 2 When the vehicle density increased to 0.8 vehicles / m², all four schemes achieved high success rates, with the scheme proposed in this paper achieving the highest task completion rate of approximately 85%. This is because the number of tasks is sparse, resulting in relatively abundant synesthetic computing resources. 2 Subsequently, the number of tasks surged, channel interference and resource contention intensified, and system performance deteriorated significantly. The success rate of the Baseline-2 scheme dropped to around 30%, while the proposed scheme maintained around 63%, demonstrating stronger robustness. Furthermore, the proposed scheme maintained the highest task success rate across all density levels, proving its superiority. Figure 9 evaluates the performance at a fixed vehicle density of 0.3 vehicles / m². 2 The impact of task latency constraints on task unloading success rate is investigated. As the maximum latency requirement increases, the task completion rate of all methods continuously improves. This is because increasing the maximum latency requirement requires fewer resources for task processing. Reducing the maximum latency requirement is a stress test of the system's processing capacity. Compared to other solutions, the proposed method improves the task completion rate through a more efficient resource slicing approach, exhibiting stronger robustness under extreme conditions.

[0180] References

[0181] [1]Qiu Y, Niu J, Zhu X, et al. Mobile edge computing in space-air-ground integrated networks: Architectures, key technologies and challenges[J]. Journal of Sensor and Actuator Networks, 2022, 11(4): 57.

[0182] [2]Gao B, Liu J, Zou H, et al. Vehicle-road-cloud collaborativeperception framework and key technologies: A review[J]. IEEE Transactions on Intelligent Transportation Systems, 2024.

[0183] [3]Wen D, Zhou Y, Li X, et al. A survey on integrated sensing,communication, and computation[J]. IEEE Communications Surveys & Tutorials,2024.

[0184] [4]Chen Y, Wang Y, Zhang J, et al. Resource allocation forintelligent reflecting surface aided vehicular communications[J]. IEEETransactions on Vehicular Technology, 2020, 69(10): 12321-12326.

[0185] [5]Pang X, Sheng M, Zhao N, et al. When UAV meets RIS: Expanding air-ground networks via passive reflection[J]. IEEE Wireless Communications,2021, 28(5): 164-170.

[0186] [6]Song W, Rajak S, Dang S, et al. Deep learning enabled RIS for 6Gintelligent transportation systems: A comprehensive study[J]. IEEETransactions on Intelligent Transportation Systems, 2022, 24(11): 12973-12990.

[0187] [7]Chen G, Qi S, Shen F, et al. Information-aware driven dynamic LEO-RAN slicing algorithm joint with communication, computing, and caching[J].IEEE Journal on Selected Areas in Communications, 2024, 42(5): 1044-1062.

[0188] [8]Cao H, Garg S, Kaddoum G, et al. Efficient resource allocation ofslicing services in softwarized space-aerial-ground integrated networks forseamless and open access services[J]. IEEE transactions on vehiculartechnology, 2023, 73(7): 9284-9295.

[0189] [9]Lyu F, Yang P, Wu H, et al. Service-oriented dynamic resourceslicing and optimization for space-air-ground integrated vehicular networks[J]. IEEE Transactions on Intelligent Transportation Systems, 2021, 23(7):7469-7483.

[0190]

[10] Zhou G, Zhao L, Zheng G, et al. Multiobjective optimization ofspace–air–ground-integrated network slicing relying on a pair of central anddistributed learning algorithms[J]. IEEE Internet of Things Journal, 2023, 11(5): 8327-8344.

[0191]

[11] Kak A, Akyildiz I F. Towards automatic network slicing for theinternet of space things[J]. IEEE Transactions on Network and ServiceManagement, 2021, 19(1): 392-412.

[0192]

[12] Rodrigues T K, Kato N. Network slicing with centralized anddistributed reinforcement learning for combined satellite / ground networks ina 6G environment[J]. IEEE Wireless Communications, 2022, 29(1): 104-110.

[0193]

[13] Li Z, Fu Y, Tian M, et al. Cooperative sensing, communication andcomputation resource allocation in mobile edge computing-enabled vehicularnetworks[J]. Journal of Information and Intelligence, 2024, 2(4): 339-354.

[0194]

[14] Wen D, Zhou Y, Li X, et al. A survey on integrated sensing,communication, and computation[J]. IEEE Communications Surveys & Tutorials,2024.

[0195]

[15] Zhao L, Wu D, Zhou L, et al. Radio resource allocation forintegrated sensing, communication, and computation networks[J]. IEEETransactions on Wireless Communications, 2022, 21(10): 8675-8687.

[0196]

[16] Xing H, Zhu G, Liu D, et al. Task-oriented integrated sensing,computation and communication for wireless edge AI[J]. IEEE Network, 2023, 37(4): 135-144.

[0197]

[17] Zhao J, Ren R, Zou D, et al. IoV-oriented integrated sensing,computation, and communication: System design and resource allocation[J].IEEE Transactions on Vehicular Technology, 2024, 73(11): 16283-16294.

[0198]

[18] Li X, Liu F, Zhou Z, et al. Integrated sensing, communication,and computation over-the-air: MIMO beamforming design[J]. IEEE Transactionson Wireless Communications, 2023, 22(8): 5383-5398.

[0199]

[19] Zhou J, Yang Q, Zhao L, et al. Mobility-aware computationoffloading in satellite edge computing networks[J]. IEEE Transactions onMobile Computing, 2024, 23(10): 9135-9149.

[0200]

[20] Shu F, Liu J, Lin Y, et al. Three high-rate beamforming methodsfor active IRS-aided wireless network[J]. IEEE Transactions on VehicularTechnology, 2023, 72(11): 15052-15056.

[0201]

[21] Peng H, Shen X. Deep reinforcement learning based resourcemanagement for multi-access edge computing in vehicular networks[J]. IEEETransactions on Network Science and Engineering, 2020, 7(4): 2416-2428.

[0202]

[22] Ye Q, Shi W, Qu K, et al. Joint RAN slicing and computationoffloading for autonomous vehicular networks: A learning-assistedhierarchical approach[J]. IEEE Open Journal of Vehicular Technology, 2021, 2:272-288.

[0203]

[23] Xu Y, Li Y, Zhang J A, et al. Joint beamforming for RIS-assistedintegrated sensing and communication systems[J]. IEEE Transactions onCommunications, 2023, 72(4): 2232-2246.

[0204]

[24] Shen H, Tian Y, Wang T, et al. Slicing-based task offloading inspace-air-ground integrated vehicular networks[J]. IEEE Transactions onMobile Computing, 2023, 23(5): 4009-4024.

[0205]

[25] C. Li, J. Wang, K. Jiang, C. Xiong and S. Wan, "MEOCI: ModelPartitioning and Early-exit point Selection Joint Optimization forCollaborative Inference in Vehicular Edge Computing," in IEEE Transactions onParallel and Distributed Systems, doi: 10.1109 / TPDS.2026.3652171.

Claims

1. A RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning, which constructs a resource slicing framework in a space-air-ground integrated vehicle-to-everything (SAGVN) network and implements resource slicing within this framework; base stations, UAVs, and satellites in the SAGVN serve as edge nodes; base stations refer to ground-based base stations, UAVs refer to UAV-borne base stations, and satellites refer to satellite-borne base stations; its characteristic is that... The SAGVN also includes multiple intelligent RIS units. First, a resource slicing framework is constructed: this framework is a RIS-ISCC functional resource slicing system based on "functional provision." Under this architecture, spectrum resources, intelligent reflector resources, and computing resources are uniformly virtualized into functional slices; RIS represents intelligent reflector, and ISCC represents integrated sensing, computing, and communication. At the sensing and communication layers of the slicing framework: the spectrum resources of the air-space-ground nodes are divided into two functional slices: sensing and detection, and communication. Correspondingly, the RIS reflector units are synchronously divided into two parts, used to enhance vehicle trajectory detection and task offloading, respectively, to enhance the vehicle trajectory detection capability of the sensing function and the task offloading capability of the communication function. At the computing layer of the slicing framework: isomorphic inference models are deployed on vehicles and edge nodes. The network structure entry / exit points of the inference model are synchronously divided, realizing the reconstruction of vehicle-edge inference functions, thereby indirectly adjusting the allocation of computing resources. Then, resource slicing is implemented: the system utility maximization problem is modeled as a constrained long-term cumulative optimization problem, and solved by a reinforcement learning algorithm for RIS-ISCC functional slicing; this algorithm, as a slicing algorithm, combines traffic flow prediction, network state, and task load to adaptively allocate spectrum resources and configure RIS functions, and determine the inference model segmentation strategy; the slicing algorithm design process is as follows: S1, construct a resource-aware vehicle trajectory detection model and a vehicle-side communication model, respectively, to quantify the impact of slicing on detection error and communication latency; S2, the long-term optimization problem with the objective of maximizing the completion rate of latency-sensitive inference tasks is modeled as a Markov decision process, and a reinforcement learning algorithm based on a dual deep Q network DDQN is designed; The algorithm dynamically and collaboratively optimizes the spectrum and the entry / exit selection of the RIS resource allocation and inference model based on ISCC functional slice status, vehicle trajectory and real-time load.

2. The RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning according to claim 1, characterized in that: The RIS-ISCC functional resource slicing framework is applied to RIS-ISCC-enabled vehicle-edge collaborative inference scenarios. In this framework, the edge controller determines resource slicing and task scheduling. The edge controller is deployed near the base station and is responsible for centralized management of the base station and drones. The ground workstation is dedicated to satellite task scheduling. System status information received by both the edge controller and the ground workstation is uniformly provided by the edge controller. It is assumed that time is divided into a series of slicing windows. The duration is Each slice window is divided into small-scale scheduling slots of equal length. The set of included scheduling slots is defined as follows: The sets of edge nodes and RIS are represented as follows: and ； 、 and These represent base stations, drones, and satellite sets, respectively. At the start of the slicing window, the edge controller formulates a slicing scheme based on collected historical mission information to segment the edge node set. and RIS collection Communication and sensing resources of mid-nodes, and segmented inference models; in time slots Within the control area, the edge controller integrates the usage of slice resources and changes in the network status of SAGVN, and uses RIS to batch schedule inference tasks generated by vehicles within the control area; the edge nodes, as collaborators, receive, process, and return the inference results to the target vehicles.

3. The RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning according to claim 2, characterized in that RIS-ISCC... In the ISCC functional resource slicing framework, at the communication and sensing levels: let the set of edge nodes be... ;node In the slice window The total number of available spectrum resource blocks is For the combination of base stations and drones Its spectrum resources are mapped into two types of functional slices, used for communication functions and sensing functions respectively, with a corresponding number of resource blocks. and And satisfy ； and Representing nodes respectively In the slice window The number of resource blocks used for communication functions and the number of resource blocks used for sensing functions; then, the RIS set The first in A RIS reflection unit Divided and Two parts, and satisfying ； and These represent the slice window respectively. The The number of reflection units used for communication and perception on each RIS; at the inference level: the vehicle and the edge node share a single unit... The pre-trained inference model consists of residual blocks, and the vehicle set is represented as... Within each slice window, the edge controller dynamically determines the inference exit point of the vehicle-side model based on the functional requirements of communication sensing and computing, as well as the SAGVN network status. And the entry point of the corresponding collaborator's edge node model. ,in Different exit options correspond to different inference function provision modes: 1) This indicates that vehicle-side full inference means that task inference relies entirely on onboard resources; 2) This indicates that the vehicle-side data is completely unloaded: the original task data is unloaded to the edge node for processing; 3) This represents vehicle-side collaborative reasoning: the vehicle completes part of the reasoning, and intermediate features are output from the vehicle model. entry point for unloading to edge nodes The edge nodes complete the remaining inference and return the calculation results.

4. The RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning according to claim 3, Its characteristic is that in S1, the vehicle-side communication model comprehensively considers the impact of the number of RIS units and phase configuration on channel gain and data transmission rate: assuming the first RIS is RIS It contains One reflective unit, wherein the reflective unit The phase is represented as The phase matrix of the reflecting surface is represented as follows: (1)RIS To edge nodes The vector representation of the channel is as follows Therefore, the vehicle via RIS To edge nodes The CC gain of the cascaded channel is calculated as (2) In the formula, Indicates vehicle To RIS The channel vector, where H represents the conjugate transpose; let the edge node... To the vehicle The distance is Then edge nodes The calculation results are sent back to the vehicle. At that time, the channel gain was calculated as (3) Among them, It is path loss. It is the path loss index. It is a small-scale Rayleigh fading channel; under maximum specific reflection (MRR), RIS Upper reflector unit The phase is determined as (4) In the formula, Represents a vector The The complex conjugate of each element, where * denotes complex conjugate; the gain of the equivalent channel is calculated as... (5) In the vehicle-edge collaborative reasoning scenario, after processing the task, the edge node, as a collaborator, must opportunistically cover the moving vehicle to complete the result delivery; RIS is only used to enhance uplink transmission; accordingly, a binary variable regarding uplink task scheduling is defined. A value of 1 indicates that in the time slot Get off the vehicle Dependent on RIS The task Unload to edge node Otherwise, it is 0; given ,vehicle Upload task To edge nodes The uplink rate at that time was quantized as In formula (6), Indicates the size of the resource block. Indicates vehicle Transmission power, Represents the noise power spectral density; in In the case of RIS assistance, equation (6) degenerates into a form that relies solely on RIS assistance. The pattern of the middle edge node; given ,vehicle Upload task To satellite The uplink speed at that time is (7) Edge nodes Return mission To the vehicle The downlink rate at that time is (8) Satellite Return mission To the vehicle The downlink rate at that time is （9） 5. The RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning according to claim 4, Its characteristic is that in S1, the addition of UAVs and RIS to the vehicle trajectory detection model enhances the detection performance of the base station from the dimensions of line-of-sight and perception range: assuming the vehicle... Located in Base station lie in ,vehicle Relative to base station The elevation angle was detected as (10) Base station Observed vehicles Azimuth angle is (11) Let the time difference between the radar transmitted wave and the echo be . Due to the two-way propagation delay of the signal, the base station With vehicles The detection range was estimated to be (12) Among them, It is the speed of light; assuming it is located at... RIS to base station and vehicles The straight-line distances are respectively and ,vehicle Signal via RIS Reaching the edge node The multipath distance was determined as (13) Based on equation (4) and the inertia rule, the vehicle exist The position of the moment is (14) Set edge nodes The received power is The resource block size is In a non-cooperative mode without RIS and drones, for the base station acting as a probe... Arrival vehicle The distance measurement error was calculated as (15) In the cooperative mode of RIS and UAV, a collaborative sensing resource detection model is constructed to characterize the impact of the number of sensing resources on the vehicle trajectory detection error. The method is as follows: assume the base station A group of drones can be dispatched drones With base station The distance between them is Then drone Due to distance and base station The attenuation of the generated sensing signal was calculated as (16) of which Reference distance; efficiency factor and It is introduced to reflect the enhanced vehicle trajectory detection performance of RIS; as an extension of Equation (15), the detection error in cooperative mode is calculated as (17) Among them, Represents the noise power spectral density; given 、 and The value in equation (7) is the same as The correlation is negative; assuming the vehicle With base station Angle measurement error meets The three-dimensional positioning error of the vehicle sensed by the base station is calculated as follows: (18) Visualize using equations (6), (17) and (18) ， and Impact on vehicle trajectory detection and communication transmission rate.

6. The RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning according to claim 5, Its characteristic is that in S2, problem modeling involves constructing an end-to-end latency model oriented towards offloading. This model comprehensively considers factors such as RIS configuration, vehicle trajectory, and node load, quantifying the impact of resource slicing schemes on task offloading latency. Based on the resource slice status (including communication / sensing resource block partitioning, RIS phase configuration, and inference model exit position), the offloading path is dynamically selected to achieve successful task delivery under latency constraints. The goal of resource slicing is to achieve the following in each slice window: Inside, it provides the optimal configuration of communication, sensing, and computing resources for downstream task offloading; the edge controller is responsible for optimizing the slicing scheme, including the allocation of communication and sensing resources, RIS unit partitioning, RIS phase configuration, and inference model exit position; Task scheduling is executed by the DDPG policy deployed on the edge controller under a given slicing scheme; therefore, the resource slicing optimization problem is independent of the specific scheduling process, and its goal is to provide the optimal resource allocation for the DDPG policy to maximize the long-term task success rate of the system; for tasks , set up vehicles Execute the task Full inference latency is Edge nodes The inference model's full inference latency is Then the task Reasoning to the reasoning model The processing delay of each export is approximated as (19) Task From the reasoning model The processing delay for each entry point to continue inference until the end is calculated as follows: (20) Vehicles The generated task set and the number of elements are respectively represented as and In the vehicle In the processing queue, the task is in order. Previously and required in the vehicle The set of task indices processed above is represented as These tasks cover three types: vehicle-side local inference, vehicle-edge collaborative inference with intermediate feature offloading, and full offloading; vehicles Reasoning task The processing delay is calculated as In the formula, Indicates in the vehicle In the processing queue, the task is in order. Previous task index set; Define binary variables The representative vehicle is in In one time slot, the vehicle The task needs to be completed. Unload to edge nodes Otherwise, it is 0; Define binary variables A value of 1 indicates that in the window Below, the exit of the reasoning model Activated, otherwise 0; Set task From the vehicle model's exit The amount of data uninstalled is Then given ,Task From vehicles Model export Uploaded to base station or drone The delay is calculated as （21） According to formula (15), the delay covers all possible scenarios where RIS can be used, demonstrating the support of resource slicing for communication capabilities; task From vehicles Model export Uploaded to satellite The delay is (22) Edge nodes The collected task set and the number of elements are represented as and ； At the edge node In the processing queue, the task is in order. The previous set of task indexes was represented as Set edge nodes The total inference latency of the inference model is Edge nodes The processing latency of tasks in the processing queue is calculated as follows: The result return delay is determined by the downlink rate and is affected by the communication resources reserved in the resource slice; base station or drone. The return task in The reasoning results to the vehicle The return delay is expressed as (23) Satellite Return mission The reasoning results to the vehicle The return delay is expressed as In equation (24), Indicates vehicle In the time slot Upload task To edge nodes Decision variables; given ,Task From generation to calculation result via nodes Delivery to vehicles The total delay was calculated as In equation (25), Indicates in window vehicle Upload task To edge nodes Decision variables; assuming the task The maximum tolerable delay is Define a binary variable with respect to the delay constraint. （26） Representative vehicle With edge nodes After the collaborative task unloading is completed, the task processing latency meets the requirements; otherwise, it is 0. (Base station configuration) Detected vehicles The position error threshold is Define a binary variable regarding the trajectory prediction accuracy constraint. (27) Combining (26) and (27), define （28） This indicates that the vehicle's trajectory is basically consistent with the predicted results, and the task... The result was successfully delivered within the specified time. Otherwise, it is 0; in the slice window Below, the total reward for completing the task is defined as... (29) Among them, Indicates vehicle and edge nodes By dividing the inference model into entry and exit points Collaborate to complete the task The profit factor; the total loss function resulting from unfinished tasks is defined as... (30) Among them, Indicates vehicle and edge nodes By dividing the inference model into entry and exit points Collaboration unfinished task The loss factor; the long-term functional resource slice utility maximization problem centered on the window is modeled as: ， (31a) (31b) (31c) (31d) (31e) (31f) Problem To maximize the completion rate of latency-sensitive inference tasks, decision-making elements including collaborative inference, RIS phase configuration, and resource slice allocation are integrated. Constraint (31a) ensures that the total amount of resources allocated to communication and radar slices on each edge node does not exceed its available spectrum resources. Under constraint (31b), the total number of all RIS units used must not exceed the total number of units held by the system. Constraint (31c) explicitly defines the RIS reflection path and its phase adjustment mechanism. Constraints (31d) and (31e) contain sufficiently small constants. make sure The unloading decision is determined according to this rule; under the constraint of constraint (31f), the uniqueness of the exit selection of the inference model is ensured.

7. The RIS-ISCC functional slicing method for vehicle-edge collaborative reasoning according to claim 6, characterized in that: In step S2, the slicing algorithm design process includes: S21) The Markov decision process is transformed into a problem to be solved. Resource allocation for functional slices is transformed into a Markov Decision Process (MDP). Within the MDP framework, the edge controller is abstracted as an agent that interacts with the dynamic environment comprised of vehicles, edge nodes, RIS (Resource Information System), and task queues. Using slice windows as the basic decision cycle, within each window, the agent observes the environmental state, executes actions (resource allocation and task scheduling decisions), and receives corresponding reward feedback from the environment. The goal is to learn an optimal policy that maximizes long-term cumulative rewards. Actions are resource allocation and task scheduling decisions. The specific elements of the MDP are described below: State Space: States Captured in the time window All key system information affecting decision-making, including the drone swarms that each base station can cooperate with, historical performance, and available resources. ， as well as Historical performance refers to the performance of the previous time window. Task success rate, average latency, and average energy consumption; the environmental state during training is represented as (31) Action space: Action Corresponding to the edge controller in each slice window The joint decisions made internally include: the number of spectrum resource blocks allocated to each base station and drone for communication and sensing. The number of reflection units used for communication and sensing on each RIS is determined. Determine the unloading exit location of the vehicle-side inference model. And to determine the real-time phase matrix configuration for the relevant RIS. The resource allocation actions performed during the training phase are: (32) Reward function: Reward Used to evaluate within a time window The immediate performance of actions taken within a window is calculated based on the processing results of all tasks within that window. The goal is to encourage successful task completion and meet latency requirements, while penalizing failed or timed-out tasks. Within each time window... At the end, the results of all tasks processed within the time window are summarized and calculated; the reward function is defined as: (33) Through MDP modeling, the coupled optimization problem of resource slicing and task scheduling is transformed into a sequence decision problem; S22) The slicing algorithm based on DDQN is designed to solve the MDP problem proposed in step S21 by using a deep reinforcement learning algorithm based on DDQN as the slicing algorithm; the slicing algorithm is deployed on the edge controller, which outputs the resource slicing policy according to the input state; the policy network architecture of the slicing algorithm consists of multiple fully connected layers, using ReLU as the activation function of the hidden layer; according to the resource slicing policy, in each slicing window Internally, the edge controller assigns parameters to each task based on the real-time accuracy requirements of the vehicle task, traffic density, and vehicle task queue status. Decisions are made on the optimal unloading strategy, and the unloading exit of the vehicle-side model is dynamically adjusted. and RIS Reflection phase matrix Based on the resource slicing strategy, the allocation of communication and sensing resource blocks between base stations and drones, as well as the working mode of the RIS (Resource Sharing System), are dynamically adjusted according to time windows. The edge controller is responsible for making decisions regarding these slices and task scheduling, and is deployed on edge nodes. Based on changes in the environment, the optimal resource slicing scheme is determined.

Images