Digital twinborn assisted collaborative awareness and resource allocation method in Internet of Vehicles

Abstract

The invention relates to a digital twin-assisted collaborative awareness and resource allocation method in the Internet of Vehicles, and belongs to the technical field of mobile communication. Aiming at the problems of high sensing delay and insufficient cooperative stability caused by low broadcast mode spectrum efficiency and poor static resource allocation adaptability in the Internet of Vehicles, a digital twin Internet of Vehicles scene is constructed, a double-layer sensing area dynamic triggering mechanism is designed, and the cooperative stability of the Internet of Vehicles is improved based on Cramer-Rao bound quantitative sensing precision and bandwidth constraint. And screening high-stability cooperative nodes by adopting an improved K-means algorithm, and establishing a deep reinforcement learning driven joint optimization model to realize dynamic allocation of communication bandwidth, power and computing resources. According to the method, the redundant communication traffic is reduced by 30%, the target positioning error is reduced by 25%, the cooperative link stability is improved by 40%, the system utility is improved by 18-22%, the task processing time delay is shortened to 65% of that of a traditional scheme, and the cooperative sensing efficiency and the resource utilization rate in the Internet of Vehicles environment are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of mobile communication technology and relates to a collaborative perception and resource allocation method in a digital twin-assisted vehicle network. Background Art

[0002] In the highly dynamic environment of the Internet of Vehicles, existing collaborative perception and resource allocation methods have technical defects such as perception blind spots, resource competition, large communication redundancy, low positioning accuracy, node instability, low static efficiency of resource allocation, and high task latency. There is an urgent need for a collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins. Summary of the Invention

[0003] In view of this, the object of the present invention is to provide a collaborative perception and resource allocation method in a digital twin-assisted Internet of Vehicles.

[0004] In order to achieve the above object, the present invention provides the following technical solutions:

[0005] A method for allocating synaesthesia computing resources and deploying vehicle digital twins in an Internet of Vehicles scenario, comprising:

[0006] S1: Build a digital twin-assisted connected vehicle scenario, covering communication, computing, and perception functions;

[0007] S2: Design a vehicle collaborative perception mechanism with dual-layer perception area division;

[0008] S3: Construct a model to quantify the constraint relationship between vehicle perception accuracy and communication bandwidth based on the Cramer-Rao bound;

[0009] S4: Build a vehicle clustering method based on digital twins to ensure the stability of the collaborative communication link;

[0010] S5: A dual-sensing task offloading scheme based on collaborative sensing strategy is proposed, and the task offloading scheme is selected according to the collaborative sensing mode;

[0011] S6: Use a multi-agent deep reinforcement learning algorithm to solve the system joint utility maximization optimization model of joint bandwidth, power and perception strategy.

[0012] Furthermore, in S1, the proposed digital twin-assisted IoV scenario consists of three layers: the physical layer, the digital twin layer, and the control layer. The physical entities include the demand vehicle, the service vehicle, and the roadside unit equipped with edge computing resources. Assume that there are M roadside units equipped with edge computing resources and service vehicles evenly distributed in the area, and the set is represented by ES = {es1, es2, ..., es MThe set of demand vehicles is denoted as N = {1, 2, ..., n}. Demand vehicles represent vehicles that require collaborative sensing, while service vehicles are responsible for providing collaborative sensing services to demand vehicles. Each vehicle is equipped with integrated intersensory sensing equipment to sense the status of its surrounding area and establish connections with roadside units or other vehicles via communication links. Let the set of time slots be K = {k1, k2, ..., k}.

[0013] Furthermore, in S2, the vehicle cooperative perception mechanism with two-layer perception area division is designed. At the beginning of each time slot k, the communication nodes exchange basic information by sending lightweight cooperative perception information. This mechanism divides the perception range between the demand vehicle and the edge service node. Assuming that the perception task is for the vehicle to perceive its perception area O S and region of interest O I For simplicity, let O S and O I The radius is R S and R I The circular area of ​​interest satisfies the range constraint between the sensitive area and the area of ​​interest. In order to represent the different perception modes of vehicle n, the binary variable set a nm (k) represents the cooperative perception strategy of vehicles in each time slot k:

[0014]

[0015] In the formula, different values ​​correspond to different perception modes. For example, when a nm = 0, it means that the vehicle has independent perception and does not perform cooperative perception; when a nm =1, it means that the vehicle performs collaborative perception, and the collaborative perception object is the service vehicle or RSU in the scene.

[0016] Furthermore, in S3, the process of constructing the constraint relationship model between the perception accuracy and communication bandwidth of the quantified vehicle based on the Cramer-Rao bound is as follows: Let the perception target o∈O in the perception interest area I The perception result is Ψ(o), where Ψ(o) = 1 indicates successful perception of the target, and Ψ(o) = 0 indicates perception failure. To simplify the expression, the generalized Cramer-Rao bound (CRB) is used to represent the estimated perception accuracy of vehicle n for the perceived target o during the perception process:

[0017]

[0018] Where, represents the distance between vehicle n and the perceived target o; represents the arrival angle between vehicle n and the perceived target o; Indicates perceived quality; W NNis the zero-to-zero beamwidth when the vehicle receives the perceived target information; is the transmission power when vehicle n perceives target o; b rms,o In order to unify the bandwidth expression, the signal spectrum is defined as S(f) = sinπfT p / πf, where the bandwidth is limited to a finite value By transmitting this waveform we can get Approximate reasonable values ​​for :

[0019]

[0020] Where, T pulse represents the pulse bandwidth; is the perceived bandwidth. We can get:

[0021]

[0022] The above CRB minimization problem is equivalently transformed into its inverse maximization problem, where the perception quality of vehicle n when perceiving target o is It can be expressed as:

[0023]

[0024] Where λ d and λ θ represents the normalization factor.

[0025] make It represents the perception quality function of vehicle n in time slot k when perceiving target o, the perception quality is related to the transmission power p and the perception bandwidth allocation b s Therefore, the weighted sum of the individual target perception gains of vehicle n in the perception area can be obtained as follows:

[0026]

[0027] The perception targets within the perception area of ​​interest need to obtain perception information through collaborative perception. Similarly, the weighted sum of the perception quality of objects outside the perception area is expressed as:

[0028]

[0029] Therefore, the gain of vehicle n from cooperative sensing in time slot k can be calculated as:

[0030]

[0031] Furthermore, in S4, a vehicle clustering method based on digital twins is constructed to ensure the stability of the cooperative communication link. Specifically, the core idea is to evaluate the motion consistency of two vehicles to measure their proximity. The specific definitions of the above parameters are as follows:

[0032] (1) Direction similarity: By checking whether the directions of the two vehicles are the same, use Indicates the direction similarity of the two vehicles, as follows:

[0033]

[0034] Where, Indicates that service vehicle n' has the same direction as vehicle n. Conversely, Indicates that vehicle n' is moving in the opposite direction to vehicle n.

[0035] (2) Speed ​​similarity: The speed similarity between vehicles n and n' is calculated by comparing whether the speeds of the two vehicles are the same. Represents the similarity of the speeds of two vehicles as follows:

[0036]

[0037] Use w n,n' represents the social trust value between the demand vehicle n and the service vehicle n'. From the perspective of the demand vehicle n, the social trust value between it and the service vehicle n' is as follows:

[0038]

[0039] Where, and are tuning parameters, representing the weights of direction and speed similarity, w n,n' The larger the value of , the higher the trust value between the demand vehicle n and the service vehicle n'. On the contrary, the smaller the trust value between vehicle n and vehicle n'. Let H = [w n,n' ]N×N is the social trust matrix of all vehicles in the area, where N represents the number of vehicles in the area. By clustering vehicles, the link links between vehicles in the same aggregation group will be more stable, and the efficiency of perception data processing will also be improved. In order to guide the aggregation of edge service vehicles, by deploying digital twin nodes in RSU, each RSU collects the computing power and communication topology of vehicles in its surrounding environment, and then transmits this data through wired transmission. At the digital twin layer, the digital twin service virtually maps the physical system, including physical entities, hardware details, real-time status, etc. The digital twin layer collects and analyzes the status of objects through real-time monitoring, dynamic adjustment, etc. to improve control and management efficiency. The digital twin model of the physical layer can be modeled as:

[0040]

[0041] Where, They represent the virtual mapping of vehicles, RSUs, and edge servers respectively.

[0042] In order to optimize the selection of cooperative objects, improve the stability of the communication link, and enhance the processing efficiency of perception data, an improved K-means algorithm based on the gravity model is proposed. The mass of the vehicle n requesting cooperative perception is defined as:

[0043]

[0044] Where C n Indicates that the vehicle completes the task n The amount of CPU cycles required; f n is the number of CPU cycles per second of the vehicle, i.e., computing resources. By applying the mass model to the gravity model, the social gravity between the demand vehicle n and the service vehicle n' is defined as:

[0045]

[0046] Where γ1, γ2, and γ3 represent the modeling parameters for the relative correlation of resources, social trust, and communication considered in digital twin network clustering. The numerator uses a maximum function to quantify the correlation between two vehicles, while the denominator replaces the original distance based on social trust and communication rate.

[0047] Furthermore, in S5, the specific steps of the binary perception task offloading scheme based on the collaborative perception strategy are as follows: when the perception target is within the vehicle perception range, that is, when a nm = 0, the vehicle performs the perception task autonomously without the need for collaborative perception and data fusion. nm =1, the vehicle requests cooperative sensing service to obtain information outside the sensing range.

[0048] (1) Local computing: In time slot k, when a vehicle chooses to execute a task locally, the local computing resources of the vehicle are represented by f n (k), the local computation delay can be expressed as:

[0049]

[0050] (2) Edge computing model: When local computing resources are insufficient to support the task completion, or when the sensing target is out of the sensing range, consider offloading the task to the edge service node for processing. Data processing delay includes task upload delay, data fusion delay, task calculation delay, and result downlink transmission delay, which are represented by t up, t mix , t comp , t down When the vehicle chooses to offload the task to the RSUm, the task upload delay model is as follows:

[0051]

[0052] After the perception point cloud data is uploaded to the RSU, data fusion is required. The size of the fused data is The delay of RSU data fusion is expressed as:

[0053]

[0054] Where Ω represents the ratio of data fusion delay to data size. After data fusion, RSU needs to perform computational processing. The computational processing delay can be expressed as:

[0055]

[0056] Where f' m represents the computing resources allocated by RSU to vehicle n, which should satisfy 0≤f' m <f m The downlink transmission delay model of the fused data task result can be expressed as:

[0057]

[0058] Where δ is the ratio of the output data size to the input data size, and δ<<1. Therefore, the data downlink transmission delay can be ignored. Therefore, in the collaborative sensing process between the demand vehicle and the edge service node RSU, the total processing delay of the sensing data can be obtained based on the above discussion:

[0059]

[0060] Next, we analyze the computational task processing model for collaborative perception between the demand vehicle and the service vehicle. When a vehicle chooses to offload the task to the collaborative vehicle n', the task upload delay model can be expressed as:

[0061]

[0062] The data fusion delay can be expressed as:

[0063]

[0064] Where, D n' (k) represents the size of the perception data of vehicle n' at time slot k. The computational delay of offloading to cooperative vehicle n' for computation can be expressed as:

[0065]

[0066] Where f' n' (k) is the computing resource provided by service vehicle n' to demand vehicle n in time slot k, which should satisfy 0≤f' n' <f n' Therefore, during the collaborative perception process between vehicle n and vehicle n', the total processing delay of the perception data is:

[0067]

[0068] In summary, when the perception task of the demand vehicle is executed on the edge service node, the total delay of task execution can be expressed as:

[0069]

[0070] By comprehensively considering the processing delay of the perception task in both local computing and collaborative computing modes, the total computing delay in different perception modes can be obtained. Specifically, it can be expressed as:

[0071]

[0072] Furthermore, in S6, the system joint utility maximization optimization model is specifically expressed as follows:

[0073]

[0074] Where p and b represent the power and bandwidth resource allocation vectors respectively; a represents the set of cooperative sensing strategies within the time slot set. Constraint C1 represents the uplink bandwidth resource allocated to the vehicle. Must not exceed the total uplink bandwidth resource b of RSU c Constraints C2 and C3 represent the collaborative sensing decision a of the demand vehicle in each time slot nm , specifies whether the perception data is calculated locally or processed in M ​​edge service nodes, N represents the set of vehicles; constraint C4 represents the local computing delay of the vehicle Edge collaborative computing latency The maximum tolerable delay of the task must not exceed T n ; Constraint C5 represents the computing resources f' allocated by RSU to the demand vehicle m Must not exceed its own available computing resources f m ; Constraint C6 represents the computing resources f' allocated by the cooperative vehicle to the demand vehicle n' Must not exceed the total amount of computing resources available to it f n' ; Constraint C7 represents the vehicle's perceived power p n Not less than the minimum perceived power P min , and must not exceed the maximum perceived power Pmax ; Constraint C8 indicates that the two weight factors φ1 and φ2 are used to match the system utility between perception and calculation, and their sum is 1; Constraint C9 indicates that the perception occupies bandwidth b s and communication bandwidth b c The sum shall not exceed the total bandwidth resource B.

[0075] Furthermore, S6 uses a deep reinforcement learning algorithm to solve the optimization problem of combining bandwidth, power, and collaborative sensing strategies. First, the optimization model is modeled as a Markov decision process, treating all demand vehicles as a single agent. The detailed definition of the Markov decision process elements is as follows:

[0076] (1) State space s: includes the perception accuracy utility, the channel gain during perception and communication, and the changes in the collaborative perception strategy, expressed as

[0077] (2) Action space: There are three key elements in the action space, namely power allocation strategy, bandwidth allocation strategy and cooperative sensing switching strategy, which are expressed as a(k) = {p(k), b(k), a(k)}.

[0078] (3) Reward function: The reward function is composed of the overall system utility minus the penalty for violation of bandwidth and delay constraints; it is specifically expressed as:

[0079]

[0080] Where η1, η2, η3 are utility scaling coefficients, task processing delay penalty coefficients, and bandwidth violation penalty coefficients, respectively; U represents system utility, represents the penalty for delay constraint violation; Indicates bandwidth violation penalty.

[0081] The TD3 algorithm is a deep reinforcement learning algorithm that uses two Critic networks and an Actor network and its corresponding target network; the Critic network is mainly used to estimate the action-state value function and assist in policy updates by optimizing Q estimation. The Critic network uses two key deep neural networks to reduce the overestimation bias during training. Overestimation and noise accumulation will produce an overestimation bias of the Q value, affecting the stability of policy learning. To solve this problem, TD3 selects the lower value from the two outputs of the Critic network. As the target value. TD3 trains the Critic network by minimizing the square loss of the time difference error. is minimized as a cost function:

[0082]

[0083] Where, Represents the parameterized state-action value function Q, with parameter θ. The target value of function Q is It can be expressed as:

[0084]

[0085] The actions used in the Critic network are defined as:

[0086]

[0087] In the formula, noise Follow the clipped normal distribution clipping This shows is a random variable, following and belongs to the interval [-c,c].

[0088] The Actor network is derived from the Critic network The output is the action, which is expressed as a k =ν κ (s k )+υ, κ represents the parameters of the Actor network, ν κ Represents the output of the Actor network, which is deterministic and continuous. To encourage exploration, a normal distribution is added. The parameter ν is adjusted by minimizing the following cost function:

[0089]

[0090] Where, d ν (s) represents the state distribution. Gradient Used to update the parameter ν. The parameters of the Actor network can be updated as follows:

[0091]

[0092] The beneficial effects of the present invention are as follows: the present invention realizes dynamic mapping of vehicle behavior characteristics and network status through digital twin technology, and combines the dual-layer perception area trigger mechanism to effectively reduce the amount of redundant communication data by more than 30%. The perception quality modeling based on the Cramer-Rao bound accurately quantifies the nonlinear relationship between communication bandwidth and perception accuracy, reducing the mean square error of target positioning by about 25%. The improved K-means clustering algorithm integrates vehicle kinematic characteristics and resource status to screen out high-trust collaborative nodes with 40% improved link stability, significantly reducing communication interruptions caused by node switching. The joint optimization model driven by deep reinforcement learning breaks through the limitations of static resource allocation, achieves synergistic gains in spectrum efficiency and computing resource utilization, and improves the overall utility of the system by 18%-22%. Through the dynamic adaptation of perception mode and task offloading strategy, the average task processing delay is reduced to 65% of the traditional solution while ensuring the integrity of perception coverage. This solution overcomes the technical difficulties of perception blind spot compensation and resource competition balance in high-dynamic scenarios, and provides a scalable solution for real-time collaborative perception in the Internet of Vehicles environment.

[0093] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0095] Figure 1 A diagram of the vehicle network architecture assisted by digital twins;

[0096] Figure 2 Schematic diagram of collaborative sensing mode;

[0097] Figure 3 Flowchart of the improved K-means algorithm for digital twin assistance;

[0098] Figure 4 Figure 2 is the block diagram of the proposed TD3 algorithm. DETAILED DESCRIPTION

[0099] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0100] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0101] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0102] In this embodiment, a digital twin-assisted vehicle network architecture diagram is shown in FIG. Figure 1 As shown, the collaborative perception mode is as follows Figure 2 As shown, the digital twin-assisted K-means algorithm process is as follows Figure 3 As shown, the block diagram of the solution algorithm TD3 is as follows Figure 4 As shown, implementing this method specifically includes the following steps:

[0103] S1: Build a digital twin-assisted connected vehicle scenario, covering communication, computing, and perception functions;

[0104] S2: Design a vehicle collaborative perception mechanism with dual-layer perception area division;

[0105] S3: Construct a model to quantify the constraint relationship between vehicle perception accuracy and communication bandwidth based on the Cramer-Rao bound;

[0106] S4: Build a vehicle clustering method based on digital twins to ensure the stability of the collaborative communication link;

[0107] S5: A dual-sensing task offloading scheme based on collaborative sensing strategy is proposed, and the task offloading scheme is selected according to the collaborative sensing mode;

[0108] S6: Use a multi-agent deep reinforcement learning algorithm to solve the system joint utility maximization optimization model of joint bandwidth, power and perception strategy.

[0109] In the above S1, if Figure 1 As shown, the digital twin-assisted Internet of Vehicles scenario involves three functions: communication, perception, and computing. The network architecture consists of three layers: the physical layer, the digital twin, and a control layer. The physical layer includes all entities, such as vehicle entities, RSUs, computing entities, and perception targets. The digital twin layer is a real-time dynamic mapping of the physical layer, which can interact with the physical layer to enhance data analysis and intelligent decision-making capabilities. The control layer is responsible for performing resource optimization and scheduling based on the state information of the digital twin layer, including the allocation of synergistic resources in synergistic integrated devices, the switching of vehicle collaborative perception strategies, and the coordination and management of physical layer resources. Assume that there are M RSUs and service vehicles equipped with edge computing resources evenly distributed in the area, and the set is represented as ES = {es1, es2, ..., es M}. The set of demand vehicles is represented as N = {1, 2, ..., n}. Demand vehicles represent vehicles that require collaborative perception, and service vehicles are responsible for providing collaborative perception services for demand vehicles. Each vehicle is equipped with an integrated sensory device to perceive the status of its surrounding area and establish a connection with the RSU or other vehicles through a communication link. Let the time slot set be K = {k1, k2, ..., k}. In each time slot, the vehicle digital twin obtains the status information of its vehicle entity based on the collected status information, including the perception status, location information, and computing resource allocation, and makes collaborative perception decisions based on this.

[0110] In the above S2, if Figure 2As shown, the demand vehicle needs to offload the perception task to detect the perception target within its area of ​​interest. The service vehicle and RSU with sufficient idle resources act as edge service nodes to provide collaborative perception and computing support for the demand vehicle. Each demand vehicle makes an offloading decision based on the real-time road status. Specifically, at the beginning of each time slot k, the communication node exchanges basic information, including location, vehicle status, and RSU status, by sending lightweight collaborative perception information. Since the collaborative perception information is relatively lightweight, its transmission overhead can be ignored. When the demand vehicle decides to calculate the perception task locally, it operates in independent perception mode. On the contrary, if it decides to offload the perception task, the demand vehicle sends its sensing data to the edge service node, which can be a static RSU or a mobile service vehicle, and switches to CP mode. The edge service node performs fusion calculations on the data offloaded by the demand vehicle and returns the enhanced detection results to the demand vehicle. The collaborative perception mode switching step mainly includes the following steps:

[0111] S21: Divide the perception range for the demand vehicle and the edge service node. Assume that the perception task is for the vehicle to perceive its perception area O. S and region of interest O I For simplicity, let O S and O I The radius is R S and R I The circular area of ​​interest satisfies the range constraint between the sensitive area and the area of ​​interest. Since vehicles are required to pay more attention to the targets around them, the collaborative perception gain under different perception modes is studied in detail.

[0112] S22: In order to represent the different perception modes of vehicle n, a binary variable set a nm (k) represents the cooperative perception strategy of vehicles in each time slot k:

[0113]

[0114] In the formula, different values ​​correspond to different perception modes. For example, when a nm = 0, it means that the vehicle has independent perception and does not perform cooperative perception; when a nm =1, it means that the vehicle performs collaborative perception, and the collaborative perception object is the service vehicle or RSU in the scene.

[0115] S23: In each time slot, the vehicle can only choose one mode for perception, so a nm The constraints must also be met:

[0116]

[0117] In S3 above, an early collaborative perception scheme based on point cloud data fusion in unicast mode is proposed. For perception information fusion, radar point cloud data has the advantage of depth information and is easier to fuse after coordinate transformation. The collaborative stage of collaborative perception is divided into early, mid-term and late fusion. Late fusion has limited perception gain and may not detect objects even after fusion. During mid-term fusion, the data features extracted by different vehicles may lack universality. In contrast, early fusion not only improves accuracy, but also allows vehicles to perform distributed fusion and calculation of raw sensor data. Therefore, early collaborative perception based on point cloud data is selected. The collaborative perception quality solution mainly includes the following steps:

[0118] S31: Assume that the point cloud data generated by the vehicle based on its sensor at time slot k is:

[0119] P k =(x k ,y k ,z k ,r k )

[0120] Where x k ,y k ,z k Represents the coordinates and intensity value of the reflection point; r k Indicates the intensity value of the reflection point.

[0121] S32: When the data is unloaded to the edge service node, the CAV point cloud data is transformed by the rotation matrix R:

[0122] R=R z (α r )R y (β r )R x (θ r )

[0123] Where R x (θ r ),R y (β r ),R z (β r ) represent the basic rotation matrices of the three-dimensional coordinate axes x, y, and z respectively. r ,β r ,θ r They represent the differences in yaw angle, roll angle, and arrival angle between the vehicle and the perceived object, respectively.

[0124] S33: Let the perception target o∈O in the perception interest area IThe perception result is Ψ(o), where Ψ(o) = 1 indicates successful perception of the target, and Ψ(o) = 0 indicates a perception failure. In many application scenarios, CRB is a key performance metric used to measure the theoretical limit of mean square error. To simplify the expression, a generalized form is used to represent the estimated CRB of vehicle n for the perceived target o during the perception process:

[0125]

[0126] Where, represents the distance between vehicle n and the perceived target o; represents the arrival angle between vehicle n and the perceived target o; Indicates perceived quality; W NN is the zero-to-zero beamwidth when the vehicle receives the perceived target information; is the transmission power when vehicle n perceives target o; b rms,o Indicates the allocated bandwidth when sensing target o.

[0127] S34: Unify the bandwidth expression and define the signal spectrum as S(f) = sinπfT p / πf, where the bandwidth is limited to a finite value By transmitting this waveform Approximate reasonable values ​​for :

[0128]

[0129] Where, T pulse represents the pulse bandwidth; is the perceived bandwidth. We can get:

[0130]

[0131] S35: The above CRB minimization problem is equivalently transformed into a maximization problem of its inverse, where the perception quality of vehicle n when perceiving target o is expressed as:

[0132]

[0133] Where λ d and λ θ represents the normalization factor.

[0134] S36: Order It represents the perception quality function of vehicle n in time slot k when perceiving target o, the perception quality is related to the transmission power p and the perception bandwidth allocation b s Therefore, the weighted sum of the individual target perception gains of vehicle n in the perception area can be obtained as follows:

[0135]

[0136] The perception targets within the perception area of ​​interest need to obtain perception information through collaborative perception. Similarly, the weighted sum of the perception quality of objects outside the perception area is expressed as:

[0137]

[0138] S37: The gain of vehicle n from cooperative sensing in time slot k can be calculated as:

[0139]

[0140] In S4 above, a social relationship model based on vehicle kinematic characteristics is constructed to quantify the matching degree between vehicles and express it as a social trust matrix. The social trust value between the demand vehicle and the service vehicle is mainly calculated by the direction similarity. and speed similarity The core idea is to evaluate the motion consistency of two vehicles to measure their proximity. The digital twin-based vehicle clustering in S4 mainly includes the following steps:

[0141] S41: Calculate direction similarity: By checking whether the directions of the two vehicles are the same, use Indicates the direction similarity of the two vehicles, as follows:

[0142]

[0143] Where, indicates that the cooperative vehicle n' has the same direction as vehicle n. Conversely, Indicates that vehicle n' is moving in the opposite direction to vehicle n.

[0144] S42 calculates speed similarity: calculates the speed similarity between vehicles n and n' by comparing whether the speeds of the two vehicles are the same, using Represents the similarity of the speeds of two vehicles as follows:

[0145]

[0146] Use w n,n' represents the social trust value between the demand vehicle n and the service vehicle n'. From the perspective of the demand vehicle n, the social trust value between it and the service vehicle n' is as follows:

[0147]

[0148] Where, and are tuning parameters, representing the weights of direction and speed similarity, w n,n'The larger the value of , the higher the trust value between the demand vehicle n and the service vehicle n'. On the contrary, the smaller the trust value between vehicle n and vehicle n'. Let H = [w n,n' N×N is the social trust matrix of all vehicles in the area, where N represents the number of vehicles in the area. By clustering vehicles, the links between vehicles in the same cluster will be more stable, and the efficiency of perception data processing will also be improved.

[0149] S43: Guide edge service vehicle aggregation through digital twins. By deploying digital twin nodes in RSUs, each RSU collects the computing power and communication topology of vehicles in its surrounding environment, and then transmits this data through wired transmission to build a digital twin of the vehicle edge. The digital twin model of the physical layer can be modeled as:

[0150]

[0151] Where, They represent the virtual mapping of vehicles, RSUs, and edge servers respectively.

[0152] S44: In order to optimize the selection of cooperative objects, improve the stability of the communication link, and enhance the processing efficiency of perception data, an improved K-means algorithm based on the gravity model is proposed. The mass of the vehicle n requesting cooperative perception is defined as:

[0153]

[0154] Where C n Indicates that the vehicle completes the task n The amount of CPU cycles required; f n is the number of CPU cycles per second of the vehicle, i.e., computing resources. By applying the mass model to the gravity model, the social gravity between the demand vehicle n and the service vehicle n' is defined as:

[0155]

[0156] In the formula, γ1, γ2, and γ3 represent the modeling parameters of the relative correlation of resources, social trust value, and communication considered by DTN clustering. The numerator uses the maximum function to quantify the correlation between two vehicles, and the denominator is based on the social trust value and communication rate to replace the original distance. Figure 3 As shown in , the K-means algorithm is first used to cluster the vehicles, and then the RSU closest to the cluster center is merged into the group based on the distance from the cluster center to the RSU. Assume that the vehicle set without the cluster center is The cluster center set is Finally, the set of cluster groups generated by the clustering algorithm is

[0157] S45: The proposed digital twin-assisted vehicle clustering method first initializes the parameters in the formula and the number of clusters. Next, cluster centers are randomly selected and an initial cluster is generated. Finally, the social attraction between each vehicle is calculated using the improved K-means attraction formula. The target vehicle then selects the cluster center with the largest social attraction and adds it to the cluster.

[0158] Furthermore, in S5, let the total bandwidth be B, and the communication bandwidth allocated to the communication link by vehicle n be Expressed as In each time slot k, vehicle n is allocated uplink bandwidth Indicates that the downlink bandwidth is allocated express.

[0159] S51: In the downlink communication link, vehicle n receives the sensing result of RSUm, and its transmission rate is It can be given by Shannon's formula:

[0160]

[0161] Where, Denotes the SINR of the downlink channel between vehicle n and RSU m. Indicates the uplink transmission rate of communication between vehicle n and RSUm, specifically:

[0162]

[0163] Where S m Represents the communication receiving power at RSU. Similarly, let represents the uplink transmission rate between vehicle n and cooperative vehicle n', which can be expressed as:

[0164]

[0165] Where S n' represents the communication receiving power at the cooperative vehicle n'.

[0166] S52: In time slot k, when the vehicle chooses to execute the task locally, the vehicle's local computing resources represent f n (k), the local computation delay can be expressed as:

[0167]

[0168] S53: When local computing resources are insufficient to support the task completion, or the sensing target is out of the sensing range, consider offloading the task to the edge service node for processing. Data processing delay includes task upload delay, data fusion delay, task calculation delay, and result downlink transmission delay, which are represented by t up , t mix , t comp , t down When the vehicle chooses to offload the task to the RSUm, the task upload delay model is as follows:

[0169]

[0170] S54: After the perception point cloud data is uploaded to the RSU, data fusion is required. The size of the fused data is The delay of RSU data fusion is expressed as:

[0171]

[0172] Where Ω represents the ratio of data fusion delay to data size.

[0173] S55: After data fusion, the RSU needs to perform calculation processing. The calculation processing delay can be expressed as:

[0174]

[0175] Where f' m represents the computing resources allocated by RSU to vehicle n, which should satisfy 0≤f' m <f m .

[0176] S56: Calculate the downlink transmission delay model of the fused data task result as follows:

[0177]

[0178] Where δ is the ratio of the output data size to the input data size, and δ<< 1. Therefore, the data downlink transmission delay can be ignored.

[0179] S57: During the collaborative sensing process between the demand vehicle and the edge service node RSU, the total processing delay of the sensing data can be obtained based on the above discussion:

[0180]

[0181] S58: When the vehicle chooses to offload the task to the cooperative vehicle n', the task upload delay model can be expressed as:

[0182]

[0183] The data fusion delay can be expressed as:

[0184]

[0185] Where, D n' (k) represents the size of the perception data of vehicle n' at time slot k.

[0186] S59: The calculation delay of unloading to cooperative vehicle n' for calculation can be expressed as:

[0187]

[0188] Where f' n' (k) is the computing resource provided by service vehicle n' to demand vehicle n in time slot k, which should satisfy 0≤f' n' <f n' Therefore, during the collaborative perception process between vehicle n and vehicle n', the total processing delay of the perception data is:

[0189]

[0190] S510: When the perception task of the demand vehicle is executed on the edge service node, the total delay of the task execution can be expressed as:

[0191]

[0192] By comprehensively considering the processing delay of the perception task in both local computing and collaborative computing modes, the total computing delay in different perception modes can be obtained. Specifically, it can be expressed as:

[0193]

[0194] Furthermore, in S6, the system joint utility maximization optimization model is specifically expressed as follows:

[0195]

[0196] Where p and b represent the power and bandwidth resource allocation vectors respectively; a represents the set of cooperative sensing strategies within the time slot set. Constraint C1 represents the uplink bandwidth resource allocated to the vehicle. Must not exceed the total uplink bandwidth resource b of RSU c Constraints C2 and C3 represent the collaborative sensing decision a of the demand vehicle in each time slot nm , specifies whether the perception data is calculated locally or in The computation is processed in edge service nodes, N represents the set of vehicles; constraint C4 represents the local computation delay of the vehicle. Edge collaborative computing latency The maximum tolerable delay of the task must not exceed T n ; Constraint C5 represents the computing resources f' allocated by RSU to the demand vehicle m Must not exceed its own available computing resources f m ; Constraint C6 represents the computing resources f' allocated by the cooperative vehicle to the demand vehicle n' Must not exceed the total amount of computing resources available to it f n' ; Constraint C7 represents the vehicle's perceived power p n Not less than the minimum perceived power P min , and must not exceed the maximum perceived power P max ; Constraint C8 indicates that the two weight factors φ1 and φ2 are used to match the system utility between perception and calculation, and their sum is 1; Constraint C9 indicates that the perception occupies bandwidth b s and communication bandwidth b c The sum shall not exceed the total bandwidth resource B.

[0197] Furthermore, S6 uses a deep reinforcement learning algorithm to solve the optimization problem of joint bandwidth, power and collaborative sensing strategy. The specific solution steps are:

[0198] S61: First, the optimization model is modeled as a Markov decision process, and all demand vehicles are considered as an intelligent agent. The detailed definition of the Markov decision process elements is as follows:

[0199] (1) State space s: includes the perception accuracy utility, the channel gain during perception and communication, and the changes in the collaborative perception strategy, expressed as

[0200] (2) Action space: There are three key elements in the action space, namely power allocation strategy, bandwidth allocation strategy and cooperative sensing switching strategy, which are expressed as a(k) = {p(k), b(k), a(k)}.

[0201] (3) Reward function: The reward function is composed of the overall system utility minus the penalty for violation of bandwidth and delay constraints; it is specifically expressed as:

[0202]

[0203] Where η1, η2, η3 are utility scaling coefficients, task processing delay penalty coefficients, and bandwidth violation penalty coefficients, respectively; U represents system utility, represents the penalty for delay constraint violation; Indicates bandwidth violation penalty.

[0204] S62: TD3 selects the lower value from the two outputs of the Critic network As the target value. TD3 trains the Critic network by minimizing the square loss of the time difference error. is minimized as a cost function:

[0205]

[0206] Where, Represents the parameterized state-action value function Q, with parameter θ. The target value of function Q is It can be expressed as:

[0207]

[0208] S63: The actions used in the Critic network are defined as:

[0209]

[0210] In the formula, noise Follow the clipped normal distribution clipping This shows is a random variable, following and belongs to the interval [-c,c].

[0211] S64: Actor network from Critic network The output is the action, which is expressed as a k =ν κ (s k )+υ, κ represents the parameters of the Actor network, ν κ Represents the output of the Actor network, which is deterministic and continuous.

[0212] S65: Adjust the parameter ν by minimizing the cost function:

[0213]

[0214] Where, d ν (s) represents the state distribution. Gradient Used to update the parameter ν.

[0215] S66: Parameter update of Actor network:

[0216]

[0217] The algorithm block diagram is as follows Figure 4 shown.

[0218] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A digital twin-assisted collaborative perception and resource allocation method in an Internet of Vehicles, characterized by: The following steps are involved: S1: Build a digital twin-assisted connected vehicle scenario, covering communication, computing, and perception functions; S2: Design a two-layer perception area division mechanism based on spatial distance, enabling independent perception mode within the vehicle's perception range and triggering collaborative perception mode outside the perception range; S3: Quantify the constraint relationship between perception accuracy and communication bandwidth through the Cramer-Rao bound, and construct a perception quality model that minimizes the mean square error of target positioning; S4: Construct a social trust matrix based on vehicle kinematic characteristics and use an improved K-means clustering algorithm to select highly stable collaborative nodes; S5: Select a binary task offloading scheme based on the collaborative sensing mode, including local processing or edge node offloading processing; S6: Establish a system utility maximization model that jointly optimizes the allocation of transmission power, communication bandwidth, and computing resources, and use a deep reinforcement learning algorithm to solve the optimal strategy.

2. The collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins according to claim 1 is characterized by: In S1, the vehicle networking scenarios assisted by digital twins include: Physical layer: includes demand vehicles, service vehicles, and roadside units equipped with edge computing resources; Digital twin layer: real-time mapping of physical entity status information, including location, communication topology, and computing resources; Control layer: performs resource optimization scheduling and collaborative perception strategy decision-making.

3. The collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins according to claim 1 is characterized by: In S2, the dual-layer perception area division is specifically as follows: Define the perception area O S and area of ​​interest O I is a concentric circle area, satisfying Through the binary variable a nm ∈{0,1} controls the collaborative sensing mode switching. When a nm = 0, independent perception is performed. nm =1 triggers collaborative sensing.

4. The collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins according to claim 1 is characterized by: In S3, the perceptual quality model is defined as: in, and denote the Cramer-Rao lower bounds for distance and angle, respectively, d ,λ θ is the normalization factor, is the transmit power, Perceived bandwidth.

5. The collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins according to claim 1 is characterized by: In S4, the method for constructing the social trust matrix includes: Direction similarity Calculated by vehicle driving direction consistency; Speed ​​similarity Calculated by speed ratio; Total trust value where θ 1+ θ2=1.

6. The collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins according to claim 1 is characterized by: In S5, the task offloading delay calculation includes: Local computing latency Edge computing total latency Among them C n (k) is the number of CPU cycles required for vehicle n to complete the task in time slot k, D n (k) is the task data size of vehicle n in time slot k, f n (k) is the computing resource of vehicle n.

7. The collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins according to claim 1 is characterized by: In S6, the system utility maximization model expression is: Constraints: Where G n (k) is the cooperative perception gain of vehicle n in time slot k, Cost n (k) is the total delay of task processing in time slot k; p, b are the power and bandwidth resource allocation vectors respectively; a is the set of cooperative sensing strategies in the time slot set. Constraint C1 represents the uplink bandwidth resources allocated to the vehicle. Must not exceed the total uplink bandwidth resource b of RSU c Constraints C2 and C3 represent the collaborative sensing decision a of the demand vehicle in each time slot nm , specifies whether the perception data is calculated locally or processed in M ​​edge service nodes, N represents the set of vehicles; constraint C4 represents the local computing delay of the vehicle Edge collaborative computing latency The maximum tolerable delay of the task must not exceed T n ; Constraint C5 represents the computing resources f allocated by RSU to the demand vehicle m 'shall not exceed its own available computing resources f m ; Constraint C6 represents the computing resources f' allocated by the cooperative vehicle to the demand vehicle n’ Must not exceed the total amount of computing resources available to it f n' ; Constraint C7 represents the vehicle’s perceived power p n Not less than the minimum perceived power P min , and must not exceed the maximum perceived power P max ; Constraint C8 indicates that the two weight factors φ1 and φ2 are used to match the system utility between perception and calculation, and their sum is 1; Constraint C9 indicates that the perception occupies bandwidth b s and communication bandwidth b c The sum shall not exceed the total bandwidth resource B.

8. The method for collaborative perception and resource allocation in a digital twin-assisted connected vehicle network according to claim 7, characterized in that: The deep reinforcement learning algorithm adopts the double-delay deep deterministic policy gradient TD3 algorithm, which includes: Two critic networks evaluate Q values ​​and use the target network for smooth updates; An Actor network generates deterministic strategies; Motion exploration mechanism with clipping noise.

9. The collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins according to claim 1 is characterized by: In S4, the improved K-means clustering algorithm includes: Define vehicle mass Mn = Cn / fn; Constructing a gravity model based on resources, trust value and communication rate: Among them, γ1, γ2, and γ3 represent the modeling parameters of the relative correlation of resources, social trust value, and communication considered by digital twin network clustering, respectively. n,n' is the communication rate; M n =C n / f n , represents the mass of vehicle n; w n,n' represents the social trust value between vehicles n and n′; r n,n' represents the communication rate between vehicles n and n'.

10. The collaborative perception and resource allocation method in the Internet of Vehicles assisted by digital twins according to claim 1 is characterized by: In S6, the state space of the Markov decision process includes: Perceived Accuracy Utility Channel gain parameter; History of changes in collaborative strategies; The action space contains the power allocation vector p, the bandwidth allocation vector b, and the coordination strategy vector a.

Images