Urban traffic ubiquitous perception method of federal auxiliary large model

Abstract

The invention provides an urban traffic ubiquitous perception method based on a federal aided large model, and belongs to the technical field of data processing, and the method specifically comprises the steps: 1, building a space-time diffusion field based on multi-source heterogeneous traffic data of a target city, and dynamically dividing traffic perception subtasks according to the space-time diffusion field, and obtaining a subtask set; 2, scene training is carried out according to each traffic perception sub-task in the sub-task set based on federated learning, and a traffic perception sub-model corresponding to each traffic perception sub-task is obtained; and step 3, performing hierarchical aggregation and splicing on all the traffic perception sub-models to obtain a hierarchical perception large model, and realizing multi-task parallel perception and dynamic management of target urban traffic. Through the scheme of the invention, the generalization ability and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a city traffic ubiquitous perception method based on federated auxiliary large model. BACKGROUND

[0002] At present, with the continuous development of intelligent transportation system, city traffic perception is facing multiple challenges such as multi-source heterogeneous data fusion difficulty, high sensor dependence and data privacy protection. The traditional method usually adopts the "data collection-cleaning-feature engineering" mode, which not only cannot meet the real-time requirement, but also cannot fully utilize the scattered traffic data resources. In addition, the city traffic environment has strong spatio-temporal non-stationarity, and the model requirements of different road network regions and different traffic scenes are quite different, which makes it difficult for a unified centralized perception model to adapt to diversified application scenarios.

[0003] Therefore, there is an urgent need for a city traffic ubiquitous perception method based on federated auxiliary large model with high generalization ability and accuracy. SUMMARY

[0004] Therefore, the present application provides a city traffic ubiquitous perception method based on federated auxiliary large model, which at least partially solves the problem of poor generalization ability and accuracy in the prior art.

[0005] The present application provides a city traffic ubiquitous perception method based on federated auxiliary large model, comprising: Step 1: constructing a spatio-temporal diffusion field based on multi-source heterogeneous traffic data of a target city, dynamically dividing traffic perception sub-tasks according to the spatio-temporal diffusion field, and obtaining a sub-task set; Step 2: performing scenario-based training according to each traffic perception sub-task in the sub-task set based on federated learning, and obtaining a traffic perception sub-model corresponding to each traffic perception sub-task; Step 3: layering and aggregating all traffic perception sub-models to obtain a layered perception large model, and realizing multi-task parallel perception and dynamic management of the target city traffic.

[0006] The city traffic ubiquitous perception method based on federated auxiliary large model in the present application comprises: step 1: constructing a spatio-temporal diffusion field based on multi-source heterogeneous traffic data of a target city, dynamically dividing traffic perception sub-tasks according to the spatio-temporal diffusion field, and obtaining a sub-task set; step 2: performing scenario-based training according to each traffic perception sub-task in the sub-task set based on federated learning, and obtaining a traffic perception sub-model corresponding to each traffic perception sub-task; step 3: layering and aggregating all traffic perception sub-models to obtain a layered perception large model, and realizing multi-task parallel perception and dynamic management of the target city traffic.

[0007] The beneficial effects of the present application are: through the scheme of the present disclosure, the context features of urban traffic are obtained by using the space-time diffusion field method, and the problem of difficult unified representation of different modal data is solved; then through the scenario training mechanism of the sub-model of federated learning, the demand of each edge node for local data privacy and diversified feature learning is met; finally, through the hierarchical perception large model aggregation, the outputs of different sub-models are flexibly spliced and fused, the multi-task parallel perception and dynamic management of urban traffic are realized, and the method of the present application can achieve efficient, accurate and safe effect in city-level traffic perception, and provides a new idea for the construction and upgrading of intelligent traffic system. BRIEF DESCRIPTION OF DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0009] Figure 1 A flowchart of a federated auxiliary large model urban traffic ubiquitous perception method provided by the present application. DETAILED DESCRIPTION

[0010] The present application will be described in detail below with reference to the drawings.

[0011] The embodiments of the present disclosure will be described below through specific concrete examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in the present description. Obviously, the described embodiments are only some embodiments of the present disclosure, not all embodiments. The present disclosure can also be implemented or applied by other different specific embodiments, and each detail in the present description can be modified or changed based on different views and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present disclosure.

[0012] It is important to note that the various aspects of the embodiments described below are within the scope of the appended claims. It should be apparent that the aspects described herein can be embodied in a wide variety of forms and that any specific structure and / or function described herein is merely illustrative. Based on the teachings provided herein one skilled in the art should appreciate that an aspect described herein can be implemented independently of any other aspects and that an aspect described herein can be implemented both as any number of software and / or hardware configurations (and equivalents thereof). For example, an aspect can be implemented as a hardware only implementation, a software only implementation, or a combination of software and hardware implementations.

[0013] In addition, in the following description, specific details are given to provide thorough understanding of examples. However, one skilled in the relevant art will understand that the aspects described herein can be practiced without these specific details.

[0014] The present application provides a city traffic ubiquitous perception method assisted by a federal large model, which can be applied to the traffic ubiquitous perception process in the city traffic management scene.

[0015] Referring to Figure 1 A flowchart of a city traffic ubiquitous perception method assisted by a federal large model provided by the present application is shown. As Figure 1 shown, the method mainly includes the following steps: Step 1, constructing a spatiotemporal diffusion field based on multi-source heterogeneous traffic data of a target city, dynamically dividing traffic perception subtasks according to the spatiotemporal diffusion field, and obtaining a subtask set; In specific implementation, this step unifies the multi-source heterogeneous traffic data into a shareable mathematical field structure through the construction of a spatiotemporal diffusion field model, then generates a context feature vector through a diffusion-de-noising process, and realizes the dynamic screening and resource optimization allocation of subtasks based on the vector. Each sub-step follows the logical chain of “defining the basic framework→modeling the core process→generating the key features→implementing the task decision”, which is specifically as follows: Sub-step 1.1, defining the construction target and the basic definition (1) determining the target of constructing a “context feature space” Traditional traffic perception relies on the “data collection—cleaning—feature engineering” process, which has problems such as difficulty in multi-modal fusion and strong dependence on sensors. This sub-step aims to unify the multi-source heterogeneous data such as traffic environment, physical scene, and semantic information into a shareable mathematical field structure through spatiotemporal diffusion field modeling, avoiding explicit dependence on specific observation / sensing methods. The finally generated context vector can directly serve downstream tasks such as path planning, cooperative perception, and accident prediction, simplifying data preprocessing and enhancing model generalization.

[0016] (2) Definition of spatiotemporal domain and latent field To uniformly describe multi-modal information in spatiotemporal continuum, we first define: 1) Definition of spatiotemporal domain: where the three-dimensional physical space and the positive half-axis of time (time variable t) constitute a continuous spatiotemporal coordinate system, which is used to describe the spatiotemporal evolution of traffic phenomena.

[0017] 2) Definition of latent field: Define a smooth function that maps any spatiotemporal point to a -dimensional feature vector . Here, is the configurable embedding dimension (e.g., 128). The field function implicitly encodes information such as traffic conditions (e.g., congestion level), physical environment (e.g., road friction coefficient), and potential semantics (e.g., accident risk level), independent of specific data sources (e.g., cameras, radars), providing a unified input carrier for subsequent diffusion processes.

[0018] 3) Diffusion time parameter To construct a reversible "noise whitening" process, we introduce a dimensionless parameter independent of physical time , constructing a reversible noise evolution process: When increases from 0 to 1 (forward diffusion), the original field is degraded to a Gaussian distribution by gradually injecting noise, providing a standardized noise starting point for reverse denoising; When decreases from 1 to 0 (reverse denoising), the original field structure is reconstructed by removing noise, achieving the fusion and enhancement of multi-modal information.

[0019] Sub-step 1.2, diffusion and denoising process modeling Based on the definition of spatiotemporal domain and latent field in sub-step 1.1, this sub-step constructs a noise reversible transformation process to provide a basis for the generation of subsequent context vectors: (1) Forward diffusion process, noise injection, establish a unified noise space: The degradation process of the latent field to Gaussian noise is described by a stochastic differential equation (SDE):

[0020] where: denotes the field at diffusion time ; and is the drift term, which contains learnable parameters , deterministic change of the control field; is the diffusion coefficient, which increases monotonically with , the control noise injection rate; is the standard Brownian motion, introducing random perturbations.

[0021] As propagates from 0 to 1, any initial field will be gradually "washed" into an isotropic multi-dimensional Gaussian distribution , forming a unified noise space independent of the original data, providing a universal reconstruction basis for the inverse process.

[0022] (2) Inverse denoising process, noise stripping, reconstruction of multi-modal features: Noise stripping is achieved through the inverse SDE, recovering the original field structure from the Gaussian noise

[0023] where is called the scoring network, which is used to estimate where is the gradient of the log-probability density function with respect to the latent field , called the score function in the diffusion model, guiding the denoising direction (e.g., suppressing noise interference, enhancing real features); is the reverse-time Brownian motion (i.e., time-reversed Wiener process) in the opposite direction of the forward process, achieving step-by-step noise stripping.

[0024] When evolving back from to , the Gaussian noise is removed layer by layer, and the field will be "denoised" to match the distribution corresponding to the original , and the final output is the original field estimate fused with multi-modal information, providing the core input for feature extraction in sub-step 1.3.

[0025] Sub-step 1.3, feature vector calculation and network training, generating task-available features Based on the diffusion-denoising framework established in sub-step 1.2, this sub-step ensures the effectiveness of the inverse process by training the scoring network and extracts context vectors that can directly drive downstream tasks: (1) Calculate the context vector, unified representation of multi-modal features After completing the inverse denoising, extract the field vector at time at the spatiotemporal point , denoted as The vector naturally integrates traffic situation, environmental characteristics and semantic information in the original field through the diffusion process, and has the following key characteristics: Sensor-free dependence: No need to focus on data sources (such as cameras or radars), only through field functions Implicit encoding; Unified dimension: All heterogeneous data are mapped to fixed dimension Vector, eliminating the dimension alignment problem in feature engineering; Noise robustness: The influence of sensor noise or data missing is filtered through the reverse denoising process.

[0026] (2) Self-supervised score matching, training score network to ensure the correctness of the reverse process To make the score network Accurately estimate the intermediate distribution of the forward diffusion , the self-supervised objective function is designed as follows:

[0027] In the formula, is the denoising score matching loss function, the objective function for training the score network, is the set of learnable parameters of the score network, is the diffusion time parameter, the value range is [0, 1], a dimensionless parameter independent of the physical time t, is the latent field state at time τ, is the marginal distribution of the forward diffusion process at time τ, is the score network used to estimate the gradient of the log probability density, is the expectation of which obeys the distribution , is the square of the L2 norm, used to measure the difference between the predicted score and the true score.

[0028] Training logic: using unlabeled historical field data Simulate the forward diffusion process, force Learn the noise evolution law, so as to accurately remove noise in the reverse process and ensure Reliability; Self-supervised advantage: no manual labeling, only rely on offline traffic scene simulation or historical data, reduce training cost, adapt to complex traffic environment.

[0029] Where, is the marginal distribution of the forward diffusion process, is the variance scheduling parameter.

[0030] Substep 1.4, task screening and mapping High-quality context vector generated in sub-step 1.3 Based on this, this sub-step solves the problem of "how to dynamically select and activate sub-tasks according to real-time scenes to achieve optimal allocation of resources": (1) Discrimination vector extraction and task-related feature screening Considering that downstream sub-tasks (such as congestion detection, accident warning) usually only focus on specific scene elements, the high-dimensional vector needs to be reduced in dimension and enhanced in features: 1) Projection transformation: Define the mapping such as a fully connected layer or a convolutional network, to convert it into a low-dimensional discrimination vector .

[0031] Design goal: retain features strongly related to the task (such as the spatiotemporal gradient of traffic flow density for congestion detection, the change in road curvature for accident warning), eliminate irrelevant dimensions, and reduce subsequent computational complexity; Flexibility: Φ can choose linear or nonlinear transformation according to task requirements, adapting to different scene feature extraction needs.

[0032] 2) Discretization indicator function: For each component of set a task-sensitive threshold (determined through historical data cross-validation or domain knowledge), define a binary activation function: .

[0033] This function converts continuous features into Boolean conditions for sub-task activation, providing a basis for preliminary task screening.

[0034] (2) Dynamic task mapping and resource optimization Predefine a set of potential sub-tasks , each task has a corresponding revenue and resource consumption (such as computing power, memory usage), the goal is to select and activate a subset M(Ct) under the resource upper limit Λ to maximize revenue: 1) Preliminary task screening: Calculate the task activation score using the scoring function , and select the candidate set:

[0035] Where is the task-specific activation threshold, excluding obviously irrelevant tasks to narrow the optimization space.

[0036] 2) Optimal task combination under resource constraints: Construct a binary optimization problem:

[0037] where, represents activating the traffic perception subtask . represents the upper limit of resources, represents the condition that must be met for each element i in the set, represents the pre-screened candidate task set; this problem is solved by integer linear programming or heuristic algorithms to ensure that high-yield, low-consumption tasks are activated first under limited resources (such as activating accident warning tasks with high real-time requirements when computing power is tight).

[0038] 3) Mapping function definition: The final activated subtask set is determined by the optimal solution : the final activated subtask set is , realizing the end-to-end mapping of “real-time scene features → task activation decision” to support dynamic resource allocation and efficient operation of the traffic perception system. As updates, will also change accordingly, thus realizing dynamic task division.

[0039] Step 2, based on federated learning, each traffic perception subtask in the subtask set is trained according to the scene, and the traffic perception submodel corresponding to each traffic perception subtask is obtained; In implementation, this step realizes the function of federated learning-based traffic perception submodel scenario training, and the specific process includes: Substep 1: Edge node data and scene division There are n edge nodes , each edge node holds a local data set . For subtask , divide into according to the scene, where represents different scenes, and satisfies , is the local training data of subtask in node . Among them, represents that after merging all L scene data sets of subtask on edge node , it is equal to the complete local training data set of the task in the node subtask at the edge node with local training data, initialize the submodel parameter set at the edge node for the subtask download initial parameters to initialize the local model .

[0040] Sub-step 2: Scenario-based local training phase At the k-th iteration, for the scenario of the subtask at the edge node , define the scenario-specific loss function as follows:

[0041] where, is the base loss function, which can adopt cross-entropy loss for classification tasks or mean square error loss for regression tasks. is the data sample and its label, is the number of samples in the scenario . is the balance coefficient, used to adjust the weight of the base loss and the scenario constraint term. is the predicted distribution of the model under the scenario . denotes divergence, used to measure the difference between the model's predicted distribution and the scenario reference distribution ; is the reference distribution of the scenario , which is generated as follows: 1) Scenario feature vector construction: Extract the spatio-temporal features related to the scenario from the context vector generated in step 1, and combine the local prior information of the edge node to construct the scenario feature vector . Let be processed by aggregation operation , while combining the node 's position coordinate component (replacing ), time period component , and historical traffic flow pattern component : .

[0042] 2) Data-driven reference distribution modeling: Based on the historical scenario data set (containing​ Scenes historical samples), estimate the reference distribution for different task types: Classification task: The reference distribution is the scene The empirical probability distribution of the following labels: ,in is a collection of categories, is the indicator function.

[0043] Regression task: reference distribution is scene The conditional Gaussian distribution of the target variable is obtained from historical data statistics: .

[0044] 3) Scene association mapping: through mapping function The scene feature vector Mapped to the reference distribution parameter space. For classification tasks, directly output the prior probability of each label; for regression tasks, output the mean and standard deviation Logarithm of (to ensure positive definiteness). By gradient calculation and parameter update: ,in, is the learning rate, is the loss function with respect to the parameters gradient.

[0045] Sub-step 3: In-scene aggregation and parameter adjustment At the edge node For the same subtask Updated parameters for different scenarios Perform in-scene aggregation: ,in It's a scene The reference distribution of With the global reference distribution JS divergence, used to measure scenario specificity. It is a scene-specific regulator that forces the model to balance between general features and scene-specific features.

[0046] Sub-step 4: Cross-node federation aggregation phase Defining aggregate functions , collect the adjusted parameters of subtask Ti on all edge nodes , perform global parameter update: ,in, It is an edge node The weight of .

[0047] Sub-step 5: Training convergence and model output Setting the convergence threshold , when satisfied When the training is stopped, the final output is a set of trained sub-models where are the optimal parameters after convergence, and these sub-models will be used for the aggregation of the hierarchical perception large model in step 3.

[0048] Step 3: The hierarchical aggregation of all traffic perception sub-models is performed to obtain a hierarchical perception large model, which realizes multi-task parallel perception and dynamic management of the target city traffic.

[0049] In specific implementation, the specific process of aggregation and construction based on the hierarchical perception large model includes: Sub-step 1: Definition of hierarchical architecture and adaptation of sub-model input The set of sub-models output in step 2 is where is the set of activated sub-tasks dynamically selected in step 1, is the scene-specific parameter obtained by training in step 2. A three-level hierarchical architecture is constructed , and each layer is defined as follows: (1) Feature adaptation layer : Define the linear transformation matrix and the bias vector , and map the context vector generated in step 1 to the adapted feature . , is the input dimension of the perception sub-model , and the inactive sub-tasks are filled with zeros to keep the structure uniform: .

[0050] (2) Sub-model parallel layer : Deploy the sub-models in parallel , and the input-output relationship is: , is the output dimension of the sub-model (classification task is the number of categories, and regression task is 1); Decision output layer : Map the fused feature vector to the final perception decision output, supporting multi-task joint inference. Define the fully connected transformation matrix and the bias vector , and the output projection matrix and the bias vector , and the input-output relationship is: where is the dynamic output dimension, is the output dimension of the sub-task (classification task is the number of categories, and regression task is 1), for the hidden layer dimension (e.g., 512), the final output is decomposed into in the order of subtask index, and the multi-task joint perception decision is realized Substep 2: Perception submodel splicing and hierarchical aggregation mechanism (1) Definition of output feature splicing operator Design a dynamic splicing operator where is the current activated subtask index set, and the specific splicing process is defined as:

[0051] where, represents a set of real numbers represents the output dimension of the submodel , and represents a vector space composed of all dimensional real number vectors, represents the total output vector space after splicing, is the current activated subtask index set, ; This operator splices along the column dimension in the order of subtask index, generating a column vector with a dimension of .

[0052] (2) Cross-task attention fusion Introduce the task attention matrix , whose element represents the contribution weight of subtask to the th dimension of the fused feature. The attention coefficient vector is generated by the scene gating function The weighted fusion process is: where is a diagonal matrix with as the diagonal element, realizing the mapping of feature dimension and task weight.

[0053] Substep 3: Formal description of dynamic structure adjustment When changes, define the activation mask ( is the predefined total number of subtasks), where if and only if . The dynamic adjustment process is implemented as follows: (1) Real-time calculation of activated subtask dimension Dynamic summation of input dimension: Activation subtask index set: .

[0054] (2) Tensor shape adaptive transformation Define dimension alignment operator , dimension expansion is performed on each activation submodel output: , where is a zero matrix of . Through this operator, the outputs of submodels with different dimensions are mapped to a unified global feature space, and finally spliced into .

[0055] (3) Dynamic reconstruction of computation graph In the dynamic computation graph, the tensor shape satisfies the relationship: , where represents element-wise multiplication, is the summation operation. When is an empty set, the default output is a zero tensor: .

[0056] (4) Gradient backpropagation compatibility The gradient transmission of the dynamic splicing operator satisfies the chain rule: , where is the joint loss function defined in step 3, ensuring the correctness of the submodel parameter update path.

[0057] Substep 4: Large model verification and error calibration (1) Multi-task joint verification dataset Construct the verification set , where is the labeled data of subtask , and the verification set dimension dynamically matches the activation subtask: ; (2) Joint perception accuracy of classification tasks: ; , where is the multiplication symbol, indicating the multiplication operation on the set I, is the indicator function symbol, which is 1 when the condition in the parentheses is true, and 0 otherwise.

[0058] Overall RMSE of regression tasks: ;

[0059] (3) Error calibration optimization equation When the verification error When , the sub-model parameters are updated through the following regularized optimization: ; in, Shared parameter space for large models, For the subtask after error calibration optimization The fine-tuning parameters of Subtasks after the federated learning training is completed The convergence parameter of To represent subtasks The general parameters of is the learning rate for error calibration, which controls the step size of parameter updates. represents the learning rate, Indicates the parameters Find the partial derivatives, represents the task loss function; The validation error Err has the following correlation with the accuracy index defined above according to the activation subtask type: The validation error of the classification task is: ; That is, the error rate is equal to 1 minus the joint perception accuracy of the classification task defined above, where: ;

[0060] Validation error for regression tasks: , that is, directly use the overall RMSE of the regression task defined above: ;

[0061] Verification error of mixed tasks: When the activation subtask set I contains both classification tasks and regression tasks: ; in is the classification task proportion weight, is the regression error reference benchmark.

[0062] in Share parameter space for large models to ensure parameter compatibility between sub-models and the overall architecture.

[0063] (4) Mathematical expression of the final large model The integrated hierarchical model can be expressed as a composite function of multilinear transformations: ; Where, represents the attention fusion mechanism, Represents the large model output, and C represents the dynamic splicing operator.

[0064] The mathematical mapping relationship of each module is: Feature adaptation: ; Sub-model reasoning: ; Dynamic splicing: ; Attention fusion: ; Decision output: .

[0065] The city traffic ubiquitous perception method of the federated auxiliary large model provided in the embodiment solves the problem that different modal data is difficult to be uniformly represented by using the space-time diffusion field method to obtain the context features of city traffic; then the demand of each edge node for local data privacy and diversified feature learning is met through the sub-model situational training mechanism of federated learning; finally, the outputs of different sub-models are flexibly spliced and fused through the hierarchical perception large model aggregation, the multi-task parallel perception and dynamic management of city traffic are realized, and the method of the embodiment can achieve efficient, accurate and safe effect in city-level traffic perception, and provides a new idea for the construction and upgrading of intelligent traffic system.

[0066] It should be understood that parts of the present disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0067] The above is only a specific implementation of the present disclosure, but the protection scope of the present disclosure is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present disclosure, which should be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A city traffic ubiquitous perception method of a federal auxiliary large model, characterized in that, The method comprises the following steps: Step 1, constructing a spatio-temporal diffusion field based on multi-source heterogeneous traffic data of a target city, dynamically dividing traffic perception sub-tasks according to the spatio-temporal diffusion field, and obtaining a sub-task set; Step 2, performing scenario-based training according to each traffic perception sub-task in the sub-task set based on federated learning, and obtaining a traffic perception sub-model corresponding to each traffic perception sub-task; Step 3, layering and splicing all traffic perception sub-models to obtain a hierarchical perception large model, and realizing multi-task parallel perception and dynamic management of the target city traffic.

2. The method of claim 1, wherein, The step 1 specifically comprises: Step 1.1, defining a spatio-temporal field and a potential field according to the multi-source heterogeneous traffic data, and introducing a diffusion time parameter; Step 1.2, constructing a noise reversible transformation process based on the spatio-temporal field, the potential field and the diffusion time parameter; Step 1.3, performing reverse diffusion based on the noise reversible transformation process and a trained scoring network, and extracting a field vector of the potential field at a spatio-temporal field node when the diffusion time parameter is 0 as a context feature vector after completing reverse denoising; Step 1.4, converting the context feature vector into a low-dimensional discriminant vector through a mapping network, defining a binary activation function for each component of the low-dimensional discriminant vector, and performing dynamic task mapping and resource optimization according to the binary activation function, and obtaining a sub-task set.

3. The method of claim 2, wherein, The step 1.4 specifically comprises: Step 1.4.1, defining the mapping network and converts the context feature vector into a low-dimensional discriminative vector ; Step 1.4.2, determining the low-dimensional discriminant vector of each component task-sensitive threshold according to which a binarization activation function is defined, according to which the low-dimensional discriminant vector is converted into a Boolean condition for the activation of the subtask: ； Step 1.4.3, predefine a set of potential subtasks each task corresponding revenue and resource consumption , denotes the total number of subtasks contained in the set of potential subtasks; Step 1.4.4, computing the score function Compute task activation scores, filter to candidate set: ； wherein, is a task-specific activation threshold; Step 1.4.5, constructing a binary optimization problem: ； wherein, indicates to activate the traffic perception subtask , indicates a resource upper bound, indicates that the following condition must be met for each element i in the set, indicates a pre-screened set of candidate tasks; Step 1.4.6, define the goal as selecting the activated subtask set under the resource upper bound Λ , achieve the maximum benefit, solve the binary optimization problem by integer linear programming or heuristic algorithm to get the optimal solution and accordingly get the final activated subtask set as .

4. The method of claim 3, wherein, The step 2 specifically comprises: Step 2.1: There are n edge nodes, each edge node Holds local datasets , for the traffic perception subtask , local dataset Get scene data by scene division ,in, Indicates different scenarios and satisfies ,in, Indicates that the edge node Previous subtask All L scene datasets After merging, it is equal to the complete local training dataset of the node and the task , Represents a subtask At the edge node Local training data, initialize the sub-model parameter set , , edge nodes For the traffic perception subtask Download initial parameters To initialize the local model ; Step 2.2, in the kth iteration, for the edge node Traffic perception subtask Scene data , define the loss function of scene data for: wherein, is a base loss function, is a data sample and its label, is a scene data sample number, is a balancing coefficient, is a scene data prediction distribution of the model, is a scene data reference distribution, denotes a KL divergence, used to measure the difference between the model prediction distribution and the scene reference distribution ; Step 2.3, at the edge node Above, for the same traffic perception sub-task Updated parameters at different scenarios Intra-scenario aggregation: ； wherein, is a reference distribution for the scene JS divergence from the global reference distribution is a scene-specific regulation factor;​​ Step 2.4, define aggregation function collect all traffic perception sub-tasks on edge nodes adjusted parameters perform global parameter update: ； wherein is the weight of the edge node . Step 2.5, set the convergence threshold When the condition is met, stop training, and finally output the trained sub-model set , where is the optimal scene-specific parameter after convergence.

5. The method of claim 4, wherein, Before the step 2.2, the method further comprises: extracting spatio-temporal features related to the scene from the contextual feature vector combining the local prior information of the edge node to construct a scene feature vector ;​ Calculating an empirical probability distribution of labels under different scenes as a reference distribution based on a scene feature vector, a data sample and a label thereof.

6. The method of claim 5, wherein, The step 3 specifically comprises: Step 3.1, constructing a three-level hierarchical architecture wherein, is a feature adaptation layer, defining a linear transformation matrix to map the context feature vector into an adaptive feature, and the inactive sub-tasks maintain the structure uniform through zero padding, is a sub-model parallel layer, and the sub-models are deployed in parallel and define their input-output relationship, is a decision output layer, mapping the fused feature vector into the final perception decision output, supporting multi-task joint inference; Step 3.2, defining an output feature splicing operator to splice all traffic perception sub-models, introducing a task attention matrix to fuse the splicing result, and obtaining a fusion large model; Step 3.3, when the sub-task set changes, defining an activation mask to adjust the structure of the fusion large model; Step 3.4, constructing a validation set to verify the validation error of the adjusted fusion large model, and when the validation error is greater than an error threshold, updating the parameters of the traffic perception sub-models through regularization optimization, integrating the traffic perception sub-models, and obtaining a hierarchical perception large model.

7. The method of claim 6, wherein, The step 3.2 specifically comprises: Step 3.2.1, design feature concatenation operator , concatenate all traffic perception sub-models along column dimension in sub-task index order according to the feature concatenation operator, generate column vector with dimension : ； wherein, , denotes a set of real numbers denotes a submodel dimension of the output of denotes a vector space consisting of all dimensional real number vectors denotes the total output vector space after concatenation is a set of current active subtask indices ; Step 3.2.2, introduce the task attention matrix , generate the attention coefficient vector through the scene gating function , and then according to the task attention matrix and the attention coefficient vector Column vector Perform weighted fusion to obtain a large fusion model.

8. The method of claim 7, wherein, The expression of the regularization optimization is: ； wherein, is a shared parameter space for large models, is a fine-tuning parameter of the subtask after error calibration optimization , is a convergence parameter of the subtask after federated learning training , is a general parameter representing the subtask , is a learning rate of error calibration, controlling the step size of parameter update, denotes the learning rate, denotes the partial derivative of the parameter , denotes the task loss function; The expression of the hierarchical perception large model is: ； In the formula, represents an attention fusion mechanism, represents a large model output, and C represents a dynamic splicing operator.

Images