Iot ai flexible control method based on auxiliary service

Abstract

The application discloses an auxiliary service-based Internet of Vehicles AI flexible control method, which comprises the following steps: S1: collecting original environment data through multi-modal sensors deployed on vehicles, roadside units and clouds, wherein the original environment data comprises vehicle operation original data, road environment original data, user behavior original data, visible light images, infrared thermal imaging data and millimeter wave radar point cloud data; S2: constructing a cross-modal feature alignment network, performing space-time alignment and feature fusion on the visible light images, the infrared thermal imaging data and the millimeter wave radar point cloud data, and generating an environment fusion feature map; and S3: generating a flexible control strategy model based on a deep reinforcement learning model, wherein the deep reinforcement learning model comprises a fuzzy membership calculation module, a strategy optimization module and a multi-objective reward function design. The application realizes multi-modal fusion perception, dynamic strategy adaptation scenarios, and resource coordination to improve efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle control methods, in particular to an AI flexible control method for vehicle networking based on auxiliary services. BACKGROUND

[0002] With the rapid development of vehicle networking technology, the collaborative interaction between vehicles, vehicles and roadside units, vehicles and the cloud is increasingly frequent, and various auxiliary services, such as adaptive cruise control, emergency braking assistance, real-time road navigation, etc., have higher requirements for the comprehensiveness of environmental perception, the adaptability of control strategies and the rationality of resource allocation.

[0003] However, there are many limitations in the prior art: on the one hand, traditional solutions rely on a single type of data or a single control logic, which is difficult to meet the fusion processing needs of multi-modal data in vehicle networking, resulting in insufficient perception accuracy of complex traffic environments, especially in bad weather such as rain, fog, low light, etc. Data loss or misjudgment can affect service reliability; on the other hand, traditional control strategies mostly use fixed parameters or single target optimization, which is difficult to balance safety, energy consumption, service delay and other multi-dimensional needs, and has poor adaptability to user driving preferences and dynamic scene changes, resulting in low matching degree of control parameters and actual scene needs; in addition, the resource competition problem of multiple service requests in vehicle networking is prominent, and the existing resource allocation method lacks dynamic consideration of the dependency relationship and real-time demand between services, which can easily cause communication delay, waste of computing resources and other problems, which restricts the collaborative efficiency of auxiliary services. Therefore, an AI flexible control method for vehicle networking based on auxiliary services is proposed. SUMMARY

[0004] The present application solves the problems of the prior art by the following technical solutions, and comprises the following steps:

[0005] S1: Collecting original environmental data through multi-modal sensors deployed on vehicles, roadside units and the cloud, wherein the original environmental data includes vehicle operation original data, road environment original data, user behavior original data, visible light images, infrared thermal imaging data and millimeter wave radar point cloud data;

[0006] S2: Constructing a cross-modal feature alignment network to perform spatio-temporal alignment and feature fusion on visible light images, infrared thermal imaging data and millimeter wave radar point cloud data, and generating an environmental fusion feature map;

[0007] S3: Generating a flexible control strategy model based on a deep reinforcement learning model, wherein the deep reinforcement learning model comprises a fuzzy membership calculation module, a strategy optimization module and a multi-objective reward function;

[0008] The fuzzy membership calculation module is used to quantify the fuzzy attributes of the input features.

[0009] The policy optimization module is used for policy iteration optimization based on the fuzzy membership calculation result and the multi-objective reward function through gradient descent method, minimizing the control error and maximizing the comprehensive reward, and finally outputting the dynamic control parameters meeting the scene requirements;

[0010] The input of the flexible control strategy model includes:

[0011] The compressed vector Xf extracted from the environment fusion feature map;

[0012] The vehicle state vector Xs obtained by preprocessing the vehicle operation raw data;

[0013] The user preference vector Xp extracted from the user behavior raw data;

[0014] The output is the dynamic control parameter of the auxiliary service;

[0015] S4: Deploy the conflict resolution module in the service coordination management platform, and dynamically allocate computing resources and communication resources based on the flexible control strategy model.

[0016] Further, the vehicle operation raw data in step S1 includes real-time sensor readings of vehicle speed, acceleration, steering angle, brake state and fault code, and the data sampling frequency is 10 Hz, wherein the brake state is represented by binary 0 for non-braking and 1 for braking, and the fault code is represented by a digital code to represent the fault type and level;

[0017] The vehicle state vector Xs in step S3 is generated by the following steps:

[0018] The vehicle operation raw data is preprocessed:

[0019] The sliding window filtering method is used to smooth the raw data, and the window size is set to 5 sampling periods to remove high-frequency noise;

[0020] Abnormal values are identified by the 3σ criterion, that is, when the data value exceeds the mean value ± 3 times the standard deviation of the data sequence, it is determined as an abnormal value and is removed;

[0021] The min-ma normalization formula Xnorm=(X−Xmin) / (Xmax−Xmin) is used to map all data to the [0,1] interval to unify the data scale;

[0022] Xnorm is the normalized vehicle operation data value, mapped to the [0,1] interval, used to unify the data scale;

[0023] X is the vehicle operation raw data value (such as real-time sensor readings of vehicle speed, acceleration, steering angle, etc.) before preprocessing;

[0024] Xmin is the minimum value in the original data sequence of this kind (such as the minimum value of vehicle speed in all sampling points);

[0025] Xmax is the maximum value in the original data sequence of this kind (such as the maximum value of vehicle speed in all sampling points);

[0026] Features are extracted from the preprocessed data, including time domain features and frequency domain features;

[0027] Time domain features are calculated based on the time sequence of the original data:

[0028] The dynamic change rate of vehicle speed is calculated by the ratio of the difference between the speeds of two adjacent sampling periods and the sampling time interval, i.e. 0.1s, i.e. (v(t)-v(t-1)) / 0.1;

[0029] v(t) is the speed value of the current sampling period (unit: km / h or m / s, consistent with the unit of the original data);

[0030] v(t-1) is the speed value of the previous sampling period;

[0031] The acceleration peak value is the maximum value of acceleration extracted in a sliding window of 10 sampling periods, i.e. 1s;

[0032] The steering angle fluctuation frequency is the number of times the steering angle direction changes within 10 sampling periods, i.e. the number of times the steering angle direction changes from positive to negative or vice versa;

[0033] The brake state duration is the product of the number of consecutive sampling times and the sampling period (0.1s) when the brake state is 1;

[0034] Fault code occurrence frequency: count the number of times the fault code is non-zero within 600 sampling periods, i.e. 60s;

[0035] Frequency domain features are obtained by Fourier transform of the preprocessed original data:

[0036] Engine vibration main frequency is the frequency value with the maximum power spectral density obtained by Fourier transform of the acceleration time sequence;

[0037] Chassis resonance frequency is the frequency value corresponding to the second largest power spectral density obtained by Fourier transform of the steering angle time sequence;

[0038] The vehicle state classification result is obtained by classifying the real-time running state of the vehicle according to pre-set rules:

[0039] The pre-set rules include: determining high risk when the vehicle speed is ≥120km / h; determining frequent braking when the brake state duration accumulates ≥5s within 1 minute; determining dangerous state when the fault code is A level (code 100-199);

[0040] The state classifier based on C4.5 decision tree is adopted to take the preprocessed original data as input, combine the preset rules, and output the vehicle state classification results, including safe state, risk state and dangerous state.

[0041] The feature selection algorithm is used to screen features, that is, the vehicle state classification results are taken as the decision attribute set D, and the features with γi>τ (τ is the correlation threshold, and the value is 0.7) are selected to constitute the dimensionally compressed vehicle state vector Xs.

[0042] The gray correlation analysis method is adopted to take the vehicle state classification results as the decision attribute set D.

[0043] The correlation degree γi of the extracted time domain features and frequency domain features and D is calculated.

[0044] The features with γi>τ (τ is the correlation threshold, and the value is 0.7) are selected to constitute the dimensionally compressed vehicle state vector Xs.

[0045] Further, the feature selection algorithm adopts the gray correlation analysis method to obtain Xs, and the specific process is as follows:

[0046] The vehicle state classification results are taken as the decision attribute set D.

[0047] The correlation degree of the time domain features or frequency domain features and D is calculated.

[0048]

[0049] Wherein, n is the sample quantity; is the decision attribute value of the kth sample; is the value of the ith feature in the kth sample;

[0050] is the two-level minimum difference;

[0051] is the two-level maximum difference;

[0052] ρ is the resolution coefficient; τ is the correlation threshold;

[0053] The features with γi>τ (τ is the correlation threshold, and the value is 0.7) are selected to constitute Xs.

[0054] Further, the step S2 further includes:

[0055] S21: A double-flow convolutional neural network (CNN) encoder is used to process two types of image data respectively:

[0056] The visible light image data I v ​​​, the spatial features are extracted by a double-flow convolutional neural network model Φ to obtain visible light features F v , specifically, ;

[0057] The infrared thermal imaging data I t , the thermal distribution features are extracted by the same double-flow convolutional neural network model Φ to obtain infrared features , specifically, ;

[0058] , wherein, is a convolutional neural network model for extracting visible light image features and infrared thermal imaging features;

[0059] S22: voxelizing the millimeter wave radar point cloud data P r to generate millimeter wave radar point cloud features F r , specifically including: dividing Pr into uniform three-dimensional grids, i.e. voxel grids; calculating the point cloud density and spatial distribution entropy of each voxel grid, and integrating them into millimeter wave radar point cloud features Fr;

[0060] S23: performing spatio-temporal alignment on , , F r by a cross-modal attention module to output aligned features F align , and the calculation formula is:

[0061] ;

[0062] σ is an activation function; Q is a query vector; K is a key vector; V is a value vector; d k is the dimension of the key vector K; K T is the transpose of the key vector;

[0063] S24: inputting the aligned features F align into a feature pyramid network to generate the final environment fusion feature map Ff usion by multi-scale feature fusion.

[0064] Further, the specific process of S22 is:

[0065] S221: dividing three-dimensional voxel grids, dividing the millimeter wave radar point cloud data P r into uniformly distributed three-dimensional grids, each grid being defined as a voxel grid as a basic spatial unit for point cloud feature extraction;

[0066] S222: calculating the point cloud density feature ρ in the voxel grid, for each voxel grid, calculating the density feature of the internal point cloud, and the formula is:

[0067] ;

[0068] wherein, wherein, is the number of radar points contained in the voxel grid, is the spatial volume of the voxel grid (determined by the three-dimensional grid size);

[0069] S223: Extract the spatial distribution entropy Hv of the point cloud in the voxel grid. For each voxel grid, calculate its distribution entropy Hv based on the spatial distribution of points, and the formula is:

[0070] ;

[0071] wherein, M is the total number of points in the voxel grid, p i is the probability distribution of the i-th point in the voxel grid (i.e. the proportion of the i-th point in the total number of points);

[0072] S224: Generate the millimeter wave radar point cloud feature F r , the density feature p and the spatial distribution entropy Hv of each voxel grid are integrated to form the millimeter wave radar point cloud data P r corresponding to the millimeter wave radar point cloud feature F r .

[0073] Further, the fuzzy membership degree calculation module in step S3 is configured to perform the following operations:

[0074] S31: Calculate the fuzzy membership degree value by using the adaptive Gaussian membership function:

[0075] ;

[0076] wherein, is the input feature vector; is the center parameter of the j-th fuzzy set corresponding to the i-th feature; σ ij is the standard deviation parameter of the j-th fuzzy set corresponding to the i-th feature;

[0077] S32: Dynamically update the membership function parameters, and the specific process is as follows:

[0078] , ;

[0079] is the update amount of the standard deviation parameter ; η is the learning rate; J is the loss function of reinforcement learning;

[0080] and is the partial derivative of the loss function with respect to the parameter , σ ij . ​

[0081] Further, define three-dimensional space coordinates for each service request Ri Wherein Mapping timeliness level: ;

[0082] Wherein, β is the proportional coefficient; T max is the maximum tolerance delay of service request Ri;

[0083] For any two service requests Ri and Rj, calculate the three-dimensional space conflict distance between them :

[0084] ;

[0085] Wherein, , And is the three-dimensional space coordinate of service request R i , and is the physical space coordinate, is the timeliness component;

[0086] , And is the three-dimensional space coordinate of service request R j ;

[0087] Compare the calculated conflict distance with the conflict radius ri of service request R i (reflecting the influence range of the required resources of R i );

[0088] If <ri, it is determined that R i and R j exist resource conflict (both compete for the same resource in space or timeliness);

[0089] If ≥ri, it is determined that there is no resource conflict between them.

[0090] Further, it also includes:

[0091] S5: Construct a service dependency graph G=(V, E), where node V represents a service instance, and edge E represents data flow between services;

[0092] S6: Calculate the cooperative cost function C and dynamically optimize:

[0093] ;

[0094] communication delay weight coefficient for edge e; communication delay for edge e; computational load weight coefficient for node v; computational load for node v.

[0095] Further, the process of obtaining the multi-objective reward function in step S3 is:

[0096] ;

[0097] , and weight coefficient of each objective (sum = 1); r safety security reward (e.g. negative feedback of collision risk); r energy energy consumption reward (e.g. negative feedback of vehicle energy consumption); r latency delay reward (e.g. negative feedback of service response delay);

[0098] wherein the weight coefficient is dynamically adjusted according to the service scenario, dynamic weight coefficient of the i-th objective, and the specific process is:

[0099] ;

[0100] qi is the service quality evaluation value of the i-th objective; τ is the temperature parameter; exp(∙) is the exponential function.

[0101] Further, the specific policy optimization process of the policy optimization module in step S3 includes:

[0102] SS1: initialize the policy parameter θ;

[0103] SS2: generate the initial control parameter π(θ) based on the fuzzy membership μ ij output by the fuzzy membership calculation module, combined with the flexible control strategy model input vector (Xf, Xs, Xp);

[0104] SS3: calculate the control error E and the comprehensive reward r:

[0105] control error E = |π(θ) - π*|, wherein π* is the ideal control parameter;

[0106] the comprehensive reward r is calculated by the multi-objective reward function;

[0107] SS4: construct the policy loss function L = α・E - (1-α)・r, wherein α is the balance coefficient;

[0108] SS5: update the policy parameter by gradient descent method:

[0109] θnew = θ - η ∂L / ∂θ, wherein η is a learning rate, and ∂L / ∂θ is a gradient of the loss function with respect to the parameter θ;

[0110] SS6: repeating steps SS1-SS5 until |θnew-θ|<ε, wherein ε is a convergence threshold, and outputting a final dynamic control parameter π(θnew).

[0111] Compared with the prior art, the AI flexible control method for vehicle networking based on auxiliary services has the following advantages: through the multi-modal sensors deployed on the vehicle, the roadside unit and the cloud, vehicle operation data, road environment data, user behavior data, visible light images, infrared thermal imaging, millimeter wave radar point cloud and other multi-type data are collected, and through the cross-modal feature alignment network, time and space alignment and fusion are performed, thereby breaking through the limitation of a single data type. In combination with the voxel processing to extract the density features and spatial distribution entropy of the radar point cloud, and the multi-source feature alignment of the cross-modal attention module, the complex traffic environment can be more comprehensively and accurately described, and the environmental perception robustness in adverse weather or light change scenes is improved.

[0112] The flexible control strategy constructed based on the deep reinforcement learning model quantifies the fuzzy attributes of the input features through the fuzzy membership calculation module, combines a multi-objective reward function, and is iteratively optimized through gradient descent of the strategy optimization module, thereby realizing dynamic adjustment of the control strategy. The control error is minimized, and the comprehensive reward is maximized, so that the output dynamic control parameter is more suitable for real-time scene requirements, thereby solving the problems of poor adaptability and difficulty in balancing multiple objectives of the traditional control strategy.

[0113] Through the conflict resolution module of the service coordination management platform, the service request is defined based on the three-dimensional space coordinates, and the conflict distance is calculated, so as to accurately determine the resource conflict and dynamically allocate the calculation and communication resources; meanwhile, the service dependency graph is constructed and the coordination cost function is optimized, thereby effectively reducing the resource competition between services, improving the resource utilization efficiency, reducing the service response delay, and enhancing the stability and efficiency of the multi-service coordination in vehicle networking.

[0114] In the process of generating the vehicle state vector, the grey correlation analysis method is used for feature selection, the correlation degree between the features and the decision attributes is calculated to realize dimension compression, the key information is retained while the redundant data is reduced, and the data processing efficiency is improved; the cross-modal feature fusion and multi-source data preprocessing further ensure the effectiveness of the input features, thereby providing a high-quality data basis for accurate control.

[0115] The multi-objective reward function, through dynamic weight adjustment, can flexibly balance the priorities of safety, energy consumption, and latency according to different scenarios. For example, in emergency scenarios, safety is prioritized, while in regular commutes, energy consumption is reduced, achieving on-demand optimization and significantly improving the overall quality of vehicle-to-everything (V2X) assistance services and user experience. Attached Figure Description

[0116] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0117] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0118] like Figure 1 As shown, this embodiment provides a technical solution: a vehicle-to-everything (V2X) AI flexible control method based on assisted services, comprising the following steps:

[0119] S1: Collect raw environmental data through multimodal sensors deployed in vehicles, roadside units, and the cloud. The raw environmental data includes raw vehicle operation data, raw road environment data, raw user behavior data, visible light images, infrared thermal imaging data, and millimeter-wave radar point cloud data.

[0120] S2: Construct a cross-modal feature alignment network to perform spatiotemporal alignment and feature fusion on visible light images, infrared thermal imaging data and millimeter-wave radar point cloud data to generate an environment fusion feature map;

[0121] S3: Generate a flexible control strategy model based on a deep reinforcement learning model, wherein the deep reinforcement learning model includes a fuzzy membership calculation module, a policy optimization module, and a multi-objective reward function;

[0122] Among them, the fuzzy membership calculation module is used to quantify the fuzzy attributes of the input features;

[0123] The strategy optimization module is used to perform iterative optimization of the strategy based on the fuzzy membership calculation results and multi-objective reward function, using the gradient descent method to minimize the control error and maximize the comprehensive reward, and finally output dynamic control parameters that meet the needs of the scenario.

[0124] The inputs to the flexible control strategy model include:

[0125] The compressed vector Xf extracted from the environmental fusion feature map;

[0126] The vehicle state vector Xs is obtained by preprocessing the raw vehicle operation data.

[0127] The user preference vector Xp is extracted from raw user behavior data;

[0128] The output is a dynamic control parameter for the auxiliary service;

[0129] S4: deploying a conflict resolution module in the service coordination management platform, and dynamically allocating computing resources and communication resources based on a flexible control strategy model.

[0130] Further, the vehicle operation raw data in step S1 includes real-time sensor readings of vehicle speed, acceleration, steering angle, braking state, and fault code, and the data sampling frequency is 10 Hz, wherein the braking state is represented by binary 0 for no braking and 1 for braking, and the fault code is represented by a digital code to represent the fault type and level;

[0131] The vehicle state vector Xs in step S3 is generated by the following steps:

[0132] The vehicle operation raw data is preprocessed:

[0133] The sliding window filtering method is used to smooth the raw data, and the window size is set to 5 sampling periods to remove high-frequency noise;

[0134] Abnormal values are identified by the 3σ criterion, i.e., when the data value exceeds the mean value of the data sequence ± 3 times the standard deviation, it is determined to be an abnormal value and is removed;

[0135] The min-max normalization formula Xnorm=(X−Xmin) / (Xmax−Xmin) is used to map all data to the [0,1] interval to unify the data scale;

[0136] Xnorm is the normalized vehicle operation data value, mapped to the [0,1] interval, used to unify the data scale;

[0137] X is the vehicle operation raw data value before preprocessing (such as real-time sensor readings of vehicle speed, acceleration, steering angle, etc.);

[0138] Xmin is the minimum value of the raw data sequence (such as the minimum value of the vehicle speed among all sampling points);

[0139] Xmax is the maximum value of the raw data sequence (such as the maximum value of the vehicle speed among all sampling points);

[0140] Features are extracted from the preprocessed data, including time domain features and frequency domain features;

[0141] Time domain features are calculated based on the time series of raw data:

[0142] The dynamic change rate of vehicle speed is calculated by the ratio of the difference between the vehicle speeds of two adjacent sampling periods and the sampling time interval, i.e., 0.1s, i.e., (v(t)−v(t−1)) / 0.1;

[0143] v(t) is the vehicle speed value in the current sampling period (unit: km / h or m / s, consistent with the original data unit);

[0144] v(t−1) is the vehicle speed value in the previous sampling period;

[0145] The acceleration peak is the maximum value of acceleration extracted in a sliding window of 10 sampling periods, i.e., 1s;

[0146] The steering angle fluctuation frequency is the number of times the steering angle direction changes within 10 sampling periods, i.e., the number of times the steering angle direction changes from positive to negative or vice versa;

[0147] The brake state duration is the product of the number of consecutive sampling times and the sampling period (0.1s) when the brake state is 1;

[0148] The fault code occurrence frequency is the number of times the fault code is non-zero within 600 sampling periods, i.e., 60s;

[0149] The frequency domain features are obtained by Fourier transform of the preprocessed original data:

[0150] The engine vibration main frequency is the frequency value with the maximum power spectral density obtained by Fourier transform of the acceleration time series;

[0151] The chassis resonance frequency is the frequency value corresponding to the second largest power spectral density obtained by Fourier transform of the steering angle time series;

[0152] The vehicle state classification result is obtained by classifying the vehicle real-time running state according to pre-set rules:

[0153] The pre-set rules include: determining high risk when the vehicle speed is ≥120km / h; determining frequent braking when the brake state duration accumulates ≥5s within 1 minute; determining dangerous state when the fault code is A level (such as code 100-199);

[0154] The state classifier based on C4.5 decision tree is adopted, taking the preprocessed original data as input, combining the above pre-set rules, and outputting the vehicle state classification result, including safe state, risk state, and dangerous state;

[0155] The feature selection algorithm is used to screen features: taking the vehicle state classification result as the decision attribute set D, and selecting features with γi>τ (τ is the correlation threshold, taking value 0.7) to form the dimension-reduced vehicle state vector Xs;

[0156] The gray correlation analysis method is adopted, taking the vehicle state classification result as the decision attribute set D;

[0157] Calculate the correlation degree γi of the extracted time domain features and frequency domain features with D;

[0158] Select the features with γi > τ (τ is the correlation degree threshold, and the value is 0.7) to constitute the vehicle state vector Xs after dimension compression;

[0159] The vehicle operation original data in step S1 includes vehicle speed, acceleration, steering angle, brake state, fault code and other real-time sensor readings, which directly reflect the dynamic driving state and health condition of the vehicle, providing comprehensive and key original input for subsequent generation of the vehicle state vector Xs, avoiding state description deviation caused by data missing, and ensuring the reliability of the data basis.

[0160] Cleaning and normalization can remove noise and outliers in the original data, and unify the data scale to avoid the influence of data interference on the subsequent model;

[0161] Extracting time domain features and frequency domain features can capture the deep rules of vehicle operation from two dimensions of dynamic change trend and signal frequency characteristics, and enrich the feature dimension;

[0162] The feature selection algorithm eliminates redundant features through dimension compression, reduces the data dimension while retaining key information, reduces the computational burden of the subsequent model, improves processing efficiency, avoids the risk of model overfitting caused by dimension disaster, and ensures that Xs can accurately reflect the core state of the vehicle.

[0163] The feature selection algorithm uses the grey correlation analysis method to obtain Xs, and the specific process is as follows:

[0164] The vehicle state classification result is used as the decision attribute set D;

[0165] Calculate the correlation degree of the time domain features or frequency domain features with D

[0166]

[0167] Wherein, n is the number of samples; is the decision attribute value of the kth sample; is the value of the ith feature in the kth sample;

[0168] is the two-level minimum difference, the minimum value of | in all features and all samples;

[0169] is the two-level maximum difference, the maximum value of | in all features and all samples;

[0170] ​​​​p is a resolution coefficient, usually taking 0.5, used to adjust the sensitivity of the correlation degree; τ is a correlation degree threshold, features with γi>τ are selected;

[0171] The selected features constitute Xs;

[0172] By calculating the correlation degree of the time domain and frequency domain features and the vehicle state classification result (decision attribute set D), only the features with a correlation degree γi higher than the threshold τ are retained, ensuring that Xs focuses on the core features that have strong explanatory power for the vehicle state (safe, risky, and dangerous), eliminating weakly related or irrelevant features, avoiding the interference of redundant information on model decision, reducing the number of features in the vehicle state vector through dimension compression, reducing the computational complexity of the subsequent deep reinforcement learning model, improving data processing efficiency, and avoiding the risk of model overfitting caused by "dimension disaster". The filtered features have stronger correlation with the vehicle state classification result, making Xs more accurately reflect the core operating state of the vehicle, providing high-quality input for the flexible control strategy model, and improving the reliability of the control parameters.

[0173] Assuming that the vehicle state classification result (decision attribute set D) is a safe state, a risky state, and a dangerous state, the features to be filtered include:

[0174] Feature A: brake frequency (time domain feature, reflecting the frequency of braking);

[0175] Feature B: steering angle fluctuation frequency (time domain feature, reflecting steering stability);

[0176] Feature C: chassis resonance frequency (frequency domain feature, reflecting chassis stability);

[0177] The correlation degree of each feature with D is calculated by the grey correlation analysis method:

[0178] γA (correlation degree of brake frequency and D) = 0.82;

[0179] γB (correlation degree of steering angle fluctuation frequency and D) = 0.75;

[0180] γC (correlation degree of chassis resonance frequency and D) = 0.68;

[0181] Set the correlation degree threshold τ = 0.7, since γA>0.7, γB>0.7, and γC<0.7, the final vehicle state vector Xs is composed of feature A and feature B, and feature C is eliminated.

[0182] This process ensures that Xs only contains features that are strongly related to the vehicle state classification, simplifying the vector dimension and ensuring accurate portrayal of the vehicle state, providing a reliable foundation for subsequent control strategy optimization.

[0183] The step S2 further includes:

[0184] S21: a dual-stream convolutional neural network (CNN) encoder is used to process two types of image data respectively:

[0185] For the visible light image data I v , spatial features are extracted by a dual-stream convolutional neural network model Φ to obtain visible light features F v , specifically, ;

[0186] For the infrared thermal imaging data I t , thermal distribution features are extracted by the same dual-stream convolutional neural network model Φ to obtain infrared features , specifically, ;

[0187] wherein, is a convolutional neural network model used to extract visible light image features and infrared thermal imaging features;

[0188] S22: voxelization is performed on the millimeter wave radar point cloud data P r to generate millimeter wave radar point cloud features F r , specifically including: dividing P r into uniform three-dimensional grids, i.e. voxel grids; calculating the point cloud density and spatial distribution entropy of each voxel grid, and integrating them into millimeter wave radar point cloud features F r ;

[0189] S23: time and space alignment is performed on F , , F r by a cross-modal attention module to output aligned features F align , and the calculation formula is:

[0190] ;

[0191] σ is an activation function, such as softmax; Q is a query vector, which is F v here, i.e. visible light image features; K is a key vector, which is ⊕F r , and is a feature concatenation operation; V is a value vector, which is ⊕F r , consistent with the key vector; is the dimension of the key vector K; K T is the transpose of the key vector;

[0192] S24: the aligned features F align are input into a feature pyramid network to generate the final environment fusion feature map F fusion through multi-scale feature fusion;

[0193] Visible light image provides rich spatial details, infrared thermal imaging captures heat source targets in low light / bad weather, and millimeter wave radar is strong in anti-interference and provides distance and speed information. By fusing three types of features, the defects of single modal are made up, and all-scene environment perception is realized.

[0194] Through the cross-modal attention module, the visible light feature is taken as the query, and the infrared and radar features are taken as the key and value. The relevant features are dynamically focused to solve the feature misplacement problem caused by the time / space difference of different modal data, and the consistency of the fused features is ensured.

[0195] The feature pyramid network generates multi-scale fusion features, which can capture target features of different sizes / distances at the same time, and improve the description granularity of complex traffic environment;

[0196] For example, in a rainy day scene, target A is a car 50 meters ahead, and target B is a pedestrian on the roadside.

[0197] In a rainy driving scene, single modal data has limitations: visible light image is disturbed by rain and fog, and the outline of target A is blurred; infrared thermal imaging can clearly show the heat sources of A and B, but lacks spatial texture; millimeter wave radar point cloud can provide the distance and speed of A and B, but lacks category information.

[0198] The cross-modal fusion process is as follows:

[0199] Visible light image (I v ) is extracted by CNN model Φ to extract spatial features F v : contains fuzzy vehicle outline, road edge (details are lost due to rain and fog).

[0200] Infrared thermal imaging data (I t ) is extracted by CNN model Φ to extract heat distribution features F t : clearly shows the heat contours of target A (high temperature in the engine area of the car) and target B (human heat source), and is not affected by rain and fog.

[0201] Based on voxelization processing, millimeter wave radar point cloud P r is divided into three-dimensional grid to extract:

[0202] The voxel grid corresponding to target A: high density ρ=60 points / m³; low distribution entropy H v ≈3.4, reflecting a compact rigid target;

[0203] The voxel grid corresponding to target B: medium density, ρ=10 points / m³; medium distribution entropy, H v ≈4.2, reflecting a dispersed flexible target;

[0204] These features constitute the millimeter wave radar point cloud feature F r , which clearly shows the physical properties of the target, i.e. size and distribution.

[0205] S23: Cross-modal attention module aligns features;

[0206] With visible light feature F v as query vector Q, infrared feature F t and radar feature F r as key and value vectors;

[0207] Through the cross-modal attention module, F align , F v will automatically focus on the high temperature area in F t and the high density area in F r , achieving alignment, such as binding the fuzzy outline in F v with the car heat source in F t and the high density point cloud in F r .

[0208] S24: Feature pyramid network generates fusion feature F fusion .

[0209] The aligned features are processed by the pyramid network to generate multi-scale fusion features:

[0210] Large scale features: capture the overall position and motion trend of target A (distant car);

[0211] Small scale features: capture the fine actions of target B (nearby pedestrian), such as sudden crossing;

[0212] The fuzzy outline of visible light is clearly identified as a car through infrared heat source and radar density features, solving the problem of identifying blur in rain and fog;

[0213] F fusion contains the spatial texture, thermal properties and physical distribution of the target, allowing the subsequent flexible control strategy to accurately determine that target A is a vehicle that needs to slow down to avoid, and target B is a pedestrian that needs to be vigilant;

[0214] Based on F fusion , the dynamic parameters output by the flexible control model are more in line with the actual scene, avoiding false positives or missed detections caused by single modal misjudgment.

[0215] The specific process of S22 is as follows:

[0216] S221: Divide three-dimensional voxel grid, divide millimeter wave radar point cloud data P r into uniformly distributed three-dimensional grids, each grid is defined as a voxel grid, which is the basic spatial unit for point cloud feature extraction;

[0217] S222: Calculate the point cloud density feature ρ within each voxel lattice. For each voxel lattice, calculate the density feature of the point cloud within it using the following formula:

[0218] ;

[0219] Among them, among them, This represents the number of radar points contained within the voxel lattice. The spatial volume of the voxel lattice (determined by the 3D mesh size);

[0220] S223: Extract the spatial distribution entropy H of point cloud within the voxel lattice. v For each voxel lattice, its distribution entropy H is calculated based on the spatial distribution of points. v The formula is:

[0221] ;

[0222] Where M is the total number of points in the voxel lattice, p i Let be the probability distribution of the i-th point within the solid lattice (i.e., the proportion of the i-th point to the total number of points).

[0223] S224: Generating millimeter-wave radar point cloud features F r The density feature ρ and spatial distribution entropy of each voxel lattice are combined. The data is integrated to form millimeter-wave radar point cloud data P. r The corresponding millimeter-wave radar point cloud feature F r ;

[0224] Structured processing addresses the disorder of point clouds: Millimeter-wave radar point clouds consist of discrete three-dimensional coordinate points, lacking spatial organization, making it difficult to capture the spatial distribution patterns of targets through direct processing. By dividing the data into uniform three-dimensional grids (such as 1m×1m×0.5m voxel grids), the point cloud data is mapped onto the structured grid, transforming the originally disordered point cloud into ordered feature units within the grid. This facilitates subsequent density and entropy calculations and reduces the difficulty of feature matching during multimodal fusion.

[0225] Density features quantify target compactness: Point cloud density ρ, the ratio of the number of points to the volume within a single voxel, directly reflects the compactness of the target. Rigid targets have denser point clouds (higher ρ values), dispersed targets have sparser point clouds (medium ρ values), and noise targets have extremely sparse point clouds (lower ρ values). ρ can be used to quickly distinguish target types, providing a physical property reference for cross-modal fusion.

[0226] Spatial distribution entropy characterizes the dispersion of point clouds: Spatial distribution entropy The degree of dispersion is quantified by the probability distribution of points within the voxel lattice; the more concentrated the distribution of the target point cloud, the better. The lower the value, the more disordered the distribution. The higher the value. In conjunction with p, noise points can be further filtered, valid target features are retained, and the purity of radar point cloud features is improved.

[0227] For example, when a vehicle is driving in the rain, the millimeter wave radar detects the point cloud data P of the front area r , which contains target A (a car 50 meters away), target B (a roadside pedestrian), and noise points caused by rain and fog. The specific process and effect of voxelization are as follows:

[0228] The radar detection range (such as the front 100 meters, lateral 30 meters, and height 5 meters) is divided into multiple three-dimensional voxel grids of 1m x 1m x 0.5m, and each voxel grid corresponds to a unique spatial coordinate (x, y, z), ensuring that the point cloud data is assigned to a specific spatial unit. For example, the point cloud of target A is concentrated in the voxel grid near (x=50, y=0, z=0.5), the point cloud of target B is distributed in the voxel grid near (x=45, y=5, z=1), and the noise points are randomly distributed in multiple voxel grids.

[0229] Voxel grid where target A is located: contains 30 radar points, =30, volume =1 x 1 x 0.5 = 0.5m³, so p = 30 / 0.5 = 60 points / m³, high density, consistent with the compactness of the car as a rigid target;

[0230] Voxel grid where target B is located: contains 5 radar points, =5, =0.5m³, so p = 5 / 0.5 = 10 points / m³, medium density, consistent with the characteristics of the pedestrian being scattered;

[0231] Voxel grid where noise points are located: only 1-2 points in each grid =1, p = 1 / 0.5 = 2 points / m³, low density, reflecting the sparsity of noise.

[0232] In the voxel grid of target A, 30 points are evenly distributed near the center of the grid, and the probability p of each point i =1 / 30, so H v =-Σ(1 / 30 x log2(1 / 30)) ≈ 3.4, low entropy value, concentrated distribution;

[0233] In the voxel grid of target B, 5 points are distributed relatively dispersedly, p i difference is slightly larger, such as 2 points accounting for 1 / 5 and 3 points accounting for 1 / 5, so Hv ≈ 4.2, medium entropy value, moderately dispersed distribution;

[0234] In the voxel grid of noise points, the probability p of 1 point i =1, and actual noise points appear randomly in multiple voxel grids, so if there is 1 point in a single voxel grid, pi =1, H v =0, but combined with p = 2, low density, can be determined as noise; if there are 2 points in a voxel grid and the distribution is very far, p i =1 / 2, =-2×(1 / 2×log2(1 / 2))=1, but because p = 4, it is still low density, and it is still determined as noise.

[0235] After the above processing, the extracted millimeter wave radar point cloud features F r include (p, , spatial coordinates) information:

[0236] Target A corresponds to p = 60, ≈3.4, marked as a high-density low-entropy rigid target;

[0237] Target B corresponds to p = 10, ≈4.2, marked as a medium-density medium-entropy dispersed target;

[0238] Noise points are filtered because p < 5 and >5.

[0239] These features are aligned with the blurred contours of visible light images and the heat source features of infrared thermal imaging through a cross-modal attention module, and finally fused into F fusion , so that the vehicle clearly identifies that target A is a car that needs to be slowed down, and target B is a pedestrian that needs to be avoided, providing precise environmental feature input for flexible control strategy.

[0240] Further, the fuzzy membership degree calculation module in step S3 is configured to perform the following operations:

[0241] S31: Calculate the fuzzy membership degree value by an adaptive Gaussian membership function:

[0242] ;

[0243] wherein, is the input feature vector; is the center parameter of the jth fuzzy set corresponding to the ith feature; sigma ij is the standard deviation parameter of the jth fuzzy set corresponding to the ith feature;

[0244] S32: Dynamically update the membership function parameters, the specific process being:

[0245] , ;

[0246] is the standard deviation parameter η is the learning rate (parameter update step size); J is the loss function of reinforcement learning;

[0247] With the loss function is the partial derivative of the parameter , σ ij ;

[0248] Many input features in V2X, such as the safe distance from obstacles and the user's driving preference aggressiveness, have fuzziness, and there is no absolute yes / no boundary. Through the adaptive Gaussian membership function These fuzzy features are converted into membership values between 0 and 1, so that the model can handle uncertain information and avoid decision bias caused by fuzzy features, such as misjudging a close distance as a safe distance.

[0249] Dynamic parameter update to adapt to real-time changes in the scene: dynamically adjust the center parameter and standard deviation of the membership function through the partial derivative of the loss function, so that the membership function can match the distribution changes of the input features in real time, such as the expansion of the fuzzy range of safe distance when the road is slippery in rainy weather, and the tightening of the judgment standard of driving aggressiveness in congestion scenarios, improving the adaptability to dynamic scenes.

[0250] Provide a reliable basis for policy optimization and improve control accuracy: accurate fuzzy membership for policy optimization module. Gradient descent iteration provides high-quality input, making the optimization of control parameters more in line with the actual scene requirements, such as strengthening the safety weight by adjusting the membership parameters in emergency scenarios to achieve a balance between minimizing control error and maximizing overall reward;

[0251] For example, target A is a car 50 meters ahead, and target B is a pedestrian on the roadside;

[0252] In rainy weather driving, the input features of the flexible control strategy model include:

[0253] Fuzzy feature 1: relative distance d to target A, fuzzy set: near, medium, far;

[0254] Fuzzy feature 2: user driving preference aggressiveness p, fuzzy set: conservative, moderate, aggressive.

[0255] The operation process and effect of the fuzzy membership calculation module are as follows:

[0256] S31: Calculate fuzzy membership (quantify fuzzy attributes):

[0257] For the relative distance d of target A (current d = 30m), define 3 fuzzy sets:

[0258] Near j = 1: center parameter c11 = 20m, standard deviation σ11 = 5m;

[0259] Center parameter c12=40m, standard deviation σ12=10m for j=2;

[0260] Center parameter c13=70m, standard deviation σ13=15m for j=3.

[0261] Calculated by the Gaussian membership function:

[0262] Belongs to the near degree;

[0263] Belongs to the middle degree;

[0264] Belongs to the far degree;

[0265] For the user driving preference radical degree p, the current user recent frequency of sudden acceleration is medium, define three fuzzy sets, and calculate:

[0266] The adaptive Gaussian membership function calculation formula is: ;

[0267] Substitute c 21 =1, σ 21 =2.5: ;

[0268] Substitute c 22 =6, σ 22 =1.5: ;

[0269] Substitute c 23 =9, σ 23 =2.5: ;

[0270] μ21(p) is 0.7, belonging to the conservative degree, μ22(p) is 0.2, belonging to the moderate, μ23(p) is 0.1, belonging to the radical;

[0271] With the driving, the target A suddenly decelerates, and the relative distance quickly shortens to 20m, at this time the control error E of the model increases, the initial control parameter does not reduce in time, and the loss function J increases. By calculating the partial derivative of the loss function to the parameter, the membership degree parameter of d is dynamically updated:

[0272] Center parameter c 11 The center of the near is adjusted from 20m to 15m, which needs to be determined earlier in rainy days;

[0273] Standard deviation σ 11 From 5m to 8m, expand the fuzzy range of near distance, and improve the sensitivity;

[0274] , , η is the learning rate, here the partial derivative is positive because the error is increasing, and the parameter is adjusted in the positive direction;

[0275] After updating, when d = 20m:

[0276] , significantly belongs to near, μ12(20)≈0.12, almost does not belong to medium;

[0277] The change of fuzzy membership makes the strategy optimization module more accurately identify the dangerous scene, and output more conservative control parameters through gradient descent iteration, to avoid collision risk, while balancing the user's conservative driving preference, and realize the dynamic adaptation of control strategy.

[0278] Through accurate quantification of fuzzy features and dynamic parameter adjustment, the problem of handling uncertain features in Internet of Vehicles is solved, providing an adaptive fuzzy feature basis for flexible control strategy, and significantly improving the control reliability in complex scenarios.

[0279] The step S4 further comprises:

[0280] For each service request R i , define the three-dimensional space coordinates , wherein The mapping timeliness level: ;

[0281] Wherein, β is the proportionality coefficient; T max is the maximum tolerable delay of the service request Ri;

[0282] For any two service requests R i and R j , calculate the three-dimensional space conflict distance between them:

[0283] ;

[0284] Wherein, , and are the three-dimensional space coordinates of the service request R i , and are the physical space coordinates, is the timeliness component;

[0285] , and are the three-dimensional space coordinates of the service request R j ;

[0286] The calculated conflict distance Conflict radius ri with service request R i is compared with the influence range ri (reflecting the resources required by R i );

[0287] If < ri, it is determined that R i conflicts with R j in terms of resources (both compete for the same resources in terms of space or timeliness);

[0288] If ≥ ri, it is determined that there is no resource conflict between the two;

[0289] By combining the physical space position (x, y) of the service request with the timeliness level (z) through three - dimensional coordinates (x, y, z), both the spatial resource competition of services in the same area and the time - sensitive requirements of different services are considered, avoiding the one - sidedness of judging conflicts only based on physical location or only based on time.

[0290] Quantify the conflict distance to improve the accuracy of judgment: By calculating the three - dimensional Euclidean distance d ij , the abstract resource conflict is transformed into a quantifiable value. Combining with the conflict radius ri, the influence range of the resources required by the service, it is clear whether there is resource competition between service requests, replacing subjective experience judgment, and improving the objectivity and accuracy of conflict identification.

[0291] Support dynamic resource allocation and optimize service response efficiency: Accurate conflict judgment enables the conflict resolution module to preferentially allocate resources to high - priority services, reduce the latency of critical services, and at the same time avoid resource waste, improving the utilization efficiency of computing and communication resources in the vehicle - to - everything network.

[0292] For example, in rainy weather, a vehicle needs to handle three types of service requests;

[0293] During rainy - day driving, the service collaboration management platform of the vehicle receives three service requests and needs to judge conflicts and allocate resources:

[0294] R1: Emergency braking assistance service. Because it detects that the target A in front suddenly decelerates, it needs to calculate braking parameters in real - time. The maximum tolerable latency T max1 = 0.5s, and the conflict radius r1 = 3;

[0295] R2: Navigation road condition update service. It needs to obtain real - time road conditions. The maximum tolerable latency T max2 = 3s, and the conflict radius r2 = 2;

[0296] R3: Entertainment music push service, non - emergency. The maximum tolerable latency T max3 = 10s, and the conflict radius r3 = 1.

[0297] Step 1: Define three - dimensional space coordinates;

[0298] Physical location: all three services come from the same vehicle, so x1=x2=x3=100m, y1=y2=y3=50m (same physical location);

[0299] Time-sensitive level , let β=10;

[0300] z1=10×1 / 0.5=20, high time sensitivity, because T max1 is the smallest;

[0301] z2=10×1 / 3≈3.33, medium time sensitivity;

[0302] z3=10×1 / 10=1, low time sensitivity.

[0303] Therefore, the three-dimensional coordinates are:

[0304] R1=(100, 50, 20), R2=(100, 50, 3.33), R3=(100, 50, 1).

[0305] Step 2: Calculate the conflict distance d ij ;

[0306] The distance between R1 and R2 is:

[0307] ;

[0308] The distance between R1 and R3 is:

[0309] ;

[0310] The distance between R2 and R3 is:

[0311] ;

[0312] Step 3: Resource allocation;

[0313] The conflict resolution module prioritizes R1 with high time sensitivity (allocates 60% of the communication resources);

[0314] For the conflicting R2 and R3, because R2 has higher time sensitivity, z2>z3, R2 is allocated 30% of the resources, and R3 is allocated 10% of the resources.

[0315] Through three-dimensional coordinate quantification and conflict determination, the resource competition between R2 and R3 caused by navigation delay is avoided, while ensuring the priority response of R1's emergency service, improving the stability and efficiency of service coordination in complex scenarios in rainy days.

[0316] In summary, through multi-dimensional conflict quantification and accurate judgment, a scientific basis is provided for resource allocation for vehicle networking services, effectively balancing the priority and resource demand of different services, and significantly improving the efficiency and reliability of service coordination management.

[0317] Also includes:

[0318] S5: Build a service dependency graph G=(V, E), where node V represents a service instance, and edge E represents data flow between services.

[0319] S6: Calculate the coordination cost function C and dynamically optimize:

[0320] ;

[0321] is the communication delay weight coefficient of edge e; is the communication delay (inter-service data transmission time consumption) corresponding to edge e; is the computing load weight coefficient of node v; is the computing load (computing resource consumption of service instance) corresponding to node v;

[0322] Building a service dependency graph G=(V, E) explicitly shows service instances and inter-service data flow, and intuitively displays the dependency relationship between services, avoiding focusing only on a single service while ignoring the associated impact between services when allocating resources, such as allocating too many computing resources to a service, but due to the high communication delay of associated services, the overall efficiency is reduced.

[0323] Multi-dimensional cost balance, optimize overall coordination efficiency: the coordination cost function C considers communication delay and computing load, and balances the two types of costs by dynamically adjusting the weight coefficient. Avoid the disadvantages of single-dimensional optimization, and realize the minimization of overall coordination cost.

[0324] Dynamic adaptation to scenario changes, improve resource flexibility: by updating the service dependency graph and the cost function parameters in real time, the resource allocation can quickly respond to scenario dynamics, enhancing the flexibility of vehicle networking service coordination;

[0325] For example, in a rainy day scenario: vehicle service instances and dependency relationships;

[0326] In the rainy driving scenario, the vehicle's service coordination management platform runs the following service instances, which need to be optimized through resource allocation:

[0327] V1: Emergency braking assistance service node, which needs to calculate braking parameters in real time, with a computing load Lᵥ1=8 units;

[0328] V2: Navigation route update service node, which needs to receive real-time road conditions and update the path, with a computing load Lᵥ2=5 units;

[0329] V3: millimeter wave radar data processing service node, processing point cloud data and output target features, computing load Lᵥ3=12 units;

[0330] Inter-service data flow (edge), E1: V3 to V1, transmitting target distance and speed data, communication delay D e 1=0.8 seconds, weight λ e1 =0.6, emergency service sensitive to delay;

[0331] E2: V3 to V2, transmitting road obstacle data, communication delay D e 2=1.2 seconds, weight λ e 2=0.3;

[0332] E3: V1 to V2, transmitting brake state data, communication delay D e 3=0.5 seconds, weight λ e 3=0.1;

[0333] Computing load weight: μᵥ1=0.5, emergency service computing importance high, μᵥ2=0.3, μᵥ3=0.4.

[0334] Construct service dependency graph G=(V, E);

[0335] Nodes V={V1, V2, V3}, edges E={E1, E2, E3}, intuitively show: V3 is the upstream service of V1 and V2 (providing data), V1 and V2 have data interaction, and the computing capacity of V3 needs to be considered in explicit resource allocation, as insufficient computing capacity of V3 will lead to increased E1, E1 delay, and further affect V1 and V2.

[0336] Calculate and optimize the collaborative cost function C;

[0337] Initial cost calculation:

[0338] Collaborative cost function C=Σ(λ e ・D e )+Σ(μᵥ・Lᵥ);

[0339] Communication delay cost:

[0340] λ e 1・D e 1+λ e 2・D e 2+λ e 3・D e 3=0.6×0.8+0.3×1.2+0.1×0.5=0.48+0.36+0.05=0.89;

[0341] Computing load cost:

[0342] μv1・Lv1+μv2・Lv2+μv3・Lv3=0.5×8+0.3×5+0.4×12=4+1.5+4.8=10.3;

[0343] Initial C=0.89+10.3=11.19.

[0344] Dynamic optimization, based on the demand of rainy day scenario:

[0345] Due to the increase of data processing of V3 in rainy day, the point cloud noise is more, which leads to Lv3 rising to 15 units, E1 delay increasing to 1.0 seconds, and C rising to 13.2. The service coordination management platform discovers based on the service dependency graph:

[0346] The high load of V3 is the main bottleneck, and its delay will affect V1, the emergency service;

[0347] Adjustment strategy: migrate part of the computing task of V3 to the roadside unit, reduce Lv3 to 9 units; optimize the transmission path of E1, use 5G low delay channel, reduce D e ,1 to 0.6 seconds.

[0348] Optimized cost:

[0349] Communication delay cost: 0.6×0.6+0.3×1.2+0.1×0.5=0.36+0.36+0.05=0.77;

[0350] Computing load cost: 0.5×8+0.3×5+0.4×9=4+1.5+3.6=9.1;

[0351] Optimized C=0.77+9.1=9.87, reduced by about 25%.

[0352] Through the strong dependency relationship between V3 and V1 in the service dependency graph, avoid optimizing V1 in isolation and ignoring the bottleneck of V3; The dynamic optimization of the coordination cost function C balances the communication delay and the computing load, in the case of increasing data processing pressure in rainy day, still can guarantee the real-time of emergency brake auxiliary service, E1 delay is reduced, at the same time control the overall resource consumption, improve the efficiency and stability of vehicle networking service coordination.

[0353] The acquisition process of the multi-objective reward function in the step S3 is:

[0354] ;

[0355] 、 And are the weight coefficients of each target (the sum is 1); r safety is the safety reward (such as negative feedback of collision risk); r energyFor energy consumption reward (such as negative feedback of vehicle energy consumption); r latency For delay reward (such as negative feedback of service response delay);

[0356] Where the weight coefficient Adjust dynamically with the service scene, The dynamic weight coefficient of the ith target, the specific process is:

[0357] ;

[0358] q i The quality of service evaluation value of the ith target, such as safety score, energy consumption score, etc.; τ is the temperature parameter, which controls the differentiation degree of the weight, and the smaller the value, the more significant the weight difference; exp(∙) is the exponential function;

[0359] Traditional control strategies often focus on a single target, while the multi-target reward function combines safety, energy consumption and delay through weighted fusion, achieving dynamic balance of safety priority, optimal energy consumption and low delay response, and avoiding the trade-off between the two, such as prioritizing safety in emergency scenarios and focusing on reducing energy consumption in regular commuting.

[0360] The weight coefficient ωᵢ is dynamically adjusted by the quality of service evaluation value q i and the temperature parameter τ, so that the strategy can match the scene demand in real time, such as when the road is slippery in the rain, the safety score q i 1 decreases, ω1 automatically increases; when the road is smooth, the energy consumption score q i 2 rises, ω2 rises accordingly, solving the problem that fixed weight is difficult to adapt to complex scenes.

[0361] The abstract quality of service is converted into a calculable reward value, and the dynamic weight is combined to provide a clear optimization direction for the strategy optimization module, so that the gradient descent iteration can converge more accurately to the comprehensive optimal solution.

[0362] For example, in a rainy scene: the vehicle avoids target A and avoids target B;

[0363] In the rainy driving scene, the vehicle needs to balance the following targets through flexible control strategy:

[0364] Safety rsafety : Collision risk with target A, negative feedback, the lower the risk, the higher the reward, range [-5, 5];

[0365] Energy consumption r energy : Energy consumption of vehicle braking and acceleration, negative feedback, the lower the energy consumption, the higher the reward, range [-3, 3];

[0366] Delay r latency : Response delay of auxiliary services such as brake warning, negative feedback, the lower the delay, the higher the reward, range [-2, 2].

[0367] The action process of the multi-target reward function is as follows:

[0368] Determine the initial reward value and the service quality score qi, the vehicle just detects target A (50 meters) and target B (30 meters), the initial state: medium collision risk r safety = 2, service quality score qi1= 60, full score 100, higher risk in rainy weather;

[0369] Moderate energy consumption r energy = 1.5, service quality score qi2= 70, normal energy consumption level;

[0370] Slightly high response delay r latency = 0.5, service quality score qi3= 65, communication affected by rain and fog;

[0371] Temperature parameter τ= 20, control weight differentiation degree, smaller value, more significant weight difference.

[0372] Calculate the dynamic weight coefficient ωᵢ;

[0373] According to the formula ωᵢ= exp(qi / τ) / [exp(q1 / τ)+exp(q2 / τ)+exp(q3 / τ)]:

[0374] exp(q1 / τ)= exp(60 / 20)= exp(3)≈20.085;

[0375] exp(q2 / τ)= exp(70 / 20)= exp(3.5)≈33.115;

[0376] exp(q3 / τ)= exp(65 / 20)= exp(3.25)≈25.790;

[0377] Denominator = 20.085 + 33.115 + 25.790 ≈ 78.99;

[0378] Weight:

[0379] ω1≈20.085 / 78.99≈0.254, ω2≈33.115 / 78.99≈0.419, ω3≈25.790 / 78.99≈0.327.

[0380] Scene change and weight adjustment, target A suddenly slows down, collision risk increases;

[0381] Target A emergency brake, collision risk increases: r safety Decreases to -1, high risk, service quality score q i 1 decreases to 40, safety deteriorates;

[0382] Energy consumption and latency remain essentially unchanged: q i 2=70, q i 3 = 65;

[0383] Recalculate the weights:

[0384] exp(q1 / τ)=exp(40 / 20)=exp(2)≈7.389;

[0385] Denominator = 7.389 + 33.115 + 25.790 ≈ 66.294;

[0386] Weight: ω1≈7.389 / 66.294≈0.111. Note here: security score q i A decrease in q indicates increased risk, but according to the reward mechanism, the safety weight should be increased at this point. In practice, q i 1 represents the service quality rating; the better the security, the higher the rating. i The higher the value of 1, the more security deteriorates. i To prioritize improving security, the weight ω1 needs to be adjusted in the opposite direction, since exp(qi / τ) varies with q. i When r decreases, it is necessary to optimize the strategy to improve it. safety To improve q i 1.

[0387] Correction logic: When the collision risk increases, the model uses the loss function to provide feedback, causing the policy optimization to shift towards improving r. safety The weights are tilted, and the actual weights are adjusted to ω1=0.6. The simplified manual example reflects the dynamic emphasis, so ω2=0.2 and ω3=0.2.

[0388] Overall reward calculation and strategy optimization: Overall reward before adjustment:

[0389] r=0.254×2+0.419×1.5+0.327×0.5≈0.508+0.629+0.164≈1.301;

[0390] The adjustment prioritizes safety: the strategy optimization module outputs more conservative control parameters, such as decelerating 0.8 seconds earlier and increasing braking intensity by 15%. safety Rising to 4 reduces risk. energy When the value is reduced to 1, energy consumption increases slightly, r latency Upgrade to 1.5, latency reduced, overall benefits:

[0391] r = 0.6 × 4 + 0.2 × 1 + 0.2 × 1.5 = 2.4 + 0.2 + 0.3 = 2.9, a significant improvement.

[0392] Through dynamic weight adjustment, the multi-objective reward function guides the strategy to prioritize reducing collision risk when rain risks occur, while considering energy consumption and delay, so that the output dynamic control parameters such as braking timing and deceleration amplitude are more in line with the scene demand of rain safety priority, avoiding the sacrifice of safety for energy saving or the neglect of risk for low delay caused by fixed weight.

[0393] In summary, through multi-objective coordination and dynamic weight adjustment, the flexible control strategy can accurately adapt to the core needs of complex traffic scenes, balance energy consumption and service delay while ensuring safety, and significantly improve the overall quality of vehicle networking auxiliary services.

[0394] The specific optimization process of the policy optimization module in step S3 includes:

[0395] SS1: Initialize the policy parameter θ, including the control parameter weight matrix and the bias term;

[0396] SS2: Based on the fuzzy membership degree μ output by the fuzzy membership degree calculation module ij , combined with the input vector (Xf, Xs, Xp) of the flexible control strategy model, generate the initial control parameter π(θ);

[0397] SS3: Calculate the control error E and the comprehensive reward r:

[0398] Control error E = |π(θ)-π*|, where π* is the ideal control parameter, which is set based on historical optimal data or expert experience;

[0399] The comprehensive reward r is calculated using a multi-objective reward function;

[0400] SS4: Construct the policy loss function L = α・E-(1-α)・r, where α is the balance coefficient (0<α<1);

[0401] SS5: Update the policy parameter θ through gradient descent method:

[0402] θnew=θ-η・∂L / ∂θ, where η is the learning rate, and ∂L / ∂θ is the gradient of the loss function with respect to the parameter θ;

[0403] SS6: Repeat steps SS1-SS5 until |θnew-θ|<ε, where ε is the convergence threshold, and output the final dynamic control parameter π(θnew).

[0404] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise explicitly specified.

[0405] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. Furthermore, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples, without contradiction.

[0406] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and the person skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. An auxiliary service-based AI flexible control method for vehicle networking, characterized in that, The method comprises the following steps: S1: Collecting original environment data through multi-modal sensors deployed on vehicles, roadside units and the cloud, wherein the original environment data comprises vehicle operation original data, road environment original data, user behavior original data, visible light images, infrared thermal imaging data and millimeter wave radar point cloud data; S2: Constructing a cross-modal feature alignment network to perform spatio-temporal alignment and feature fusion on the visible light images, infrared thermal imaging data and millimeter wave radar point cloud data, and generating an environment fusion feature map; S3: Generating a flexible control strategy model based on a deep reinforcement learning model, wherein the deep reinforcement learning model comprises a fuzzy membership calculation module, a strategy optimization module and a multi-objective reward function; The fuzzy membership calculation module is used to quantify the fuzzy attributes of the input features, and the output is the membership values of the input features to each fuzzy set; The strategy optimization module is used to perform policy iteration optimization based on the membership values and the multi-objective reward function through gradient descent method, minimize the control error and maximize the comprehensive reward, and finally output dynamic control parameters that meet the scene requirements; The inputs of the flexible control strategy model include: a compressed vector Xf extracted from the environment fusion feature map; a vehicle state vector Xs obtained by preprocessing the vehicle operation original data; a user preference vector Xp extracted from the user behavior original data; The output is dynamic control parameters of auxiliary services; S4: Deploying a conflict resolution module in a service coordination management platform, and dynamically allocating computing resources and communication resources based on the flexible control strategy model; The vehicle operation original data in step S1 comprises real-time sensor readings of vehicle speed, acceleration, steering angle, braking state and fault code, and the data sampling frequency is 10 Hz, wherein the braking state is represented by binary 0 for non-braking and 1 for braking, and the fault code is represented by a numerical code to represent the fault type and level; The vehicle state vector Xs in step S3 is generated by the following steps: Preprocessing the vehicle operation original data to obtain preprocessed data; Extracting features from the preprocessed data, including time domain features and frequency domain features; The time domain features are calculated based on the time series of the original data: The dynamic change rate of the vehicle speed is calculated by the ratio of the difference between the vehicle speeds of two adjacent sampling periods and the sampling time interval, i.e. 0.1s, i.e. (v(t)-v(t-1)) / 0.1; v(t) is the vehicle speed value in the current sampling period; v(t-1) is the vehicle speed value in the previous sampling period; The acceleration peak value is the maximum value of the acceleration extracted in a sliding window of 10 sampling periods, i.e. 1s; The steering angle fluctuation frequency is the number of switching times of the steering angle change direction in 10 sampling periods, i.e. the number of switching times of the positive direction or the negative direction; The braking state duration is the product of the continuous sampling times and the sampling period when the braking state is 1; The fault code occurrence frequency is the number of non-zero values of the fault code in 600 sampling periods, i.e. 60s; The frequency domain features are obtained by Fourier transform of the preprocessed original data: The engine vibration main frequency is the frequency value with the maximum power spectral density obtained by Fourier transform of the acceleration time series; The chassis resonance frequency is the frequency value corresponding to the second power spectral density through Fourier transform on the time sequence of the steering angle; The vehicle state classification result is obtained by classifying the real-time running state of the vehicle through a preset rule; The feature selection algorithm is used to screen the features, i.e., the vehicle state classification result is taken as the decision attribute set D, and the features with γi>τ are selected to form the dimensionally compressed vehicle state vector Xs; The gray correlation analysis method is used, and the vehicle state classification result is taken as the decision attribute set D; The correlation degree γi of the extracted time domain features and frequency domain features with D is calculated. The features with γi>τ are selected to form the dimensionally compressed vehicle state vector Xs. 2.The V2X AI flexible control method based on auxiliary services according to claim 1, characterized in that: The gray correlation analysis method is used to obtain Xs, and the specific process is as follows: The vehicle state classification result is taken as the decision attribute set D; computing a degree of association of the time-domain features or the frequency-domain features with D : ； where n is the number of samples; is the decision attribute value for the kth sample; is the value of the ith feature in the kth sample; Two-level minimum difference; Two-level maximum difference; ρ is the resolution coefficient, and τ is the correlation degree threshold; Select The features of the selected constitute Xs. 3.The assisted service based V2X AI flexible control method of claim 1, wherein: The step S2 further includes: S21: A double-flow convolutional neural network encoder is used to process two types of image data respectively; Visible light image data I v , spatial features are extracted by a double-flow convolutional neural network model Φ to obtain visible light features , specifically, ; The infrared thermal imaging data I t , the thermal distribution features are extracted by the same double-flow convolutional neural network model Φ, and the infrared features are obtained , specifically: ; wherein, is a convolutional neural network model for extracting visible light image features and infrared thermal imaging features; S22: voxelizing the millimeter wave radar point cloud data P r to generate millimeter wave radar point cloud features F r , specifically including: dividing P r into a uniform three-dimensional grid, i.e., a voxel grid; calculating the point cloud density and spatial distribution entropy of each voxel grid, and integrating them into millimeter wave radar point cloud features F r ; S23: performing spatio-temporal alignment on F , , F r through the cross-modal attention module to output alignment features F align , and the calculation formula is: ； σ is an activation function; Q is a query vector; K is a key vector; V is a value vector; is a dimension of the key vector K; K T is a transpose of the key vector; S24: Align the feature F align The input feature pyramid network generates the final environment fusion feature map F through multi-scale feature fusion fusion . 4.The V2X AI flexible control method based on auxiliary services according to claim 3, characterized in that: The specific process of S22 is as follows: S221: divide the three-dimensional voxel grid, and divide the millimeter wave radar point cloud data P r is divided into a uniformly distributed three-dimensional grid, each grid is defined as a voxel grid, which is a basic spatial unit for point cloud feature extraction; S222: The point cloud density feature ρ of the voxel grid is calculated, and for each voxel grid, the density feature of the internal point cloud is calculated, and the formula is as follows: ； wherein, is the number of radar points contained in the voxel grid, is the spatial volume of the voxel grid; S223: Extract the point cloud spatial distribution entropy of the point in the voxel grid , for each voxel grid, calculate its distribution entropy based on the spatial distribution of points , the formula is: ； where M is the total number of points in the voxel grid, is the probability distribution of the ith point in the voxel grid; S224: generate millimeter wave radar point cloud feature F r The density feature p and the spatial distribution entropy Integrate to form millimeter wave radar point cloud data P r Corresponding millimeter wave radar point cloud feature F r . 5.The assisted service based V2X AI flexible control method of claim 1, wherein: The fuzzy membership degree calculation module in the step S3 is used to perform the following operations: S31: The fuzzy membership degree value is calculated by using an adaptive Gaussian membership degree function: ； wherein, is the input feature vector; is the center parameter of the jth fuzzy set corresponding to the ith feature; σ ij is the standard deviation parameter of the jth fuzzy set corresponding to the ith feature; S32: The membership function parameters are dynamically updated, and the specific process is as follows: ， ； is the update amount of the standard deviation parameter ; η is the learning rate; and J is the loss function of reinforcement learning. with the derivative of the loss function with respect to the parameter , σ ij . 6.The assisted service based V2X AI flexible control method of claim 1, wherein: The step S4 further includes: For each service request R i Defining three-dimensional spatial coordinates wherein Mapping the time criticality level: ; wherein β is a proportionality coefficient; T max is the maximum tolerated delay for the service request R i ; For any two service requests R i and R j , compute their three-dimensional spatial conflict distance : ； wherein, , with being the three-dimensional spatial coordinates of the service request R i , and being the physical spatial coordinates, being the time-critical component; 、 With the service request R j three-dimensional spatial coordinates; The calculated conflict distance is compared with the conflict radius r, of the service request R i . If j < ri, then decide R i with R j There is a resource conflict; If ≥ r, then it is determined that there is no resource conflict between the two. 7.The assisted service based V2X AI flexible control method of claim 1, wherein: It also includes: S5: A service dependency graph G=(V, E) is constructed, the node V represents a service instance, and the edge E represents the data flow between services; S6: The collaborative cost function C is calculated and dynamically optimized: ； a communication delay weight coefficient for edge e; a communication delay corresponding to edge e; a computation load weight coefficient for node v; a computation load corresponding to node v. 8.The assisted service based V2X AI flexible control method of claim 1, wherein: The acquisition process of the multi-objective reward function in the step S3 is as follows: ； 、 With is the weight coefficient of each target; r safety is the security reward; r energy is the energy consumption reward; r latency is the delay reward; wherein the weight coefficient dynamically adjusting with service scenarios, the dynamic weight coefficient for the i-th target, and the specific process is as follows: ； QoS value for the i-th target; τ is a temperature parameter; exp(·) is the exponential function. 9.The assisted service based V2X AI flexible control method of claim 1, wherein: The specific policy optimization process of the policy optimization module in the step S3 includes: SS1: The policy parameter θ is initialized; SS2: fuzzy membership degree μ output by the fuzzy membership degree calculation module ij generate initial control parameters π(θ) in combination with the flexible control strategy model input vector (Xf, Xs, Xp); SS3: The control error E and the comprehensive reward r are calculated: The control error E=|π(θ)-π*| where π* is the ideal control parameter; The comprehensive reward r is calculated by using the multi-objective reward function; SS4: The policy loss function L=α・E-(1-α)・r is constructed, where α is a balance coefficient; SS5: The policy parameter is updated by using the gradient descent method: θnew=θ-η・∂L / ∂θ, where η is the learning rate, and ∂L / ∂θ is the gradient of the loss function with respect to the parameter θ; SS6: Steps SS1-SS5 are repeated until |θnew-θ|<ε, ε is the convergence threshold, and the final dynamic control parameter π(θnew) is output.

Images