Decision-making method for intelligent predictive driving of connected vehicles based on machine learning

Abstract

The present invention relates to the field of predictive driving technology, and in particular to a decision-making method for intelligent predictive driving of connected vehicles based on machine learning. The method comprises: acquiring and preprocessing multi-source data, real-time data from the Internet of Vehicles, and historical driving data from connected vehicles; extracting features from the preprocessed multi-source data and real-time data from the Internet of Vehicles to obtain preliminary feature data; introducing a multi-layer recursive feature fusion algorithm to fuse the preliminary feature data to obtain final fused data; constructing and training a predictive driving model using a deep cross-adaptive prediction algorithm for intelligent predictive driving of connected vehicles; and using the predictive driving model to perform predictive processing on the final fused data to obtain driving prediction results and generate driving decisions. The method addresses the technical issues of inaccurate processing of multi-source heterogeneous data and poor predictive adaptability and low accuracy in driving decision making in intelligent predictive driving of connected vehicles.

Description

Technical Field

[0001] The present invention relates to the field of predictive driving technology, and in particular to a decision-making method for intelligent predictive driving of connected vehicles based on machine learning. Background Art

[0002] With the rapid development of intelligent connected vehicle (ICV) technology, vehicles are becoming increasingly intelligent. Driving assistance and autonomous driving systems based on onboard sensors and the Internet of Vehicles (IoV) are gradually being applied in real-world traffic scenarios. ICVs must make fast and accurate driving decisions in complex and dynamic traffic environments to ensure driving safety, enhance driving comfort, and optimize energy efficiency. However, current driving decision-making systems still face numerous technical challenges in data acquisition, information fusion, driving prediction, and real-time decision-making. Connected vehicles rely on a variety of sensors, including LiDAR, millimeter-wave radar, cameras, and ultrasonic sensors, to perceive their surroundings in real time. They also interact with other vehicles, infrastructure, and cloud platforms through IoV technology to obtain critical data such as traffic signals, road conditions, and weather conditions. This multi-source, heterogeneous data exhibits significant differences in numerical range, timestamp accuracy, and data structure. This can lead to information loss, inconsistency, and redundancy during the data fusion process, impacting the accuracy of subsequent driving predictions and decision-making.

[0003] However, existing intelligent predictive decision-making methods for connected vehicles have the following technical problems: the processing of multi-source heterogeneous data in intelligent predictive driving of connected vehicles is not accurate enough, and the prediction adaptability and accuracy of driving decisions are poor. Summary of the Invention

[0004] The present invention provides a decision-making method for intelligent predictive driving of connected vehicles based on machine learning to solve the technical problems of inaccurate processing of multi-source heterogeneous data in intelligent predictive driving of connected vehicles, as well as poor predictive adaptability and low accuracy of driving decisions.

[0005] The decision-making method for intelligent predictive driving of connected vehicles based on machine learning of the present invention specifically includes the following technical solutions:

[0006] The decision-making method for intelligent predictive driving of connected vehicles based on machine learning includes the following steps:

[0007] S1. Acquire and preprocess multi-source data, real-time data from the Internet of Vehicles, and historical driving data of connected vehicles to obtain preprocessed multi-source data, real-time data from the Internet of Vehicles, and historical driving data of connected vehicles; perform feature extraction on the preprocessed multi-source data and real-time data from the Internet of Vehicles to obtain preliminary feature data; and fuse the preliminary feature data using a multi-layer recursive feature fusion algorithm to obtain final fused data.

[0008] S2. Utilize the connected vehicle intelligent predictive driving prediction algorithm based on deep cross-adaptation to build and train a predictive driving model based on preprocessed historical driving data. Utilize the predictive driving model to perform predictive processing on the final fused data to obtain driving prediction results, generate driving decisions, and drive according to the driving decisions.

[0009] Preferably, the S1 specifically includes:

[0010] The multi-layer recursive feature fusion algorithm performs multi-level, recursive feature fusion on preliminary feature data, adopts a multi-level nested structure, and combines nonlinear mapping with recursive optimization strategy, and introduces an adaptive weight adjustment mechanism to fuse the preliminary feature data.

[0011] Preferably, the S1 specifically includes:

[0012] In the implementation process of the multi-layer recursive feature fusion algorithm, the preliminary feature data is standardized to obtain the standardized feature data, and the standardized feature data is used as input to enter the recursive feature transformation stage to obtain the final recursive feature transformed feature data. The specific implementation formula is:

[0013]

[0014] in, It is the feature data after the recursive feature transformation of the i+1th layer; is the feature data after recursive feature transformation of the i-th layer, W i is the weight of the recursive feature transformation at the i-th layer; b i is the bias term during the recursion of the i-th layer; α is the adaptive balance coefficient; γ is the frequency adjustment factor of the sinusoidal nonlinear transformation; X norm It is the standardized feature data.

[0015] Preferably, the S1 specifically includes:

[0016] In the implementation process of the multi-layer recursive feature fusion algorithm, after completing the recursive feature transformation, the feature data after the final recursive feature transformation enters the orthogonal feature fusion stage, and the feature data after the final recursive feature transformation is processed to obtain an orthogonal projection matrix; the orthogonal projection matrix is ​​used to perform redundancy removal operations on the feature data after the final recursive feature transformation to obtain preliminary non-redundant feature data.

[0017] Preferably, the S1 specifically includes:

[0018] In the implementation process of the multi-layer recursive feature fusion algorithm, after the orthogonal feature fusion, the initial non-redundant feature data enters the adaptive weight fusion stage for processing. The specific implementation formula is:

[0019]

[0020] in, It is The weight of the initial non-redundant feature data in the final fusion result; β is the weight smoothing factor; Representative The bi-norm of the initial non-redundant feature data; Represents the bi-norm of the j-th initial non-redundant feature data; is the total number of initial non-redundant feature data.

[0021] Preferably, the S1 specifically includes:

[0022] In the implementation process of the multi-layer recursive feature fusion algorithm, after the weight calculation, the initial non-redundant feature data is weighted and summed to obtain the fused feature data, and the fused feature data is subjected to feature data dimensionality reduction processing to obtain the final fused data.

[0023] Preferably, the S2 specifically includes:

[0024] In the process of implementing the intelligent predictive driving prediction algorithm of connected vehicles based on deep cross-adaptability, the corresponding environmental perception data and vehicle network data are obtained based on the acquired historical driving data. After preprocessing, the environmental perception data and vehicle network data are merged with the preprocessed historical driving data to obtain comprehensive preprocessed data.

[0025] Preferably, the S2 specifically includes:

[0026] After completing data preprocessing, a predictive driving model is constructed using multi-layer recursive cross-perception modeling. The modeling process is divided into a local perception layer, a global trend layer, and a cross-feedback layer.

[0027] Preferably, the S2 specifically includes:

[0028] During the training of the predictive driving model, a hierarchical game strategy is used to infer and optimize the driving decision-making strategy of the predictive driving model. After the predictive driving model training is completed, the final fused data is processed to generate driving prediction results.

[0029] The beneficial effects of the technical solution of the present invention are:

[0030] 1. A multi-layer recursive feature fusion algorithm is used to perform deep optimization and adaptive adjustment at the feature level. This can effectively resolve inconsistencies in scale, modality, and time synchronization among multi-source data, improve the accuracy and robustness of data fusion, and provide high-quality input for predictive driving models.

[0031] 2. Through a hierarchical modeling strategy, combining the local perception layer, the global trend layer, and the cross-feedback layer, it is possible to comprehensively capture short-term and long-term driving behavior characteristics. The local perception layer quickly extracts the vehicle's immediate driving behavior, such as acceleration, braking, and lane changes, within a short time window, while the global trend layer can monitor the driver's habits and the influence of the external environment over a long period of time, thereby enhancing the generalization ability of the predictive driving model in different scenarios and improving the stability and adaptability of driving prediction results.

[0032] 3. A multi-level recursive cross-update strategy is adopted to quickly adjust the prediction strategy when the real-time traffic environment changes, reduce short-term prediction errors, and avoid erroneous driving behaviors caused by sudden environmental changes. In complex traffic scenarios, such as sudden traffic accidents and temporary road closures, the driving prediction results can be quickly corrected, providing accurate and immediate decision-making support for the driving system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the decision-making method for intelligent predictive driving of connected vehicles based on machine learning as described in the present invention. DETAILED DESCRIPTION

[0034] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0035] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0036] The specific scheme of the decision-making method for intelligent predictive driving of a connected vehicle based on machine learning provided by the present invention is described in detail below with reference to the accompanying drawings.

[0037] Refer to the attached Figure 1 , which shows a flow chart of a decision-making method for intelligent predictive driving of a connected vehicle based on machine learning provided by one embodiment of the present invention, the method comprising the following steps:

[0038] S1. Acquire and preprocess multi-source data, real-time data from the Internet of Vehicles, and historical driving data of connected vehicles to obtain preprocessed multi-source data, real-time data from the Internet of Vehicles, and historical driving data of connected vehicles; perform feature extraction on the preprocessed multi-source data and real-time data from the Internet of Vehicles to obtain preliminary feature data; and fuse the preliminary feature data using a multi-layer recursive feature fusion algorithm to obtain final fused data.

[0039] Acquire multi-source data of connected vehicles from data acquisition devices such as LiDAR, millimeter-wave radar, cameras, and ultrasonic sensors, including vehicle dynamic data such as sensor data, vehicle speed, acceleration, steering angle, tire pressure, accelerator pedal position, and brake status, as well as location data such as GPS positioning information and vehicle latitude, longitude, and elevation. Acquire real-time Internet of Vehicles data from weather and traffic management systems through existing integrated API interfaces, including traffic signal information, traffic flow, road condition information, road closures, accident alerts, weather data, construction information, temporary road closures, vehicle-to-vehicle information exchange (such as vehicle speed and location), and communication between vehicles and roadside units (RSUs) (such as traffic light status and real-time road conditions). Acquire historical driving data from existing databases, including driving habit data such as sudden acceleration, sudden braking, and lane changes, environmental data such as current weather, road conditions, and traffic density, and driving trajectories such as the vehicle's historical driving paths.

[0040] The obtained multi-source data, real-time data from the Internet of Vehicles, and historical driving data from the connected vehicles are preprocessed, such as by data cleaning, denoising, normalization, and standardization, to obtain preprocessed multi-source data, real-time data from the Internet of Vehicles, and historical driving data from the connected vehicles. The technical means used in the preprocessing process are well known to those skilled in the art and are not described in detail here.

[0041] The pre-processed multi-source data of connected cars and real-time data of the Internet of Vehicles are further subjected to feature extraction using existing feature engineering technology to obtain time feature data such as time period and day-night transition, spatial feature data such as distance to obstacles and lane deviation, and driving behavior feature data such as acceleration mutation and braking frequency. The obtained time feature data, spatial feature data and driving behavior feature data are used as preliminary feature data. A multi-layer recursive feature fusion algorithm is introduced to fuse the preliminary feature data. The multi-layer recursive feature fusion algorithm performs multi-level and recursive feature fusion on the preliminary feature data, adopts a multi-level nested structure, combines nonlinear mapping and recursive optimization strategy, and introduces an adaptive weight adjustment mechanism to achieve gradual optimization and fusion of preliminary feature data, ensuring that the preliminary feature data after feature extraction is optimally mapped in high-dimensional space. The specific implementation process is as follows:

[0042] The preliminary feature data is expressed as: Among them, X t represents time feature data, j s Represents spatial feature data, X d Represents the driving behavior characteristic data, n is the number of characteristic data, and m is the dimension of the characteristic data. Due to the differences in the numerical scale and distribution of characteristic data from different sources, in order to ensure the stability of subsequent calculations, it is necessary to standardize the preliminary characteristic data to obtain the standardized characteristic data X norm The standardized feature data is used as input to enter the recursive feature transformation stage. The technical goal of this stage is to extract more discriminative deep features from the interactive relationship between the preliminary feature data by constructing a multi-layer nonlinear mapping structure. In a recursive form, the output of each layer of feature transformation is used as the input of the next layer to form an iteratively optimized feature representation. The mathematical expression of recursive feature transformation is as follows:

[0043]

[0044] in, It is the feature data after the recursive feature transformation of the i+1th layer; is the feature data after recursive feature transformation of the i-th layer, W i is the weight of the recursive feature transformation at the i-th layer, representing the influence weight of each feature data on the final fusion result, controlling the contribution of different types of preliminary feature data (time feature data, spatial feature data, driving behavior feature data) in the high-dimensional space, and is determined by expert experience; b i is the bias term during the i-th recursion, determined by experimental methods; α is the adaptive balance coefficient, which determines the weight of the influence of the sinusoidal nonlinear component in the recursive feature transformation on the feature data after the final recursive feature transformation. It is used to control the importance of the nonlinear component in the feature fusion process. Its value is between [0,1] to adapt to the needs of different scenarios. It is set according to the specific scenario and is not limited here; γ is the frequency adjustment factor of the sinusoidal nonlinear transformation, which is used to control the degree of nonlinear change of the standardized feature data in the high-dimensional space and improve the discrimination ability of the standardized feature data. It is determined according to expert experience.

[0045] The above recursive operation continuously updates the feature representation until it reaches the convergence condition preset according to the expert experience method, such as reaching the number of layers preset according to the expert experience method, thereby obtaining the feature data X after the final recursive feature transformation. trans .

[0046] After the recursive feature transformation is completed, the feature data X after the final recursive feature transformation is converted to transEntering the orthogonal feature fusion stage, the technical goal of this stage is to eliminate the redundant information between the feature data after the final recursive feature transformation, ensure that the feature data after different final recursive feature transformations remain independent during the fusion process, and thus improve the effectiveness of data fusion. This step uses the mathematical method of orthogonal projection to process the feature data after the final recursive feature transformation and calculate the orthogonal projection matrix:

[0047]

[0048] Where P is the orthogonal projection matrix; T is the transpose of the matrix;

[0049] Then, the orthogonal projection matrix P is used to perform redundancy removal on the feature data after the final recursive feature transformation to obtain the initial non-redundant feature data X orth :

[0050] X orth =X trans -PX trans

[0051] After orthogonal feature fusion, the initial non-redundant feature data X orth , there is still a certain degree of feature redundancy, so further operations are required in the adaptive weight fusion stage. The goal of this stage is to assign appropriate weights to different initial non-redundant feature data dimensions to ensure that each feature data contributes optimally to the final decision. The calculation method of adaptive weight fusion is based on the exponential function of the feature norm, and is specifically implemented as follows:

[0052]

[0053] in, It is The weight of a preliminary non-redundant feature data in the final fusion result reflects the relative importance of the feature data in the decision-making; β is the weight smoothing factor, which is used to control the sensitivity of the exponential function to different feature norms. The larger the value, the more sensitive the weight is to the change of the feature norm. When the value is small, the weight distribution tends to be balanced. It is determined according to the expert experience method; Representative The bi-norm of the initial non-redundant feature data; Represents the bi-norm of the j-th initial non-redundant feature data; is the total number of initial non-redundant feature data;

[0054] After weight calculation, the initial non-redundant feature data is weighted and summed to obtain the fused feature data F fuse d :

[0055]

[0056] Furthermore, the existing principal component analysis method is used to perform dimension reduction processing on the fused feature data to obtain the final fused data;

[0057] S2. Utilize the connected vehicle intelligent predictive driving prediction algorithm based on deep cross-adaptation to build and train a predictive driving model based on preprocessed historical driving data. Utilize the predictive driving model to perform predictive processing on the final fused data to obtain driving prediction results, generate driving decisions, and drive according to the driving decisions.

[0058] A predictive driving model is constructed and trained based on preprocessed historical driving data using a deep cross-adaptive connected vehicle intelligent predictive driving prediction algorithm. The specific process is as follows: First, based on the historical driving data, corresponding environmental perception data is obtained, including high-dimensional unstructured data such as lidar point clouds, camera images, and meteorological data, as well as Internet of Vehicles data, including surrounding vehicle status for V2V and traffic signal information for V2I. After preprocessing the environmental perception data and Internet of Vehicles data, they are merged with the preprocessed historical driving data to obtain comprehensive preprocessed data.

[0059] After data preprocessing, a predictive driving model is constructed using multi-layer recursive cross-perception modeling. This modeling process is divided into a local perception layer, a global trend layer, and a cross-feedback layer. The technical purpose of the local perception layer is to analyze the vehicle's immediate driving behavior through a short time window and use gated temporal convolution to capture driver behavior patterns, such as short-term characteristics such as sudden acceleration and emergency braking. This is achieved by slicing the preprocessed data into time windows and inputting them into a gated temporal convolutional network. Gated units are used for feature extraction to obtain local perception results, and short-term decision weights are adjusted using an adaptive learning rate.

[0060] After local perception, the global trend layer is entered. The technical purpose of the global trend layer is to establish a long-term driving behavior model and identify the driver's long-term driving preferences and environmental influence trends. A hierarchical temporal recurrent neural network is used. Its implementation process includes: first, defining a long-term observation window and associating local perception results with historical driving data. Then, a recurrent neural network is used to build long-term dependencies and a gated unit is used to capture long-term changes in the driver's behavior. Finally, a weighted attention mechanism is used to give higher weights to high-frequency behaviors to enhance the predictive ability of the predictive driving model.

[0061] After the global trend layer, we enter the cross-feedback layer. The technical purpose of the cross-feedback layer is to introduce a feedback correction mechanism between local and global trends, and adjust local and global weights through a multi-level recursive cross-update strategy. The implementation steps of the multi-level recursive cross-update strategy include: first, calculating the dynamic weight difference between the local prediction error and the global trend error, and then establishing a loss function optimization target to reduce the interference of short-term decisions on long-term strategies. Finally, using the dynamic weight adjustment strategy, we optimize the driving prediction results so that driving decisions can take into account both short-term accuracy and long-term stability.

[0062] During the training of a predictive driving model, a hierarchical game strategy inference is employed to optimize the model's driving decision-making strategy. This technology aims to enhance the intelligence and collaboration of driving decisions and prevent global decision-making bias caused by local optima. The specific implementation involves first stratifying the driving scenario into a short-term local game layer and a long-term global game layer. The short-term local game layer is responsible for immediate driving maneuvers, such as lane changes and overtaking, and selects the optimal decision among multiple possible options based on the Nash equilibrium principle of game theory. The long-term global game layer models the long-term driving strategy and utilizes existing deep reinforcement learning methods, combined with driver behavioral preferences, to gradually optimize long-term driving decisions. Specifically, in the initial state, a random strategy is used to collect environmental feedback data. An experience replay buffer is used to store historical state-action-reward sequences. Subsequently, a deep Q-network is used to update the long-term driving strategy, and existing value iteration methods are used to select the action with the highest long-term benefit. The innovation of this hierarchical strategy lies in the interaction of driving decisions at different time scales, avoiding the limitations of traditional single-time-domain methods.

[0063] After the predictive driving model is trained, the final fused data is processed to generate driving prediction results. These include short-term behavior predictions, mid-term driving strategies, and long-term path planning results. Short-term behavior predictions primarily focus on driving actions within the next 1-3 seconds, such as acceleration, braking, and lane changes. Mid-term driving strategies focus on driving plans within the next 3-10 seconds, combining predictions from the global trend layer with environmental variables such as road conditions and traffic flow to determine the optimal lane and speed control scheme. Long-term path planning results are based on long-term predictions of more than 10 seconds, combined with historical driving data and existing traffic flow models to generate recommended routes with the best energy consumption and shortest travel time.

[0064] To ensure the reliability and sustainable optimization of the predictive driving model, a real-time feedback loop is introduced to capture actual vehicle driving data, compare it with the predicted driving results, calculate deviations, and adjust model parameters. This implementation involves first collecting acceleration, speed, path deviation, and other data in real time during actual driving, then smoothing the data using the existing weighted sliding average method. The error between the processed data and the model predictions is then calculated, and the model parameters are fine-tuned using the existing Bayesian online update method. The goal of this real-time feedback loop is to enable the predictive driving model to adapt to the dynamically changing traffic environment and continuously optimize its decision-making capabilities. By introducing an incremental learning mechanism, new driving data can be continuously absorbed without losing historical experience, and the robustness and adaptability of the predictive driving model can be maintained by adjusting the neural network structure.

[0065] In summary, the decision-making method for intelligent predictive driving of connected vehicles based on machine learning has been completed.

[0066] The order in which the embodiments of the invention are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0067] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0068] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A decision-making method for intelligent predictive driving of connected vehicles based on machine learning, characterized in that: The following steps are involved: S1. Acquire and preprocess multi-source data, real-time data from the Internet of Vehicles, and historical driving data of connected vehicles to obtain preprocessed multi-source data, real-time data from the Internet of Vehicles, and historical driving data of connected vehicles; Feature extraction is performed on the pre-processed multi-source data of connected vehicles and real-time data of the Internet of Vehicles to obtain preliminary feature data. A multi-layer recursive feature fusion algorithm is introduced to fuse the preliminary feature data to obtain the final fused data. S2. Introducing a connected vehicle intelligent predictive driving prediction algorithm based on deep cross-adaptability. During implementation of this algorithm, corresponding environmental perception data and vehicle network data are obtained based on historical driving data. These environmental perception data and vehicle network data are preprocessed and then merged with the preprocessed historical driving data to obtain comprehensive preprocessed data. After data preprocessing, a predictive driving model is constructed using multi-layer recursive cross-perception modeling. This modeling process is divided into a local perception layer, a global trend layer, and a cross-feedback layer. The predictive driving model is then trained. During the training process, a hierarchical game strategy is used to infer and optimize the driving decision-making strategy of the predictive driving model. After completing the predictive driving model training, the final fused data is processed for prediction to obtain the driving prediction results, generate driving decisions, and drive according to the driving decisions.

2. The decision-making method for intelligent predictive driving of connected vehicles based on machine learning according to claim 1, characterized in that: Said S1 specifically includes: The multi-layer recursive feature fusion algorithm performs multi-level, recursive feature fusion on preliminary feature data, adopts a multi-level nested structure, and combines nonlinear mapping with recursive optimization strategy, and introduces an adaptive weight adjustment mechanism to fuse the preliminary feature data.

3. The decision-making method for intelligent predictive driving of connected vehicles based on machine learning according to claim 2, characterized in that: Said S1 specifically includes: In the implementation process of the multi-layer recursive feature fusion algorithm, the preliminary feature data is standardized to obtain the standardized feature data, and the standardized feature data is used as input to enter the recursive feature transformation stage to obtain the final recursive feature transformed feature data. The specific implementation formula is: ， in, It is Feature data after layer-wise recursive feature transformation; For the Feature data after layer-wise recursive feature transformation, is the recursive feature transformation in the Weights during layer recursion; For the Bias term during layer recursion; is the adaptive balance coefficient; is the frequency adjustment factor of the sinusoidal nonlinear transformation; It is the standardized feature data.

4. The decision-making method for intelligent predictive driving of connected vehicles based on machine learning according to claim 3, characterized in that: Said S1 specifically includes: In the implementation process of the multi-layer recursive feature fusion algorithm, after completing the recursive feature transformation, the feature data after the final recursive feature transformation enters the orthogonal feature fusion stage, and the feature data after the final recursive feature transformation is processed to obtain an orthogonal projection matrix; the orthogonal projection matrix is ​​used to perform redundancy removal operations on the feature data after the final recursive feature transformation to obtain preliminary non-redundant feature data.

5. The decision-making method for intelligent predictive driving of connected vehicles based on machine learning according to claim 4, characterized in that: Said S1 specifically includes: In the implementation process of the multi-layer recursive feature fusion algorithm, after the orthogonal feature fusion, the initial non-redundant feature data enters the adaptive weight fusion stage for processing. The specific implementation formula is: ， in, It is The weight of the initial non-redundant feature data in the final fusion result; is the weight smoothing factor; Representative The bi-norm of the initial non-redundant feature data; Representative The bi-norm of the initial non-redundant feature data; is the total number of initial non-redundant feature data.

6. The decision-making method for intelligent predictive driving of connected vehicles based on machine learning according to claim 5, characterized in that: Said S1 specifically includes: In the implementation process of the multi-layer recursive feature fusion algorithm, after the weight calculation, the initial non-redundant feature data is weighted and summed to obtain the fused feature data, and the fused feature data is subjected to feature data dimensionality reduction processing to obtain the final fused data.

Images