New energy automobile vehicle-mounted data processing method and system based on intelligent network connection

Abstract

The invention relates to the technical field of new energy vehicle-mounted data processing, and discloses a new energy vehicle vehicle-mounted data processing method and system based on intelligent networking, and the method comprises the steps: collecting and preprocessing data through a multi-source sensor, carrying out the slicing fusion of the data based on time-space correlation, and extracting features through a lightweight convolutional neural network. Constructing an inter-vehicle communication topology model to realize collaborative feature learning, adopting a federal learning framework to train the model and protect privacy, applying a reinforcement learning algorithm to make a decision, establishing a multi-objective optimization model, deploying a dynamic priority scheduling module to manage resources, and optimizing the model based on an incremental learning algorithm. The system covers a plurality of modules such as data acquisition and preprocessing, data dynamic slicing and fusion and the like. According to the method, the data processing efficiency and accuracy are improved, intelligent decision and multi-objective optimization are realized, data privacy is guaranteed, computing resources are reasonably distributed, and the overall performance of the new energy automobile in an intelligent network connection environment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy vehicle data processing, and specifically to a method and system for processing new energy vehicle on-vehicle data based on intelligent networking. Background Art

[0002] With the development of the automotive industry towards intelligence, networking, and electrification, the proportion of new energy vehicles in the market has gradually increased. Intelligent networking technology has brought richer functions and better user experiences to new energy vehicles, but at the same time, it has also caused a sharp increase in the scale and complexity of on-vehicle data.

[0003] Traditional automotive data processing methods are no longer sufficient to meet the requirements of new energy vehicles in an intelligent networking environment. In terms of data collection, new energy vehicles need to obtain data on various aspects such as vehicle operating status, battery status, and surrounding environment. A single type of sensor cannot meet the requirement of comprehensive perception, and the data collected by multiple sensors often suffer from problems such as noise interference, time asynchronization, inconsistent units, and data loss, which bring great difficulties to subsequent data processing and analysis. For example, in complex road conditions, the data collected by sensors is easily affected by external environmental factors, resulting in inaccurate data, which in turn affects the vehicle's judgment of the surrounding environment.

[0004] In the data processing and analysis stage, due to the spatio-temporal characteristics of the data generated during vehicle driving, traditional data processing methods cannot effectively mine the spatio-temporal correlation between data, making it difficult to reasonably slice and fuse the data, resulting in low information utilization rate. In addition, energy management and driving decision-making for new energy vehicles are crucial and require comprehensive consideration of multiple factors such as energy consumption, path planning, and battery life. However, existing decision-making models often cannot make full use of on-vehicle data and cannot achieve multi-objective optimization, resulting in high energy consumption, unreasonable driving paths during vehicle operation, shortening of battery life, and affecting the overall performance of the vehicle and the user's usage cost.

[0005] In a vehicle networking environment, vehicles need to exchange information to achieve cooperative driving and intelligent traffic management. However, the current construction of the vehicle-to-vehicle communication topology model is not flexible enough and cannot be dynamically adjusted according to factors such as real-time vehicle position, communication delay, and signal strength, affecting the effect of cooperative feature learning and resulting in insufficient cooperative ability of vehicles in complex traffic scenarios. At the same time, there is a risk of data privacy leakage during the distributed model training process, which makes each vehicle hesitant to share data for model training, restricting the development and application of vehicle networking technology.

[0006] In the aspect of in-vehicle computing resource management, with the continuous increase of in-vehicle applications, the demand for computing resources is also growing. The existing resource scheduling methods lack dynamism and flexibility, and cannot reasonably allocate computing resources according to the real-time requirements of data processing tasks, resulting in some critical tasks not being processed in time, affecting the safety and stability of the vehicle. In addition, the in-vehicle model cannot be adaptively adjusted in a timely manner according to new data and environmental changes, causing the accuracy and adaptability of the model to gradually decline and unable to meet the changing driving needs of the vehicle. Summary of the Invention

[0007] The purpose of the present invention is to provide an in-vehicle data processing method and system for new energy vehicles based on intelligent networking to solve the problems raised in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solutions: An in-vehicle data processing method for new energy vehicles based on intelligent networking, the method includes:

[0009] Step S1: Collect vehicle operation data, environmental perception data, and battery status data through in-vehicle multi-source sensors, and perform noise filtering and data standardization processing;

[0010] Step S2: Dynamically slice the data processed in Step S1 based on spatio-temporal correlation, generate multi-dimensional data blocks, and fuse each data block using an adaptive weight allocation algorithm;

[0011] Step S3: Construct an in-vehicle edge computing node, and use a lightweight convolutional neural network to extract high-dimensional features from the fused data to generate a set of feature vectors;

[0012] Step S4: Establish an inter-vehicle communication topology model based on a dynamic graph neural network, and update the feature vectors of each node through a neighborhood propagation mechanism to achieve collaborative feature learning in the vehicle networking environment;

[0013] Step S5: Use a federated learning framework to perform distributed model training on the collaborative features generated in Step S4, and introduce a differential privacy mechanism to perturb the gradient parameters;

[0014] Step S6: Use a reinforcement learning algorithm to construct a real-time decision-making model, take the high-dimensional features extracted in Step S3 and the model parameters trained in Step S5 as inputs, and output vehicle control instructions and energy scheduling strategies;

[0015] Step S7: According to the output results of Step S6, establish a multi-objective optimization model, take the minimum energy consumption, the optimal path, and the maximum battery life as the objective functions, and use a genetic particle swarm hybrid algorithm to solve the optimal solution set;

[0016] Step S8: Deploy a dynamic priority scheduling module in the in-vehicle computing unit, allocate computing resources according to the real-time requirements of data processing, and ensure that high-priority tasks are executed first;

[0017] Step S9: Update the lightweight convolutional neural network parameters in Step S3 based on the incremental learning algorithm, and combine the federated model parameters in Step S5 to achieve online adaptive adjustment of the in-vehicle model.

[0018] Preferably, in Step S1, the multi-source sensors include lidar, millimeter-wave radar, cameras, inertial measurement units, and battery management system sensors; the data normalization processing includes timestamp alignment, unit unification, and missing value imputation.

[0019] Preferably, in Step S2, the adaptive weight allocation algorithm dynamically adjusts the fusion weights by calculating the information entropy and correlation coefficient of each data block. The specific formula is:

[0020]

[0021] where W i is the weight of the i-th data block, N is the total number of data blocks generated after dynamic slicing, E i is the information entropy, C i is the average correlation coefficient with other data blocks, and α and β are adjustment factors.

[0022] Preferably, in Step S4, the inter-vehicle communication topology model dynamically constructs a graph structure with vehicle position, communication delay, and signal strength as edge weights; the neighborhood propagation mechanism updates the current node features by aggregating the feature vectors and edge weights of adjacent nodes.

[0023] Preferably, in Step S5, the differential privacy mechanism ensures data privacy during model training by adding noise that satisfies the Laplace distribution to the gradient parameters. The noise addition formula is:

[0024]

[0025] where is the perturbed gradient, g is the original gradient, Δf is the gradient sensitivity, ∈ is the privacy budget, and Lap(·) is the Laplace distribution generator.

[0026] Preferably, in Step S7, the multi-objective optimization model adopts the Pareto front screening strategy, combines the global search ability of the genetic algorithm and the local optimization ability of the particle swarm algorithm to generate a non-dominated solution set.

[0027] Preferably, in step S8, the dynamic priority scheduling module dynamically divides the task priorities according to the task latency tolerance, data freshness, and computational complexity, and adopts a preemptive scheduling strategy to allocate in-vehicle computing resources.

[0028] Preferably, in step S9, the incremental learning algorithm adjusts the network parameter update step size by calculating the feature distribution difference between the new data and the historical data. The specific formula is as follows:

[0029]

[0030] where θ t is the neural network parameter at the current moment, θ t+1 is the neural network parameter at the next moment after update, is the gradient of the loss function with respect to the parameter, indicating the parameter optimization direction, D is the distribution difference metric value, k is the decay coefficient, η is the initial learning rate, and exp(·) is the Sigmoid function, which maps the distribution difference to the interval (0, 1).

[0031] Preferably, the present invention further includes an in-vehicle data processing system for new energy vehicles based on intelligent network connection. The system includes:

[0032] Data acquisition and preprocessing module: Collect vehicle operation data, environment perception data, and battery state data through in-vehicle multi-source sensors, and perform noise filtering and data standardization processing.

[0033] Data dynamic slicing and fusion module: Dynamically slice the preprocessed data based on spatio-temporal correlation to generate multi-dimensional data blocks, and fuse each data block using an adaptive weight allocation algorithm.

[0034] Feature extraction module: Build an in-vehicle edge computing node, and use a lightweight convolutional neural network to extract high-dimensional features from the fused data to generate a set of feature vectors.

[0035] Collaborative feature learning module: Establish a vehicle-to-vehicle communication topology model based on a dynamic graph neural network, and update the feature vectors of each node through a neighborhood propagation mechanism to achieve collaborative feature learning in the vehicle network environment.

[0036] Distributed model training module: Use a federated learning framework to perform distributed model training on collaborative features, and introduce a differential privacy mechanism to perturb the gradient parameters.

[0037] Real-time decision-making module: Use a reinforcement learning algorithm to build a real-time decision-making model, and take the extracted high-dimensional features and the trained model parameters as inputs to output vehicle control instructions and energy scheduling strategies.

[0038] Multi-objective optimization module: According to the output results of the real-time decision-making module, establish a multi-objective optimization model, with the minimum energy consumption, the optimal path, and the maximum battery life as the objective functions, and use the genetic particle swarm hybrid algorithm to solve the optimal solution set;

[0039] Resource scheduling module: Deploy a dynamic priority scheduling module in the vehicle-mounted computing unit, allocate computing resources according to the real-time requirements of data processing, and ensure that high-priority tasks are executed first;

[0040] Model adaptive adjustment module: Update the parameters of the lightweight convolutional neural network based on the incremental learning algorithm, and combine the federal model parameters to achieve online adaptive adjustment of the vehicle-mounted model.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] Rich data is collected by multi-source sensors, and noise filtering and standardization processing are performed, effectively improving the accuracy and availability of the data. For example, Kalman filtering is used to process radar data noise, and time synchronization technology is used to align timestamps, ensuring the reliability of the data in subsequent processing and providing a solid data foundation for the precise decision-making of the vehicle.

[0043] Dynamically slice data based on spatio-temporal correlation, and fuse data blocks in combination with the adaptive weight allocation algorithm, which can more accurately reflect the actual state of the vehicle and the environment. At the same time, using the lightweight convolutional neural network to extract high-dimensional features not only reduces the computational complexity but also improves the efficiency and quality of feature extraction, which helps the vehicle to quickly perceive and understand complex environments.

[0044] The vehicle-to-vehicle communication topology model constructed by the dynamic graph neural network is dynamically adjusted according to the vehicle position, communication delay, and signal strength, and combined with the neighborhood propagation mechanism to achieve efficient collaborative feature learning. The differential privacy mechanism is introduced in the distributed model training. By adding noise to the gradient parameters, on the premise of ensuring data privacy and security, the data of each vehicle is fully utilized for model training, improving the generalization ability of the model.

[0045] The real-time decision-making model constructed by the reinforcement learning algorithm comprehensively considers various features and model parameters such as vehicle operation, environment, and battery state, and outputs reasonable vehicle control instructions and energy scheduling strategies. The multi-objective optimization model aims at the minimum energy consumption, the optimal path, and the maximum battery life, and is solved by the genetic particle swarm hybrid algorithm, effectively balancing the various requirements of the vehicle in different driving scenarios, reducing energy consumption, extending battery life, and improving driving efficiency.

[0046] The dynamic priority scheduling module of the in-vehicle computing unit dynamically divides task priorities according to task latency tolerance, data freshness, and computational complexity, and adopts a preemptive scheduling strategy to ensure the priority execution of critical tasks, improving the utilization efficiency of in-vehicle computing resources. Update the neural network parameters based on the incremental learning algorithm, and combine the federated model parameters to achieve the online adaptive adjustment of the in-vehicle model, enabling the model to continuously optimize with changes in the environment and data and maintain good performance.

[0047] The present invention innovates in multiple aspects from data collection, processing, analysis to decision-making, resource management, and model optimization, comprehensively improving the performance, safety, energy efficiency, and user experience of new energy vehicles in the intelligent networked environment, which is of great significance for promoting the development of the new energy vehicle industry. Brief Description of the Drawings

[0048] Figure 1 It is the working principle diagram of the in-vehicle data processing method for new energy vehicles described in the present invention;

[0049] Figure 2 It is the working flowchart of the adaptive weight allocation algorithm;

[0050] Figure 3 It is the working flowchart of the vehicle-to-vehicle communication topology model and the neighborhood propagation mechanism.

[0051] Figure 4 It is the working flowchart of the solution of the multi-objective optimization model. Detailed Embodiments

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0053] Please refer to Figures 1-4 , the present invention provides a technical solution: an in-vehicle data processing method for new energy vehicles based on intelligent networking, the method includes:

[0054] Step S1: Use in-vehicle multi-source sensors to collect vehicle operation data, environmental perception data, and battery status data, and then perform noise filtering and data standardization processing.

[0055] Step S2: Dynamically slice the data processed in Step S1 according to spatio-temporal correlation to generate multi-dimensional data blocks, and fuse each data block by means of an adaptive weight allocation algorithm.

[0056] Step S3: Build an in-vehicle edge computing node, and use a lightweight convolutional neural network to extract high-dimensional features from the fusion data to form a set of feature vectors.

[0057] Step S4: Build an inter-vehicle communication topology model based on a dynamic graph neural network, and update the feature vectors of each node through a neighborhood propagation mechanism to achieve collaborative feature learning in the vehicle networking environment.

[0058] Step S5: Use a federated learning framework to perform distributed model training on the collaborative features generated in Step S4, and at the same time introduce a differential privacy mechanism to perturb the gradient parameters.

[0059] Step S6: Use a reinforcement learning algorithm to build a real-time decision-making model, take the high-dimensional features extracted in Step S3 and the model parameters trained in Step S5 as inputs, and output vehicle control instructions and energy scheduling strategies.

[0060] Step S7: According to the output results of Step S6, establish a multi-objective optimization model, use the minimum energy consumption, the optimal path, and the maximum battery life as the objective functions, and use a genetic particle swarm hybrid algorithm to solve the optimal solution set.

[0061] Step S8: Deploy a dynamic priority scheduling module in the in-vehicle computing unit, and allocate computing resources according to the real-time requirements of data processing to ensure that high-priority tasks are executed first.

[0062] Step S9: Update the parameters of the lightweight convolutional neural network in Step S3 based on an incremental learning algorithm, and combine the federated model parameters in Step S5 to achieve online adaptive adjustment of the in-vehicle model.

[0063] The present invention will be further described below in conjunction with Embodiments 1 to 5:

[0064] Embodiment 1:

[0065] In actual application scenarios, the operating environment of new energy vehicles is complex and variable, and accurate data collection is the key to subsequent processing. The multi-source sensors in step S1 cover lidar, millimeter-wave radar, cameras, inertial measurement units, and battery management system sensors. Lidar obtains three-dimensional point cloud data of the surrounding environment by emitting laser beams and receiving reflected light, and can accurately identify the position, shape, and distance of obstacles, providing key information for vehicle obstacle avoidance and path planning. Millimeter-wave radar uses electromagnetic waves in the millimeter-wave frequency band, has the advantages of high measurement accuracy and being unaffected by light and weather, can monitor the speed, distance, and angle changes of the vehicle in real time, and is used for functions such as adaptive cruise control and collision warning. As a visual sensor, the camera can collect rich image information, such as road signs, traffic signal states, and the behaviors of other vehicles and pedestrians, and provides intuitive environmental perception data for the vehicle through image recognition algorithms. The inertial measurement unit is used to measure the attitude information of the vehicle, such as acceleration and angular velocity, which helps to determine the motion state and driving trajectory of the vehicle and plays an important role in vehicle navigation and stability control. The battery management system sensor is responsible for monitoring the state parameters of the battery, such as voltage, current, and temperature, ensuring the safe operation of the battery, and providing necessary data support for energy management.

[0066] During the data collection process, sensors will inevitably be interfered by various noises. To improve data quality, a variety of noise filtering techniques are adopted. For example, for the data of lidar and millimeter-wave radar, the Kalman filtering algorithm is used for noise processing. Kalman filtering is an optimal recursive filtering algorithm based on the state space model of a linear system. It can effectively estimate the true signal by predicting and updating the system state, thereby removing noise interference. For the image data collected by the camera, the median filtering algorithm is adopted. Median filtering is a non-linear filtering method that replaces the value of each pixel point in the image with the median of its neighboring pixel values, and can effectively remove impulse interference such as salt-and-pepper noise while maintaining the edge information of the image.

[0067] Data standardization is an important step to ensure data consistency and availability. Among them, timestamp alignment is used to unify the time reference of different sensor data. Due to the different sampling frequencies and data transmission delays of each sensor, before data fusion, the timestamps of all data need to be adjusted to the same time scale. This can be achieved through precise clock synchronization technology. For example, the time signal of the Global Positioning System (GPS) can be used as a reference to calibrate the timestamps of each sensor. Unit unification is to perform unit conversion on the data collected by different sensors so that they have the same dimension. For example, unify the distance unit measured by the millimeter-wave radar to meters, and unify the voltage unit measured by the battery management system sensor to volts, etc. Missing value imputation is to process the missing data that may occur during the data collection process. For a small number of discrete missing values, linear interpolation can be used to linearly calculate based on the values of adjacent data points to fill in the missing values. For continuous missing values, machine learning-based methods can be used, such as the missing value filling method based on the K-nearest neighbor algorithm. The K-nearest neighbor algorithm predicts and fills in the missing value by finding the K nearest neighbor data points to the data point where the missing value is located and using the feature values of these neighbor data points.

[0068] Example 2:

[0069] In step S2, dynamically slicing the preprocessed data based on spatio-temporal correlation is to better capture the characteristics of the data in the time and space dimensions. Spatio-temporal correlation considers the mutual relationship between data at different time points and different spatial positions. For example, during the driving of a vehicle, there is a certain correlation between the position information at different times and the changes in the surrounding environment. By analyzing these correlations, the continuous data stream is divided according to time intervals and spatial regions to generate multi-dimensional data blocks. These data blocks not only contain the operation state information of the vehicle within a specific time and space range, but also integrate the corresponding environmental perception data and battery state data, providing rich data units for subsequent data fusion.

[0070] The adaptive weight allocation algorithm plays a key role in data fusion. This algorithm dynamically adjusts the fusion weights by calculating the information entropy and correlation coefficient of each data block. The specific formula is:

[0071]

[0072] Among them, W i is the weight of the i-th data block, N is the total number of data blocks generated after dynamic slicing, E i is the information entropy, which is used to measure the uncertainty of the information in the data block. The larger the information entropy, the richer the information in the data block, and its importance in the fusion process may be higher. C iis the average correlation coefficient with other data blocks, reflecting the degree of association between this data block and other data blocks. The higher the correlation coefficient, the better the coordination between this data block and other data blocks, and a higher weight should be given during fusion. α and β are adjustment factors, which are adjusted according to the actual application scenario and data characteristics, and are used to balance the relative importance of information entropy and correlation coefficient in weight calculation.

[0073] In the actual calculation process, taking the vehicle operation data, environmental perception data, and battery status data collected at a certain moment as an example. First, calculate the information entropy of each data block. Assume that data block A contains operation parameters such as the speed and acceleration of the vehicle. By analyzing the probability distribution of these parameters and using the information entropy calculation formula E = -∑ k p k log2 p k (where p k is the probability of the k-th state appearing in the data block), the information entropy E A of data block A is obtained. Then calculate the average correlation coefficient C A of data block A with other data blocks, which is obtained by calculating the Pearson correlation coefficient of the corresponding data elements in data block A and other data blocks and taking the average. According to the above formula, the weight W A of data block A is calculated. Similar calculations are performed on all data blocks, and finally, the data blocks are fused according to the calculated weights. Through this adaptive weight allocation algorithm, the characteristics of each data block can be fully considered, enabling the fused data to more accurately reflect the actual situation of the vehicle and the environment, and providing more reliable data support for subsequent data analysis and decision-making.

[0074] Example 3:

[0075] This example deeply illustrates the construction of the vehicle-to-vehicle communication topology model and the neighborhood propagation mechanism, as well as the application of the differential privacy mechanism in distributed model training, to improve the collaborative feature learning effect and ensure data privacy and security.

[0076] In step S4, establishing a vehicle-to-vehicle communication topology model based on a dynamic graph neural network is crucial for realizing collaborative feature learning in the vehicle networking environment. This model dynamically constructs a graph structure with vehicle position, communication delay, and signal strength as edge weights. In the actual vehicle networking scenario, the vehicle position information can be obtained through the Global Positioning System (GPS) or other high-precision positioning technologies. Communication delay is an important indicator to measure the data transmission time between vehicles, which is affected by factors such as network congestion and signal interference. Signal strength reflects the quality of the communication signal between vehicles and can be obtained through the Received Signal Strength Indicator (RSSI) of in-vehicle communication devices.

[0077] Taking multiple new energy vehicles in a certain area as an example, each vehicle serves as a node in the graph structure. According to the real-time position information of the vehicles, the distances between the vehicles are calculated. Edge connections are established between vehicles with relatively short distances, and the weights of the edges are set according to communication latency and signal strength. For example, if the communication latency between vehicle A and vehicle B is low and the signal strength is high, the weight of the edge between them is relatively large, indicating that the communication quality between these two vehicles is good, and during the collaborative feature learning process, the information interaction between them is more important.

[0078] The neighborhood propagation mechanism updates the current node features by aggregating the feature vectors and edge weights of adjacent nodes. Suppose the current feature vector of vehicle A is F A , and its adjacent nodes are vehicle B and vehicle C, with corresponding feature vectors F B and F C respectively, and the edge weights are W AB and W AC . Then the updated feature vector of vehicle A is calculated by the following formula: (where neighbors(A) represents the set of adjacent nodes of vehicle A). Through this neighborhood propagation mechanism, vehicles can continuously absorb the feature information of adjacent vehicles, realize collaborative feature learning in the vehicle networking environment, and improve the perception and understanding ability of complex environments.

[0079] In step S5, a federated learning framework is adopted for distributed model training, and a differential privacy mechanism is introduced to perturb the gradient parameters to protect data privacy. The differential privacy mechanism ensures data privacy during model training by adding noise that satisfies the Laplace distribution to the gradient parameters. The noise addition formula is:

[0080]

[0081] where, is the perturbed gradient, g is the original gradient, Δf is the gradient sensitivity, is the privacy budget, and Lap(·) is the Laplace distribution generator. In practical applications, first, the gradient sensitivity Δf is calculated, which reflects the sensitivity of the model to changes in individual data points. By making small perturbations to all data points in the training dataset and calculating the maximum change in the model gradient, the gradient sensitivity is obtained. The privacy budget is then a parameter set according to actual requirements, used to control the intensity of privacy protection. The smaller the value, the greater the added noise and the higher the degree of privacy protection. However, it may have a certain impact on the training accuracy of the model. During the model training process, after each vehicle calculates the gradient locally, noise is added to the gradient parameters according to the above formula, and then the perturbed gradient is uploaded to the server for model aggregation. In this way, both distributed model training can be achieved, making full use of the data resources of each vehicle, and the privacy data of the vehicle can be effectively protected to ensure the security of data during sharing.

[0082] Example 4:

[0083] In step S6, a real-time decision-making model is constructed using the reinforcement learning algorithm, with the high-dimensional features extracted in step S3 and the model parameters trained in step S5 as inputs, and vehicle control instructions and energy scheduling strategies are output. The reinforcement learning algorithm continuously learns the optimal behavior strategy through the interaction between the agent and the environment. In the application scenario of new energy vehicles, the agent is the vehicle, and the environment includes the running state of the vehicle, surrounding environment information, and battery state, etc.

[0084] Taking the example of a vehicle encountering traffic congestion during driving, the real-time decision-making model selects the optimal behavior action according to high-dimensional features such as the current vehicle speed, acceleration, battery power, and the distribution of surrounding vehicles, combined with the trained model parameters, through the policy selection mechanism of the reinforcement learning algorithm. For example, the decision-making model may choose to reduce the vehicle speed to reduce energy consumption, and at the same time adjust the charging and discharging strategy of the battery, preferentially using the remaining battery power for driving to avoid energy loss caused by frequent charging in congested sections. During the reinforcement learning process, a reasonable reward function is set to guide the model to learn the optimal strategy. If the vehicle can reduce energy consumption, reduce driving time, or extend battery life while maintaining safe driving, a positive reward is given; otherwise, a negative reward is given. Through continuous trial and error and learning, the real-time decision-making model can gradually optimize the decision-making strategy and improve the vehicle's decision-making ability in various complex environments.

[0085] In step S7, according to the output result of step S6, a multi-objective optimization model is established, with the minimum energy consumption, the optimal path, and the maximum battery life as the objective functions, and the genetic particle swarm hybrid algorithm is used to solve the optimal solution set. The multi-objective optimization model adopts the Pareto front screening strategy, combining the global search ability of the genetic algorithm and the local optimization ability of the particle swarm algorithm to generate a non-dominated solution set. The Pareto front screening strategy is to find a balance between multiple objectives, so that it is impossible to further optimize a certain objective without sacrificing other objectives.

[0086] Suppose there are multiple candidate paths during the vehicle's driving process, and each path corresponds to different energy consumption, driving time, and battery loss. The genetic algorithm performs selection, crossover, and mutation operations on the candidate paths by simulating the biological evolution process, conducts global search in a large solution space, and seeks potential optimal solutions. The particle swarm algorithm is based on the idea of swarm intelligence. Each particle represents a solution and conducts fine search in a local area by tracking its own historical optimal solution and the group's historical optimal solution. In practical applications, first, the candidate paths are encoded into a form that can be processed by the genetic algorithm and the particle swarm algorithm, such as binary encoding or real-number encoding. Then, the fitness value of each solution is calculated according to objective functions such as energy consumption, path length, and battery life. Through the alternating iteration of the genetic algorithm and the particle swarm algorithm, the quality of the solution is continuously optimized, and finally, a set of non-dominated solution sets is obtained on the Pareto front. The vehicle can select the most suitable solution from these non-dominated solution sets according to actual needs, achieve the comprehensive optimization of multiple objectives such as energy consumption, path, and battery life, and improve the overall performance and operation efficiency of the vehicle.

[0087] Embodiment 5:

[0088] In step S8, a dynamic priority scheduling module is deployed in the in-vehicle computing unit to allocate computing resources according to the real-time requirements of data processing, ensuring that high-priority tasks are executed first. The dynamic priority scheduling module dynamically divides task priorities according to task delay tolerance, data freshness, and computational complexity, and adopts a preemptive scheduling strategy to allocate in-vehicle computing resources. Task delay tolerance reflects the sensitivity of a task to processing time. For example, the task of processing the vehicle's emergency braking signal has extremely low delay tolerance and must be completed within a short time; otherwise, it may lead to safety accidents. Therefore, such tasks have a higher priority. Data freshness refers to the timeliness of data. For data with high real-time requirements, such as current road conditions and the vehicle's real-time speed, the higher the data freshness, the higher the corresponding task priority. Computational complexity measures the amount of computing resources required for task processing. For tasks with low computational complexity, their priorities can be appropriately reduced in the case of limited resources, and tasks with high computational complexity and urgency should be processed first.

[0089] Taking the vehicle during driving as an example, when the vehicle detects an obstacle suddenly appearing ahead, the processing tasks of environmental perception data and the emergency braking decision-making tasks have a relatively high priority. The dynamic priority scheduling module sets their priorities to the highest according to the latency tolerance, data freshness, and computational complexity of these tasks. At this time, if the in-vehicle computing unit is processing other low-priority tasks, such as the data update task of the vehicle entertainment system, the dynamic priority scheduling module will adopt a preemptive scheduling strategy, suspend the execution of the low-priority task, and allocate computing resources to the high-priority emergency task to ensure that the vehicle can respond in a timely manner and avoid collision accidents. Through this dynamic priority scheduling mechanism, the utilization efficiency of in-vehicle computing resources can be effectively improved, the timely processing of key vehicle tasks can be guaranteed, and the safety and reliability of the vehicle can be improved.

[0090] In step S9, based on the incremental learning algorithm, update the lightweight convolutional neural network parameters in step S3, and combine the federated model parameters in step S5 to realize the online adaptive adjustment of the in-vehicle model. The incremental learning algorithm adjusts the network parameter update step size by calculating the feature distribution difference between the new data and the historical data. The specific formula is:

[0091]

[0092] where, θ t is the neural network parameter at the current moment, θ t+1 is the neural network parameter at the next moment after update, is the gradient of the loss function with respect to the parameter, indicating the parameter optimization direction, D is the distribution difference metric value, k is the decay coefficient, η is the initial learning rate, and exp(·) is the Sigmoid function, which maps the distribution difference to the interval (0,1).

[0093] In practical applications, as the vehicle is driving, new data is continuously generated. For example, when the vehicle enters a new road environment, different road condition information and environmental data are collected. The incremental learning algorithm first calculates the feature distribution difference D between the new data and the historical data. Methods such as KL divergence can be used to measure the distribution difference. If the difference between the new data and the historical data is large, it indicates that the environment where the vehicle is located has changed greatly. At this time, the difference is mapped to the interval (0,1) through the Sigmoid function, so that the update step size Increase and accelerate the update speed of neural network parameters, enabling the model to quickly adapt to the new environment. Conversely, if the difference is small, appropriately reduce the update step size to avoid over-adjustment of the model. By combining the federal parameters of the model, the in-vehicle model can make full use of the data information and training results of other vehicles to further optimize its own performance. Through this incremental learning algorithm, the in-vehicle model can adaptively adjust online, continuously improve its adaptability to vehicle operation and environmental changes, and maintain good prediction and decision-making performance.

[0094] The present invention also includes an in-vehicle data processing system for new energy vehicles based on intelligent networking, and the system includes:

[0095] Data acquisition and preprocessing module: Collect vehicle operation data, environmental perception data, and battery state data through in-vehicle multi-source sensors, and perform noise filtering and data standardization processing;

[0096] Data dynamic slicing and fusion module: Dynamically slice the preprocessed data based on spatio-temporal correlation, generate multi-dimensional data blocks, and fuse each data block using an adaptive weight allocation algorithm;

[0097] Feature extraction module: Build an in-vehicle edge computing node, and use a lightweight convolutional neural network to extract high-dimensional features from the fused data to generate a set of feature vectors;

[0098] Collaborative feature learning module: Establish a vehicle-to-vehicle communication topology model based on a dynamic graph neural network, and update the feature vectors of each node through a neighborhood propagation mechanism to achieve collaborative feature learning in the vehicle networking environment;

[0099] Distributed model training module: Use a federated learning framework to perform distributed model training on collaborative features, and introduce a differential privacy mechanism to perturb the gradient parameters;

[0100] Real-time decision-making module: Use a reinforcement learning algorithm to build a real-time decision-making model, take the extracted high-dimensional features and the trained model parameters as inputs, and output vehicle control instructions and energy scheduling strategies;

[0101] Multi-objective optimization module: According to the output results of the real-time decision-making module, establish a multi-objective optimization model, take the minimum energy consumption, the optimal path, and the maximum battery life as objective functions, and use a genetic particle swarm hybrid algorithm to solve the optimal solution set;

[0102] Resource scheduling module: Deploy a dynamic priority scheduling module in the in-vehicle computing unit, allocate computing resources according to the real-time requirements of data processing, and ensure that high-priority tasks are executed first;

[0103] Model adaptive adjustment module: Update the parameters of the lightweight convolutional neural network based on the incremental learning algorithm, and combine the federal model parameters to achieve online adaptive adjustment of the in-vehicle model.

[0104] The implementation manner of this system refers to the above-mentioned embodiments and will not be elaborated in the description.

[0105] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0106] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for processing vehicle-mounted data of new energy vehicles based on intelligent networking, characterized in that: The following steps are involved: Step S1: Collect vehicle operation data, environmental perception data and battery status data through on-board multi-source sensors, and perform noise filtering and data standardization processing; Step S2: dynamically slice the data processed in step S1 based on spatiotemporal correlation to generate multi-dimensional data blocks, and fuse the data blocks using an adaptive weight allocation algorithm; Step S3: construct an on-vehicle edge computing node, use a lightweight convolutional neural network to extract high-dimensional features in the fused data, and generate a feature vector set; Step S4: Establishing a vehicle-to-vehicle communication topology model based on a dynamic graph neural network, updating the feature vector of each node through a neighborhood propagation mechanism, and realizing collaborative feature learning in a connected vehicle environment; Step S5: Use the federated learning framework to perform distributed model training on the collaborative features generated in step S4, and introduce a differential privacy mechanism to perturb the gradient parameters; Step S6: construct a real-time decision model using a reinforcement learning algorithm, taking the high-dimensional features extracted in step S3 and the model parameters trained in step S5 as input, and outputting vehicle control instructions and energy scheduling strategies; Step S7: Based on the output result of step S6, a multi-objective optimization model is established, with minimum energy consumption, optimal path and maximum battery life as the objective functions, and a genetic particle swarm hybrid algorithm is used to solve the optimal solution set; Step S8: deploy a dynamic priority scheduling module in the vehicle computing unit to allocate computing resources according to the real-time requirements of data processing to ensure that high-priority tasks are executed first; Step S9: Update the lightweight convolutional neural network parameters in step S3 based on the incremental learning algorithm, and combine them with the federated model parameters in step S5 to achieve online adaptive adjustment of the vehicle model.

2. The method according to claim 1, characterized in that: In step S1, the multi-source sensor includes a laser radar, a millimeter-wave radar, a camera, an inertial measurement unit, and a battery management system sensor; the data standardization processing includes timestamp alignment, unit unification, and missing value interpolation.

3. The method according to claim 1, characterized in that: In step S2, the adaptive weight allocation algorithm dynamically adjusts the fusion weight by calculating the information entropy and correlation coefficient of each data block. The specific formula is: Among them, W i is the weight of the i-th data block, N is the total number of data blocks generated after dynamic slicing, and E i is information entropy, C i is the average correlation coefficient with other data blocks, and α and β are adjustment factors.

4. The method according to claim 1, characterized in that: In step S4, the inter-vehicle communication topology model dynamically constructs a graph structure using vehicle positions, communication delays, and signal strengths as edge weights; the neighborhood propagation mechanism updates the current node features by aggregating feature vectors and edge weights of adjacent nodes.

5. The method according to claim 1, characterized in that: In step S5, the differential privacy mechanism ensures data privacy during model training by adding noise that satisfies the Laplace distribution to the gradient parameters. The noise addition formula is: in, is the perturbed gradient, g is the original gradient, Δf is the gradient sensitivity, ∈ is the privacy budget, and Lap(·) is the Laplace distribution generator.

6. The method according to claim 1, characterized in that: In step S7, the multi-objective optimization model adopts a Pareto front screening strategy, combining the global search capability of the genetic algorithm with the local optimization capability of the particle swarm algorithm to generate a non-dominated solution set.

7. The method according to claim 1, characterized in that: In step S8, the dynamic priority scheduling module dynamically divides task priorities according to task delay tolerance, data freshness and computational complexity, and allocates on-board computing resources using a preemptive scheduling strategy.

8. The method according to claim 1, characterized in that: In step S9, the incremental learning algorithm adjusts the network parameter update step by calculating the difference in feature distribution between new data and historical data. The specific formula is: Among them, θ t is the neural network parameter at the current moment, θ t+1 is the updated neural network parameter for the next moment, is the gradient of the loss function with respect to the parameters, indicating the direction of parameter optimization, D is the distribution difference measure, k is the decay coefficient, η is the initial learning rate, and exp (·) is the Sigmoid function, which maps the distribution difference to the (0,1) interval.

9. A vehicle-mounted data processing system for new energy vehicles based on intelligent networking, characterized in that: include: Data acquisition and preprocessing module: collects vehicle operation data, environmental perception data and battery status data through on-board multi-source sensors, and performs noise filtering and data standardization processing; Data dynamic slicing and fusion module: dynamically slices the preprocessed data based on spatiotemporal correlation, generates multi-dimensional data blocks, and fuses the data blocks using an adaptive weight allocation algorithm; Feature extraction module: Build an on-board edge computing node, use a lightweight convolutional neural network to extract high-dimensional features from fused data, and generate a set of feature vectors; Collaborative feature learning module: It builds a vehicle-to-vehicle communication topology model based on a dynamic graph neural network, updates the feature vector of each node through a neighborhood propagation mechanism, and implements collaborative feature learning in a connected vehicle environment; Distributed model training module: The federated learning framework is used to perform distributed model training on collaborative features, and a differential privacy mechanism is introduced to perturb the gradient parameters; Real-time decision-making module: Use reinforcement learning algorithm to build a real-time decision-making model, take the extracted high-dimensional features and trained model parameters as input, and output vehicle control instructions and energy scheduling strategies; Multi-objective optimization module: Based on the output results of the real-time decision-making module, a multi-objective optimization model is established, with minimum energy consumption, optimal path and maximum battery life as the objective functions, and the genetic particle swarm hybrid algorithm is used to solve the optimal solution set; Resource scheduling module: Deploy a dynamic priority scheduling module in the vehicle computing unit to allocate computing resources according to the real-time requirements of data processing to ensure that high-priority tasks are executed first; Model adaptive adjustment module: Updates the lightweight convolutional neural network parameters based on the incremental learning algorithm and combines the federated model parameters to achieve online adaptive adjustment of the vehicle model.

Images