Vehicle driving mode intelligent decision model construction method and system
By constructing an intelligent decision-making model that comprehensively considers parameters such as vehicle load and geographical location, and adopting a two-stage training strategy, the problems of automation and systematization of vehicle driving mode decision-making are solved, achieving personalized energy-saving and emission-reduction effects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-07
AI Technical Summary
Current vehicle driving mode decisions rely on driver experience, lack automation and systematization, and fail to comprehensively consider multi-dimensional dynamic parameters, resulting in limited energy conservation and emission reduction effects.
An intelligent decision-making model is constructed based on vehicle load, geographical location, and environmental parameters. It recommends the optimal driving mode through contextual Bandit algorithm and neural network training. Combined with a two-stage training strategy, it achieves personalized and automated decision-making.
It achieves the globally optimal driving mode, ensuring safety and energy saving, reducing parameter space complexity, adapting to different terrains and load conditions, and is highly applicable and easy to use on a large scale for commercial applications.
Smart Images
Figure CN121799424A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of vehicle energy-saving control technology, specifically relating to a method and system for constructing intelligent decision-making models for vehicle driving modes. Background Technology
[0002] With the increasing severity of energy security and climate change issues, energy conservation and emission reduction in vehicles are receiving more and more attention. In particular, heavy-duty commercial vehicles (such as long-haul logistics vehicles) have always been major energy consumers, making energy conservation and emission reduction a focus of industry attention. Currently, some advanced vehicles are equipped with multiple preset driving modes (such as economy mode and power mode) for drivers to choose from based on experience, aiming to achieve energy savings under different operating conditions. However, existing technologies have the following shortcomings: 1. Reliance on driver experience and lack of automation. Manual mode switching relies on the driver's subjective judgment and experience, making continuous optimization difficult in complex road conditions, resulting in low automation. 2. Limited decision-making dimensions and optimization effects. Some existing automatic mode switching methods typically only consider one or a few parameters, failing to systematically and individually consider multi-dimensional dynamic parameters such as real-time vehicle load, precise terrain, environmental climate, and the vehicle's own equipment status, leading to limited effectiveness and an inability to achieve global optimization. Summary of the Invention
[0003] The purpose of this invention is to provide a method and system for constructing an intelligent decision-making model for vehicle driving modes. The design covers at least the state and environmental parameter space of vehicle load and geographical location information. The intelligent decision-making model is trained based on real vehicle data and can automatically and personalizedly determine the optimal driving mode for the vehicle to meet dynamic constraints according to design goals (such as energy saving).
[0004] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:
[0005] In a first aspect, the present invention provides a method for constructing an intelligent decision-making model for vehicle driving modes, comprising the following steps:
[0006] Acquire sample data, including vehicle status and environmental parameters, driving mode, optimization target parameters, and dynamic constraint parameters; the status and environmental parameters include the vehicle's geographical grid, as well as one or more of the vehicle's load range, ambient temperature range, mileage range, maintenance records, and battery health status; the geographical grid is obtained by dividing the road map according to terrain based on a preset terrain granularity; the optimization target parameters are obtained by statistically analyzing the set targets based on a preset time or distance granularity;
[0007] An intelligent decision-making model is trained based on training sample data. The intelligent decision-making model takes the state and environmental parameters as input features and outputs a driving mode that optimizes the target parameters and satisfies the dynamic constraints.
[0008] Furthermore, the optimization target parameters include one or more of energy consumption per unit mileage and vehicle power.
[0009] Furthermore, the method of dividing the road map according to terrain based on a preset terrain granularity includes:
[0010] The road map is first divided according to administrative regions, and then divided according to terrain within the same administrative region. The terrain includes one or more of plains, plateaus, mountains, hills and basins. A continuous sub-region belonging to the same terrain attribute within the same administrative region is a geographic grid. The plateau type geographic grid is further divided according to altitude, and a continuous sub-region within the same altitude range is a geographic grid.
[0011] Alternatively, it may further include merging geographic grids with the same topographic attribute within adjacent administrative regions.
[0012] Furthermore, the dynamic constraint parameters are determined based on the vehicle's throttle opening distribution and vehicle acceleration; if the vehicle's actual acceleration is consistently lower than a preset acceleration threshold within a preset commonly used throttle opening range, then insufficient power is determined.
[0013] When training an intelligent decision-making model, select sample data that meets the dynamic constraints as the training sample data for the intelligent decision-making model, or impose penalties on sample data that does not meet the dynamic constraints.
[0014] Furthermore, the sample data is collected through experimental vehicles and / or actual vehicles on the road. The vehicles upload the collected state and environmental parameters, driving modes, optimization target parameters, and dynamic constraint parameters to the server. The server performs data cleaning, anomaly detection based on preset intervals and value ranges, and dynamic constraint checks to obtain training sample data for training the intelligent decision-making model. Alternatively, the vehicles preprocess the collected data locally before uploading it to the server. The preprocessing includes data cleaning, anomaly detection, and / or dynamic constraint checks.
[0015] Furthermore, the training of the intelligent decision-making model employs either a contextual bandit algorithm or a supervised learning algorithm. The contextual bandit algorithm treats each driving mode as an arm, the state and environmental parameters as context, and the optimized target parameters as rewards. The supervised learning algorithm uses a neural network, whose input includes the state and environmental parameters, and whose output is the optimized target parameters for each driving mode, in order to obtain the driving mode that optimizes the target. During training, the observed optimized target values are used to minimize the prediction error.
[0016] Furthermore, the intelligent decision-making model employs a two-stage training process. In the first stage, the contextual Bandit algorithm is used for online exploration and utilization, and updated model parameters are periodically or as needed sent to the vehicle to guide it in making decisions based on the latest model parameters. Once the amount of data in the sample data consisting of state and environmental parameters-driving mode combinations reaches a preset threshold, the model switches to a neural network-based supervised learning algorithm for offline training and optimization.
[0017] Secondly, the present invention provides an intelligent decision-making method for vehicle driving modes, comprising the following steps:
[0018] The intelligent decision-making model obtained according to the intelligent decision-making model construction method for vehicle driving mode described in the first aspect is deployed on the vehicle terminal to form an in-vehicle intelligent agent.
[0019] The in-vehicle intelligent agent recommends the optimal driving mode based on the real-time acquisition of the current vehicle status and environmental parameters;
[0020] Control the vehicle to switch to the optimal driving mode.
[0021] Furthermore, the vehicle starts according to a preset standard driving mode and periodically executes real-time decisions and applications at a preset frequency; and / or, the vehicle provides a human-machine interface for users to enable or disable the intelligent decision-making function of the driving mode.
[0022] Furthermore, when the onboard intelligent agent periodically decides the optimal driving mode based on the vehicle's state and environmental parameters at a preset frequency, if the decided optimal driving mode is different from the current driving mode, the vehicle will only be controlled to switch to the optimal driving mode when the same optimal driving mode is recommended for a continuous number of predictions.
[0023] Thirdly, the present invention provides a vehicle driving mode intelligent decision-making system, including an in-vehicle terminal and a server, for implementing the vehicle driving mode intelligent decision-making method according to the second aspect, wherein the in-vehicle terminal includes:
[0024] The parameter acquisition module is used to acquire the current vehicle status and environmental parameters in real time.
[0025] The in-vehicle intelligent agent is used to recommend the optimal driving mode based on the real-time acquisition of the current vehicle status and environmental parameters;
[0026] The vehicle control module is used to control the vehicle to switch to the optimal driving mode;
[0027] The communication module is used for data interaction with the cloud server;
[0028] The server includes:
[0029] The spatiotemporal database is used to store the collected vehicle driving status and environmental parameters, driving modes, optimization target parameters, and dynamic constraint parameters.
[0030] The model training module is used to train intelligent decision-making models periodically or on demand based on selected sample data.
[0031] The model management and distribution module is used to distribute the trained intelligent decision-making model to the vehicle terminal.
[0032] Fourthly, the present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the vehicle driving mode intelligent decision-making model construction method according to the first aspect or the steps of the vehicle driving mode intelligent decision-making method according to the second aspect.
[0033] Compared with the prior art, the present invention has the following significant advantages:
[0034] 1. Compared to traditional mode selection strategies that rely on a single or a few explicit parameters, this invention comprehensively considers state and environmental parameters, driving modes, optimization target parameters, and dynamic constraint parameters to conduct systematic parameter design. Furthermore, based on the core parameters of the vehicle's geographical grid, it can expand parameters such as the vehicle's load range, ambient temperature range, and vehicle mileage range, thus systematically constructing decision factors and achieving the globally optimal driving mode.
[0035] 2. When collecting sample data, this invention gives equal importance to dynamic constraint parameters and optimization target parameters. By excluding or penalizing samples with insufficient power, it can ensure that the trained model avoids undesirable patterns that would lead to weak vehicle acceleration and affect driving safety and efficiency when pursuing optimization goals such as energy saving, thus achieving a balance between safety, efficiency and energy saving.
[0036] 3. This invention uses a geographic grid to represent vehicle location information and divides various parameters such as vehicle load, ambient temperature, and vehicle mileage into intervals. It discretizes and structures continuous variables based on physical meaning, which greatly reduces the parameter space. The model only needs to learn the features of a limited number of geographic grids (such as 50) and numerical intervals, so that complex real-world optimization problems such as energy saving are transformed into a modelable and trainable intelligent decision-making model learning problem.
[0037] 4. This invention divides road maps into geographic grids based on preset terrain granularity, injecting prior knowledge of the different energy consumption characteristics of different terrains into the data representation. This effectively helps the model quickly and accurately capture the inherent correlation between geographical location and energy consumption indicators. Furthermore, the geographic grid division strategy adopted by this invention, which combines pre-division by administrative region with subdivision by terrain, offers both management convenience and refined adaptation. In addition, the grid merging strategy can reduce the total number of grids while ensuring accuracy, thus mitigating the cold start problem of data collection (i.e., insufficient data in a new grid).
[0038] 5. The intelligent decision-making model in this invention adopts a two-stage training strategy. The first stage uses the contextual Bandit algorithm for online exploration and utilization, which solves the problem of data scarcity in the early stages of system deployment. The system can be put into use immediately from scratch. The Bandit algorithm can quickly establish a preliminary but effective decision-making strategy based on limited real-time feedback. Furthermore, the Bandit algorithm can strategically try patterns that have not yet been fully validated through its inherent exploration and trade-off mechanism, efficiently collecting diverse data and generating high-quality training samples. The second stage utilizes neural networks to learn complex mapping functions, thereby making more accurate and stable optimization target predictions and driving mode recommendations. This phased design cleverly combines the advantages of both approaches, ensuring that the intelligent decision-making model of this invention is not only a high-performance algorithm model, but also a highly available, implementable, and evolving complete industrial solution.
[0039] 6. This invention can automatically select the optimal driving mode based on the unique state and environmental parameters of each vehicle, achieving truly personalized energy-saving optimizations. The in-vehicle intelligent agent used for intelligent decision-making in this invention can be deployed as an upper-level software strategy on existing vehicles with control and communication capabilities without requiring large-scale modifications to the vehicle hardware. It has strong applicability and is easy to apply on a large scale for commercial use. Attached Figure Description
[0040] Figure 1 The overall flowchart of the intelligent decision-making model construction method for vehicle driving mode provided in the embodiments of the present invention is shown.
[0041] Figure 2 The flowchart illustrates the two-stage training of the intelligent decision-making model provided in this embodiment of the invention.
[0042] Figure 3 The overall flowchart of the intelligent decision-making method for vehicle driving modes provided in the embodiments of the present invention is shown.
[0043] Figure 4 This is a block diagram of a vehicle driving mode intelligent decision-making system provided in an embodiment of the present invention. Detailed Implementation
[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0045] like Figure 1 As shown in the figure, the intelligent decision-making model construction method for vehicle driving mode provided by the embodiment of the present invention may include the following steps S11 to S12.
[0046] S11. Acquire sample data, including vehicle driving status and environmental parameters, driving mode, optimization target parameters, and dynamic constraint parameters; the status and environmental parameters include the vehicle's geographical grid, and one or more of the vehicle's load range, ambient temperature range, mileage range, maintenance records, and battery health status; the geographical grid is obtained by dividing the road map according to terrain based on a preset terrain granularity; the optimization target parameters are obtained by statistically analyzing the set targets based on a preset time or distance granularity;
[0047] S12. An intelligent decision-making model is trained based on training sample data. The intelligent decision-making model takes the state and environmental parameters as input features and outputs a driving mode that optimizes the target parameters and satisfies the dynamic constraints.
[0048] In this embodiment, the vehicle's driving mode refers to a set of various vehicle parameter curves during vehicle operation, including throttle sensitivity, vehicle power distribution (such as engine speed control; the power distribution ratio between the hybrid engine and the electric motor; and energy recovery intensity, battery discharge capacity, etc.), and gear control. When implementing this embodiment, for a specific vehicle model, the driving mode is a pre-set set of parameters. It is understood that this invention is not limited to a fixed set of driving modes; during implementation, the driving modes can be expanded as needed.
[0049] In this embodiment, considering the vehicle's driving mode, it is necessary to ensure that the vehicle carrying a certain weight of cargo has sufficient acceleration capability, while also taking into account the impact of various external conditions such as load, terrain, and environment (e.g., altitude and ambient temperature) on energy consumption. Furthermore, road terrain varies greatly across my country; for example, the North China Plain and Northeast Plain are predominantly plains, the eastern coastal areas are mainly hilly, and the southwest region is predominantly mountainous. Different terrains significantly affect the energy consumption of different vehicle models and load capacities. For example, for fuel-powered vehicles, in plains areas, the engine speed can be appropriately reduced to decrease energy consumption while maintaining vehicle acceleration. For hybrid vehicles, in hilly terrain, appropriately optimizing the energy recovery intensity and fully utilizing energy recovery during downhill driving can effectively reduce energy consumption. How to comprehensively consider the influence of external conditions and achieve optimal energy consumption by switching appropriate driving modes is the focus of this invention.
[0050] It should be noted that the implementation of this invention is not limited to my country, but is applicable globally. The optimization objectives considered in determining the optimal driving mode are not limited to energy conservation and emission reduction; the optimization target parameters may include one or more of energy consumption per unit mileage and vehicle power, determined according to actual design requirements.
[0051] In practical applications, step S11, which divides the road map according to terrain based on a preset terrain granularity, can include: first dividing the road map according to administrative regions, and then dividing it according to terrain within the same administrative region. Here, terrain can be one or more of plains, plateaus, mountains, hills, and basins. Continuous sub-regions within the same administrative region belonging to the same terrain attribute are treated as a geographic grid. For example, Sichuan Province can be further divided into the Chengdu Plain grid and the Western Sichuan Mountain grid. Plateau-type geographic grids can be further divided according to altitude, with continuous sub-regions within the same altitude range treated as a geographic grid. For example, Gansu Province can be divided into the Gannan Plateau grid and the Longzhong Loess Plateau grid. Directly training the model using the continuous latitude and longitude coordinates of collected vehicles would result in an infinitely large parameter space and would fail to effectively represent the key semantic information of terrain. This embodiment, by dividing the map into geographic grids containing terrain attributes, is equivalent to injecting prior knowledge of the different energy consumption characteristics of different terrains into the data representation, which can effectively help the model quickly and accurately capture the inherent relationship between geographical location and energy consumption indicators.
[0052] In one possible implementation, geographic grids with the same topographical attributes within adjacent administrative regions can be merged. For example, geographic grids belonging to neighboring provinces within the North China Plain can be merged. Specific merging rules can be flexibly determined manually based on management and data collection conditions, reducing the total number of grids and mitigating the cold start problem in data collection while ensuring accuracy.
[0053] Practical research has revealed that the vehicle's geographical location (including terrain) and load range (for logistics vehicles) significantly impact optimization objectives (especially energy consumption). In practical applications, the state and environmental parameters described in step S11 prioritize the vehicle's geographical grid location as the first priority, followed by the vehicle's load range. Further consideration of ambient temperature range, vehicle mileage range, maintenance records, and battery health status enables more accurate and personalized intelligent decision-making. The load range is determined by dividing the vehicle from empty to fully loaded based on a preset load granularity; the ambient temperature range is determined by dividing the ambient temperature range based on a preset temperature granularity; and the vehicle mileage range is determined by dividing the vehicle mileage range based on a preset mileage granularity. The specific granularity can be determined by comprehensively considering decision-making effectiveness and data collection. By dividing each parameter into intervals, the continuous variables are discretized and structured based on physical meaning, significantly reducing the parameter space. The model only needs to learn the features of a limited number of geographical grids and numerical intervals, transforming the complex real-world optimization problem into a modelable and trainable intelligent decision-making model learning problem.
[0054] For example, specific methods for setting and obtaining parameter space may include:
[0055] S111. Select a vehicle model and set K driving modes for it.
[0056] S112. Construct a parameter space to describe the vehicle's state and environmental parameters. This space includes:
[0057] (1) Divide the national road map into multiple grids according to the terrain (e.g., divide the national roads into 50 grids, some grids can be divided by province; some provinces have diverse terrain and can be divided into multiple grids). The vehicle determines which grid it is currently in through the positioning system.
[0058] (2) According to the vehicle’s load capacity, divide it into M load ranges (e.g., if the maximum load capacity of a vehicle is 40 tons, it can be divided into 20 load ranges from empty to fully loaded).
[0059] (3) Optionally, the ambient temperature can be divided into multiple ranges (e.g., from -30 to 50 degrees, divided into 8 ranges).
[0060] (4) Optionally, vehicle mileage can be divided into multiple ranges (e.g., from 0 to 1 million kilometers, divided into 10 ranges). Since the vehicle's power system (e.g., engine, battery, etc.) will affect the vehicle's overall energy consumption as the vehicle ages, vehicle mileage can be taken into account.
[0061] S113. Parameter Acquisition. Grid location can be acquired through a positioning module (such as BeiDou / GPS); vehicle load can be acquired through human-machine interface input, sensor acquisition, or vehicle power acceleration intelligent algorithm, etc.; ambient temperature, vehicle acceleration, and vehicle mileage can be acquired by onboard intelligent devices. Ambient temperature can be acquired through vehicle internal bus (such as CAN bus) data from vehicle temperature sensors, or through the temperature sensor of the onboard intelligent device itself; vehicle acceleration can be acquired through the acceleration sensor of the onboard intelligent device itself, or through a sensor interface (such as 485 bus); vehicle mileage can be acquired through the vehicle internal bus.
[0062] The parameter space can be expanded as needed to add different parameters. For example, vehicle maintenance records or battery health status (SoH) can also be expanded as external parameters. Vehicle maintenance records can be obtained from the manufacturer's database, and battery health status can be obtained through the vehicle's internal bus.
[0063] During sample data collection, the experimental vehicle conducted on-site experiments according to the above parameters, recording its parameter space coordinates in real time, including the current geographic grid Gi, load range Wj, ambient temperature range Tk, vehicle mileage range Ol, and the currently activated driving mode Mn. This process acquires specific state and environmental parameters, optimization target parameters under the driving mode, and dynamic constraint parameters. For example, in scenarios where the optimization objective is optimal energy consumption (for gasoline and hybrid vehicles, this means the lowest fuel consumption; for electric vehicles, it means the lowest electricity consumption; the target could also be optimal power, to adapt to power tuning scenarios), the vehicle's internal bus is connected via onboard intelligent devices to acquire the energy consumption per unit mile E of the vehicle over a fixed distance (e.g., 100 kilometers) or a fixed time period under the current state and environmental parameter input conditions. (The value of the fixed distance can be changed, such as changing it to 1 kilometer or 10 kilometers, which reduces the parameter space granularity; the length of the fixed time period can be increased, which expands the parameter space granularity.) Simultaneously, the accelerator pedal opening distribution and vehicle acceleration are also recorded synchronously over this distance or time period via onboard intelligent devices connected to the vehicle's internal bus. If the vehicle's actual acceleration remains below the safety and efficiency threshold within the commonly used throttle opening range (e.g., 30%-70%), the driving mode is deemed to have insufficient power. This data will be marked as invalid or subject to an extremely high energy consumption penalty. In other words, when training the intelligent decision-making model, only sample data that meets the power constraints will be selected as training sample data for the intelligent decision-making model, or a penalty will be imposed on sample data that does not meet the power constraints.
[0064] In practical applications, sample data can be collected from experimental vehicles or vehicles actually on the road. The experimental vehicles and / or vehicles actually on the road upload the collected state and environmental parameters, driving modes, optimization target parameters, and dynamic constraint parameters to the server. The server performs data cleaning, anomaly detection based on preset intervals and value ranges, and dynamic constraint checks to obtain training sample data for training the intelligent decision-making model. Alternatively, the vehicle can preprocess the collected data locally before uploading it to the server. The preprocessing includes data cleaning, anomaly detection, and / or dynamic constraint checks.
[0065] For example, the vehicle uploads all collected data tuples {Gi, Wj, Tk, Ol, Mn, E, power constraint flag} to a cloud server via an onboard communication module (such as 4G / 5G). Data collection and cleaning can be performed on the vehicle or server side. Taking cloud server processing as an example, the vehicle continuously uploads data, including: vehicle ID, timestamp, location grid Gi, load range Wj, temperature range Tk, mileage range Ol, current driving mode Mn, actual energy consumption E, and power performance indicators (relationship between throttle opening and acceleration). The cloud server performs missing value and outlier processing, deleting records with missing key fields. Outliers are detected based on preset ranges and value ranges. For example, negative energy consumption or energy consumption exceeding a reasonable range is considered abnormal. The system checks whether the load range, temperature range, and mileage range are within preset ranges to determine if there are any parameter anomalies; it also checks whether the location grid is in a known grid list and whether the driving mode is in a preset mode list. In addition, dynamic constraint checks are performed. For example, for each data point, the vehicle's acceleration should be greater than a certain threshold (e.g., 0.5 m / s²) under commonly used throttle opening. If it is below the threshold, the current mode is considered to have insufficient power, and this data cannot be used for training. The cloud server can also re-merge the vehicle-uploaded data according to preset time or distance granularity, calculating the merged data values based on the mean (continuous values) or mode (discrete values).
[0066] In step S12, the cloud server uses the preprocessed multidimensional parameters {Gi, Wj, Tk, Ol} as input features to train an intelligent decision-making model with the optimization objective of finding the driving mode Mn that minimizes energy consumption E and satisfies dynamic constraints. This intelligent decision-making model is essentially an optimal mapping function from the multidimensional parameter space to a finite set of driving modes.
[0067] like Figure 2As shown, in practical applications, considering the variability of samples, the training of the intelligent decision-making model adopts a two-stage training approach. The first stage (small sample method) involves experimental vehicles. Due to the limited number of vehicles and data, the contextual bandit algorithm can be used for online exploration and utilization, periodically or on-demand dissemination of updated model parameters to the experimental vehicles to guide them in making decisions based on the latest model parameters. The second stage (large sample method) involves actual on-road testing, for example, with 10,000 similar vehicles operating in different locations and under different conditions. This stage utilizes a supervised learning algorithm based on neural networks for offline training and optimization.
[0068] Specifically, the contextual Bandit algorithm treats each driving mode as an arm, the state and environmental parameters as context, and the optimized target parameters as rewards; the supervised learning algorithm uses a neural network, the input of which includes the state and environmental parameters, and the output is the optimized target parameters of each driving mode, so as to obtain the driving mode that optimizes the target. During training, the observed optimized target values are used to minimize the prediction error.
[0069] For example, the first phase uses the contextual bandit algorithm for exploration and utilization, collecting data. For instance, the LinUCB algorithm is used to update the model online, once daily or weekly, and then distributed to the vehicles. The specific process is as follows:
[0070] Assuming there are K driving modes (i.e. K arms), at each time t, the observed context feature vector... (Features such as geographic grid, load range, ambient temperature range, and mileage range), then select an arm (driving mode). Receive a reward (For example, negative energy consumption, because we want to minimize energy consumption, so maximizing the reward is equivalent to minimizing energy consumption).
[0071] The LinUCB algorithm maintains a ridge regression model for each arm to predict the reward in a given context. Specific steps:
[0072] S1211, Initialization.
[0073] For each arm Initialize the model parameter matrix Given a d-dimensional identity matrix and a parameter vector Let be a d-dimensional zero vector, where d is the dimension of the context feature vector.
[0074] S1212. For each time step t (i.e., each time the vehicle needs to select a mode), execute:
[0075] a. Observe the current context feature vector (d-dimensional vector);
[0076] b. For each arm Calculate ridge regression parameters And UCB score :
[0077]
[0078] in These are exploration parameters;
[0079] c. Select the one with the highest arm ;
[0080] d. Perform the action Observation Rewards ;
[0081] e. Update: ; .
[0082] In practical applications, vehicle data is centralized in the cloud, and the model is then updated periodically (daily or weekly). The updated model parameters are then... , The data is then sent to the vehicles, which make decisions based on the latest model parameters.
[0083] Phase 2: When the amount of data is large enough, and the amount of data for each state and environmental parameter-driving mode combination reaches a certain threshold, switch to supervised learning, train a neural network model, update it regularly, and recommend the driving mode with the lowest energy consumption and sufficient power.
[0084] Specifically, the neural network model takes state features (geographic grid, load, temperature, mileage, etc.) as input and outputs the predicted energy consumption for each driving mode. Then, it selects the driving mode with the lowest energy consumption. During training, actual observed energy consumption is used as the target value to minimize prediction error. In this way, the model can predict the energy consumption of each driving mode and then select the driving mode corresponding to the lowest energy consumption.
[0085] For example, suppose the collected data is in the form of: contextual features x, selected driving mode a, and actual energy consumption y. In supervised learning, the energy consumption for each driving mode is needed to train a model to predict the energy consumption of each driving mode given the contextual features. The specific steps are as follows:
[0086] S1221. Data Collection: The dataset D collected from the Bandit phase, each sample is (x, a, y, p), where p is the dynamic index when selecting mode a. In this example, only samples that meet the dynamic requirements are retained (i.e., p is true).
[0087] S1222. Data preprocessing: Standardize the context features x. Discrete features such as geographic grids, load ranges, and temperature ranges can be encoded using one-hot encoding or embedded representation.
[0088] S1223. Construct a neural network structure, where:
[0089] Input layer: Input context features x.
[0090] Hidden layers: Several fully connected layers. In this example, two layers are used, each with 128 neurons, using the ReLU activation function.
[0091] Output layer: K neurons. In this example, linear activation is used, and each neuron corresponds to the predicted energy consumption of a driving mode.
[0092] S1224. Construct the loss function: Use mean squared error, but only calculate the error of the output node corresponding to the actual selected driving mode. That is, for sample (x, a, y), only calculate the mean squared error between the predicted value of the a-th output node and y.
[0093] S1225. Model Training: Training is performed using gradient descent optimization.
[0094] S1226. Model Usage: Using the trained model, given input x, output K predicted energy consumptions, and select the driving mode with the lowest predicted energy consumption.
[0095] This embodiment combines online learning and offline training, enabling it to adapt to large-scale vehicle trials and gradually optimize recommendation performance. The first stage employs the contextual Bandit algorithm for online exploration and utilization, addressing the data scarcity issue in the early stages of system deployment. The system can be immediately deployed from scratch. The Bandit algorithm can quickly establish a preliminary but effective decision-making strategy based on limited real-time feedback. Furthermore, through its inherent exploration and trade-off mechanism, the Bandit algorithm strategically tries modes that have not yet been fully validated, efficiently collecting diverse data and generating high-quality training samples. The second stage utilizes neural networks to learn complex mapping functions, thereby making more accurate and stable optimization target predictions and driving mode recommendations.
[0096] like Figure 3 As shown in the figure, the intelligent decision-making method for vehicle driving mode provided by the embodiment of the present invention may include the following steps S21 to S23.
[0097] S21. Deploy the intelligent decision-making model on the vehicle terminal to form an in-vehicle intelligent agent;
[0098] S22. The in-vehicle intelligent agent recommends the optimal driving mode based on the real-time acquisition of the current vehicle status and environmental parameters.
[0099] S23. Control the vehicle to switch to the optimal driving mode.
[0100] In this embodiment, a lightweight intelligent decision-making model trained and converged in the cloud can be deployed to an in-vehicle intelligent terminal or vehicle domain controller to form an in-vehicle intelligent agent. In actual use, the vehicle first starts in standard driving mode, and then periodically executes real-time decisions and applications at a preset frequency. Specifically, the vehicle acquires various input parameters at a certain frequency (e.g., every 10 minutes), including load range, location grid, or further information such as ambient temperature range and mileage range. These real-time parameters are input into the in-vehicle intelligent agent, which performs forward inference and instantly outputs the recommended optimal driving mode M_optimal. The vehicle controller or power domain controller receives this instruction and automatically switches the vehicle's driving mode to M_optimal.
[0101] Furthermore, to prevent fluctuations in state and environmental parameters (e.g., the vehicle being at the edge of a geographic grid), if the optimal driving mode determined by the decision differs from the current driving mode, the vehicle will only switch to M_optimal after the same optimal driving mode M_optimal has been recommended for a consecutive number of predictions (e.g., 3 times). Additionally, the automatic switching of the optimal driving mode can also be manually disabled through the vehicle's human-machine interface; if disabled, the vehicle will no longer automatically switch to the optimal driving mode.
[0102] In practical applications, cloud servers can also continuously collect the actual operating data of all online vehicles, regularly retrain and optimize the intelligent decision-making model, and push model updates to vehicles via OTA, enabling the system to adapt to long-term dynamics such as road changes and vehicle performance degradation.
[0103] like Figure 4 As shown in the figure, an intelligent decision-making system for vehicle driving modes provided by an embodiment of the present invention includes an in-vehicle terminal and a server, used to implement the intelligent decision-making method for vehicle driving modes. The in-vehicle terminal includes: a parameter acquisition module for acquiring the current vehicle status and environmental parameters in real time; an in-vehicle intelligent agent for recommending the optimal driving mode based on the real-time acquired current vehicle status and environmental parameters; a vehicle control module for controlling the vehicle to switch to the optimal driving mode; and a communication module for data interaction with the cloud server. The server includes: a spatiotemporal database for storing the acquired vehicle driving status and environmental parameters, driving mode, optimization target parameters, and dynamic constraint parameters; a model training module for periodically or on-demand training of the intelligent decision-making model based on selected sample data; and a model management and distribution module for distributing the trained intelligent decision-making model to the in-vehicle terminal.
[0104] The present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the vehicle driving mode intelligent decision-making model construction method or the steps of the vehicle driving mode intelligent decision-making method in any of the above embodiments.
[0105] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of the present invention to be performed.
[0106] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.
[0107] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for constructing an intelligent decision-making model for vehicle driving modes, characterized in that, Includes the following steps: Acquire sample data, including vehicle status and environmental parameters, driving mode, optimization target parameters, and dynamic constraint parameters; the status and environmental parameters include the vehicle's geographical grid, as well as one or more of the vehicle's load range, ambient temperature range, mileage range, maintenance records, and battery health status; the geographical grid is obtained by dividing the road map according to the terrain based on a preset terrain granularity; The optimization target parameters are obtained by statistical analysis of the set target based on a preset time or distance granularity. An intelligent decision-making model is trained based on training sample data. The intelligent decision-making model takes the state and environmental parameters as input features and outputs a driving mode that optimizes the target parameters and satisfies the dynamic constraints.
2. The method for constructing an intelligent decision-making model for vehicle driving modes according to claim 1, characterized in that, The optimization target parameters include one or more of the following: energy consumption per unit mileage and vehicle power.
3. The method for constructing an intelligent decision-making model for vehicle driving modes according to claim 1, characterized in that, The method of dividing the road map according to terrain based on a preset terrain granularity includes: The road map is first divided according to administrative regions, and then divided according to terrain within the same administrative region. The terrain includes one or more of plains, plateaus, mountains, hills and basins. A continuous sub-region belonging to the same terrain attribute within the same administrative region is a geographic grid. The plateau type geographic grid is further divided according to altitude, and a continuous sub-region within the same altitude range is a geographic grid. Alternatively, it may further include merging geographic grids with the same topographic attribute within adjacent administrative regions.
4. The method for constructing an intelligent decision-making model for vehicle driving modes according to claim 1, characterized in that, The dynamic constraint parameters are determined based on the vehicle's throttle opening distribution and vehicle acceleration; if the vehicle's actual acceleration is consistently lower than the preset acceleration threshold within the preset commonly used throttle opening range, then insufficient power is determined. When training an intelligent decision-making model, select sample data that meets the dynamic constraints as the training sample data for the intelligent decision-making model, or impose penalties on sample data that does not meet the dynamic constraints.
5. The method for constructing an intelligent decision-making model for vehicle driving modes according to claim 1, characterized in that, The sample data is collected through experimental vehicles and / or actual vehicles on the road. The vehicles upload the collected state and environmental parameters, driving modes, optimization target parameters, and dynamic constraint parameters to the server. The server performs data cleaning, anomaly detection based on preset intervals and value ranges, and dynamic constraint checks to obtain training sample data for training the intelligent decision-making model. Alternatively, the vehicles preprocess the collected data locally before uploading it to the server. The preprocessing includes data cleaning, anomaly detection, and / or dynamic constraint checks.
6. The method for constructing an intelligent decision-making model for vehicle driving modes according to claim 1, characterized in that, The training of the intelligent decision-making model employs either a contextual bandit algorithm or a supervised learning algorithm. The contextual bandit algorithm treats each driving mode as an arm, the state and environmental parameters as context, and the optimized target parameters as rewards. The supervised learning algorithm uses a neural network, whose input includes the state and environmental parameters, and whose output is the optimized target parameters for each driving mode, in order to obtain the driving mode that optimizes the target. During training, the observed optimized target values are used to minimize the prediction error.
7. The method for constructing an intelligent decision-making model for vehicle driving modes according to claim 6, characterized in that, The intelligent decision-making model employs a two-stage training process. In the first stage, the contextual Bandit algorithm is used for online exploration and utilization, periodically or as needed, to send updated model parameters to the vehicle, guiding the vehicle to make decisions based on the latest model parameters. Once the amount of data in the sample data consisting of state and environmental parameters-driving mode combinations reaches a preset threshold, the model switches to a neural network-based supervised learning algorithm for offline training and optimization.
8. A vehicle driving mode intelligent decision-making method, characterized in that, Includes the following steps: The intelligent decision-making model obtained by the intelligent decision-making model construction method for vehicle driving mode according to any one of claims 1-7 is deployed on the vehicle terminal to form an in-vehicle intelligent agent; The in-vehicle intelligent agent recommends the optimal driving mode based on the real-time acquisition of the current vehicle status and environmental parameters; Control the vehicle to switch to the optimal driving mode.
9. The intelligent decision-making method for vehicle driving modes according to claim 8, characterized in that, The vehicle starts according to a preset standard driving mode and periodically executes real-time decisions and applications at a preset frequency; and / or, the vehicle provides a human-machine interface for users to enable or disable the intelligent decision-making function of driving modes.
10. The intelligent decision-making method for vehicle driving modes according to claim 8, characterized in that, When the vehicle-mounted intelligent agent periodically decides the optimal driving mode based on the vehicle's status and environmental parameters at a preset frequency, if the optimal driving mode is different from the current driving mode, the vehicle will only switch to the optimal driving mode after the same optimal driving mode has been recommended a certain number of times in a row.
11. A vehicle driving mode intelligent decision-making system, comprising an on-board terminal and a server, for implementing the method according to any one of claims 8-10, characterized in that, The vehicle-mounted terminal includes: The parameter acquisition module is used to acquire the current vehicle status and environmental parameters in real time. The in-vehicle intelligent agent is used to recommend the optimal driving mode based on the real-time acquisition of the current vehicle status and environmental parameters; The vehicle control module is used to control the vehicle to switch to the optimal driving mode; The communication module is used for data interaction with the cloud server; The server includes: The spatiotemporal database is used to store the collected vehicle driving status and environmental parameters, driving modes, optimization target parameters, and dynamic constraint parameters. The model training module is used to train intelligent decision-making models periodically or on demand based on selected sample data. The model management and distribution module is used to distribute the trained intelligent decision-making model to the vehicle terminal.
12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-10.