EPS adaptive calibration method and device based on AI intelligence

Abstract

The invention discloses an EPS adaptive calibration method and device based on AI intelligence, and the method comprises the steps: obtaining the information collected by a vehicle and the voice information of a driver, and constructing an EPS calibration sequence sample data set; constructing a parameter calibration model based on reinforcement learning; performing offline training based on a reinforcement learning parameter calibration model according to the sequence sample data set between EPS calibration to obtain a preliminary self-adaptive calibration model; in the actual running process of the vehicle. The beneficial effects are that the system can learn and adapt to the personalized preference of the current driver by introducing the driver voice feedback as a user satisfaction award item.

Description

Technical Field

[0001] This invention relates to the field of vehicle calibration, and more particularly to an AI-based adaptive EPS calibration method. Background Technology

[0002] With the continuous maturation of artificial intelligence technology and the ongoing improvement in the electrification and intelligence of automobiles, the vehicle development cycle has been significantly compressed, while the standards for vehicle matching performance and technical requirements are constantly increasing. In the development of electric power steering systems, the vehicle matching calibration of the steering system is particularly critical, and it is imperative to complete the calibration of a high-performance and robust system within a limited time.

[0003] Currently, traditional EPS calibration methods rely primarily on human experience. It typically requires at least two experienced calibration engineers to complete the calibration work through module-by-module, item-by-item debugging and iterative processes. This process is not only time-consuming and labor-intensive but also highly dependent on the engineers' subjective judgment, lacking objective data support. During the delivery phase, differences in subjective evaluation often lead to multiple reworks, further lengthening the development cycle and increasing labor costs.

[0004] Furthermore, traditional calibration methods struggle to quantify system robustness, posing certain performance risks. The lack of systematic accumulation and intelligent analysis of historical calibration data prevents the formation of effective experience feedback and continuous optimization mechanisms, thus limiting the adaptive capabilities and performance stability of the EPS system in diverse driving scenarios. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to solve the technical problems in the prior art and provide an AI-based adaptive calibration method and device for EPS.

[0006] Technical solution:

[0007] Firstly, this application proposes an AI-based adaptive EPS calibration method, including:

[0008] Acquire information collected from vehicles and voice information from drivers to construct an EPS calibration sequence sample dataset;

[0009] Construct a parameter calibration model based on reinforcement learning;

[0010] A preliminary adaptive calibration model was obtained by offline training based on the parameter calibration model of reinforcement learning using the EPS calibration sequence sample dataset.

[0011] Based on the preliminary adaptive calibration model, information collected by the vehicle in real time is acquired and combined with driver voice feedback to perform online reinforcement learning training on the parameter calibration model, dynamically adjust and output EPS calibration parameters to the ECU.

[0012] Preferably, the acquisition of vehicle-collected information includes different functional modules for current vehicle-collected information and historical vehicle-collected information, including:

[0013] Torque, angle, temperature, power steering current, steering wheel torque, angle, as well as speed, torque, vehicle speed, steering angle, angular velocity, and lateral acceleration.

[0014] Preferably, the parameter calibration model based on reinforcement learning includes, in sequence, an input layer, a state encoding layer, a policy network layer, a value network layer, and a parameter output layer.

[0015] Preferably, the input layer is used to concatenate historical vehicle data into a complete state vector;

[0016] The state coding layer uses parallel fully connected sub-networks to extract local features and segment and encode the state information required by the four functional modules of basic assist, self-alignment, damping, and friction compensation, forming local state feature vectors for each module.

[0017] The strategy network layer configures an improved Actor network for each of the four functional modules: basic assist, self-centering, damping, and friction compensation. Each Actor network takes the corresponding local state feature vector as input and outputs the continuous adjustment amount of the calibration parameters of the functional module.

[0018] The value network layer deploys a dual Critic network for each functional module to output a state-action value function, thereby evaluating the action output by the Actor network and calculating the policy gradient error.

[0019] The parameter output layer performs amplitude limiting and smoothing filtering on the calibration parameter adjustment amounts output by each Actor network, integrates them into a complete set of EPS calibration parameters, and writes them into the ECU through the calibration interface.

[0020] Preferably, an initial adaptive calibration model is obtained by offline training based on a reinforcement learning-based parameter calibration model using the EPS calibration interval sequence sample dataset, including:

[0021] From the historical vehicle data, the corresponding state-action-report sequences are extracted according to the tags of different functional modules and stored in their respective experience playback buffers.

[0022] In each training iteration, the Actor network and Critic network of each functional module are trained independently in turn.

[0023] The Actor network updates its data by maximizing the expected reward using a deterministic policy gradient ascent algorithm.

[0024] The Critic network is updated by minimizing the Bellman error, which is caused by the temporal difference error;

[0025] The target network parameters of each functional module are synchronized using a soft update method until the reward function of each module converges, forming the initial policy library.

[0026] Preferably, the offline training uses a combined loss function, the expression of which is:

[0027] ;

[0028] in, For the clip target of the PPO algorithm, Let be the mean square error of the value function Critic. Let c be the policy entropy, and c1 and c2 be the weight coefficients.

[0029] Preferably, based on the preliminary adaptive calibration model, information collected by the vehicle in real time is acquired, and combined with driver voice feedback, the parameter calibration model is trained online through reinforcement learning, dynamically adjusted, and the EPS calibration parameters are output to the ECU, including:

[0030] During vehicle operation, the system enters the online enhancement phase based on the initial policy library obtained through offline training.

[0031] Real-time collection of vehicle information;

[0032] Simultaneously collect the driver's voice feedback, and convert the voice feedback into a quantitative user satisfaction score through the voice recognition module;

[0033] Construct an online objective function, which adds a user satisfaction reward item based on the user satisfaction score to the existing built-in rewards of the module.

[0034] The corresponding Actor-Critic network is locally updated using mini-batch gradient descent and empirical replay methods.

[0035] Preferably, the Actor-Critic network adopts the improved Deep Deterministic Policy Gradient (DDPG) framework, which includes a policy network layer and a value network layer. The policy network layer configures an independent Actor network for each functional module, and the value network layer deploys a dual Critic network.

[0036] Preferably, the online objective function is in the form of:

[0037] ;

[0038] In the formula, The average reward for all possible driving conditions. Weighting of the basic support module. To correct the performance module weights, For the damping module weights, For the weights of the friction compensation module, Basic support sub-rewards, As a reward for Zhengzi, As a reward for damping, For friction compensation sub-rewards;

[0039] Each sub-reward is calculated in real time based on measurable signals, and the expression is as follows:

[0040] ;

[0041] Where T represents the steering wheel torque. The torque-angle linear residual is ΔT, representing the difference in left and right steering torques.

[0042] ;

[0043] in, The remaining angle after correction. To correct the overshoot angle;

[0044] ;

[0045] in, For the high-frequency standard deviation of torque, The high-frequency standard deviation of the motor current;

[0046] ;

[0047] In the formula, This is the gain for small-angle yaw.

[0048] Secondly, embodiments of the present invention provide an AI-based adaptive calibration device for EPS, comprising:

[0049] The acquisition unit is used to acquire information collected by the vehicle and the driver's voice information to construct an EPS calibration sequence sample dataset.

[0050] Building units are used to construct parameter calibration models based on reinforcement learning;

[0051] The offline model unit is used to train a preliminary adaptive calibration model offline based on a reinforcement learning parameter calibration model using an EPS calibration sequence sample dataset.

[0052] The online model unit is used to collect information collected by the vehicle in real time based on the preliminary adaptive calibration model, and combine it with the driver's voice feedback to perform online reinforcement learning training on the parameter calibration model, dynamically adjust and output EPS calibration parameters to the ECU.

[0053] Thirdly, embodiments of the present invention provide an electronic device, including a processor and a memory. The memory stores one or more computer programs; when the one or more computer programs stored in the memory are executed by the processor, the electronic device is able to implement any of the possible design methods described in the first aspect.

[0054] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in any of the above embodiments.

[0055] Fifthly, embodiments of the present invention also provide a computer program product that, when run on an electronic device, causes the electronic device to perform any possible design method of any of the above aspects.

[0056] Beneficial effects:

[0057] 1. The system automatically optimizes and iterates parameters through a reinforcement learning model, achieving online lifelong learning of "driving, calibrating, and optimizing simultaneously." This shortens the traditional lengthy process of repeated manual road tests and adjustments into a one-time automated learning process, significantly reducing the calibration cycle. It eliminates the need for two or more senior calibration engineers throughout the process; a single driver can complete the calibration work under specific operating conditions. This significantly reduces reliance on manpower and the demand for specialized experience, thereby reducing the company's labor and training costs.

[0058] 2. It abandons the limitations of traditional methods that rely on the subjective feelings of engineers. By constructing objective and quantifiable multi-objective reward functions (such as assist smoothness, self-centering performance, and damping feel) through multi-sensor data (such as torque, angle, and acceleration), calibration decisions are based on real data and algorithm models, resulting in more scientific, consistent, and reliable results. The reinforcement learning model can comprehensively consider the balance of multiple performance objectives, automatically finding the globally optimal or near-optimal parameter combinations under different operating conditions. This avoids the potential for overlooking certain aspects in manual calibration, thus significantly improving the adaptability and robustness of the EPS system in various complex driving scenarios.

[0059] 3. The system can continuously record data and strategies during the calibration process, forming a valuable experience database. This not only provides data support for subsequent offline model training but also enables the system to continuously learn and evolve, becoming increasingly intelligent with use. By introducing driver voice feedback as a reward for user satisfaction, the system can learn and adapt to the current driver's personalized preferences (such as a preference for a light or firm feel), achieving customized steering feel for each individual driver, thus enhancing user experience and product added value. Attached Figure Description

[0060] Figure 1 A schematic diagram of the method framework for this invention is provided;

[0061] Figure 2 This is a flowchart of the EPS calibration process of the present invention;

[0062] Figure 3 This is a schematic diagram of the system modules of the present invention;

[0063] Figure 4 This is a block diagram of the calibration interface for each module of the present invention;

[0064] Figure 5 The diagram shows the calibration parameters of the basic power assist module at various vehicle speeds.

[0065] Figure 6 This is a diagram showing the correspondence between the basic power assist module data of this invention and the ECU software;

[0066] Figure 7 This is a block diagram of a device structure provided in one embodiment of this application;

[0067] Figure 8 This is a block diagram of an electronic device structure provided in one embodiment of this application. Detailed Implementation

[0068] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0069] Example 1

[0070] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed following the word and its equivalents, but do not exclude other elements or objects.

[0071] In response to the problems existing in the current technology, such as Figure 1-6 As shown, an AI-based adaptive EPS calibration method is proposed, characterized by including:

[0072] S101. Obtain vehicle-collected information and driver's voice information to construct an EPS calibration sequence sample dataset;

[0073] In some specific embodiments, the acquisition of vehicle-collected information includes different functional modules for current vehicle-collected information and historical vehicle-collected information, including:

[0074] Torque, angle, temperature, power steering current, steering wheel torque, angle, as well as speed, torque, vehicle speed, steering angle, angular velocity, and lateral acceleration.

[0075] In some specific embodiments, the process also includes obtaining the vehicle's calibration request and the obtained vehicle status information and uploading them to the cloud platform.

[0076] In some real-world test cases, the system collects vehicle status information from various sources, including the vehicle's various sensor modules. This information includes, but is not limited to:

[0077] Steering angle, current, and temperature sensor data.

[0078] Dynamic parameters such as torque, vehicle speed, steering angle, angular velocity, and lateral acceleration.

[0079] The system also collects voice feedback data from the driver, which is used to further analyze the driver's feelings about the current steering system (such as whether it is too heavy or too light, and whether the return to center is smooth).

[0080] Voice information is used to evaluate the steering system's performance based on the driver's subjective feelings, thus serving as a basis for adjusting calibration parameters.

[0081] The collected vehicle information and driver feedback were integrated to construct a sample dataset containing current and historical vehicle information. This dataset was used for subsequent EPS adaptive calibration.

[0082] S102. Construct a parameter calibration model based on reinforcement learning;

[0083] In some specific embodiments, the parameter calibration model based on reinforcement learning includes, in sequence, an input layer, a state encoding layer, a policy network layer, a value network layer, and a parameter output layer.

[0084] In some specific embodiments, the input layer is used to concatenate historical vehicle data into a complete state vector. The function of the input layer is to merge multiple vehicle dynamics-related parameters (such as steering wheel torque, vehicle speed, steering angle, angular velocity, motor current, lateral acceleration, temperature, and power assist current) into a single state vector. This data comes from the vehicle's sensors and represents the vehicle's current operating conditions and dynamic state. By merging this data, the system can gain a comprehensive understanding of the vehicle's current operating status.

[0085] The state encoding layer, through parallel fully connected sub-networks, extracts local features and segments the state information required by the four functional modules—basic assist, self-alignment, damping, and friction compensation—to form local state feature vectors for each module. The state encoding layer also extracts local features and segments the raw data from the input layer through parallel fully connected sub-networks. Each functional module (basic assist, self-alignment, damping, friction compensation, etc.) has an independent encoding process. The purpose of this layer is to transform complex input data into local state feature vectors required by each module. Each module's feature vector is specifically used for subsequent processing and decision-making within that module.

[0086] The strategy network layer configures an improved Actor network for each of the four functional modules: basic power assist, self-centering, damping, and friction compensation. Each Actor network takes its corresponding local state feature vector as input and outputs the continuous adjustment amount of the calibration parameters for that functional module. Each functional module (such as basic power assist, self-centering, damping, and friction compensation) is configured with an improved Actor network. The input of this Actor network is the local state feature vector of each module, and the output is the continuous adjustment amount of the calibration parameters for that module. The Actor network calculates the corresponding adjustment amount based on the current state characteristics, with the goal of optimizing the steering system's performance. For example, how to adjust the basic power assist for easier handling, or how to optimize the self-centering torque, etc.

[0087] The value network layer deploys a dual Critic network for each functional module, outputting a state-action value function to evaluate the actions output by the Actor network and calculate the policy gradient error. Each functional module has a dual Critic network. The Critic network evaluates the Actor network's output actions and calculates the policy gradient error, thus helping to optimize the Actor network's policy. The Critic network assesses the Actor network's performance based on the current state-action value and reduces errors through a policy gradient update algorithm, gradually improving the calibration results.

[0088] The parameter output layer limits and smooths the calibration parameter adjustments output by each Actor network, then integrates them into a complete set of EPS calibration parameters, which are then written to the ECU via the calibration interface. These adjustments are then written to the vehicle's electronic control unit (ECU) for actual application in controlling the vehicle's steering system.

[0089] S103. Based on the EPS calibration sequence sample dataset, a preliminary adaptive calibration model is obtained through offline training using a parameter calibration model based on reinforcement learning.

[0090] In some specific embodiments, an initial adaptive calibration model is obtained by offline training based on a reinforcement learning-based parameter calibration model using an EPS calibration sequence sample dataset, including:

[0091] From the historical vehicle data, the corresponding state-action-report sequences are extracted according to the tags of different functional modules and stored in their respective experience playback buffers.

[0092] In each training iteration, the Actor network and Critic network for each functional module are trained independently. Training data: Based on the EPS calibration sequence sample dataset, the reinforcement learning model is trained using historical vehicle data. This data includes vehicle state and action labels, which are used for subsequent calibration model training.

[0093] The process includes the following steps:

[0094] Actor Network Training: The Actor Network maximizes the expected reward through gradient ascent using a deterministic policy, adjusting its calibration parameters accordingly. Through continuous training, the Actor Network can optimize the policy and progressively update the calibration parameters of each functional module to improve system performance.

[0095] Critic Network Training: The Critic network updates state-action values ​​by minimizing the Bellman error, evaluates the actions output by the Actor network, and helps optimize the training process. By minimizing the error of the current policy, the Critic network continuously adjusts the objective function, helping the Actor network improve.

[0096] The Actor network updates its data by maximizing the expected reward using a deterministic policy gradient ascent algorithm.

[0097] The Critic network is updated by minimizing the Bellman error, which is caused by the temporal difference error;

[0098] The target network parameters of each functional module are synchronized using a soft update method until the reward function of each module converges, forming the initial policy library.

[0099] In some specific embodiments, the offline training employs a combined loss function, the expression of which is: the loss function L(θ) is calculated by combining the loss functions of the PPO algorithm and the Critic network:

[0100] ;

[0101] in, The clip objective of the PPO algorithm aims to limit the policy update magnitude and prevent over-updates from causing training failure. The mean squared error of the value function Critic helps calculate and optimize the value of the current strategy. is the policy entropy, used to increase the exploratory nature of the policy and avoid local optima; c1 and c2 are weight coefficients.

[0102] By optimizing the loss function described above, the system can continuously improve the policy and optimize the calibration parameters of the steering system during training. Especially in the PPO algorithm, by limiting the policy update magnitude and introducing policy entropy, the system avoids overly aggressive policies or getting trapped in local optima.

[0103] S104. Based on the preliminary adaptive calibration model, the vehicle information is collected in real time, and combined with the driver's voice feedback, the parameter calibration model is trained online through reinforcement learning, and the EPS calibration parameters are dynamically adjusted and output to the ECU.

[0104] In some specific embodiments, based on the preliminary adaptive calibration model, information collected by the vehicle in real time is acquired, and combined with driver voice feedback, the parameter calibration model is trained online through reinforcement learning, dynamically adjusted, and the EPS calibration parameters are output to the ECU, including:

[0105] During vehicle operation, the system first uses the initial strategy library obtained through offline training as a benchmark to enter the online enhancement phase. This strategy library, acquired through historical data and experiments in the offline phase, is then used for real-time adjustment and optimization in the online phase. During vehicle operation, the system collects various vehicle information in real time (such as steering wheel torque, vehicle speed, and steering angle) to dynamically adjust calibration parameters. Simultaneously, the system collects driver voice feedback and converts it into quantitative user satisfaction scores using a voice recognition module. The user satisfaction score, as a quantitative feedback, reflects the driver's actual experience with the steering system (such as steering weight and return-to-center speed). An online objective function is constructed, which adds a user satisfaction reward item based on the user satisfaction score to the existing built-in rewards. This new reward item makes the model more attentive to driver feedback, thereby more accurately adjusting calibration parameters to improve the driver's driving experience.

[0106] Mini-batch gradient descent and empirical replay are used to locally update the corresponding Actor-Critic network during training. This method improves the model's convergence speed and stability with each update.

[0107] In some specific embodiments, the Actor-Critic network adopts an improved deep deterministic policy gradient (DDPG) framework, which includes a policy network layer and a value network layer. The policy network layer configures an independent Actor network for each functional module, and the value network layer deploys a dual Critic network.

[0108] The improved DDPG framework (Deep Deterministic Policy Gradient) is used. DDPG is an Actor-Critic-based reinforcement learning algorithm suitable for continuous action space problems. The framework consists of two layers: a policy network (Actor network) and a value network (Critic network). Each functional module (e.g., basic assist, homing, damping) has an independent Actor network. Each Actor network outputs corresponding calibration parameter adjustments based on the current state characteristics. Each module has a dual Critic network used to evaluate the Actor network's output and update the policy through error updates, further optimizing the calibration process.

[0109] In some specific embodiments, the online objective function takes the form of:

[0110] ;

[0111] In the formula, The average reward for all possible driving conditions. Weighting of the basic support module. To correct the performance module weights, For the damping module weights, For the weights of the friction compensation module, Basic support sub-rewards, As a reward for Zhengzi, As a reward for damping, For friction compensation sub-rewards;

[0112] Each sub-reward is calculated in real time based on measurable signals, and the expression is as follows:

[0113] ;

[0114] Where T represents the steering wheel torque. The torque-angle linear residual is ΔT, representing the difference in left and right steering torques.

[0115] ;

[0116] in, The remaining angle after correction. To correct the overshoot angle;

[0117] ;

[0118] in, For the high-frequency standard deviation of torque, The high-frequency standard deviation of the motor current;

[0119] ;

[0120] in, This is the gain for small-angle yaw.

[0121] In some specific embodiments, the method also includes collecting the output setting information of the vehicle calibration engineer (the calibration values ​​written into the ECU when the engineer manually adjusts the parameters (such as the power assist gain table, damping coefficient, etc.)), calculating the deviation between the engineer's set parameters and the online model calculation, and continuously fine-tuning the parameters and adjusting the model using the real-time deviation during the online phase.

[0122] In other embodiments of the invention, combined with Figure 7 This invention discloses an AI-based adaptive EPS calibration device, comprising:

[0123] The acquisition unit 201 is used to acquire information collected by the vehicle and the driver's voice information to construct an EPS calibration sequence sample dataset.

[0124] Construction unit 202 is used to construct a parameter calibration model based on reinforcement learning;

[0125] The offline model unit 203 is used to train the parameter calibration model based on reinforcement learning offline based on the EPS calibration sequence sample dataset to obtain a preliminary adaptive calibration model.

[0126] The online model unit 204 is used to collect information collected by the vehicle in real time based on the preliminary adaptive calibration model, and combine it with the driver's voice feedback to perform online reinforcement learning training on the parameter calibration model, dynamically adjust and output EPS calibration parameters to the ECU.

[0127] In some specific embodiments, the calibration interface is a calibration protocol between AI data and the ECU calibration host computer software, which can be XCP or CCP, etc. Through the calibration interface, the data that the AI ​​needs to modify can be changed online via the protocol to modify the relevant parameters of the calibration module. The calibration module includes a basic power assist module for different vehicle speeds, a self-centering module, friction compensation, damping compensation, system ATF parameters, etc.

[0128] In some specific embodiments, the model of this application, through demand analysis and data analysis, precisely modifies the parameters of the calibration module based on a big data learning model. This AI module needs to repeatedly learn from calibration data matching different vehicle models and projects to accurately determine which module's parameters, calibrated through the ECU, and in what way, can solve or improve the target performance of existing needs. (For example, if it's necessary to lighten the effort required for stationary adjustment, the AI ​​will systematically optimize the stationary assist parameters, increasing the current equivalent value to reduce the effort required.) Figure 5 )

[0129] In some specific embodiments, such as Figure 6 The ECU controls the steering motor's movement and output performance through internal control software to ensure the EPS system meets the vehicle's performance requirements. The electric steering system primarily relies on the motor as the power assist output, amplified through reduction mechanisms such as worm gears, thus replacing manual steering with electric assistance. Furthermore, due to the ECU's control over the motor, it can assess the vehicle's overall condition and adapt the steering performance to meet different operating conditions. Figure 4 The calibration interface block diagram for each module.

[0130] In other embodiments of the present invention, an electronic device 400 is disclosed, such as... Figure 8As shown, the electronic device may include: one or more processors 401; a memory 402; a display 403; one or more application programs (not shown); and one or more computer programs 404. These devices can be connected via one or more communication buses 405. The one or more computer programs 404 are stored in the memory 402 and configured to be executed by the one or more processors 401. The one or more computer programs 404 include instructions that can be used to perform actions such as... Figures 1 to 2 And the various steps in the corresponding embodiments.

[0131] Through the above description of the embodiments, those skilled in the art will clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0132] In the various embodiments of this invention, the functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0133] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as flash memory, portable hard disk, read-only memory, random access memory, magnetic disk, or optical disk.

[0134] The above description is merely a specific implementation of the embodiments of the present invention, but the protection scope of the embodiments of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed in the embodiments of the present invention should be covered within the protection scope of the embodiments of the present invention. Therefore, the protection scope of the embodiments of the present invention should be determined by the protection scope of the claims.

Claims

1. An AI intelligence-based EPS adaptive calibration method, characterized in that, The method comprises the following steps: acquiring information collected by a vehicle and voice information of a driver, and constructing an EPS calibration interval sequence sample data set; constructing a parameter calibration model based on reinforcement learning; performing offline training on the parameter calibration model based on reinforcement learning according to the EPS calibration interval sequence sample data set to obtain a preliminary adaptive calibration model; based on the preliminary adaptive calibration model, acquiring information collected by the vehicle in real time, combining voice feedback of the driver, performing online reinforcement learning training on the parameter calibration model, and dynamically adjusting and outputting EPS calibration parameters to an ECU. 2.The AI intelligent-based EPS adaptive calibration method according to claim 1, characterized in that, The acquisition of the information collected by the vehicle comprises different functional modules of current vehicle collection information and historical vehicle collection information, and comprises: torque angle temperature, assist current, steering wheel torque, angle, and rotation speed, torque, vehicle speed, rotation angle, angular velocity, and lateral acceleration.

3. The AI intelligent-based EPS adaptive calibration method according to claim 2, wherein the parameter calibration model based on reinforcement learning comprises an input layer, a state coding layer, a policy network layer, a value network layer, and a parameter output layer in sequence.

4. The AI intelligent-based EPS adaptive calibration method according to claim 3, wherein the input layer is used for splicing the historical vehicle collection information into a complete state vector; the state coding layer performs local feature extraction and segmented coding on state information required by four functional modules of basic assist, return-to-zero, damping, and friction compensation through parallel full-connection sub-networks to form local state feature vectors of the modules; the policy network layer configures an improved Actor network for each of the four functional modules of basic assist, return-to-zero, damping, and friction compensation, each Actor network taking a corresponding local state feature vector as input and outputting a continuous adjustment amount of a calibration parameter of the functional module; the value network layer deploys a double Critic network for each functional module, which is used to output a state-action value function to evaluate the action output by the Actor network and calculate a policy gradient error; the parameter output layer limits and smooths the calibration parameter adjustment amounts output by the Actor networks, integrates them into a complete set of EPS calibration parameters, and writes them into the ECU through a calibration interface. 5.The AI intelligent-based EPS adaptive calibration method according to claim 2, characterized in that, The offline training of the parameter calibration model based on reinforcement learning according to the EPS calibration interval sequence sample data set comprises the following steps: from the historical vehicle collection information, corresponding sequences are extracted according to the labels of different functional modules and stored in corresponding experience replay buffers; in each training iteration, the Actor network and the Critic network of each functional module are trained independently in sequence; wherein the Actor network is updated by maximizing the expected return through a deterministic policy gradient ascent algorithm; the Critic network is updated by minimizing Bellman error formed by a time difference error; the target network parameters of each functional module are synchronized in a soft update manner until the reward functions of the modules converge to form an initial policy library. 6.The AI intelligent-based EPS adaptive calibration method according to claim 5, characterized in that, The offline training adopts a combined loss function, and the expression is: ； wherein, is the clip target for the PPO algorithm, is the mean squared error of the value function Critic, is the policy entropy, and c1, c2 are weight coefficients. 7.The AI intelligent-based EPS adaptive calibration method according to claim 5, characterized in that, Based on the preliminary adaptive calibration model, real-time acquisition of vehicle information collected by the vehicle is performed, and the parameter calibration model is trained by online reinforcement learning combined with driver voice feedback, and the EPS calibration parameters are dynamically adjusted and output to the ECU, including: During vehicle operation, the initial policy library obtained by offline training is used as a benchmark to enter the online enhancement stage; Real-time acquisition of vehicle information collected by the vehicle; Synchronous acquisition of driver voice feedback, and conversion of voice feedback to quantitative user satisfaction scores through a voice recognition module; An online objective function is constructed, which adds a user satisfaction reward item based on the user satisfaction score on the basis of the original module built-in reward; A small batch gradient descent and experience replay method is used to update the corresponding Actor-Critic network locally. 8.The AI intelligent-based EPS adaptive calibration method according to claim 7, characterized in that, The Actor-Critic network uses an improved deep deterministic policy gradient (DDPG) framework, including a policy network layer and a value network layer, where the policy network layer configures an independent Actor network for each functional module, and the value network layer deploys a double Critic network. 9.The AI intelligent-based EPS adaptive calibration method according to claim 7, characterized in that, The form of the online objective function is: ； wherein is the average reward for all possible driving conditions, is the base assistance module weight, is the returnability performance module weight, is the damping module weight, is the friction compensation module weight, is the base assistance sub-reward, is the returnability sub-reward, is the damping sub-reward, is the friction compensation sub-reward; Where each sub-reward is calculated in real time based on measurable signals, and the expression is as follows: ； where T is the steering wheel hand torque, is the linear residual of the torque angle, and AT is the difference between the left and right steering torques. ； wherein, is the post-return residual angle, is the post-return overshoot angle; ； wherein is the high frequency standard deviation of the torque, is the high frequency standard deviation of the motor current; ； wherein is the small angle yaw gain.

10. An AI intelligence-based EPS adaptive calibration device, characterized in that, Including: An acquisition unit is configured to acquire information collected by the vehicle and voice information of the driver, and to construct an EPS calibration sequence sample data set; A construction unit is configured to construct a parameter calibration model based on reinforcement learning; An offline model unit is configured to obtain a preliminary adaptive calibration model based on offline training of the parameter calibration model based on reinforcement learning according to the EPS calibration sequence sample data set; An online model unit is configured to, based on the preliminary adaptive calibration model, real-time acquisition of vehicle information collected by the vehicle, and combined with driver voice feedback, train the parameter calibration model by online reinforcement learning, and dynamically adjust and output the EPS calibration parameters to the ECU.

Images