User preference model training method, device and equipment

Abstract

The invention provides a user preference model training method, device and equipment, and is applied to the technical field of Internet of Vehicles. The method comprises the following steps: acquiring operation record data of a user on vehicle functions and vehicle state data matched with the operation record data at a vehicle end; and according to the operation record data and the vehicle state data, constructing an initial personalized model of the user at the vehicle end. And receiving global model parameters issued by the cloud, and fusing the global model parameters with personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle end. And predicting preferred vehicle function parameters of the user according to the current vehicle state data of the vehicle and the target personalized model, so as to provide personalized services for the user based on the preferred vehicle function parameters. The problems that in traditional model training, the data privacy leakage risk is high, the vehicle-mounted resource adaptability is poor, and the personalized modeling precision is insufficient can be solved.

Description

Technical Field

[0001] This application relates to the field of vehicle networking technology, and in particular to a user preference model training method, apparatus and device. Background Technology

[0002] With the development of intelligent vehicle technology, personalized settings for cabin environment, driving modes, and other functions via vehicle control apps have become crucial for enhancing user experience. Among related technologies, the mainstream solution for achieving personalized services relies on centralized cloud training: each vehicle terminal needs to upload raw data, including sensitive user operating habits and geographical location, to a cloud server for centralized training of a global recommendation model. While this architecture can utilize massive amounts of data to optimize model performance, it poses a significant risk of user privacy data leakage and does not comply with increasingly stringent data security regulations. Furthermore, in-vehicle terminal hardware suffers from resource constraints such as limited computing power, power sensitivity, and unstable network bandwidth, making it difficult to support complex local model training. In addition, the average model obtained through centralized training cannot accurately characterize the unique preferences of each user in different driving scenarios.

[0003] Therefore, there is an urgent need for a user preference model training method to achieve high-precision personalized user preference modeling for intelligent vehicle control systems, while ensuring user data privacy and security and without exceeding the limitations of in-vehicle hardware resources. Summary of the Invention

[0004] The purpose of this application is to provide a user preference model training method, apparatus, and device to solve the problems of high risk of data privacy leakage, poor adaptability to vehicle resources, and insufficient accuracy of personalized modeling in traditional model training.

[0005] Firstly, embodiments of this application provide a user preference model training method. This method includes acquiring user operation record data of vehicle functions and vehicle status data matching the operation record data on the vehicle side. Based on the operation record data and vehicle status data, an initial personalized model of the user is constructed on the vehicle side. The personalized model is used to learn the correlation between user operations and vehicle status. Global model parameters are received from the cloud and fused with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle side. The global model parameters are determined by the cloud based on personalized model parameters uploaded from multiple vehicle sides. Based on the current vehicle status data and the target personalized model, the user's preferred vehicle function parameters are predicted to provide personalized services to the user based on these parameters.

[0006] The user preference model training method provided in this application obtains user operation records of vehicle functions and vehicle status data matching these operation records on the vehicle side. Then, based on the operation record data and vehicle status data, an initial personalized model of the user is constructed. Next, comprehensive model parameters are received from the cloud, and these global model parameters are fused with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle. Finally, based on the current vehicle status data and the optimized target personalized model, the user's preferred vehicle function parameters are predicted. Based on these preferred vehicle function parameters, the current vehicle functions are implemented to provide personalized services to the user. This achieves a low-latency, high-precision personalized service closed loop in a resource-constrained in-vehicle environment, improving the accuracy, usability, and user experience of the personalized vehicle control system while ensuring user data privacy and in-vehicle resource security boundaries.

[0007] One possible implementation involves constructing an initial personalized model for the user on the vehicle based on operation log data and vehicle status data. This includes combining operation log data and vehicle status data into sample pairs and storing these sample pairs in a secure storage area on the vehicle. When the number of sample pairs in the secure storage area reaches a preset number, an initial personalized model is constructed based on this preset number of sample pairs, and the personalized model parameters corresponding to the initial personalized model are determined.

[0008] One possible implementation involves combining operation log data and vehicle status data into sample pairs, including: de-identifying the operation log data; and combining the de-identified operation log data with the vehicle status data into sample pairs.

[0009] One possible implementation, as provided in this application embodiment, is a user preference model training method that further includes: detecting whether the vehicle is in a turned-off state and whether the vehicle's remaining battery power is higher than a preset battery power. If so, in a preset low-power mode, the steps of constructing an initial personalized model based on a preset number of samples and determining the personalized model parameters corresponding to the initial personalized model are performed. The low-power mode is used to limit the resource utilization of the onboard processor.

[0010] One possible implementation involves fusing global model parameters with the personalized model parameters of an initial personalized model to obtain a target personalized model suitable for the current vehicle. This includes: loading the global model parameters into the corresponding structure of the initial personalized model to replace the personalized model parameters of the initial personalized model, thus obtaining a basic personalized model; acquiring current sample pairs of vehicles collected within the current preset time period; adjusting the basic personalized model using the current samples to obtain the model parameter increments of the basic personalized model; and superimposing the model parameter increments with the model parameters of the basic personalized model to obtain the target personalized model.

[0011] One possible implementation, the user preference model training method provided in this application embodiment, further includes: after constructing an initial personalized model of the user on the vehicle side, encrypting the personalized model parameters of the initial personalized model; and uploading the encrypted personalized model parameters, anonymized operation record data, and vehicle status data to the cloud.

[0012] One possible implementation involves predicting the user's preferred vehicle function parameters based on the vehicle's current state data and a target personalization model, in order to provide personalized services to the user based on these parameters. This includes: inputting the current vehicle state data into the target personalization model to obtain the predicted vehicle function parameters; generating function control commands for the corresponding vehicle functions based on the predicted parameters; and sending the function control commands to the vehicle's corresponding in-vehicle system to cause the in-vehicle system to execute the function control commands.

[0013] One possible implementation involves acquiring user operation logs of vehicle functions and vehicle status data matching the operation logs at the vehicle end. This includes: monitoring the data stream from the vehicle's bus; recording the control command as operation log data when a user triggers a control command for a corresponding vehicle function; and extracting at least one vehicle status signal synchronized with the control command from the vehicle bus data stream to obtain vehicle status data matching the operation logs.

[0014] Secondly, embodiments of this application provide a user preference model training device, the device comprising: an acquisition module, a construction module, a processing module, and a prediction model.

[0015] The acquisition module is used to acquire user operation records of vehicle functions and vehicle status data that match the operation records on the vehicle side.

[0016] The module builds an initial personalized model of the user on the vehicle based on operation log data and vehicle status data. This personalized model learns the correlation between user actions and vehicle status.

[0017] The processing module receives global model parameters from the cloud and merges them with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle. The global model parameters are determined by the cloud based on personalized model parameters uploaded from multiple vehicles.

[0018] The prediction module is used to predict the user's preferred vehicle function parameters based on the vehicle's current vehicle status data and the target personalization model, so as to provide personalized services to the user based on the preferred vehicle function parameters.

[0019] One possible implementation involves a module that, when building an initial personalized model for the user on the vehicle based on operation log data and vehicle status data, specifically combines the operation log data and vehicle status data into sample pairs and stores these sample pairs in a secure storage area on the vehicle. When the number of sample pairs in the secure storage area reaches a preset number, the initial personalized model is built based on this preset number of sample pairs, and the personalized model parameters corresponding to the initial personalized model are determined.

[0020] One possible implementation involves building a module that, when combining operation log data and vehicle status data into sample pairs, specifically performs anonymization processing on the operation log data. The anonymized operation log data is then combined with the vehicle status data into sample pairs.

[0021] In one possible implementation, the user preference model training device provided in this application embodiment is further used to detect whether the vehicle is in a turned-off state and whether the vehicle's remaining battery power is higher than a preset battery power. If so, in a preset low-power mode, the steps of constructing an initial personalized model based on a preset number of samples and determining the personalized model parameters corresponding to the initial personalized model are performed. The low-power mode is used to limit the resource utilization of the onboard processor.

[0022] One possible implementation involves a processing module that, when fusing global model parameters with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle, specifically loads the global model parameters into the corresponding structure of the initial personalized model to replace the personalized model parameters of the initial personalized model, thus obtaining the basic personalized model. This involves acquiring current sample pairs of vehicles collected within the current preset time period; adjusting the basic personalized model using the current samples to obtain the model parameter increments of the basic personalized model; and then superimposing the model parameter increments with the model parameters of the basic personalized model to obtain the target personalized model.

[0023] In one possible implementation, the user preference model training device provided in this application embodiment is further used to encrypt the personalized model parameters of the initial personalized model after building the user's initial personalized model on the vehicle side. The encrypted personalized model parameters, the anonymized operation record data, and the vehicle status data are then uploaded to the cloud.

[0024] One possible implementation involves a prediction module that, when predicting a user's preferred vehicle function parameters based on the vehicle's current state data and a target personalization model to provide personalized services, specifically inputs the current vehicle state data into the target personalization model to obtain the predicted vehicle function parameters. Based on these parameters, it generates corresponding vehicle function control commands. These commands are then sent to the vehicle's onboard system to execute them.

[0025] One possible implementation involves an acquisition module that, when acquiring user operation log data for vehicle functions and vehicle status data matching the operation log data at the vehicle end, specifically monitors the data stream from the vehicle's vehicle bus. When a user triggers a control command for a corresponding vehicle function, the control command is recorded as operation log data. From the vehicle bus data stream, at least one vehicle status signal synchronized with the control command is extracted to obtain vehicle status data matching the operation log data.

[0026] Thirdly, embodiments of this application provide a user preference model training device, which has the function of implementing a user preference model training method that achieves the first aspect or any possible implementation of the first aspect. This function can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above-described function.

[0027] Fourthly, embodiments of this application provide a computer-readable storage medium storing instructions that, when executed on a computer, enable the computer to perform the user preference model training method described in the first aspect or any possible implementation thereof.

[0028] Fifthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, enable the computer to execute the user preference model training method described in the first aspect or any possible implementation thereof.

[0029] The technical effects of any of the design methods in aspects two through five can be found in aspect one or in different possible implementations of aspect one, and will not be repeated here. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0031] Figure 1 A schematic diagram of the structure of a vehicle-mounted terminal provided in an embodiment of this application; Figure 2 A system architecture diagram of an in-vehicle control system provided in this application embodiment; Figure 3 A flowchart illustrating a user preference model training method provided in this application embodiment; Figure 4 A schematic diagram of a user preference model training device provided in this application embodiment; Figure 5 This is a system architecture diagram of a user preference model training system provided in an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0033] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0034] In related technologies, the mainstream solution for achieving personalized services relies on centralized cloud training: each vehicle terminal needs to upload raw data, including users' sensitive operating habits and geographical location, to a cloud server for centralized training of a global recommendation model. While this architecture can utilize massive amounts of data to optimize model performance, it poses a significant risk of user privacy data leakage and does not comply with increasingly stringent data security regulations. Furthermore, in-vehicle terminal hardware suffers from resource constraints such as limited computing power, power sensitivity, and unstable network bandwidth, making it difficult to support complex local model training. In addition, the average model obtained through centralized training is insufficient to accurately characterize the unique preferences of each user in different driving scenarios.

[0035] Specifically, the vast amounts of data generated during daily vehicle operation, such as navigation tracks accurate to the meter, timestamped driving behavior records, and even voice and image information collected through microphones or cameras, may contain highly sensitive personal information. Continuously transmitting this raw data to the cloud not only faces the risk of interception during transmission but also constitutes a massive data aggregation target in cloud storage and processing, easily triggering large-scale data breaches. While related technologies can employ data encryption for transmission, they cannot fundamentally change the fact that the raw data is outside the user's control.

[0036] Secondly, the vehicle's computing unit is designed with driving safety and core functions as primary priorities, thus its available computing resources (such as NPU / CPU computing power), memory capacity, and persistent storage space are strictly limited. Complex machine learning model training tasks are typically computationally intensive and memory-intensive; if deployed directly on the vehicle, they could easily compete for resources with critical processes such as vehicle instruments and driver assistance systems, impacting driving safety. Furthermore, vehicles primarily rely on battery power; high-intensity local computing significantly increases energy consumption, potentially preventing the vehicle from starting after prolonged parking due to battery depletion. Simultaneously, cellular network signals are unstable in scenarios such as tunnels and underground parking garages, making solutions relying on stable, high-bandwidth continuous data uploads impractical in real-world driving environments.

[0037] Furthermore, driving behavior is highly context-dependent and varies greatly from person to person. For example, the same user might prefer strong air conditioning during a hot summer commute, while prioritizing seat heating for short trips in winter. Centralized models learn common trends from massive amounts of data, aiming to reduce global average error. This results in predictive results that are often too general for specific user scenarios (such as a combination of actions at a specific time, on a specific route, or under specific weather conditions), failing to accurately reflect the user's true preferences at that particular moment. Consequently, automatic adjustment or recommendation functions often become biased, requiring manual user intervention and reducing the practical value of intelligent systems.

[0038] Based on this, embodiments of this application provide a user preference model training method, apparatus, and device. The method includes acquiring user operation record data of vehicle functions and vehicle status data matching the operation record data at the vehicle end. Based on the operation record data and vehicle status data, an initial personalized model of the user is constructed at the vehicle end. The personalized model is used to learn the correlation between user operations and vehicle status. Global model parameters are received from the cloud and fused with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle end. The global model parameters are determined by the cloud based on personalized model parameters uploaded from multiple vehicle ends. Based on the current vehicle status data and the target personalized model, the user's preferred vehicle function parameters are predicted to provide personalized services to the user based on these parameters.

[0039] The user preference model training method provided in this application obtains user operation records of vehicle functions and vehicle status data matching these operation records on the vehicle side. Then, based on the operation record data and vehicle status data, an initial personalized model of the user is constructed. Next, comprehensive model parameters are received from the cloud, and these global model parameters are fused with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle. Finally, based on the current vehicle status data and the optimized target personalized model, the user's preferred vehicle function parameters are predicted. Based on these preferred vehicle function parameters, the current vehicle functions are implemented to provide personalized services to the user. This achieves a low-latency, high-precision personalized service closed loop in a resource-constrained in-vehicle environment, improving the accuracy, usability, and user experience of the personalized vehicle control system while ensuring user data privacy and in-vehicle resource security boundaries.

[0040] The methods provided in the embodiments of this application will now be described in conjunction with the specific accompanying drawings.

[0041] On the one hand, embodiments of this application provide a vehicle-mounted terminal. For example... Figure 1 As shown, the vehicle terminal 100 may include a vehicle bus interface 101, a sensor 102, a user input interface 103, a processor 104, a memory 105, and a communication module 106.

[0042] The vehicle bus interface 101 can exchange data with the vehicle's internal network (such as the CAN bus, vehicle Ethernet, etc.). The vehicle bus interface 101 can be used to continuously monitor and parse the data stream on the bus, extracting vehicle status data reflecting the vehicle's operating status, such as vehicle speed, engine speed, ambient temperature sensor readings, air conditioning system status, seat position motor feedback signals, geolocation (GPS) information, and timestamps.

[0043] Sensor 102 can be deployed inside the vehicle cabin. Sensor 102 can be a collection of hardware used to directly or indirectly sense the presence and behavior of a user. Sensor 102 may include an in-cabin camera for capturing user gestures or facial orientation, non-contact sensors (such as millimeter-wave radar or infrared sensors) for detecting hand proximity, and a microphone array for recognizing voice commands.

[0044] User input interface 103 is a human-machine interface for users to directly operate vehicle functions. User input interface 103 may include: physical buttons and knobs installed on the steering wheel or center console; in-vehicle central control touch screen; and software controls integrated into the intelligent vehicle control mobile application (APP). When the user triggers an operation through these interfaces (such as clicking the "Air Conditioning 24℃" button on the screen or rotating the volume knob), corresponding operation log data will be generated and recorded. This data may include the function identifier and operation value of the operation.

[0045] The processor 104 is the computing core of the vehicle terminal 100 and can be an automotive-grade system-on-a-chip (SoC), such as a chip integrating a central processing unit (CPU), a graphics processing unit (GPU), and / or a neural network processing unit (NPU). The processor 104 can be used to run intelligent vehicle control applications and underlying system services; execute vehicle-side steps in the user preference model training method provided in this application embodiment, such as controlling data acquisition, building an initial personalized model in a local secure environment, encrypting and decrypting model parameters, and making predictions and generating control commands based on the final model.

[0046] The memory 105 includes volatile memory (such as RAM) and non-volatile storage (such as eMMC / UFS flash memory). A portion of the memory 105 can be designated as a secure storage area for securely storing anonymized operation log data and vehicle status data (combined into sample pairs) collected by the vehicle bus interface 101 and user input interface 103, as well as personalized model parameters generated during training. This area is access-isolated through hardware or software mechanisms to protect user data privacy.

[0047] The communication module 106 is a hardware module for the vehicle terminal 100 to communicate with an external network. The communication module 106 can be, for example, a cellular network module (4G / 5G), a Wi-Fi module, or a C-V2X module. The communication module 106 can be used to upload encrypted local model parameters to the server on the network side; and to receive aggregated and optimized global model parameters sent by the server from the network side.

[0048] On the one hand, embodiments of this application provide a vehicle-mounted control system. For example... Figure 2As shown, the vehicle control system 200 may include multiple vehicle terminals 100, multiple edge computing nodes 201, and a cloud server 202. Each edge computing node 201 can connect to multiple vehicle terminals 100. Each vehicle terminal 100 can also serve as an edge computing node 201.

[0049] The edge computing node 201 can serve as a data source and local computing unit, responsible for local processing of privacy data and initial model training. The edge computing node 201 can be deployed on server equipment in a regional data center (such as a city or a car manufacturer's service area). The edge computing node 201 can receive encrypted personalized model parameters uploaded from multiple in-vehicle terminals 100 within its jurisdiction; verify the integrity and legality of the parameters; execute the user preference model training method provided in this application embodiment to aggregate multiple local model parameters into one or more regional optimized model parameters; and forward the aggregated regional model parameters to the cloud server 202.

[0050] The cloud server 202 can be a server cluster deployed in a central data center, forming the core of the vehicle-to-everything (TSP) service platform. The cloud server 202 can manage federated learning tasks, such as: sending initial global model parameters to each in-vehicle terminal 100; receiving regional aggregated model parameters uploaded from each edge computing node 201; executing a global model optimization algorithm to further fuse and optimize model parameters from different regions, generating a new generation of global models; compressing and encrypting the optimized global model; and finally, sending the final global model parameters to each in-vehicle terminal 100 via edge computing nodes 201 or directly, to complete the collaborative updating and knowledge sharing of all participating vehicle models.

[0051] For example, as shown in Table 1, the specific requirements for each category of equipment in the vehicle control system 200 can be as shown in Table 1.

[0052] Table 1

[0053] It should be noted that the above Figure 1 The vehicle terminal 100 shown is as follows Figure 2 The vehicle control system 200 shown is merely an example illustrating the application scenario of the solution in this application, and is not intended to limit the application scenario of the solution in this application.

[0054] On the one hand, embodiments of this application provide a user preference model training method, which can be deployed by, for example, Figure 1 The vehicle terminal shown is as follows Figure 2 The vehicle control system 200 shown is executed. For example... Figure 3 As shown, the method may include the following steps.

[0055] S301 acquires user operation records of vehicle functions and vehicle status data that match the operation records on the vehicle side.

[0056] Among them, vehicle functions refer to various electronic function modules on the vehicle that can be controlled or adjusted by the user, such as the temperature, fan speed, and air direction settings of the air conditioning system, the heating, ventilation, massage, and position memory functions of the seats, the color and brightness of the ambient lighting, the volume and sound effect modes of the audio system, the selection of driving modes (such as economy, comfort, and sport), and the setting of navigation destinations, etc.

[0057] Operation log data is a record representing a user's specific control intent regarding one or more vehicle functions. Operation log data can include function identifiers and operation values. The function identifier uniquely identifies the operated vehicle function, such as "AC_Temperature_Setpoint". The operation value can be a user-defined target parameter, such as the temperature parameter 24, typically in degrees Celsius. Essentially, operation log data records when, which function, and what value the user adjusted.

[0058] Vehicle status data refers to a series of parameters that reflect the overall operating status and environmental information of a vehicle at a certain moment, such as real-time vehicle speed, engine speed, outside ambient temperature, current inside temperature, vehicle geographical location (GPS coordinates), time (including time and day of the week), weather information (such as whether it is raining), and the status of other vehicle systems (such as whether the windows are closed).

[0059] One possible implementation involves acquiring user operation logs of vehicle functions and vehicle status data matching the operation logs at the vehicle end. This includes: monitoring the data stream from the vehicle's bus; recording the control command as operation log data when a user triggers a control command for a corresponding vehicle function; and extracting at least one vehicle status signal synchronized with the control command from the vehicle bus data stream to obtain vehicle status data matching the operation logs.

[0060] Specifically, the system first continuously listens for and receives raw data frames transmitted on the bus at a preset sampling frequency (e.g., 10Hz or 100Hz). When a user triggers a control command for a corresponding vehicle function, the control command is recorded as operation log data. This process involves identifying commands generated by direct user interaction from the complex bus data stream. This can be achieved by monitoring data frames sent from specific ECU addresses that contain function setting values. For example, when a user rotates the air conditioning temperature knob, the corresponding Body Control Module (BCM) or Air Conditioning Control Module sends a data frame containing the newly set temperature value to the bus. When parsing the bus data, the data acquisition module identifies the identifier (ID) and data structure of the data frame to determine that it is a user operation. It then extracts the function identifier and operation value, adds a precise timestamp (accurate to milliseconds), forms a complete operation log data, and temporarily caches or stores it directly.

[0061] Finally, at least one vehicle status signal synchronized with the control command is extracted from the vehicle bus data stream to obtain vehicle status data matching the operation record data. At the same time the operation command is recorded, the data acquisition module needs to extract the vehicle status signal that is closest in time to the operation command from the historical or real-time bus data stream it is continuously monitoring. For example, this could be the most recent vehicle speed broadcast value or the ambient temperature sensor reading before the operation command was triggered. These status signals together constitute a vehicle status data vector.

[0062] This state vector is associated with previously recorded operation log data to form a sample pair. This sample pair identifies the user's action described by the operation log data when the vehicle is in the scenario described by the state vector. Through continuous data collection, the system accumulates a large number of such sample pairs locally on the vehicle, forming the original training dataset for the user-personalized model. All raw data undergoes necessary anonymization processing on the vehicle (such as removing precise GPS coordinates and generalizing to regions) and is stored in a secure storage area on the vehicle, preparing for subsequent local model training.

[0063] S302 builds an initial personalized model of the user on the vehicle based on operation record data and vehicle status data.

[0064] The personalized model is used to learn the relationship between user actions and vehicle status.

[0065] One possible implementation involves constructing an initial personalized model for the user on the vehicle based on operation log data and vehicle status data. This includes combining operation log data and vehicle status data into sample pairs and storing these sample pairs in a secure storage area on the vehicle. When the number of sample pairs in the secure storage area reaches a preset number, an initial personalized model is constructed based on this preset number of sample pairs, and the personalized model parameters corresponding to the initial personalized model are determined.

[0066] Specifically, each operation record (including function identifier and operation value) is first precisely bound to vehicle status data (such as vehicle speed, temperature, time, etc.) synchronized with the collection time. For example, a sample pair can be represented as: {Status: [Vehicle speed = 60km / h, Outside temperature = 32°C, Time = 14:30], Operation: {Function: "AC_Temperature", Value: 22°C}}. In this case, the sample pair records the user's action of setting the air conditioning temperature to 22°C at "2:30 pm, vehicle speed 60km / h, outside temperature 32°C".

[0067] Furthermore, the operation log data and vehicle status data are combined into sample pairs, including: de-identifying the operation log data; and then combining the de-identified operation log data with the vehicle status data into sample pairs.

[0068] Specifically, considering data privacy and security, the operation log data undergoes anonymization processing before combining sample pairs. This anonymization aims to remove or obscure sensitive information in the data that may directly or indirectly identify a specific individual. For example, precise GPS coordinates are replaced with the administrative division code of the city; or, when the operation log data contains raw text from voice commands, personally identifiable information such as names and specific addresses is removed. The anonymized operation log data is then combined with vehicle status data to form sample pairs. The combined sample pairs are stored in a secure storage area on the vehicle. This secure storage area is a dedicated space in the vehicle's onboard memory that is isolated and protected through hardware encryption or a secure operating system (such as a TEE, Trusted Execution Environment) mechanism, ensuring that the original training data cannot be accessed by unauthorized applications, thus protecting user privacy from both physical and logical perspectives.

[0069] Furthermore, the system checks whether the vehicle is off and whether its remaining battery power is higher than a preset level. If so, in a preset low-power mode, it executes steps to construct an initial personalized model based on a preset number of samples and determine the personalized model parameters corresponding to the initial personalized model. The low-power mode is used to limit the resource usage of the onboard processor.

[0070] Specifically, vehicle-side computing resources are limited, and it's necessary to avoid impacting normal vehicle operation. Therefore, model training is typically performed when the vehicle is idle. For example, the system detects whether the vehicle is off and whether its remaining battery power is higher than a preset threshold. The system continuously monitors the vehicle's power mode (e.g., IGN OFF) and the remaining battery power (SOC). If both conditions are met—the vehicle is off (indicating it's parked and not in motion) and the remaining battery power is higher than a preset safety threshold (e.g., above 30% to ensure training won't prevent the vehicle from starting)—then it's determined that the conditions for safely starting local training are met.

[0071] When the above conditions are met and the number of sample pairs accumulated in the secure storage area reaches a preset number (e.g., 500 or 1000 valid sample pairs have been accumulated to ensure that the model has enough data for learning), the system starts the training process. Training is performed in a preset low-power mode.

[0072] Low-power mode refers to the operating state controlled by the vehicle's operating system or power management module. Its core feature is limiting the resource utilization of the vehicle's processor (such as CPU / GPU / NPU). For example, the processor frequency is locked at a low level, training tasks are limited to using only a portion of the computing cores, and strict power limits are set (such as total power consumption not exceeding 5W) to ensure that the training process does not excessively consume battery power and affect the vehicle's next normal start-up.

[0073] In low-power mode, an initial personalized model is constructed based on a preset number of sample pairs, and the corresponding personalized model parameters are determined. The set of sample pairs is read from a secure storage area, using vehicle state data from the sample pairs as input features and operation values ​​(or encoded target values) from operation record data as the expected output. Then, a lightweight machine learning model structure (e.g., a small fully connected neural network or decision tree model) is selected, and a training algorithm (such as gradient descent) is run using the vehicle-side processor (especially its NPU module to improve efficiency). The model's internal weights and biases, i.e., the personalized model parameters, are iteratively adjusted to ensure that the model's predicted output for the training set samples is as close as possible to the actual user operation values. After training, an initial personalized model that can preliminarily characterize the user's behavioral habits and its corresponding personalized model parameters are produced locally on the vehicle.

[0074] Furthermore, after building the initial personalized model of the user on the vehicle side, the personalized model parameters of the initial personalized model are encrypted. The encrypted personalized model parameters, the anonymized operation record data, and the vehicle status data are then uploaded to the cloud.

[0075] Specifically, the vehicle-side system uses an encryption algorithm to encrypt the personalized model parameters. This encryption algorithm can be a standard symmetric encryption algorithm (such as AES-256) or an asymmetric encryption algorithm (such as one based on the SM2 national cryptographic algorithm or the RSA algorithm).

[0076] Subsequently, the encrypted personalized model parameters, along with the previously anonymized operation log data and vehicle status data (i.e., sample pairs used for model training), are packaged together and uploaded to the cloud server via the vehicle communication module. Uploading the anonymized original data (sample pairs) is optional; its main purpose is to provide the cloud with further analysis, model performance evaluation, or other compliance research that does not involve user identification. This data has already been anonymized before uploading, removing direct identifiers. Uploading the encrypted model parameters is the core of federated learning collaborative optimization. After receiving a large number of encrypted parameters from vehicles, the cloud server needs to perform a secure aggregation operation either by decryption (if symmetric encryption is used and the cloud possesses the key) or in its ciphertext form (if homomorphic encryption technology supporting ciphertext computation is used) to generate updated global model parameters.

[0077] Another possible implementation, in order to cope with more stringent privacy protection requirements (e.g., resisting reverse inference attacks targeting the model parameters themselves), is that the user preference model training method provided in this application embodiment can also, after completing local model training on the vehicle to obtain personalized model parameters, add random noise conforming to a specific distribution to the original parameters before uploading them to the cloud or edge node, instead of directly uploading the original parameters.

[0078] For example, the update amount (gradient) of the local model parameters is calculated, and the sensitivity (Δf) of this gradient update is determined, i.e., the maximum parameter change that a single sample data point may cause. Then, random noise following a Laplace distribution (Laplace(Δf / ε)) is generated, where ε is a preset privacy budget used to control the strength of privacy protection (a smaller ε value results in stronger privacy protection, but may introduce more noise, affecting model accuracy). A recommended value for ε is between 0.1 and 1.0. Finally, the noisy model parameters (or gradients) are encrypted and uploaded. The cloud or edge nodes aggregate the received noisy parameters to generate global model parameters protected by differential privacy.

[0079] The process only modifies the parameter uploading and aggregation stages, retaining the original overall architecture of local vehicle data collection, anonymized storage, model training, and data distribution. Experimental data shows that using the ε=0.5 scheme on specific vehicle models can significantly reduce the probability of user trajectory data leakage to an extremely low level (e.g., 0.0021%), while only causing a slight decrease in recommendation accuracy (e.g., 1.2%), achieving an effective balance between meeting stringent regulatory requirements such as GDPR and maintaining service usability.

[0080] Another possible implementation is that, when uploading to the cloud or edge nodes, multiple vehicle-side devices (as computing participants) can collaboratively train a shared, personalized model without the need for a trusted third party (i.e., without centralized aggregation by a cloud server).

[0081] Specifically, a secret sharing protocol is employed. Each vehicle divides its locally trained personalized model parameters into multiple secret shares (e.g., n shares) and distributes these shares to other participants (such as other vehicles or designated computing nodes). The protocol is designed so that complete parameter information can only be recovered when at least k shares are collected (k ≤ n), and a single or few shares cannot reveal any valid information. For nonlinear calculations involved in model training (such as the ReLU activation function in neural networks), secure computation in an encrypted state can be achieved by combining techniques such as obfuscated circuits. Each participant completes a round of collaborative training steps in the form of ciphertext or secret shares, updating their respective secret shares.

[0082] Ultimately, by exchanging and combining their shares, the parties obtain updated model parameters, without exposing any party's original data or intermediate parameters in plaintext to the others. This process completely avoids the role of cloud servers, eliminating single points of failure and single point of trust risks, making it particularly suitable for scenarios where multiple untrusted entities (such as different car manufacturers) need to perform joint modeling. Experiments show that this solution can achieve high accuracy in specific collaborative tasks (such as collision warning) while keeping additional communication latency within an acceptable range.

[0083] S303 receives global model parameters from the cloud and merges them with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle.

[0084] The global model parameters are determined by the cloud based on personalized model parameters uploaded from multiple vehicle terminals.

[0085] One possible implementation involves fusing global model parameters with the personalized model parameters of an initial personalized model to obtain a target personalized model suitable for the current vehicle. This includes: loading the global model parameters into the corresponding structure of the initial personalized model to replace the personalized model parameters of the initial personalized model, thus obtaining a basic personalized model; acquiring current sample pairs of vehicles collected within the current preset time period; adjusting the basic personalized model using the current samples to obtain the model parameter increments of the basic personalized model; and superimposing the model parameter increments with the model parameters of the basic personalized model to obtain the target personalized model.

[0086] Specifically, since the global model distributed from the cloud and the initial personalized model trained locally on the vehicle typically use the same model architecture (e.g., both are neural networks with the same number of layers and neurons), their parameter structures are aligned. The vehicle directly loads the received global model parameters (a set of weights and biases) into the computation graph of the local model, overwriting the original personalized model parameters that were trained based on only a small amount of data from the vehicle itself.

[0087] Then, to quickly adapt to the latest user behavior, the current sample pairs of vehicles collected within the current preset time period are obtained. Next, newly collected sample pairs that have not yet been used for training are extracted from the local secure storage area (e.g., the past week or since the last training).

[0088] Next, the base personalized model is adjusted using the current samples to obtain the parameter increments of the base personalized model. Starting with the base personalized model, the current sample pairs are used as training data to perform a small number of training epochs (e.g., 1-5 epochs) to calculate the small amount of parameter adjustment required to make the model more accurate in predicting these new samples.

[0089] Finally, the model parameter increments are superimposed with the model parameters of the basic personalized model to obtain the target personalized model. The calculated model parameter increments are then added to the original parameters of the basic personalized model using a vector addition operation to obtain the target personalized model parameters. These target personalized model parameters constitute the final target personalized model applicable to the current vehicle.

[0090] S304 predicts the user's preferred vehicle function parameters based on the vehicle's current vehicle status data and the target personalization model, in order to provide personalized services to the user based on the preferred vehicle function parameters.

[0091] Among them, the preferred vehicle function parameters refer to the specific settings of one or more vehicle functions that the user is most likely to expect, as predicted by the target personalization model for the current specific situation (such as air conditioning temperature, seat heating level, ambient lighting color, etc.).

[0092] One possible implementation involves predicting the user's preferred vehicle function parameters based on the vehicle's current state data and a target personalization model, in order to provide personalized services to the user based on these parameters. This includes: inputting the current vehicle state data into the target personalization model to obtain the predicted vehicle function parameters; generating function control commands for the corresponding vehicle functions based on the predicted parameters; and sending the function control commands to the vehicle's corresponding in-vehicle system to cause the in-vehicle system to execute the function control commands.

[0093] Specifically, when the vehicle is running, it continuously collects real-time vehicle status data through the vehicle bus interface, including current vehicle speed, interior / exterior temperature, time, geographical location, and weather conditions. This current vehicle status data is then used as input to the loaded target personalized model. This model performs calculations and inferences based on its internally learned "scene-preference" mapping. Its output is the parameter value predicted by the model for the user's possible preferences for specific vehicle functions in that input scene.

[0094] For example, the target personalized model might output: {"AC_Temperature": 23.5, "Seat_Heating_Level": 2}.

[0095] Subsequently, the abstract predicted vehicle function parameters output by the model are transformed into specific physical control commands that the vehicle's electronic control system can recognize and execute. Based on the function identifier and predicted value in the predicted vehicle function parameters, a preset instruction mapping table is queried or the corresponding control function is called to generate standard function control commands that conform to the vehicle's internal communication protocol (such as the CAN bus protocol). For example, for the predicted "air conditioning temperature 23.5°C", a complete CAN data frame containing the target address (air conditioning control module), instruction type (temperature setting), and set value (23.5) will be generated. Similarly, for the seat heating level, another corresponding control command is generated.

[0096] Finally, the generated function control command data frames are sent to the vehicle's internal network (such as the CAN bus) via the vehicle bus interface. The corresponding vehicle function control modules (such as the air conditioning control module and seat control module), upon receiving the commands on the bus, parse the command content and drive the actuators (such as the air conditioning compressor and heating elements) to operate, thereby automatically adjusting the vehicle functions to the state predicted by the model and conforming to user preferences. This process requires no active user intervention, achieving seamless personalized services with scene perception and intelligent pre-adjustment, enhancing the user experience. It ensures that the vehicle environment is automatically adjusted to a comfortable state when the user enters the vehicle or when the scene changes while driving, enhancing the convenience and attractiveness of the smart cockpit. Furthermore, because the model is a target personalized model optimized through federated learning, its predictions are generally more accurate than methods based on general rules or simple historical averaging, better reflecting the user's real habits and immediate needs.

[0097] The above primarily describes the solutions provided in the embodiments of this application from the perspective of the device's working principle. It is understood that, in order to achieve the above functions, the user preference model training device includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the algorithm steps of the examples described in the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0098] This application embodiment can divide the user preference model training device into functional modules according to the above method example. For example, each function can be divided into a separate functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module.

[0099] It should be noted that the module division in this embodiment is illustrative and represents only one logical functional division; in actual implementation, other division methods may be used. When dividing functional modules according to their respective functions, Figure 4 A schematic diagram illustrating a possible configuration of the user preference model training apparatus described above and in the embodiments is shown. Figure 4 As shown, the user preference model training device 400 may include: an acquisition module 401, a construction module 402, a processing module 403, and a prediction module 404.

[0100] The acquisition module 401 is used to support the execution of the user preference model training device 400. Figure 3 S301 in the illustrated user preference model training method.

[0101] Module 402 is used to support the execution of the user preference model training device 400. Figure 3 S302 in the illustrated user preference model training method.

[0102] Processing module 403 is used to support the execution of the user preference model training device 400. Figure 3 S303 is shown in the illustrated user preference model training method.

[0103] Prediction module 404 is used to support the execution of user preference model training device 400. Figure 3 S304 is shown in the illustrated user preference model training method.

[0104] One possible implementation involves a module that, when building an initial personalized model for the user on the vehicle based on operation log data and vehicle status data, specifically combines the operation log data and vehicle status data into sample pairs and stores these sample pairs in a secure storage area on the vehicle. When the number of sample pairs in the secure storage area reaches a preset number, the initial personalized model is built based on this preset number of sample pairs, and the personalized model parameters corresponding to the initial personalized model are determined.

[0105] One possible implementation involves building a module that, when combining operation log data and vehicle status data into sample pairs, specifically performs anonymization processing on the operation log data. The anonymized operation log data is then combined with the vehicle status data into sample pairs.

[0106] In one possible implementation, the user preference model training device provided in this application embodiment is further used to detect whether the vehicle is in a turned-off state and whether the vehicle's remaining battery power is higher than a preset battery power. If so, in a preset low-power mode, the steps of constructing an initial personalized model based on a preset number of samples and determining the personalized model parameters corresponding to the initial personalized model are performed. The low-power mode is used to limit the resource utilization of the onboard processor.

[0107] One possible implementation involves a processing module that, when fusing global model parameters with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle, specifically loads the global model parameters into the corresponding structure of the initial personalized model to replace the personalized model parameters of the initial personalized model, thus obtaining the basic personalized model. This involves acquiring current sample pairs of vehicles collected within the current preset time period; adjusting the basic personalized model using the current samples to obtain the model parameter increments of the basic personalized model; and then superimposing the model parameter increments with the model parameters of the basic personalized model to obtain the target personalized model.

[0108] In one possible implementation, the user preference model training device provided in this application embodiment is further used to encrypt the personalized model parameters of the initial personalized model after building the user's initial personalized model on the vehicle side. The encrypted personalized model parameters, the anonymized operation record data, and the vehicle status data are then uploaded to the cloud.

[0109] One possible implementation involves a prediction module that, when predicting a user's preferred vehicle function parameters based on the vehicle's current state data and a target personalization model to provide personalized services, specifically inputs the current vehicle state data into the target personalization model to obtain the predicted vehicle function parameters. Based on these parameters, it generates corresponding vehicle function control commands. These commands are then sent to the vehicle's onboard system to execute them.

[0110] One possible implementation involves an acquisition module that, when acquiring user operation log data for vehicle functions and vehicle status data matching the operation log data at the vehicle end, specifically monitors the data stream from the vehicle's vehicle bus. When a user triggers a control command for a corresponding vehicle function, the control command is recorded as operation log data. From the vehicle bus data stream, at least one vehicle status signal synchronized with the control command is extracted to obtain vehicle status data matching the operation log data.

[0111] It should be noted that all relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0112] The user preference model training device 400 provided in this application embodiment is used to perform the above-mentioned... Figure 3 The user preference model training method shown can achieve the same effect as the user preference model training method described above.

[0113] This application also provides a user preference model training device, which can execute the user preference model training method and related steps in the above method embodiments.

[0114] This application also provides a computer-readable storage medium storing instructions thereon, which, when executed, perform the user preference model training method and related steps described in the above method embodiments.

[0115] This application also provides a computer program product that, when run on a computer, causes the computer to execute the user preference model training method and related steps described in the above method embodiments.

[0116] In some embodiments, the methods shown in this application can be implemented as computer program instructions encoded in a machine-readable format on a computer-readable storage medium or on other non-transitory media or articles of art.

[0117] This application also provides a user preference model training system 500, such as... Figure 5 As shown, the user preference model training system 500 includes at least one processor 501 and at least one interface circuit 502.

[0118] As an example, when the user preference model training system 500 includes a processor and an interface circuit, the processor can be... Figure 5 The processor 501 shown in the solid box (or the processor 501 shown in the dashed box) can be an interface circuit. Figure 5 The interface circuit 502 is shown in the solid box (or the dashed box). When the user preference model training system 500 includes two processors and two interface circuits, the two processors include... Figure 5 The processor 501 shown in the solid box and the processor 501 shown in the dashed box, these two interface circuits include Figure 5 Interface circuit 502 is shown in both solid and dashed boxes. No limitations are imposed on this.

[0119] The processor 501 and the interface circuit 502 can be interconnected via a line. For example, the interface circuit 502 can be used to receive signals. Alternatively, the interface circuit 502 can be used to send signals to other devices (such as the processor 501). For instance, the interface circuit 502 can read computer instructions stored in memory and send those instructions to the processor 501. The processor 501 executes the instructions and, in conjunction with input / output devices, implements the various steps in the above embodiments, such as implementing... Figure 3The methods illustrated are the steps performed in the embodiments shown. Of course, this user preference model training system may also include other discrete components, and this application does not specifically limit this.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual 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.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0122] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units; that is, it can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] Furthermore, the functional units in the various embodiments of this application 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.

[0124] 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 readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to it, or all or part of the technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A user preference model training method, characterized in that, The method includes: The system acquires user operation records of vehicle functions and vehicle status data that matches the operation records in the vehicle. Based on the operation record data and the vehicle status data, an initial personalized model of the user is constructed on the vehicle side; the personalized model is used to learn the correlation between user operations and vehicle status. The system receives global model parameters from the cloud and merges them with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle. The global model parameters are determined by the cloud based on personalized model parameters uploaded from multiple vehicle terminals. Based on the vehicle's current vehicle status data and the target personalization model, the user's preferred vehicle function parameters are predicted, so as to provide personalized services to the user based on the preferred vehicle function parameters.

2. The method according to claim 1, characterized in that, The step of constructing the user's initial personalized model on the vehicle based on the operation record data and the vehicle status data includes: The operation record data and the vehicle status data are combined into a sample pair, and the sample pair is stored in the secure storage area of ​​the vehicle. When the number of sample pairs in the secure storage area reaches a preset number, the initial personalized model is constructed based on the preset number of sample pairs, and the personalized model parameters corresponding to the initial personalized model are determined.

3. The method according to claim 2, characterized in that, The step of combining the operation record data and the vehicle status data into a sample pair includes: The operation record data is anonymized; The anonymized operation record data and the vehicle status data are combined to form the sample pair.

4. The method according to claim 2, characterized in that, The method further includes: Detect whether the vehicle is in a turned-off state and whether the remaining battery power of the vehicle is higher than the preset battery power. If so, in a preset low-power mode, the steps of constructing the initial personalized model based on a preset number of samples and determining the personalized model parameters corresponding to the initial personalized model are executed; wherein, the low-power mode is used to limit the resource occupancy of the vehicle processor.

5. The method according to claim 1, characterized in that, The step of fusing the global model parameters with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle includes: The global model parameters are loaded into the corresponding structure of the initial personalized model to replace the personalized model parameters of the initial personalized model, thereby obtaining the basic personalized model. Obtain the current sample pairs of the vehicles collected within the current preset time period; The basic personalized model is adjusted using the current sample to obtain the model parameter increment of the basic personalized model; The incremental model parameters are superimposed with the model parameters of the basic personalized model to obtain the target personalized model.

6. The method according to claim 1, characterized in that, The method further includes: After constructing the user's initial personalized model on the vehicle side, the personalized model parameters of the initial personalized model are encrypted; The encrypted personalized model parameters, the anonymized operation record data, and the vehicle status data are uploaded to the cloud.

7. The method according to claim 1, characterized in that, The step of predicting the user's preferred vehicle function parameters based on the vehicle's current vehicle status data and the target personalization model, and providing personalized services to the user based on the preferred vehicle function parameters, includes: The current vehicle status data is input into the target personalized model to obtain the predicted vehicle function parameters of the vehicle; Based on the predicted vehicle function parameters, generate function control commands corresponding to the vehicle functions; The function control command is sent to the vehicle's corresponding in-vehicle system so that the in-vehicle system executes the function control command.

8. The method according to claim 1, characterized in that, The step of acquiring user operation record data for vehicle functions on the vehicle side, and vehicle status data matching the operation record data, includes: Listen to the data stream from the vehicle's bus; When a user triggers a control command for a corresponding vehicle function, the control command is recorded as the operation log data. From the data stream of the vehicle bus, at least one vehicle status signal that is time-synchronized with the control command is extracted to obtain vehicle status data that matches the operation record data.

9. A user preference model training device, characterized in that, The device includes: The acquisition module is used to acquire user operation record data of vehicle functions and vehicle status data that matches the operation record data at the vehicle end. The construction module is used to build an initial personalized model of the user on the vehicle based on the operation record data and the vehicle status data; the personalized model is used to learn the correlation between user operations and vehicle status. The processing module is used to receive global model parameters sent from the cloud and fuse the global model parameters with the personalized model parameters of the initial personalized model to obtain a target personalized model suitable for the current vehicle terminal; the global model parameters are determined by the cloud based on personalized model parameters uploaded from multiple vehicle terminals. The prediction module is used to predict the user's preferred vehicle function parameters based on the vehicle's current vehicle status data and the target personalization model, so as to provide personalized services to the user based on the preferred vehicle function parameters.

10. A user preference model training device, characterized in that, The user preference model training device includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, and the processor executing the machine-executable instructions to implement the user preference model training method according to any one of claims 1 to 8.

Images