In-vehicle display personalized control method and system based on deep learning

Abstract

The present invention provides a method and system for personalized control of an in-vehicle display based on deep learning, which relates to the technical field of in-vehicle display control, including obtaining multi-dimensional data of a driver's state, and adaptively determining personalized configuration parameters of an in-vehicle display interface; obtaining the driver's historical preference data from a cloud knowledge base, generating an in-vehicle display optimization strategy based on driving scenarios and driver preferences, and generating a personalized in-vehicle display interface according to the in-vehicle display optimization strategy; obtaining the driver's visual behavior and human-computer interaction behavior, as well as subjective and objective feedback data on the personalized in-vehicle display interface, and continuously optimizing a state detection model and personalized display rules using an adaptive optimization algorithm, wherein the adaptive optimization algorithm uses the driver's subjective and objective feedback data as reward and punishment signals, updates weight parameters of the state detection model and personalized rules of a display template library, and realizes adaptive evolution of the in-vehicle display system.

Description

Technical Field

[0001] The present invention relates to vehicle-mounted display control technology, and in particular to a vehicle-mounted display personalized control method and system based on deep learning. Background Art

[0002] With the rapid development of smart car technology, the in-vehicle display system, as an important interface for human-computer interaction, has a direct impact on driving experience and safety due to its personalization and intelligence level. Traditional in-vehicle display systems usually use fixed interface layouts and display modes, which are difficult to adapt to the personalized needs of different drivers and changing driving environments. Although some advanced systems have introduced limited personalized setting options, they still lack the ability to perceive the driver's status in real time and intelligently adjust it.

[0003] In the existing technology, some studies have tried to evaluate fatigue by monitoring the driver's physiological indicators and adjust the display interface accordingly. However, these methods often only consider the driver's state in a single dimension, ignoring important factors such as attention level and emotional state. In addition, some systems have introduced interface customization functions based on user preferences, but most of them are limited to static configuration and cannot be dynamically adjusted according to real-time driving scenarios.

[0004] In addition, existing in-vehicle display personalization methods generally lack effective adaptive optimization mechanisms. They usually rely on preset rules or limited user feedback for adjustment, which makes it difficult to adapt to the changes in drivers' long-term usage habits and complex and changing driving environments. At the same time, most systems fail to make full use of advanced technologies such as deep learning to improve personalization and intelligence.

[0005] Therefore, there is an urgent need for a personalized vehicle display control method based on deep learning, which can fully perceive the driver's status, integrate multi-dimensional data, realize intelligent and personalized dynamic adjustment of the display interface, and have continuous optimization capabilities to provide a safer and more comfortable driving experience. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for personalized control of an in-vehicle display based on deep learning, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] A method for personalized control of vehicle display based on deep learning is provided, including:

[0009] Acquire multi-dimensional data of the driver's state, and input the multi-dimensional data into a pre-trained state detection model to judge the driver's fatigue, attention level and emotional state, output a quantitative score of the driver's state, match the optimal display template from a preset display template library according to the quantitative score of the driver's state, wherein each display template contains a predefined personalized display rule, and adaptively determine personalized configuration parameters of the vehicle display interface according to the personalized display rule of the matched display template;

[0010] The driver's historical preference data is obtained from the cloud knowledge base, and the personalized configuration parameters of the vehicle display interface are optimized by using a collaborative filtering algorithm to generate personalized configuration parameters preferred by the driver. The vehicle driving status information and road traffic information are obtained in real time, and dynamically integrated with the personalized configuration parameters preferred by the driver to generate a vehicle display optimization strategy based on the driving scenario and the driver's preference. According to the vehicle display optimization strategy, the interface layout, display content, information hierarchy, color matching and interaction mode of the vehicle display interface are dynamically adjusted in real time to generate a personalized vehicle display interface;

[0011] The driver's visual behavior and human-computer interaction behavior, as well as subjective and objective feedback data on the personalized in-vehicle display interface, are obtained, and the driver's visual behavior, human-computer interaction behavior and subjective and objective feedback data are stored in a user portrait database, and an adaptive optimization algorithm is used to continuously optimize the state detection model and personalized display rules, wherein the adaptive optimization algorithm uses the driver's subjective and objective feedback data as reward and punishment signals, updates the weight parameters of the state detection model and the personalized rules of the display template library, and realizes the adaptive evolution of the in-vehicle display system.

[0012] In an optional embodiment,

[0013] The step of obtaining multi-dimensional data of the driver's state and inputting the multi-dimensional data into the pre-trained state detection model also includes:

[0014] Collect multi-dimensional data of driver status on the vehicle-mounted devices of multiple participants, and perform local preprocessing and feature extraction to obtain local feature data. Each vehicle-mounted device trains a state detection model based on the local feature data to generate local model parameters.

[0015] Using multiple encryption communication protocols to encrypt and sign the local model parameters to form encrypted model parameters, and uploading the encrypted model parameters to the cloud server;

[0016] The cloud server receives and decrypts and verifies the encrypted model parameters uploaded by each participant, obtains the decrypted local model parameters, and aggregates the decrypted local model parameters into global model parameters by introducing a secure aggregation algorithm with a differential privacy mechanism;

[0017] An asynchronous aggregation strategy is adopted. When a preset number of participants upload encrypted model parameters, global aggregation is triggered. After global aggregation is triggered, the cloud server calculates the time decay factor of each participant, and weights the decrypted local model parameters of each participant based on the time decay factor. At the same time, a secure aggregation algorithm with a differential privacy mechanism is introduced to aggregate the weighted local model parameters into global model parameters.

[0018] The global model parameters are encrypted and sent to each participant. Each participant receives and decrypts the global model parameters, replaces the local model parameters, and fine-tunes the replaced local model parameters based on the new local feature data to obtain the final state detection model.

[0019] In an optional embodiment,

[0020] According to the personalized display rules of the matched display template, the personalized configuration parameters of the vehicle display interface are adaptively determined, including:

[0021] Construct a multi-level domain knowledge graph, including three core entity types: driver status, environmental factors, and display elements, as well as multiple relationship edges representing complex relationships between entities, extract entities and relationships from vehicle system logs and driver feedback using natural language processing technology, and update the multi-level domain knowledge graph through a timestamp mechanism;

[0022] Encode the entities in the updated multi-level domain knowledge graph into low-dimensional dense vectors, and construct a hierarchical graph neural network, including bottom-level, middle-level, and top-level network structures, and use the hierarchical graph neural network to perform multi-hop reasoning on the encoded entities to obtain multi-hop reasoning results;

[0023] The multi-hop reasoning result and the quantitative score of the driver's state are used as feature vectors, input into a pre-trained multi-layer perceptron model, and initial display configuration parameters are generated. The quantitative score of the driver's state, environmental factor parameters, and initial display configuration parameters are input into a pre-trained random forest model to generate configuration rules for display elements. Based on the obtained configuration rules for display elements, combined with the personalized display rules in the matched display template and the initial display configuration parameters, final personalized display rules are dynamically generated, and based on the final personalized display rules, personalized configuration parameters of the vehicle display interface are adaptively determined.

[0024] In an optional embodiment,

[0025] Using the multi-hop reasoning result and the quantized score of the driver state as feature vectors, inputting a pre-trained multi-layer perceptron model to generate initial display configuration parameters, inputting the driver state quantized score, environmental factor parameters and initial display configuration parameters into a pre-trained random forest model, and generating configuration rules for display elements includes:

[0026] The quantitative scores of the driver's status are normalized, the environmental factor parameters are one-hot encoded, and the initial display configuration parameters are standardized to obtain preprocessed input data;

[0027] Using a permutation-based feature importance analysis algorithm to evaluate the preprocessed input data, select a target feature subset, and construct an integrated learning model including a random forest model, a gradient boosted decision tree model, and a support vector machine model;

[0028] Using a grid search combined with a cross-validation algorithm to optimize the hyperparameters of the ensemble learning model to obtain an optimized base model, using a stacking algorithm to combine the optimized base models to obtain an initial ensemble model, and designing a dynamic weight adjustment mechanism to collect display configuration effect evaluation information in actual driving scenarios, and regularly adjusting the weights of each base model in the initial ensemble model according to the display configuration effect evaluation information to obtain an adjusted ensemble model;

[0029] The preprocessed input data is input into the adjusted integrated model to generate preliminary display element configuration rules, a rule-based post-processing system is designed, the driver's historical preference data is input into the post-processing system, the preliminary display element configuration rules are personalized, and the final display element configuration rules are generated.

[0030] In an optional embodiment,

[0031] Acquire vehicle driving status information and road traffic information in real time, and dynamically integrate them with the personalized configuration parameters preferred by the driver, generate an in-vehicle display optimization strategy based on driving scenarios and driver preferences, and dynamically adjust the interface layout, display content, information hierarchy, color matching and interaction mode of the in-vehicle display interface in real time according to the in-vehicle display optimization strategy. Generate a personalized in-vehicle display interface including:

[0032] Acquire and share multi-source heterogeneous data including vehicle driving status information and road traffic information, represent and store the received multi-source heterogeneous data in a unified semantic level, build a multi-source information fusion framework, perceive the global dynamic traffic environment in real time, and obtain the fused global traffic status information;

[0033] Based on the integrated global traffic state information, a vehicle-to-vehicle interaction model and a vehicle-to-road interaction model are established. Each vehicle is regarded as an intelligent agent, and a multi-agent reinforcement learning framework is introduced. Through the collaboration between intelligent agents, a multi-objective optimization function is constructed. The objective function includes information transmission efficiency, driving safety, and cross-vehicle display coordination. Combined with the multi-objective optimization function, a distributed strategy generation algorithm is used to generate an in-vehicle display optimization strategy.

[0034] According to the generated in-vehicle display optimization strategy, the driving status information of each vehicle and the personalized preference parameters preset by the driver are obtained in real time, and the interface layout, display content, information hierarchy, color matching and interaction mode of the in-vehicle display interface are adjusted dynamically in real time;

[0035] Through cross-vehicle information interaction, the display content of different vehicles is synchronized, the display conflicts of adjacent vehicles are coordinated, the vehicle group display is collaboratively optimized according to the vehicle formation status and driving characteristics, and a personalized in-vehicle display interface that is adaptive to the scenario is generated.

[0036] In an optional embodiment,

[0037] Based on the integrated global traffic status information, a vehicle-to-vehicle interaction model and a vehicle-to-road interaction model are established. Each vehicle is regarded as an intelligent agent, and a multi-agent reinforcement learning framework is introduced. Through the collaboration between agents, a multi-objective optimization function is constructed. The objective function includes information transmission efficiency, driving safety, and cross-vehicle display coordination. Combined with the multi-objective optimization function, a distributed strategy generation algorithm is used to generate the vehicle display optimization strategy, including:

[0038] Based on the fused global traffic state information, the interaction relationship between vehicles is determined according to the preset distance threshold and angle threshold, and the vehicles are used as nodes of the first directed graph. Directed edges are connected between vehicle pairs that meet the interaction conditions to obtain a vehicle-vehicle interaction model.

[0039] The vehicles and road infrastructure are taken as nodes of the second directed graph, and directed edges are connected between relevant nodes according to the spatial relationship between the vehicles and the road infrastructure to obtain the vehicle-road interaction model;

[0040] Generate an agent state according to the first directed graph and the second directed graph, wherein the agent state includes a vehicle motion state, environmental perception information, and interaction information between the vehicle and road infrastructure;

[0041] According to the state of the intelligent agent, three optimization objectives are defined: information transmission efficiency, driving safety, and cross-vehicle display coordination. The information transmission efficiency is calculated by the delay and packet loss rate of display request transmission between vehicles. The driving safety is calculated by the influence of the display layout on the driver's attention allocation. The cross-vehicle display coordination is calculated by the difference of different vehicle interfaces. The three optimization objectives are combined to form a multi-objective optimization function. By solving the multi-objective optimization function, an initial display strategy is obtained.

[0042] Based on the initial display strategy, the distributed strategy generation algorithm is used to learn and optimize the display strategy in parallel according to the local state and neighbor information of the agent. The agent Q function is updated by introducing a penalty term representing the difference in neighbor strategies to obtain the optimized display strategy. According to the optimized display strategy, key features are extracted and mapped to the semantic space for strategy conflict detection, and the conflicting strategy pairs are identified using the graph matching algorithm.

[0043] For the detected conflicting strategy pairs, the strategy priority is determined by priority scoring. The conflicting strategies with different strategy priorities are coordinated by priority decision-making method. The conflicting strategies with the same strategy priority are coordinated by consensus negotiation method, and finally a consistent in-vehicle display optimization strategy is obtained.

[0044] In an optional embodiment,

[0045] The calculation formula for updating the agent Q function is as follows:

[0046] ;

[0047] Among them, Q * represents the updated Q function of agent i, Q(·) represents the Q function, represents the strategy adopted by agent i at the kth step, α represents the learning rate, represents the immediate reward value returned by the environment after the agent i takes action a in the kth step, γ represents the reward decay factor, represents the maximum Q value that can be obtained by taking the optimal strategy π' in the next state, β represents the weight of the penalty term for the difference in neighbor strategies, and N i represents the neighbor set of agent i, represents the Q value of the strategy of agent i’s neighbor j at the kth step.

[0048] According to a second aspect of the embodiments of the present invention,

[0049] Provided is a vehicle display personalized control system based on deep learning, including:

[0050] The first unit is used to obtain multi-dimensional data of the driver's state, and input the multi-dimensional data into a pre-trained state detection model to judge the driver's fatigue, attention level and emotional state, output a quantitative score of the driver's state, match the optimal display template from a preset display template library according to the quantitative score of the driver's state, wherein each display template contains a predefined personalized display rule, and adaptively determine the personalized configuration parameters of the vehicle display interface according to the personalized display rule of the matched display template;

[0051] The second unit is used to obtain the driver's historical preference data from the cloud knowledge base, optimize the personalized configuration parameters of the vehicle display interface by using the collaborative filtering algorithm, generate the personalized configuration parameters preferred by the driver, obtain the vehicle driving status information and road traffic information in real time, and dynamically integrate them with the personalized configuration parameters preferred by the driver to generate a vehicle display optimization strategy based on the driving scene and the driver's preference. According to the vehicle display optimization strategy, the interface layout, display content, information hierarchy, color matching and interaction mode of the vehicle display interface are dynamically adjusted in real time to generate a personalized vehicle display interface;

[0052] The third unit is used to obtain the driver's visual behavior and human-computer interaction behavior, as well as the subjective and objective feedback data on the personalized vehicle display interface, and store the driver's visual behavior, human-computer interaction behavior and subjective and objective feedback data in the user portrait database, and use an adaptive optimization algorithm to continuously optimize the state detection model and personalized display rules. The adaptive optimization algorithm uses the driver's subjective and objective feedback data as reward and punishment signals, updates the weight parameters of the state detection model and the personalized rules of the display template library, and realizes the adaptive evolution of the vehicle display system.

[0053] According to a third aspect of the embodiments of the present invention,

[0054] An electronic device is provided, comprising:

[0055] processor;

[0056] a memory for storing processor-executable instructions;

[0057] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0058] According to a fourth aspect of the embodiments of the present invention,

[0059] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0060] In this embodiment, the system can provide a customized display interface according to the driver's status, historical preferences and real-time feedback, greatly improving the user experience. The display interface can be adjusted in real time according to the driver's status, vehicle driving status and road traffic information to maintain the best display effect and interactive experience. By real-time monitoring of the driver's status and optimizing information presentation, the driver's visual burden and distraction are reduced, and driving safety is improved. The ability to learn and predict the driver's behavior pattern and actively adjust the display strategy reflects a high degree of intelligence. Through the adaptive optimization algorithm, the system can continuously improve the state detection model and personalized display rules to achieve continuous evolution. Combined with the driver's status, historical preferences, real-time behavior and environmental information, the effective integration and utilization of multi-dimensional data is achieved. Through continuous optimization and personalized adjustment, the driver's needs can be better met and user satisfaction can be improved. Personalized interface layout and interaction methods can improve the interaction efficiency between the driver and the vehicle system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a process of a method for personalized control of an in-vehicle display based on deep learning according to an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram of the structure of a vehicle-mounted display personalized control system based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0065] Figure 1 FIG. 1 is a flow chart of a method for personalized vehicle display control based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] S101. Acquire multi-dimensional data of the driver's status, and input the multi-dimensional data into a pre-trained status detection model to determine the driver's fatigue, attention level and emotional state, and output a quantitative score of the driver's status. According to the quantitative score of the driver's status, match the optimal display template from a preset display template library, wherein each display template contains predefined personalized display rules, and according to the personalized display rules of the matched display template, adaptively determine the personalized configuration parameters of the vehicle display interface.

[0067] First, the multi-dimensional data of the driver's status is collected through on-board sensors and cameras, including physiological signals (such as heart rate, breathing rate, skin galvanic response, etc.), facial expressions, eye tracking, head posture, voice semantics, etc. These multi-dimensional data can fully reflect the driver's real-time status.

[0068] The acquired multi-dimensional data is input into a pre-trained state detection model, which can be a multimodal fusion network based on deep learning, such as a long short-term memory network (LSTM), a convolutional neural network (CNN), an attention mechanism, etc. Through model reasoning, the driver's fatigue, attention level, and emotional state are judged, and the corresponding quantitative scores are output. For example, the fatigue score ranges from 0 to 100, and the higher the score, the more fatigued the driver is.

[0069] According to the quantitative score of the driver's state, the optimal display template is matched from the preset display template library. By calculating the similarity between the quantitative score of the driver's state and each display template, the display template with the highest similarity is selected as the optimal template under the current driver's state. The similarity calculation can adopt common methods such as Euclidean distance and cosine similarity. Each template in the display template library contains predefined personalized display rules. For example, when the driver's fatigue is higher than 80, a simple interface layout is adopted to reduce unnecessary information display, and voice is used as the main interaction method. By matching the quantitative score with the template rules, the optimal display template is selected.

[0070] In an optional embodiment,

[0071] The step of obtaining multi-dimensional data of the driver's state and inputting the multi-dimensional data into the pre-trained state detection model also includes:

[0072] Collect multi-dimensional data of driver status on the vehicle-mounted devices of multiple participants, and perform local preprocessing and feature extraction to obtain local feature data. Each vehicle-mounted device trains a state detection model based on the local feature data to generate local model parameters.

[0073] Using multiple encryption communication protocols to encrypt and sign the local model parameters to form encrypted model parameters, and uploading the encrypted model parameters to the cloud server;

[0074] The cloud server receives and decrypts and verifies the encrypted model parameters uploaded by each participant, obtains the decrypted local model parameters, and aggregates the decrypted local model parameters into global model parameters by introducing a secure aggregation algorithm with a differential privacy mechanism;

[0075] An asynchronous aggregation strategy is adopted. When a preset number of participants upload encrypted model parameters, global aggregation is triggered. After global aggregation is triggered, the cloud server calculates the time decay factor of each participant, and weights the decrypted local model parameters of each participant based on the time decay factor. At the same time, a secure aggregation algorithm with a differential privacy mechanism is introduced to aggregate the weighted local model parameters into global model parameters.

[0076] The global model parameters are encrypted and sent to each participant. Each participant receives and decrypts the global model parameters, replaces the local model parameters, and fine-tunes the replaced local model parameters based on the new local feature data to obtain the final state detection model.

[0077] Exemplarily, first, multi-dimensional data of the driver's status is collected on the in-vehicle devices of multiple participants. These data include visual data, physiological data, behavioral data, environmental data, and voice data. Visual data collects the driver's facial expressions, blinking frequency, and head posture through the in-vehicle camera. Physiological data monitors the driver's heart rate, skin conductivity, and body temperature in real time through embedded sensors on the steering wheel. Behavioral data records steering wheel operation, pedal use, and shifting mode. Environmental data includes in-vehicle temperature, humidity, and noise level. Voice data collects the driver's voice samples through the in-vehicle microphone system.

[0078] Next, the collected multi-dimensional data is locally preprocessed and features are extracted. Preprocessing includes data cleaning, time alignment, and feature extraction. Data cleaning uses a moving median filter to process physiological signals and a Kalman filter to smooth behavioral data. Time alignment uses a dynamic time warping algorithm to process data with different sampling rates. Feature extraction includes extracting facial feature points from visual data, calculating blink frequency and percentage of eye closure time; extracting heart rate variability features from physiological data; calculating the standard deviation of steering wheel angles and the frequency of pedal operations from behavioral data; and extracting acoustic features such as Mel-frequency cepstral coefficients and fundamental frequency from speech data. Finally, the extracted features are Z-score standardized to obtain local feature data.

[0079] Based on local feature data, each vehicle-mounted device trains a state detection model and generates local model parameters. The state detection model uses a deep neural network with a multi-task learning framework, including an input layer, a shared layer, a task-specific layer, and an output layer. During model training, batch normalization is used to accelerate training convergence, a cosine annealing learning rate strategy is used, L2 regularization and Dropout are applied to prevent overfitting, and an early stopping strategy is used to avoid overtraining.

[0080] Then, multiple encryption communication protocols are used to encrypt and sign the local model parameters. Specifically, a combination of homomorphic encryption and secure multi-party computation is used to ensure the privacy of model parameters during transmission and aggregation. The digital signature uses the elliptic curve digital signature algorithm (ECDSA) to ensure data integrity and non-repudiation. The encrypted model parameters are uploaded to the cloud server.

[0081] After receiving the encrypted model parameters, the cloud server first verifies the digital signature to ensure the authenticity and integrity of the data source. Then it uses the private key to decrypt and obtain the local model parameters. In order to improve the aggregation efficiency, an asynchronous aggregation strategy is adopted. When the preset number (such as 100) of participants upload the encrypted model parameters, global aggregation is triggered.

[0082] After triggering global aggregation, the cloud server calculates the time decay factor of each participant. The time decay factor is calculated based on the time interval between participants uploading model parameters, using an exponential decay function. For example, if a participant is the last to upload model parameters, its time decay factor is 1; if it is the first to upload, its time decay factor may be 0.8.

[0083] The decrypted local model parameters of each participant are weighted based on the time decay factor. During the weighting process, the local model parameters of each participant are multiplied by their corresponding time decay factor. This can reduce the impact of the earlier uploaded model parameters on the global model and improve the timeliness of the model.

[0084] During the aggregation process, a differential privacy mechanism is introduced to further protect the privacy of the participants. The specific implementation method is to add random noise that conforms to the Laplace distribution to the weighted local model parameters of each participant before aggregation. The size of the noise is determined by the preset privacy budget ε. A smaller ε provides stronger privacy protection, but may reduce model performance.

[0085] The weighted and noisy local model parameters are aggregated into global model parameters using a secure aggregation algorithm. The secure aggregation algorithm is based on a secret sharing scheme to ensure that the model parameters of any single participant are not leaked during the aggregation process.

[0086] After the aggregation is completed, the cloud server encrypts the global model parameters. The encryption method uses an asymmetric encryption algorithm, such as the RSA algorithm, and uses the public key of each participant for encryption. The encrypted global model parameters are sent to each participant.

[0087] After receiving the encrypted global model parameters, each participant uses its own private key to decrypt them. The decrypted global model parameters are used to replace the local model parameters. Subsequently, the replaced local model parameters are fine-tuned based on the newly collected local feature data. The fine-tuning process uses a smaller learning rate, such as 1 / 10 of the original learning rate, to retain the universality of the global model while adapting to the local data distribution.

[0088] After the above steps, all parties involved get the final state detection model. The model can simultaneously predict the driver's fatigue, attention level and emotional state, and output the corresponding quantitative scores. For example, fatigue can be expressed by a score of 0-100, with 0 indicating full wakefulness and 100 indicating extreme fatigue; attention level can be expressed by a score of 0-10, with 0 indicating complete distraction and 10 indicating high concentration; emotional state can be divided into three categories: positive, neutral and negative.

[0089] Finally, based on the quantitative scores output by the state detection model, the interface of the in-vehicle display system is dynamically adjusted. For example, when the driver's fatigue level is detected to be over 80 points, the system will automatically adjust the brightness and contrast of the display interface and may trigger a voice reminder; when the attention level is below 3 points, the system will simplify the display information and highlight important warning information; when the emotional state is negative, the system may play soothing music or adjust the interior lighting to improve the driver's emotional state.

[0090] In this embodiment, a federated learning framework is adopted, and the original data of each participant is always stored locally and does not need to be uploaded to the central server. Through multiple encrypted communication protocols, differential privacy mechanisms and secure aggregation algorithms, the data privacy and model parameter security of the participants are effectively protected. This enables different car companies, fleets or individual car owners to participate in model training together without leaking sensitive information. By aggregating the data training experience of multiple participants, the scale and diversity of the model's training data are significantly expanded. This helps to improve the generalization ability of the model so that it can adapt to different drivers, different models and different driving environments. By allocating most of the computing tasks to the local devices of each participant, the computing burden of the central server is reduced. At the same time, the asynchronous aggregation strategy avoids waiting for all participants to upload data, improving the overall efficiency of the system. This distributed computing method allows the system to be better expanded and supports the participation of more participants. By integrating multi-dimensional data such as vision, physiology, behavior, environment and voice, a comprehensive assessment of the driver's status is achieved. Compared with the detection method of a single data source, this solution can more accurately identify the driver's fatigue, attention level and emotional state, reducing the possibility of misjudgment. Through the innovative federated learning framework and multi-dimensional data fusion, the accuracy, real-time and adaptability of driver status detection are significantly improved while protecting privacy, providing strong technical support for improving road traffic safety.

[0091] In an optional embodiment,

[0092] According to the personalized display rules of the matched display template, the personalized configuration parameters of the vehicle display interface are adaptively determined, including:

[0093] Construct a multi-level domain knowledge graph, including three core entity types: driver status, environmental factors, and display elements, as well as multiple relationship edges representing complex relationships between entities, extract entities and relationships from vehicle system logs and driver feedback using natural language processing technology, and update the multi-level domain knowledge graph through a timestamp mechanism;

[0094] Encode the entities in the updated multi-level domain knowledge graph into low-dimensional dense vectors, and construct a hierarchical graph neural network, including bottom-level, middle-level, and top-level network structures, and use the hierarchical graph neural network to perform multi-hop reasoning on the encoded entities to obtain multi-hop reasoning results;

[0095] The multi-hop reasoning result and the quantitative score of the driver's state are used as feature vectors, input into a pre-trained multi-layer perceptron model, and initial display configuration parameters are generated. The quantitative score of the driver's state, environmental factor parameters, and initial display configuration parameters are input into a pre-trained random forest model to generate configuration rules for display elements. Based on the obtained configuration rules for display elements, combined with the personalized display rules in the matched display template and the initial display configuration parameters, final personalized display rules are dynamically generated, and based on the final personalized display rules, personalized configuration parameters of the vehicle display interface are adaptively determined.

[0096] Exemplarily, first, a multi-level domain knowledge graph is constructed. The knowledge graph contains three core entity types: driver status, environmental factors, and display elements. Driver status entities include fatigue, attention level, and emotional state; environmental factor entities include weather conditions, light intensity, road type, etc.; display element entities include font size, color scheme, information layout, etc. Entities are connected by multiple relationship edges to represent complex interactive relationships. For example, the "influence" relationship connects the driver status and display elements, and the "subject to" relationship connects environmental factors and display elements.

[0097] In order to keep the knowledge graph up to date, the system uses natural language processing technology to continuously extract entities and relationships from vehicle system logs and driver feedback. Specifically, a named entity recognition model is used to identify key entities in the text, such as "fatigue" and "direct sunlight", and a relationship extraction model is used to extract relationships between entities, such as "direct sunlight affects display brightness". Each extracted entity and relationship is timestamped and used to update the knowledge graph. When a new entity or relationship is identified, the system adds it to the knowledge graph; when new evidence appears for an existing relationship, the system updates its confidence score.

[0098] Next, the entities in the updated multi-level domain knowledge graph are encoded into low-dimensional dense vectors. The TransE algorithm is used for entity encoding, which maps both entities and relations into the same low-dimensional vector space. For example, the "fatigue" entity is encoded into a 128-dimensional vector.

[0099] Then, a hierarchical graph neural network is constructed, including the bottom layer, middle layer and top layer network structures. The bottom layer is responsible for processing basic entity features; the middle layer captures local interactions between entities; and the top layer integrates global information. In specific implementation, the bottom layer uses a graph convolutional network (GCN) to process local features of nodes; the middle layer uses a graph attention network (GAT) to learn the importance weights between nodes; and the top layer applies a graph pooling operation to summarize the full graph features.

[0100] The constructed hierarchical graph neural network is used to perform multi-hop reasoning on the encoded entities. The multi-hop reasoning process simulates the human reasoning chain and obtains deep-level related information by wandering in the knowledge graph multiple times. For example, starting from "high fatigue", it reaches "font size" through the "affect" relationship, and then reaches "light intensity" through the "subject to" relationship, forming a reasoning chain. The system performs weighted aggregation on reasoning chains of different lengths to obtain the final multi-hop reasoning result.

[0101] Next, the multi-hop reasoning results and the quantitative scores of the driver's state are concatenated as a feature vector. For example, the feature vector may contain a 128-dimensional vector obtained by multi-hop reasoning, as well as three scalar values ​​representing fatigue (0-100), attention level (0-10), and emotional state (-1 for negative, 0 for neutral, and 1 for positive).

[0102] The feature vector is input into the pre-trained multi-layer perceptron model to generate the initial display configuration parameters. The multi-layer perceptron model contains 3 hidden layers, each layer uses the ReLU activation function, and the last layer uses the Sigmoid function to map the output to between 0 and 1. The initial display configuration parameters include font size ratio, brightness adjustment coefficient, contrast adjustment coefficient, etc.

[0103] Subsequently, the quantitative scores of the driver's state, environmental factor parameters, and initial display configuration parameters are combined into a new feature vector and input into a pre-trained random forest model. The random forest model consists of 100 decision trees, each with a maximum depth of 10. The model outputs configuration rules for display elements, such as "IF fatigue > 80 AND light intensity < 50 THEN increase font size by 20% AND increase brightness by 30%".

[0104] Finally, based on the obtained display element configuration rules, combined with the personalized display rules in the display template obtained by matching and the initial display configuration parameters, the final personalized display rules are dynamically generated. The personalized display rules in the display template may contain user preference information, such as "keep a simple layout" or "highlight speed information". The system integrates these rules with the previously generated configuration rules, resolves possible conflicts, and forms the final personalized display rules.

[0105] Based on the final personalized display rules, the system adaptively determines the personalized configuration parameters of the vehicle display interface. These parameters may include specific font size (such as 12pt), background color (such as #F5F5F5), information layout method (such as two-column layout), etc. The system applies these parameters to the vehicle display interface to achieve dynamic adjustment of the interface.

[0106] In this embodiment, by integrating the driver's state, environmental factors and personal preferences, the refined personalized configuration of the vehicle display interface is realized, and the driving experience and comfort are improved. The system can respond to the driver's state and environmental changes in real time, dynamically adjust the display interface, and ensure the clarity, readability and timeliness of information. The continuous update mechanism of the multi-level domain knowledge graph enables the system to continuously learn and accumulate new knowledge, improving the long-term adaptability of the system. Using graph neural networks and multi-hop reasoning, the system can capture and utilize complex relationships between entities to provide more intelligent decision support. By timely adjusting the display interface to adapt to the driver's state, information interference is reduced, which helps to improve driving safety. Combined with knowledge graphs and rule generation, the decision-making process of the system has a certain degree of explainability, which is conducive to the debugging and optimization of the system. Through the intelligent and personalized display interface configuration, the user experience and safety of the vehicle information system are comprehensively improved, providing innovative solutions for the field of intelligent driving.

[0107] In an optional embodiment,

[0108] Using the multi-hop reasoning result and the quantized score of the driver state as feature vectors, inputting a pre-trained multi-layer perceptron model to generate initial display configuration parameters, inputting the driver state quantized score, environmental factor parameters and initial display configuration parameters into a pre-trained random forest model, and generating configuration rules for display elements includes:

[0109] The quantitative scores of the driver's status are normalized, the environmental factor parameters are one-hot encoded, and the initial display configuration parameters are standardized to obtain preprocessed input data;

[0110] Using a permutation-based feature importance analysis algorithm to evaluate the preprocessed input data, select a target feature subset, and construct an integrated learning model including a random forest model, a gradient boosted decision tree model, and a support vector machine model;

[0111] Using a grid search combined with a cross-validation algorithm to optimize the hyperparameters of the ensemble learning model to obtain an optimized base model, using a stacking algorithm to combine the optimized base models to obtain an initial ensemble model, and designing a dynamic weight adjustment mechanism to collect display configuration effect evaluation information in actual driving scenarios, and regularly adjusting the weights of each base model in the initial ensemble model according to the display configuration effect evaluation information to obtain an adjusted ensemble model;

[0112] The preprocessed input data is input into the adjusted integrated model to generate preliminary display element configuration rules, a rule-based post-processing system is designed, the driver's historical preference data is input into the post-processing system, the preliminary display element configuration rules are personalized, and the final display element configuration rules are generated.

[0113] Exemplarily, first, the input data is preprocessed. The quantitative scores of the driver's state include indicators such as fatigue, attention level and emotional state, and the scales of these indicators may be different. For example, fatigue may be a score of 0-100, and the attention level may be a level of 1-10. In order to make these indicators comparable, the minimum-maximum normalization method is used to uniformly map them to the range of 0-1. Specifically, for the original value x, the minimum value is min, the maximum value is max, and the normalized value x' is calculated as (x-min) / (max-min). For example, if the original value of fatigue is 75, the minimum value is 0, and the maximum value is 100, the normalized value is 0.75.

[0114] Environmental factor parameters usually include discrete category variables such as weather conditions, light intensity, and road type. These parameters are processed by one-hot encoding to convert each category into a binary vector. For example, the weather conditions may be sunny, cloudy, and rainy, which are encoded as [1, 0, 0], [0, 1, 0], and [0, 0, 1] respectively.

[0115] The initial display configuration parameters include continuous variables such as font size ratio, brightness adjustment coefficient, and contrast adjustment coefficient. These parameters are standardized so that their mean is 0 and their standard deviation is 1. Standardization can eliminate the dimensional differences between different parameters, which is beneficial for subsequent model training. For example, if the original value of the font size ratio is 1.2, the mean is 1.0, and the standard deviation is 0.2, the standardized value is 1.0.

[0116] Next, the preprocessed input data is evaluated using a permutation-based feature importance analysis algorithm. This algorithm evaluates the importance of a feature by randomly shuffling the value of a feature and observing the change in model performance. The specific steps are as follows: First, use all features to train a baseline random forest model and record its performance indicators (such as mean square error); then, randomly permute each feature, retrain the model and calculate the performance indicators; finally, compare the performance changes before and after the permutation. The greater the change, the more important the feature. Based on the analysis results, select the top 80% of the features in importance as the target feature subset.

[0117] Subsequently, an integrated learning model including a random forest model, a gradient boosted decision tree model, and a support vector machine model was constructed. The random forest model consists of multiple decision trees, each of which is trained independently, and the prediction results are finally synthesized by voting or averaging. The gradient boosted decision tree model is an iterative integration method, and each iteration generates a new decision tree to correct the errors of the previous model. The support vector machine model classifies data points by finding the best hyperplane. These three models have their own advantages, and they can complement each other through integration.

[0118] In order to optimize the performance of the ensemble learning model, the grid search combined with the cross-validation algorithm was used to optimize the model's hyperparameters. For the random forest model, the number of trees (such as 50 to 500, with a step size of 50) and the maximum depth of each tree (such as 5 to 20, with a step size of 5) were optimized; for the gradient boosting decision tree model, the learning rate (such as 0.01 to 0.1, with a step size of 0.01) and the maximum depth were optimized; for the support vector machine model, the kernel function type (such as linear kernel, RBF kernel) and the regularization parameter C (such as 0.1 to 10, with a logarithmic step size of 10) were optimized. The performance of each set of hyperparameters was evaluated using 5-fold cross-validation, and the hyperparameter combination with the best average performance was selected.

[0119] The optimized base models are combined into the initial ensemble model through a stacking algorithm. The stacking algorithm consists of two layers: the first layer is the optimized base model, and the second layer is a meta-learner (such as logistic regression) that learns how to best combine the predictions of the base models. Specifically, k-fold cross validation is first used to generate the predictions of the base models on the training set, and then these predictions are used as features to train the meta-learner.

[0120] In order to enable the model to adapt to changes in actual driving scenarios, a dynamic weight adjustment mechanism is designed. The system regularly collects information on the effect evaluation of the display configuration, including subjective feedback from the driver (such as comfort score) and objective indicators (such as gaze time, operation error rate). Based on this information, the prediction accuracy of each base model is calculated, and its weight in the integrated model is adjusted accordingly. For example, if the prediction accuracy of the random forest model in the recent period is 85%, the gradient boosting decision tree is 80%, and the support vector machine is 75%, their weights may be adjusted to 0.35, 0.33, and 0.32.

[0121] The preprocessed input data is fed into the tuned ensemble model to generate preliminary display element configuration rules. These rules may include "if fatigue is greater than 0.7 and light intensity is low, increase font size by 20%".

[0122] Finally, a rule-based post-processing system is designed to input the driver's historical preference data into the system to personalize the preliminary display element configuration rules. The post-processing system contains a set of predefined adjustment rules, such as "if the driver has historically preferred high contrast, then increase the contrast by 5% on the original basis." Through these rules, the system can meet the driver's personal preferences to the greatest extent while ensuring safety and readability.

[0123] In this embodiment, by integrating multiple machine learning models, the advantages of each model are fully utilized to improve the generation accuracy of display configuration rules. The dynamic weight adjustment mechanism enables the system to be continuously optimized according to the actual effect and adapt to the changes of different drivers and environments. The post-processing system takes into account the driver's historical preferences and realizes the refined customization of the display configuration. The rule-based post-processing system provides a certain degree of interpretability, which helps to understand and debug system behavior. The optimized model structure enables the system to quickly generate configuration rules to meet the needs of real-time adjustment. By comprehensively considering the driver's status and environmental factors, the system can minimize interference to the driver while ensuring that the information is clear and readable. A highly intelligent and personalized in-vehicle display interface configuration is achieved, providing strong support for improving driving safety and user experience.

[0124] S102. Obtain the driver's historical preference data from the cloud knowledge base, use the collaborative filtering algorithm to optimize the personalized configuration parameters of the vehicle display interface, generate the personalized configuration parameters preferred by the driver, obtain the vehicle driving status information and road traffic information in real time, and dynamically integrate them with the personalized configuration parameters preferred by the driver to generate a vehicle display optimization strategy based on the driving scenario and the driver's preference. According to the vehicle display optimization strategy, the interface layout, display content, information hierarchy, color matching and interaction method of the vehicle display interface are dynamically adjusted in real time to generate a personalized vehicle display interface.

[0125] First, the system obtains the driver's historical preference data from the cloud knowledge base. This data includes the driver's past choices and evaluations of various configurations of the vehicle display interface. For example, driver A prefers a light background during the day and a dark background at night; the font size is usually set to 120% of the default size; the information display prefers a simple mode, etc. The system uses these historical data as a basis to provide a basis for subsequent personalized optimization.

[0126] Next, the system uses collaborative filtering algorithms to optimize the personalized configuration parameters of the vehicle display interface. The core idea of ​​collaborative filtering algorithms is to predict the preferences of the current user by analyzing user groups with similar preferences. The specific implementation can adopt a user-based collaborative filtering method. The system first calculates the similarity between the current driver and other drivers, and methods such as cosine similarity or Pearson correlation coefficient can be used. For example, if driver A and driver B both like large fonts and high contrast, while driver C likes small fonts and low contrast, then the system will consider A and B to have a higher similarity.

[0127] Then, the system predicts and optimizes the configuration parameters of the current driver based on the preferences of the driver groups with high similarity. For example, if the driver group with high similarity to Driver A generally likes to highlight road conditions in the navigation interface, the system will take this feature into consideration even if Driver A has not explicitly expressed this preference before. In this way, the system generates a more comprehensive and accurate set of driver preference personalized configuration parameters.

[0128] In an optional embodiment,

[0129] Acquire vehicle driving status information and road traffic information in real time, and dynamically integrate them with the personalized configuration parameters preferred by the driver, generate an in-vehicle display optimization strategy based on driving scenarios and driver preferences, and dynamically adjust the interface layout, display content, information hierarchy, color matching and interaction mode of the in-vehicle display interface in real time according to the in-vehicle display optimization strategy. Generate a personalized in-vehicle display interface including:

[0130] Acquire and share multi-source heterogeneous data including vehicle driving status information and road traffic information, represent and store the received multi-source heterogeneous data in a unified semantic level, build a multi-source information fusion framework, perceive the global dynamic traffic environment in real time, and obtain the fused global traffic status information;

[0131] Based on the integrated global traffic state information, a vehicle-to-vehicle interaction model and a vehicle-to-road interaction model are established. Each vehicle is regarded as an intelligent agent, and a multi-agent reinforcement learning framework is introduced. Through the collaboration between intelligent agents, a multi-objective optimization function is constructed. The objective function includes information transmission efficiency, driving safety, and cross-vehicle display coordination. Combined with the multi-objective optimization function, a distributed strategy generation algorithm is used to generate an in-vehicle display optimization strategy.

[0132] According to the generated in-vehicle display optimization strategy, the driving status information of each vehicle and the personalized preference parameters preset by the driver are obtained in real time, and the interface layout, display content, information hierarchy, color matching and interaction mode of the in-vehicle display interface are adjusted dynamically in real time;

[0133] Through cross-vehicle information interaction, the display content of different vehicles is synchronized, the display conflicts of adjacent vehicles are coordinated, the vehicle group display is collaboratively optimized according to the vehicle formation status and driving characteristics, and a personalized in-vehicle display interface that is adaptive to the scenario is generated.

[0134] For example, first, the system acquires and shares multi-source heterogeneous data including vehicle driving status information and road traffic information. These data come from a wide range of sources, including on-board sensors, roadside units, traffic management systems, and the Internet. For example, vehicle driving status information can be obtained from the vehicle's CAN bus, including speed, acceleration, steering wheel angle, etc.; road traffic information can be obtained from the roadside unit through V2X communication, including traffic flow, signal light status, road construction information, etc.; weather information can be obtained in real time through the Internet API.

[0135] The system uniformly represents and stores the received multi-source heterogeneous data at the semantic level. This step uses a semantic ontology model to map data from different sources and formats into a unified semantic space. For example, speed data from different sources (such as GPS speed measurement and speed calculated by wheel speed) are uniformly represented as a standardized speed concept. This unified representation method facilitates subsequent data fusion and processing.

[0136] Next, the system builds a multi-source information fusion framework to perceive the global dynamic traffic environment in real time and obtain the fused global traffic status information. This framework adopts a hierarchical fusion structure, including data layer, feature layer and decision layer. In the data layer, algorithms such as Kalman filtering are used to initially fuse the raw data from different sensors; in the feature layer, deep learning methods are used to extract high-level features; in the decision layer, methods such as Bayesian reasoning are used to obtain the final fusion result. For example, by fusing on-board radar, camera and V2X communication data, the system can more accurately perceive the location and dynamics of surrounding vehicles.

[0137] Based on the fused global traffic status information, the system establishes vehicle-to-vehicle interaction models and vehicle-to-road interaction models. These models describe the interactions between vehicles and between vehicles and road infrastructure. For example, the vehicle-to-vehicle interaction model can describe the distance changes and relative speeds between adjacent vehicles; the vehicle-to-road interaction model can describe the relationship between vehicles and traffic lights, lane lines, etc.

[0138] The system treats each vehicle as an agent and introduces a multi-agent reinforcement learning framework. Each agent (i.e., each vehicle) has its own state space, action space, and reward function. The state space includes the vehicle's own state (such as speed, position) and the state of the surrounding environment; the action space includes various possible adjustments to the display interface; and the reward function reflects the effect of the adjustment.

[0139] Through the collaboration between agents, the system constructs a multi-objective optimization function. This function comprehensively considers multiple objectives such as information transmission efficiency, driving safety, and cross-vehicle display coordination. Information transmission efficiency can be measured by the driver's response time to the display content; driving safety can be evaluated by the degree of reduction of potential dangerous events; and cross-vehicle display coordination considers the consistency and complementarity of the display content of adjacent vehicles.

[0140] Combined with the multi-objective optimization function, the system uses a distributed strategy generation algorithm to generate an in-vehicle display optimization strategy. This distributed algorithm allows each agent (vehicle) to make decisions based on local information and limited global information, while coordinating overall behavior through communication between agents. For example, when a traffic accident is detected ahead, the system will generate a coordinated display strategy for all related vehicles to ensure effective dissemination of information and safe response.

[0141] Based on the generated in-vehicle display optimization strategy, the system obtains the driving status information of each vehicle and the personalized preference parameters preset by the driver in real time. These parameters may include the driver's settings for information display density, color preferences, font size, etc. The system combines these personalized parameters with the current traffic conditions and optimization strategy to dynamically adjust the in-vehicle display interface in real time.

[0142] Specifically, the system adjusts the layout, display content, information hierarchy, color matching and interaction mode of the vehicle display interface. For example, on highways, the system may increase the display ratio of navigation information and vehicle speed information and reduce the display of entertainment information; at complex intersections, the system may highlight turn instructions and surrounding vehicle location information; when driving at night, the system will automatically adjust to a dark theme to reduce glare.

[0143] Through cross-vehicle information interaction, the system synchronizes the display content of different vehicles and coordinates display conflicts of adjacent vehicles. For example, when two adjacent vehicles need to display important warning information, the system will coordinate their display methods to avoid information duplication or mutual interference. The system also collaboratively optimizes the display of vehicle groups based on the vehicle formation status and driving characteristics. When the convoy is driving, the leading vehicle may display more navigation and road conditions information, while the following vehicles may focus more on displaying information such as the distance and speed difference with the front vehicle.

[0144] Through this series of steps, the system finally generates a personalized in-vehicle display interface that is adaptive to the situation. This interface not only takes into account the preferences of individual drivers and the current driving environment, but also coordinates the information display of the entire vehicle group, achieving a higher level of intelligence and personalization.

[0145] For example, suppose there is a convoy of three cars driving on the highway. The driver of the lead car A prefers a simple interface, the driver of car B likes detailed information, and the driver of car C is sensitive to color. The system detects a traffic accident 5 kilometers ahead and generates a coordinated display strategy for the three cars through a multi-agent learning framework:

[0146] Car A: As the leading car, its display interface highlights the accident location, recommended detour route and estimated delay time. Although the driver prefers a simple interface, the system weighs safety factors and moderately increases the information density.

[0147] Vehicle B: The display contains more detailed accident information, such as accident type, estimated cleanup time, etc., while highlighting the distance and speed difference with the preceding vehicle A.

[0148] Vehicle C: Considering the driver’s sensitivity to color, a high-contrast color scheme is used to highlight the warning information while displaying the distance to the preceding vehicle B and the overall status of the current convoy.

[0149] In this embodiment, the efficiency and accuracy of information transmission are improved, so that the driver can obtain key information more quickly and accurately. Driving safety is enhanced by timely and effective transmission of dangerous information to help drivers react in advance. Information coordination across vehicles is achieved, information redundancy and conflict are avoided, and the information acquisition experience of the entire vehicle group is optimized. Under the premise of ensuring safety and efficiency, the display preferences of individual drivers are met to the greatest extent, and the user experience is improved. Through multi-agent learning, the system can continuously optimize and adapt to different traffic scenarios, and has strong scalability and adaptability.

[0150] In an optional embodiment,

[0151] Based on the integrated global traffic status information, a vehicle-to-vehicle interaction model and a vehicle-to-road interaction model are established. Each vehicle is regarded as an intelligent agent, and a multi-agent reinforcement learning framework is introduced. Through the collaboration between agents, a multi-objective optimization function is constructed. The objective function includes information transmission efficiency, driving safety, and cross-vehicle display coordination. Combined with the multi-objective optimization function, a distributed strategy generation algorithm is used to generate the vehicle display optimization strategy, including:

[0152] Based on the fused global traffic state information, the interaction relationship between vehicles is determined according to the preset distance threshold and angle threshold, and the vehicles are used as nodes of the first directed graph. Directed edges are connected between vehicle pairs that meet the interaction conditions to obtain a vehicle-vehicle interaction model.

[0153] The vehicles and road infrastructure are taken as nodes of the second directed graph, and directed edges are connected between relevant nodes according to the spatial relationship between the vehicles and the road infrastructure to obtain the vehicle-road interaction model;

[0154] Generate an agent state according to the first directed graph and the second directed graph, wherein the agent state includes a vehicle motion state, environmental perception information, and interaction information between the vehicle and road infrastructure;

[0155] According to the state of the intelligent agent, three optimization objectives are defined: information transmission efficiency, driving safety, and cross-vehicle display coordination. The information transmission efficiency is calculated by the delay and packet loss rate of display request transmission between vehicles. The driving safety is calculated by the influence of the display layout on the driver's attention allocation. The cross-vehicle display coordination is calculated by the difference of different vehicle interfaces. The three optimization objectives are combined to form a multi-objective optimization function. By solving the multi-objective optimization function, an initial display strategy is obtained.

[0156] Based on the initial display strategy, the distributed strategy generation algorithm is used to learn and optimize the display strategy in parallel according to the local state and neighbor information of the agent. The agent Q function is updated by introducing a penalty term representing the difference in neighbor strategies to obtain the optimized display strategy. According to the optimized display strategy, key features are extracted and mapped to the semantic space for strategy conflict detection, and the conflicting strategy pairs are identified using the graph matching algorithm.

[0157] For the detected conflicting strategy pairs, the strategy priority is determined by priority scoring. The conflicting strategies with different strategy priorities are coordinated by priority decision-making method. The conflicting strategies with the same strategy priority are coordinated by consensus negotiation method, and finally a consistent in-vehicle display optimization strategy is obtained.

[0158] Exemplarily, first, based on the fused global traffic state information, the system establishes a vehicle-to-vehicle interaction model. In this step, the system determines the interaction relationship between vehicles based on the preset distance threshold and angle threshold. Specifically, the system regards each vehicle as a node in the first directed graph. For any two vehicles A and B, if the distance between them is less than the preset distance threshold (for example, 50 meters), and vehicle B is located in the 120-degree fan-shaped area in front of vehicle A, it is considered that there is an interaction relationship between vehicle A and vehicle B. In this case, the system connects a directed edge between the nodes representing vehicle A and vehicle B, with the direction pointing from A to B. In this way, the system obtains the first directed graph describing the vehicle-to-vehicle interaction relationship, namely, the vehicle-to-vehicle interaction model.

[0159] For example, suppose there are three cars A, B, and C at a certain intersection. Car A is 20 meters behind car B, and car C is 30 meters to the left of car B. If the distance threshold is set to 50 meters and the angle threshold is set to 120 degrees, the system will connect a directed edge from A to B between A and B, and a directed edge from B to C between B and C.

[0160] Next, the system builds a vehicle-road interaction model. In this step, the system treats both vehicles and road infrastructure as nodes of the second directed graph. Road infrastructure includes fixed facilities such as traffic lights, road signs, speed bumps, and dynamic facilities such as variable information signs and intelligent roadside units. The system connects directed edges between related nodes based on the spatial relationship between vehicles and road infrastructure. For example, when a vehicle enters the control range of a traffic light, the system connects a directed edge between the node representing the vehicle and the node representing the traffic light, with the direction from the vehicle to the light.

[0161] Taking the above intersection as an example, suppose there is a traffic light D at the intersection. Then the system will connect directed edges between the nodes corresponding to the three vehicles A, B, and C and the node representing the traffic light D, with the direction pointing from the vehicle to the traffic light.

[0162] Based on the first directed graph (vehicle-vehicle interaction model) and the second directed graph (vehicle-road interaction model), the system generates the agent state. The agent state includes the vehicle motion state, environmental perception information, and the interaction information between the vehicle and the road infrastructure. The vehicle motion state includes speed, acceleration, direction, etc.; the environmental perception information includes the relative position and speed of surrounding vehicles, road conditions, etc.; the interaction information between the vehicle and the road infrastructure includes the distance to the traffic light, the current signal light status, etc.

[0163] For example, for vehicle B, its agent state may include: speed 40km / h, acceleration 0m / s², driving direction east; there is vehicle A 20 meters behind, and vehicle C 30 meters to the left front; it is 50 meters away from traffic light D, and the current traffic light state is green.

[0164] According to the state of the agent, the system defines three optimization objectives: information transmission efficiency, driving safety, and cross-vehicle display coordination, and combines them to form a multi-objective optimization function. Information transmission efficiency is calculated by the delay and packet loss rate of display requests transmitted between vehicles. For example, if vehicle A sends a display request to vehicle B, the system records the time from sending to receiving and whether it is successfully received, thereby evaluating the efficiency of information transmission. Driving safety is calculated by the impact of display layout on driver attention allocation. The system can analyze the driver's line of sight distribution through eye tracking technology and evaluate the impact of different display layouts on driving safety. Cross-vehicle display coordination is calculated by the difference between different vehicle interfaces. The system compares the display content and layout of adjacent vehicles. The smaller the difference, the higher the coordination.

[0165] By solving the multi-objective optimization function, the system obtains an initial display strategy, which provides a basic display solution for each vehicle, including interface layout, information priority, etc.

[0166] Next, the system optimizes the initial display policy using a distributed policy generation algorithm. Each agent (i.e., each vehicle) learns and optimizes the display policy in parallel based on its local state and neighbor information. The system updates the agent's Q function by introducing a penalty term that represents the difference in neighbor strategies, thereby obtaining the optimized display policy. This approach allows each vehicle to independently improve its own display strategy while taking into account the surrounding environment.

[0167] For example, if vehicle B finds that its display strategy is significantly different from that of vehicles A and C, the system will adjust B's strategy to make it more coordinated. The adjustment may include changing the display location of the information, adjusting the priority of the information, etc.

[0168] After obtaining the optimized display strategy, the system extracts key features and maps them into semantic space for strategy conflict detection. The system uses a graph matching algorithm to identify conflicting strategy pairs. For example, if both vehicles A and B plan to display important warning information in the upper right corner of the screen, the system will identify this as a potential conflict.

[0169] For the conflicting policy pairs detected, the system determines the policy priority through priority scoring. Priority scoring may consider factors such as the urgency of the information and security relevance. For conflicting policies with different priorities, the system coordinates them using priority decision-making, that is, retaining the policy with higher priority and adjusting the policy with lower priority. For conflicting policies with the same priority, the system coordinates them using consensus negotiation, which may combine the two policies or choose a compromise solution.

[0170] For example, if vehicles A and B both intend to display information about the road conditions ahead, but A's information is updated more promptly, the system may prioritize A's display strategy and adjust B's display content to other important information.

[0171] Through this series of steps, the system finally obtains a consistent in-vehicle display optimization strategy. This strategy not only takes into account the needs of a single vehicle, but also ensures the coordination and consistency of the display of the entire vehicle group.

[0172] In an optional embodiment,

[0173] The calculation formula for updating the agent Q function is as follows:

[0174] ;

[0175] Among them, Q * represents the updated Q function of agent i, Q(·) represents the Q function, represents the strategy adopted by agent i at the kth step, α represents the learning rate, represents the immediate reward value returned by the environment after the agent i takes action a in the kth step, γ represents the reward decay factor, represents the maximum Q value that can be obtained by taking the optimal strategy π' in the next state, β represents the weight of the penalty term for the difference in neighbor strategies, and N i represents the neighbor set of agent i, represents the Q value of the strategy of agent i’s neighbor j at the kth step.

[0176] In this embodiment, the intelligence level of the vehicle information system is improved, and the display strategy can be dynamically adjusted according to the complex traffic environment. The information coordination between vehicles is enhanced, the information redundancy and conflict are reduced, and the overall information transmission efficiency is improved. By considering driving safety, the display layout is optimized, and the unnecessary distraction of the driver's attention is reduced. The coordination and consistency of the cross-vehicle display is achieved, and the driver's overall perception of the surrounding environment is improved. The distributed algorithm is adopted to improve the scalability and robustness of the system and adapt to large-scale vehicle networking scenarios. The real-time and adaptability of the display strategy are ensured through dynamic adjustment and conflict resolution mechanisms.

[0177] S103. Obtain the driver's visual behavior and human-computer interaction behavior, as well as the subjective and objective feedback data on the personalized vehicle-mounted display interface, and store the driver's visual behavior, human-computer interaction behavior and subjective and objective feedback data in a user portrait database, and use an adaptive optimization algorithm to continuously optimize the state detection model and personalized display rules, wherein the adaptive optimization algorithm uses the driver's subjective and objective feedback data as reward and punishment signals, updates the weight parameters of the state detection model and the personalized rules of the display template library, and realizes the adaptive evolution of the vehicle-mounted display system.

[0178] Exemplarily, first, the system obtains the driver's visual behavior data. This step is achieved through an eye tracking device installed in the car. The eye tracking device usually includes an infrared camera and a processing unit, which can capture the driver's eye movements, pupil size changes and other information in real time. The system records data such as the driver's gaze position, duration, and scanning path. For example, the system may record that the driver switched his gaze between the center console display, the road ahead, and the rearview mirror 5 times in 10 seconds, and the average time for each gaze at the display was 0.8 seconds.

[0179] Next, the system acquires the driver's human-computer interaction behavior data. This includes all interactions between the driver and the in-vehicle infotainment system, such as touching the screen, pressing buttons, voice commands, etc. The system records the type, time, duration, and results of each interaction. For example, the system may record that the driver adjusted the air conditioning through the voice command "turn up the temperature" while driving, and the entire interaction process took 3 seconds and the operation was successful.

[0180] At the same time, the system collects subjective and objective feedback data from drivers on the personalized in-vehicle display interface. Subjective feedback is collected through short questionnaires or voice evaluations that pop up regularly, including satisfaction scores on interface layout, information presentation, and ease of operation. Objective feedback is automatically obtained by analyzing indicators such as the driver's frequency of use, operation speed, and error rate. For example, a driver may give a satisfaction score of 4 points (out of 5 points) for the interface layout, and the system also records that after the driver uses the new interface, the average time to complete common operations has been reduced by 15%.

[0181] The system stores the collected driver visual behavior, human-machine interaction behavior, and subjective and objective feedback data in the user portrait database. The user portrait database establishes a unique behavior model for each driver, including habitual sight allocation mode, preferred interaction method, attention to different types of information, etc. These data will be used for subsequent personalized optimization.

[0182] Based on the stored user data, the system uses an adaptive optimization algorithm to continuously optimize the state detection model and personalized display rules. The adaptive optimization algorithm uses the driver's subjective and objective feedback data as reward and punishment signals to update the weight parameters of the state detection model and the personalized rules of the display template library.

[0183] Specifically, for the optimization of the state detection model, the system uses the driver's feedback as a supervisory signal to adjust the weights of each feature in the model. For example, if the system finds that a driver often dims the display brightness when driving at night, it will increase the weight of the "ambient light" feature in the night mode judgment. In this way, the state detection model gradually adapts to each driver's personal preferences and habits.

[0184] For the optimization of personalized display rules, the system dynamically adjusts the priority of each rule in the display template library according to the driver's usage frequency and satisfaction score. For example, if a driver is found to frequently use the navigation function and highly appreciates the large-size display of navigation information, the system will increase the weight of the rule "navigation information priority display". At the same time, the system will also optimize the presentation position of information based on the driver's visual behavior data, placing important information in the area where the driver habitually looks.

[0185] The adaptive optimization process is ongoing. The system analyzes the new data collected regularly (such as weekly or monthly) and updates the state detection model and display rules. This ensures that the in-vehicle display system can continue to evolve as the driver's usage habits change.

[0186] For example, suppose that in the initial system settings, "driving information" and "entertainment information" have equal display priority. After one month of use, the system finds that a driver spends 80% of his screen gaze time on driving information while driving on the highway, and subjectively expresses that he wants driving information to be more eye-catching. Based on this data, the system will significantly increase the display priority of driving information when driving at high speeds, perhaps placing it in the center of the screen and increasing the font size.

[0187] In this embodiment, the system can provide a customized display interface based on each driver's unique usage habits and preferences, greatly improving the user experience. As the driver's habits change, the system can automatically adjust to always maintain the best display effect and interactive experience. By optimizing the presentation method and location of information, the driver's visual burden and distraction are reduced, and driving safety is improved. The continuous optimization process ensures that the system can respond to user feedback in a timely manner and continuously improve user satisfaction. The ability to learn and predict driver behavior patterns and actively adjust display strategies reflects a high degree of intelligence. Through the accumulation and analysis of large amounts of user data, the system can continuously improve its algorithms and rules to achieve continuous improvement in overall performance.

[0188] Figure 2 Schematic diagram of the structure of the in-vehicle display personalized control system based on deep learning according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0189] The first unit is used to obtain multi-dimensional data of the driver's state, and input the multi-dimensional data into a pre-trained state detection model to judge the driver's fatigue, attention level and emotional state, output a quantitative score of the driver's state, match the optimal display template from a preset display template library according to the quantitative score of the driver's state, wherein each display template contains a predefined personalized display rule, and adaptively determine the personalized configuration parameters of the vehicle display interface according to the personalized display rule of the matched display template;

[0190] The second unit is used to obtain the driver's historical preference data from the cloud knowledge base, optimize the personalized configuration parameters of the vehicle display interface by using the collaborative filtering algorithm, generate the personalized configuration parameters preferred by the driver, obtain the vehicle driving status information and road traffic information in real time, and dynamically integrate them with the personalized configuration parameters preferred by the driver to generate a vehicle display optimization strategy based on the driving scene and the driver's preference. According to the vehicle display optimization strategy, the interface layout, display content, information hierarchy, color matching and interaction mode of the vehicle display interface are dynamically adjusted in real time to generate a personalized vehicle display interface;

[0191] The third unit is used to obtain the driver's visual behavior and human-computer interaction behavior, as well as the subjective and objective feedback data on the personalized vehicle display interface, and store the driver's visual behavior, human-computer interaction behavior and subjective and objective feedback data in the user portrait database, and use an adaptive optimization algorithm to continuously optimize the state detection model and personalized display rules. The adaptive optimization algorithm uses the driver's subjective and objective feedback data as reward and punishment signals, updates the weight parameters of the state detection model and the personalized rules of the display template library, and realizes the adaptive evolution of the vehicle display system.

[0192] According to a third aspect of the embodiments of the present invention,

[0193] An electronic device is provided, comprising:

[0194] processor;

[0195] a memory for storing processor-executable instructions;

[0196] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0197] According to a fourth aspect of the embodiments of the present invention,

[0198] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0199] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0200] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for personalized control of vehicle display based on deep learning, characterized in that: include: Acquire multi-dimensional data of the driver's state, and input the multi-dimensional data into a pre-trained state detection model to judge the driver's fatigue, attention level and emotional state, output a quantitative score of the driver's state, match the optimal display template from a preset display template library according to the quantitative score of the driver's state, wherein each display template contains a predefined personalized display rule, and adaptively determine personalized configuration parameters of the vehicle display interface according to the personalized display rule of the matched display template; The driver's historical preference data is obtained from the cloud knowledge base, and the personalized configuration parameters of the vehicle display interface are optimized by using a collaborative filtering algorithm to generate personalized configuration parameters preferred by the driver. The vehicle driving status information and road traffic information are obtained in real time, and dynamically integrated with the personalized configuration parameters preferred by the driver to generate a vehicle display optimization strategy based on the driving scenario and the driver's preference. According to the vehicle display optimization strategy, the interface layout, display content, information hierarchy, color matching and interaction mode of the vehicle display interface are dynamically adjusted in real time to generate a personalized vehicle display interface; Acquire the driver's visual behavior and human-computer interaction behavior, as well as subjective and objective feedback data on the personalized in-vehicle display interface, and store the driver's visual behavior, human-computer interaction behavior and subjective and objective feedback data in a user portrait database, and continuously optimize the state detection model and personalized display rules using an adaptive optimization algorithm, wherein the adaptive optimization algorithm uses the driver's subjective and objective feedback data as a reward and punishment signal, updates the weight parameters of the state detection model and the personalized rules of the display template library, and realizes the adaptive evolution of the in-vehicle display system; According to the personalized display rules of the matched display template, the personalized configuration parameters of the vehicle display interface are adaptively determined, including: Construct a multi-level domain knowledge graph, including three core entity types: driver status, environmental factors, and display elements, as well as multiple relationship edges representing complex relationships between entities, extract entities and relationships from vehicle system logs and driver feedback using natural language processing technology, and update the multi-level domain knowledge graph through a timestamp mechanism; Encode the entities in the updated multi-level domain knowledge graph into low-dimensional dense vectors, and construct a hierarchical graph neural network, including bottom-level, middle-level, and top-level network structures, and use the hierarchical graph neural network to perform multi-hop reasoning on the encoded entities to obtain multi-hop reasoning results; Using the multi-hop reasoning result and the quantified score of the driver's state as feature vectors, inputting a pre-trained multi-layer perceptron model to generate initial display configuration parameters, inputting the quantified score of the driver's state, environmental factor parameters and the initial display configuration parameters into a pre-trained random forest model to generate configuration rules for display elements, dynamically generating final personalized display rules based on the obtained configuration rules for display elements, combining the personalized display rules in the display template obtained by matching and the initial display configuration parameters, and adaptively determining personalized configuration parameters of the vehicle display interface according to the final personalized display rules; Acquire vehicle driving status information and road traffic information in real time, and dynamically integrate them with the personalized configuration parameters preferred by the driver, generate an in-vehicle display optimization strategy based on driving scenarios and driver preferences, and dynamically adjust the interface layout, display content, information hierarchy, color matching and interaction mode of the in-vehicle display interface in real time according to the in-vehicle display optimization strategy. Generate a personalized in-vehicle display interface including: Acquire and share multi-source heterogeneous data including vehicle driving status information and road traffic information, represent and store the received multi-source heterogeneous data in a unified semantic level, build a multi-source information fusion framework, perceive the global dynamic traffic environment in real time, and obtain the fused global traffic status information; Based on the integrated global traffic state information, a vehicle-to-vehicle interaction model and a vehicle-to-road interaction model are established. Each vehicle is regarded as an intelligent agent, and a multi-agent reinforcement learning framework is introduced. Through the collaboration between intelligent agents, a multi-objective optimization function is constructed. The objective function includes information transmission efficiency, driving safety, and cross-vehicle display coordination. Combined with the multi-objective optimization function, a distributed strategy generation algorithm is used to generate an in-vehicle display optimization strategy. According to the generated in-vehicle display optimization strategy, the driving status information of each vehicle and the personalized preference parameters preset by the driver are obtained in real time, and the interface layout, display content, information hierarchy, color matching and interaction mode of the in-vehicle display interface are adjusted dynamically in real time; Through cross-vehicle information interaction, the display content of different vehicles is synchronized, the display conflicts of adjacent vehicles are coordinated, the display of vehicle groups is collaboratively optimized according to the vehicle formation status and driving characteristics, and a personalized vehicle display interface that is adaptive to the scenario is generated; Based on the integrated global traffic status information, a vehicle-to-vehicle interaction model and a vehicle-to-road interaction model are established. Each vehicle is regarded as an intelligent agent, and a multi-agent reinforcement learning framework is introduced. Through the collaboration between agents, a multi-objective optimization function is constructed. The objective function includes information transmission efficiency, driving safety, and cross-vehicle display coordination. Combined with the multi-objective optimization function, a distributed strategy generation algorithm is used to generate the vehicle display optimization strategy, including: Based on the fused global traffic state information, the interaction relationship between vehicles is determined according to the preset distance threshold and angle threshold, and the vehicles are used as nodes of the first directed graph. Directed edges are connected between vehicle pairs that meet the interaction conditions to obtain a vehicle-vehicle interaction model. The vehicles and road infrastructure are taken as nodes of the second directed graph, and directed edges are connected between relevant nodes according to the spatial relationship between the vehicles and the road infrastructure to obtain the vehicle-road interaction model; Generate an agent state according to the first directed graph and the second directed graph, wherein the agent state includes a vehicle motion state, environmental perception information, and interaction information between the vehicle and road infrastructure; According to the state of the intelligent agent, three optimization objectives are defined: information transmission efficiency, driving safety, and cross-vehicle display coordination. The information transmission efficiency is calculated by the delay and packet loss rate of display request transmission between vehicles. The driving safety is calculated by the influence of the display layout on the driver's attention allocation. The cross-vehicle display coordination is calculated by the difference of different vehicle interfaces. The three optimization objectives are combined to form a multi-objective optimization function. By solving the multi-objective optimization function, an initial display strategy is obtained. Based on the initial display strategy, the distributed strategy generation algorithm is used to learn and optimize the display strategy in parallel according to the local state and neighbor information of the agent. The agent Q function is updated by introducing a penalty term representing the difference in neighbor strategies to obtain the optimized display strategy. According to the optimized display strategy, key features are extracted and mapped to the semantic space for strategy conflict detection, and the conflicting strategy pairs are identified using the graph matching algorithm. For the detected conflicting strategy pairs, the strategy priority is determined by priority scoring. Conflicting strategies with different strategy priorities are coordinated by priority decision-making. Conflicting strategies with the same strategy priority are coordinated by consensus negotiation, and finally a consistent vehicle display optimization strategy is obtained. The calculation formula for updating the agent Q function is as follows: Among them, Q * represents the updated Q function of agent i, Q(·) represents the Q function, represents the strategy adopted by agent i at the kth step, α represents the learning rate, represents the immediate reward value returned by the environment after the agent i takes action a in the kth step, γ represents the reward decay factor, represents the maximum Q value that can be obtained by taking the optimal strategy π' in the next state, β represents the weight of the penalty term for the difference in neighbor strategies, and N i represents the neighbor set of agent i, represents the Q value of the strategy of agent i’s neighbor j at the kth step.

2. The method according to claim 1, characterized in that The step of obtaining multi-dimensional data of the driver's state and inputting the multi-dimensional data into the pre-trained state detection model also includes: Collect multi-dimensional data of driver status on the vehicle-mounted devices of multiple participants, and perform local preprocessing and feature extraction to obtain local feature data. Each vehicle-mounted device trains a state detection model based on the local feature data to generate local model parameters. Using multiple encryption communication protocols to encrypt and sign the local model parameters to form encrypted model parameters, and uploading the encrypted model parameters to the cloud server; The cloud server receives and decrypts and verifies the encrypted model parameters uploaded by each participant, obtains the decrypted local model parameters, and aggregates the decrypted local model parameters into global model parameters by introducing a secure aggregation algorithm with a differential privacy mechanism; An asynchronous aggregation strategy is adopted. When a preset number of participants upload encrypted model parameters, global aggregation is triggered. After global aggregation is triggered, the cloud server calculates the time decay factor of each participant, and weights the decrypted local model parameters of each participant based on the time decay factor. At the same time, a secure aggregation algorithm with a differential privacy mechanism is introduced to aggregate the weighted local model parameters into global model parameters. The global model parameters are encrypted and sent to each participant. Each participant receives and decrypts the global model parameters, replaces the local model parameters, and fine-tunes the replaced local model parameters based on the new local feature data to obtain the final state detection model.

3. The method according to claim 1, characterized in that Using the multi-hop reasoning result and the quantized score of the driver state as feature vectors, inputting a pre-trained multi-layer perceptron model to generate initial display configuration parameters, inputting the driver state quantized score, environmental factor parameters and initial display configuration parameters into a pre-trained random forest model, and generating configuration rules for display elements includes: The quantitative scores of the driver's status are normalized, the environmental factor parameters are one-hot encoded, and the initial display configuration parameters are standardized to obtain preprocessed input data; Using a permutation-based feature importance analysis algorithm to evaluate the preprocessed input data, select a target feature subset, and construct an integrated learning model including a random forest model, a gradient boosted decision tree model, and a support vector machine model; Using a grid search combined with a cross-validation algorithm to optimize the hyperparameters of the ensemble learning model to obtain an optimized base model, using a stacking algorithm to combine the optimized base models to obtain an initial ensemble model, and designing a dynamic weight adjustment mechanism to collect display configuration effect evaluation information in actual driving scenarios, and regularly adjusting the weights of each base model in the initial ensemble model according to the display configuration effect evaluation information to obtain an adjusted ensemble model; The preprocessed input data is input into the adjusted integrated model to generate preliminary display element configuration rules, a rule-based post-processing system is designed, the driver's historical preference data is input into the post-processing system, the preliminary display element configuration rules are personalized, and the final display element configuration rules are generated.

4. A vehicle display personalized control system based on deep learning, used to implement the method described in any one of claims 1 to 3, characterized in that: include: The first unit is used to obtain multi-dimensional data of the driver's state, and input the multi-dimensional data into a pre-trained state detection model to judge the driver's fatigue, attention level and emotional state, output a quantitative score of the driver's state, match the optimal display template from a preset display template library according to the quantitative score of the driver's state, wherein each display template contains a predefined personalized display rule, and adaptively determine the personalized configuration parameters of the vehicle display interface according to the personalized display rule of the matched display template; The second unit is used to obtain the driver's historical preference data from the cloud knowledge base, optimize the personalized configuration parameters of the vehicle display interface by using the collaborative filtering algorithm, generate the personalized configuration parameters preferred by the driver, obtain the vehicle driving status information and road traffic information in real time, and dynamically integrate them with the personalized configuration parameters preferred by the driver to generate a vehicle display optimization strategy based on the driving scene and the driver's preference. According to the vehicle display optimization strategy, the interface layout, display content, information hierarchy, color matching and interaction mode of the vehicle display interface are dynamically adjusted in real time to generate a personalized vehicle display interface; The third unit is used to obtain the driver's visual behavior and human-computer interaction behavior, as well as the subjective and objective feedback data on the personalized vehicle display interface, and store the driver's visual behavior, human-computer interaction behavior and subjective and objective feedback data in the user portrait database, and use an adaptive optimization algorithm to continuously optimize the state detection model and personalized display rules. The adaptive optimization algorithm uses the driver's subjective and objective feedback data as reward and punishment signals, updates the weight parameters of the state detection model and the personalized rules of the display template library, and realizes the adaptive evolution of the vehicle display system.

5. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 3.

6. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 3 is implemented.

Images