Vehicle-mounted interaction system based on cross-scene behavior learning and service prediction method
By using a cross-scenario behavior learning-based in-vehicle interaction system, which generates contextual models through four-dimensional behavior modeling and federated learning, and combines a trust mechanism, the limitations of single-dimensional learning and the neglect of dynamic context in existing technologies are solved, thereby achieving precise user interaction services and enhanced security.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- RIVOTEK TECH (JIANGSU) CO LTD
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-12
AI Technical Summary
Existing in-vehicle interaction systems suffer from single-dimensional learning limitations, are unable to correlate user behavior across scenarios, ignore dynamic contextual changes, lack proactive prediction capabilities, rely on explicit command responses, and are unable to provide accurate user interaction services.
The system adopts a cross-scenario behavior learning approach. It uses a four-dimensional behavior modeling engine (spatial habits, emotional state, decision-making patterns, and environmental associations) combined with federated learning and trust mechanisms to collect and analyze user behavior data in real time, generate contextual models, and predict and execute services.
It achieves unified modeling of behavior across scenarios, accurately integrates dynamic context prediction, adopts a progressive trust mechanism to solve the interference of initial misprediction, improves user experience and security, and meets privacy-sensitive requirements.
Smart Images
Figure CN122019627A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automotive artificial intelligence technology, and in particular to an in-vehicle interaction system and service prediction method based on cross-scenario behavior learning. Background Technology
[0002] With the intelligent development of vehicle cabin experience, users' requirements are no longer limited to mechanical performance such as power and acceleration, but they are pursuing a more refined and comfortable intelligent cabin environment and interactive experience. With the intelligent trend of multi-screen and large-screen cabins, proactive interaction helps to reduce distraction, improve driving safety, and enhance user experience.
[0003] However, existing in-vehicle interaction systems have limitations in single-dimensional learning, only analyzing preferences in a single scenario and failing to correlate cross-scenario behaviors (e.g., CN114567832A); and traditional recommendation systems ignore contextual changes (e.g., mood / weather affecting decision-making), rely on explicit command responses, and lack proactive prediction capabilities (e.g., US2023015334A1). Summary of the Invention
[0004] The purpose of this invention is to provide an in-vehicle interaction system and service prediction method based on cross-scenario behavior learning, which breaks through cross-dimensional association, integrates dynamic context and adopts a progressive trust mechanism.
[0005] This invention is implemented as follows: an in-vehicle interaction system based on cross-scenario behavior learning, comprising: The data layer collects user behavior data within the vehicle, including movement information, voice information, body language information, and environmental information. Based on this behavior data, a four-dimensional behavior modeling engine is constructed, encompassing spatial habits, emotional states, decision-making patterns, and environmental associations. Simultaneously, the collected behavior data is compared with behavior data stored in a cloud-based knowledge base. If no matching data is found, the behavior data is encrypted, stored in the cloud-based knowledge base, and sent to the learning layer. If matching data is found, the corresponding service action in the cloud-based knowledge base is directly invoked. The learning layer receives behavioral data and a four-dimensional behavioral modeling engine collected by the data layer, runs a federated learning framework on the vehicle edge computing unit, and learns the user's behavioral data by combining the output features of the four-dimensional behavioral modeling engine, generates a contextual model of the user's current comprehensive state, and sends it to the interaction layer. The interaction layer establishes a service execution hierarchy mechanism based on trust. It predicts the services that users may need based on the received context model, generates prediction results, and sends the prediction results to the execution layer. It calculates the system's prediction accuracy based on historical interaction data, determines the service execution method based on the number of user interactions with the system and the prediction accuracy, and sends the execution results, prediction results, and user feedback on the execution method back to the learning layer for context model updates and accuracy calculation. The execution layer establishes a prediction service matrix, which stores the mapping relationship between prediction results and corresponding execution actions. It receives prediction results and execution methods sent by the interaction layer, queries the prediction service matrix based on the prediction results to execute the corresponding service actions, and feeds back the execution status to the interaction layer.
[0006] Preferably, the method for constructing the four-dimensional behavior modeling engine includes: Spatial Habit Dimension: Collect users' mobility information, namely location and route selection preferences, use the ST-DBSCAN algorithm to analyze spatial data, evaluate the entropy value of locations, prioritize locations with high entropy values when executing service actions, mine users' habitual patterns in spatial movement, and output users' spatial habit feature vectors. Emotional state dimension: Collect user voice information, user voice and music played, analyze pitch, speech rate, keywords and music type played by the user, use LSTM emotion transfer model to learn the user's emotional state, calculate emotion index in real time, determine whether the user is currently in a low, happy, irritable or angry emotional state, and output emotional state label and emotion index. Decision-making pattern dimension: Collect users' body information, i.e., users' driving actions, and use Bayesian behavioral networks to build and dynamically update users' decision-making pattern models, thereby understanding the behavioral logic behind users' various in-vehicle function settings and driving operations, and outputting the probability distribution of decision preferences in different situations; Environmental Relationship Dimension: Integrate environmental information, including weather, time, and road conditions, and use graph neural networks to analyze the relationship between environmental factors and user behavior, outputting a weight matrix of the impact of environmental factors on user behavior.
[0007] Preferably, the spatiotemporal weighting function used in the ST-DBSCAN algorithm is: ; in, The weighting of historical time data to the current analysis. The attenuation coefficient; For the current time, This refers to historical time.
[0008] Preferably, the formula for calculating the entropy value of a location in the spatial habit dimension is: ; in, The location entropy value. For location Frequency of appearance in user history.
[0009] Preferably, the LSTM emotion transfer model is used to learn the user's emotional state and calculate the voice emotion index in real time, as shown in the following formula: ; in, The mood index is graded from 0 to 10, where 1-3 is low, 4-6 is happy, 7-9 is irritable, and 10 is angry. , , and These represent the sentiment probabilities corresponding to tone of voice, speech rate, keywords, and music genre, respectively. , , and The feature weights are for pitch, speech rate, keywords, and music genre, respectively.
[0010] Preferably, the user's decision-making pattern model constructed and dynamically updated using a Bayesian behavioral network is calculated as follows: ; in, Let A be the conditional probability that decision factor A leads to behavior B after the update. Let A be the conditional probability that decision factor A leads to behavior B before the update. This refers to the historical sample size. For indicator functions, new behavior The value is 1 if it occurs, and 0 otherwise. Preferably, the operation of the federated learning framework includes: Local model update: The in-vehicle edge computing unit uses a local dataset. For global model parameters Update: ; in, For learning rate, ; The loss function; For the updated local model parameters; These are model parameters; Represents the loss function Regarding parameters The gradient; Cloud aggregation: The cloud server collects local model parameters from each participating device. and the number of local samples Perform a weighted average: ; in, This represents the number of local samples. The total number of samples from all participating devices; The new global model parameters after aggregation; For the first Updated local model parameters for each vehicle-mounted edge computing unit; Differential privacy protection when updating the local model Add Laplace noise before uploading to the cloud: ; in, Sensitivity to updating vectors for the model; Budget for privacy and control the level of privacy protection.
[0011] Preferably, the step of predicting the services the user might need based on the received context model and generating a prediction result is achieved through multi-dimensional feature weighted fusion, calculated as follows: ; ; in, For the predicted service type score, For the first Weights of each dimension For the first The feature values of each dimension; the service with the highest score is the prediction result.
[0012] Preferably, the establishment of a trust-based service execution tiered mechanism, which determines the service execution method based on the number of user interactions with the system and the prediction accuracy, includes: Trust score T is the prediction accuracy based on the number of historical interactions N and the statistics within the sliding time window. The calculation formula is as follows: ; in , These are the weighting coefficients. ; It is an increasing function with respect to the number of interactions; If trust rating It only executes explicit instructions, that is, instructions actively entered by the user; If trust rating The system then provides a prediction result option for the user to confirm and execute, and updates the prediction record and accuracy based on the user's selection. ; If trust rating If the prediction is not executed automatically, the result will be returned to the user. The system will then adjust the update based on whether the user cancels the update of the prediction record and the accuracy rate. ; The prediction accuracy The calculation method is as follows: within the set statistical period or sample size window, the proportion of successful predictions to the total number of prediction attempts, where successful prediction is defined as: the user confirms the execution after the system provides the option, or the user does not cancel after the system automatically executes; failed prediction is defined as: the user rejects the option or cancels the automatic execution.
[0013] Preferably, the prediction accuracy The calculation method is as follows: within the set statistical period or sample size window, the proportion of successful predictions to the total number of prediction attempts, where successful prediction is defined as: the user confirms the execution after the system provides the option, or the user does not cancel after the system automatically executes; failed prediction is defined as: the user rejects the option or cancels the automatic execution.
[0014] A service prediction method based on cross-scenario behavior learning includes: Real-time collection of user in-vehicle behavior data; the behavior data includes: voice information, body information and environmental information, and a four-dimensional behavior modeling engine is constructed based on the user's behavior data, including spatial habit dimension, emotional state dimension, decision-making mode dimension and environmental association dimension; The collected behavioral data is compared with the behavioral data stored in the cloud knowledge base. If no matching data is found, the behavioral data is encrypted, stored in the cloud knowledge base, and sent to the learning layer. If matching data is found, the service action corresponding to the behavioral data in the cloud knowledge base is directly invoked. Run the federated learning framework and combine it with the four-dimensional behavior modeling engine to learn from user behavior data and generate a contextual model that represents the user's current comprehensive state. Establish a service execution hierarchy mechanism based on trust. Based on the received context model, predict the services that users may need, generate prediction results, and send the prediction results to the execution layer. Calculate the system's prediction accuracy based on historical interaction data. Determine the service execution method based on the number of user interactions with the system and the prediction accuracy. Send the execution results, prediction results, and user feedback on the execution method back to the learning layer for context model updates and accuracy calculations. The service execution method is determined based on the number of user interactions with the system and the system's prediction accuracy: if trust score is used... Only execute explicit instructions; if trust score It provides a prediction result option for the user to confirm execution and updates the trust score and prediction accuracy; if the trust score... It automatically executes the prediction and returns the result to the user.
[0015] Update prediction records, prediction accuracy, and trust scores based on the user's selection of provided options or undoing of the automatic execution results.
[0016] Update prediction records, prediction accuracy, and trust scores based on the user's selection of provided options or undoing of the automatic execution results.
[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. Unify the modeling of driving behavior, entertainment preferences, and decision-making patterns to break through the single-dimensional correlation of CN114567832A; 2. It integrates real-time prediction of physical environment and psychological state, incorporating dynamic context, and is more accurate than the pure geolocation prediction of US2023015334A1. 3. A progressive trust mechanism is adopted, which dynamically opens system permissions in stages based on the number of interactions to solve the problem of initial misprediction interference. Users can freeze their learning at any time to meet privacy-sensitive needs. Attached Figure Description
[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a system structure block diagram of Embodiment 1 of the present invention; Figure 2 This is a flowchart of the prediction method in Embodiment 1 of the present invention. Detailed Implementation
[0020] To better understand the technical content of this invention, the technical solutions of this invention are further described and explained below with reference to specific embodiments, but are not limited thereto. The technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0021] Example 1 refer to Figures 1 to 2 This in-vehicle interaction system based on cross-scenario behavior learning includes: The data layer collects user behavior data within the vehicle, including movement information, voice information, body language information, and environmental information. Based on this behavior data, a four-dimensional behavior modeling engine is constructed, encompassing spatial habits, emotional states, decision-making patterns, and environmental associations. Simultaneously, the collected behavior data is compared with behavior data stored in a cloud-based knowledge base. If no matching data is found, the behavior data is encrypted, stored in the cloud-based knowledge base, and sent to the learning layer. If matching data is found, the corresponding service action in the cloud-based knowledge base is directly invoked. The construction method of the four-dimensional behavior modeling engine includes: Spatial Habit Dimension: Collect users' mobility information, namely location and route selection preferences, use the ST-DBSCAN algorithm to analyze spatial data, evaluate the entropy value of locations, prioritize locations with high entropy values when executing service actions, mine users' habitual patterns in spatial movement, and output users' spatial habit feature vectors. The spatiotemporal weighting function used in the ST-DBSCAN algorithm is: ; in, The weighting of historical time data to the current analysis. This is the attenuation coefficient, with a value ranging from 0.1 to 1.0; For the current time, For historical time; The formula for calculating the entropy value of a location in the spatial habit dimension is as follows: ; in, The location entropy value; For location The lower the entropy value, the more concentrated the location preference is, indicating the frequency of occurrence in the user's historical trajectory.
[0022] Emotional state dimension: Collect user voice information, user voice and music played, analyze pitch, speech rate, keywords and music type played by the user, use LSTM emotion transfer model to learn the user's emotional state, calculate emotion index in real time, determine whether the user is currently in a low, happy, irritable or angry emotional state, and output emotional state label and emotion index. The LSTM emotion transfer model is used to learn the user's emotional state and calculate the voice emotion index in real time. The calculation formula is as follows: ; in, The mood index is graded from 0 to 10, where 1-3 is low, 4-6 is happy, 7-9 is irritable, and 10 is angry. , , and These represent the sentiment probabilities corresponding to tone of voice, speech rate, keywords, and music genre, respectively. , , and The feature weights are for pitch, speech rate, keywords, and music genre, respectively.
[0023] Decision-making pattern dimension: Collect user body information, i.e., user driving actions, such as decision-related data on air conditioning settings, driving style and seat adjustment, and use Bayesian behavioral networks to build and dynamically update the user's decision-making pattern model, thereby understanding the behavioral logic behind the user's various in-vehicle function settings and driving operations, and outputting the probability distribution of decision preferences in different situations. The user's decision-making pattern model, constructed and dynamically updated using a Bayesian behavioral network, is calculated as follows: ; in, Let A be the conditional probability that decision factor A leads to behavior B after the update. Let A be the conditional probability that decision factor A leads to behavior B before the update. This refers to the historical sample size. For indicator functions, new behavior The value is 1 if it occurs, and 0 otherwise.
[0024] Environmental Relationship Dimension: Integrate environmental information, including weather, time, and road conditions, and use graph neural networks to analyze the relationship between environmental factors and user behavior, outputting a weight matrix of the impact of environmental factors on user behavior.
[0025] The learning layer receives behavioral data and a four-dimensional behavioral modeling engine collected by the data layer, runs a federated learning framework on the vehicle edge computing unit, and learns the user's behavioral data by combining the output features of the four-dimensional behavioral modeling engine, generates a contextual model of the user's current comprehensive state, and sends it to the interaction layer.
[0026] The operation of the federated learning framework includes: Local model update: The in-vehicle edge computing unit uses a local dataset. For global model parameters Update: ; in, The learning rate; The loss function; For the updated local model parameters; These are model parameters; Represents the loss function Regarding parameters The gradient; Cloud aggregation: The cloud server collects local model parameters from each participating device. and the number of local samples Perform a weighted average: ; in, This represents the number of local samples. The total number of samples from all participating devices; The new global model parameters after aggregation; For the first Updated local model parameters for each vehicle-mounted edge computing unit; Differential privacy protection when updating the local model Add Laplace noise before uploading to the cloud: ; in, Sensitivity to updating vectors for the model; For privacy budgeting, control the strength of privacy protection, with a value ranging from 0.1 to 10.0.
[0027] The interaction layer establishes a service execution hierarchy mechanism based on trust. It predicts the services a user might need based on the received context model, generates prediction results, and sends these results to the execution layer. The system's prediction accuracy is calculated based on historical interaction data. The service execution method is determined based on the number of user interactions and the prediction accuracy. The execution results, prediction results, and user feedback are sent back to the learning layer for context model updates and accuracy calculations. The prediction of potential user services based on the received context model is achieved through multi-dimensional feature weighted fusion, calculated as follows: ; ; in, For the predicted service type score, For the first Weights of each dimension For the first The feature values of each dimension; the service with the highest score is the prediction result; The establishment of a trust-based service execution tiered mechanism, which determines the service execution method based on the number of user interactions with the system and prediction accuracy, includes: Trust score T is the prediction accuracy based on the number of historical interactions N and the statistics within the sliding time window. The calculation formula is as follows: ; in , These are the weighting coefficients. ; It is an increasing function with respect to the number of interactions; If trust rating It only executes explicit instructions, that is, instructions actively entered by the user; If trust rating If the current prediction result has a confidence level >75%, then a prediction result option will be provided for the user to confirm and execute, and the prediction record and accuracy will be updated according to the user's selection. ; If trust rating If the current prediction result has a confidence level >90%, the prediction will be executed automatically, and the result will be fed back to the user. The system will then adjust the update based on whether the user cancels the update of the prediction record and the accuracy. ; The prediction accuracy The calculation method is as follows: within the set statistical period or sample size window, the proportion of successful predictions to the total number of prediction attempts, where successful prediction is defined as: the user confirms the execution after the system provides the option, or the user does not cancel after the system automatically executes; failed prediction is defined as: the user rejects the option or cancels the automatic execution.
[0028] The execution layer establishes a prediction service matrix, which stores the mapping relationship between prediction results and corresponding execution actions. It receives prediction results and execution methods sent by the interaction layer, queries the prediction service matrix based on the prediction results, executes the corresponding service actions, and feeds back the execution status to the interaction layer. For example: When the mood index is 1-3 (low), play soothing music to help alleviate the low mood; when the mood index is 4-6 (happy), maintain the current in-car environment; when the mood index is 7-9 (irritable), turn on the ventilation and play soft music to help soothe the irritable mood; when the mood index is 10 (angry), actively reduce noise and dim the lights to reduce irritation. When the weather is sunny, open the sun visor to prevent glare from affecting driving; when the weather is cloudy, play upbeat music to avoid the gloom affecting your mood; when the weather is rainy, turn on the windshield wipers to clear the rain from the windshield; when the weather is windy or thundering, turn on the noise reduction to prevent wind and thunder noise from affecting driving; when the weather is snowy, turn on the air conditioner to blow cold air into the windshield to prevent the warm air inside the car from fogging up the windshield; When it is morning, lower the car windows so the wind blows towards the driver to help them stay awake and avoid drowsiness; when it is noon, keep the interior as it is in spring and autumn, use the air conditioning to cool in summer and heat in winter; when it is evening, turn on the headlights. When the traffic conditions on the frequently used route are clear, take no action; when the traffic conditions on the frequently used route are congested or under construction, select the shortest and clearest route for navigation.
[0029] For example: Table 1 Execution Layer Runtime Examples ; Table 2. Evolution Curve of Prediction Accuracy .
[0030] Example 2 A service prediction method based on cross-scenario behavior learning includes: Real-time collection of user in-vehicle behavior data; the behavior data includes: voice information, body information and environmental information, and a four-dimensional behavior modeling engine is constructed based on the user's behavior data, including spatial habit dimension, emotional state dimension, decision-making mode dimension and environmental association dimension; The collected behavioral data is compared with the behavioral data stored in the cloud knowledge base. If no matching data is found, the behavioral data is encrypted, stored in the cloud knowledge base, and sent to the learning layer. If matching data is found, the service action corresponding to the behavioral data in the cloud knowledge base is directly invoked. Run the federated learning framework and combine it with the four-dimensional behavior modeling engine to learn from user behavior data and generate a contextual model that represents the user's current comprehensive state. Establish a service execution hierarchy mechanism based on trust. Based on the received context model, predict the services that users may need, generate prediction results, and send the prediction results to the execution layer. Calculate the system's prediction accuracy based on historical interaction data. Determine the service execution method based on the number of user interactions with the system and the prediction accuracy. Send the execution results, prediction results, and user feedback on the execution method back to the learning layer for context model updates and accuracy calculations. The service execution method is determined based on the number of user interactions with the system and the system's prediction accuracy: if trust score is used... Only execute explicit instructions; if trust score It provides a prediction result option for the user to confirm execution and updates the trust score and prediction accuracy; if the trust score... It automatically executes the prediction and returns the result to the user.
[0031] Update prediction records, prediction accuracy, and trust scores based on the user's selection of provided options or undoing of the automatic execution results.
[0032] Update prediction records, prediction accuracy, and trust scores based on the user's selection of provided options or undoing of the automatic execution results.
[0033] Example 3 1. Initial stage (after 1 week) User said, "It's a bit stuffy" → System suggestion: "Turn on external circulation? Play some light music?" 2. Mid-term (3 months later) User sighs → System automatically: Navigate to the viewing platform at Jiangbin Park; Play the saved playlist of "Seaside Piano Music"; Set the fragrance to "Ocean Breeze" mode; 3. Long-term (after 1 year) The user says, "Let's go." → The system is based on a Friday scenario at 5:30 PM: Automatically navigate to your favorite hot pot restaurant; Play the weekend party playlist; Reserve a seat in the store.
[0034] The above-described embodiments are only some embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. An in-vehicle interaction system based on cross-scenario behavior learning, characterized in that, include: The data layer collects user behavior data within the vehicle, including movement information, voice information, body language information, and environmental information. Based on this behavior data, a four-dimensional behavior modeling engine is constructed, encompassing spatial habits, emotional states, decision-making patterns, and environmental associations. Simultaneously, the collected behavior data is compared with behavior data stored in a cloud-based knowledge base. If no matching data is found, the behavior data is encrypted, stored in the cloud-based knowledge base, and sent to the learning layer. If matching data is found, the corresponding service action in the cloud-based knowledge base is directly invoked. The learning layer receives behavioral data and a four-dimensional behavioral modeling engine collected by the data layer, runs a federated learning framework on the vehicle edge computing unit, and learns the user's behavioral data by combining the output features of the four-dimensional behavioral modeling engine, generates a contextual model of the user's current comprehensive state, and sends it to the interaction layer. The interaction layer establishes a service execution hierarchy mechanism based on trust. It predicts the services that users may need based on the received context model, generates prediction results, and sends the prediction results to the execution layer. It calculates the system's prediction accuracy based on historical interaction data, determines the service execution method based on the number of user interactions with the system and the prediction accuracy, and sends the execution results, prediction results, and user feedback on the execution method back to the learning layer for context model updates and accuracy calculation. The execution layer establishes a prediction service matrix, which stores the mapping relationship between prediction results and corresponding execution actions. It receives prediction results and execution methods sent by the interaction layer, queries the prediction service matrix based on the prediction results to execute the corresponding service actions, and feeds back the execution status to the interaction layer.
2. The in-vehicle interaction system based on cross-scenario behavior learning as described in claim 1, characterized in that, The construction method of the four-dimensional behavior modeling engine includes: Spatial Habit Dimension: Collect users' mobility information, namely location and route selection preferences, use the ST-DBSCAN algorithm to analyze spatial data, evaluate the entropy value of locations, prioritize locations with high entropy values when executing service actions, mine users' habitual patterns in spatial movement, and output users' spatial habit feature vectors. Emotional state dimension: Collect user voice information, user voice and music played, analyze pitch, speech rate, keywords and music type played by the user, use LSTM emotion transfer model to learn the user's emotional state, calculate emotion index in real time, determine whether the user is currently in a low, happy, irritable or angry emotional state, and output emotional state label and emotion index. Decision-making pattern dimension: Collect users' body information, i.e., users' driving actions, and use Bayesian behavioral networks to build and dynamically update users' decision-making pattern models, thereby understanding the behavioral logic behind users' various in-vehicle function settings and driving operations, and outputting the probability distribution of decision preferences in different situations; Environmental Relationship Dimension: Integrate environmental information, including weather, time, and road conditions, and use graph neural networks to analyze the relationship between environmental factors and user behavior, outputting a weight matrix of the impact of environmental factors on user behavior.
3. The in-vehicle interaction system based on cross-scenario behavior learning as described in claim 2, characterized in that, The spatiotemporal weighting function used in the ST-DBSCAN algorithm is: in, The weighting of historical time data to the current analysis. The attenuation coefficient; For the current time, This refers to historical time.
4. The in-vehicle interaction system based on cross-scenario behavior learning as described in claim 3, characterized in that, The formula for calculating the entropy value of a location in the spatial habit dimension is as follows: in, The location entropy value. For location Frequency of appearance in user history.
5. The in-vehicle interaction system based on cross-scenario behavior learning as described in claim 2, characterized in that, The LSTM emotion transfer model is used to learn the user's emotional state and calculate the voice emotion index in real time. The calculation formula is as follows: in, The mood index is graded from 0 to 10, where 1-3 is low, 4-6 is happy, 7-9 is irritable, and 10 is angry. , , and These represent the sentiment probabilities corresponding to tone of voice, speech rate, keywords, and music genre, respectively. , , and The feature weights are for pitch, speech rate, keywords, and music genre, respectively.
6. The in-vehicle interaction system based on cross-scenario behavior learning as described in claim 2, characterized in that, The user's decision-making pattern model, constructed and dynamically updated using a Bayesian behavioral network, is calculated as follows: in, Let A be the conditional probability that decision factor A leads to behavior B after the update. Let A be the conditional probability that decision factor A leads to behavior B before the update. This refers to the historical sample size. For indicator functions, new behavior The value is 1 if it occurs, and 0 otherwise.
7. The in-vehicle interaction system based on cross-scenario behavior learning as described in claim 1, characterized in that, The operation of the federated learning framework includes: Local model update: The in-vehicle edge computing unit uses a local dataset. For global model parameters Update: in, The learning rate; The loss function; For the updated local model parameters; These are model parameters; Represents the loss function Regarding parameters The gradient; Cloud aggregation: The cloud server collects local model parameters from each participating device. and the number of local samples Perform a weighted average: in, This represents the number of local samples. The total number of samples from all participating devices; These are the parameters for the new global model after aggregation. For the first Updated local model parameters for each vehicle-mounted edge computing unit; Differential privacy protection when updating the local model Add Laplace noise before uploading to the cloud: in, Sensitivity to updating vectors for the model; Budget for privacy and control the level of privacy protection.
8. The in-vehicle interaction system based on cross-scenario behavior learning as described in claim 1, characterized in that, The process of predicting the services a user might need based on the received context model and generating a prediction result is achieved through multi-dimensional feature weighted fusion, calculated as follows: in, For the predicted service type score, For the first Weights of each dimension For the first The feature values of each dimension; the service with the highest score is the prediction result.
9. The in-vehicle interaction system based on cross-scenario behavior learning as described in claim 1, characterized in that, The establishment of a trust-based service execution tiering mechanism determines the service execution method based on the number of user interactions with the system and the prediction accuracy. This includes: a trust score T based on the historical interaction count N and the prediction accuracy calculated within a sliding time window. The calculation formula is as follows: in , These are the weighting coefficients. ; It is an increasing function with respect to the number of interactions; If trust rating It only executes explicit instructions, that is, instructions actively entered by the user; If trust rating The system then provides a prediction result option for the user to confirm and execute, and updates the prediction record and accuracy based on the user's selection. ; If trust rating If the prediction is not executed automatically, the result will be returned to the user. The system will then adjust the update based on whether the user cancels the update of the prediction record and the accuracy rate. ; The prediction accuracy The calculation method is as follows: within the set statistical period or sample size window, the proportion of successful predictions to the total number of prediction attempts, where successful prediction is defined as: the user confirms the execution after the system provides the option, or the user does not cancel after the system automatically executes; failed prediction is defined as: the user rejects the option or cancels the automatic execution.
10. A service prediction method based on cross-scenario behavior learning, characterized in that, include: Real-time collection of user in-vehicle behavior data; The behavioral data includes: voice information, body information and environmental information. A four-dimensional behavioral modeling engine is constructed based on the user's behavioral data, which includes spatial habit dimension, emotional state dimension, decision-making mode dimension and environmental association dimension. The collected behavioral data is compared with the behavioral data stored in the cloud knowledge base. If no matching data is found, the behavioral data is encrypted, stored in the cloud knowledge base, and sent to the learning layer. If matching data is found, the service action corresponding to the behavioral data in the cloud knowledge base is directly invoked. Run the federated learning framework and combine it with the four-dimensional behavior modeling engine to learn from user behavior data and generate a contextual model that represents the user's current comprehensive state. Establish a service execution hierarchy mechanism based on trust. Based on the received context model, predict the services that users may need, generate prediction results, and send the prediction results to the execution layer. Calculate the system's prediction accuracy based on historical interaction data. Determine the service execution method based on the number of user interactions with the system and the prediction accuracy. Send the execution results, prediction results, and user feedback on the execution method back to the learning layer for context model updates and accuracy calculations. The service execution method is determined based on the number of user interactions with the system and the system's prediction accuracy: if trust score is used... Only execute explicit instructions; if trust score It provides a prediction result option for the user to confirm execution and updates the trust score and prediction accuracy; if the trust score... It automatically executes the prediction and returns the result to the user. Update prediction records, prediction accuracy, and trust scores based on the user's selection of provided options or undoing of the automatic execution results.