A control method for assisting human-vehicle interaction based on a lightweight large model

By building intention recognition modules and a lightweight big model with dual-stage pre-training in smart cars, combining the interaction between cloud and on-board systems, the problem of insufficient emotional and intention understanding in smart cars is solved, and more accurate driver needs are achieved Perceived and intelligent driving experience.

CN117944703BActive Publication Date: 2024-11-12CHONGQING UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311806692.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-26
Publication Date
2024-11-12
Estimated Expiration
2043-12-26

AI Technical Summary

Technical Problem

The existing intelligent car human-vehicle interaction technology lacks an effective understanding of the driver's emotions and intentions, which leads to the vehicle being unable to accurately perceive the driver's needs and emotional state, affecting the intelligent driving experience.

Method used

The auxiliary human-vehicle interaction control method based on lightweight large models is adopted. By building an intention recognition module, the T5 model is used for dual-stage pre-training and continuous optimization, combining the interaction between the cloud and the on-board system, the recognition and response of emotions and intentions are achieved.

Benefits of technology

It improves the intelligence level of human-vehicle interaction, can more accurately identify the driver's emotional state and intentions, and improves the personalization of the driving experience and the accuracy of response.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117944703B_ABST
    Figure CN117944703B_ABST
Patent Text Reader

Abstract

The present invention discloses a control method for assisting human-vehicle interaction based on a lightweight large model, and relates to the technical field of human-vehicle interaction. The present invention includes at least the following steps: S1: construction and application of a lightweight student model; S2: two-stage pre-training of the model; S3: continuous optimization and update mechanism; S4: interaction between the cloud and the vehicle-mounted system. The present invention can more accurately identify the driver's emotional state and intention through a model optimized for driving scenarios, thereby improving the level of intelligence in human-vehicle interaction, and by continuously collecting data under driving scenarios for optimization, the model can better adapt to specific driving environments and user needs, achieve scene adaptability, and use a high-capability model on the cloud to compare and optimize the lightweight student model, so that the vehicle-mounted model can improve the accuracy and personalization of the response while maintaining its lightweight.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of human-vehicle interaction, and in particular to a control method for assisting human-vehicle interaction based on a lightweight large model. Background Art

[0002] Human-vehicle interaction refers to the interaction between people and cars. It not only includes intelligent interaction such as voice recognition, but also covers the emotional interaction between vehicles and people. With the rapid development of smart cars, human-vehicle interaction is becoming more and more important for smart cars.

[0003] First, human-vehicle interaction can improve driving safety and comfort. Smart cars can automatically adjust the temperature and music in the car according to the driver's emotional state and needs, provide a more personalized driving experience, and enhance driving comfort. Second, human-vehicle interaction can improve the level of intelligent driving. Through the emotional interaction between the vehicle and the person, the smart car can perceive the driver's emotions and needs, and automatically adjust the vehicle's operating status and functions to provide a more intelligent driving experience. For example, when the driver feels tired, the vehicle can automatically reduce the speed, play relaxing music, etc. to relieve the driver's fatigue. In the autonomous driving mode, the vehicle needs to be able to accurately perceive the intentions and needs of people so as to make correct driving decisions. For example, when the vehicle encounters a dangerous situation during autonomous driving, the vehicle needs to interact with passengers through voice prompts, light prompts, etc. to ensure the safety of passengers.

[0004] In the existing patents, such as the application publication number CN116913270A, the name is a voice interaction method, system and vehicle, which determines the voice command through the user's voice; generates a trigger signal according to the voice command, and obtains the interaction result corresponding to the trigger signal.

[0005] This patent generates trigger signals through voice commands, but lacks understanding of the driver's emotions and intentions during the interaction process. Such interaction is insufficient for intelligent human-vehicle interaction.

[0006] Therefore, it can be seen that the current human-vehicle interaction of smart cars lacks emotional understanding and intention understanding, which means that the vehicle cannot truly understand the driver's needs and emotional state. In the design of smart cars, the understanding of emotions and intentions will become the key. However, due to the limitations of the vehicle scene, it is difficult to deploy more models to analyze the driver's language, voice and behavior to understand their emotional state and driving intentions, resulting in a poor human-vehicle interaction experience. Therefore, a new solution to the above problems is needed. Summary of the invention

[0007] The object of the present invention is to provide a control method for assisting human-vehicle interaction based on a lightweight large model.

[0008] A control method for assisting human-vehicle interaction based on a lightweight large model includes at least the following steps:

[0009] S1: Construction and application of lightweight student model, using T5 model as the basic model, by using it to distill the lightweight T5 student model obtained from the large model, specifically for the intention recognition of driving scenes, to build an intention recognition module, the intention recognition module is used to recognize the user's intention from the user's language input, and may also include the recognition of emotional state. In order to achieve the purpose of improving the quality of control interaction;

[0010] S2: Two-stage pre-training of the model, combining emotion recognition and intent recognition datasets for initial pre-training of the T5 model, and pre-training enhancement through simulated scenarios;

[0011] S3: Continuous optimization and update mechanism, which continuously collects data in the vehicle system and continuously optimizes and updates parameters through the large model on the cloud server, including at least subsequent model optimization, which includes at least cloud data collection, model optimization and quantitative evaluation of the new model, and the subsequent model optimization is a unified improvement for all users in the entire interaction scenario;

[0012] S4: Interaction between the cloud and in-vehicle systems, using the high-capability model on the cloud to compare and optimize with the lightweight student model on the vehicle.

[0013] The two-stage pre-training of the model includes at least initial pre-training and pre-training enhancement;

[0014] The initial pre-training adopts the T5 model as a basic model, and the model is initially pre-trained by using a public emotion recognition dataset and an intention recognition dataset. The pre-training enhancement is used to further enhance the initial pre-training.

[0015] Preferably, the initial pre-training includes at least the following steps:

[0016] First, we will collect and organize a large amount of text data, including data from various languages ​​and fields, to build a large-scale corpus;

[0017] Said corpus contains a variety of different types of text, including at least news articles, social media posts, product reviews, and conversations;

[0018] Subsequently, the T5 model is pre-trained using the data in the corpus so that the model can learn the ability to extract emotions and intentions from text. The model uses the cross entropy loss function as the loss function, as shown in the following formula:

[0019]

[0020] L(θ) is the loss function itself, usually represented by L(.), where θ represents the model parameters. The purpose of the loss function is to measure the performance of the model, where θ represents the model parameters;

[0021] p data (x,y) is the true distribution of the data, which indicates the probability of the correct label y given the input x;

[0022] p model (y|x; θ) is the conditional probability distribution predicted by the model, which represents the probability of predicting the correct label y given the input x under the model parameters θ;

[0023] The pre-training process is to minimize the loss function L(θ) by adjusting θ;

[0024] Finally, the model after the initial pre-training is evaluated and optimized, and the performance of the model is measured using standard evaluation indicators, including at least accuracy, recall, and F1 score, as shown in the following formula:

[0025]

[0026]

[0027]

[0028]

[0029] Among them, TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives respectively. According to the evaluation results, the model will be further adjusted and optimized to improve its performance;

[0030] Accuracy is a measure of the proportion of correct predictions made by the classifier, that is, the proportion of all correct predictions (positive and negative) to all predictions;

[0031] Recision is a measure of the proportion of positive classes that are actually predicted to be positive, also known as the precision rate;

[0032] Recall is a measure of the proportion of actually positive classes that are correctly predicted to be positive, also known as recall or sensitivity. It is the harmonic mean of precision and recall, and is used when we need to consider the balance between precision and recall.

[0033] Preferably, the pre-training enhancement comprises at least the following steps:

[0034] By using two large models, one of which simulates the role of the driver and the other is responsible for responding to the driver's requests;

[0035] Whenever the driver character model makes a request, the request is passed to the other large model and the lightweight T5 student model at the same time;

[0036] The outputs of the two models constitute a bigram, which includes at least the emotional state and the user intention regarding the driving demand;

[0037] By comparing the answers of the large model with those of the lightweight T5 student model, the lightweight T5 student model was optimized, and the pre-training enhancement process was achieved through simulation scenarios;

[0038] In this process we did not pre-train P data (y|x), but instead have a dataset with two large examples of model inputs and correct outputs. In this case, the loss function used is as follows:

[0039]

[0040] Where N is the total number of samples;

[0041] i is the index, used to represent the i-th sample;

[0042] x (i) is the input feature representing the i-th sample;

[0043] y (i) represents the actual label of the i-th sample;

[0044] P model is the probability distribution function predicted by the model, which gives the probability of predicting the output y given the input x and model parameters θ.

[0045] The following formula is used to minimize the loss:

[0046]

[0047] Among them, α is the learning rate, is the gradient of the loss function with respect to the parameter θ;

[0048] θ new is the updated model parameter. old are the model parameters before updating.

[0049] After two stages of pre-training, the model can enter the deployment stage.

[0050] Preferably, the subsequent optimization of the model comprises at least the following steps:

[0051] Collect user-vehicle interaction data in driving scenarios. The data content includes user requests, speech recognition results, responses after recognition, and the next user request. Through the next user request, we can predict and judge the response result and whether the previous response meets the user's needs.

[0052] Transmitting the collected data to the cloud means transferring the data stored onboard the vehicle to a remote server;

[0053] A large, advanced model is deployed in the cloud. This model uses user data to generate corresponding responses and compares and optimizes them with a lightweight Student T5 model. This large model has a very high ability to quickly generate accurate and personalized responses. At the same time, the lightweight Student T5 model updates and optimizes its own parameters based on the difference in responses with the large model.

[0054] Different from the pre-training of the model, a dynamic adjustment function is introduced for the continuous optimization process to simulate the real-time update of the cloud model parameters. The formula is as follows:

[0055]

[0056] Among them, θ t+1 These are the model parameters at time step or iteration t+1;

[0057] θ t These are the model parameters at time step or iteration t;

[0058] η t This is the learning rate at time step or iteration t, a positive scalar that controls the parameter update step size;

[0059] This is the gradient of the loss function J with respect to the parameter θ at time step t, which represents the local slope or derivative of the loss function with respect to the parameter;

[0060] ∈ t This is the noise term added at time step t, which is usually used to help the model jump out of the local minimum or add randomness to the training process;

[0061] After the optimization, we use the area under the ROC curve to quantify the model performance. The formula is as follows:

[0062]

[0063] Where TPR(d) is the true positive rate, FDR(d) is the false positive rate, and d is the decision threshold;

[0064] After a period of time, the T5 model parameters in the cloud are transmitted to the on-board T5 to update the parameters of the on-board T5 model;

[0065] The T5 lightweight model is deployed between the speech recognition module and the control module, responsible for identifying user emotions and intentions to improve response quality;

[0066] The big model in the cloud generates answers based on the collected user data and compares and optimizes the response of the T5 lightweight model;

[0067] The optimized T5 lightweight model parameters will be transmitted back to the model deployed in the driving scenario to update the parameters of the model in the driving scenario.

[0068] Preferably, the application of the intention recognition module includes at least the following steps:

[0069] Receiving voice input: The system collects the user's voice commands, converts the voice input into text data through the voice recognition module, and cleans and preprocesses the converted text;

[0070] Feature extraction: extract features from the cleaned text for model training and inference;

[0071] Model inference: Use a lightweight student model to process features and output the probability distribution of intent prediction;

[0072] Post-processing and response: Select the most likely intent based on the model output and convert it into a command that the interactive system can understand and respond to;

[0073] Through the optimization of this part, it is possible to use a lightweight large vehicle-mounted model to recognize the user's intentions and emotions during the human-vehicle interaction process, so as to improve the recognition rate of user commands in the human-vehicle interaction and improve the quality of the human-vehicle interaction process.

[0074] Preferably, the model inference process is as follows:

[0075]

[0076] The posterior probability is as above;

[0077]

[0078] The posterior probability is as above;

[0079]

[0080] The posterior probability is as above;

[0081]

[0082] The posterior probability is as above.

[0083] Compared with the prior art, the present invention has the following beneficial effects:

[0084] 1. This invention introduces a dedicated intent recognition module through the design of the intent recognition module. The module uses a lightweight student model distilled from a large model, which is specifically targeted at driving scenarios. It is used to improve the current human-vehicle interaction;

[0085] 2. The present invention uses the T5 model to perform initial pre-training and pre-training enhancement on the model by designing a pre-training method, combined with emotion recognition and intention recognition datasets. This is different from the single pre-training model that may be used in traditional methods or a general model that is not optimized for a specific scenario;

[0086] 3. The present invention continuously collects data in the vehicle system through the design of continuous optimization process, and continuously optimizes it through the large model on the cloud server. This method can achieve continuous evolution and adaptability improvement of the model;

[0087] 4. The present invention can more accurately identify the driver's emotional state and intention through a model optimized for driving scenarios, thereby improving the level of intelligence of human-vehicle interaction;

[0088] 5. The present invention continuously collects data under driving scenarios for optimization, so that the model can better adapt to specific driving environments and user needs, and achieve scene adaptability;

[0089] 6. The present invention uses a high-capability model in the cloud to compare and optimize the lightweight student model, so that the on-board model can improve the accuracy and personalization of the response while maintaining its lightweight. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0091] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0092] Figure 2 A schematic diagram of the process flow for identifying the present invention. DETAILED DESCRIPTION

[0093] The technical solutions in the embodiments of the present invention will be described clearly and completely 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, rather than all the embodiments.

[0094] Embodiment 1:

[0095] See also Figure 1 , a control method for assisting human-vehicle interaction based on a lightweight large model, comprising at least the following steps:

[0096] S1: Construction and application of lightweight student model, using T5 model as the basic model, and using it to distill the lightweight T5 student model obtained from the large model, specifically for the intention recognition of driving scenes, to build an intention recognition module, which is used to recognize the user's intention from the user's language input, and may also include the recognition of emotional state. In order to achieve the purpose of improving the quality of control interaction;

[0097] Importance: This is the core of achieving efficient human-vehicle interaction, ensuring the efficient operation and accuracy of the model in the vehicle environment.

[0098] S2: Two-stage pre-training of the model, combining emotion recognition and intent recognition datasets for initial pre-training of the T5 model, and pre-training enhancement through simulated scenarios;

[0099] Importance: Ensure that the model has a high degree of accuracy and adaptability, especially in complex human-vehicle interaction environments.

[0100] S3: Continuous optimization and update mechanism, which continuously collects data in the vehicle system and continuously optimizes and updates parameters through the large model on the cloud server, including at least subsequent model optimization. The subsequent model optimization process includes at least cloud data collection, model optimization and quantitative evaluation of the new model. The subsequent model optimization is a unified improvement for all users in the entire interaction scenario.

[0101] Why it matters: This allows the model to adapt to changing user needs and driving conditions, keeping it at the cutting edge of technology.

[0102] S4: Interaction between the cloud and in-vehicle systems, using the high-capability model on the cloud to compare and optimize with the lightweight student model on the vehicle.

[0103] Importance: This design ensures the real-time and personalized response of the system while maintaining the lightweight of the vehicle system.

[0104] Embodiment 2:

[0105] See also Figure 1 This embodiment is used to further disclose specific steps based on the premise of the above embodiment 1;

[0106] The two-stage pre-training of the model includes at least initial pre-training and pre-training enhancement;

[0107] The initial pre-training uses the T5 model as the basic model. The model is pre-trained for the first time by using a public emotion recognition dataset and an intent recognition dataset. The pre-training enhancement is used to further enhance the initial pre-training.

[0108] Initial pre-training includes at least the following steps:

[0109] First, we will collect and organize a large amount of text data, including data from various languages ​​and fields, to build a large-scale corpus;

[0110] Said corpus contains a variety of different types of text, including at least news articles, social media posts, product reviews, and conversations;

[0111] Subsequently, the T5 model is pre-trained using the data in the corpus so that the model can learn the ability to extract emotions and intentions from text. The model uses the cross entropy loss function as the loss function, as shown in the following formula:

[0112]

[0113] Where θ represents the parameters of the model, and the pre-training process is to minimize the loss function L(θ) by adjusting θ;

[0114] Finally, the model after the initial pre-training is evaluated and optimized, and the performance of the model is measured using standard evaluation indicators, including at least accuracy, recall, and F1 score, as shown in the following formula:

[0115]

[0116]

[0117]

[0118]

[0119] Among them, TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives respectively. According to the evaluation results, the model will be further adjusted and optimized to improve its performance.

[0120] Pre-training enhancement includes at least the following steps:

[0121] By using two large models, one of which simulates the role of the driver and the other is responsible for responding to the driver's requests;

[0122] Whenever the driver character model makes a request, the request is passed to the other large model and the lightweight T5 student model at the same time;

[0123] The outputs of the two models form a bigram, which includes at least the emotional state and the user intention regarding driving needs;

[0124] By comparing the answers of the large model with those of the lightweight T5 student model, the lightweight T5 student model was optimized, and the pre-training enhancement process was achieved through simulation scenarios;

[0125] In this process we did not pre-train P data (y|x), but instead have a dataset with two large examples of model inputs and correct outputs. In this case, the loss function used is as follows:

[0126]

[0127] The following formula is used to minimize the loss:

[0128]

[0129] α is the learning rate, is the gradient of the loss with respect to the parameter θ;

[0130] After two stages of pre-training, the model can enter the deployment phase, after which it can be continuously optimized in driving scenarios.

[0131] Subsequent optimization of the model includes at least the following steps:

[0132] Collect user-vehicle interaction data in driving scenarios. The data content includes user requests, speech recognition results, responses after recognition, and the next user request. Through the next user request, we can predict and judge the response result and whether the previous response meets the user's needs.

[0133] Transmitting the collected data to the cloud means transferring the data stored onboard the vehicle to a remote server;

[0134] A large, advanced model is deployed in the cloud. This model uses user data to generate corresponding responses and compares and optimizes them with a lightweight Student T5 model. This large model has a very high ability to quickly generate accurate and personalized responses. At the same time, the lightweight Student T5 model updates and optimizes its own parameters based on the difference in responses with the large model.

[0135] Different from the pre-training of the model, a dynamic adjustment function is introduced for the continuous optimization process to simulate the real-time update of the cloud model parameters. The formula is as follows:

[0136]

[0137] After the optimization, we use the area under the ROC curve (AUC) to quantify the model performance. The formula is as follows:

[0138]

[0139] Where TPR(d) is the true positive rate, FDR(d) is the false positive rate, and d is the decision threshold;

[0140] After a period of time, the T5 model parameters in the cloud are transmitted to the on-board T5 to update the parameters of the on-board T5 model;

[0141] The T5 lightweight model is deployed between the speech recognition module and the control module, responsible for identifying user emotions and intentions to improve response quality;

[0142] The big model in the cloud generates answers based on the collected user data and compares and optimizes the response of the T5 lightweight model;

[0143] The optimized T5 lightweight model parameters will be transmitted back to the model deployed in the driving scenario to update the parameters of the model in the driving scenario.

[0144] The application of the intent recognition module includes at least the following steps:

[0145] Receiving voice input: The system collects the user's voice commands, converts the voice input into text data through the voice recognition module, and cleans and preprocesses the converted text, such as removing noise and standardizing the text;

[0146] Feature extraction: extract features from the cleaned text for model training and inference, such as word embeddings or other NLP features;

[0147] Model inference: Use a lightweight student model to process features and output the probability distribution of intent prediction;

[0148] Post-processing and response: Select the most likely intent based on the model output and convert it into a command that the interactive system can understand and respond to;

[0149] Through the optimization of this part, it is possible to use a lightweight large vehicle-mounted model to recognize the user's intentions and emotions during the human-vehicle interaction process, so as to improve the recognition rate of user commands in the human-vehicle interaction and improve the quality of the human-vehicle interaction process.

[0150] The model inference process is as follows:

[0151]

[0152] The posterior probability is as above;

[0153]

[0154] The posterior probability is as above;

[0155]

[0156] The posterior probability is as above;

[0157]

[0158] The posterior probability is as above.

[0159] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A control method for assisting human-vehicle interaction based on a lightweight large model, characterized in that: At least the following steps are included: S1: Construction and application of lightweight student model. Using the T5 model as the basic model, the lightweight T5 student model obtained by distilling it from the large model is used to build an intention recognition module specifically for intention recognition in driving scenarios. The intention recognition module is used to recognize the user's intention from the user's language input, and also includes the recognition of emotional state, so as to achieve the purpose of improving the quality of control interaction; S2: Two-stage pre-training of the model, combining emotion recognition and intent recognition datasets for initial pre-training of the T5 model, and pre-training enhancement through simulated scenarios; S3: Continuous optimization and update mechanism, which continuously collects data in the vehicle system and continuously optimizes and updates parameters through the large model on the cloud server, including at least subsequent model optimization, which includes at least cloud data collection, model optimization and quantitative evaluation of the new model, and the subsequent model optimization is a unified improvement for all users in the entire interaction scenario; The subsequent optimization of the model includes at least the following steps: Collect user-vehicle interaction data in driving scenarios. The data content includes user requests, speech recognition results, responses after recognition, and the next user request. Through the next user request, we can predict and judge the response result and whether the previous response meets the user's needs. Transmitting the collected data to the cloud means transferring the data stored onboard the vehicle to a remote server; A large, advanced model is deployed in the cloud. This model uses user data to generate corresponding responses and compares and optimizes them with a lightweight Student T5 model. This large model has a very high ability to quickly generate accurate and personalized responses. At the same time, the lightweight Student T5 model updates and optimizes its own parameters based on the difference in responses with the large model. Different from the pre-training of the model, a dynamic adjustment function is introduced for the continuous optimization process to simulate the real-time update of the cloud model parameters. The formula is as follows: , Among them, θ t+1 These are the model parameters at time step or iteration t+1; θ t These are the model parameters at time step or iteration t; η t This is the learning rate at time step or iteration t, a positive scalar that controls the parameter update step size; ∇ θ J(θ t ) is the gradient of the loss function J with respect to the parameter θ at time step t, which represents the local slope or derivative of the loss function with respect to the parameter; ϵ t This is the noise term added at time step t, which is usually used to help the model jump out of the local minimum or add randomness to the training process; After the optimization, we use the area under the ROC curve to quantify the model performance. The formula is as follows: , Where TPR(d) is the true positive rate, FDR(d) is the false positive rate, and d is the decision threshold; After a period of time, the T5 model parameters in the cloud are transmitted to the on-board T5 to update the parameters of the on-board T5 model; The T5 lightweight model is deployed between the speech recognition module and the control module, responsible for identifying user emotions and intentions to improve response quality; The big model in the cloud generates answers based on the collected user data and compares and optimizes the response of the T5 lightweight model; The optimized T5 lightweight model parameters will be transmitted back to the model deployed in the driving scenario to update the parameters of the model in the driving scenario; S4: Interaction between the cloud and in-vehicle systems, using the high-capability model on the cloud to compare and optimize with the lightweight student model on the vehicle.

2. The control method for assisting human-vehicle interaction based on a lightweight large model according to claim 1, characterized in that: The two-stage pre-training of the model includes at least initial pre-training and pre-training enhancement; The initial pre-training adopts the T5 model as a basic model, and the model is initially pre-trained by using a public emotion recognition dataset and an intention recognition dataset. The pre-training enhancement is used to further enhance the initial pre-training.

3. The control method for assisting human-vehicle interaction based on a lightweight large model according to claim 2 is characterized in that: The initial pre-training at least comprises the following steps: First, we will collect and organize a large amount of text data, including data from various languages ​​and fields, to build a large-scale corpus; Said corpus contains a variety of different types of text, including at least news articles, social media posts, product reviews, and conversations; Subsequently, the T5 model is pre-trained using the data in the corpus so that the model can learn the ability to extract emotions and intentions from text. The model uses the cross entropy loss function as the loss function, as shown in the following formula: , L(θ) is the loss function itself, usually represented by L(.), where θ represents the model parameters. The purpose of the loss function is to measure the performance of the model. Represents the parameters of the model; p data (x,y) is the true distribution of the data, which indicates the probability of the correct label y given the input x; p model (y|x;θ) is the conditional probability distribution predicted by the model, which represents the probability of predicting the correct label y given the input x under the model parameters θ; The pre-training process is to adjust To minimize the loss function L( ); Finally, the model after the initial pre-training is evaluated and optimized, and the performance of the model is measured using standard evaluation indicators, including at least accuracy, recall, and F1 score, as shown in the following formula: , Among them, TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives respectively. According to the evaluation results, the model will be further adjusted and optimized to improve its performance; Accuracy is a measure of the proportion of correct predictions made by the classifier, that is, the proportion of all correct predictions to all predictions; Recision is a measure of the proportion of positive classes that are actually predicted to be positive, also known as the precision rate; Recall is a measure of the proportion of actually positive classes that are correctly predicted to be positive, also known as recall or sensitivity. It is the harmonic mean of precision and recall, and is used when we need to consider the balance between precision and recall.

4. The control method for assisting human-vehicle interaction based on a lightweight large model according to claim 3 is characterized by: The pre-training enhancement comprises at least the following steps: By using two large models, one of which simulates the role of the driver and the other is responsible for responding to the driver's requests; Whenever the driver character model makes a request, the request is passed to the other large model and the lightweight T5 student model at the same time; The outputs of the two models constitute a bigram, which includes at least the emotional state and the user intention regarding the driving demand; By comparing the answers of the large model with those of the lightweight T5 student model, the lightweight T5 student model was optimized, and the pre-training enhancement process was achieved through simulation scenarios; We did not pre-train in this process , but instead there is a dataset with two large examples of model inputs and correct outputs. In this case, the loss function used is as follows: , Where N is the total number of samples; i is the index, used to represent the i-th sample; x (i) is the input feature representing the i-th sample; y (i) represents the actual label of the i-th sample; P model is the probability distribution function predicted by the model, which gives the probability of predicting the output y given the input x and the model parameters θ; The following formula is used to minimize the loss: , in, gradient; θ new is the updated model parameter; θ old are the model parameters before updating; After two stages of pre-training, the model can enter the deployment stage.

5. The control method for assisting human-vehicle interaction based on a lightweight large model according to claim 1, characterized in that: The application of the intention recognition module includes at least the following steps: Receiving voice input: The system collects the user's voice commands, converts the voice input into text data through the voice recognition module, and cleans and preprocesses the converted text; Feature extraction: extract features from the cleaned text for model training and inference; Model inference: Use a lightweight student model to process features and output the probability distribution of intent prediction; Post-processing and response: Select the most likely intent based on the model output and convert it into a command that the interactive system can understand and respond to; Through the optimization of this part, it is possible to use a lightweight large vehicle-mounted model to recognize the user's intentions and emotions during the human-vehicle interaction process, so as to improve the recognition rate of user commands in the human-vehicle interaction and improve the quality of the human-vehicle interaction process.

6. The control method for assisting human-vehicle interaction based on a lightweight large model according to claim 1, characterized in that: The model inference process is as follows: , The posterior probability is as above; , The posterior probability is as above; , The posterior probability is as above; , The posterior probability is as above.