Vehicle-mounted psychological counseling service system and method based on large language model

Abstract

The invention discloses a vehicle-mounted AI psychological counseling system based on a large language model, which constructs a lightweight psychological counseling model special for a vehicle-mounted environment through a multi-stage fine tuning and quantitative optimization technology. The model can analyze the emotion of a driver in real time, provide personalized suggestions and dynamic psychological support, effectively relieve driving pressure, reduce traffic accident risks caused by psychological factors, make remarkable breakthrough in the aspects of model size, energy consumption and scene adaptability, and solve the problems of computing power and privacy of a vehicle-mounted environment.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically, the application of natural language processing (NLP) and machine learning in in-vehicle systems. In particular, it relates to an in-vehicle psychological counseling service system and method based on a Large Language Model (LLM), designed to provide real-time psychological support and intervention for drivers, applicable to various driving scenarios. Background Art

[0002] In modern driving scenarios, drivers often face issues such as driving anxiety, road rage, fatigue, and post-accident trauma. These issues not only impact the driving experience but can also lead to accidents. Traditional in-vehicle systems primarily focus on navigation and entertainment functions, lacking intelligent support for the driver's psychological state. Existing solutions often rely on preset voice prompts or simple emotion recognition algorithms, failing to provide professional and personalized psychological counseling.

[0003] Although large-scale language models (LLMs) have demonstrated powerful understanding and generation capabilities in natural language processing and have made significant progress in multiple verticals, directly applying their existing capabilities to in-vehicle psychological counseling services still faces numerous challenges. First, typical in-vehicle computing platforms have relatively limited computing resources, making it difficult to efficiently support the real-time, low-latency inference of existing high-performance, parameter-heavy LLMs, which are crucial for driving scenarios requiring immediate responses. Second, specialized psychological counseling datasets tailored to driving scenarios are extremely scarce. Existing general-purpose datasets fail to fully cover the psychological challenges unique to driving, such as stress, anxiety, and fatigue, and their complex interactions. There is an urgent need to develop highly specialized and scenario-specific datasets. Furthermore, the in-vehicle environment involves a large amount of personal privacy information. Effectively balancing model performance with user data privacy requirements, ensuring data security and compliance, is a key issue that needs to be addressed in the application of large-scale models. Currently, existing technologies focus more on developing general LLM capabilities. However, the application of large language models specifically optimized and deployed for in-vehicle psychological services, particularly those that can provide systematic, lightweight, and high-precision solutions, remains in the exploratory stage and has yet to be fully developed. Therefore, there is an urgent need to develop a lightweight, high-precision specialized model and its deployment method that can overcome the above challenges, operate efficiently in a vehicle environment with limited computing resources, and provide professional psychological support. Summary of the Invention

[0004] The purpose of the present invention is to provide an in-vehicle psychological counseling service system and method based on a large language model. By constructing a dedicated data set and efficient fine-tuning technology, it can provide real-time professional psychological support on the in-vehicle device side, thereby improving driving safety and user experience.

[0005] The technical solution of the present invention includes the following steps: S1. Based on a public psychology dataset, it is formatted into "command-response" pairs through program code, covering common psychological counseling scenarios such as emotion recognition and stress relief, and Qlora is fine-tuned; S2. Paragraphs are extracted from relevant professional books such as "Driving Psychology" to construct a domain dataset, and the domain dataset is used for secondary fine-tuning; S3. Whether the model performance meets the requirements of human dialogue is judged, and a human-computer dialogue is conducted with the fine-tuned model to determine whether the model performance is normal; S4. A simulated dialogue scenario of a driving scenario is constructed, and the fine-tuned model is evaluated in multiple aspects; S5. The model is quantified; S6. The model is deployed on the edge device.

[0006] In the preferred solution, the base model used for fine-tuning S1 and S2 is Qwen2.5-7B.

[0007] In the preferred solution, S2 also includes the following steps: S21, extracting knowledge from professional books and generating driving scenario dialogue data as the second round of fine-tuning data; S22, data cleaning and manual correction: removing redundant content, adjusting the tone to conform to psychological counseling standards, and forming a dedicated data set containing tens of thousands of dialogues.

[0008] In the preferred solution, step S4 specifically includes: using a large language model such as deepseek to simulate and generate driving scenario dialogue questions, simulating multiple rounds of interaction scenarios such as traffic jam anxiety and post-accident psychological intervention.

[0009] In the preferred solution, step S5 specifically includes: merging the fine-tuned QLora adapter weights with the base model, converting to 8-bit GGUF format through . ollama, and compressing the model size to <8GB.

[0010] In the preferred solution, step S6 is: deploying the parameter file in the GGUF format generated in step S5 locally through the ollama library.

[0011] Interaction method: The driver's voice is collected through the vehicle microphone, and the ASR module realizes speech-to-text conversion.

[0012] The present invention has at least the following beneficial effects: (1) Domain adaptability: After two rounds of fine-tuning and professional data sets, the model scored 4.5 points (out of 5 points) in the psychological counseling test of driving scenarios. In comparison, the general models of the same scale, gemma-7B scored 3.25, openchat3.5-7B scored 3.8, mistral-7B scored 2.8, and qwen2.5-7B scored 4.1, indicating that this solution can better understand and alleviate the psychological pressure of users in driving scenarios. (2) Lightweight deployment: After 8-bit quantization, the model only requires 7.6GB of storage space, and the CPU inference speed reaches 8 tokens / s, which meets the requirements of vehicle-mounted equipment; (3) Real-time and privacy protection: Local deployment eliminates cloud dependence, and data is encrypted throughout the process, complying with relevant privacy standards; (4) Multimodal interaction: Combining voice and physiological signals, it provides more natural and personalized psychological support.

[0013] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the invention. DETAILED DESCRIPTION

[0014] The implementation process of the present invention is described below with reference to specific embodiments: First, we generated the first round of training data from open psychology datasets using a custom Alpaca command format conversion script. For the second round of data generation, we extracted passages from eight professional books, including "Driving Psychology," and used DeepSeek to generate approximately 50,000 conversations in scenarios such as traffic anxiety and post-accident stress. After manual correction, we created the final dataset.

[0015] Next, we fine-tuned Qwen2.5-7B using 4-bit QLora on dual GeForce RTX 2080 Ti GPUs, setting rank to 8, and achieved a training loss of 0.32. After merging the weights, we converted the model to 8-bit GGUF format using the .ollama tool, compressing the model to 7.6 GB.

[0016] Finally, the model was deployed on a mobile AMD CPU, and the test scenarios included: Fatigue driving: After simulating 4 hours of continuous driving, the system detected the driver's fatigue and automatically triggered dialogue guidance such as "deep breathing exercises"; Road rage intervention: When the driver frequently used negative words (such as "traffic jam"), the system responded with "Try to pay attention to your breathing rhythm, anger cannot change the road conditions" and played soothing music; Post-accident support: When the driver described a minor rear-end collision, the system provided "Safety Handling Affirmation" and subsequent psychological counseling suggestions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1This is the overall flow chart of the system of the present invention, showing the overall process from data set construction to deployment on edge devices.

[0018] Figure 2 Build a flowchart for the dataset, showing how to transform the relevant book data into a high-quality dataset.

Claims

1. An in-vehicle psychological counseling service system based on a large language model, characterized in that: It includes the following components: a dedicated dataset construction module for constructing a two-round fine-tuning dataset containing psychological consultation dialogues in driving scenarios; The model fine-tuning module uses QLora technology to quantize and fine-tune the pre-trained large language model; The edge device module deploys the fine-tuned model on the vehicle edge device; the multimodal interaction module integrates voice input, text output and physiological signal monitoring functions.

2. The system according to claim 1, wherein: The dedicated dataset construction module includes: a first-round fine-tuning data unit, which formats the public psychology dataset into "command-response" pairs; a second-round fine-tuning data unit, which contains driving psychology knowledge extracted from professional books and generated scenario dialogue data; and a data cleaning unit, which manually corrects and adjusts the tone of the generated data.

3. The system according to claim 1, wherein: The model fine-tuning module specifically includes: the base model uses the Qianwen large language model (Qwen2.5-7B); a 4-bit QLora fine-tuning unit, which adjusts parameters through a low-rank adapter; and a model quantization unit, which converts the fine-tuned model into an 8-bit GGUF format.

4. The system according to claim 1, wherein: The local deployment module meets the following requirements: model size is less than 8GB; pure CPU inference latency is less than 1 second; and inference speed reaches 8 tokens / s.

5. The system according to claim 1, wherein: The multimodal interaction module includes: a voice input unit that collects the driver's voice through an on-board microphone; a customized output unit that provides text replies and soothing music recommendations; and a real-time monitoring unit that integrates heart rate sensor data to dynamically adjust consulting strategies.

6. A vehicle-based psychological consultation service method based on a large language model, characterized in that: The following steps are involved: S1. Fine-tune LoRa using a general cognitive dataset. S2, use the driving psychology related dataset for secondary fine-tuning; S3. Determine whether the model performance meets the requirements for human dialogue; S4. Evaluate the model's performance in multiple dimensions. S5. Quantify the model; S6. Deploy the model on the edge device.

7. The method according to claim 6, characterized in that Step S2 specifically includes: S21: Extract knowledge from professional books and generate driving scenario dialogue data as the second round of fine-tuning data; S22. Manually correct and clean the generated data.

Images