AI partner system and method based on personal data

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an AI partner system and method based on personal data, and the system comprises a terminal and a server which are in communication connection through an encrypted network. The terminal is provided with a personal basic data acquisition module, a dynamic data acquisition module and a result feedback module; the server comprises a data processing module, an intelligent agent processing module, a passive / active response module and a parameter optimization module. Static preference and multi-source dynamic sensing data of a user are collected through multiple channels, desensitization, feature extraction and edge calculation are locally completed on a terminal, and privacy security is guaranteed. And the server side fuses the multi-modal data to generate a comprehensive state vector, instantiates an independent AI agent for each user, and supports passive response and active care dual-mode interaction. And based on user dominant / implicit feedback, incremental optimization is performed on the personal model by adopting a parameter efficient fine tuning technology, so that continuous personalized evolution is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an AI companion system and method based on personal data. Background Technology

[0002] With breakthroughs in generative AI and large language models, intelligent dialogue systems have evolved from simple question-and-answer tools to virtual assistants. However, existing AI interaction systems still have many key limitations, making it difficult to achieve deep personalization and human-like companionship.

[0003] Existing systems generally lack the ability to integrate multi-dimensional data, relying solely on the context of a single conversation or limited explicit preferences. This fails to construct a complete digital profile covering both long-term static traits and short-term dynamic intentions, resulting in superficial AI understanding of users. Traditional AI models are mostly static and general-purpose, and even after fine-tuning, they are often geared towards specific groups. They struggle to achieve personalized and continuously evolving adaptations, failing to evolve with changes in user interests, knowledge structures, and life stages, thus lacking growth potential and adaptability.

[0004] Meanwhile, existing systems primarily rely on passive responses, lacking proactive care capabilities based on a deep understanding of user states, making it difficult to provide forward-looking services in scenarios such as health management and emotional support. Furthermore, some systems suffer from issues such as limited data collection dimensions, insufficient multi-terminal collaboration and adaptation, and inadequate privacy protection mechanisms, further restricting the application value of AI as a companion. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides an AI companion system and method based on personal data.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution: an AI companion system based on personal data, comprising a terminal and a server, wherein the terminal and the server are connected via encrypted network communication, the terminal includes a personal basic data collection module, a personal dynamic data collection module, and a result feedback module, for data collection and feedback capture, and the server includes a data processing module, an intelligent agent processing module, a passive response module, an active response module, and a parameter optimization module, for data processing, model inference, interactive response, and parameter optimization;

[0009] The outputs of the personal basic data collection module and the personal dynamic data collection module are both connected to the input of the data processing module;

[0010] The output of the data processing module is connected to the input of the agent processing module. The output of the agent processing module is connected to the input of the passive response module and the active response module, respectively. The outputs of the passive response module and the active response module are both connected to the terminal. The output of the result feedback module is connected to the input of the parameter optimization module. The output of the parameter optimization module is connected to the input of the agent processing module.

[0011] Preferably, the personal basic data collection module includes a multi-channel collection unit, a localized preprocessing engine, and an encrypted transmission unit;

[0012] The multi-channel data collection unit collects data through user registration guidance, historical data import, device information reading, and initial interactive Q&A. The collected data includes personal identity data, long-term preference and trait data, historical records, and knowledge base data.

[0013] The localized preprocessing engine performs desensitization, standardization, and tagging on sensitive data, and the encrypted transmission unit uploads the processed data to the data processing module through an encrypted channel.

[0014] More preferably, the personal dynamic data acquisition module includes a multi-source interface adaptation unit, a lightweight edge computing unit, a cross-device collaboration unit, and a low-latency synchronization unit;

[0015] The multi-source interface adaptation unit includes system-level interfaces, application-level interfaces, and protocol-level interfaces, which respectively access the terminal system's native API, third-party software open interfaces, and IoT communication protocols; the cross-device collaboration unit includes a user identity association unit, a data timing alignment unit, and a priority scheduling unit.

[0016] Preferably, the system-level interface interfaces with Android's SensorManager and iOS's CoreMotion native API to obtain raw data from the accelerometer, gyroscope, heart rate sensor, and GPS positioning module; the protocol-level interface supports Bluetooth BLE, Wi-Fi, and NFC protocols to interface with IoT sensors to obtain environmental context data.

[0017] Preferably, the lightweight edge computing unit performs data processing locally on the terminal: it uses Kalman filtering and moving average algorithms to remove random noise; it extracts key features such as speech tone, gait features, and emotional intensity; it detects specific events such as emotional fluctuations and sleep state switching; and it transforms unstructured raw data into standardized feature vectors.

[0018] More preferably, the data processing module adopts a three-layer serial architecture, specifically including:

[0019] Data access and cleaning layer: performs format parsing, outlier detection, and redundancy removal on encrypted data streams;

[0020] Multimodal fusion and context reconstruction layer: Generates user, context, and behavioral event sequences through timestamp alignment and spatial association techniques;

[0021] Feature engineering and vectorization layer: Extract high-dimensional semantic features and convert them into numerical vectors, then weight and concatenate long-term static features and real-time dynamic features to output a comprehensive state vector.

[0022] Preferably, the agent processing module is a parallelized, containerized model hosting and inference platform, including a model instantiation and resource scheduling unit, a personalized inference and decision engine, and a dynamic memory management unit.

[0023] The model instantiation unit allocates an independent running environment to each user and loads a pre-trained multimodal base model and a dedicated parameter adapter.

[0024] The dynamic memory management unit allocates a dedicated vectorized memory library to each model to store short-term context and long-term key events.

[0025] Preferably, the parameter optimization module employs efficient parameter fine-tuning and incremental learning techniques to maintain lightweight adaptation parameters for each user and run them in combination with a shared basic model;

[0026] The parameter optimization module receives explicit and implicit feedback from the result feedback module, processes it, and then performs targeted incremental training on the personal intelligent agent model. The optimization processes of each user are isolated from each other to avoid data leakage.

[0027] A further preferred method for creating an AI companion based on personal data includes the following steps:

[0028] S1: Initial data collection and preprocessing. Basic user data is collected through multiple channels, and after local de-identification and standardization, it is encrypted and uploaded to the server.

[0029] S2: Dynamic data acquisition and integration. By adapting to multiple source interfaces, the sensor data channels are opened up. After edge computing processing, multi-terminal data integration is achieved through cross-device collaborative units.

[0030] S3: Data fusion and vectorization. The data processing module cleans and fuses multimodal data to generate a comprehensive state vector and stores it in isolation.

[0031] S4: Personal intelligent agent instantiation and reasoning, assigning an independent model to each user, and generating personalized response strategies based on comprehensive state vectors;

[0032] S5: Dual-mode interactive response, the passive response module responds to explicit user commands, and the active response module initiates interaction autonomously, both outputting results through an encrypted channel;

[0033] S6: Feedback Acquisition and Model Optimization. Capture multimodal user feedback and perform incremental parameter optimization on the personal intelligent agent model based on the feedback.

[0034] In step S2, multiple terminal devices are bound through the user identity identification association unit, the data time sequence alignment unit synchronizes data based on timestamps, the priority scheduling unit allocates collection resources according to importance, and the core data is preferentially transmitted to the server through WebSocket and MQTT protocols.

[0035] (III) Beneficial Effects

[0036] Compared with existing technologies, the present invention provides an AI companion system and method based on personal data, which has the following beneficial effects:

[0037] This technical solution collects basic user data, such as identity, preferences, historical knowledge, and multi-source dynamic data, such as sensor data, behavior data, and environmental context data, through multiple channels on the terminal side. It then performs anonymization, feature extraction, and edge computing locally, effectively protecting user privacy and reducing transmission load. The server-side adopts a three-layer data processing architecture to achieve multimodal data fusion and comprehensive state vectorization, building a precise contextual profile for each user.

[0038] The intelligent agent processing module instantiates an independent AI model for each user in a containerized manner. Combined with a dedicated memory bank and lightweight adaptation parameters, it achieves highly personalized inference while ensuring data isolation. The system supports dual-mode interaction: passively responding to user commands and proactively initiating care or suggestions based on contextually predicted needs, enhancing the sense of companionship and practicality. The parameter optimization module utilizes explicit and implicit feedback, employing efficient parameter fine-tuning and incremental learning techniques to continuously optimize the individual model, achieving adaptive evolution that becomes increasingly intuitive with use.

[0039] The entire communication chain employs encrypted communication, ensuring that sensitive data does not leave the terminal. The model training process is completely isolated between users, complying with privacy regulations such as GDPR. A cross-device collaboration mechanism ensures data timing alignment and priority scheduling across multiple terminals, guaranteeing a low-latency and highly consistent user experience.

[0040] This technical solution breaks through the limitations of traditional general-purpose AI assistants that are all the same, and truly realizes a safe, considerate, and growable personalized digital companion, which is suitable for high-value scenarios such as mental health support, elderly care, and educational guidance. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0042] Figure 2 This is a schematic diagram of the process flow of the method of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see Figure 1-2 This invention is an AI companion system based on personal data, comprising a terminal and a server. The terminal and the server are connected via encrypted network communication. The terminal includes a personal basic data collection module, a personal dynamic data collection module, and a result feedback module for data collection and feedback capture. The server includes a data processing module, an intelligent agent processing module, a passive response module, an active response module, and a parameter optimization module for data processing, model inference, interactive response, and parameter optimization.

[0045] The outputs of the personal basic data collection module and the personal dynamic data collection module are both connected to the input of the data processing module;

[0046] The output of the data processing module is connected to the input of the agent processing module. The output of the agent processing module is connected to the input of the passive response module and the active response module, respectively. The outputs of the passive response module and the active response module are both connected to the terminal. The output of the result feedback module is connected to the input of the parameter optimization module. The output of the parameter optimization module is connected to the input of the agent processing module.

[0047] The working principle of this technical solution is dual-end collaboration, data closed loop, and personalized models for each user: the terminal focuses on multi-dimensional data collection and feedback capture, while the server is responsible for in-depth data processing, dedicated model inference, and interactive response. Secure data flow between the two ends is achieved through an encrypted network. A complete user digital profile is constructed using static basic data and dynamic sensor data, and a high-dimensional state vector is generated through layered processing. An independent and continuously evolving personal intelligent agent model is instantiated for each user, satisfying both immediate and potential user needs through passive and active response dual-mode interaction. Finally, user feedback drives incremental model optimization, forming a closed-loop process of collection, processing, inference, interaction, feedback, and optimization, achieving deep personalization and growth potential for the AI ​​companion. The entire solution is based on mature sensing technology, AI model framework, and data security mechanisms, making it highly feasible.

[0048] As the core of system data initialization, the personal basic data collection module integrates long-term stable static data of users through multiple channels. After local preprocessing and encrypted transmission, it provides the initial digital genes for the personal intelligent agent model, ensuring that the model has the user's unique characteristics from the initial stage.

[0049] Multi-channel data collection unit: Through user registration guidance, such as filling in identity and preference information, importing historical data, such as authorizing the import of social media accounts, cloud document data, and reading device information such as terminal model, system version, and initial interactive Q&A, such as collecting values ​​and lifestyle habits through targeted questions, four channels are used to comprehensively collect personal identity data, long-term preference and trait data, historical records and knowledge base data, covering the user's static core dimensions.

[0050] Localized preprocessing engine: Sensitive data such as ID card numbers and home addresses are anonymized using algorithms, such as replacing some fields with asterisks. Raw data in different formats, such as text, tables, and images, are standardized and tagged according to identity, preferences, and history to ensure a consistent data structure.

[0051] Encrypted transmission unit: Data is uploaded using an HTTPS+AES-256 encrypted channel to prevent data leakage during transmission and comply with relevant personal information protection regulations.

[0052] The multi-channel data collection logic is based on the existing APP registration and data authorization process, which is easy to develop; the localized preprocessing engine can reuse mature de-identification and standardized algorithm libraries; the encrypted transmission adopts the industry-standard HTTPS and AES encryption schemes, and developers can directly integrate the relevant SDKs without any technical barriers.

[0053] The personal dynamic data acquisition module serves as a real-time sensing sensor for the system. It integrates the multi-source sensing capabilities of terminal devices, third-party software, and IoT sensors to continuously capture user dynamic behavior, physiological state, and environmental information. After local preprocessing, the data is synchronized to the server to provide fresh data input for real-time model inference.

[0054] Multi-source interface adapter unit:

[0055] System-level interfaces: Connect to Android's SensorManager and iOS's CoreMotion native API to directly obtain raw hardware sensor data such as accelerometer gait data, gyroscope motion data, heart rate sensor physiological data, and GPS positioning module location data, adapting to mainstream mobile terminal systems.

[0056] Application-level interfaces: Access open authorized interfaces of social apps, health software, and office tools to capture software-level data such as user application usage time, social interaction trajectory, and health monitoring records.

[0057] Protocol-level interfaces: Supports Bluetooth BLE, Wi-Fi, and NFC communication protocols, and can interface with smart home sensors, such as temperature and humidity sensors, human infrared sensors, and wearable devices, such as smartwatches and wristbands, to obtain environmental context and physiological supplementary data.

[0058] Lightweight edge computing unit: Data preprocessing is completed locally on the terminal without uploading raw data: Sensor noise, such as signal interference in heart rate data, is removed through Kalman filtering and moving average algorithms; key features such as voice tone, gait frequency, and emotional intensity are extracted; emotional fluctuations, such as sudden changes in voice emotion, sleep state switching, and specific events such as prolonged periods without data from motion sensors, are detected; unstructured voice streams and motion sequences are transformed into standardized feature vectors, reducing transmission bandwidth consumption and privacy risks.

[0059] Cross-device collaboration unit:

[0060] User identification association unit: Through account binding, device fingerprinting, or hardware serial number matching, the user's mobile phone, smartwatch, smart home and other terminal devices are associated to form a unified data collection network.

[0061] Data time-series alignment unit: Based on high-precision timestamps, it synchronizes sensor data from different devices, such as mobile phone interaction behavior data and smartwatch heart rate data, aligned along the timeline to ensure context integrity.

[0062] Priority scheduling unit: Data is classified according to importance, such as physiological abnormality data > real-time interactive data > environmental data. Collection resources are dynamically allocated, and core data is transmitted first through WebSocket and MQTT protocols to ensure the real-time nature of critical information.

[0063] Low-latency synchronization unit: Using WebSocket, such as real-time interactive data, and MQTT, such as low-latency protocol for batch sensor data, the preprocessed feature vectors are synchronized to the server, with transmission latency controlled within 100ms.

[0064] The multi-source interface adaptation unit can reuse the terminal system's native APIs and third-party open interfaces. Development only requires completing interface calls and data format adaptation. The lightweight edge computing unit can integrate open-source filtering and feature extraction algorithm libraries, which are fully supported by the terminal's computing power. The cross-device collaboration unit has simple identity association and timing alignment logic, and uses mature protocols for priority scheduling and low-latency transmission, which can be directly implemented.

[0065] The results feedback module captures user feedback signals in real time to the system interaction response, such as explicit and implicit signals. After preprocessing, these signals are synchronized to the parameter optimization module on the server side, providing a direct basis for incremental optimization of the personal intelligent agent model.

[0066] Explicit feedback capture: By setting up entry points for likes / dislikes, star ratings, and text / voice feedback through the terminal interface, we can directly collect users' explicit evaluations of the response results.

[0067] Implicit feedback capture: By monitoring terminal behavior, obtain indirect feedback signals such as user's ignoring / replying to / following up on messages, duration of dwell time on recommended content, and whether to execute system suggestions.

[0068] Feedback preprocessing: Clean the raw feedback data to remove erroneous data, standardize and unify the scoring range, and label and associate the corresponding interaction events to ensure the validity of the feedback data.

[0069] Explicit feedback entry points can be implemented through conventional UI component development; implicit feedback capture can be achieved using a reusable terminal behavior statistics SDK; the preprocessing logic is simple, requiring only basic data cleaning and format conversion, making development difficult.

[0070] The data processing module receives multimodal data, such as static and dynamic data, transmitted from the terminal. After hierarchical processing, it transforms the data into a structured, time-series high-dimensional comprehensive state vector, providing high-quality data input for personalized reasoning of the personal intelligent agent model.

[0071] Data access and cleaning layer: Receives encrypted data streams from terminals, performs format parsing (compatible with formats such as JSON and Protocol Buffers), outlier detection, removes data exceeding reasonable ranges (e.g., heart rate > 200 bpm), fills missing values ​​using mean / median, and removes duplicate data to ensure data integrity and consistency.

[0072] Multimodal fusion and context reconstruction layer: Apply timestamp alignment to synchronize the time dimension and spatial correlation of different types of data, associate location data with environmental data and semantic mapping, map text, voice and physiological signals into a unified semantic space technology, and fuse discrete multimodal data into user, context and behavioral event sequences, such as user in the office: context, initiating work consultation: behavior, stable heart rate: state, reconstructing the complete context graph of user activities.

[0073] Feature Engineering and Vectorization Layer: Utilizing pre-trained language models, such as BERT and domain knowledge graphs, high-dimensional semantic features, such as user intent, emotional intensity, and behavioral patterns, are extracted from the fused data and transformed into numerical vectors. Through a weighted concatenation algorithm, long-term static features, such as value orientation, are fused with real-time dynamic features, such as current emotional state, to output a comprehensive state vector that combines historical depth and immediacy.

[0074] The data access and cleaning layer can reuse mature data processing frameworks such as Spark and Flink; the timestamp alignment and spatial association of the multimodal fusion and context reconstruction layer are conventional data processing logic, and the semantic mapping can reuse pre-trained models; the feature engineering and vectorization layer can integrate open-source feature extraction algorithms and pre-trained models, which are fully supported by server computing power, and developers can implement them according to standard procedures.

[0075] The agent processing module, as the core reasoning unit of the system, adopts a parallel and containerized architecture. It instantiates an independent personal agent model for each user, completes personalized reasoning based on the comprehensive state vector, and generates response strategies that conform to the user's characteristics and current needs.

[0076] Model instantiation and resource scheduling unit: Utilizes containerized scheduling platforms such as Kubernetes to dynamically allocate independent model runtime environments, such as CPU / GPU resources, to each user; Each model instance uses a pre-trained multimodal base model, such as the GPT series of large language models or CLIP vision-language models, as a general cognitive foundation, and loads a user-specific parameter adapter and personalized weights to avoid model redundancy and reduce storage overhead.

[0077] Personalized Reasoning and Decision Engine: Receives a comprehensive state vector output from the data processing module and, through the model's proprietary parameters, completes intent recognition (user need assessment), emotion inference (interpretation of user emotional state), and context-related memory retrieval (calling historical interaction data from a dynamic memory bank). Ultimately, it generates a response strategy, which includes not only text / voice replies but also action suggestions, such as a 10-minute rest, resource recommendations (like links to related articles), and virtual avatar interaction commands (like facial expressions and actions).

[0078] Dynamic Memory Management Unit: Each individual intelligent agent model is allocated a dedicated vectorized memory bank, using a vector database such as Milvus to store: short-term context, such as key information in the current dialogue, long-term key events, such as important life milestones of the user, preference changes, and decision feedback results, such as the user's evaluation of past responses, ensuring that the model remains consistent in multi-turn dialogues and continuously accumulating user-related knowledge.

[0079] Containerized scheduling platforms (Kubernetes) are a common industry solution that can be deployed directly; pre-trained multimodal base models can be selected from open-source models such as Llama, ChatGLM, or commercial APIs; dynamic memory can be integrated with mature vector databases, and developers only need to complete the integration of model loading, inference logic and memory calls, making the technical solution mature and controllable.

[0080] The passive response module focuses on responding to explicit user interaction requests, such as questions and commands. Based on the exclusive parameters and dynamic memory of the individual intelligent agent model, it generates accurate and personalized real-time responses to meet the user's proactive needs.

[0081] It continuously listens for user commands from the terminal, such as text, voice, and touch operations, and receives structured request data transformed by the data processing module, such as merging user profiles and real-time context. It then schedules the corresponding personal intelligent agent model based on the user's identity and generates customized response content through model inference. For example, if a user asks "recommend a suitable travel destination for me", the model combines the user's long-term preferences and current location to generate a personalized recommendation, which is then fed back to the terminal through an encrypted channel.

[0082] Request listening can be implemented through WebSocket long connections. The model scheduling logic is simple, based on user ID matching. Response generation is completed by the personal intelligent agent model. Development only requires completing the interface connection between modules, without any complex technical difficulties.

[0083] The proactive response module does not rely on user requests. By continuously analyzing the comprehensive state vector, it autonomously initiates interactions based on the user's historical preferences and real-time status, such as care reminders and topic recommendations, to achieve human-like proactive companionship.

[0084] Periodically, such as hourly, or based on specific events, such as detecting a user's low mood or arriving at a frequently visited gym, an internal inference request is sent to the corresponding personal intelligent agent model. The model uses dedicated parameters and a dynamic memory bank to analyze the user's current state, such as mood, location, and historical preferences, such as fitness habits and hobbies, to generate proactive interactive decisions, such as initiating comforting topics when the user is in a low mood or reminding them of exercise precautions when at a gym. The decisions are then transformed into natural interactive behaviors, such as text messages, voice reminders, and push notifications, which are pushed to the terminal through an encrypted channel.

[0085] Periodic triggering can be implemented through a timed task framework such as Quartz. Event triggering is based on the event detection results of the edge computing unit. The interactive behavior generation and passive response module reuse the same set of output logic, which is easy to develop and can be quickly integrated.

[0086] The parameter optimization module, as the core of system evolution, receives user feedback data and uses efficient parameter fine-tuning technology to perform targeted incremental training on the personal intelligent agent model, ensuring that the model continues to evolve with user interaction and adapts to changes in user needs.

[0087] Feedback reception and processing: Receive explicit feedback from the result feedback module, such as ratings and evaluations, and implicit feedback, such as behavioral data. Transform the feedback signals into reward / loss values ​​that the model can recognize, such as recording a like as a positive reward and ignoring as a negative loss.

[0088] Efficient parameter fine-tuning: Employing lightweight fine-tuning techniques such as LoRA (Low-Rank Adaptation), it does not maintain a complete model copy for each user. Instead, it trains and maintains only a set of lightweight adaptation parameters, which are 1%-5% the size of the complete model. During inference, it is combined with a shared pre-trained base model, greatly reducing storage and computational overhead.

[0089] Isolation optimization and privacy protection: The model optimization process of each user is isolated from each other, and data and parameters are not shared to avoid leakage of user privacy; the optimized parameters only update the dedicated adapter of the corresponding personal intelligent agent model, without affecting the base model and other user models.

[0090] The LoRa (Local Area Reduction) technique has mature open-source implementations, such as HuggingFace's peft library, which developers can directly integrate. The feedback signal conversion logic is simple, and isolation optimization can be achieved through user ID partitioning, which fully meets the feasibility requirements.

[0091] Detailed Workflow

[0092] Initialization phase: basic data collection and model building;

[0093] Users install the system client on their terminals, complete the registration process, and authorize the system to collect basic data.

[0094] The personal basic data collection module acquires static data such as user identity, long-term preferences, and historical records through multi-channel collection units;

[0095] The localized preprocessing engine desensitizes, standardizes, and tags sensitive data, while the encrypted transmission unit uploads data to the server via an HTTPS+AES channel.

[0096] The server-side data processing module cleans and vectorizes the basic data to generate an initial comprehensive state vector.

[0097] The agent processing module allocates an independent operating environment for users, loads the pre-trained multimodal base model and the initial dedicated parameter adapter, constructs a personal agent model instance, and initializes the dynamic memory library.

[0098] Operational phase: Dynamic data acquisition and dual-mode interaction;

[0099] The personal dynamic data acquisition module continuously captures real-time behavioral interaction data of users through a multi-source interface adaptation unit, such as dialogue, application usage, physiological and status data, such as heart rate and emotion, and environmental context data, such as location, temperature and humidity.

[0100] The lightweight edge computing unit performs data filtering, feature extraction, and event detection locally on the terminal, transforming them into standardized feature vectors.

[0101] The cross-device collaboration unit associates data from multiple terminals through identity identifiers, the time sequence alignment unit synchronizes data timestamps, the priority scheduling unit allocates resources according to importance, and core data is synchronized to the server via WebSocket / MQTT protocol;

[0102] The data processing module receives dynamic data, merges it with basic data, and generates a real-time comprehensive state vector through a three-layer architecture.

[0103] The agent processing module, based on a comprehensive state vector and dynamic memory, completes personalized reasoning.

[0104] If a clear user instruction is detected, such as a question or operation, the passive response module generates a precise response and pushes it to the terminal.

[0105] If no user actively requests it, the proactive response module will initiate interactions based on the user's status and preferences, such as care reminders and topic recommendations, and push them to the terminal.

[0106] Optimization phase: Feedback collection and model iteration;

[0107] The results feedback module captures explicit user feedback on interactive responses in real time, such as likes and comments, as well as implicit feedback, such as replies and dwell time.

[0108] After preprocessing, the feedback data is synchronized to the parameter optimization module on the server through an encrypted channel;

[0109] The parameter optimization module converts the feedback signal into reward / loss values ​​and uses LoRA technology to incrementally train the personalized adaptation parameters of the personal agent model.

[0110] The optimized parameters are updated to the corresponding personal intelligent agent model, and the dynamic memory bank synchronously stores the interaction and feedback information to complete the model evolution.

[0111] The system continuously repeats the dynamic data collection, dual-mode interaction, and feedback optimization process to achieve continuous model growth and personalized adaptation.

[0112] Device replacement / expansion phase: cross-device collaboration and data synchronization;

[0113] When a user installs the client on a new terminal and logs in with the same account, the personal dynamic data collection module automatically associates the new device with the existing device through the user's identity identifier association unit.

[0114] The new device connects to local sensing capabilities through a multi-source interface adapter unit, forming a collaborative data acquisition network with the existing device.

[0115] The server synchronizes the user's personal intelligent agent model, dynamic memory, and adaptation parameters to the operating environment of the new device to ensure a consistent user experience across multiple devices.

[0116] The cross-device collaboration unit continuously synchronizes data from multiple terminals to ensure context integrity and interaction continuity.

[0117] All modules and preferred technologies in this technical solution are based on existing mature technologies and industry standards, making it highly feasible. Specifically, this is reflected in:

[0118] The core technologies are mature and controllable: multi-source interface adaptation, edge computing, data encryption, efficient parameter fine-tuning (LoRA), containerized deployment and other core technologies are all industry-standard solutions with mature open source libraries, SDKs and deployment tools that developers can directly integrate and reuse.

[0119] Wide hardware compatibility: Supports mainstream smartphones (iOS / Android), wearable devices, and IoT sensors. The interface logic is standardized, requiring no customized development, and can cover the vast majority of user terminal scenarios.

[0120] Low development and deployment costs: Lightweight design, such as edge computing preprocessing and efficient parameter fine-tuning, reduces terminal computing power consumption and server storage / computing overhead; containerized deployment supports elastic scaling to adapt to different user volumes, and operation and maintenance costs are controllable.

[0121] Privacy and security are guaranteed: Multiple privacy protection measures are adopted, including data anonymization, encrypted transmission, local preprocessing, and user data isolation, which comply with the requirements of the Personal Information Protection Law and other relevant regulations, and there is no compliance risk.

[0122] Highly adaptable and scalable: Standardized interfaces between modules support the addition of new sensing devices, expansion of data types, and upgrade of model architecture, flexibly adapting to various scenarios such as emotional companionship, health monitoring, and personalized education.

[0123] In summary, this technical solution is fully feasible in terms of technology selection, architecture design, development and deployment, and can be quickly transformed into actual products and put into market application.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI companion system based on personal data, characterized in that, It includes a terminal and a server, which are connected via encrypted network communication. The terminal includes a personal basic data collection module, a personal dynamic data collection module, and a result feedback module for data collection and feedback capture. The server includes a data processing module, an intelligent agent processing module, a passive response module, an active response module, and a parameter optimization module for data processing, model inference, interactive response, and parameter optimization. The outputs of the personal basic data collection module and the personal dynamic data collection module are both connected to the input of the data processing module; The output of the data processing module is connected to the input of the agent processing module. The output of the agent processing module is connected to the input of the passive response module and the active response module, respectively. The outputs of the passive response module and the active response module are both connected to the terminal. The output of the result feedback module is connected to the input of the parameter optimization module. The output of the parameter optimization module is connected to the input of the agent processing module.

2. The AI ​​companion system based on personal data according to claim 1, characterized in that, The personal basic data collection module includes a multi-channel collection unit, a local preprocessing engine, and an encrypted transmission unit; The multi-channel data collection unit collects data through user registration guidance, historical data import, device information reading, and initial interactive Q&A. The collected data includes personal identity data, long-term preference and trait data, historical records, and knowledge base data. The localized preprocessing engine performs desensitization, standardization, and tagging on sensitive data, and the encrypted transmission unit uploads the processed data to the data processing module through an encrypted channel.

3. The AI ​​companion system based on personal data according to claim 1, characterized in that, The personal dynamic data acquisition module includes a multi-source interface adaptation unit, a lightweight edge computing unit, a cross-device collaboration unit, and a low-latency synchronization unit. The multi-source interface adaptation unit includes system-level interfaces, application-level interfaces, and protocol-level interfaces, which respectively access the terminal system's native API, third-party software open interfaces, and IoT communication protocols; the cross-device collaboration unit includes a user identity association unit, a data timing alignment unit, and a priority scheduling unit.

4. The AI ​​companion system based on personal data according to claim 3, characterized in that, The system-level interface connects to Android's SensorManager and iOS's CoreMotion native API to obtain raw data from the accelerometer, gyroscope, heart rate sensor, and GPS positioning module; the protocol-level interface supports Bluetooth BLE, Wi-Fi, and NFC protocols to connect to IoT sensors and obtain environmental context data.

5. The AI ​​companion system based on personal data according to claim 3, characterized in that, The lightweight edge computing unit performs data processing locally on the terminal: it uses Kalman filtering and moving average algorithms to remove random noise; and extracts key features such as speech intonation, gait characteristics, and emotional intensity. It detects specific events such as mood swings and sleep state transitions; and transforms unstructured raw data into standardized feature vectors.

6. The AI ​​companion system based on personal data according to claim 1, characterized in that, The data processing module adopts a three-layer serial architecture, specifically including: Data access and cleaning layer: performs format parsing, outlier detection, and redundancy removal on encrypted data streams; Multimodal fusion and context reconstruction layer: Generates user, context, and behavioral event sequences through timestamp alignment and spatial association techniques; Feature engineering and vectorization layer: Extract high-dimensional semantic features and convert them into numerical vectors, then weight and concatenate long-term static features and real-time dynamic features to output a comprehensive state vector.

7. The AI ​​companion system based on personal data according to claim 1, characterized in that, The intelligent agent processing module is a parallelized, containerized model hosting and inference platform, including a model instantiation and resource scheduling unit, a personalized inference and decision engine, and a dynamic memory management unit. The model instantiation unit allocates an independent running environment to each user and loads a pre-trained multimodal base model and a dedicated parameter adapter. The dynamic memory management unit allocates a dedicated vectorized memory library to each model to store short-term context and long-term key events.

8. The AI ​​companion system based on personal data according to claim 1, characterized in that, The parameter optimization module employs efficient parameter fine-tuning and incremental learning techniques to maintain lightweight adaptation parameters for each user and run them in combination with a shared basic model. The parameter optimization module receives explicit and implicit feedback from the result feedback module, processes it, and then performs targeted incremental training on the personal intelligent agent model. The optimization processes of each user are isolated from each other to avoid data leakage.

9. A method for using an AI companion system based on personal data, as described in claims 1-8, characterized in that... Includes the following steps: S1: Initial data collection and preprocessing. Basic user data is collected through multiple channels, and after local de-identification and standardization, it is encrypted and uploaded to the server. S2: Dynamic data acquisition and integration. By adapting to multiple source interfaces, the sensor data channels are opened up. After edge computing processing, multi-terminal data integration is achieved through cross-device collaborative units. S3: Data fusion and vectorization. The data processing module cleans and fuses multimodal data to generate a comprehensive state vector and stores it in isolation. S4: Personal intelligent agent instantiation and reasoning, assigning an independent model to each user, and generating personalized response strategies based on comprehensive state vectors; S5: Dual-mode interactive response, the passive response module responds to explicit user commands, and the active response module initiates interaction autonomously, both outputting results through an encrypted channel; S6: Feedback Acquisition and Model Optimization. Capture multimodal user feedback and perform incremental parameter optimization on the personal intelligent agent model based on the feedback.

10. The AI ​​companion method based on personal data according to claim 9, characterized in that, In step S2, multiple terminal devices are bound through the user identity identification association unit, the data time sequence alignment unit synchronizes data based on timestamps, the priority scheduling unit allocates collection resources according to importance, and core data is preferentially transmitted to the server through WebSocket and MQTT protocols.

Images