Multi-carrier lightweight personality digital twin system based on behavior logic replication

Abstract

This invention discloses a lightweight, multi-carrier personality digital twin system based on behavioral logic replication, belonging to the interdisciplinary fields of artificial intelligence, digital twins, and service robots. The system comprises five core modules: an AI behavioral logic training module, a structured personality data storage module, a user authorization and data encryption module, a multimodal simulation interaction module, and a multi-carrier adaptive operation module. This invention trains the behavioral logic model through routine user interactions using the LoRA lightweight fine-tuning algorithm, encrypts and stores scalable structured personality data, and incorporates a personality consistency verification mechanism to ensure output matching accuracy. Simultaneously, it sets up a full-process user authorization and end-to-end encryption mechanism to ensure privacy compliance. One personality model can simultaneously drive the operation of multiple carriers, including physical simulation robots, electronic virtual digital humans, and pure voice dialogue software. This invention requires no cutting-edge technological breakthroughs, can be operated independently by ordinary users, and boasts advantages such as lightweight design, low cost, high replicability, multi-scenario adaptability, and strong compliance. It solves the problems of poor implementation, low replicability, high privacy risks, and inconsistent output across multiple carriers in existing technologies, making it suitable for civilian-grade personality digital twin and multi-scenario interactive applications.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of artificial intelligence, digital twins and service robots. Specifically, it relates to a lightweight personality digital twin system based on deep learning of user behavior logic, encrypted storage of structured personality data, and multi-carrier collaborative operation. It is suitable for digital replication and multi-scenario interactive applications of user language style, voice features, behavioral logic and decision preferences in civilian scenarios. Background Technology

[0002] Existing digital personality replication technologies mainly suffer from two core technical flaws that prevent them from simultaneously achieving both widespread civilian adoption and high fidelity: The first category is practicality defects: existing solutions mostly focus on uncommercialized cutting-edge directions such as consciousness uploading and full brain data collection, or require full fine-tuning of large models, which have high technical thresholds, high computing power requirements, and high implementation costs. They require professional technicians to operate and cannot be popularized to the general public without technical background. The second category is technical defects: existing civilian-grade solutions are mainly single-form digital human or voice clone products, which can only replicate the user's appearance or superficial language habits, and cannot achieve deep replication of the user's behavioral logic and decision-making preferences, resulting in low personality matching. At the same time, they generally do not set up full-process user authorization and data encryption mechanisms, which pose legal risks of personal information leakage and non-compliant collection and use of personal data, and cannot guarantee the consistency of personality output when running on multiple carriers, and cannot meet the stable interaction needs of multiple scenarios.

[0003] Currently, there is no civilian-grade system in the industry that can be implemented based on existing mature technologies, with lightweight training, low cost, multi-carrier adaptability, strong compliance, and high-fidelity personality replication under consumer-grade hardware conditions. Summary of the Invention

[0004] I. Technical problems to be solved This invention addresses the shortcomings of existing technologies by resolving the core pain point that current digital personality replication technologies cannot simultaneously achieve high fidelity and widespread civilian use. It also solves the technical problems of poor applicability, limited application scenarios, low replication accuracy, insufficient privacy compliance, and inconsistent output across multiple carriers. The invention provides a digital personality twin system that requires no cutting-edge technological breakthroughs, can be operated independently by ordinary users, features lightweight training, multi-carrier compatibility, and full-process compliant encryption.

[0005] II. Core Technology Solution The system of this invention comprises five core modules, which communicate and operate collaboratively with each other in both directions. A single personality model can drive the synchronous operation of multiple interactive platforms, as detailed below: 1. AI Behavioral Logic Training Module: Relying on mature open-source AI models, this module collects user language content, word preferences, tone features, decision-making logic, and emotional feedback data through daily routine interactions. It then completes model training based on a lightweight low-rank adaptation (LoRA) fine-tuning algorithm to generate a user-specific personality decision-making model. A built-in personality consistency verification submodule extracts text features from a pre-trained language model and uses cosine similarity calculation to verify the match between the model's output content and the user's behavioral logic and decision-making preferences in real time, ensuring that the output content conforms to the user's native interaction characteristics.

[0006] 2. Structured Personality Data Storage Module: Employing end-to-end encryption, this module lightweightly stores 50-300 scalable pieces of structured personality data for each user. Data types include core memories, social relationships, personal behavioral guidelines, decision-making preferences, emotional tendencies, and lifestyle habits. The data supports multi-device synchronization, secure migration, and authorized access, eliminating the need to store the user's entire life data and significantly reducing storage costs and data processing pressure.

[0007] 3. User Authorization and Data Encryption Module: A hierarchical user authorization verification mechanism is set up throughout the entire process of data collection, storage, retrieval, and migration. The AES256-GCM end-to-end encryption algorithm is used to encrypt all personal data, and a unique key is generated and stored locally by the user. Without the user's active written / electronic authorization and key verification, the system cannot collect, read, retrieve, or transmit any user's personal data, thus ensuring the security of user personal information from a technical perspective and complying with relevant laws and regulations on personal information protection.

[0008] 4. Multimodal simulation interaction module: including a voice tone cloning submodule and an appearance and body posture simulation submodule; the voice tone cloning submodule is based on an open-source voice cloning model, which completes model training through effective voice materials from multiple user scenarios, accurately replicating the user's timbre, speech rate, tone and emotional vocal characteristics; the appearance and body posture simulation submodule is based on an open-source digital human modeling tool, which completes 1:1 digital modeling through user facial photos, restoring the user's facial appearance and physical characteristics.

[0009] 5. Multi-carrier adaptation and operation module: The trained personality decision-making model, encrypted structured personality data, and multimodal simulation model are synchronously adapted and deployed to dual-type multi-form operation carriers. The dual-type carriers are divided into physical carriers and digital carriers. The personality data, model parameters, and interaction logic of all carriers are completely unified and can run independently or synchronously.

[0010] III. Beneficial Technical Effects Compared with the prior art, the present invention has the following core beneficial technical effects: 1. Low barrier to entry and high accessibility: It does not require cutting-edge technologies such as brain-computer interfaces and full brain data collection. It can be implemented by relying on existing mature open-source technologies. Ordinary users without professional technical backgrounds can independently complete system training and deployment. The training cycle is only 3-4 months. There are no high service costs, and it can be popularized to the general public.

[0011] 2. High degree of personality replication and strong interaction consistency: Through the dual mechanism of LoRA lightweight fine-tuning training and personality consistency verification, it can achieve a high degree of replication of user behavior logic, language style and decision preferences; the model parameters and data are completely consistent when running on multiple carriers, ensuring that the interaction output in different scenarios is consistent with the user's original characteristics, and solving the problem of inconsistent output in multiple scenarios in existing technologies.

[0012] 3. Lightweight and highly compatible: The structured personality data supports 50-300 scalable configurations, eliminating the need for full-scale life data collection and storage, significantly reducing storage and computing costs; it is also compatible with three types of carriers: physical, virtual digital humans, and pure voice, covering all scenarios such as offline physical interaction, online visual interaction, and lightweight interaction on screenless terminals.

[0013] 4. Full-process compliance and strong privacy security: Through full-process hierarchical authorization and end-to-end encryption mechanisms, the system achieves controllable collection, secure storage and authorized access to user personal data from a technical perspective, avoiding unauthorized data leakage and abuse, complying with relevant laws and regulations on personal information protection, and avoiding compliance risks of existing technologies.

[0014] 5. This invention addresses a core technological pain point in the industry and demonstrates outstanding innovation: Existing technologies typically require hundreds of hours of user dialogue data or full-scale fine-tuning of large models, placing extremely high demands on computing power and data volume, making widespread adoption on consumer-grade hardware impossible. This invention, by jointly optimizing LoRA low-rank adaptation parameters and personality consistency verification thresholds, achieves for the first time a personality matching accuracy of over 85% under conditions of 36 hours of accumulated interaction data and a single consumer-grade GPU. This solves the rigid dependence of high-fidelity personality replication on computing power and data volume, possessing outstanding substantive features and significant technological advancements. Detailed Implementation

[0015] This embodiment details the complete implementation process of the present invention. Those skilled in the art can reproduce the technical solution of the present invention based on this embodiment. The specific steps are as follows: Step 1: Training the AI ​​Behavioral Logic Model Users interact with the system daily through the system's interface on mobile phones, computers, and other terminals, engaging in routine chats, decision-making consultations, and expressing opinions, with a cumulative daily interaction time of no less than 30 minutes. The system automatically collects users' language content, word preferences, tone characteristics, decision-making logic, and emotional feedback data. The collected dialogue data is organized into a pair format of "user question - user's actual reply," and after standardized preprocessing, it is used as a training dataset. Based on mature open-source AI models, the LoRA lightweight fine-tuning algorithm is used for training.

[0016] The LoRA rank r=8, scaling factor α=16, trainable parameters are controlled at 0.1%, learning rate is set to 2e-4, training period is 3-4 months, and cumulative effective interaction time is not less than 36 hours. After training, a user-specific personality decision model is generated, and the model output matches the user's original behavioral logic by no less than 85%.

[0017] During system operation, the built-in personality consistency verification submodule extracts 768-dimensional feature vectors from the user's original text and the model's output text based on the pre-trained BERT model, calculates the cosine similarity between the two, and presets a similarity threshold of 0.8-0.9. When the similarity of the output content is lower than the threshold, the model is automatically triggered for secondary fine-tuning and calibration to ensure that the output content always conforms to the user's original interaction characteristics.

[0018] In a preferred embodiment, the base model of the AI ​​behavior logic training module adopts Llama 3-8B, which has a 32-layer Transformer architecture, 32 attention heads, and a hidden layer dimension of 4096. During training, the cross-entropy loss function is used to calculate the difference between the model output and the user's actual response. The optimizer adopts AdamW, with β1 set to 0.85-0.95, β2 set to 0.99-0.9999, weight decay of 0.005-0.015, and the learning rate decaying from 3e-4 to 1e-5 using a cosine annealing strategy. The LoRA module is only applied to the query and value weight matrices of the attention layer and does not modify the parameters of other layers of the model, further reducing the training computational power requirements.

[0019] Step 2: Structured Personality Data Input and Encrypted Storage Users can input 50-300 pieces of expandable structured personality data through the system's visual input interface. The data includes, but is not limited to: information on family members and core social relationships, memories of important life milestones, occupational rules, living habits, likes and dislikes, core decision-making preferences, and value statements. The system uses a user authorization and data encryption module to encrypt the above data on the device side using AES256-GCM, and simultaneously generates a unique 256-bit key that is stored locally by the user. The encrypted data is stored on the local terminal or in a storage service designated by the user. It supports multi-device synchronization, authorization migration, and access. Without user authorization and key verification, no entity can read any data content.

[0020] Step 3: Multimodal simulation modeling 1. Voice and Tone Cloning: Users record more than 10 minutes of effective voice material in a quiet environment, covering different scenarios such as daily chat, serious expression, and positive / negative emotional expression. The system is based on an open-source voice cloning model. After noise reduction, frame segmentation, and acoustic feature extraction of the material, the system is trained. The training rounds are set to 10-30 rounds, and the batch size is set to 2-8, accurately replicating the user's timbre, speech rate, tone, and emotional voice characteristics.

[0021] 2. Facial Appearance and Body Simulation: Users upload 3-5 clear, headshot photos from different angles. The system uses open-source digital human modeling tools to extract facial feature points, generate meshes, and map textures, achieving a 1:1 digital model of the user's facial features and body shape, thus restoring the user's core facial features.

[0022] Step 4: Multi-carrier adaptation and deployment The system will simultaneously adapt and deploy the trained personality decision-making model, encrypted structured personality data, and multimodal simulation model to three types of runtime platforms. These three platforms can run independently or synchronously, and all personality data, model parameters, and interaction logic will be completely unified. 1. Physical carrier: Physical simulation robot, which adopts a general humanoid hardware platform, is equipped with a consumer-grade independent computing unit and system running program, replicates the user's appearance and physical characteristics, and is used for offline face-to-face physical interaction; 2. Digital Virtual Carrier: Electronic virtual digital human, packaged as a software program that can run on electronic devices such as computers, tablets, and mobile phones, to achieve online visual interaction through virtual image; 3. Lightweight Voice Carrier: Pure voice dialogue software, streamlined and packaged into a program that can run independently on various smart terminals. It only integrates an AI behavior logic training module, a structured personality data storage module, and a voice tone cloning module. It has no appearance or body posture simulation or visual presentation, and achieves lightweight interaction only through voice. It is suitable for screenless terminal scenarios such as in-vehicle and smart speakers.

[0023] Step 5: System Operation and Synchronization Management During system operation, users can switch to the corresponding operating platform according to the usage scenario. The interaction data and personality data of all platforms are synchronized in real time through an encrypted channel to ensure the consistency of multi-scenario interaction. Users can update the structured personality data and adjust the model parameters at any time after authorization verification. The system automatically completes the synchronous update of multiple platforms without the need for repeated training and deployment.

Claims

1. A multi-carrier lightweight personality digital twin system based on behavioral logic replication, characterized in that, It includes an AI behavior logic training module, a structured personality data storage module, a user authorization and data encryption module, a multimodal simulation interaction module, and a multi-carrier adaptation and operation module. The AI ​​behavior logic training module is used to collect users' language, tone, decision-making logic, and emotional feedback data through users' daily routine interactions, and generate a user-specific personality decision-making model through lightweight fine-tuning training. It also has a built-in personality consistency verification sub-module to verify the matching degree between the model output and the user's original behavior logic. The structured personality data storage module is used to encrypt and store 50-300 scalable structured personality data entries of a user. The structured personality data includes core memories, social relationships, behavioral norms, decision-making preferences, and emotional tendencies. The data supports multi-device synchronization and authorized access. The user authorization and data encryption module is used to set up hierarchical user authorization verification throughout the entire process of data collection, storage, retrieval, and migration, and to perform end-to-end encryption processing on all user personal data, preventing the collection and reading of user personal data without user authorization; the multimodal simulation interaction module includes a voice tone cloning submodule and an appearance and body posture simulation submodule, which are used to realize user-specific voice replication and 1:1 digital modeling of appearance and body posture, respectively; the multi-carrier adaptation and operation module is used to synchronously adapt and deploy the personality decision model, structured personality data, and multimodal simulation model to physical carriers and digital carriers, realizing independent or synchronous operation of multiple carriers, and ensuring that the personality data and interaction logic of all carriers are completely unified.

2. The system according to claim 1, characterized in that, The AI ​​behavior logic training module relies on a mature open-source AI model and uses the low-rank adaptation LoRA lightweight fine-tuning algorithm to complete the training. The training cycle is 3-4 months, the cumulative effective interaction time is no less than 36 hours, and the matching degree between the model output and the user's native behavior logic is no less than 85%.

3. The system according to claim 2, characterized in that, The LoRA lightweight fine-tuning algorithm has a rank r=8, a scaling factor α=16, a trainable parameter count of 0.1%, and a learning rate of 2e-4.

4. The system according to claim 1, characterized in that, The personality consistency verification submodule extracts feature vectors from the user's original text and the model's output text based on a pre-trained language model, calculates the matching degree through cosine similarity, and presets a similarity threshold of 0.8-0.

9. When the similarity of the output content is lower than the threshold, the model is automatically triggered for secondary calibration.

5. The system according to claim 1, characterized in that, The structured personality data storage module uses edge encryption to store data, eliminating the need to store the user's full life data. The number of data entries can be expanded and adjusted within the range of 50-300 according to user needs.

6. The system according to claim 1, characterized in that, The user authorization and data encryption module uses the AES256-GCM encryption algorithm to encrypt user personal data end-to-end, and generates a unique key that is stored locally by the user. The entire process is set up with dual verification of user authorization and key. The corresponding data collection, reading, calling and migration operations can only be performed after the user actively authorizes and passes the key verification.

7. The system according to claim 1, characterized in that, The voice tone cloning submodule of the multimodal simulation interaction module is trained using more than 10 minutes of effective voice materials from multiple scenarios, with 10-30 training rounds, to accurately replicate the user's timbre, speech rate, tone, and emotional vocal characteristics. The appearance and body shape simulation submodule completes a 1:1 digital model using 3-5 front-facing photos of the user from different angles, restoring the user's facial features and physical characteristics.

8. The system according to claim 1, characterized in that, The physical carrier of the multi-carrier adaptation and operation module is a physical simulation robot, which adopts a general humanoid hardware platform and carries a system running program for offline face-to-face physical interaction.

9. The system according to claim 1, characterized in that, The digital carrier of the multi-carrier adaptation and operation module includes an electronic virtual digital human, which is encapsulated as a software program that can run on electronic devices and realizes online visual interaction through a virtual image.

10. The system according to claim 1, characterized in that, The digital carrier of the multi-carrier adaptation and operation module also includes pure voice dialogue software. The pure voice dialogue software only integrates the AI ​​behavior logic training module, the structured personality data storage module, and the voice tone cloning module. It has no visual presentation and achieves lightweight interaction only through voice, adapting to screenless smart terminal scenarios.

11. The system according to claim 1, characterized in that, All operating carriers of the multi-carrier adaptation operation module can run independently or synchronously. All personality data, model parameters, and interaction logic are completely unified. When switching carriers, data is synchronized in real time through an encrypted channel.