Management system for intelligent digital human screen-crossing explanation based on spatial positioning exhibition hall

Abstract

The invention belongs to the technical field of intelligent guide systems, and particularly relates to a management method for intelligent digital human cross-screen explanation based on a spatial positioning exhibition hall, and the method comprises the following specific steps: a system starting and initializing stage: a dynamic knowledge engine obtains latest exhibition item data through a block chain node, synchronizes academic database information under a federal learning framework, and stores the information in a database; automatically generating a digital human script and adapting the digital human script to a 3D action; in the visiting whole-process management stage, tourist layered services are activated; real-time space positioning and interaction are carried out, and multi-mode positioning real-time tracking and self-adaptive interaction based on intention recognition are adopted; performing cross-device content execution; ethical risk dynamic intervention is carried out, error information is rectified, and an anti-addiction mechanism is adopted; and a closed-loop optimization and maintenance stage. According to the method, high-precision space positioning and multi-device cooperation can be realized, personalized interaction of generative AI driving is realized, a dynamic knowledge engine and hierarchical service are realized, and ethical compliance and risk prevention and control are realized.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent tour guide systems, specifically relating to a management system for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning. Background Technology

[0002] The exhibition hall digital human screen-crossing guide is an intelligent tour guide system that integrates artificial intelligence, multimodal interaction and spatial perception technologies. It allows virtual digital human images to seamlessly switch between multiple screen devices (such as sliding screens, transparent screens and holographic cabins) to achieve dynamic explanations and immersive interactive experiences. Its core is to break the boundaries of physical screens, allowing digital humans to "travel" through different display interfaces to provide visitors with a coherent and intelligent tour guide service.

[0003] Problems with existing technology: Positioning accuracy and stability defects: Existing UWB (Ultra-Wideband) or infrared positioning technologies are easily affected by complex exhibition hall environments (such as metal display cases and dense crowds), causing position drift or motion delay when the digital human moves through the screen; in scenarios where multiple people interact at the same time, the system has difficulty accurately distinguishing the individual positions, resulting in confusion about the objects being explained by the digital human. Insufficient natural interaction experience: Digital human voice recognition has a high false trigger rate in noisy environments and lacks the ability to adapt to dialects and professional terms; gesture interaction relies on a preset action library and cannot understand spontaneous gestures of tourists in real time (such as pointing to non-exhibit areas). Low efficiency in content generation and updating: The knowledge base relies on manual input, and remodeling is required when temporary exhibits are adjusted, resulting in a delayed response; RAG (Retrieval Enhanced Generation) technology has weak generalization ability for uncommon questions and is prone to returning "No relevant information found"; When digital humans walk through screens, there is a delay in data synchronization between devices such as sliding screens, transparent screens, and holographic cabins, resulting in visual fragmentation; There is no deep integration with the exhibition hall's IoT system (such as lighting and temperature control), which weakens the sense of immersion in the environment. Summary of the Invention

[0004] The purpose of this invention is to provide a management system for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning. It can achieve high-precision spatial positioning and multi-device collaboration, has generative AI-driven personalized interaction, and also has a dynamic knowledge engine and layered services, while being ethically compliant and risk-controlled.

[0005] The specific technical solution adopted by this invention is as follows: The management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning involves the following steps: System startup and initialization phase: Environmental self-inspection and equipment linkage; knowledge base synchronization and verification: the dynamic knowledge engine obtains the latest exhibit data through blockchain nodes, synchronizes academic database information under the federated learning framework, and automatically generates digital human scripts and adapts 3D actions. Visit the entire process management stage: The system features tiered visitor services: a fun guided tour mode for general visitors, with digital avatars providing animated explanations of core exhibits; a "deep mode" option for professional visitors, supporting one-click access to expert AR remote collaboration; real-time spatial positioning and interaction, employing multimodal positioning for real-time tracking and adaptive interaction based on intent recognition; cross-device content execution, achieving seamless visual transitions and multimodal output across sliding screens, transparent screens, and holographic cabins; and dynamic intervention for ethical risks, correcting erroneous information and employing anti-addiction mechanisms. Closed-loop optimization and maintenance phase.

[0006] The specific operation process of the dynamic knowledge engine is as follows: Data injection and preprocessing: Real-time data is obtained through API / web crawler, and noise is removed using median filtering and spatiotemporal alignment techniques; Knowledge fusion and storage: dynamically weighted entities based on user historical behavior; Kafka message queues enable 15-minute-level synchronization of financial data, and Flink stream processing handles millions of relationship changes; Intelligent query and generation: For ordinary queries, standardized answers are returned based on keyword matching; for complex queries, RAG is triggered to call academic databases to generate in-depth analysis; and text / video is automatically adapted according to the request type. Feedback optimization and security control: Negative user feedback automatically triggers knowledge review, and the Drools rule engine detects logical conflicts.

[0007] The specific operation process of the tiered service model is as follows: Level determination phase: After a user logs in, the system obtains department and role information through LDAP / OAuth2 and matches it with preset rules; it calculates the value of the user's historical interactions in real time and dynamically upgrades the service level; Knowledge scheduling phase: For the standard layer, answers are retrieved from the public knowledge base, and a response is generated using the RAG+BM25 algorithm; for the professional layer, information is verified by linking IEEE / CNKI, and a technical interpretation report is generated using GPT-4; for the VIP layer, a professional case library with blockchain evidence is called, and an expert video conferencing interface is integrated. Feedback optimization loop: User ratings of answers, dwell time, and other data are fed back into the federated learning pool; departmental knowledge weights are updated regularly.

[0008] The specific operation process of the multimodal positioning system is as follows: Data synchronization and preprocessing: unifying the timestamps of each sensor using hardware clocks or software protocols; determining the coordinate transformation relationship between sensors through a calibration board, such as the extrinsic parameter matrix between the camera and the LiDAR; and eliminating sensor noise using filtering algorithms. Multi-source data fusion: Sensor weights are adjusted according to environmental conditions; LiDAR point clouds and visual images are bound together through target IDs to achieve joint "location-semantic" tracking; Real-time positioning and map updates: Calculate short-distance displacements and build local high-precision maps; match pre-loaded high-precision maps or receive UWB base station signals to correct drift; use deep learning to predict environmental changes and dynamically update maps. Decision and control closed loop: Generate the optimal path and adjust the trajectory in real time through predictive control; switch to the offline lightweight model in weak network environments to ensure basic functions; use user interaction data to train the model in reverse to achieve system self-evolution.

[0009] The specific operation process of the adaptive interaction model is as follows: Data Acquisition and Fusion: Raw data such as user voice, facial expressions, and gestures are captured through devices such as cameras and microphones; spatiotemporal synchronization technology aligns multi-source signals, and Kalman filtering eliminates noise; Dynamic intent recognition: Parsing voice commands and combining visual data to determine user emotions; Personalized content generation: Use large models to generate images, videos, or 3D presentations; Cross-modal execution and feedback: Content is generated by matching digital human motion libraries and presented through holographic projection; IoT devices are linked; data such as user dwell time and question type are fed back to the knowledge base to optimize subsequent interactions.

[0010] The specific process for dynamic intervention of ethical risks is as follows: Data collection compliance: Collect de-identified behavioral data through a federated learning framework and refuse to upload raw biometric features; when a user interacts for the first time, obtain graded permissions based on a dynamic authorization model; Real-time ethical risk monitoring: Multimodal sensors analyze the user's emotional state. If discomfort is detected, the digital human is triggered to switch to eye protection mode or suspend the service; the knowledge engine connects to academic databases to verify the scientific accuracy of the content in real time and block erroneous or biased information. Risk classification and early warning: A risk matrix is ​​used to assess threats. For low-risk threats, local logging is performed without interrupting service. For high-risk threats, the connection is immediately blocked and an early warning is pushed to the administrator's terminal. A "Risk Heat Map" is generated to mark high-frequency violation areas. Intervention and feedback optimization: For algorithmic biases, the RAG engine is called to correct the output in real time; major ethical conflicts are transferred to the human ethics committee for review; based on user complaints and audit logs, the risk model threshold is optimized through reinforcement learning.

[0011] The specific process for the closed-loop optimization and maintenance phase is as follows: Data such as visitor dwell time, question type, and interaction frequency are collected to generate an "Exhibition Popularity Report" to guide content iteration; AI models are optimized under a federated learning framework to avoid privacy leaks; When the positioning signal is lost, it automatically switches to visual SLAM single-mode positioning and issues an alarm; after a system crash, it activates the offline voice guide to ensure basic services.

[0012] A management system for intelligent digital human-based guided tours in exhibition halls based on spatial positioning, the management system comprising: The intelligent sensing layer includes a fusion positioning system and a multimodal interaction device; The decision-making central layer includes a generative AI engine and a dynamic knowledge engine; The generative AI engine integrates GPT-5 / Claude 3 multimodal large models, supports RAG enhanced retrieval and federated learning optimization, and adopts a dynamic adjustment explanation strategy; the dynamic knowledge engine integrates blockchain notarization and federated learning to build a knowledge base, which is used to access CNKI / IEEE in real time to verify accuracy, automatically generate digital human scripts and adapt 3D actions, and respond to the update of exhibits. An interaction layer is executed, which includes a cross-screen device cluster and a hierarchical service terminal. The screen-through device cluster is used to achieve seamless switching between sliding screen / transparent screen / holographic cabin; the layered service terminal includes a normal mode and a professional mode. The normal mode is used for fun interactive tours and automatic loop playback, while the professional mode enables one-click access to expert AR remote collaboration and linkage with a large data visualization screen. The operation and maintenance management layer includes an ethics and security module and a self-evolving system. The ethical and safety module employs an AI ethical protocol to embed an accountability mechanism for erroneous information, and automatically switches to an eye-protection mode to prevent children from becoming addicted. Sensitive data is processed locally using edge computing nodes. The self-evolving system is used to record the duration of a visitor's stay and the type of question asked.

[0013] The fusion positioning system employs multimodal positioning using lidar, UWB, and visual SLAM; it uses millimeter-wave radar to compensate for dynamic target tracking defects, and an IMU to provide motion state calibration; the multimodal interactive device uses an infrared camera for micro-expression recognition, a microphone array for voice emotion analysis, an EEG interface for EEG interest prediction, and a wearable vibration bracelet for visually impaired assisted navigation.

[0014] An electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the management method for intelligent digital human screen explanation.

[0015] The technical effects achieved by this invention are as follows: I. High-precision spatial positioning and multi-device collaboration Employing multimodal technology combining LiDAR, UWB, and visual SLAM, it achieves a positioning accuracy of ±5cm, supports real-time tracking of over 200 people, and exhibits strong anti-interference capabilities. The digital human can smoothly traverse devices such as sliding screens, transparent screens, and holographic cabins, with natural transitions between movements and light and shadow, eliminating visual discontinuities. Sensitive data is processed locally to avoid the risk of cloud-based data leakage.

[0016] II. Generative AI-Driven Personalized Interaction Integrating large models such as GPT-5 and Claude 3, it can recognize tourists' micro-expressions and tone of voice in real time and dynamically adjust the explanation strategy (such as switching to popular language when confused); it can capture tourists' attention focus through EEG technology and trigger digital humans to demonstrate relevant exhibits in advance; it supports real-time translation between multiple languages ​​such as Chinese, English, Japanese and Spanish to adapt to international tourists.

[0017] III. Dynamic Knowledge Engine and Layered Services After the exhibit data is updated, new scripts and 3D animations for digital humans are automatically generated, and the accuracy of the content is verified in real time by accessing academic databases such as CNKI and IEEE. The intelligent layered guide provides fun and interactive tours for ordinary visitors, and enhances the experience by combining gesture recognition and AR games. Professional visitors can switch to depth mode with one click, access remote AR collaborative explanations from experts, and directly inquire about technical principles.

[0018] IV. Ethical Compliance and Risk Prevention An "AI tour guide ethics agreement" has been established, making it possible to trace and hold accountable those responsible for erroneous information and ensuring the reliability of the content; children's interaction timeouts are automatically switched to eye protection mode to guide them to rest; vibration guide wristbands assist visually impaired tourists in receiving voice information simultaneously; and a lightweight offline model is used in weak network environments to ensure that basic services are not interrupted. Attached Figure Description

[0019] Figure 1 This is a flowchart of the management method provided in an embodiment of the present invention; Figure 2 This is an architecture diagram of the management system provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0021] like Figure 1 As shown, the management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning is as follows: Step 1: System Startup and Initialization Phase Step 1.1: Environmental self-test and equipment linkage When the system starts up, it automatically detects the status of multimodal sensors (LiDAR / UWB / visual cameras), digital human projection devices (sliding screen, holographic cabin), and IoT environmental devices (lighting, temperature control); edge computing nodes load local lightweight models to ensure the availability of basic services in weak network environments; Step 1.2: Knowledge Base Synchronization and Verification The dynamic knowledge engine obtains the latest exhibit data through blockchain nodes, synchronizes information from academic databases (IEEE / CNKI) under the federated learning framework, and automatically generates digital human scripts and adapts them to 3D motion.

[0022] Step 2: Visit the full-process management stage Step 2.1: Activate Tiered Tourist Services Regular visitors: The fun guided tour mode is automatically activated by facial recognition or ticket QR code, and the digital human explains the core exhibits in an animated way; Professional visitors: The permission engine recognizes identification information, opens the "deep mode" option, and supports one-click access to expert AR remote collaboration (such as real-time anatomy demonstration in the medical exhibition area). Step 2.2: Real-time Spatial Positioning and Interaction Multimodal positioning: LiDAR + UWB + visual SLAM fusion positioning (accuracy ±5cm), real-time tracking of 200+ people; Example: When a visitor approaches the display case, the positioning system triggers a digital human to pass through the screen to the corresponding transparent screen; Intent recognition: Micro-expression analysis (CNN algorithm) detects confused expressions and automatically switches to colloquial language; EEG interface predicts focus of interest and preloads relevant holographic presentations.

[0023] Step 2.3: Cross-device content execution Cross-screen collaboration: 6G network schedules digital human motion data to achieve seamless visual transition between sliding screens, transparent screens, and holographic cabins; Multimodal output: RAG engine retrieves knowledge base, GPT-5 generates explanatory text, and Stable Diffusion renders principle animations in real time; IoT linkage: lights automatically brighten when focusing on exhibits, and digital human voice output is enhanced in noisy areas.

[0024] Step 2.4: Dynamic Intervention for Ethical Risks Error correction: The blockchain-based evidence records the content of the explanation. When the audience raises questions, the knowledge base is triggered to perform real-time verification, and errors are automatically marked and responsibilities are traced. Anti-addiction mechanism: After a child's interaction time exceeds 20 minutes, the digital human switches to eye protection mode and guides them to the rest area, with the vibration guide wristband providing a simultaneous reminder.

[0025] Step 3: Closed-loop optimization and maintenance phase Step 3.1: Data Feedback Loop Data such as visitor dwell time, question type, and interaction frequency are collected to generate an "Exhibition Popularity Report" to guide content iteration; AI models are optimized under a federated learning framework to avoid privacy leaks (such as anonymizing location data). Step 3.2: Emergency Troubleshooting When the positioning signal is lost, it automatically switches to visual SLAM single-mode positioning and issues an alarm; after a system crash, it activates the offline voice guide to ensure basic services.

[0026] Example 1: A fusion-sensing positioning system is adopted: It utilizes multimodal positioning combining LiDAR, UWB, and visual SLAM, dynamically compensating for signal distortion through AI algorithms to improve accuracy to within ±5cm, supporting real-time tracking of over 200 people. Additionally, edge computing nodes are embedded to process sensitive data locally, mitigating the risk of cloud-based data leakage. The core technologies required are as follows: 1. Multimodal sensor array Perception layer: The system needs to integrate LiDAR (3D point cloud modeling), visual sensors (RGB / infrared images), millimeter-wave radar (velocity measurement and anti-interference), UWB / IMU (inertial navigation and short-range positioning), etc. Complementary design: LiDAR solves the problem of vision failure in low light / rain and fog, millimeter-wave radar makes up for the shortcomings of LiDAR in tracking dynamic targets, and IMU provides motion state compensation.

[0027] 2. Data Fusion and Computing Engine Algorithm layer: Low-level fusion: Directly merge raw sensor data (such as point cloud and image pixel alignment), and use Kalman filtering (linear system), particle filtering (non-linear scene) or extended Kalman filtering (EKF) to process noise; High-level fusion: Feature-based (such as target contour extraction) or decision layer (weighted voting of multiple algorithm results) fusion to improve robustness; AI models: Convolutional Neural Networks (CNNs) are used for image target recognition, Graph Neural Networks (GNNs) are used to process multi-sensor correlated data, and reinforcement learning is used to optimize path planning.

[0028] 3. High-precision positioning and mapping SLAM technology: Visual SLAM (such as ORB-SLAM) and laser SLAM (such as LOAM) are combined with IMU data to achieve real-time localization and map building in dynamic environments; Global calibration: Provides absolute position anchor points via GPS / BeiDou signals (outdoors) or UWB base stations (indoors) to correct accumulated errors.

[0029] 4. Edge-Cloud Collaborative Computing Edge computing: FPGA / embedded GPUs process real-time data (such as obstacle detection) to reduce latency; Cloud-based: Distributed clusters train AI models and update global maps and decision-making strategies.

[0030] The specific operation process is as follows: Step 1: Data Synchronization and Preprocessing Time synchronization: Hardware clocks or software protocols (such as PTP) unify the timestamps of each sensor; Spatial calibration: Determine the coordinate transformation relationship between sensors through a calibration board, such as the extrinsic parameter matrix between the camera and the lidar; Noise reduction: Filtering algorithms (such as median filtering) are used to eliminate sensor noise.

[0031] Step 2: Multi-source data fusion Dynamic weight allocation: Adjust sensor weights according to environmental conditions (e.g., increase radar weights in rainy or foggy weather). Cross-modal association: LiDAR point clouds and visual images are bound together through target IDs to achieve joint tracking of "location and semantics".

[0032] Step 3: Real-time positioning and map updates Local positioning: IMU + wheel velocity meter calculate short-distance displacement, SLAM algorithm constructs local high-precision map; Global correction: Matching pre-loaded high-precision maps (such as parking space coordinates) or receiving UWB base station signals to correct drift; Adaptive mapping: Deep learning predicts environmental changes (such as moving obstacles) and dynamically updates the map.

[0033] Step 4: Decision-making and control closed loop Path planning: A or RRT algorithms generate the optimal path, and Model Predictive Control (MPC) adjusts the trajectory in real time; Disaster recovery mechanism: Switch to offline lightweight model in weak network environments to ensure basic functions; Feedback optimization: User interaction data (such as obstacle avoidance success rate) is used to train the model in reverse, enabling the system to self-evolve.

[0034] Example 2: Adaptive interaction driven by generative AI: Integrating multimodal large models (such as GPT-5 + Claude 3) to achieve real-time analysis of tourists' micro-expressions and adjust explanation strategies (such as automatically switching to colloquial language when confusion is detected), and supporting tourists to customize question paths (such as skipping historical background and directly inquiring about technical principles); In addition, introducing an electroencephalogram (EEG) interface to predict focus of interest and trigger digital human cross-screen demonstrations in advance. The core technologies required are as follows: 1. Multimodal sensing layer Sensor array: Integrates visual cameras (micro-expression recognition), microphone arrays (voice emotion analysis), infrared sensors (body tracking), and other devices to capture user behavior data in real time; Biometric recognition: Predicting user interest focus through brainwave interface (EEG) or eye tracker; Technical support: Computer vision (CNN) analyzes user facial expressions and gestures; speech recognition (ASR) combines with NLP to analyze semantics and emotions.

[0035] 2. Generative AI Engine Large model foundation: Employs pre-trained models such as GPT-4 and Claude 3 to achieve contextual understanding and content generation; Adaptive Mechanism: Dynamic Knowledge Base: Based on RAG (Retrieval Enhanced Generation) technology, it accesses academic databases (such as IEEE) in real time to verify information; Personalized Fine-tuning: Under the federated learning framework, it uses user historical data to optimize the local model and avoid privacy leaks.

[0036] 3. Decision-making and interactive control Intent reasoning engine: Graph Neural Network (GNN) analyzes the correlation of multimodal data to predict user needs; Multi-device collaboration: Synchronizes digital human movements, lighting, temperature control and other IoT devices through 6G network to enhance immersion; Disaster recovery design: Switch to a lightweight model in weak network environments to ensure basic services.

[0037] The specific operation process is as follows: Step 1: Data Acquisition and Fusion The system captures raw data such as user voice, facial expressions, and gestures through devices such as cameras and microphones; spatiotemporal synchronization technology (such as the PTP protocol) aligns multi-source signals, and Kalman filtering eliminates noise.

[0038] Step 2: Dynamic Intent Recognition NLP analyzes voice commands and combines them with visual data to determine the user's emotions (such as confusion / excitement); Example: When a user frowns, supplementary explanation materials are automatically retrieved.

[0039] Step 3: Personalized Content Generation Scene adaptation: Call large models to generate text, images, videos or 3D demonstrations, such as using Stable Diffusion to generate principle diagrams.

[0040] Step 4: Cross-modal execution and feedback The digital human motion library is matched to generate content, which is then presented through holographic projection; IoT devices are linked (such as automatically brightening the lights when focusing on exhibits); data such as user dwell time and question type are fed back to the knowledge base to optimize subsequent interactions.

[0041] Example 3: Employing a dynamic knowledge engine: A "blockchain + federated learning" knowledge base is constructed. After the exhibit data is updated, a new script for the digital human is automatically generated and 3D motion is adapted. Academic databases (such as CNKI and IEEE) are accessed to verify the accuracy of the explanations in real time. The core technologies required are as follows: 1. Multimodal data acquisition and analysis Heterogeneous data access: Supports access to structured data (databases) and unstructured data (PDF / images / videos) through multiple channels such as APIs, web crawlers, and IoT devices, and extracts text information through OCR (Optical Character Recognition) and NLP technologies; Dynamic document parsing: Employs deep learning models (such as CLIP and 3D-VAE) to parse complex elements (tables / charts / formulas) while preserving semantic relationships (e.g., TextIn ParseX enables lossless conversion of cross-page tables).

[0042] 2. Knowledge Graph Construction and Update Dynamic graph engine: Builds an entity relationship network based on graph database (Neo4j / GraphDB) and updates nodes in real time with an event-driven architecture (such as automatically triggering graph corrections when policies change). Federated learning and blockchain: Distributed training ensures data privacy, while blockchain notarization ensures the traceability of knowledge versions (avoiding information conflicts).

[0043] 3. Intelligent Retrieval and Generation Layer RAG Enhanced Retrieval: Integrating the vector database (Milvus) with semantic retrieval algorithms, the BM25+ANN hybrid model improves query accuracy (300% increase in answer acceptance rate in medical scenarios). Multimodal large model: Integrates models such as GPT-4 and Claude 3, and supports cross-modal content generation (such as the schematic diagram of the generation principle of StableDiffusion).

[0044] 4. Adaptive Decision Making and Feedback User intent reasoning: Graph Neural Networks (GNNs) analyze user behavior data (click-through rate / stay time) to dynamically optimize knowledge push strategies; Self-evolution mechanism: knowledge weights are adjusted through reinforcement learning, and content iteration is guided by the "Exhibition Item Popularity Report" (the knowledge reuse rate of a certain e-commerce platform has increased by 90%).

[0045] The specific operation process is as follows: Step 1: Data Injection and Preprocessing Acquire real-time data (such as academic journals and IoT sensors) via API / web crawler, and use median filtering and spatiotemporal alignment techniques to remove noise; Example: TextIn ParseX parses industrial manuals, compresses 3D models to 2% of their original size while retaining key parameters.

[0046] Step 2: Knowledge Integration and Storage Entity disambiguation: dynamically weighting entities based on user historical behavior; Incremental updates: Kafka message queues enable 15-minute synchronization of financial data, and Flink stream processing handles millions of relationship changes.

[0047] Step 3: Intelligent Query and Generation Hierarchical retrieval strategy: Standard query: Returns standardized answers based on keyword matching; Complex query: Triggers RAG to call academic databases (IEEE / CNKI) to generate in-depth analysis; Multimodal output: Automatically adapts text / images / videos based on request type (e.g., linking medical Q&A with pathology diagrams and medication instructions).

[0048] Step 4: Feedback Optimization and Security Control User negative feedback automatically triggers knowledge review, and the Drools rule engine detects logical conflicts (such as the contradiction between "free" and "subscription"); the RBAC permission model (read-only / edit / administrator) + AES-256 encryption meets financial-grade compliance requirements.

[0049] Based on the dynamic knowledge engine, a tiered service model is established: For general tourists, digital humans provide engaging and interactive guided tours; for professional audiences, a one-click switch to in-depth mode allows for remote AR collaborative explanations from experts. The core technologies required are as follows: 1. Fine-grained permission engine Three-dimensional access control model: Based on the hierarchical design of "role-department-user" (e.g., the R&D department can only access technical documents, and the sales department can focus on product selling points), combined with attribute-based encryption (ABE) to dynamically control the scope of data access; Blockchain-based evidence storage: Permission change records are uploaded to the blockchain to ensure traceability of operations (e.g., a medical platform uses smart contracts to restrict unauthorized personnel from accessing patient data).

[0050] 2. Federated Learning Layered Optimization Differentiated training mechanisms: Basic layer: General knowledge federated training (such as industry terminology database); Specialized layer: Vertical domain model fine-tuning (such as financial risk control models trained only with compliance department data); Gradient-weighted aggregation: Dynamically adjust federated learning weights based on departmental data quality to improve the accuracy of models for high-value departments.

[0051] 3. Dynamic resource scheduling system Knowledge service tiers: Refer to Table 1 below. Table 1 Service level Computing power quota Response delay Knowledge coverage depth Standard layer 1 CPU core <500ms Basic Question and Answer Database Professional level GPU cluster <200ms Academic database + real-time verification VIP Floor edge nodes <50ms Private knowledge graph + expert interface Elastic scaling: Based on Kubernetes' HPA strategy, computing resources are automatically allocated according to the amount of concurrent requests.

[0052] 4. Personalized Service Engine User profile building: Generate interest tags through behavioral tracking (such as clicking on hot topics, question types) to drive knowledge push strategies (such as prioritizing the push of patent databases to R&D personnel). Multimodal interaction adaptation: Dynamically render knowledge content (text / 3D model / speech summary) based on terminal type (PC / mobile / AR glasses).

[0053] The specific operation process is as follows: Step 1: Hierarchical Determination Stage After a user logs in, the system obtains department and role information through LDAP / OAuth2 and matches it with preset rules (such as "Compliance Department → Professional Level"); it calculates the value of the user's historical interactions in real time (such as the complexity of the consultation questions) and dynamically upgrades the service level (such as triggering VIP service after 10 consecutive in-depth questions).

[0054] Step 2: Knowledge Scheduling Phase Standard Layer: Retrieves answers from public knowledge bases and generates responses using the RAG+BM25 algorithm; Professional Layer: Links with IEEE / CNKI to verify information and generates technical interpretation reports using GPT-4; VIP Layer: Calls a professional case library with blockchain-based evidence and integrates expert video conferencing interfaces (such as in medical consultation scenarios).

[0055] Step 3: Feedback optimization closed loop User ratings and dwell time data are fed back into the federated learning pool; departmental knowledge weights are updated regularly (e.g., the sales department frequently asks "competitive analysis" → which increases the priority of the market intelligence database).

[0056] Example 4: Ethical and risk prevention measures are adopted: A "Digital Rights and Responsibilities Delineation AI Guideline" is established, clearly defining the accountability mechanism for erroneous information; an anti-addiction design is implemented, automatically switching to eye protection mode and guiding children to rest after excessive interaction time. The core technologies required are as follows: 1. Data privacy protection technology Federated Learning + Blockchain: Employing a distributed federated learning framework ensures localized processing of user data and prevents leakage of raw data; combined with blockchain notarization technology, it records data usage trajectories to achieve traceability and auditability. Dynamic desensitization and encryption: Real-time desensitization of sensitive information (such as location and biometrics) (such as differential privacy algorithms) and the security of data during the computation process are ensured through homomorphic encryption technology.

[0057] 2. Ethical Decision Engine Multimodal risk perception: Integrating visual analysis (micro-expression recognition), voice emotion detection, and EEG signal analysis, it monitors user discomfort or ethical conflicts in real time (such as excessive inducement to consume). Compliance Knowledge Graph: Construct a graph library based on industry ethical standards (such as IEEE ethical guidelines) to automatically verify whether digital human behavior complies with norms such as privacy protection and fairness.

[0058] 3. Dynamic Risk Early Warning System AI-driven risk modeling: Utilize time series analysis models (such as LSTM) to predict potential risks (such as data breach trends), combine risk matrices to quantify threat levels, and trigger graded early warnings (low / medium / high). Edge-cloud collaborative computing: Edge devices process sensitive data in real time, while the cloud performs global analysis of risk patterns, reducing response latency.

[0059] 4. Disaster recovery and emergency control Graceful degradation mechanism: In the event of a system failure, the system automatically switches to a lightweight model, retains basic navigation functions, and provides emergency guidance (historical dialogue) to visually impaired users via a vibrating wristband. Blockchain notarization + smart contracts: automatically execute rollback of illegal operations (such as unauthorized data collection) and freeze high-risk accounts [[8]9].

[0060] The specific operation process is as follows: Step 1: Compliant Data Collection Collect anonymized behavioral data (such as dwell time and question type) through a federated learning framework, and refuse to upload raw biometric features; when a user interacts for the first time, obtain hierarchical permissions based on a dynamic authorization model (such as context-based access control).

[0061] Step 2: Real-time ethical risk monitoring Multimodal sensors analyze users' emotional states (such as frowning in anxiety). If discomfort is detected (such as children becoming addicted), the digital human is triggered to switch to eye protection mode or suspend the service. The knowledge engine connects to academic databases (IEEE / CNKI) to verify the scientific accuracy of the content in real time and block erroneous or biased information.

[0062] Step 3: Risk Classification and Early Warning Threats are assessed using a risk matrix: Low risk (e.g., temporary network delays): Local logging is used to ensure service is not interrupted; High risk (e.g., abnormal data transmission): Connections are immediately blocked and an alert is pushed to the administrator's terminal; A "Risk Heat Map" is generated to mark high-frequency violation areas (e.g., excessive collection of facial data in a certain exhibition area).

[0063] Step 4: Intervention and Feedback Optimization Automatic intervention: In case of algorithmic bias (such as regionally discriminatory explanations), the RAG engine is invoked to correct the output in real time; Human intervention: Major ethical conflicts (such as brain-computer interface mind leakage) are transferred to an artificial ethics committee for review; Continuous evolution: Based on user complaints and audit logs, the risk model threshold is optimized through reinforcement learning.

[0064] like Figure 2 As shown, the management system for the intelligent digital human walk-through narration in the exhibition hall based on spatial positioning is used to implement the management method for the intelligent digital human walk-through narration. The management system includes: The intelligent sensing layer includes a fusion positioning system and multimodal interaction devices; The fusion positioning system adopts multimodal positioning of "LiDAR + UWB + visual SLAM" with an accuracy of ±5cm and supports real-time tracking of 200+ people; millimeter-wave radar compensates for the defects in dynamic target tracking, and IMU provides motion state calibration; Multimodal interactive devices use infrared cameras for micro-expression recognition, microphone arrays for voice emotion analysis, EEG interfaces for EEG interest prediction, and wearable vibration wristbands for visually impaired assisted navigation.

[0065] The decision-making central layer includes a generative AI engine and a dynamic knowledge engine. The generative AI engine integrates GPT-5 / Claude 3 multimodal large models, supports RAG enhanced retrieval and federated learning optimization, and adopts dynamic adjustment of explanation strategies (such as switching to colloquial language when identifying confusion). The dynamic knowledge engine integrates blockchain notarization and federated learning to build a knowledge base, which is used to access CNKI / IEEE in real time to verify accuracy, automatically generate digital human scripts and adapt 3D actions, and respond to item updates.

[0066] The execution interaction layer includes a cross-screen device cluster and a layered service terminal. The screen-through device cluster is used to achieve seamless switching between sliding screen / transparent screen / holographic cabin. The 6G network ensures that the light and shadow transition latency is <10ms. The naked-eye 3D display technology (such as Chenjing Optoelectronics solution) is used to enhance the sense of spatial depth. The tiered service terminal includes a normal mode and a professional mode. The normal mode is used for fun interactive tours and automatic loop playback, while the professional mode enables one-click access to expert AR remote collaboration and linkage with a large data visualization screen.

[0067] The operation and maintenance management layer includes an ethics and security module and a self-evolving system. The ethics and safety module adopts an AI ethics protocol to embed an accountability mechanism for error information, and automatically switches to eye protection mode to prevent children from becoming addicted. It also uses edge computing nodes to process sensitive data locally and transmits data with AES-256 encryption. The self-evolutionary system is used to record visitor dwell time and question types, generating an "Exhibition Popularity Report" to guide curation.

[0068] The specific implementation plan for this management system is as follows: Phase 1: Basic Deployment (0-3 months) Hardware deployment: Install positioning base stations (UWB anchor points + LiDAR grid) to cover a 2000㎡ exhibition hall, and deploy AI interactive large screens (starting from 86 inches) and holographic projection equipment; Knowledge base construction: Upload exhibit materials to the blockchain knowledge base, train the federated learning model, connect to the academic database API, and set content update thresholds (such as weekly automatic verification).

[0069] Phase 2: System Integration and Testing (4-6 months) Multimodal interaction testing: calibrating sensor spatiotemporal synchronization (PTP protocol), optimizing noise filtering algorithms, constructing an EEG interest prediction model, with a prediction accuracy of ≥85%; Cross-screen collaboration verification: Test the continuity of digital human movements across devices (such as the transition from sliding screen to holographic cabin), and configure IoT linkage rules (such as automatically increasing the brightness of the lights by 30% when focusing on the exhibit).

[0070] Phase 3: Scenario Implementation (7-12 months) Tiered services launched: General audience channel: interactive gesture / voice guidance available; Professional audience channel: remote expert collaboration via AR glasses enabled; Disaster recovery mechanism activated: Switch to lightweight model in weak network environment to ensure offline basic services, and trigger emergency broadcast + vibration guide wristband when system crash.

[0071] Phase 4: Continuous Optimization (Long-Term) Monthly iteration: Optimize knowledge graph weights based on visitor interaction data (such as pre-answering high-frequency questions); Annual upgrade: Introduce quantum sensing to improve positioning stability in extreme environments.

[0072] An electronic device, comprising: at least one processor; and memory that is communicatively connected to at least one processor; The memory stores a computer program that can be executed by at least one processor, which enables the at least one processor to implement the management method for intelligent digital human screen explanation.

[0073] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning, characterized in that: The specific steps are as follows: System startup and initialization phase: Environmental self-inspection and equipment linkage; knowledge base synchronization and verification: the dynamic knowledge engine obtains the latest exhibit data through blockchain nodes, synchronizes academic database information under the federated learning framework, and automatically generates digital human scripts and adapts 3D actions. Visit the entire process management stage: The visitor tiered service is activated, with a fun guided tour mode for general visitors, where digital avatars explain the core exhibits in animated form; a "deep mode" option is available for professional visitors, supporting one-click access to expert AR remote collaboration; real-time spatial positioning and interaction are achieved through multimodal positioning and real-time tracking, and adaptive interaction based on intent recognition. Cross-device content execution enables seamless visual transitions and multimodal output across sliding screens, transparent screens, and holographic cabins; dynamic intervention for ethical risks corrects erroneous information and employs anti-addiction mechanisms. Closed-loop optimization and maintenance phase.

2. The management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning as described in claim 1, characterized in that: The specific operation process of the dynamic knowledge engine is as follows: Data injection and preprocessing: Real-time data is obtained through API / web crawler, and noise is removed using median filtering and spatiotemporal alignment techniques; Knowledge fusion and storage: dynamically weighted entities based on user historical behavior; Kafka message queues enable 15-minute-level synchronization of financial data, and Flink stream processing handles millions of relationship changes; Intelligent query and generation: For ordinary queries, standardized answers are returned based on keyword matching; for complex queries, RAG is triggered to call academic databases to generate in-depth analysis; and text / video is automatically adapted according to the request type. Feedback optimization and security control: Negative user feedback automatically triggers knowledge review, and the Drools rule engine detects logical conflicts.

3. The management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning as described in claim 1, characterized in that: The specific operation process of the tiered service model is as follows: Level determination phase: After a user logs in, the system obtains department and role information through LDAP / OAuth2 and matches it with preset rules; it calculates the value of the user's historical interactions in real time and dynamically upgrades the service level; Knowledge scheduling phase: For the standard layer, answers are retrieved from the public knowledge base, and the RAG+BM25 algorithm is used to generate responses; For the professional level, information is verified through IEEE / CNKI, and a technical interpretation report is generated using GPT-4; for the VIP level, a professional case library of blockchain-based evidence is accessed, and an expert video conferencing interface is integrated. Feedback optimization loop: User ratings of answers, dwell time, and other data are fed back into the federated learning pool; Regularly update the department's knowledge weights.

4. The management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning as described in claim 1, characterized in that: The specific operation process of the multimodal positioning system is as follows: Data synchronization and preprocessing: unifying the timestamps of each sensor using hardware clocks or software protocols; determining the coordinate transformation relationship between sensors through a calibration board, such as the extrinsic parameter matrix between the camera and the LiDAR; and eliminating sensor noise using filtering algorithms. Multi-source data fusion: Sensor weights are adjusted according to environmental conditions; LiDAR point clouds and visual images are bound together through target IDs to achieve joint "location-semantic" tracking; Real-time positioning and map updates: Calculate short-distance displacements and build local high-precision maps; match pre-loaded high-precision maps or receive UWB base station signals to correct drift; use deep learning to predict environmental changes and dynamically update maps. Decision and control closed loop: Generate the optimal path and adjust the trajectory in real time through predictive control; switch to the offline lightweight model in weak network environments to ensure basic functions; use user interaction data to train the model in reverse to achieve system self-evolution.

5. The management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning as described in claim 1, characterized in that: The specific operation process of the adaptive interaction model is as follows: Data Acquisition and Fusion: Raw data such as user voice, facial expressions, and gestures are captured through devices such as cameras and microphones; spatiotemporal synchronization technology aligns multi-source signals, and Kalman filtering eliminates noise; Dynamic intent recognition: Parsing voice commands and combining visual data to determine user emotions; Personalized content generation: Use large models to generate images, videos, or 3D presentations; Cross-modal execution and feedback: Content is generated by matching digital human motion libraries and presented through holographic projection; IoT devices are linked; data such as user dwell time and question type are fed back to the knowledge base to optimize subsequent interactions.

6. The management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning as described in claim 1, characterized in that: The specific process for dynamic intervention of ethical risks is as follows: Data collection compliance: Collect de-identified behavioral data through a federated learning framework and refuse to upload raw biometric features; when a user interacts for the first time, obtain graded permissions based on a dynamic authorization model; Real-time ethical risk monitoring: Multimodal sensors analyze the user's emotional state. If discomfort is detected, the digital human is triggered to switch to eye protection mode or suspend the service; the knowledge engine connects to academic databases to verify the scientific accuracy of the content in real time and block erroneous or biased information. Risk classification and early warning: Threats are assessed using a risk matrix. For low-risk threats, local logging is performed to ensure uninterrupted service. For high-risk cases, immediately block the connection and push an alert to the administrator's terminal; Generate a "Risk Heat Map" and mark areas with high frequency of violations; Intervention and feedback optimization: To correct algorithmic biases, the RAG engine is invoked to correct the output in real time; Major ethical conflicts will be transferred to the artificial ethics committee for review. Based on user complaints and audit logs, the risk model threshold is optimized through reinforcement learning.

7. The management method for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning as described in claim 1, characterized in that: The specific process for the closed-loop optimization and maintenance phase is as follows: Data such as visitor dwell time, question type, and interaction frequency are collected to generate an "Exhibition Popularity Report" to guide content iteration; AI models are optimized under a federated learning framework to avoid privacy leaks; When the positioning signal is lost, it automatically switches to visual SLAM single-mode positioning and issues an alarm; after a system crash, it activates the offline voice guide to ensure basic services.

8. A management system for intelligent digital human walk-through narration in exhibition halls based on spatial positioning, used to execute the management method for intelligent digital human walk-through narration as described in any one of claims 1-7, characterized in that: The management system includes: The intelligent sensing layer includes a fusion positioning system and a multimodal interaction device; The decision-making central layer includes a generative AI engine and a dynamic knowledge engine; The generative AI engine integrates GPT-5 / Claude 3 multimodal large models, supports RAG enhanced retrieval and federated learning optimization, and adopts a dynamic adjustment explanation strategy; the dynamic knowledge engine integrates blockchain notarization and federated learning to build a knowledge base, which is used to access CNKI / IEEE in real time to verify accuracy, automatically generate digital human scripts and adapt 3D actions, and respond to the update of exhibits. An interaction layer is executed, which includes a cross-screen device cluster and a hierarchical service terminal. The screen-through device cluster is used to achieve seamless switching between sliding screen / transparent screen / holographic cabin; the layered service terminal includes a normal mode and a professional mode. The normal mode is used for fun interactive tours and automatic loop playback, while the professional mode enables one-click access to expert AR remote collaboration and linkage with a large data visualization screen. The operation and maintenance management layer includes an ethics and security module and a self-evolving system. The ethical and safety module employs an AI ethical protocol to embed an accountability mechanism for erroneous information, and automatically switches to an eye-protection mode to prevent children from becoming addicted. Sensitive data is processed locally using edge computing nodes. The self-evolving system is used to record the duration of a visitor's stay and the type of question asked.

9. The management system for intelligent digital human walk-through explanation in exhibition halls based on spatial positioning as described in claim 8, characterized in that: The fusion positioning system employs multimodal positioning using "LiDAR, UWB, and visual SLAM"; it uses millimeter-wave radar to compensate for dynamic target tracking defects, and an IMU to provide motion state calibration; the multimodal interaction device uses an infrared camera for micro-expression recognition, a microphone array for voice emotion analysis, an EEG interface for EEG interest prediction, and a wearable vibration bracelet for visually impaired assisted navigation.

10. An electronic device, characterized in that: The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the management method for intelligent digital human screen explanation as described in any one of claims 1 to 7.

Images