Holographic projection system based on AI large model

Abstract

The invention discloses a holographic projection system based on an AI large model. According to the invention, through the nano-scale optical coating and environment self-adaptive adjustment technology, harmful blue light is effectively filtered, environment light parameters are dynamically matched, the problem that eyes are injured by a common screen is solved, and the method is especially suitable for long-time learning scenes of children. The holographic projection core module adopts a naked eye 3D light field technology, presents complex models such as dinosaur, marine organisms and the like in a three-dimensional suspension manner through high-fidelity color restoration and dynamic depth-of-field adjustment capability, and is matched with gesture touch control and voice instructions of multi-mode interaction hardware, so that a user can directly touch a virtual object, and the real-time performance of the virtual object is improved. Experience upgrading from plane observation to space interaction is achieved, and knowledge such as English learning and animal and plant cognition is converted into interactive games and story narration through virtual teacher images and scene scripts. The autistic children can complete social dialogue training under the guidance of the virtual role, the system analyzes emotion fluctuation in real time and adjusts a guidance strategy, and social fear is gradually eliminated.

Description

Technical Field

[0001] The present invention belongs to the technical field of educational products, and specifically relates to a holographic projection system based on an AI large model. Background Art

[0002] Educational home products and devices refer to intelligent devices that integrate educational functions with modern home technologies, designed to provide convenient and efficient learning support in the home environment. These devices typically include smart learning machines, educational robots, and supporting software applications. By integrating rich curriculum resources, interactive teaching modules, and personalized learning path planning, they can assist students with academic tutoring, interest development, and skill development. Furthermore, these devices often include voice interaction, remote monitoring, and learning data feedback, allowing parents to participate in supervision and guidance. Deeply integrating technology and education, educational home products and devices are dedicated to optimizing home learning spaces and promoting the development of personalized learning and smart education.

[0003] However, the prior art has the following disadvantages:

[0004] Traditional display technology: Ordinary screens rely on a single light source and lack a mechanism to filter out harmful light, which can easily lead to visual fatigue after long-term use. Traditional 3D displays require wearing special glasses, have limited interaction methods, and cannot achieve naked-eye stereo touch. The content is dull and lacks realism.

[0005] Limitations of educational tools: Existing educational products mostly adopt a one-way knowledge infusion model, with lagging content updates and lacking emotional guidance, making it difficult to meet the social training needs of autistic children. At the same time, they are unable to dynamically adjust teaching strategies based on user behavior, and lack personalization and fun.

[0006] Fragmentation of smart devices: Existing smart home devices have scattered functions, low voice interaction accuracy and reliance on fixed instruction sets, fragmented multimodal operations, and a lack of cloud collaboration and scenario-based service integration. Users need to frequently switch devices to complete different tasks, which makes the experience cumbersome.

[0007] High threshold for technology application: Traditional AI systems rely on complex operating interfaces and professional terminology. Generative technology and emotional computing are mostly limited to laboratory environments and fail to be transformed into intuitive home interaction forms. Ordinary users find it difficult to enjoy the benefits of technology, and the technology is not accessible enough. Summary of the Invention

[0008] The purpose of the present invention is to provide a holographic projection system based on an AI large model in order to solve the above-mentioned problems.

[0009] The technical solution adopted by the present invention is as follows: a holographic projection system based on an AI large model, the system comprising: multimodal interactive hardware, a health eye protection module, an AI interaction and processing module, a dynamic content cloud platform, a service and operation and maintenance module, a holographic projection core module, and an education and companionship module;

[0010] The AI ​​interaction and processing module is internally configured with: an emotional computing and social guidance submodule, a multimodal instruction parsing submodule, an intelligent question-answering and content generation submodule, an audio signal conversion and driving module, and an immersive service management submodule;

[0011] The voice input interface and touch signal output end of the multimodal interaction hardware are respectively connected to the instruction parsing unit and the emotion calculation unit of the AI ​​interaction and processing module;

[0012] The ambient light sensor data output end of the healthy eye protection module is connected to the light source control interface of the holographic projection core module, and its device status monitoring end is connected to the hardware health management interface of the service and operation and maintenance module;

[0013] The content request interface of the AI ​​interaction and processing module obtains the 3D model and educational script through the data distribution gateway of the dynamic content cloud platform;

[0014] The content push interface of the dynamic content cloud platform is directly connected to the rendering engine of the holographic projection core module and the course loading unit of the education and companionship module;

[0015] The remote control interface of the service and operation and maintenance module is embedded in the firmware upgrade channel of the holographic projection core module and the Bluetooth protocol stack of the multimodal interaction hardware;

[0016] The social training data output of the education and companionship module is fed back to the strategy optimization unit of the AI ​​interaction and processing module, forming a closed-loop iteration from behavior analysis to guidance strategy.

[0017] In a preferred embodiment, the multimodal interaction hardware consists of a pressure-sensitive projection screen, a ring-shaped microphone array, and a low-power Bluetooth chip. The pressure-sensitive projection screen utilizes a distributed pressure sensor network. The ring-shaped microphone array includes six sets of highly sensitive sound pickup units. Using a beamforming algorithm, it focuses on the user's voice source. The adaptive noise reduction module eliminates ambient noise interference, ensuring clear capture of voice commands. The low-power Bluetooth chip supports dual-mode protocols. The hardware layer also integrates an edge computing unit.

[0018] In a preferred embodiment, the healthy eye protection module is based on a nano-scale optical coating layer and an ambient light sensing system. The nano-coating layer, applied to the projection screen surface, utilizes multi-layer interference filtering technology to selectively absorb high-energy blue light in the 420-455nm band. The ambient light sensing system, comprised of a multispectral sensor and a dynamic adjustment algorithm, monitors indoor light intensity and color temperature in real time. The module also includes a built-in eye protection strategy library, including preset options such as a children's learning mode and a nighttime entertainment mode.

[0019] In a preferred embodiment, the emotion calculation and social guidance submodule consists of three parts: a multimodal emotion recognition engine, a dynamic strategy adjustment unit, and a virtual character interaction system. The multimodal emotion recognition engine collects voice signals and facial expression data through a high-precision microphone array and an optional camera, respectively, and uses a deep neural network fusion architecture to synchronously analyze intonation, speech speed, facial action units, and body language to generate a three-dimensional emotion vector containing pleasure, tension, and participation. The dynamic strategy adjustment unit matches the preset social guidance strategy library in real time based on the emotion vector, and optimizes the guidance intensity in combination with the user response pattern in the historical interaction data. The virtual character interaction system integrates a lightweight digital human engine and a scenario script library, renders the natural expressions and movements of the virtual image through skeletal drive technology and lip synchronization algorithm, and calls the social scene scripts updated in the cloud;

[0020] Social guidance priority index = basic emotional sensitivity × real-time emotional fluctuation value + historical adaptation attenuation coefficient × scene complexity weight;

[0021] Among them: Basic emotional sensitivity: the system's preset initial response sensitivity level, with a value range of 0.2 to 1.0. The higher the value, the more sensitive it is to capturing subtle emotional changes.

[0022] Real-time emotion fluctuation value: The comprehensive emotion change rate calculated based on the current voice and expression data, including the instantaneous amplitude of a sudden rise in tone or a frown, normalized to a scalar between 0 and 1.

[0023] Historical adaptation attenuation coefficient: A dynamic parameter that reflects the user's long-term interaction habits. The initial value is 0.8. If the user ignores the guidance instructions three times in a row, the coefficient decreases in steps of 0.1 to a minimum of 0.3.

[0024] Scenario complexity weight: A fixed value preset based on the difficulty of the current social training scenario.

[0025] In a preferred embodiment, the multimodal instruction parsing submodule consists of four parts: a voice wake-up unit, a multimodal input adapter, an instruction classification engine, and a context management unit. The voice wake-up unit uses a low-power chip and beamforming technology to accurately identify preset wake-up words in a noisy environment. The multimodal input adapter integrates multiple interaction methods such as voice, touch, handwriting, and file upload, and supports real-time signal conversion and format unification. The instruction classification engine is based on a deep neural network model to perform intent recognition and entity extraction on user instructions, and generate multiple candidate results for fuzzy input. The context management unit dynamically maintains the dialogue status and historical interaction records to achieve multi-round instruction connection;

[0026] Entity display priority = voice confidence × default weight + historical preference coefficient × interaction frequency factor

[0027] Among them: Speech confidence: The target entity recognition accuracy score output by the speech recognition model, ranging from 0 to 1. The higher the value, the more reliable the recognition result.

[0028] Default weights represent preset entity priority parameters.

[0029] The historical preference coefficient indicates the preference strength calculated based on the user’s past selection records;

[0030] The interaction frequency factor represents the normalized value of the number of times an entity is actively called by users in recent interactions.

[0031] In a preferred embodiment, the intelligent question-answering and content generation submodule consists of three parts: a knowledge-enhanced retrieval engine, a dynamic generation model cluster, and a user portrait adapter. The knowledge-enhanced retrieval engine adopts a hybrid retrieval architecture, integrating local structured knowledge graphs with cloud-based unstructured data sources, modeling entity relationship weights through a graph attention network, prioritizing the extraction of answer fragments that match the user's age and historical behavior, automatically filtering complex ecological terms, and returning storytelling explanations. The dynamic generation model cluster deploys a multimodal generative model, based on retrieval results and real-time conversation context, and optimizes the synchronous output of text, voice, and 3D animation through an adversarial training framework. The user portrait adapter analyzes users' long-term interest preferences and short-term interactive feedback through a temporal convolutional network, dynamically adjusts the complexity and expression of generated content, and inserts English vocabulary annotations for learning users.

[0032] To ensure content diversity, the system introduces a random latent space perturbation mechanism, injecting controllable noise into the latent variable layer of the generative model. This allows the same semantic input to produce content with consistent themes but differentiated presentation. Furthermore, a question-and-answer quality monitor calculates the comprehensibility and knowledge accuracy of generated content in real time, fine-tuning the generative model online through reinforcement learning strategies to ensure that the output meets educational psychology standards.

[0033] in:

[0034] Content generation diversity score = topic consistency coefficient × semantic perturbation intensity + user freshness factor × historical exposure decay rate;

[0035] Where:

[0036] The topic consistency coefficient indicates the degree of matching between the generated content calculated based on the twin network and the core topic of the retrieved knowledge, ranging from 0 to 1. The higher the value, the lower the risk of the content deviating from the preset topic.

[0037] The semantic perturbation intensity represents a hyperparameter that controls the amplitude of latent variable noise injection. It constrains the deviation between the generated distribution after perturbation and the original distribution through the KL divergence constraint, and the preset range is 0.1 to 0.5.

[0038] The user freshness factor represents a diversity demand indicator calculated based on the homogeneity of the content the user has recently been exposed to. If a user receives the same type of content three times in a row, the factor decreases linearly from 1.0 to 0.3 to force diversity enhancement.

[0039] The historical exposure decay rate represents a content repetition suppression parameter that decays exponentially over time.

[0040] In a preferred embodiment, the audio signal conversion and driving module consists of three parts: a high-fidelity audio acquisition unit, a multi-dimensional feature analysis engine, and a hardware driver controller. The high-fidelity audio acquisition unit uses a circular microphone array and an adaptive beamforming algorithm to capture user voice commands in a complex sound field environment. At the same time, it uses nonlinear filtering technology to suppress environmental noise and device background noise to ensure the time-frequency domain integrity of the original signal. The multi-dimensional feature analysis engine deploys a hybrid signal processing model, combining Mel-spectrogram and time-domain envelope features to extract speech content, emotional intonation, and background voiceprint information, and performs parallel analysis of multiple task targets based on a gated convolutional neural network.

[0041] In a preferred embodiment, the immersive service management submodule consists of a virtual customer service engine, a dynamic scenario adapter, and a user profile linkage system. The virtual customer service engine utilizes a multi-layered dialogue management architecture, integrating a large language model with a domain knowledge graph to support contextual understanding and intent prediction for multimodal interactions. A reinforcement learning framework optimizes dialogue strategies, enabling intelligent service switching from device operation guidance to educational content recommendations. The dynamic scenario adapter deploys a state machine based on a hidden Markov model to analyze user behavior sequences and environmental parameters in real time, triggering seamless switching of service modes.

[0042] The user profile linkage system uses a time-series graph neural network to model long-term behavioral data, integrating historical interaction records, content preferences, and emotional feedback into a multi-dimensional feature vector to drive personalized service strategy customization. This system is deeply integrated with the cloud-based content library. The module also includes a built-in abnormal state rollback mechanism. By comparing the deviation between real-time interaction data and historical patterns, it automatically corrects scene freezes caused by network delays or recognition errors, ensuring the integrity and stability of the immersive experience.

[0043] In a preferred embodiment, the dynamic content cloud platform is internally configured with:

[0044] Distributed content library, intelligent recommendation engine, and differential update server. The distributed content library stores 3D models, educational courses, and interactive scripts by topic. The intelligent recommendation engine uses collaborative filtering and knowledge graph fusion algorithms to analyze click-through rates, dwell time, and interaction frequency in user behavior logs to dynamically generate personalized content lists. The differential update server uses binary difference comparison technology to transmit only incremental data related to content version changes.

[0045] The service and operations module consists of a virtual customer service cluster and remote monitoring nodes. The virtual customer service cluster deploys a multi-instance conversation engine, managing the user consultation process based on a domain-driven state machine. The remote monitoring node collects real-time hardware operating metrics, including temperature, memory usage, and network latency, and triggers automatic maintenance through threshold warning mechanisms. The module also integrates log analysis tools, utilizing a time-series database to store user operation records and generate device usage reports and abnormal event tracing reports.

[0046] In a preferred embodiment, the optical system of the holographic projection core module consists of a laser diffraction array and a dynamic rendering engine. The laser diffraction array adopts a multi-angle interference optical path design, generates naked-eye 3D images through sub-micron phase modulation, and supports 360-degree rotation of objects and dynamic adjustment of depth of field. The dynamic rendering engine is based on a physical optics simulation algorithm to calculate the light field propagation path and material reflection characteristics in real time. The color calibration system is deeply bound to the Pantone color library, and the projection color gamut data is fed back through the spectrum analyzer to automatically correct the color difference to the professional standard of ΔE<1.5 to ensure the color consistency of virtual content and real objects.

[0047] The education and companionship module includes a graded course library and a scenario simulation engine. The graded course library divides knowledge units by age and subject. The scenario simulation engine incorporates multi-role interactive scripts and optimizes the virtual teacher's conversational strategies through reinforcement learning. During social training, the avatar adjusts the difficulty of questions based on the complexity of the user's responses, gradually guiding autistic children through complete conversation chains. The module also integrates a cognitive assessment system that analyzes user learning progress through voice recognition and behavioral logs, generates a multidimensional competency map, and delivers customized practice content.

[0048] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0049] 1. This invention uses nano-scale optical coating and environmental adaptive adjustment technology to effectively filter harmful blue light and dynamically match ambient light parameters, solving the eye damage problem of ordinary screens. It is particularly suitable for children in scenarios where they study for long periods of time. The core holographic projection module uses naked-eye 3D light field technology, with high-fidelity color reproduction and dynamic depth of field adjustment capabilities, to present complex models such as dinosaurs and marine life in a suspended three-dimensional form. Combined with gesture touch and voice commands of multimodal interactive hardware, users can directly "touch" virtual objects, realizing an upgraded experience from flat observation to spatial interaction. This technology not only makes content appear lifelike, but also builds a sustainable and healthy usage ecosystem by reducing the risk of visual fatigue.

[0050] 2. In this invention, through the use of virtual teacher images and situational scripts, knowledge such as English learning and plant and animal cognition is transformed into interactive games and storytelling. Autistic children can complete social dialogue training under the guidance of virtual characters. The system analyzes emotional fluctuations in real time and adjusts guidance strategies, gradually eliminating social phobias. At the same time, the dynamic content cloud platform continuously updates thematic content adapted to different age groups based on user behavior data, such as dinosaur evolution animations and marine ecological simulations, making the learning process both interesting and systematic. This "learning through play" model not only improves the efficiency of knowledge absorption, but also stimulates the desire to explore through ever-changing content formats, breaking the one-way indoctrination limitations of traditional education.

[0051] 3. This invention seamlessly integrates into everyday life through natural language interaction and cloud-based collaboration. Users can quickly complete operations such as setting alarms, playing music, and checking the time through voice commands. The Bluetooth speaker linkage function extends the entertainment experience from auditory to visual, presenting virtual dancer animations simultaneously with music playback. The online service module uses digital human technology and real-time dialogue management to provide immersive support such as device usage guidance and content recommendations, significantly lowering the threshold for technical use. In addition, dynamically updated eye protection strategies and device self-checking mechanisms ensure long-term stable operation of the system.

[0052] 4. In the present invention, cutting-edge technologies such as generative AI and emotional computing are transformed into intuitive interactive forms. For example, a user's casual display request can trigger a complete link from 3D modeling to ecological knowledge explanation. The AI ​​intelligent question-and-answer module converts complex knowledge into popular answers through big data screening and natural language generation. This "what you think is what you get" interactive mode concretizes abstract technology into perceptible virtual images and scenes, allowing ordinary home users to enjoy the benefits of technology. By reducing the cost of using technology and the cognitive threshold, the system not only meets current functional requirements, but also reserves interfaces for future expansion of scenarios such as IoT linkage and smart home control, continuously empowering digital life innovation and improving the convenience of using the overall system. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a block diagram of the overall system of the present invention;

[0054] Figure 2 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] Example:

[0057] Reference Figure 1-2 , a holographic projection system based on AI large model, the system includes: multimodal interactive hardware, health eye protection module, AI interaction and processing module, dynamic content cloud platform, service and operation and maintenance module, holographic projection core module and education and companionship module;

[0058] The internal configuration of the AI ​​interaction and processing module includes: emotional computing and social guidance sub-module, multimodal instruction parsing sub-module, intelligent question answering and content generation sub-module, audio signal conversion drive module and immersive service management sub-module;

[0059] The voice input interface and touch signal output of the multimodal interaction hardware are connected to the instruction parsing unit and emotion calculation unit of the AI ​​interaction and processing module respectively;

[0060] The data output of the ambient light sensor of the healthy eye protection module is connected to the light source control interface of the holographic projection core module, and its device status monitoring end is connected to the hardware health management interface of the service and operation and maintenance module;

[0061] The content request interface of the AI ​​interaction and processing module obtains 3D models and educational scripts through the data distribution gateway of the dynamic content cloud platform;

[0062] The content push interface of the dynamic content cloud platform is directly connected to the rendering engine of the holographic projection core module and the course loading unit of the education and companionship module;

[0063] The remote control interface of the service and operation and maintenance module is embedded in the firmware upgrade channel of the holographic projection core module and the Bluetooth protocol stack of the multimodal interaction hardware;

[0064] The social training data output of the education and companionship module is fed back to the strategy optimization unit of the AI ​​interaction and processing module, forming a closed-loop iteration from behavior analysis to guidance strategy.

[0065] The multimodal interaction hardware consists of a pressure-sensitive projection screen, a circular microphone array, and a low-power Bluetooth chip. The pressure-sensitive projection screen uses a distributed pressure sensor network that can accurately identify touch position and force level, with a light touch triggering basic commands and a long press activating advanced functions. The circular microphone array has six built-in high-sensitivity pickup units, which use a beamforming algorithm to focus on the user's sound source, and cooperate with the adaptive noise reduction module to eliminate environmental noise interference, ensuring clear capture of voice commands. The low-power Bluetooth chip supports dual-mode protocols and is compatible with A2DP audio transmission and HFP call control, achieving seamless connection with external devices such as mobile phones and tablets. For example, it automatically synchronizes device volume adjustment when playing music. The hardware layer also integrates an edge computing unit, equipped with a dedicated NPU to accelerate voice wake-up and local command parsing, reducing cloud dependence to improve response speed.

[0066] The healthy eye protection module is based on a nano-level optical coating layer and an ambient light sensing system. The nano-coating layer covers the surface of the projection screen and uses multi-layer interference filtering technology to selectively absorb high-energy blue light in the 420-455nm band while maintaining a color reproduction rate of more than 95%. The ambient light perception system consists of a multi-spectral sensor and a dynamic adjustment algorithm to monitor the indoor light intensity and color temperature in real time. For example, it automatically reduces the projection brightness in low-light environments and switches to a warm color temperature mode to avoid glare stimulation. The module has an embedded eye protection strategy library, which includes preset schemes such as children's learning mode and night entertainment mode. For example, in learning mode, a reading-friendly color gamut is enabled and an eye protection reminder is triggered every 20 minutes to ensure eye health through hardware-level regulation.

[0067] The affective computing and social guidance submodule consists of a multimodal emotion recognition engine, a dynamic strategy adjustment unit, and a virtual character interaction system. The multimodal emotion recognition engine uses a high-precision microphone array and an optional camera to collect voice signals and facial expression data, respectively. Using a deep neural network fusion architecture, it simultaneously analyzes intonation, speech rate, facial action units, and body language to generate a three-dimensional emotion vector that reflects pleasure, tension, and engagement. The dynamic strategy adjustment unit uses this emotion vector to match a preset social guidance strategy library in real time. For example, if a user is nervous, it automatically switches to a lower-difficulty conversation task and enables encouraging voice feedback. It also optimizes guidance intensity based on user response patterns from historical interaction data. The virtual character interaction system integrates a lightweight digital human engine and a scenario script library. Using skeletal drive technology and lip-sync algorithms, it renders the avatar's natural expressions and movements. It also calls upon social scenario scripts updated in the cloud, such as simulated park socializing or classroom question-and-answer scenarios, to guide users through progressive tasks to complete social training. To enhance realism, the system also allows users to customize their avatar's appearance and interaction style, ensuring a personalized and adaptable social guidance process.

[0068] Social guidance priority index = basic emotional sensitivity × real-time emotional fluctuation value + historical adaptation attenuation coefficient × scene complexity weight;

[0069] Among them: Basic emotional sensitivity: the system's preset initial response sensitivity level, with a value range of 0.2 to 1.0. The higher the value, the more sensitive it is to capturing subtle emotional changes.

[0070] Real-time emotion fluctuation value: The comprehensive emotion change rate calculated using the current voice and expression data, such as the instantaneous amplitude of a sudden rise in voice tone or a frown, normalized to a scalar between 0 and 1.

[0071] Historical adaptation attenuation coefficient: A dynamic parameter that reflects the user's long-term interaction habits. The initial value is 0.8. If the user ignores the guidance instructions three times in a row, the coefficient decreases in steps of 0.1 to a minimum of 0.3.

[0072] Scenario complexity weight: A fixed value preset based on the difficulty of the current social training scenario. For example, the weight of the "simple greeting" scenario is 0.5, and the weight of the "multi-person conversation simulation" scenario is 1.0.

[0073] This formula dynamically adjusts the intensity of social guidance by balancing immediate emotional feedback with long-term behavioral patterns. For example, during a new user's initial use, the system prioritizes real-time emotional fluctuations for a quick response, avoiding guidance delays caused by insufficient historical data. As user usage increases, the historical adaptation decay coefficient gradually reduces the priority of mechanical repetitive guidance, instead adapting to a more natural interaction rhythm based on scenario complexity and user habits, thus balancing guidance efficiency with user comfort.

[0074] The multimodal command parsing submodule consists of a voice wake-up unit, a multimodal input adapter, a command classification engine, and a context management unit. The voice wake-up unit utilizes a low-power chip and beamforming technology to accurately recognize preset wake-up words, such as "Xiao Hai," or custom keywords, even in noisy environments. It also uses a noise reduction algorithm to filter out background interference. The multimodal input adapter integrates multiple interaction methods, including voice, touch, handwriting, and file upload, supporting real-time signal conversion and format unification. For example, this can convert hand-drawn animal outlines into vector graphics or parse keywords from uploaded documents. The command classification engine, based on a deep neural network model, performs intent recognition and entity extraction on user commands. For example, it can separate the action command "play" from the target entity "African lion" in the query "play a documentary about African lions." It also generates multiple candidate results for ambiguous input, such as converting the ambiguous polyphonetic character "giraffe" into graphic and text options for user confirmation. The context management unit dynamically maintains conversation state and historical interaction records to facilitate the connection between multiple rounds of commands. For example, when a user first asks "What do lions eat" and then follows up with "Where do they live," the system automatically associates the context entity "lion" to ensure coherence in the answer.

[0075] Entity display priority = voice confidence × default weight + historical preference coefficient × interaction frequency factor

[0076] Among them: Speech confidence: The target entity recognition accuracy score output by the speech recognition model, ranging from 0 to 1. The higher the value, the more reliable the recognition result.

[0077] Default weight: The preset entity priority parameter. For example, in "lion, tiger, dog", the weight of "lion" is set to 0.6, and the weight of the remaining entities decreases in sequence.

[0078] Historical preference coefficient: The preference strength is calculated based on the user's past selection records. For example, if the user has selected "Tiger" three times, the coefficient increases to 0.8, and the initial value is 0.5.

[0079] Interaction frequency factor: The normalized value of the number of times an entity is actively called by users in recent interactions. For example, if "dog" is selected 10 times in the past week, the factor is 0.9.

[0080] This formula dynamically optimizes entity display strategies by integrating real-time recognition credibility with user behavior data. For example, when a new user first says "show lion," the system prioritizes displaying "lion" based on voice confidence and default weighting. If the user subsequently switches to "tiger" multiple times, the historical preference coefficient and interaction frequency factor will gradually increase the display priority of "tiger," achieving a smooth transition from system defaults to personalized recommendations while reducing interaction friction caused by voice recognition errors.

[0081] The intelligent question-answering and content generation submodule consists of a knowledge-enhanced search engine, a dynamic generation model cluster, and a user profile adapter. The knowledge-enhanced search engine utilizes a hybrid search architecture, integrating a local structured knowledge graph with unstructured cloud data sources. Using a graph attention network to model entity relationship weights, it prioritizes answer fragments that match the user's age and historical behavior. For example, for a child's question, "How do lions hunt?", the engine automatically filters out complex ecological terms and returns a storytelling explanation. The dynamic generation model cluster deploys a multimodal generative model. Based on search results and real-time conversation context, it uses an adversarial training framework to optimize the simultaneous output of text, speech, and 3D animation. For example, it transforms an encyclopedia description of an African lion into an interactive narrative featuring onomatopoeia and body movements. The user profile adapter uses a temporal convolutional network to analyze users' long-term interests and short-term interaction feedback, dynamically adjusting the complexity and presentation of generated content. For example, it adds holographic predation scene animations for users who prefer visual stimulation, while also inserting English vocabulary annotations for learning-oriented users.

[0082] To ensure content diversity, the system introduces a random latent space perturbation mechanism, injecting controllable noise into the latent variable layer of the generative model. This allows the same semantic input to produce content with consistent themes but differentiated presentations. For example, a video titled "A Day in the Life of a Lion" can generate different versions, such as a grassland stroll, hunting training, or family interactions. Furthermore, a question-and-answer quality monitor calculates the comprehensibility and knowledge accuracy of generated content in real time, fine-tuning the generative model online through a reinforcement learning strategy to ensure that the output meets educational psychology standards.

[0083] in:

[0084] Content generation diversity score = topic consistency coefficient × semantic perturbation intensity + user freshness factor × historical exposure decay rate;

[0085] Where:

[0086] Topic consistency coefficient: The degree of matching between the generated content and the core topic of the retrieved knowledge calculated based on the twin network, ranging from 0 to 1. The higher the value, the lower the risk of the content deviating from the preset topic.

[0087] Semantic perturbation strength: A hyperparameter that controls the amplitude of latent variable noise injection. It constrains the deviation between the generated distribution after perturbation and the original distribution through KL divergence. The preset range is 0.1 to 0.5.

[0088] User freshness factor: A diversity demand indicator calculated based on the homogeneity of the content the user has recently been exposed to. If a user receives the same type of content three times in a row, the factor decreases linearly from 1.0 to 0.3 to force diversity enhancement.

[0089] Historical exposure decay rate: A content repetition suppression parameter that decays exponentially over time. For example, the decay rate of a "lion hunting" scene shown three days ago is 0.6, while the decay rate of the scene shown today is 1.0.

[0090] This formula achieves a dynamic balance between educational value and fun by quantifying the conflicting relationship between topic consistency and user needs. For example, when the system detects that the user's freshness factor has dropped below a threshold due to content duplication, it automatically increases the semantic perturbation intensity to generate a variant version, while simultaneously ensuring that core knowledge is not lost through the topic consistency coefficient. The algorithm layer uses the diversity score as a reward signal for reinforcement learning, driving the generative model to maximize content innovation while ensuring accuracy, thereby meeting the dual requirements of "constant variety" and "gradual cognitive improvement" in patent materials.

[0091] The audio signal conversion and driver module consists of a high-fidelity audio acquisition unit, a multi-dimensional feature analysis engine, and a hardware driver controller. The high-fidelity audio acquisition unit utilizes a circular microphone array and adaptive beamforming algorithms to capture user voice commands in complex sound environments. Nonlinear filtering techniques are also used to suppress ambient noise and device background noise, ensuring the time-frequency integrity of the original signal. The multi-dimensional feature analysis engine employs a hybrid signal processing model that combines mel-spectrograms with time-domain envelope features to extract speech content, emotional intonation, and background voiceprint information. It also uses a gated convolutional neural network to perform parallel analysis of multiple tasks, for example, separating explicit commands such as "play music" and implicit requests such as "increase volume" from the same audio stream. The hardware driver controller uses dynamic impedance matching circuits and a low-latency protocol stack to convert the analysis results into hardware control signals. These signals can, for example, adjust the frame rate of a holographic animation in sync with the music rhythm or trigger the speaker's low-frequency enhancement mode via the Bluetooth A2DP protocol.

[0092] The immersive service management submodule consists of three parts: a virtual customer service engine, a dynamic scenario adapter, and a user portrait linkage system. The virtual customer service engine adopts a multi-layer dialogue management architecture, integrating a large language model and a domain knowledge graph to support contextual understanding and intent prediction for multimodal interactions. Through the optimization of dialogue strategies through a reinforcement learning framework, intelligent service switching from device operation guidance to educational content recommendations is achieved. For example, based on the user's question "How to connect to Bluetooth", the user can automatically jump to the technical support process, or start the English training mode after identifying the learning intent. The dynamic scenario adapter deploys a state machine based on the hidden Markov model to analyze the user behavior sequence and environmental parameters in real time, triggering seamless switching of service modes. For example, when three consecutive entertainment instructions are detected, the complexity of the virtual character's dance movements is automatically enhanced to match the user's increased interest.

[0093] The user portrait linkage system uses a time-series graph neural network to model long-term behavioral data, integrating historical interaction records, content preferences, and emotional feedback into a multi-dimensional feature vector to drive personalized service strategy customization. The system is deeply coupled with the cloud-based content library. For example, when a user prefers a "dinosaur theme," the virtual customer service conversation case library prioritizes loading relevant knowledge points, while the holographic projection scene simultaneously switches to a Jurassic forest environment. To ensure service continuity, the module has a built-in abnormal state rollback mechanism. By comparing the deviation between real-time interaction data and historical patterns, it automatically repairs scene freezes caused by network delays or recognition errors, ensuring the integrity and stability of the immersive experience.

[0094] The internal settings of the dynamic content cloud platform are:

[0095] Distributed content library, intelligent recommendation engine and differential update server. The distributed content library stores 3D models, educational courses and interactive scripts by subject classification. For example, the dinosaur encyclopedia contains skeletal animations and habitat scene data of different species. The intelligent recommendation engine uses collaborative filtering and knowledge graph fusion algorithms to analyze the click-through rate, dwell time and interaction frequency in user behavior logs, and dynamically generates personalized content lists. For example, it prioritizes whale migration simulation content for users who prefer marine life. The differential update server uses binary difference comparison technology to transmit only incremental data of content version changes. For example, a 10MB dinosaur model update only requires downloading a 2MB difference package, which greatly reduces bandwidth consumption and update time.

[0096] The service and operations module consists of a virtual customer service cluster and remote monitoring nodes. The virtual customer service cluster deploys a multi-instance conversation engine and manages the user consultation process based on a domain-driven state machine. For example, the technical support branch includes a built-in device pairing guide and troubleshooting knowledge base, while the education branch links to the national textbook outline and course schedule. The remote monitoring node collects hardware operating indicators in real time, including temperature, memory usage, and network latency, and triggers automatic maintenance through a threshold warning mechanism. For example, if the projection module is detected to be overheating, it immediately starts the cooling fan and pushes a maintenance notification to the operation and maintenance terminal. The module also integrates log analysis tools, using a time series database to store user operation records and generate device usage reports and abnormal event tracing reports.

[0097] The optical system of the core module of holographic projection consists of a laser diffraction array and a dynamic rendering engine. The laser diffraction array adopts a multi-angle interference optical path design, generates naked-eye 3D images through sub-micron phase modulation, and supports 360-degree rotation of objects and dynamic adjustment of depth of field. The dynamic rendering engine is based on a physical optics simulation algorithm to calculate the light field propagation path and material reflection characteristics in real time. For example, the metal surface presents high light reflection and the hair model simulates the sub-light scattering effect. The color calibration system is deeply bound to the Pantone color library, and the projection color gamut data is fed back through the spectrum analyzer to automatically correct the color difference to the professional standard of ΔE<1.5 to ensure the color consistency between virtual content and real objects.

[0098] The Education and Companionship module is internally configured with a graded course library and a scenario simulation engine. The graded course library divides knowledge units by age and subject. For example, preschool English courses focus on word cards and nursery rhymes, while elementary school courses incorporate grammar explanations and scenario dialogues. The scenario simulation engine features built-in multi-character interactive scripts, optimizing the virtual teacher's dialogue strategies through reinforcement learning. For example, during social training, the SpongeBob SquarePants character adjusts the difficulty of questions based on the complexity of the user's answers, gradually guiding autistic children through complete conversation chains. The module also integrates a cognitive assessment system that analyzes user learning progress through voice recognition and behavior logs, generates a multidimensional ability map, and pushes customized practice content, such as automatically increasing the frequency of spelling games for vocabulary weaknesses.

[0099] From the above we can know:

[0100] In this invention, nano-level optical coating and environmental adaptive adjustment technology are used to effectively filter harmful blue light and dynamically match ambient light parameters, solving the eye damage problem of ordinary screens. It is especially suitable for children studying for a long time. The core module of holographic projection uses naked-eye 3D light field technology, with high-fidelity color reproduction and dynamic depth of field adjustment capabilities, to present complex models such as dinosaurs and marine life in a suspended three-dimensional form. With the gesture touch and voice commands of multimodal interactive hardware, users can directly "touch" virtual objects, realizing an upgraded experience from plane observation to spatial interaction. This technology not only makes the content appear lifelike, but also builds a sustainable and healthy usage ecosystem by reducing the risk of visual fatigue.

[0101] In this invention, knowledge such as English learning and plant and animal cognition is transformed into interactive games and storytelling through virtual teacher images and situational scripts. Autistic children can complete social dialogue training under the guidance of virtual characters. The system analyzes emotional fluctuations in real time and adjusts guidance strategies to gradually eliminate social phobias. At the same time, the dynamic content cloud platform continuously updates thematic content adapted to different age groups based on user behavior data, such as dinosaur evolution animations and marine ecological simulations, making the learning process both interesting and systematic. This "learning through play" model not only improves the efficiency of knowledge absorption, but also stimulates the desire to explore through ever-changing content forms, breaking the one-way indoctrination limitations of traditional education.

[0102] This system seamlessly integrates into everyday life through natural language interaction and cloud-based collaboration. Users can quickly set alarms, play music, and query the time using voice commands. The Bluetooth speaker linkage function extends the entertainment experience from auditory to visual, for example, by simultaneously presenting virtual dancer animations while playing music. The online service module utilizes digital human technology and real-time conversation management to provide immersive support such as device usage guidance and content recommendations, significantly lowering the barrier to entry for technology users. Furthermore, dynamically updated eye protection strategies and device self-checking mechanisms ensure long-term stable operation of the system.

[0103] In the present invention, cutting-edge technologies such as generative AI and emotional computing are transformed into intuitive interactive forms. For example, a user's casual display requirement can trigger a complete link from 3D modeling to ecological knowledge explanation. The AI ​​intelligent question-answering module converts complex knowledge into popular answers through big data screening and natural language generation. This "what you think is what you get" interactive mode concretizes abstract technology into perceptible virtual images and scenes, allowing ordinary home users to enjoy the benefits of technology. By reducing the cost of using technology and the cognitive threshold, the system not only meets current functional requirements, but also reserves interfaces for future expansion of scenarios such as IoT linkage and smart home control, continuously empowering digital life innovation and improving the convenience of using the overall system.

[0104] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0105] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A holographic projection system based on an AI large model, characterized by: The system includes: multimodal interactive hardware, healthy eye protection module, AI interaction and processing module, dynamic content cloud platform, service and operation and maintenance module, holographic projection core module and education and companionship module; The AI ​​interaction and processing module is internally configured with: an emotional computing and social guidance submodule, a multimodal instruction parsing submodule, an intelligent question-answering and content generation submodule, an audio signal conversion and driving module, and an immersive service management submodule; The voice input interface and touch signal output end of the multimodal interaction hardware are respectively connected to the instruction parsing unit and the emotion calculation unit of the AI ​​interaction and processing module; The ambient light sensor data output end of the healthy eye protection module is connected to the light source control interface of the holographic projection core module, and its device status monitoring end is connected to the hardware health management interface of the service and operation and maintenance module; The content request interface of the AI ​​interaction and processing module obtains the 3D model and educational script through the data distribution gateway of the dynamic content cloud platform; The content push interface of the dynamic content cloud platform is directly connected to the rendering engine of the holographic projection core module and the course loading unit of the education and companionship module; The remote control interface of the service and operation and maintenance module is embedded in the firmware upgrade channel of the holographic projection core module and the Bluetooth protocol stack of the multimodal interaction hardware; The social training data output of the education and companionship module is fed back to the strategy optimization unit of the AI ​​interaction and processing module, forming a closed-loop iteration from behavior analysis to guidance strategy.

2. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The multimodal interaction hardware consists of a pressure-sensitive projection screen, a circular microphone array and a low-power Bluetooth chip; the pressure-sensitive projection screen adopts a distributed pressure sensor network; the circular microphone array has six built-in high-sensitivity pickup units, which use a beamforming algorithm to focus on the user's sound source, and cooperate with an adaptive noise reduction module to eliminate environmental noise interference, ensuring clear capture of voice commands; the low-power Bluetooth chip supports dual-mode protocols; the hardware layer also integrates an edge computing unit.

3. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The healthy eye protection module is based on a nano-level optical coating layer and an ambient light sensing system. The nano-coating layer covers the surface of the projection screen and uses multi-layer interference filtering technology to selectively absorb high-energy blue light in the 420-455nm band. The ambient light sensing system consists of a multi-spectral sensor and a dynamic adjustment algorithm to monitor indoor light intensity and color temperature in real time. The module has an embedded eye protection strategy library, which includes preset plans for children's learning mode and night entertainment mode.

4. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The emotional computing and social guidance submodule consists of three parts: a multimodal emotion recognition engine, a dynamic strategy adjustment unit, and a virtual character interaction system. The multimodal emotion recognition engine collects voice signals and facial expression data through a high-precision microphone array and an optional camera, and uses a deep neural network fusion architecture to synchronously analyze intonation, speech rate, facial action units, and body language to generate a three-dimensional emotion vector containing pleasure, tension, and engagement. The dynamic strategy adjustment unit matches the preset social guidance strategy library in real time based on the emotion vector, and optimizes the guidance intensity based on the user response pattern in the historical interaction data. The virtual character interaction system integrates a lightweight digital human engine and a scenario script library, renders the natural expressions and movements of the virtual image through skeletal drive technology and lip synchronization algorithm, and calls the social scene script updated in the cloud. Social guidance priority index = basic emotional sensitivity × real-time emotional fluctuation value + historical adaptation attenuation coefficient × scene complexity weight; Among them: Basic Emotional Sensitivity: The system's preset initial response sensitivity level, ranging from 0.2 to 1.

0. The higher the value, the more sensitive it is to capturing subtle emotional changes; Real-time emotion fluctuation value: The comprehensive emotion change rate calculated from the current voice and expression data, including the instantaneous amplitude of a sudden rise in tone or a frown, normalized to a scalar between 0 and 1; Historical Adaptation Decay Coefficient: This is a dynamic parameter that reflects the user's long-term interaction habits. The initial value is 0.

8. If the user ignores the guidance instructions three times in a row, the coefficient decreases in steps of 0.1 to a minimum of 0.

3. Scenario complexity weight: A fixed value preset based on the difficulty of the current social training scenario.

5. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The multimodal command parsing submodule consists of four parts: a voice wake-up unit, a multimodal input adapter, a command classification engine, and a context management unit. The voice wake-up unit uses a low-power chip and beamforming technology to accurately identify preset wake-up words in noisy environments. The multimodal input adapter integrates multiple interaction methods such as voice, touch, handwriting, and file upload, and supports real-time signal conversion and format unification; The command classification engine, based on a deep neural network model, identifies user commands' intent and extracts entities, generating multiple candidate results for fuzzy inputs. The context management unit dynamically maintains conversation status and historical interaction records to enable multi-round command linkage. Entity display priority = voice confidence × default weight + historical preference coefficient × interaction frequency factor Among them: Speech confidence: The target entity recognition accuracy score output by the speech recognition model, ranging from 0 to 1. The higher the value, the more reliable the recognition result; Default weights represent preset entity priority parameters; The historical preference coefficient indicates the preference strength calculated based on the user’s past selection records; The interaction frequency factor represents the normalized value of the number of times an entity is actively called by users in recent interactions.

6. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The intelligent question-answering and content generation submodule consists of three parts: a knowledge-enhanced retrieval engine, a dynamic generation model cluster, and a user portrait adapter. The knowledge-enhanced retrieval engine adopts a hybrid retrieval architecture, integrating local structured knowledge graphs with cloud-based unstructured data sources. It models entity relationship weights through a graph attention network, prioritizes answer fragments that match the user's age and historical behavior, automatically filters complex ecological terms, and returns storytelling explanations. The dynamic generation model cluster deploys a multimodal generative model, which optimizes the synchronous output of text, voice, and 3D animation through an adversarial training framework based on retrieval results and real-time conversation context. The user portrait adapter analyzes users' long-term interest preferences and short-term interaction feedback through a temporal convolutional network, dynamically adjusts the complexity and expression of generated content, and inserts English vocabulary annotations for learning users. To ensure content diversity, the system introduces a random latent space perturbation mechanism, injecting controllable noise into the latent variable layer of the generative model. This allows the same semantic input to output content with consistent themes but differentiated presentation. Furthermore, a question-and-answer quality monitor calculates the comprehensibility and knowledge accuracy of generated content in real time, fine-tuning the generative model online through reinforcement learning strategies to ensure that the output meets educational psychology standards. in: Content generation diversity score = topic consistency coefficient × semantic perturbation intensity + user freshness factor × historical exposure decay rate; Where: The topic consistency coefficient indicates the degree of matching between the generated content calculated based on the twin network and the core topic of the retrieved knowledge, ranging from 0 to 1. The higher the value, the lower the risk of the content deviating from the preset topic. The semantic perturbation intensity is a parameter that controls the amplitude of latent variable noise injection. It constrains the deviation of the generated distribution after perturbation from the original distribution through the KL divergence constraint, and the preset range is 0.1 to 0.

5. The user freshness factor represents a diversity demand indicator calculated based on the homogeneity of the content the user has recently been exposed to. If a user receives the same type of content three times in a row, the factor decreases linearly from 1.0 to 0.3 to force diversity enhancement; The historical exposure decay rate represents a content repetition suppression parameter that decays exponentially over time.

7. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The audio signal conversion and driving module consists of three parts: a high-fidelity audio acquisition unit, a multi-dimensional feature analysis engine, and a hardware drive controller. The high-fidelity audio acquisition unit uses a circular microphone array and an adaptive beamforming algorithm to capture user voice commands in a complex sound field environment. At the same time, it uses nonlinear filtering technology to suppress environmental noise and device background noise, ensuring the time and frequency domain integrity of the original signal. The multi-dimensional feature analysis engine deploys a hybrid signal processing model, combines the Mel-spectrogram and time-domain envelope features to extract speech content, emotional intonation, and background voiceprint information, and performs parallel analysis of multi-task targets based on a gated convolutional neural network.

8. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The immersive service management submodule consists of three parts: a virtual customer service engine, a dynamic scenario adapter, and a user portrait linkage system. The virtual customer service engine adopts a multi-layer dialogue management architecture, integrating a large language model and domain knowledge graph to support context understanding and intent prediction in multimodal interactions. Optimize conversation strategies through a reinforcement learning framework to achieve intelligent service switching from device operation guidance to educational content recommendations; The dynamic scenario adapter deploys a state machine based on the Hidden Markov Model to analyze user behavior sequences and environmental parameters in real time, triggering seamless switching of service modes; The user portrait linkage system uses a time-series graph neural network to model long-term behavioral data, integrating historical interaction records, content preferences, and emotional feedback into multi-dimensional feature vectors to drive personalized customization of service strategies; the system is deeply coupled with the cloud content library; the module has a built-in abnormal state rollback mechanism, which automatically repairs scene freezes caused by network delays or recognition errors by comparing the deviation between real-time interaction data and historical patterns, ensuring the integrity and stability of the immersive experience.

9. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The internal configuration of the dynamic content cloud platform includes: Distributed content library, intelligent recommendation engine, and differential update server; the distributed content library stores 3D models, educational courses, and interactive scripts by topic; the intelligent recommendation engine uses collaborative filtering and knowledge graph fusion algorithms to analyze click-through rate, dwell time, and interaction frequency in user behavior logs to dynamically generate personalized content lists; The differential update server uses binary difference comparison technology to transmit only the incremental data of content version changes; The service and operation and maintenance module consists of a virtual customer service cluster and remote monitoring nodes. The virtual customer service cluster deploys a multi-instance conversation engine and manages the user consultation process based on a domain-driven state machine. The remote monitoring node collects hardware operating indicators in real time, including temperature, memory usage, and network latency, and triggers automatic maintenance through a threshold warning mechanism. The module also integrates log analysis tools, uses a time series database to store user operation records, and generates device usage reports and abnormal event tracing reports.

10. The holographic projection system based on the AI ​​large model according to claim 1, characterized in that: The optical system of the holographic projection core module consists of a laser diffraction array and a dynamic rendering engine. The laser diffraction array uses a multi-angle interference optical path design to generate naked-eye 3D images through submicron phase modulation, supporting 360-degree rotation of objects and dynamic adjustment of depth of field. The dynamic rendering engine uses a physical optics simulation algorithm to calculate the light field propagation path and material reflection characteristics in real time. The color calibration system is deeply integrated with the Pantone color library, and through the feedback of projection color gamut data from the spectrum analyzer, it automatically corrects color differences to the professional standard of ΔE<1.5, ensuring color consistency between virtual content and real objects. The internal settings of the education and companionship module include: a graded course library and a scenario simulation engine; the graded course library divides knowledge units according to age and subject; the scenario simulation engine has a built-in multi-role interactive script, and optimizes the virtual teacher's dialogue strategy through reinforcement learning. In social training, the virtual image adjusts the difficulty of questions according to the complexity of the user's answers, and gradually guides autistic children to complete a complete dialogue chain; the module also integrates a cognitive assessment system, analyzes the user's learning progress through voice recognition and behavior logs, generates a multi-dimensional ability map and pushes customized practice content.

Images