Intelligent psychological health education system and method for realizing home-school interaction and subject fusion based on education robot

Abstract

The invention discloses an intelligent psychological health education system based on an education robot. The system comprises (1) AI family-school cooperation: an improved YOLOv7 algorithm captures classroom behaviors, dynamic Gaussian blur protects privacy, directionally pushes student participation images containing desensitization IDs, supports AI to generate personalized communication verbal skills, and improves the degree of participation of parents; (2) an interdisciplinary fusion engine: constructing a knowledge graph to associate psychological courses with mathematics, sports and other subjects, predicting an optimal fusion opportunity in combination with educational administration data, and generating a double-teacher collaborative teaching plan; (3) interest and emotion driven daily penetration: analyzing student interest labels based on a latent semantic model (LFM), and triggering personalized family task pushing in combination with multi-modal emotion recognition to realize intelligent penetration of psychological health education in a family scene; and (4) full-link privacy protection: adopting differential privacy, block chain and federated learning technologies to realize'all anonymity 'and'availability and invisibility' of data, and ensuring information security through three-level desensitization and quantum encryption, so that the system meets the requirements of'personal information protection method '.

Description

Technical Field

[0001] This invention belongs to the application technology of artificial intelligence in the field of psychological education, and complies with the requirements of the "Personal Information Protection Law" and the "Education Data Security Management Specification". Background Art

[0002] 1. Students face prominent mental health issues: A report from the Institute of Psychology of the Chinese Academy of Sciences shows that 14.8% of primary and secondary school students are at risk of depression (23.0% for junior high school students and 13.0% for primary school students), and the depression risk detection rate for college students is as high as 24.6%. There is an urgent need for mental health education.

[0003] 2. Shortage of psychological education resources in schools: A 2024 survey in Nanjing, Jiangsu Province pointed out that 57.4% of schools had "insufficient professional training of psychological teachers" and 53.7% lacked professional psychological teachers. Physical educational robots have become an important solution to alleviate the pressure on teachers.

[0004] 3. Pain points in home-school interaction are prominent: A 2023 survey by the Chinese Academy of Educational Sciences showed that 83% of parents prioritized "photos or videos of their children" when participating in school activities. For example, parents at one Zhejiang elementary school downloaded five times more photos of activities than text reports. Approximately 12% of parents opposed mental health programs due to their prioritization of academics, reflecting a divergent philosophy between home and school.

[0005] 4. Policy and Theoretical Support: The "Guidelines for Mental Health Education in Primary and Secondary Schools" call for cultivating students' ability to withstand setbacks and emphasize the integration of "topical discussion" and "disciplinary integration" (e.g., discussions on mental health in sports). Emotional management skills are a core protective factor for mental health, and emotional intelligence (EQ) is crucial for psychological growth.

[0006] 5. Existing technology is flawed and interdisciplinary integration is insufficient: Traditional mental health education curricula are separate from subjects like mathematics and physical education, making core content like "properly viewing winning and losing" difficult to incorporate into daily instruction. Inefficient home-school collaboration: Unified content push leads to low parent engagement, with 65% of parents lacking effective response strategies for scenarios like "test failure." High privacy risks: Student behavioral data is not effectively anonymized, and traditional storage methods present a 73% risk of privacy leakage ("Education Data Security White Paper 2024").

[0007] 6. Background of the Invention Upgrade: This invention is an upgraded version of an existing patent (CN106558254B) by the same applicant. During testing, the patent (CN106558254B) exposed issues with resource integration and personalized adaptation. For example, the system used text messaging to push information about joint home-school mental health education, but parental participation was low. Some parents even openly expressed opposition in parent groups, believing that this would interfere with learning of the core curriculum. This invention achieves a breakthrough by leveraging artificial intelligence technology. It does not involve medical diagnosis, but focuses on supporting mental health education.

[0008] 7. Definitions: Physical Education Robot: An intelligent terminal device (humanoid or non-humanoid) deployed in key campus areas (such as the school gate, dormitory entrances, and teaching building entrances). Based on artificial intelligence algorithms and a database of mental health education resources, it provides support for mental health education. Both the classroom and home versions feature a miniaturized design for portability and companionship. Software Education Robot: An intelligent program system running on computers, mobile terminals, and smart wearable devices, integrating natural language processing (NLP), machine learning (ML), and robotic process automation (RPA) technologies, supporting deployment on multiple client types. Summary of the Invention

[0009] In order to solve the technical problems in the background, the system of the present invention includes a physical educational robot and a basic system, a software educational robot, a home-school interactive psychological education module, an interdisciplinary integration module for mental health education, a parent collaborative penetration module and an information security assurance module.

[0010] Physical educational robots (including miniaturized versions for each classroom) are deployed in key areas of the campus to support mental health education courses and assist psychology teachers in improving educational effectiveness. The system strictly protects student privacy as a third party and does not store personal information such as names and ID numbers, eliminating the risk of leakage at the source.

[0011] 1. Physical educational robots and basic systems This invention provides an intelligent educational robot system and data management method that combines campus mental health education capabilities with full lifecycle privacy protection. The core contents are as follows: 1.1 System Architecture and Core Technologies Core Innovation: A multidisciplinary model integrating mental health education, home-school interaction, and parent-teacher collaborative education. Anthropomorphic Interactive Hardware: Integrating a fully laminated touchscreen, a 3D bionic face, micro-expression motors, and pressure sensors, the system enables natural human-computer interaction through flexible touch and dynamic expression simulation. Mental Health Education Basic Module: Built-in expert courses on mental health education, supporting live streaming and playback of videos from experts and in-school teachers.

[0012] 1.2 Hot-swappable hardware expansion architecture: A layered expansion architecture that combines hot-swappable and non-hot-swappable hardware is used to achieve dynamic configuration of hardware resources.

[0013] 1.2.1 Composite physical interface design: The anti-misinsertion structure integrates multi-protocol interfaces such as Universal Serial Bus (USB), PCI-e (PCI-Express), and Serial ATA 4.0 (SATA 4.0), and is compatible with IEEE 1394b interface.

[0014] 1.2.2 High reliability technology system Hardware protection: Gold-plated contacts and guide key design ensure a plug-in / plug-out lifespan of more than 100,000 times; integrated overvoltage / overcurrent protection circuits and electrostatic discharge (ESD) protection diodes enable power sequencing and signal integrity protection.

[0015] Software Mechanism: Based on the Linux kernel's udev dynamic device management system, it supports graceful shutdown when the device is hot removed; the microkernel architecture supports dynamic loading / unloading of modules and is equipped with a role-based access control (RBAC) security mechanism to prevent unauthorized access.

[0016] 1.2.3 Standardized modular specifications The mechanical interface complies with the International Electrotechnical Commission IEC 61076-4-101 standard dimensions; a unified three-layer communication protocol stack (physical layer, data link layer, application layer) ensures cross-platform compatibility; and a multi-parameter real-time monitoring system is established to ensure module operation stability.

[0017] 1.2.4 Open scalability: Supports over-the-air (OTA) upgrade technology and dynamically adapts to future new interface protocols; the module complies with CE / FCC certification standards and meets EN 55032 Class B electromagnetic compatibility requirements.

[0018] 1.3 Functional Advantages and Applications, Multi-Scenario Coverage: Supports core scenarios such as campus mental health education and home-school interaction. Typical Applications: Schools can choose hot-swappable modules such as cameras, displays, 3D and standard projectors, players, VR glasses, and more as needed, enabling flexible device expansion.

[0019] 1.4 Data Integration and Privacy Protection Multi-source data fusion technology (five-layer hybrid cloud architecture): Data integration layer: Access to heterogeneous data sources such as security systems, smart classrooms, and electronic blackboards through the Real-Time Streaming Protocol (RTSP) and the Education Private Network Protocol (EDU-VPN); Spatiotemporal alignment layer: The Beidou timing module (±10 nanosecond accuracy) is combined with the Apache Beam window function to achieve spatiotemporal alignment of multi-source data; AI analysis layer: Uses a three-dimensional convolutional network (I3D) to process video features and a Wav2Vec 2.0 model to parse audio data, and integrates the visual transformer (ViT-Large / 16) and speech representation model (HuBERT-Large) to build a multimodal pre-training model.

[0020] 2. Physical / software educational robots for home-school interactive psychological education This module achieves home-school collaboration in mental health education through intelligent recording, privacy protection, and personalized content generation. Its core functions are as follows: 2.1 Overview of Core Functions Physical / software educational robots have built-in psychological education modules (including expert-level courses) or access them through a server. They support learning through display screens, projection devices, or virtual reality (VR) glasses. They can livestream and replay expert and class teacher lessons in real time, and simultaneously display them on an electronic blackboard. Taking into account the psychological connections within families, the system pushes course content to parents' mobile devices, enabling simultaneous home-school discussions and learning.

[0021] The system intelligently generates images, videos, or animations of each student participating in a course (or mental health activity), along with the student's voice, to attract parents to the program. For resistant parents, information substitution and educational materials are used to attract them.

[0022] 2.2 Mechanisms to Improve Parental Participation 83% of parents only care about their children's participation in the course. In addition, 12% of parents are opposed to taking mental health education classes (see Background Technology Document 3, etc.). This invention achieves the following through AI upgrades: Intelligent Image Capture: With parent / student authorization, a physical educational robot can capture photos / videos of the class, documenting every moment of student participation and interaction. If a physical educational robot is unavailable, teachers can place their smartphone on a rotatable stand, which the software-controlled educational robot can use to capture the entire class. Alternatively, teachers can ask volunteer parents to use their smartphones in conjunction with the software-controlled educational robot to capture the class.

[0023] Desensitized storage: The original information is desensitized (ID is used instead of name), and only the desensitized ID (such as "CLS2024001_STU03") is stored without personalized identification information such as name; Precision push: The system pushes desensitized children's course images (real-time / historical) to the corresponding parents' mobile devices. It supports teachers to add virtual decorations (such as window frames and cartoon hats) or synthesize interactive voice through AI editing, including: "Mom and Dad, I hope you will volunteer to attend the mental health education class with me." “Mom and Dad, I hope you will volunteer to take a positive psychology class with me.” "Mom and Dad, do you know how to help me deal with winning and losing correctly?" Different scripts can be generated for each class to further attract parents to voluntarily participate in interaction and learning.

[0024] 2.3 Technical Solution of Multimodal Mental Health Teaching System This system supports multi-terminal immersive learning, and its core design is as follows: Multi-device access: Course learning can be achieved through display screens, projection devices or virtual reality (VR) glasses, and devices can be expanded through hot-swappable interfaces; Intelligent live broadcast / playback: Using AI video encoding technology, synchronize expert courses to the electronic blackboard (supporting breakpoint resumption and knowledge point annotation), or synchronize the teacher's courses in the class to the home-school interaction platform.

[0025] 2.4 AI-driven home-school collaboration optimization solution 2.4.1 Intelligent Visual Capture and Dynamic Privacy Protection With parent / student authorization, the system uses the following technologies to achieve image acquisition and privacy protection: AI vision sensor: Equipped with a binocular camera (1920×1080 resolution, 30fps frame rate), an infrared depth sensor (0.5-4m effective range), and a 6-channel MEMS microphone array to construct a 3D classroom coordinate system and collect audio. Improved YOLOv7 object detection algorithm: Based on YOLOv7, it adds a spatial attention module (SAM). After training with more than 200,000 labeled samples (including 12 types of behaviors such as raising hands and standing), the detection speed is increased by 20%, and the mean average prediction accuracy reaches 94.3%. Based on the Visual Transformer (ViT) facial processing engine: The 12-layer encoder extracts global facial features (1024-dimensional vectors), and the cosine similarity between target and non-target pupil features is less than 0.3, achieving accurate differentiation; Dynamic blur algorithm: Only the target student's face is clearly presented (recognition ≥ 85%). The remaining student faces are differentially protected through dynamic radius Gaussian blur (formula: \( R = 5\sqrt{d} \), \( d \) is the relative distance) + Laplace noise injection (privacy budget ε = 0.1). The peak signal-to-noise ratio (PSNR) of non-target faces is ≤ 20dB.

[0026] 2.4.2 Hierarchical Authorization and Data Collection Architecture Two-way authorization mechanism: Parents / students sign the "Classroom Image Collection Authorization Agreement" through fingerprint / face biometric authentication, clarifying the scope, term, and purpose of authorization; the system generates a unique desensitized ID, encrypts and stores real information through blockchain nodes, and realizes irreversible mapping of "ID-name"; Sensitive information filtering: Using OCR technology to blur student names, student ID numbers, and other information within the classroom, hiding indirect identifying features such as clothing brands and school emblems; Desensitized storage specification: The original data format is "STU_{UUID}@YYYY-MM-DD_HH:MM:SS.jpg", and the metadata includes the desensitized ID, location coordinates, and behavior tags (such as "writing").

[0027] 2.4.3 Personalized Content Generation and Adaptive Interaction Parent profile construction: Integrate 15-dimensional behavioral data (including basic attributes, interactive behaviors, and resistance characteristics) and use the XGBoost classifier (with an accuracy rate of 93.2%) to identify "resistant" parents; Dynamic wording adjustment: Based on the mental health knowledge graph (containing 5,000+ triples), automatically replace "mental health education" with "emotional intelligence education" for "resistant" parents. Example: "Invite you to participate in the mental health education class" → "Invite you to participate in the emotional intelligence education class."

[0028] With the authorization of the parents, we can record and analyze the calling language and scene that the parents like based on their participation in interactive courses, and replace different calling languages ​​and backgrounds.

[0029] Popular Science Content Generation: We fine-tuned the GPT-3.5 model to generate explanatory text (e.g., "Emotional intelligence is an important component of mental health"), achieving a BLEU score of 0.68 and an artificial naturalness score of 4.2 / 5. AI editing function: supports virtual scene addition (generating cartoon decorations through StyleGAN3, SSIM ≥ 0.92) and voice and action synthesis (VITS model generates voice, MOS ≥ 4.0; MAESTRO algorithm synthesizes action, synchronization error ≤ 80ms).

[0030] 2.4.4 Technical Test Results and Innovations |Indicators|Achievement Results|Test Standards| |Object detection accuracy|mAP≥94.3% |Classroom scene dataset| |Single frame processing delay|<80ms (30-person classroom test)|Real-time performance test| Privacy protection level: Meets ε-DP (ε=0.1) and NIST SP 800-200-2. | Parent course acceptance rate | Increased from 17.3% to 73.8% | Test of 100 "resistant" parents | Innovation: (1) Integration of privacy and individual focus: The world's first "single student clarity" processing of classroom images (targets are identifiable, non-targets are not identifiable); (2) Intelligent guidance of cognitive bias: combining knowledge graphs with NLP technology to dynamically convert educational terms and generate explanatory content to address parental resistance; (3) Full-process AI closed loop: Building an intelligent collaborative system of "learning-recording-pushing-interaction", the average parent participation time increased from 4.2 minutes to 28.5 minutes.

[0031] 2.5 Closed-loop real-time interaction and data security After class, a "parent remote interaction session" is launched, where parents can watch course clips and voice discussions in real time through mobile devices, and retrieve their children's historical participation records; blockchain technology is used to encrypt data to ensure that only authorized parents can access exclusive content (passed NIST SP 800-22 randomness test).

[0032] This solution uses core technologies such as NLP, knowledge graphs, generative adversarial networks, and multimodal synchronization to systematically address problems such as "parent resistance" and "monotonous content" in home-school interaction. It combines technological innovation and market application value in terms of privacy protection and increased participation.

[0033] 3. Intelligent interdisciplinary integration and daily penetration of mental health education This module builds a multimodal knowledge graph and an intelligent recommendation engine to achieve intelligent association between mental health education and subjects such as mathematics and physical education, dynamic integration of teaching content, and personalized penetration, solving the technical difficulties of traditional mental education such as isolation, formalization, and coping with inspections. At the same time, it adopts differential privacy, blockchain and other technologies to ensure the privacy security of students.

[0034] The system intelligently analyzes the syllabus and themes of each psychological education course and matches them with other subjects. For example, referring to the Ministry of Education's syllabus, a middle school designed its first positive psychology lesson, "How to Correctly View Winning and Losing," which focuses on the positive psychology chapter in its primary and secondary school psychological education curriculum. The system intelligently searches for content related to winning and losing in other subjects, conducting an expanded intelligent search, including methods of winning and losing, the probability of winning and losing, the consequences of winning and losing, and sports or activities involving winning and losing. By analyzing the course schedules in the academic administration system, it determines how to integrate psychological education into teaching. If a math course includes probability content, the system sends an integrated teaching message to the math teacher's smartphone app. When the math teacher teaches probability, the psychological education teacher integrates the math content into the course and explains how students should correctly view winning and losing. If a physical education course includes competitive content such as basketball, the system sends an integrated teaching message to the physical education teacher's smartphone app. After the physical education teacher organizes a class basketball game, the psychological education teacher integrates the physical education content into the course and explains how students should correctly view winning and losing.

[0035] A case study of interdisciplinary penetration: Math class: Using a data chart showing the probability of athletes winning or losing, explain that "winning or losing is a probabilistic event, and continuous effort can improve the odds of winning." This can be explained by a psychology teacher on-site, or by live broadcasting or replaying online expert explanations. Physical Education Class: Introduce the "fair competition principle" in the textbooks into basketball games. After the game, ask players from both teams to tell each other three "honorable" performances of the other team. The explanations can be given on the spot by a psychology teacher, or by live broadcast or replay of the explanations of online experts.

[0036] Building class culture: The system intelligently reminds the class teacher to establish the "Best Progress Award" and "Courage to Attempt Award", and the award speeches quote the core ideas of the textbook (for example: "Having lost 10 times and still daring to challenge for the 11th time, this is the victory of growth").

[0037] The system focuses on protecting student privacy. The information stored in this module is desensitized, and personalized information such as names are replaced by ID numbers. Encryption and privacy protection technologies are used to meet the requirements of various laws and regulations.

[0038] 3.1 Innovative Paths to Achieve Multidisciplinary Integration Breakthroughs 3.1.1. Intelligent Association: Leveraging natural language processing (NLP) and graph neural networks (GNNs), we automatically analyze psychology course topics and match them with subject content (e.g., linking "correctly viewing winning and losing" with mathematical probability and sports). 3.1.2. Dynamic Penetration: Combined with academic data and time series analysis, accurately predict the optimal timing for psychological education penetration (e.g., triggering push notifications two hours before a math probability lecture); 3.1.3. Model Innovation: Design standardized penetration templates (mathematical data charts, sports competition evaluations) and personalized recommendations (adapting to teachers' teaching preferences) to improve the efficiency of interdisciplinary teaching; 3.1.4. Security and Compliance: Integrate technologies such as differential privacy and blockchain to make student data “available but invisible”, in compliance with the requirements of the Personal Information Protection Law and the Education Data Security Management Specification.

[0039] 3.2 System Architecture Innovation 3.2.1 Intelligent Semantic Parsing Module Function: Perform fine-grained analysis of the psychological education curriculum outline to extract core keywords (such as "win-lose perspective") and their associated dimensions (win-lose method / probability / result).

[0040] AI technology: Using a bidirectional long short-term memory network (Bi-LSTM) combined with the attention mechanism (Attention Mechanism), the course text is segmented, entity recognized (such as "win or lose", "probability") and sentiment analyzed to extract the core teaching objectives; based on the BERT pre-training model, high-dimensional semantic vectors are generated to construct a multi-level feature space for psychological themes (such as breaking down "positive psychology" into sub-dimensions such as "emotional management", "resistance to frustration", and "correct view of winning and losing").

[0041] 3.2.2 Construction of Interdisciplinary Knowledge Graph Three-layer mapping structure: 1. Conceptual layer: Use word embedding (Word2Vec) to establish semantic associations between "win or lose" and "mathematical statistics" and "sports competition," and calculate cosine similarity (e.g., the similarity between "win or lose probability" and "binomial distribution" is ≥ 0.92). 2. Case layer: Stores over 500 interdisciplinary cases (e.g., "Athlete win-loss rate statistics" in mathematics, "Post-game evaluation of basketball games" in sports), and supports intelligent retrieval of multimedia resources such as videos, charts, and test questions. 3. Strategy layer: A pre-set fusion strategy matrix is ​​built, which includes 12 parameters, such as the timing of collaboration (before / during / after class) and presentation method (live broadcast with two teachers / recorded playback / augmented reality (AR) demonstration). The strategy combination is dynamically optimized through a graph neural network (GNN).

[0042] 3.3 Core Technology 3.3.1 Dynamic Course Matching and Timing Prediction Data connection: Real-time access to the course plan database of the academic affairs system to extract course schedules (such as "Mathematics Probability Class in Week 3" and "Physical Education Basketball Class in Week 5").

[0043] Algorithm implementation: The correlation between psychological topics and subject content is calculated based on an improved collaborative filtering algorithm, and a threshold of >0.85 is set to trigger a recommendation. The temporal characteristics of course scheduling are extracted through a time series analysis model, and the optimal fusion time is predicted by combining the support vector machine (SVM) classification model, with the error controlled within ±1 class hour.

[0044] 3.3.2 Multi-terminal intelligent push system Context-aware push notifications: Based on parameters such as the teacher's geographic location and course progress, precise push notifications are implemented through instant messaging application programming interfaces (APIs, such as WeChat / DingTalk) (e.g., sending a "post-match review template" 10 minutes after a sports match).

[0045] Personalized adaptation: Combined with teachers' historical lesson preparation preferences (such as "Mathematics teacher A prefers chart teaching"), the push content format (data charts / case videos / interactive tasks) is dynamically adjusted through natural language generation (NLG) technology.

[0046] 3.4 Innovation in teaching implementation 3.4.1 Intelligent teaching auxiliary module Automatic generation of lesson plans: Arrange a timeline for dual-teacher collaboration (psychology teachers explain theory → subject teachers demonstrate cases). For example, insert a data chart on "Statistical Probability of Athletes Winning or Losing" into a math probability class, and have the psychology teacher explain "Continuous Efforts to Improve Winning Rates" on the spot or through live broadcast. Integrate augmented reality (AR) visualization tools to support 3D dynamic presentation of the relationship between probability statistics and win-loss concepts (such as the visual mapping of the winning rate calculation formula and the degree of effort), and support various devices such as 3D projection.

[0047] Standardized penetration template: Mathematics: Automatically generate charts for "NBA player season win percentage" and "student intramural game data," and combine them with the probability formula (win percentage = number of wins / total number of games) to instill the concept that "winning and losing are probabilistic events"; Physical Education: After a basketball game, a "glorious defeat" evaluation template (including dimensions such as "teamwork" and "defensive enthusiasm") is provided to guide students to praise each other's strengths.

[0048] 3.4.2 Cultivating Class Culture Intelligence Based on the student behavior analysis model (such as detecting that a student has not given up after 10 attempts), the rule engine triggers reminders of the "Courage to Attempt Award" and "Best Progress Award"; combined with natural language generation (NLG) technology, personalized award words are automatically filled in (such as "Having lost 10 times and still daring to challenge for the 11th time, this is the victory of growth - Xiao Ming improved by 5 places in the mathematics competition"), and the core ideas of the textbook are quoted to strengthen growth thinking.

[0049] 3.5 Summary of Innovation Plan 3.5.1. Intelligent Matching of Multi-Technology Integration: For the first time, it integrates technologies such as natural language processing (NLP), graph neural networks (GNN), and bidirectional long short-term memory networks (Bi-LSTM) to achieve semantic-level association between psychological themes and subject content (matching accuracy increased to 91.3%), solving the problem of low efficiency of traditional manual matching; 3.5.2. Dynamic Prediction and Personalized Recommendation: Through time series analysis and support vector machine (SVM) models, the prediction error of fusion timing was controlled within ±1 class hour, and collaborative filtering algorithms were used to adapt to teacher preferences (the usage rate of recommended content increased by 78%). 3.5.3. Scenario-based and standardized teaching implementation: Build a penetration template library covering 12 subjects such as mathematics and physical education, support diversified teaching methods such as dual-teacher collaboration and augmented reality (AR) demonstrations, and achieve a natural transformation of psychological education from theory to practice; 3.5.4. Privacy Protection Technology Integration: Integrate technologies such as differential privacy, blockchain, and federated learning to establish a full-chain security system for "collection-storage-application". The privacy protection level complies with the ISO / IEC 27001 standard, and the data anonymization rate is 100%; 3.5.5. Closed-loop feedback and effect optimization: Deploy the Affective Computing module to analyze students' micro-expressions through classroom videos (for example, concentration increased by 25% when explaining the "probability of winning or losing"), continuously optimize the fusion strategy, and form a closed loop of "data collection-effect analysis-strategy iteration."

[0050] 3.6 Beneficial effects Testing environment, teaching effectiveness improved: students' understanding of "correctly looking at winning and losing" increased from 38% to 78% (data from a certain experimental school in 2024), and the average score on the anti-frustration ability assessment increased by 32.6%; the coverage rate of interdisciplinary teaching increased from 22% to 89%, and the natural integration rate of psychological education and subject scenarios reached 76%.

[0051] Improved teaching efficiency: The speed of generating collaborative teaching plans for two teachers has increased by 87%, supporting full-process penetration before class (teaching plan push), during class (AR assistance), and after class (cultural cultivation).

[0052] Data security assurance: The anonymization processing rate of student behavior data is 100%, no privacy leakage incidents have occurred during the nearly one-year testing period, and it has passed the national information security level protection level 3 certification; the teaching data tampering detection rate is 100%, which complies with the requirements of the "Personal Information Protection Law" and the "Education Data Security Management Specifications".

[0053] 4. Mental health education intelligence and parents’ daily penetration Psychological research has found that fostering positive psychological thinking based on students' diverse interests and hobbies is more effective. After completing psychological education courses like "How to Correctly View Winning and Losing," it's important to incorporate the concepts of the course into daily family education, translating them into behavioral habits and truly cultivating a correct perspective on winning and losing.

[0054] With the consent of parents, the system intelligently matches family activities based on student interests and themes of winning and losing. For example, for students who enjoy playing chess, the system pushes personalized intensive training courses to the parent's smartphone client and generates a "Family Win-Loss Challenge" manual containing the following content: Interest-based tasks: If a student is good at checkers and rarely loses, parents are advised to deliberately let their child lose a few times during parent-child games, observe their reactions, and guide them to reflect: "Although you lost, the strategy of this move was very special. Let's try to optimize it again?" Ask students to record their "losing" experiences and redefine failure with "I gained ___ experience" or "I found ___ needs improvement" to weaken the self-labeling of the results.

[0055] Cross-media learning tasks: If students like watching movies, we recommend watching films such as "Moneyball" and "The Pursuit of Happyness" and discussing the character experiences in combination with the "win-lose values" in the textbook.

[0056] The system also combines test scores from the academic affairs system to generate a guide checklist for parents: providing scientific communication skills on "how to respond to children's emotions about winning or losing" (for example, when a child does poorly on a test, parents are advised to say, "Mom saw that you got 3 fewer wrong questions in this unit than last time. That's progress!"), and reminding parents to avoid an all-out blaming approach to education.

[0057] Psychological research shows that students are more receptive to the "correct values ​​of winning and losing" when they are in a positive emotional state. When parents activate the educational robot software on their smartphone, the system uses technologies such as eye tracking and voice analysis to identify their child's emotional state, ensuring that educational content is delivered during a time when they are most likely to absorb it.

[0058] The emotion module only identifies students' positive emotional states and does not store any information to avoid the leakage of personal privacy.

[0059] In response to possible psychological education questions or mental health issues that parents may have, the system supports parents to seek anonymous consultation through software education robots or connect to online psychology experts for anonymous consultation after authorization, building a two-way support system of "student psychological development and parent capacity building".

[0060] 4.1 Innovative routes to achieve technological breakthroughs in home-school collaboration Interest-based precise matching: Generate personalized home tasks based on students' interests, solving the adaptation problem of traditional education that is "one size fits all"; Deep integration of academic data: Dynamically generate parent communication strategies based on test scores and other academic data to improve the practicality of "post-secondary education"; Closed loop of parental psychological support: Through software-based educational robots and remote expert systems, we achieve two-way empowerment of "student psychological development and parental capacity building"; Privacy protection throughout the entire process: Using technologies such as federated learning and k-anonymity to ensure that home scenario data is "available but not visible," complying with the requirements of the Personal Information Protection Law.

[0061] 4.2 Deep Mining Module for Students’ Interest Characteristics 4.2.1 Multimodal Interest Modeling Technology Data fusion: Natural language processing (NLP) is used to analyze students' speeches in mental health classes and texts of home-school interactions. Combined with classroom behavior data collected by intelligent educational robots (such as raising hands to answer "favorite sports") and images of student works analyzed by image recognition (CV), an interest vector containing 12 dimensions (such as "board games", "inspirational movies", and "team competition") is constructed.

[0062] Dynamic tag updating: Using the latent semantic model (LFM) to learn students' preference in the "Family Win or Lose Challenge" in real time, the weight of interest tags is dynamically adjusted (for example, a special task is triggered when the weight of the "playing chess" tag is ≥0.7). The interest recognition accuracy rate reaches 92.3% (an increase of 23.8% over traditional methods).

[0063] 4.2.2 Intelligent Matching Algorithm for Winning and Losing Scenario A knowledge graph for "interest-based win-lose activities" was constructed (including over 500 family-implementable scenarios, such as checkers, puzzle competitions, and parent-child sports). A graph neural network (GNN) was used to calculate the correlation between interest tags and activity types (threshold ≥ 0.85). For example, the "board games" interest tag was automatically matched to the "checkers parent-child competition" task, generating a standardized process for "deliberately making children lose games + strategic guidance" and linking it to the curriculum's "process over results" philosophy.

[0064] 4.3 Dynamic Task Generation Engine 4.3.1 Personalized Task Generation System Dynamic template library: 30+ preset family task templates (such as "Reflection on Game Wins and Losses" and "Discussion on Movie Values"), and call corresponding templates based on interest tags (such as "Movie Lover" automatically matches the "Moneyball" movie viewing task).

[0065] AI text generation: By fine-tuning the GPT-3.5 model and integrating student classroom interaction data, personalized task descriptions are generated. For example, for the "Always Winning at Chess" task, the following student description is generated: "This week's task: Play three games of checkers, intentionally winning one. Instruct the child to record their 'strategies attempted when losing' and complete the 'Failure Experience Gain Sheet': I gained ____ experience, and I found ____ needs improvement." 4.3.2 Multimodal Content Output Rich media resource library: Integrates over 500,000 inspirational film clips and over 200,000 parent-child game guides. It uses a generative adversarial network (GAN) to automatically extract film and television clips that match educational goals (such as the "persistence after losing" theme clip), allowing parents to receive tasks in the form of pictures, text, and videos.

[0066] Progress tracking: Task completion is recorded through blockchain technology to generate a "family psychological growth file" for students, which only displays desensitized ID-related data to ensure privacy and security.

[0067] 4.4 Parent Collaborative Agent and Dynamic Intervention System 4.4.1 Deep Integration of Educational Affairs Data and Script Generation Academic performance analysis: Real-time access to the academic administration system allows machine learning models to identify "exam failure" scenarios (e.g., grade fluctuations >20%) and build a knowledge base of 200+ scientific parent response strategies.

[0068] Adapting teaching methods: Support vector machines (SVMs) are used to match the best response. For example, if a student performs poorly on a test and their parents are critical, a recommended response might be: "Mom noticed you made three fewer mistakes this unit than last time. That's progress! How about we analyze the mistakes together, combining this with adjustments to our chess strategy?" Real-time intervention: By pushing a "post-loss education guide" through the instant messaging application programming interface (API, such as WeChat service account), which includes emotional soothing steps and specific improvement suggestions, parents' response timeliness has increased by 40%.

[0069] 4.4.2 Parent Psychological Support Module Software educational robot: Based on natural language processing (NLP) technology, it supports parents' questions such as "What should I do if my child can't afford to lose?" and retrieves matching answers through knowledge graphs (such as "It is recommended to set up a 'Strategy Innovation Award' so that even losing a game can earn recognition"), with a response time of less than 15 seconds.

[0070] Remote expert connection: After parental authorization, the system synchronizes student interest data and task completion status, connects to online experts through video calls, and the accuracy rate of abnormal psychological state warning reaches 89.7% (false alarm rate 6.2%).

[0071] Parent Mental Health Assessment: Deploy the simplified version of the PHQ-9 assessment tool, analyze parent emotional texts through machine learning, and push mindfulness meditation audio, parent mutual support community links, and other resources when depressive tendencies are detected (score ≥ 10 points).

[0072] 4.5 Multimodal Emotional Intelligence Adaptation System 4.5.1 Multimodal Emotional State Recognition Module Eye tracking and concentration analysis: The front-facing camera of the parent's smartphone is equipped with an active infrared eye tracking module to collect 6-dimensional data such as pupil diameter and gaze trajectory; A lightweight convolutional neural network (MobileNetV3-Small) is used to locate the eye area, and a Kalman filter is used to track pupil movement to generate an eye movement concentration score (0-1, with a threshold ≥0.7 indicating high concentration).

[0073] Facial expression recognition technology: The MediaPipe face detection framework locates 68 facial key points and extracts 12-dimensional expression features. The ResNet-50 neural network is used to classify expression categories (happy, neutral, and depressed), with an accuracy rate of 91.2% for positive expression recognition.

[0074] Speech Tone Analysis System: This system collects speech signals and extracts 20-dimensional acoustic features, including Mel-Frequency Cepstral Coefficients (MFCCs) and fundamental frequency. It uses the Transformer-based speech emotion model (Speech-Transformer) to analyze tone tendencies and output a positivity score (0-1, with a threshold of ≥0.6 indicating a positive tone). Single-frame processing time is less than 50ms.

[0075] Multimodal fusion decision: Build a weighted fusion model (eye movement 0.4 + facial expression 0.3 + voice 0.3) and set positive emotion judgment rules: When eye movement concentration is greater than 0.7, facial expression pleasure is greater than 0.6, and voice positivity is greater than 0.6, a "positive emotional state" is triggered.

[0076] 4.5.2 Dynamic Adaptive Educational Content Generation System Positive Emotional Response Strategy: Build a resource pool of "correct values ​​of winning and losing," including gamified task templates (such as a "checkers strategy review table") and a library of film and television clips, which are dynamically accessed based on students' interest tags; When positive emotions are detected, customized guidance words are generated through the GPT-3.5 fine-tuning model, and the content is dynamically filled in based on classroom knowledge points.

[0077] Effect-enhancing feedback mechanism: Establish an "emotional engagement" feedback loop, analyze historical data through federated learning, dynamically adjust task push frequency and content type, and improve the efficiency of educational content matching.

[0078] 4.6 Innovation Interest-driven personalized penetration technology: For the first time, it integrates natural language processing (NLP), image recognition (CV) and graph neural network (GNN) to build a dynamic interest graph, achieve precise matching of "interest-based winning and losing activities" (participation rate increased to 78%), and solve the problem of low adaptability of traditional psychological education.

[0079] Semantic-level fusion technology for educational data: Dynamically generates parent communication scripts through support vector machines (the scientific response rate has increased from 28% to 69%), transforming academic data into actionable family education strategies, filling the technological gap in "post-education".

[0080] Closed-loop architecture for parental psychological support: Integrating software educational robots and remote expert systems to build a two-way empowerment model of "student behavior analysis, parent strategy recommendation, and psychological state warning", which improves the efficiency of resolving parents' psychological confusion by more than 50%.

[0081] Innovative application of multimodal emotion recognition: For the first time, this system combines eye tracking, expression recognition, and speech analysis technologies to build a positive emotion recognition system for family scenarios, increasing the acceptance of content on "correct values ​​of winning and losing" by 40%.

[0082] Privacy protection technology system: Federated learning, k-anonymity, and blockchain technology are applied to make household data “available but invisible.” The privacy protection level complies with the ISO / IEC 27001 standard, and no data leakage occurred during the pilot period.

[0083] 4.7 Beneficial Effects Results of the educational experiment: The frequency of students applying the "correct concept of winning and losing" in family scenarios increased by 3.2 times, and the incidence of anti-frustration behavior increased from 19% to 65%; the duration of parents' "positive interaction after losing" increased by 2.8 hours per week, and family communication conflicts decreased by 41%.

[0084] Technical efficiency: Task package generation takes <10 seconds per time, and the matching accuracy of the parent guidance list reaches 92.3%, which is 80 times higher than the efficiency of manual formulation; the software education robot responds 24 hours a day, 7 days a week, and the success rate of remote expert docking is 98%.

[0085] Compliance and security: The anonymization rate of student interest data is 100%, the encrypted storage of parent consultation records meets the national Level 3 security protection requirements, and has passed the NIST SP 800-22 randomness test (pass rate 100%).

[0086] 5. Privacy protection and data security system This invention has built a privacy protection system covering the entire process of data collection, storage, and analysis. Through user voluntary authorization, multi-level desensitization processing, dynamic quantum encryption technology and refined authority management, it ensures the personal information security of students and parents, and complies with the "Personal Information Protection Law", "Data Security Law" and the third-level security requirements.

[0087] 5.1 Voluntary Authorization and User Self-Control Mechanism Prerequisites for enabling features: All privacy-related features (such as multimodal emotion recognition and home task data collection) require the user (parent / student) to sign an "Electronic Privacy Protection Agreement" through biometric authentication (fingerprint / face). This agreement specifies the scope of authorization (e.g., "only classroom participation images are allowed"), the usage period (one semester / school year), and the purpose of the data (limited to mental health education). After signing the agreement, the system generates a unique authorization code (valid for the same period as the agreement). Users who have not signed the agreement cannot use the relevant features.

[0088] The emotion recognition module only recognizes emotions and does not store them, thus protecting user privacy to the greatest extent possible.

[0089] Dynamic permission management: Parents can withdraw authorization at any time through the mobile terminal to trigger the data deletion mechanism: Locally stored data: In accordance with the DoD 5220.22-M standard, SSD storage media is overwritten 3 times + physically shredded, and HDD is overwritten 35 times using the Gutmann algorithm; Blockchain evidence data: Marked as "deleted" through smart contracts, it cannot be traced later, ensuring that user data is "forgettable".

[0090] 5.2 Full-link three-level desensitization processing technology 5.2.1 Collection Layer: Source Anonymization Sensitive information replacement technology: Differential Privacy technology is used to irreversibly anonymize sensitive information such as student names, ID numbers, mobile phone numbers, classes, and home addresses, generating irregular ID numbers (such as "STU-20240715-AB32").

[0091] The ID and real identity are mapped using the SM4 encryption algorithm certified by the National Security Administration, ensuring that no personalized information is retained at the source. The real identity information only contains the facial / fingerprint feature code, and no personalized information such as name is retained.

[0092] Multimodal data desensitization: For image data, the student's face is located using the improved YOLOv7 object detection algorithm, and only the clear image of the student associated with the desensitized ID is retained. The remaining faces are dynamically Gaussian blurred (PSNR ≤ 20dB). For voice data, Laplace noise (privacy budget ε = 0.1) is injected when extracting acoustic features such as Mel-Frequency Cepstral Coefficients (MFCCs) to meet the ε-DP differential privacy standard.

[0093] 5.2.2 Storage Layer: Encryption Isolation and Tamper-Proofing Dynamic ID mapping and blockchain evidence storage establish a dynamic mapping table of "ID-behavior data", which is stored on the campus local server or the server of a third-party authoritative organization. Blockchain technology is used to perform SHA-256 hash encryption on the mapping relationship, and the data tampering detection rate reaches 100%.

[0094] Hierarchical storage strategy: Basic data (such as desensitized IDs and task completion records): encrypted and stored in a local database, not uploaded to the public network; Core associated data (such as ID-interest tags, ID and real identity face or fingerprint feature codes): Dynamic keys are generated through quantum encryption technology (following the national secret SM9 algorithm). Once an intrusion is detected (such as abnormal login IP > 3 times), the key replacement mechanism is automatically triggered (replacement cycle < 1 minute).

[0095] 5.2.3 Application Layer: Data Available but Not Visible Federated learning-driven privacy computing utilizes federated learning technology to analyze the effectiveness of win-lose values ​​education. Each participant (school and parent) only uploads encrypted model parameters. Raw data (such as student test scores and homework records) is anonymized and stored locally, ensuring that data never leaves the school or home. A cross-domain collaboration mechanism utilizes zero-knowledge proof technology, allowing parents to prove to psychology teachers that their students have completed a task without disclosing the specific task content. Parents only provide an encrypted hash of the completion status.

[0096] 5.3 Refined Access Control System Role-Based Access Control (RBAC) defines three types of access roles with increasing permissions: Parents can only access their own children's data associated with desensitized IDs (such as classroom images and task completion records), and cannot view other students' information; Psychology teachers can access the overall desensitized data of their classes (such as class resilience statistics), but are prohibited from obtaining the mapping between individual student IDs and real identities; System administrators can only perform server maintenance, and data access requires two-person two-factor authentication (biometrics + dynamic token).

[0097] Real-time auditing and risk warning, deploying an intrusion detection system (IDS), real-time monitoring of data access behavior, and immediately freezing the account and sending an early warning notification (response time <10 seconds) when unauthorized access is detected (such as a psychology teacher trying to query data of a class he or she is not in charge of).

[0098] 5.4 Technological Innovations 5.4.1. Collaborative architecture of voluntary authorization and dynamic desensitization: For the first time, biometric authentication, quantum encryption technology, and differential privacy are combined in a mental health education system to achieve autonomous control of the entire "authorization-collection-storage" process, eliminating the risk of privacy leakage at the source. The entire information flow of the system is designed to be anonymous.

[0099] 5.4.2. Application of Federated Learning and Zero-Knowledge Proof in Educational Scenarios: Through privacy computing technology, collaborative analysis of home-school data can be achieved without sharing the original data, ensuring that all information is anonymous, thus breaking through the dilemma of "data silos" and "privacy leaks" in the traditional education system.

[0100] 5.4.3. Full-link security technology integration: Integrate blockchain anti-tampering, quantum key dynamic replacement, RBAC permission management and other technologies to build a privacy protection system that meets Level 3 security protection requirements, with technical indicators reaching internationally advanced levels (referring to ISO / IEC 27001 standards).

[0101] 5.5 Beneficial effects Privacy protection compliance: The anonymization rate of students' personalized information is 100%, the data collection compliance rate is 100%, and it has passed the national information security level protection level 3 certification, meeting the "anonymization processing" requirements of Article 27 of the "Personal Information Protection Law".

[0102] Data security and reliability: Blockchain evidence storage ensures that data tampering can be traced, and quantum encryption technology reduces the probability of key cracking to less than 1×10⁻¹ 00 Federated learning has achieved “data available but invisible”, and no privacy leaks have occurred in the pilot program over the past year.

[0103] User autonomy: The response time for parents to withdraw authorization is less than 30 seconds, data deletion complies with the "minimum necessary" principle, and user satisfaction with privacy functions reaches 98% (survey data from a certain experimental school).

[0104] This invention builds a privacy protection system covering all scenarios of mental health education through a technical chain of "voluntary authorization - three-level desensitization - privacy computing - permission management." Its core innovation lies in the deep integration of artificial intelligence and data technologies, which not only ensures the effectiveness of collaborative education between home and school, but also addresses the core issue of personal information protection from the perspectives of legal compliance and technical implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] 1. Figure 1 : Flowchart of the Intelligent Mental Health Education System Flowchart description: This shows the core innovative modules and internal processes of the intelligent mental health education system based on educational robots. (1) AI home-school collaboration: Using the improved YOLOv7 algorithm to capture classroom behavior, AI generates personalized communication scripts to enhance parent participation. (2) Interdisciplinary fusion engine: Generates dual-teacher collaborative lesson plans and supports visual teaching. (3) Daily penetration driven by personalized interests and emotions: Makes full use of students' personalized interests and positive emotions, and collaborates with parents to intelligently infiltrate mental health education into daily life. (4) Full-link privacy protection: Ensures information security through three-level desensitization and quantum encryption, and complies with the requirements of the Personal Information Protection Law.

[0106] 2. Figure 2 :Technical flow chart of physical educational robots and basic systems Description: Demonstrate the hardware architecture, hot-swap mechanism, and full-link privacy protection process of the physical educational robot. (1) The hot-swap layered architecture supports dynamic hardware expansion and is compatible with future devices; (2) The five-layer hybrid cloud architecture enables spatiotemporal calibration of multi-source data and AI analysis; Federated learning + blockchain technology ensures privacy and security throughout the entire life cycle.

[0107] 3. Figure 3 :Flowchart of the home-school interactive psychological education module Description: Demonstrates mechanisms for increasing parent engagement, focusing on AI image acquisition, dynamic privacy processing, and personalized content push. (1) The world's first "single student clarity" image processing technology balances privacy and parental concerns; (2) Dynamic language adjustment based on knowledge graphs increases parent course acceptance rate from 17.3% to 73.8%; (3) Multimodal content generation technology enhances interactive appeal, increasing parent engagement time by 6 times.

[0108] 4. Figure 4 : Flowchart of interdisciplinary integration of mental health education Description: Demonstrates the intelligent association and dynamic penetration process of psychology courses with subjects such as mathematics and physical education. (1) NLP+GNN technology achieves semantic-level association between psychology and subjects (matching accuracy of 91.3%); dynamic timing prediction and personalized recommendation increase the cross-disciplinary teaching coverage rate from 22% to 89%; (2) Standardized penetration templates reduce teachers' operating costs and promote the daily use of psychological education.

[0109] 5. Figure 5 : Parent Collaborative Penetration Module Flowchart Description: Demonstrates interest-driven personalized task generation, emotion adaptation, and parent support mechanisms in family scenarios. (1) Interest-driven precision matching technology increases family task participation to 78%; (2) Multimodal emotion recognition system increases educational content acceptance by 40%; (3) Deep integration of educational data and psychological education increases parental scientific response rate from 28% to 69%.

[0110] 6. Figure 6 :Flowchart of the full-link privacy protection system Description: Demonstrates privacy protection mechanisms throughout the entire data collection, storage, and application process, highlighting compliance and technical security. (1) Voluntary authorization and dynamic desensitization collaborative architecture eliminate privacy leaks at the source; (2) Federated learning + zero-knowledge proof achieves "data availability without visibility" and breaks through data silos; (3) Full-link security technology integration achieves privacy protection levels that meet ISO / IEC27001 standards.

[0111] Figure 7 :Improved YOLOv7 algorithm flow chart (including spatial attention module SAM) Illustration: (1) Input layer: receives classroom images with a resolution of 1920×1080 and a frame rate of 30 frames per second (FPS). (2) Backbone network: A cross-stage partial network (CSPDarknet53) is used to extract multi-scale feature maps (80×80, 40×40, and 20×20). After the spatial attention module (SAM) is inserted into the 3rd, 6th, and 9th cross-stage partial (CSP) modules, the structure is as follows: Channel splitting: The input feature map is divided into two groups (Group1 and Group2). Attention weight generation: Group1 is subjected to a 3×3 convolution → Sigmoid activation function to generate a spatial attention matrix (0-1 weight). Group2 is multiplied element-by-element by the attention matrix to enhance the features of high-response areas; Feature fusion: Group1 is concatenated with the weighted Group2 to output the optimized feature map; (3) Neck network: Path Aggregation Network (PANet) is used to fuse multi-scale features to improve small target detection capabilities; (4) Head network: The output layer predicts the target box (Bounding Box), category (12 types of behavior) and confidence through 1×1 convolution; the complete intersection over union loss function (CIoU Loss) is used to optimize the detection box accuracy.

[0112] (5) Technical annotation: SAM innovation: By focusing on dense student areas (such as raising hands and standing) through the spatial attention mechanism, the mean average precision (mAP) of target detection in a classroom scenario with 30 students is improved from 89.1% of the original YOLOv7 to 94.3%; real-time indicator: single frame processing time is less than 80 milliseconds (ms) (tested on NVIDIA Jetson Xavier NX platform). DETAILED DESCRIPTION DETAILED DESCRIPTION

[0113] 1. Physical Education Robot Hardware Architecture Anthropomorphic interactive hardware: Integrates a 10.1-inch fully laminated capacitive touchscreen (resolution 1920×1200, touch sampling rate 120Hz), supports 10-point touch and glove operation modes, a 3D bionic face (supporting dynamic simulation of 20+ micro-expressions), a pressure sensor, and a six-axis gyroscope for natural human-computer interaction; supports hot-swappable expansion of 4K cameras, VR glasses, projectors, and other devices; the interface complies with the IEC 61076-4-101 standard, and has a plug-in lifespan of ≥100,000 cycles.

[0114] Layered hybrid expansion architecture: Adopts composite interfaces such as USB 3.2 Gen 2 and PCI-e 4.0, combined with the Linux kernel's udev system for dynamic hardware management, supports graceful shutdown during hot module removal, and is equipped with overvoltage / overcurrent protection circuits and electrostatic discharge (ESD) protection diodes to ensure hardware stability and security.

[0115] 2. Implementation details of the home-school interaction module AI visual capture technology: The robot's head binocular camera (1920×1080, 30fps) constructs a three-dimensional classroom coordinate system based on the principle of parallax, with a positioning accuracy of ±5cm. It uses an improved YOLOv7 algorithm to detect student behavior at a rate of 0.5 seconds per frame, prioritizing the capture of high-value interactive behaviors such as raising hands (weight 0.95) and standing (weight 0.90). The single-frame processing time is less than 80ms, supporting real-time detection of classes with more than 30 students.

[0116] Dynamic privacy protection technology: The Visual Transformer (ViT) facial processing engine extracts 1024-dimensional facial features from single-frame images. Clustering is performed using an improved DBSCAN algorithm (density threshold ε = 0.6, minimum sample size MinPts = 2). Laplace noise (noise factor Δf = 255, privacy budget ε = 0.1) is injected into non-target student faces, meeting ε-differential privacy (ε-DP) standards. Targeted push notifications for parents: Image hash values ​​are stored on the blockchain. Parents access exclusive content after verification and authorization using zero-knowledge proof (ZKP) technology. Response time is less than 2 seconds. AI editing features are supported, such as adding virtual scenes and voice-action synthesis, with synchronization error controlled within 80ms.

[0117] 3. Implementation of interdisciplinary integration modules Intelligent semantic analysis: A bidirectional long short-term memory (Bi-LSTM) network combined with an attention mechanism is used to analyze psychology course outlines, extracting keywords such as "win-lose perspective" and "positive psychology" and their associated dimensions (probability, effort, and growth). High-dimensional semantic vectors are generated through the BERT pre-training model to construct a multi-level feature space for psychological themes.

[0118] Knowledge graph and dynamic matching: Build an interdisciplinary knowledge graph that includes concept layer, case layer, and strategy layer.

[0119] Conceptual layer: Using word embedding (Word2Vec), semantic associations are established between “win or lose”, “mathematical statistics”, and “sports competition”, with a cosine similarity ≥ 0.92 (e.g., “win or lose probability” and “binomial distribution”). Case layer: stores 500+ interdisciplinary cases (e.g., "athlete win rate statistics" in mathematics, "post-game evaluation of basketball" in sports); Strategy layer: Real-time access to the academic affairs system, using time series analysis and support vector machine (SVM) models to predict the timing of integration (error ±1 class hour). For example, a data chart of "NBA players' season win rate" is pushed two classes before the mathematics probability class, and the psychology teacher explains "continued efforts to improve the win rate" through a live broadcast on the electronic blackboard.

[0120] 3D Visualization Teaching: Integrates augmented reality (AR) tools to dynamically present the relationship between the probability formula (win rate = number of wins / total number of games) and the concept that "winning and losing are probabilistic events," supporting 3D visualization of the mapping between "win rate calculation and effort level."

[0121] 4. Penetration of psychological education in family settings Interest-driven task generation: Use computer vision (CV) to identify students' classroom paintings and natural language processing (NLP) to parse speeches, and construct an interest vector containing 12 dimensions (for example, the label weight of "board games" is ≥ 0.7); call on 30+ family task templates such as "checkers parent-child game" and "movie viewing task" to generate personalized tasks such as "deliberately losing the game + strategy review" and "discussing the winning and losing of characters in "Moneyball"", and combine with generative adversarial networks (GAN) to intercept and match film and television clips.

[0122] Multimodal emotion adaptation: The front camera collects eye movement data (pupil diameter, gaze trajectory), combines it with ResNet-50 to analyze facial expressions (the degree of mouth corner lift, the distance between eyebrows and eyes), and the microphone extracts 20-dimensional acoustic features such as Mel-frequency cepstral coefficients (MFCC). A weighted fusion model (eye movement 0.4 + expression 0.3 + voice 0.3) is used to determine positive emotional states (threshold ≥ 0.65), triggering content push such as "wrong question attribution analysis" and "post-loss strategy adjustment", with a synchronization rate error of less than 80ms.

[0123] 5. Privacy protection module implemented Three-level desensitization mechanism: Collection layer: Differential privacy technology is used to generate irregular IDs (such as "ID-20240601-001") to replace real identities. OCR technology automatically blurs sensitive information such as student names and student numbers on classroom electronic screens and blackboards; Storage layer: Blockchain evidence image hash value, dynamic ID mapping table uses SHA-256 encryption, local deployment or storage in a third-party authority, not uploaded to the public network; Application layer: When Federated Learning analyzes teaching effectiveness, each participant only uploads encrypted model parameters, the original data is retained locally, and the data integrity is verified through zero-knowledge proof (ZKP) technology.

[0124] Permission management and data security: Parents sign the "Data Use Agreement" through biometric authentication, the system generates a unique desensitized ID, and the associated relationship is encrypted and stored in the blockchain node; sensitive data is cleared in accordance with the DoD 5220.22-M standard, the data anonymization rate is 100%, and it has passed the national information security level 3 protection certification.

[0125] 6. Software educational robot function Parent Psychological Support System: Based on the XGBoost classifier (accuracy ≥ 93.2%), it identifies "resistant" parents, triggers dynamic term replacement rules (such as "mental health education class" → "emotional intelligence education class"), and generates popular science content (BLEU value 0.68) by fine-tuning the GPT-3.5 model to explain that "emotional intelligence is an important part of mental health"; it supports 24 / 7 real-time consultation with a response time of less than 15 seconds.

[0126] Closed-loop feedback on educational effectiveness: Deploy an emotional computing module to analyze changes in students' micro-expressions in class (for example, concentration increased by 25% when explaining the "probability of winning or losing"), and continuously optimize interdisciplinary integration strategies and family task push models.

[0127] 7. Technical effect verification Educational effectiveness has improved. Data from an experimental school show that: parent participation has increased from 32.7% to 80%, and the duration of a single interaction has increased from 4.2 minutes to 28.5 minutes; the incidence of students' anti-frustration behavior has increased from 19% to 65%, and the understanding of "correctly looking at winning and losing" has increased from 38% to 78%, and the average score on the anti-frustration ability assessment has increased by 32.6%; the coverage rate of interdisciplinary teaching has increased from 22% to 89%, and the natural integration rate of psychological education and subject scenarios has reached 76%.

[0128] Privacy and security protection: The data anonymization processing rate is 100%, and blockchain evidence storage ensures a 100% data tampering detection rate; no privacy leaks have occurred in the pilot program in the past year, which complies with the requirements of the "Personal Information Protection Law" and the "Education Data Security Management Specifications".

[0129] Through flexible hardware expansion, AI-driven home-school collaboration, deep cross-disciplinary integration, and full-link privacy protection technology, this invention has built a complete mental health education system covering "curriculum analysis-scenario penetration-effect feedback", effectively solving the efficiency and safety pain points of traditional education, and has significant innovation and practical value.

Claims

1. An intelligent mental health education system, characterized in that: include: Physical educational robots, deployed in key areas of the campus, integrate a fully laminated touch screen, a 3D bionic face, and hot-swappable expansion interfaces; The software educational robot runs on mobile terminals and supports natural language processing (NLP) and federated learning technologies. The home-school interaction module captures students' classroom images through AI visual sensors, uses a dynamic Gaussian blur algorithm to protect the privacy of non-target students, and generates exclusive push content for parents. The interdisciplinary fusion module associates psychology courses with subject teaching content based on knowledge graphs and graph neural networks (GNNs) to generate dual-teacher collaborative teaching plans. The privacy protection module uses differential privacy and blockchain technology to achieve data desensitization storage.

2. The system according to claim 1, wherein: The physical educational robot adopts a layered hybrid expansion architecture, including: a composite physical interface that integrates USB, PCI-e, and SATA 4.0 protocols and supports the IEEE 1394b interface; a gold-plated contact and guide key structure design that supports a plug-in and pull-out life of more than 100,000 times; and a udev dynamic device management system based on the Linux kernel to achieve elegant exit of hot-swappable devices and dynamic loading of the microkernel.

3. The system according to claim 1, wherein: The home-school interaction module includes: an improved YOLOv7 target detection algorithm that captures student classroom behavior at a rate of 0.5 seconds per frame; a visual Transformer (ViT) facial processing engine feature encoder that generates differentiated facial blurring strategies for target and non-target students; and a blockchain-based image hash storage and zero-knowledge proof (ZKP) authorization access mechanism.

4. The system according to claim 1, wherein: The interdisciplinary fusion module is implemented through the following steps: using a bidirectional long short-term memory network (Bi-LSTM) to parse psychology course topics and generate semantic vectors; for example, it includes constructing an "interest-win-lose activity" knowledge graph to match mathematical probability classes with sports competition scenarios; and dynamically pushing AR visualization lesson plans to demonstrate the connection between the winning rate calculation formula and psychological education goals.

5. A mental health education method, characterized in that: The following steps are involved: The physical educational robot's hardware can be expanded through hot-swappable interfaces to collect student classroom behavior data in real time, or through software educational robots. The improved YOLOv7 algorithm is used to identify high-weight behaviors and generate classroom images with desensitized ID associations. Based on parent type profiles (resistance / acceptance), educational terminology is dynamically replaced to generate customized home-school interaction content. Through federated learning, we analyze the effectiveness of interdisciplinary teaching and provide feedback to optimize the knowledge graph association strategy.

6. The method according to claim 5, characterized in that The dynamic term replacement steps include: constructing a mental health education knowledge graph that includes the semantic association between "emotional intelligence education" and "mental health"; automatically replacing "mental health courses" with "emotional intelligence education courses" when parents are identified as resistant; and generating explanatory text by fine-tuning the GPT-3.5 model to strengthen the term conversion logic.

7. The method according to claim 6, characterized in that The classroom image processing steps include: using a density clustering algorithm (DBSCAN) to locate the target student's face and retain the clear area; dynamically calculating the blur radius of the non-target student's face according to the formula 𝑅=5√̅d (d is the relative distance); injecting Laplace noise to meet ε-differential privacy (ε=0.1) and peak signal-to-noise ratio (PSNR) ≤ 20dB.

8. The method according to claim 6, characterized in that The steps for generating home tasks include: analyzing student interest tags through latent semantic models (LFM) to match win-loss scenarios such as checkers and movie watching; generating film and television clips that match the course theme based on GAN, and capturing the key frame of "persistence after loss"; and deploying multimodal emotion recognition to trigger task push when positive emotions are detected.

9. The system according to claim 1, wherein: The privacy protection module includes: three-level desensitization processing: Laplace noise injection is used in the collection layer, the SM4 encryption algorithm is used in the storage layer, and federated learning is deployed in the application layer; a dynamic quantum key replacement mechanism, and the key replacement cycle after responding to abnormal login is less than 1 minute.

10. The method according to claim 6, characterized in that The privacy protection steps include: parents sign an electronic agreement through biometric authentication to generate a unique desensitized ID; sensitive data is encrypted using the SM9 algorithm and stored in fragments, and local data is overwritten and cleared using the DoD 5220.22-M standard; and the task completion status is verified through zero-knowledge proof to avoid leakage of original data.

Images