Intelligent system and method for realizing man-machine interaction based on education robot and generating psychological counseling strategy based on record

Abstract

The invention discloses a man-machine interaction method based on an education robot and a system for intelligently generating a psychological counseling strategy based on interaction records, and constructs a progressive interaction path of tool use-emotional interaction-trust deepening by integrating high-frequency interaction scenes such as learning tutoring and home-school interaction. The system collects multi-modal interaction data through an entity and a software education robot, constructs a trust-intimacy two-dimensional evaluation model (T = 0.6 F1 + 0.4 E, I = 0.5 F2 + 0.5 E), and automatically generates a differentiated psychological counseling scheme based on a Becker T-type model and a four-quadrant strategy matching algorithm. The innovation points comprise a triple authorization privacy protection mechanism (block chain evidence storage and dynamic anonymous ID generation), a multi-modal sentiment analysis engine (a bidirectional coded representation model (BERT), a long short-term memory network (LSTM) and an FACS), and a rule engine and XGBoost mixed strategy generation architecture, and the whole process intelligence from interaction and data analysis to precise psychological counseling strategy generation is realized.

Description

Technical Field

[0001] This invention relates to the application of artificial intelligence in the field of psychological education, involving the intersection of psychological health analysis and data encryption. It complies with the requirements of the Personal Information Protection Law and the Education Data Security Management Specification. Background Art

[0002] 1. Existing research shows that psychological counseling is essentially a process of building trust. However, if robots only provide counseling services, lacking the support of daily interaction scenarios, it will be difficult to break users' inherent perception of them as "cold machines." According to social penetration theory, deepening interpersonal relationships requires continuous information exchange and emotional investment, and this principle also applies to human-computer interaction scenarios.

[0003] 2. The current mainstream psychological counseling robots have the following major technical bottlenecks: Lack of trust-building mechanism: Most systems rely on static rules or simple questionnaires for psychological assessments and lack a quantitative model for trust in dynamic interaction processes. Single functional scenarios: Single-function robots focus only on psychological intervention and lack deep integration with educational scenarios. Studies have shown that students' trust in robots is positively correlated with the frequency of their use in daily scenarios such as learning assistance and home-school communication. Weak emotional connection: Existing technologies often ignore the key dimension of intimacy. According to media equivalence theory, users will regard robots as interactive objects with social attributes, and their emotional involvement directly affects their willingness to accept psychological counseling services.

[0004] The educational robot proposed in this paper integrates daily functions such as learning guidance and home-school interaction, creating a progressive interaction path from "tool use" to "emotional interaction" to "deepening trust." Specifically, trust is accumulated in multiple scenarios: Through high-frequency interaction scenarios such as problem-solving and homework assistance, the robot utilizes reinforcement learning algorithms to dynamically optimize guidance strategies, guiding students to leverage positive emotions to overcome learning difficulties, and gradually building user trust through practical problem-solving.

[0005] The intimacy index of the intimacy quantification model is obtained by calculating parameters such as the frequency of parent-school interaction and the proportion of emotional conversations. Compared with the traditional single trust assessment model, user acceptance is increased by 35%.

[0006] Dynamic Strategy Generation: Based on Beck's T-Model cognitive behavioral therapy framework and incorporating user trust and intimacy metrics, we automatically match differentiated psychological counseling strategies. For example, by initiating a deep conversation mode for high-trust and high-intimacy users and employing Rogers' humanistic therapy guidance techniques, the adoption rate of these interventions increased by 28% compared to traditional approaches.

[0007] 3. Comparison table of existing technologies |Comparison Dimensions| Prior Art| Innovation of the Present Invention| |Trust Assessment| Single Dimension | Two-Dimensional Model (Trust, Intimacy), Integrating Learning, Home-School, and Emotional Data | |Policy generation|Static or single|Rule engine + machine learning, supporting dynamic scenario adaptation| |Privacy Protection|Static Anonymity + Centralized Storage|Triple Authorization (Blockchain Evidence) + Dynamic ID Encryption + Federated Learning| | Sentiment Analysis | Single Modality | Multimodal Fusion (Text + Voice + Vision, Output 0-1 Sentiment Value E | |Interactive scenarios|Independent consultation lacks linkage|Learning-home-school-psychological progression, physical robot deployment, software support for mobile terminals| In English, BERT stands for Bidirectional Encoder Representations from Transformers. LSTM stands for Long Short-Term Memory Network. FACS stands for Facial Action Coding System. Summary of the Invention

[0008] Glossary Software Educational Robot: An intelligent program system that runs on computers, mobile devices, and smart wearable devices, integrating natural language processing (NLP), machine learning (ML), and robotic process automation (RPA) technologies. It supports installation and deployment on various mobile clients, including teachers, students, parents, and members of the public.

[0009] Physical Educational Robot: A physical educational robot is an intelligent terminal device deployed in key campus areas (including but not limited to the school gate, dormitory entrances, teaching building entrances, and risk monitoring points). This device integrates sensor technology (covering modules such as visual perception, voice recognition, and environmental parameter monitoring) with artificial intelligence algorithms, and integrates an educational resource database, providing core functions such as teaching assistance and support.

[0010] Multimodal emotion composite value (E): A 0-1 value calculated by weighting text emotion intensity (60%), voice emotion classification (30%), and visual micro-expression features (10%). See Section 4.4.2 for the specific formula.

[0011] Learning problem interaction frequency (F1): The number of times students receive learning guidance through the robot per unit time, normalized to [0,1].

[0012] Frequency of home-school interaction (F2): The number of effective message interactions between students and parents via the robot per unit time, normalized to [0,1].

[0013] Multimodal sentiment composite value (E): A 0-1 value that integrates text (60%), speech (30%), and visual (10%) sentiment analysis. For detailed calculation, see Section 4.4.2.

[0014] To address the technical issues mentioned above, the system of the present invention includes a privacy and information security protection module that complies with legal regulations, an encyclopedic education module for educational robots based on emotion-learning dual-drive, a home-school interaction tool module that can analyze emotions, and an interactive data analysis and psychological counseling strategy intelligent generation module. The steps include: S1: Activate data collection through a triple authorization agreement (students / parents / school); S2: The physical robot collects biological signals and interaction data and calculates the TI value; S3: The policy engine matches the four-quadrant intervention plan and generates a JSON policy package; S4: Execute psychological counseling dialogues, record interaction data and update the user model; S5: The federated learning node aggregates model parameters every month and updates the global policy library; S6: After the data life cycle expires, the SM4 encryption destruction procedure is triggered.

[0015] 1. Have legal and compliant privacy and information security protection functions.

[0016] 1.1. Technical Solution Overview This system can only collect and analyze student data after parents or students who meet the legal age have authorized and voluntarily signed an agreement.

[0017] The entire system of this invention does not store personalized information such as names, ID cards, student ID numbers, phone numbers, addresses, or parent information. Student and parent names are replaced with unique, non-standard ID numbers, ensuring student privacy and information security from the source. Student and parent ID numbers must be linked, a process that integrates information from the school information system. Once the link is established and verified, the data is deleted, disconnecting the information integration with the school information system.

[0018] The present invention provides a student data privacy protection system based on privacy priority. Through a technical combination of authorization control, multi-level anonymization processing, dynamic association verification, multi-dimensional encrypted storage and real-time monitoring, it realizes the privacy protection of the entire process of student data from collection, processing to storage, solves the risk of personalized information leakage in existing education data systems, and meets the requirements of laws and regulations such as the "Personal Information Protection Law" and the "Data Security Law".

[0019] 1.2. Specific technical implementation steps 1.2.1 Triple Authorization Verification and Blockchain Evidence Storage Mechanism This system only starts data collection after the three-party authorization confirmation is completed. The specific process is as follows: (1) Authorization subject and agreement signing: Authorization subject and agreement signing: The student (those over 14 years old can sign independently, and those under 14 years old must be represented by their legal guardian), legal guardian and school respectively sign the "Data Collection and Use Authorization" through the electronic signing module, which includes the "Data Use Scope Commitment Letter", "Data Lifecycle Management Terms" and "Third-Party Sharing Restriction Terms", and clearly specifies the type of data collected (such as psychological behavioral data, homework completion and other common data), scope of use (limited to educational analysis), storage period (no more than 3 years after the authorization is terminated) and revocation method; (2) Generation of authorization validity certificate: After verification, an "Authorization Validity Certificate" containing a unique authorization number and effective timestamp is generated as the basis for data collection access. For students under the age of 14, the authorization agreement must be signed separately by the legal guardian, and students can only perform biometric verification confirmation. 1.2.2 Multi-level anonymization and dynamic ID generation technology The system follows the principle of "minimum necessary" and performs three-level de-identification processing on personalized information: (1) Unregulated unique ID generation: Student ID is generated using a combination of "hash algorithm + random number + national secret algorithm" with the formula: StudentID = SHA-256 (school code + year of admission + random number) ⊕ AES-128 (system key) (the school code is the unified code of the Ministry of Education, and the random number ensures that there are no duplicates in the same school and class); Parent ID is generated through an association algorithm: ParentID = StudentID + HMAC-SHA1 (the last 4 digits of the parent's mobile phone number + system salt value) to ensure a unique correspondence. (2) Hierarchical mapping and obfuscation processing: First layer: sensitive fields are located through regular matching (name matches Chinese characters, ID card number matches the legal format) and converted into irreversible metadata through hash operation; Second layer: dynamic obfuscation factors are generated using a quantum random number generator to perform secondary perturbation on the metadata; Third layer: an improved Bloom filter is applied to fuzzify the eigenvalues, an ID association matrix library is constructed, and the association relationship is encrypted and stored using lattice cryptography technology. (3) Original data clearing: After the replacement is completed, the system uses a file shredder (such as the Eraser tool) to overwrite and erase the storage medium seven times (in compliance with DoD5220.22-M standards) to ensure that sensitive information cannot be recovered.

[0020] 1.2.3 Dynamic Integration and Automated Disconnection System To verify the ID association, the system is integrated with the school information system in a limited way: (1) Secure interface connection: Interact with the school standard API interface (supporting RESTful / SOAP protocols) through an HTTPS encrypted channel (TLS 1.3). The transmitted data only contains the school code, the student / parent name (temporarily used before de-identification), and the last 4 digits of the parent contact information (JSON format). (2) Association verification and data cleanup: The school verifies the correspondence between StudentID and ParentID based on the local relational table and returns the verification result and timestamp. After the verification is completed, the zero-knowledge proof process is triggered to confirm the data integrity, and the multi-copy erasure program is started to clear all temporary data (including interface logs and caches). The network connection is disconnected through physical one-way transmission technology, and a blockchain evidence record (including data fingerprint and cleanup timestamp) is generated.

[0021] 1.2.4 Four-Dimensional Secure Storage and Federated Learning Enhanced Architecture (1) Multi-layer encrypted storage system: Storage layer: Structured data blocks are encrypted using the SM4 symmetric encryption algorithm (128-bit key). The key is encrypted using the SM2 asymmetric algorithm (256-bit public key) and stored in a hardware security module (HSM). Critical data is encrypted using a one-time key generated by a quantum key distribution (QKD) device based on the "one-time, one-pad" principle (devices such as ID Quantique Clavis4). Architecture layer: Construct a four-dimensional storage model: Spatial dimension: Distributed storage in nationally certified government cloud nodes (passed the Level 3 security assessment); Temporal dimension: Implement dynamic lifecycle management and automatically trigger the data incineration process (starting three years after authorization termination); Permission dimension: Role-based access control (RBAC) to achieve fine-grained permission management; Encryption dimension: Adopt SM2 / SM9 combined encryption strategy, and sensitive operations are processed within the Trusted Execution Environment (TEE, such as Intel SGX encrypted enclave).

[0022] Protection layer: Data blocks are stored in a distributed file system (HDFS) in a dispersed manner, protected against data loss through a RAID 6 redundancy mechanism, and a database audit system is deployed to monitor abnormal operations.

[0023] (2) Privacy-preserving computing technology: Using a horizontal federated learning framework (such as TensorFlow Federated), each school trains model parameters locally, and a central server aggregates the parameters to generate a global model; During training, Laplace noise (ε=0.5, δ=1e-5) that satisfies (ε, δ) differential privacy is injected, where δ is the maximum allowed privacy leakage probability, to prevent gradient backpropagation of individual data. It supports heterogeneous federated learning (horizontal / vertical / migration hybrid applications), processes parameter updates through homomorphic encryption technology, and ultimately outputs statistical conclusions (such as class accuracy distribution), prohibiting the disclosure of individual details.

[0024] 1.2.5 Real-time monitoring and emergency response system (1) Abnormal behavior detection: Build an AI detection model based on the long short-term memory neural network (LSTM) to analyze access patterns in real time and identify risky behaviors such as abnormal logins and high-frequency data retrieval. (2) Data lineage tracking: Completely record the entire trajectory of data from generation to flow to use, forming a visual lineage map to support real-time auditing and traceability. (3) Three-level circuit breaker mechanism: Level 1 response: Automatically trigger data sandbox isolation and limit abnormal account operation permissions; Level 2 response: Start the data self-destruction countdown (complete encryption and destruction within 10 minutes); Level 3 response: Physically cut off the power supply of the storage medium to ensure that the data cannot be recovered.

[0025] 1.3. Technical solution features 1.3.1. Full-process privacy-enhancing design: Data invisibility is achieved through triple authorization and dynamic ID generation at the data collection source. Data irreversibility is ensured during the processing phase through layered obfuscation and federated learning. Data unavailability is achieved during the storage phase through four-dimensional encryption and a TEE environment. At the architectural level, system disconnection is achieved through automated disconnection and physical isolation.

[0026] 1.3.2. Multi-Technology Integration and Innovation: Integrates national secret algorithms (SM2 / SM4 / SM9), quantum key distribution (QKD), federated learning, lattice cryptography, blockchain evidence storage, and other technologies to cover the entire data lifecycle security and meet the highest protection level of GB / T 35273 "Information Security Technology Personal Information Security Specification".

[0027] 1.3.3. Dynamic Adaptive Architecture: A heterogeneous federated learning framework supports collaborative data analysis in different educational scenarios. Combining differential privacy and homomorphic encryption technologies, it ensures model accuracy while protecting privacy (≤2% error compared to centralized training).

[0028] 1.4. Technical solution verification and results 1.4.1. Security Verification: A third-party organization was commissioned to conduct 500,000 penetration tests, and no personal information leakage was found; the national encryption algorithm was verified by the GMT0018-2012 test suite, and the quantum encryption equipment meets the QKD technical standards; the federated learning model was trained 100,000 times, with an accuracy rate of over 98% (the error compared to centralized training is < 1.5%).

[0029] 1.4.2. Compliance and Performance: The system has passed the National Information Security Level Protection Level 3 certification, with a data encryption transmission rate of ≥10Gbps, a single-node failure recovery time of <30 seconds, and supports millions of concurrent data processing, meeting the needs of large-scale educational data applications.

[0030] 2. Encyclopedia-style educational function based on emotional learning dual drive In order to enhance students' trust and intimacy in robot psychological counseling, the educational robot system described in the present invention has both learning assistance functions. Students can ask educational robots for help in learning. The educational robots have encyclopedia functions and also have intelligent learning tutoring functions: they support homework photo recognition and problem-solving guidance, formulate personalized learning plans based on cognitive load theory, and integrate common educational robot functions such as time management and learning reminders. The innovation of this system is that students' learning difficulties and key points can be learned through positive emotions under the guidance of the educational robot.

[0031] By using positive emotional learning, students can use software educational robots or physical educational robots to learn. Physical educational robots can be placed in dormitories, school gates, teaching buildings, etc., and can also be placed at home.

[0032] The physical robot integrates extended facilities such as a smart bracelet (hereinafter referred to as the bracelet) and a brain-computer interface helmet (hereinafter referred to as the helmet) through a hot-swappable interface. Students can voluntarily use the bracelet or helmet after signing an authorization agreement.

[0033] All data of the interaction between students and educational robots can only be stored after authorization by legitimate users (parents or students who meet the age requirements). If there is no authorization, the data will not be stored.

[0034] 2.1. Dual-Drive Intelligent Interaction Architecture (1) Learning drive module Based on natural language processing (NLP) technology, it enables subject knowledge query, case analysis, and extended learning, and supports homework photo recognition and problem-solving guidance. Develop personalized learning plans based on cognitive load theory, integrate time management and learning reminder functions, and provide scenario-based services such as wrong question organization and eye exercise push.

[0035] (2) Emotion-driven module After the user voluntarily agrees and authorizes, the facial micro-expression recognition model (FERNet) and the speech emotion analysis model (ProsodyNet) can be used through the smartphone camera to evaluate the emotional state in real time. If the user wears a helmet, in addition to facial expression and speech emotion analysis, bio-indicators such as brain waves and skin electricity can also be analyzed to obtain emotional conditions. If the user voluntarily wears a hot-swappable modular bracelet, the system can detect his or her emotional state through bio-characteristics such as heart rate, skin electricity, and blood pressure. The system dynamically recommends adaptive content: relaxation training courses are pushed when the stress is high, and deep learning tasks (including the difficulty of the course) are pushed when the mood is positive. The software robot uses the smartphone camera to analyze the user's emotions, and the accuracy rate is slightly lower. The cost of the helmet is relatively high, so users are mainly recommended to use the bracelet to obtain emotional information.

[0036] 2.2. Progressive Trust Building Mechanism Adopt a three-level progressive service opening system (users’ voluntary consent and authorization are required): Primary Trust (0-40 points): Provides basic learning assistance functions, including error correction, time management reminders, and subject knowledge query; Intermediate Trust (40-60 points): Open emotional cloud visualization (based on biological signals collected by smart bracelets / brain-computer interfaces, requiring secondary authorization), showing status trends such as stress and concentration; Advanced Trust (>60 points): Launch multimodal psychological assessment services, integrating physiological indicators (heart rate, EEG), conversation content, and behavioral characteristics to generate psychological profiles.

[0037] 2.3. Multimodal emotion-based learning system Emotional Assessment: Based on the cognitive behavioral therapy (CBT) framework, this model integrates physiological signals collected by wearable devices (users must voluntarily and authorize the use of wearable devices, including wristbands and helmets), conversation semantic sentiment analysis, and learning behavior data (including fluctuations in homework completion time) to build a three-dimensional emotion assessment model. Differentiated intervention: In a positive mood, users are reminded that this is the best time to learn; in a fluctuating mood, mindfulness meditation and deep breathing training are triggered; in a sad mood, customized guidance voice messages (with family-friendly dialogue) and relaxation scenes are pushed. Provide a guided conversation process when using it for the first time, and gradually build an emotional connection.

[0038] 2.4. Compliance and Privacy Protection by Design Data minimization: Only desensitized student ID numbers and anonymized feature data are stored. Personal identifiers such as names and ID numbers are not collected. This complies with the data forgetting requirements of Article 47 of the Personal Information Protection Law. Selective authorization: Wearable device data collection, emotion recognition and other functions require users to sign a consent form to activate, ensuring transparency in data use, and these data are not stored; Security Certification: The system complies with the GB / T 35273-2020 Personal Information Security Specification, and data processing has passed the ISO / IEC 27001 Information Security Management System Certification.

[0039] 2.5. Core innovative features Emotional learning dual-engine coupling mechanism: FERNet and ProsodyNet enable real-time perception of emotional states, dynamically adapting learning content and psychological counseling solutions, breaking through the one-way knowledge output model of the traditional education system; Trust-based tiered service system: Based on learning-assisted behaviors, psychological functions are gradually opened up, and a progressive interactive logic of "learning-assisted emotional visualization psychological intervention" is constructed to improve user acceptance; Multimodal composite assessment model: Integrates physiological indicators, conversation content, and learning behavior data for emotion analysis, and combines it with the CBT framework to generate personalized counseling plans, achieving a deep integration of education and psychological services; Privacy-enhanced design: Through anonymous storage and selective authorization mechanisms, a balance is established between functional expansion and data security, complying with domestic and international data protection regulations.

[0040] 3. Home-school interaction tools with emotional analysis In addition to using emotions for learning guidance, in order to increase students' trust in physical educational robots and increase mutual intimacy, physical educational robots also play a role in daily home-school interaction.

[0041] Specifically for closed campus management scenarios, the physical educational robot provides a home-school interactive system with multimodal authentication, emotional intelligence interaction, and secure communication. Parents can use their smartphones to leave two-way messages to the educational robot at the dormitory entrance. After the physical educational robot recognizes the student, it transmits the parent's message to the student and also supports the reverse operation. The core innovative technical solutions are as follows: 3.1. Privacy-enhanced bidirectional identity association technology Anonymous binding mechanism: Establishes an encrypted association between student ID and parent ID, only stores the desensitized ID mapping, and does not retain sensitive information such as name and ID card, in compliance with the data minimization principle of the Personal Information Protection Law; Triple authentication system: Biometric verification: facial recognition (based on ArcFace / FaceNet model, 1:1 comparison accuracy > 99.8%) and voiceprint recognition (Deep Speaker / X-Vector model); physical credential assistance: campus card (RFID / NFC) swiping as a backup authentication solution, building a "biometric + physical credential" two-factor authentication system.

[0042] 3.2. Emotional Intelligence Interaction Module Message intent analysis: The BERT+CRF model is used to extract entities (parent titles, event keywords) and analyze sentiment (emotion classification such as anxiety, urgency, and encouragement). For example, it automatically tags messages like "Good luck on your business trip, Mom!" as encouraging. Contextual feedback design: When the robot reads the parents' message, a warm yellow light flashes synchronously (simulating a warm family scene), and combined with voice tone adjustment to achieve emotional interaction and enhance psychological comfort effect.

[0043] 3.3. Highly reliable real-time communication architecture Hybrid communication technology: Real-time channel: WebSocket persistent connection (implemented by Netty / Spring WebFlux) ensures millisecond-level message synchronization and supports real-time interaction between the robot and the parent end (iOS / Android native app + mini program); Asynchronous processing: Message queues (RabbitMQ / Kafka) carry high-concurrency requests, support 3,000+ terminals online simultaneously, and ensure offline message storage and asynchronous notifications. Secure data transfer: Student voice input is parsed by natural language processing (NLP) to generate encrypted message cards with timestamps, which are transmitted through the HTTPS channel. Sensitive data is encrypted using the SM4 national encryption algorithm.

[0044] 3.4. Parent-side Functional Innovation and Compliance Design Multi-scenario support: provides voice / text / picture message sending, unread reminder (push + SMS dual protection), emergency contact (one-click transfer to the school duty room) and other functions; Strict compliance: Data storage complies with the GB / T 35273-2020 Personal Information Security Specification.

[0045] Core innovative features Anonymous two-way binding technology: associates student and parent identities through desensitized ID mapping, ensuring communication security while protecting user privacy; Three-factor composite authentication system: Integrates facial recognition, voiceprint recognition, and campus card-assisted authentication to create a highly accurate identity verification mechanism (>99.8%). Sentiment analysis and feedback system: Analyzes message sentiment based on the BERT+CRF model and integrates lighting interaction to achieve intelligent linkage between the physical environment and emotional content. Dedicated communication architecture for educational scenarios: A hybrid communication solution combining WebSocket and message queues to meet the high-concurrency, low-latency home-school interaction needs of closed campuses.

[0046] All data of the interaction between students and parents can only be stored after authorization by the legitimate user (parent or student who meets the age requirement). If there is no authorization, the data will not be stored at all and will be completely deleted after the communication ends.

[0047] 4. Interactive data analysis and intelligent generation of psychological counseling strategies Only after authorization by students or parents can data be stored anonymously and analyzed. By processing and analyzing the interactive behaviors between students and educational robots, personalized psychological counseling strategies can be formulated.

[0048] 4.1. System Architecture Design The system constructed by this module includes four core components: user authorization module, data processing module, data analysis module and policy generation module. Each module realizes data flow through encrypted data interface to ensure the security and controllability of data throughout the entire process.

[0049] 4.2. Authorization and licensing mechanism A dynamic authorization chain protocol and a two-factor authorization system establish a dual authorization mechanism for parents and students: students confirm authorization using biometrics (fingerprint / face / voiceprint) and a dynamic password. Parents simultaneously receive an authorization request via SMS containing a timestamp and a random verification code, requiring a second confirmation within 15 minutes. Authorization authenticity is verified using lightweight zero-knowledge proof protocols (zk-SNARKs). Authorization agreements are stored on-chain using smart contract technology to ensure immutability, with a configurable validity period (72 hours by default, with manual renewal supported). Authorization is limited to three levels of permissions: "Behavior Frequency Analysis," "Emotional Feature Extraction," and "Trust Assessment." A role-based access control (RBAC) system isolates permissions for different modules, establishing three levels of data access permissions: anonymized basic data (only anonymous IDs), desensitized interactive content (removing key identity information), and raw data (requires a separate application, with each access recorded and stored on the blockchain).

[0050] 4.3. Data Desensitization and Encrypted Storage Multi-dimensional desensitization processing adopts a hybrid desensitization solution of "rule engine + AI model": Basic desensitization: Perform SHA-256 hashing on direct identification information such as name and ID number to generate a 32-bit fixed-length anonymous ID (such as AN_20250514_0001), and append a salt value to prevent rainbow table attacks.

[0051] Content desensitization: Natural language processing (NLP) entity recognition technology is used to mask sensitive information such as school names and home addresses in interactive texts (for example, "XX Middle School" is replaced with "[School]"), retaining key semantic information.

[0052] Audio desensitization: Using voice feature parameter replacement technology, the fundamental frequency and formant parameters are modified while maintaining the emotional characteristics of the voice to achieve voiceprint de-identification.

[0053] Visual desensitization: Micro-expression data obtained through non-contact physiological measurement technology is feature-desensitized based on the Facial Action Coding System (FACS), retaining only non-identifying features related to emotional intensity.

[0054] Hierarchical encrypted storage architecture builds a hot and cold tiered storage system: Hot data (interaction records for the past 30 days): Stored in a distributed database that supports the national encryption SM4 algorithm, it uses dynamic key management (encryption keys are updated every hour, and the keys are generated by the hardware security module HSM), and feature extraction is completed within the Intel Trusted Execution Environment (Intel SGX).

[0055] Cold data (historical data): After being encrypted with Paillier homomorphic encryption, it is encrypted and stored in the blockchain distributed storage system. Each data block generates a Merkle hash tree to support data integrity verification. The federated transfer learning framework is used for model training. Each educational robot participates as an independent node. The output results are anonymized by k (k≥5) and then stored as evidence.

[0056] 4.4. Multimodal Behavioral Feature Analysis Model 4.4.1. Interaction frequency analysis module designs a two-dimensional frequency statistical model: Frequency of learning problem interactions (F1): This method counts the number of times students use educational robots to query knowledge points, provide homework guidance, and analyze exam questions, etc., on a weekly / monthly basis, with a weight of 60%. The modified Holt-Winters triple exponential smoothing model is used to calculate the interaction density in the time series. Combined with a sliding time window algorithm (the window size is configurable, with a default of 7 days), the average frequency over the last three periods is calculated to eliminate the impact of short-term fluctuations.

[0057] Frequency of home-school interaction (F2): Counts the number of times students use educational robots to send messages to parents, receive home-school notifications, participate in online parent-teacher conferences, and other interactive behaviors, with a weight of 40%. The criteria for defining "effective interaction" are: the number of single dialogue rounds ≥ 3, the duration > 30 seconds, and the inclusion of a complete problem-solving closed loop.

[0058] 4.4.2. Emotion Analysis Module Constructs a Multimodal Emotion Recognition Model: Text Sentiment Analysis: Based on a pre-trained Bidirectional Encoding Representation Model (BERT) model and fine-tuned using an education corpus (containing 200,000 teacher-student conversations), this model identifies positive, negative, and neutral emotions in text and outputs an emotion intensity value between 0 and 1 (e.g., 0.8 indicates a strong positive emotion). A pre-trained BERT-wwm model is used for semantic analysis, outputting multi-dimensional emotion labels such as joy, anxiety, and confusion.

[0059] Speech emotion recognition: Mel-frequency cepstral coefficients (MFCCs) are used to extract speech features, which are then fed into a bidirectional long short-term memory (LSTM) neural network. Combined with an attention mechanism, this technology identifies eight emotion categories, including anxiety, depression, and excitement, with an accuracy rate of ≥92%. Emotional engagement is analyzed through speech fundamental frequency variance, and wavelet packet energy spectrum technology is used to quantify emotional fluctuations.

[0060] Visual emotion recognition: Analyzes psychological states through micro-expression recognition and pupil diameter changes. Based on the Facial Action Coding System (FACS), 36 action unit (AU) features, such as the degree of mouth corner lift and eyebrow wrinkling frequency, are extracted. Emotional states are classified using a Gaussian mixture model (GMM).

[0061] Comprehensive interaction emotion value (E): Calculated by weighted text (60%), voice (30%), and visual (10%), and dynamically adjusted based on the interaction scenario (for example, the emotion weight during homework tutoring is higher than that in the notification reception scenario).

[0062] The two-dimensional evaluation model of trust and intimacy establishes a Cartesian coordinate system, with the horizontal axis representing trust (T) and the vertical axis representing intimacy (I). The calculation formula is as follows: \(T = 0.6 \times F1 + 0.4 \times E \\ I = 0.5 \times F2 + 0.5 \timesE\) If students come to the educational robot for psychological counseling, they will be divided into four quadrants based on the analysis of the interaction records: high trust and high intimacy zone, high trust and low intimacy zone, low trust and high intimacy zone, and low trust and low intimacy zone.

[0063] 4.4.3. The technical solutions for each quadrant are detailed as follows: (1) Quadrant 1: High Trust and High Intimacy Zone (T ≥ 0.7, I ≥ 0.7) If students in this quadrant come for psychological counseling, we will give priority to recommending an in-depth psychological counseling plan, starting a deep dialogue mode, using Rogers' humanistic therapy guidance technology, opening up more personalized data query permissions (such as academic trend analysis), and generating a progressive psychological counseling plan (weekly intensity + 15%).

[0064] Strategic goal: Deepen the effect of psychological intervention and establish a long-term and stable psychological support relationship. Core technical solution: The deep dialogue model is constructed using the core technology of Rogers' humanistic therapy: Empathy Response Engine: Builds an emotional mirroring system based on the BERT-wwm model, automatically identifying emotional keywords (such as "stress" and "loneliness") in user expressions and generating a three-layer response structure: Emotional validation ("I understand you're feeling anxious right now") + contextual repetition ("You mentioned that you recently failed an exam and are feeling uncertain about your future") + open-ended guidance ("Can you tell me specifically when this feeling started?") Non-directive conversation strategy: The consultant (robot) actively speaks ≤ 30% of the time in the conversation round, and encourages users to explore themselves by asking questions such as "What do you think of this solution?" Personalized data service opening Dynamic permission configuration: Dynamically open 3 types of data query interfaces based on the T value (all permissions are unlocked when T≥0.7): | Data Type | Open Content Examples | Technical Implementation | |Academic Trend Analysis|Distribution of Knowledge Points by Wrong Answers in the Last Three Months|Visualization of Prediction Model Based on XGBoost| |Interaction behavior report|Weekly / monthly effective interaction duration|Encrypted interface (HTTPS+AES-256 encryption)| |Psychological state assessment|Recent E-value trend chart|Interactive visualization component based on D3.js| Progressive intervention program design Dynamic adjustment algorithm for intervention intensity: S_t = S_{t-1} \times (1 + 15\%) \quad (S_0=basic intervention intensity), basic intensity definition: 2 active psychological care conversations per week (each 10-15 minutes), Intensity components: Conversation duration (increase by 2-3 minutes each time), question openness (using a Likert-7 scale, with the proportion of open questions increasing by 10% each week), and privacy data sharing level (gradually opening up non-sensitive self-disclosure guidance) Phased intervention goals: Week 1-2: Complete 3 in-depth emotion profiling conversations (based on FACS micro-expression recognition to verify emotional consistency), Week 3-4: Build a unique psychological profile (including 20+ personalized intervention indicators).

[0065] (2) The second quadrant: high trust and low intimacy zone (T≥0.7, I<0.7). If students in this quadrant come for psychological counseling, intervention strategies should focus on establishing emotional connections and increasing emotional support interaction content.

[0066] Strategic goal: Strengthen emotional ties and balance the interaction between instrumental rationality and emotional support. Core technical solutions: Parent-robot collaborative communication mechanism Three-level information synchronization system: |Synchronization level|Trigger conditions|Parent-side output content|Technical implementation| |Basic Level|Weekly I value < 0.5|Weekly Interaction Frequency Report|Encrypted PDF file of blockchain evidence| |Advanced|I value has not increased for two consecutive weeks|Emotional fluctuation heat map|WebRTC real-time data sharing channel| |Emergency Level|E value < 0.3 for 3 consecutive days|SMS gateway + AI|Dual voice notification| Parent-side interactive interface: Integrated emotional support knowledge base, providing 50+ scenario-based communication skills (such as "A guide to responding when your child says 'I don't want to go to school'") Emotional support interactive content system Multimodal emotional companionship module: Voice interaction: Customized encouraging audio is pushed every morning (based on Tacotron2 speech synthesis technology, with emotional parameters dynamically adjusted based on the previous day's E-value) Text interaction: Built-in "emotional tree hole" function, supports 24-hour anonymous text confession, and automatically generates emotional feedback reports (including emotional keyword cloud map) Relationship strengthening algorithm, through the emotional engagement formula: A=0.3E+0.2F emotional conversation+0.5ΔI, where F emotional conversation is the weekly frequency of emotional conversations.

[0067] Dynamically adjust interaction strategies to ensure weekly I value increases ≥5% (3) The third quadrant: low trust and high intimacy zone (T < 0.7, I ≥ 0.7). If students in this quadrant receive psychological counseling, they should strengthen the professional ability demonstration and guidance program, adopt the cognitive behavioral therapy (CBT) structured question and answer template, and enhance trust by solving practical problems.

[0068] Strategic goal: To transform emotional closeness into cognitive trust through concrete demonstration of professional capabilities. Core technical solutions: Implementation of CBT structured question and answer template Four-stage psychological counseling process engine: Problem identification (5 minutes) → Cognitive restructuring (10 minutes) → Behavior planning (8 minutes) → Effect rehearsal (5 minutes) Problem identification: Using a problem classification model based on dependency syntax analysis (91% accuracy), quickly identify cognitive biases (such as "catastrophizing" and "overgeneralization"). Cognitive Restructuring: Use the 200+ pre-set cognitive logic check rules (e.g., "If this thing has a 10% chance of happening, what is the other 90% chance?") Behavior plan: Generate a list of SMART goals (e.g., "Complete 30 minutes of physical exercise every day this week") and use blockchain smart contracts to record goal completion.

[0069] Professional ability visualization module Knowledge graph empowerment: Build an education knowledge graph containing over 1,200 knowledge points, supporting real-time generation of problem-solving path diagrams (e.g., a three-level derivation chain of "concept definition → typical examples → error-prone analysis" for mathematical function problems). Success Case Library Call: Automatically matches historical solutions to similar problems based on the user's question type (displayed after desensitization, such as "2024 senior high school student Xiao Wang improved his math score by 20 points using XX method") Trust enhancement mechanism: After each professional problem is solved, the trust correction algorithm \(T' = T + 0.1 \times (1 T) \times S\) is triggered (S is the satisfaction with the problem solution, ranging from 0 to 1) (4) The fourth quadrant: low trust and low intimacy zone (T<0.7, I<0.7). If students in this quadrant come for psychological counseling, the ice-breaking strategy based on trust building is used, the sandbox dialogue mechanism is activated, and all sensitive questions are responded to after differential privacy processing, gradually guiding users to establish interactive habits.

[0070] Strategic goal: break down interaction barriers and establish a foundation of trust. Core technology solutions: Sandbox dialogue mechanism construction Three-layer security isolation architecture: |Isolation level|Data processing method|Interaction function restrictions| |First layer|Add ε-differential privacy noise (ε=0.5) to the data|Prohibit topics such as family / body| |Second layer|Conversation content is sanitized by NLP semantics|Limit the number of conversation rounds to ≤5| |Third layer|Real-time encrypted storage of interaction logs|Only open basic learning problem consultation interface| Progressive unblocking strategy: After 10 cumulative effective interactions, the first layer of isolation will be automatically lifted; when the I value is ≥0.5, the second layer of isolation will be lifted Interactive guidance strategy design Icebreaker Script Library: Contains 200+ initial conversation templates across three categories, using a reinforcement learning-based script recommendation model (average icebreaker success rate 68%): Learning about association classes: "I heard you're studying quadratic functions lately. Do you need to learn about image transformation techniques in this area?" Interest Exploration: "What's your favorite subject? I can help you organize some interesting knowledge." Mild Emotion: "It's a nice day today. What do you usually do on a day like this?" Interactive incentive mechanism: establish an interactive growth system, and the cumulative effective interaction time reaches 30 minutes to unlock personalized functions such as "exclusive nickname setting" and "conversation background change" Comparison table of technical characteristics of four quadrant strategies |Dimension|High Trust High Intimacy Zone|High Trust Low Intimacy Zone|Low Trust High Intimacy Zone|Low Trust Low Intimacy Zone| |Technology|Rogers Therapy + Data|Parent Collaboration + Emotional Companionship|CBT Template + Knowledge Graph|Sandbox Mechanism + Icebreakers| |Data|Three-level data query|Parent-side limited data synchronization|Only open problem solving|Interaction under differential privacy protection| |Intervention|Progressive reinforcement (weekly +15%)|Emotional interaction ≥40%|Problem-solving orientation ≥70%|Lightweight interaction ≤5 rounds| |Trust|Emotional resonance → Self-exploration|Family support → Emotional transfer|Ability recognition → Cognitive trust|Interaction → Habit formation| |Implementation|Deep semantics + visualization |Multimodal emotions + parent interface |Logical reasoning + case matching |Privacy protection + speech recommendation| 4.5. Psychological Counseling Strategy Generation Engine This module uses a hybrid architecture of "rules engine + machine learning" to achieve intelligent transformation from assessment data to intervention plans. It can be divided into three subsystems: strategy generation logic layer, decision engine implementation layer, and human-machine collaborative optimization layer. The specific technical details are as follows: 4.5.1. Rule Engine Subsystem (1) Multi-dimensional rule modeling system A three-level rule classification system is built, covering basic trigger rules, scenario adaptation rules, and risk control rules. It supports "addition, deletion, modification, and query" of rules and priority configuration through a visual interface (the rule engine architecture is shown in Figure 1): Basic trigger rules (60%): Based on the quadrant division of the two-dimensional coordinates of trust and intimacy, 12 core trigger rules are preset: IF (T≥0.7 AND I≥0.7) THEN Activate "Deep Psychological Intervention Mode" IF (T≥0.7 AND I<0.7 AND E-week decrease>20%) THEN trigger the "Emotional Connection Strengthening" task package (including 3 exclusive caring conversations + parent-side emotional warning) IF (T<0.5 AND F1 monthly growth >30%) THEN Activate the "Trust Breaking - Knowledge Empowerment" strategy (prioritize learning-related questions and embed 5% psychological counseling techniques) Scenario adaptation rules (30%): Generate differentiated strategies based on dynamic parameters of educational scenarios (such as exam week and the start of the school year). WHEN The "Midterm / Final Exam Week" tag is detected AND (T≥0.6) THEN The strategy library automatically loads the "Exam Anxiety Relief" submodule (including pre-exam relaxation audio and guidance on attributing wrong answers) WHEN Identify the "New Students Enrollment" scenario AND I < 0.5 THEN Insert the "Campus Environment Adaptation" topic dialogue (trigger once every 3 days, for 2 weeks) Risk control rules (10%): Establish a three-level early warning response mechanism: |Risk Level|Trigger Condition|Response Strategy|Notification Mechanism| | Level 1 | E < 0.3 for 3 consecutive times for < 10 seconds | End the conversation and push a contact card | Notify parents within 5 minutes | |Level 2|T<0.4 Sensitive topic rate>60%|Enable sandbox to block data|Generate risk report on the same day| |Level 3|Strategy execution fails more than 5 times|Automatically switches to manual review mode|Triggers a strategy optimization ticket| (2) Rule engine technology implementation Rule expression language: Adopts the DRL (Drools Rule Language) syntax of the Drools decision engine, supports complex conditional expressions (such as time windows and numerical ranges), and the compilation time of a single rule is ≤10ms.

[0071] Efficient matching algorithm: Based on the Rete algorithm, rule matching is optimized, and an index network is built through the type nodes and attribute nodes of fact objects, which controls the matching delay of millions of rules to within 50ms.

[0072] Visual configuration tool: provides a web-based rule editing interface (as shown in Figure 2), supports drag-and-drop rule components (condition nodes, action nodes, branch nodes), has built-in 20+ educational psychology rule templates, and supports one-click import / export.

[0073] 4.5.2. ML-Based Recommendation Subsystem (1) Deep processing of feature engineering Construct an input vector space containing 53-dimensional features. The specific processing logic is as follows: Basic features (12 dimensions): Static features: grade (one-hot encoding for grades 1-12), gender (binarized as 0 / 1), and whether or not the student lives on campus (Boolean value).

[0074] Dynamic features: historical consultation times (past 30 days), academic performance segmentation (A / B / C / D grades, using ordered coding).

[0075] Interaction features (25 dimensions): Frequency category: F1 (accumulated value per week / month / semester), F2 (same as above), maximum daily interaction duration (minutes).

[0076] Quality category: effective interaction rate (number of effective interactions / total number of interactions), problem-solving satisfaction (0-10 points, collected through a questionnaire after the conversation).

[0077] Time series: entropy of interaction periods (reflecting the regularity of interaction time), distribution of interaction days within a week (7-dimensional vector).

[0078] Environmental characteristics (16 dimensions): External environment: daily temperature / humidity (obtained from API, normalized), air quality index (AQI, 5-level one-hot encoding).

[0079] School events: whether it is an exam day (Boolean value), whether an event is being held (one-hot encoding of the event type, such as sports meeting / art performance / lecture).

[0080] Seasonal factors: Season labels (spring / summer / autumn / winter, using sine-cosine encoding to handle periodicity).

[0081] (2) XGBoost model optimization strategy Data enhancement: To address the problem of insufficient samples in low-confidence groups, we use SMOTE oversampling technology (oversampling ratio 1:1.5) and Tomek Links to remove noise samples.

[0082] Parameter tuning: The Bayesian optimization algorithm is used to search for the optimal parameter combination and finally determine: max_depth=6, learning_rate=0.05, n_estimators=300, subsample=0.8,colsample_bytree=0.7 Incremental learning: Automatically load the latest interaction data (nearly 30 days) every month for model fine-tuning, using warm start technology to reduce training time by 40% compared to full training.

[0083] (3) Strategy generation and dynamic optimization Multi-strategy fusion mechanism: The model outputs the top three high-probability strategies (probability threshold ≥ 0.15) and generates a combination plan through linear weighting (weights: first strategy 0.6, second strategy 0.3, third strategy 0.1); Introduce domain expert knowledge for policy filtering, such as prohibiting the recommendation of "privacy data disclosure" policies for users with T<0.5.

[0084] Q-learning reinforcement learning closed loop: Q(s,a) = (1-γ)Q(s,a) + γ(max_a' Q(s',a')); State space s: User TI coordinates + current interaction scene label; Action space a: 20 preset strategy encodings (0-19) Reward function r: Behavior improvement indicator within 3 days after the strategy is executed (e.g., effective interaction rate increase × 0.6 + E-value increase × 0.4). Every 50 strategy executions trigger a Q-value table update (ε-greedy strategy, initial exploration rate ε = 0.3, decaying by 0.1 every 100 times) 4.5.3. Full process control of strategy generation (1) Three-stage processing pipeline Data input (TI coordinates + feature vectors) → Rule engine pre-screening (excluding invalid strategies) → XGBoost model ranking (output probability matrix) → Strategy combination optimization (generate 3 sets of alternative plans) → Manual review interface (psychology teachers can adjust the intensity of the speech by ±20%) → Plan packaging (JSON format output).

[0085] Single-user policy generation latency ≤ 400ms (50ms for rule matching + 200ms for model inference + 150ms for combination optimization) (2) Structured output of strategy content The generated psychological counseling plan contains a standardized JSON structure with the following fields: {"strategy_id": "STR_20250514_001", "user_id": "AN_20250514_0001", "strategy_type": "Strengthening emotional connection", "execution_plan": [ {"step": 1, "trigger_condition": "User initiates a conversation", "response_template": "I noticed that you've been having some challenges with your math homework lately. This might be stressful for you, right? Let's analyze the specific issues together." "priority":3,"expected_effect":{"E_increase": "0.1-0.2","F2_weekly":"+2-3 times"}} ], "evaluation_metrics": [ "Effective interaction rate within 3 days", "E value fluctuation within 1 week"]} (3) Human-machine collaborative optimization mechanism Psychological teacher intervention interface: provides a strategy details editing page, supporting: Fine-tuning the words: Modify keywords based on the preset template (e.g., replace "pressure" with "trouble"), and the system automatically performs semantic consistency verification Strategy weight adjustment: reorder the top 3 strategies recommended by the model (requires input of adjustment reason, which is recorded in the blockchain audit log) Strategy performance review: Generate strategy execution reports every week, including: Hit rate analysis: rule trigger accuracy (actual trigger count / should-have trigger count) Decay curve: The effect retention rate after 1 / 3 / 7 days of strategy execution Anomaly detection: Strategies that fail to achieve the expected effect for 5 consecutive times will automatically enter the manual review queue. 4.5.4. Key technological innovations Rule-based model collaborative decision-making architecture: This architecture achieves rapid response (50ms level) through a rule engine and combines machine learning to process complex patterns (such as the nonlinear correlation between environmental factors and emotions), improving strategy accuracy by 28% compared to a single decision-making model.

[0086] Deep adaptation to educational scenarios: By incorporating dedicated scenario variables such as "exam week" and "new student enrollment" into feature engineering, strategy generation is more closely aligned with the school's actual teaching rhythm. Actual measurements have shown a 45% increase in strategy adoption during exam periods.

[0087] Dynamic security policy injection: Privacy protection rules are built into the rule engine (such as automatically adding differential privacy noise for users with T < 0.6) to ensure that the policy generation process complies with the "right to be forgotten" requirements of Article 17 of the GDPR. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] 1. Figure 1 : Overall interaction and strategy generation main process High-frequency interactive scenarios: covering core scenarios such as learning guidance and home-school interaction, and building a progressive interactive path.

[0089] Three-stage interaction path: Tool usage: Educational robots provide basic functional services (such as homework tutoring); Emotional interaction: Establish emotional connections through multimodal interactions such as voice and expressions; Deepening trust: Improving user trust and intimacy based on continuous interaction.

[0090] Data-driven process: Collect multimodal data (text, voice, emoticons, etc.), process it for privacy protection, and input it into the evaluation model to calculate the trust value (T) and intimacy (I), and finally generate a personalized consulting strategy.

[0091] 2. Figure 2 : Multimodal sentiment analysis and strategy generation architecture Multimodal analysis engine: Text: Extract semantic sentiment features through BERT; Speech: Using LSTM to identify emotional fluctuations in intonation; Expression: FACS-based analysis of the emotional state corresponding to micro-expressions.

[0092] Hybrid Strategy Architecture: Rule engine: quickly match preset policy templates; XGBoost model: Optimizes strategy weights based on historical data to improve personalization accuracy.

[0093] 3. Figure 3 :System overall architecture diagram Physical educational robots: Deployed in teaching and living areas, they integrate vision (HD camera), voice (omnidirectional microphone array), and touch interaction modules, and are equipped with a real-time operating system (RTOS) and edge computing engine.

[0094] Software educational robot: developed based on a cross-platform framework, supporting lightweight model deployment.

[0095] Data Center: Using distributed database (hot data) and blockchain storage (cold data), computing clusters support model training and real-time reasoning.

[0096] Data flow: The physical robot transmits data to the edge node via Wi-Fi 6, and the software robot connects to the cloud via an encrypted channel.

[0097] 4. Figure 4 : Flowchart of triple authorization and dynamic ID generation Authorization chain: Students / parents complete electronic signing through biometrics (fingerprint / face recognition) and two-factor verification, in compliance with the Electronic Signature Law.

[0098] Dynamic ID generation: Generate anonymous IDs through hash encryption and blockchain technology to ensure that identity information cannot be traced.

[0099] Evidence storage mechanism: Authorization records are stored on the chain, including key information such as timestamps and hash values.

[0100] 5. Figure 5 : Rule engine and machine learning hybrid strategy generation architecture diagram Rule Engine: Based on conditional triggering strategies (such as "automatically turning on deep intervention mode for high-trust users"), it supports scenario adaptation (such as automatically pushing relaxation tasks during exam weeks).

[0101] Machine learning input: Integrates basic user information, interactive behavior, and environmental characteristics, and optimizes strategy recommendations through the XGBoost model.

[0102] 6. Figure 6 : Schematic diagram of the home-school interaction emotion analysis interface Emotion visualization: High-frequency emotional words are displayed through emotion cloud maps, and interactive heat maps show the distribution of communication time and emotion intensity.

[0103] Intelligent assistance: Recommends communication techniques based on students' trust-intimacy coordinates, and supports one-click contact with the psychological duty room in emergencies. DETAILED DESCRIPTION

[0104] 1. Hardware deployment and interaction scenarios 1.1. Campus Scenario Full Coverage Solution Teaching area: One physical robot (model: AI-EDU B1) is deployed at the elevator entrance of each teaching building. It is equipped with a wide-angle camera (horizontal viewing angle of 120°) and a homework scanning table (A4 size, 300dpi resolution). It supports taking photos of homework and answering questions after class (response time ≤ 15 seconds). It has a hot-swappable interface and can be expanded to smart bracelets, smart brain-computer interface helmets, etc.

[0105] Living area: A robot (model: AI-EDU D1) is deployed at the dormitory entrance.

[0106] Mobile scenario: Develop a hot-swappable smart bracelet (model: EDU-Band V1) with a built-in three-axis accelerometer (ADXL345), a heart rate sensor (MAX30102), a biosignal acquisition module (AD8232 electrocardiogram module, gain 1000, bandwidth 0.5-40Hz), and an environmental perception module (BME680 temperature, humidity, air pressure, and VOC four-in-one sensor). It supports active authorization to collect motion trajectory, sleep data, and physiological indicators (sampling rate: 100Hz).

[0107] 1.2. Hardware Configuration Details Physical robot terminal: The main control chip is NVIDIA Jetson AGX Xavier (32 TOPS AI computing power), which integrates multimodal interaction capabilities.

[0108] 2. Interaction Record Collection and Trust Building 2.1. Learning assistance scenarios Students use the robot to scan wrong questions on math test papers. The OCR module recognizes the question text (with an accuracy rate of 98.7%), calls the knowledge graph to generate a solution path (such as "trigonometric function → angle formula → example question analysis"), and records data such as interaction duration and question category (F1 count +1).

[0109] The robot dynamically adjusts the difficulty of explanations based on cognitive load theory: if a student repeatedly asks about the same knowledge point 3 or more times, the robot automatically reduces the depth of the explanation and inserts analogies (such as using "reservoir water storage" to illustrate "points accumulation").

[0110] 2.2. Home-school interaction scenarios When parents send voice messages through the app, the BERT+CRF model analyzes the emotional category (such as "encouragement" with a confidence level of 0.92). The robot adjusts the voice tone when reading (reducing the speaking speed by 10% and increasing the pitch by 5Hz) and triggers the light module to flash warm yellow (color temperature 2700K).

[0111] Special handling for low-trust scenarios: After sentiment analysis of parents' messages, if anxiety is detected (e.g., "Remember to take cold medicine," with a confidence level of 0.78), the robot provides contextual feedback by flashing a warm yellow LED (2Hz frequency, 40% duty cycle) and adjusting the voice tone (increasing pitch by 15% and reducing speech speed by 10%). An interaction log is also generated to record the intensity of the emotion and the response delay.

[0112] 3. Psychological Counseling Strategy Generation and Implementation 3.1. High Trust, High Intimacy Zone Case (T ≥ 0.7, I ≥ 0.7) User coordinates: T=0.82, I=0.75 (Li, a senior high school student, used the robot to answer questions every day for two consecutive months, and his parents left messages four times a week. Because of his academic performance, the student went to the physical educational robot for psychological counseling.).

[0113] 3.1.1. Data collection and preprocessing: The smart bracelet collects heart rate (HR: 60-120bpm) and skin conductance (GSR: 0-20μS) data and calculates the emotional baseline value every 5 seconds: \[ E_{bio} = 0.4 \times (1 \frac{\beta_{avg}}{\alpha_{avg} + \theta_{avg}}) + 0.6 \times \frac{GSR GSR_{min}}{GSR_{max} GSR_{min}} \] (The α / β / θ wave power spectra were calculated by fast Fourier transform (FFT), frequency range: α[8-12Hz], β[13-30Hz], θ[4-7Hz]).

[0114] Learning interaction data: homework tutoring frequency F1 = 24 times (weight 60%), home-school interaction frequency F2 = 15 times (weight 40%), conversation text emotional intensity E_text = 0.85 (confidence > 90%).

[0115] 3.1.2. Trust Assessment: \[ T = 0.6 \times \frac{F1}{F1_{max}} + 0.4 \times E_{bio} = 0.768, \quad I = 0.5 \times \frac{F2}{F2_{max}} + 0.5 \times E_{text} = 0.725 \] 3.1.3. Strategy Execution: In-depth dialogue: Using Rogers' humanistic therapy, the dialogue structure is "emotional confirmation (30%) + content repetition (20%) + open-ended questions (50%)", calling the pre-trained template: json{"response_template": [ "Emotional validation": "It sounds like your recent test results have made you anxious." "Content recap": "You mentioned that the error rate in math function questions increased by 20% compared to last month." "Open Guidance": "We can analyze the commonalities of these incorrect answers together. What do you think?"]} Open data: The "Mathematics Wrong Question Knowledge Point Distribution Radar Chart" (based on the XGBoost model, with a prediction accuracy of ≥92%) is available to students, showing that the error rate of the "Solid Geometry" module is 45%, and recommends 3 typical questions + video analysis every day.

[0116] Effect evaluation: After 2 weeks, the emotional value E increased from 0.58 to 0.72, and the number of active consultations increased by 50%; after 4 consecutive weeks of intervention, the effective interaction rate increased from 65% to 82%, and the weekly average E increased to 0.81.

[0117] 3.2. Low Trust and Low Intimacy Zone Case (T<0.5, I<0.5) User coordinates: T=0.41, I=0.38 (Student Wang, a first-year junior high school student, only used the robot to query the answers to his homework twice, and his parents did not leave any messages).

[0118] 3.2.1. Strategy Execution: Sandbox conversation: All questions are added with differential privacy noise of ε=0.5 to hide sensitive words (for example, replacing "not wanting to go to school" with "feelings of adapting to the new environment").

[0119] Ice-breaking words: Recommend "interest exploration" words based on reinforcement learning (such as generating relevant interesting knowledge based on the search history of "dinosaur extinction").

[0120] Permission unblocking: After accumulating 10 effective interactions, the isolation will be lifted and the "Emotional Tree Hole" text confession function will be opened.

[0121] 3.2.2. Trust Correction: Trust update formula after a single effective interaction: \[ T' = T + 0.1 \times (1-T) \times S \] (The satisfaction score S is obtained through a questionnaire after the conversation. For example, when S=0.8, T increases from 0.65 to 0.678).

[0122] 4. Implementation of privacy protection technology 4.1. Data anonymization process Initial data: Student Zhang San, student number 20250301, parent’s mobile phone number 1385678.

[0123] First-level processing: Name → Hash value (SHA-256(Zhang San) → a1b2c3...), student number → StudentID = SHA-256(01+2025+random number) ⊕ AES-128(key) → "ST_2503_01".

[0124] Secondary processing: The last 4 digits of the parent’s mobile phone number “5678” → HMAC-SHA1 (5678 + salt value) → “d4e5f6”, generating ParentID = “ST_2503_01_d4e5”.

[0125] Storage rules: Only the anonymous ID association table (ST_2503_01 ↔ ST_2503_01_d4e5) is retained, and the original information is overwritten and erased seven times using the Eraser tool.

[0126] 4.2. Federated Learning Training Process Participants: Three schools (School A, School B, and School C), each providing 1,000 anonymized interaction records.

[0127] 4.2.1. Training steps: (1) Each school trained the XGBoost model locally (parameters: max_depth=6, learning_rate=0.05), and input anonymized features (such as F1, F2, E); (2) Encrypt and upload model gradients (homomorphic encryption technology) to the central server; (3) The server aggregates the gradients to generate a global model (error ≤ 1.2%) and returns the updated model to each school.

[0128] 4.2.2. Privacy protection: No other party’s data can be obtained during the training process, and the gradient aggregation results only contain statistical parameters (such as weight mean and variance).

[0129] 4.3. Emergency Circuit Breaker and Data Destruction 4.3.1. Abnormal response: Level 1 response (within 10 seconds): triggering data sandbox isolation and blocking the attacking IP address; Level 2 response (within 5 minutes): Initiate data self-destruction program and quickly format the attacked partition according to DoD 5220.22-M standard; Level 3 response (within 15 minutes): Physically disconnect the storage array power supply and generate blockchain evidence (including attack time and signature of the person who processed the attack).

[0130] 4.3.2. Data destruction audit: After the storage period expires, the data will be overwritten and erased seven times to generate a blockchain evidence record (including hash value and destruction time).

[0131] 4.4. Special Scenarios for Data Desensitization Voice message processing: After the parent's voice is recognized through voiceprint recognition (Deep Speaker model, 1:1 comparison score 0.92), the fundamental frequency (±15%) and formant (standard deviation σ = 50Hz) are adjusted to generate an anonymous message card (e.g., "Please attend the [school] meeting on time").

[0132] 5. System Security and Compliance Penetration Testing A simulated attack was commissioned by a third-party company, and the test conclusions were: the success rate of inferring the real name from the ID association table was 0%; the failure rate of cracking AES-256-GCM encrypted communications was 100%; the system has the ability to pass the third-level security certification, which complies with Article 27 of the Personal Information Protection Law and the EU GDPR requirements.

[0133] Performance indicators Real-time Response: End-to-end voice interaction latency ≤ 300ms (including speech recognition (ASR) + natural language processing (NLP) + text-to-speech synthesis (TTS)), policy generation latency ≤ 400ms. Encryption Efficiency: SM4 algorithm encryption rate ≥ 15Gbps (AES-NI instruction set acceleration). Biometric Accuracy: Face recognition False Rejection Rate (FRR) = 0.1%, False Acceptance Rate (FAR) = 0.01%.

[0134] 6. Experimental data support 6.1. Trust Building Efficiency The trust level of the experimental group (this system) increased from 38% to 68% in 4 weeks, while that of the control group (traditional robots) only increased from 35% to 42%, resulting in an efficiency improvement of 428%.

[0135] 6.2. Strategy Adoption Rate Adoption rate in high-trust zones: 89% for the present invention vs. 56% for traditional methods (a 33% increase); adoption rate in low-trust zones: 62% for the present invention vs. 31% for traditional methods (a 31% increase). 6.3. Privacy Protection Effect Differential privacy: After injecting ε=0.5 noise, the data re-identification rate dropped from 98% to 3.2%; federated learning accuracy: the error compared with centralized training is ≤1.4%.

[0136] Privacy and security: Dynamic ID generation and federated learning technology reduce the risk of data leakage and meet the requirements of the Data Security Law.

Claims

1. An intelligent psychological consultation system based on an educational robot, characterized in that: include: (1) Multi-scenario interaction module: Physical educational robots are deployed in public areas of the campus to provide high-frequency interactive functions such as learning guidance (homework photo recognition), two-way communication between home and school, etc. Software robots run on mobile terminals and support voice Q&A and emotional counseling; (2) Interaction data acquisition module: Collect multimodal interaction data, of which the normalized learning problem interaction frequency (F1∈[0,1]) accounts for 60% of the trust calculation weight, and the normalized home-school interaction frequency (F2∈[0,1]) accounts for 50% of the intimacy calculation weight (3) Two-dimensional evaluation and strategy generation module: Construct a trust / intimacy coordinate system (TI coordinate), based on the formula: T=0.6×F1+0.4×E, I=0.5×F2+0.5×E, generate user coordinates, and automatically match psychological counseling strategies in combination with the four-quadrant strategy model (high trust and high intimacy zone / low trust and low intimacy zone, etc.).

2. The system according to claim 1, wherein: The privacy protection module includes: (1) a triple authorization mechanism: students, guardians, and school authorities sign a "Data Collection Authorization Letter" to generate a timestamp blockchain-based authorization certificate (including a unique authorization number); (2) dynamic ID generation technology: using the SHA-256 hash algorithm to generate an irregular student ID (StudentID = SHA-256 (school code + year of admission + random number) ⊕ AES-128 (system key)), and the parent ID is encrypted and associated with the student ID using the HMAC-SHA1 algorithm.

3. The system according to claim 1, wherein: The multimodal sentiment analysis module includes: (1) text sentiment analysis: based on the bidirectional encoding pre-trained model-whole word mask (BERT-wwm) model, it identifies emotional labels such as anxiety / encouragement in interactive texts and outputs an emotion intensity value (E) of 0-1; (2) speech emotion recognition: through MFCC feature extraction combined with LSTM network, it recognizes 8 emotion categories (accuracy ≥ 92%); (3) visual sentiment analysis: based on the facial action coding system (FACS) to extract 36 action unit (AU) features, combined with the Gaussian mixture model (GMM) to classify emotional states.

4. A psychological counseling strategy generation method, characterized in that: The method includes the following steps: (1) Interaction data preprocessing: k-anonymity (k=5) + differential privacy (ε=0.5) desensitization of learning problem interaction, home-school messages and other data to generate anonymous ID-associated interaction feature vectors; (2) Two-dimensional coordinate calculation: Calculate user trust T and intimacy I based on F1 (learning interaction frequency), F2 (home-school interaction frequency), and E (multimodal emotion comprehensive value); Four-quadrant strategy matching: High trust and high intimacy zone (T≥0.7, I≥0.7): Enable Rogers' humanistic therapy deep dialogue mode and open academic trend analysis permissions; Low trust and low intimacy zone (T<0.7, I<0.7): A sandbox dialogue mechanism is used to handle sensitive issues through differential privacy.

5. The method according to claim 4, characterized in that The strategy generation includes the use of a "rule engine + machine learning" hybrid architecture: (1) The rule engine is implemented based on Drools, with 12 core trigger rules preset (including "IF T≥0.7 AND I≥0.7 THEN deep psychological intervention"), and the rule matching delay is ≤50ms; (2) The XGBoost model inputs a 53-dimensional feature vector (including grade, interaction frequency, exam week label, etc.), and determines parameters through Bayesian optimization (max_depth=6, learning_rate=0.05), and the strategy generation delay is ≤400ms.

6. The system according to claim 3, wherein: The workflow of the strategy generation engine includes: input layer: receiving TI coordinates and feature vectors; processing layer: the rule engine filters invalid strategies within 50ms, and the XGBoost model outputs the probability matrix within 200ms; output layer: generating a standardized JSON strategy package, including an execution plan, expected effects, and evaluation indicators; control layer: adjusting the trigger probability weight of the preset script (such as increasing the probability of the appearance of "encouragement-type" script by 20%).

7. A privacy-protected data storage system, characterized by: (1) Hot data (interaction records from the past 30 days) is stored in a distributed database that supports SM4 encryption, with the key updated every hour (HSM generation). (2) Cold data (historical data) is stored on the blockchain after being homomorphically encrypted using Paillier encryption. The model is trained using federated transfer learning, and the output is k-anonymized (k ≥ 5). (3) Federated learning unit: Gradient parameters are transmitted using Paillier homomorphic encryption, and Laplace noise (ε = 0.5, δ = 1e-5) is injected.

8. The method according to claim 4, characterized in that The execution effect of the strategy is optimized through Q-learning reinforcement learning: (1) the state space s contains TI coordinates and scene labels (such as exam week / new student enrollment); (2) the reward function r = 0.6× effective interaction rate increase + 0.4× E value increase, and the Q value table is updated every 50 executions (ε-greedy strategy, initial exploration rate 0.3).

9. The system according to claim 1, wherein: The educational robot supports the "emotion-learning dual-drive" function: (1) pushing deep learning tasks when in a positive emotional state (E>0.7), and triggering mindfulness meditation guidance when under high stress (E<0.4); (2) progressive trust opening: basic trust (0-40 points) provides error summary, and advanced trust (>60 points) generates multimodal psychological profiles. The trust score is based on a linear mapping of the trust value T (trust score = T × 100).

10. An emergency response mechanism, characterized by: (1) Level 1 risk response: The emotional composite value (E<0.3) triggers parent notification for three consecutive days and pushes a link to a psychological institution; (2) Data fuse mechanism: When abnormal logins exceed five times, the power supply of the storage medium is physically cut off, and the records are destroyed through blockchain data storage; (3) Compliance: Data collection complies with the "Personal Information Protection Law" and uses differential privacy (ε=0.5) + k-anonymity (k=5) technology for desensitization to meet GDPR requirements.

Images