A method and device for constructing an AI intelligent body of a smart campus heart platform

Abstract

The application discloses a kind of wisdom campus heart platform AI intelligent body construction method and device, belong to monitoring technical field.The application is under the campus private network to collect communication, psychological research, three kinds of multi-source data of campus activities, obtains structured feature set by washing desensitization standardization;Multi-modal fusion recognition model including multilayer structure is built, and school customized psychological recognition model is obtained by relying on data set training;Student psychological risk score is calculated in real time by using training model inference calculation and generates psychological portrait, and risk level is divided by referring to threshold value and double determination standard;According to risk level, start graded early warning and match individualized intervention scheme, after intervention data back, iteration optimization model parameters and grading threshold value are assisted by reinforcement learning.The application realizes student psychological dynamic monitoring, graded early warning and model autonomous optimization, and improves the intelligent level of campus psychological screening.

Description

Technical Field

[0001] This invention relates to the field of monitoring technology, and in particular to a method and apparatus for constructing an AI intelligent agent for a smart campus mental health education platform. Background Technology

[0002] Smart campus mental health management, as an important component of educational informatization, is widely used in the field of student mental state monitoring and intervention. Among related technologies, the collaborative operation of 5G base stations, campus private networks, and mobile terminal signaling has enabled the continuous and seamless collection of behavioral data such as student location trajectories, online duration, and communication frequency. Specifically, existing mental health education systems cover the entire process from questionnaire configuration and text storage to general large-scale model sentiment analysis and manual report viewing, including key aspects such as mental health education questionnaire management, text data storage, sentiment analysis, and manual screening, forming a static management model based on subjective reporting.

[0003] However, existing methods of school mental health education rely heavily on periodic questionnaires and subjective reporting, without connecting to telecommunications data sources. This can lead to poor data accuracy and significant distortion, or delayed monitoring and passive early warning, thus affecting the accuracy of psychological risk assessment and the timeliness of intervention. Specifically, because they depend entirely on students' subjective responses, there is a risk of deliberate concealment and emotional faking, and the lack of objective behavioral data corroborates the findings, resulting in a high rate of misjudgments and missed judgments. Furthermore, static batch processing on a monthly or quarterly basis cannot capture sudden psychological fluctuations, and the high degree of reliance on manual intervention leads to low intervention efficiency and an inability to address the hidden and dynamic nature of student psychological risks. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose a method for constructing an AI intelligent agent for a smart campus mental health education platform.

[0006] Another objective of this invention is to propose an AI intelligent body construction device for a smart campus mental health education platform.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention proposes a method for constructing an AI agent for a smart campus mental health education platform, comprising: S10, in the campus private network environment, collects students' objective telecommunications behavior data, subjective psychological data on mental health education, and campus auxiliary behavior data in real time, and cleans, desensitizes, and standardizes the collected multi-source raw data to generate a structured feature dataset. S20, Construct a multimodal fusion psychological recognition model, the model including a communication feature extraction layer, a psychological text feature extraction layer, a cross-modal feature fusion layer and a psychological state classification output layer, and train the model using the structured feature dataset to generate a campus-specific psychological recognition model; S30 utilizes a trained campus-specific psychological recognition model to perform real-time reasoning on the structured feature dataset collected in real time, dynamically generating a psychological risk score and psychological profile for each student, and automatically determining the psychological risk level based on preset grading thresholds and dual judgment rules. S40, based on the determined psychological risk level, automatically trigger a graded early warning and generate a personalized intervention plan, and after the intervention is completed, backfill the intervention effect data, and use the intervention effect data to iteratively update the parameters and graded threshold of the model through reinforcement learning.

[0010] In one embodiment of the present invention, S10 includes: After removing abnormal jitter data caused by base station signal drift or brief network outages, duplicate data, and invalid data with a missing rate greater than 30%, cleaned multi-source data is obtained. The sensitive fields of mobile phone number, ID card number, class and name in the cleaned multi-source data were masked and replaced by the national cryptographic SM4 telecom-grade encryption algorithm, and only the anonymous student ID was retained for correlation analysis, thus obtaining the de-identified multi-source data. The numerical features of traffic, duration and frequency in the desensitized multi-source data are normalized by maximum and minimum values, compressed to the range of 0 to 1, and the difference in units is eliminated to obtain the normalized numerical features. The psychological diaries and questionnaire texts in the desensitized multi-source data are segmented, stop words are removed, and semantic errors are corrected to generate Embedding feature vectors, converting the text into computable numerical features, thus obtaining the structured feature dataset.

[0011] In one embodiment of the present invention, the step of segmenting, removing stop words, and semantically correcting the psychological diaries and questionnaire texts in the desensitized multi-source data to generate an embedding feature vector includes: Perform word segmentation on the psychological diary and questionnaire text to obtain a word segmentation sequence; Stop words are removed from the segmented sequence to obtain a valid word sequence; Perform semantic error correction on the effective word sequence to obtain the corrected text data; The corrected text data is input into a pre-trained Embedding model to generate the Embedding feature vector.

[0012] In one embodiment of the present invention, S20 includes: Construct a dedicated extraction layer for communication features using CNN or GRU networks to automatically extract objective behavioral features such as sudden changes in base station trajectory, surges in nighttime traffic, sharp drops in call frequency, and abnormal online duration; A psychological text feature extraction layer using a BERT pre-trained model is constructed to extract subjective semantic features of emotions such as positive and negative, repression, anxiety, depression, and loneliness. A cross-modal feature fusion layer is constructed by combining feature concatenation and attention weights to deeply fuse the objective behavioral features with the subjective semantic features, highlighting the weights of highly correlated features. Construct a psychological state classification output layer that outputs a continuous psychological risk quantification score from 0 to 1 through a fully connected layer. The higher the score, the higher the psychological risk.

[0013] In one embodiment of the present invention, it further includes: Using a publicly available pre-trained large model as a foundation, historical psychological annotation data of students from this school is input, including questionnaires, consultation records and intervention results. Fine-tuning and training are performed until the model recognition accuracy is greater than 92% and convergence occurs, and then the model is solidified into the campus-specific psychological recognition model.

[0014] In one embodiment of the present invention, S30 includes: The real-time psychological risk score for each student is recalculated hourly using a sliding window method. By combining the latest changes in communication behavior and psychological data, the individual psychological profile curve is dynamically updated; It continuously tracks the trend of psychological state fluctuations, updates scores dynamically, and captures short-term emotional fluctuations and long-term psychological risks.

[0015] In one embodiment of the present invention, S40 includes: When the risk level is mild emotional fluctuation, the class teacher will be notified via a pop-up on the platform, and it is recommended to pay attention to the situation through daily communication. When the risk level is moderate anxiety risk, an alert will be issued via both SMS and the campus platform, and an analysis of the reasons for the student's abnormal behavior will be pushed to the system. When the risk level is classified as a severe psychological crisis, an urgent warning is issued, and the information is simultaneously sent to school administrators, the psychological counseling center, and homeroom teachers to initiate an emergency intervention process.

[0016] To achieve the above objectives, a second aspect of the present invention provides an AI intelligent agent construction device for a smart campus mental health education platform, comprising: The multi-source data acquisition and preprocessing module is used to collect students' objective telecommunications behavior data, subjective psychological data on mental health education, and campus auxiliary behavior data in real time under the campus private network environment, and to clean, desensitize, and standardize the collected multi-source raw data to generate a structured feature dataset. A multimodal model construction and training module is used to construct a multimodal fusion psychological recognition model. The model includes a communication feature extraction layer, a psychological text feature extraction layer, a cross-modal feature fusion layer, and a psychological state classification output layer. The model is trained using the structured feature dataset to generate a campus-specific psychological recognition model. The psychological risk assessment and profiling module is used to perform real-time reasoning on the structured feature dataset collected in real time using a trained campus-specific psychological recognition model, dynamically generate a psychological risk score and psychological profile for each student, and automatically determine the psychological risk level according to preset grading thresholds and dual judgment rules. The graded early warning iterative optimization module is used to automatically trigger graded early warnings and generate personalized intervention plans based on the determined psychological risk level. After the intervention is completed, the intervention effect data is backfilled, and the parameters and graded thresholds of the model are updated iteratively through reinforcement learning using the intervention effect data.

[0017] This invention discloses a method and apparatus for constructing an AI intelligent agent for a smart campus mental health education platform. By integrating objective behavioral data from telecommunications and subjective data from mental health education, a multimodal AI model is constructed to achieve seamless real-time monitoring, automatic hierarchical early warning, and closed-loop iterative optimization. Technical effects: Improves data authenticity and accuracy of judgment, enabling proactive risk assessment.

[0018] To achieve the above objectives, a third aspect of this application provides a computer device comprising: a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing the method described in the first aspect embodiment.

[0019] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the first aspect.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for constructing an AI intelligent agent for a smart campus mental health education platform according to an embodiment of the present invention; Figure 2 This is a system architecture diagram of an AI intelligent agent construction system for a smart campus mental health education platform according to an embodiment of the present invention; Figure 3 This is a structural diagram of an AI intelligent agent construction device for a smart campus mental health education platform according to an embodiment of the present invention; Figure 4 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] The following description, with reference to the accompanying drawings, describes a method and apparatus for constructing an AI intelligent agent for a smart campus mental health education platform according to an embodiment of the present invention.

[0025] Figure 1 This is a flowchart of a method for constructing an AI intelligent agent for a smart campus mental health education platform according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S10, in the campus private network environment, collects students' objective telecommunications behavior data, subjective psychological data on mental health education, and campus auxiliary behavior data in real time, and cleans, desensitizes, and standardizes the collected multi-source raw data to generate a structured feature dataset. S20, Construct a multimodal fusion psychological recognition model, the model including a communication feature extraction layer, a psychological text feature extraction layer, a cross-modal feature fusion layer and a psychological state classification output layer, and train the model using the structured feature dataset to generate a campus-specific psychological recognition model; S30 utilizes a trained campus-specific psychological recognition model to perform real-time reasoning on the structured feature dataset collected in real time, dynamically generating a psychological risk score and psychological profile for each student, and automatically determining the psychological risk level based on preset grading thresholds and dual judgment rules. S40, based on the determined psychological risk level, automatically trigger a graded early warning and generate a personalized intervention plan, and after the intervention is completed, backfill the intervention effect data, and use the intervention effect data to iteratively update the parameters and graded threshold of the model through reinforcement learning.

[0026] Furthermore, such as Figure 2 As shown. Step S1: Compliant real-time collection of multi-source data. In the closed-loop environment of the telecom campus private network, the data does not go to the public network or leave the campus. The three types of data are collected synchronously, frequently, and seamlessly. All data are timestamped and associated with anonymous student IDs. The entire process does not touch plain text names, mobile phone numbers, or ID card numbers, ensuring privacy and security from the source.

[0027] Furthermore, objective behavioral data on telecommunications communication (collected every 5 minutes): location trajectory of 5G base stations on campus, cell handover records; online / offline duration of mobile terminals, daily online time distribution; number of daily calls, call duration, and call time characteristics; data usage from 23:00 to 6:00 the next day, and types of active online apps; Furthermore, subjective psychological data for mental health education (synchronized in real time): psychological questionnaire responses and scale scores; psychological diaries, emotional reflections, and psychological counseling appointment / consultation records; psychological course check-ins and feedback on participation in psychological activities; campus auxiliary behavior data (synchronized in real time): attendance arrival / departure times and absence records; participation in campus activities, club activity levels, and library access records.

[0028] Step S2: Telecom-grade data cleaning, desensitization, and standardization preprocessing. The collected multi-source raw data undergoes four layers of standardization to remove noise, desensitize for privacy, and unify the format, forming a structured feature dataset that AI can directly compute. Data cleaning: Data with abnormal jitter caused by base station signal drift, brief network outages, duplicate data, and invalid data with a missing rate >30% is removed to ensure data quality. National cryptographic-grade desensitization: The national cryptographic SM4 telecom-grade encryption algorithm is used to mask and replace sensitive fields such as mobile phone numbers, ID card numbers, classes, and names, retaining only anonymous student IDs for correlation analysis. Numerical normalization: Numerical features such as traffic, duration, and frequency are normalized using maximum and minimum values, compressed to the 0-1 range, and the difference in units of measurement is eliminated. Text vector conversion: The psychological diary and questionnaire texts are segmented, stop words are removed, and semantic errors are corrected to generate embedding feature vectors, converting the text into numerical features that AI can compute.

[0029] Step S3: Construction and dedicated training of multimodal fusion AI psychological recognition model. Construct a four-layer fixed structure multimodal fusion model specifically adapted to campus mental health education scenarios. Use transfer learning + fine-tuning training with labeled data from the school to form a campus-specific psychological recognition model.

[0030] The first layer is a communication feature extraction layer, employing a CNN / GRU network to automatically extract objective behavioral features such as sudden changes in base station trajectory, surges in nighttime traffic, sharp drops in call frequency, and abnormal online duration. The second layer is a psychological text feature extraction layer, using a BERT pre-trained model to extract subjective semantic features such as positive / negative emotions, repression, anxiety, depression, and loneliness. The third layer is a cross-modal feature fusion layer, deeply fusing objective communication features with subjective text features through feature concatenation and attention-weighted fusion, highlighting highly correlated feature weights. The fourth layer is a psychological state classification output layer, outputting a continuous psychological risk quantification score of 0-1 through a fully connected layer; the higher the score, the higher the psychological risk. Training method: Using a publicly available pre-trained large model as a foundation, historical psychological data (questionnaires, consultation records, and intervention results) from students of this school are input for fine-tuning and training. Iteration continues until the model's recognition accuracy is >92% and converges, solidifying it into a campus-specific model.

[0031] Step S4: The AI ​​agent performs real-time reasoning and dynamically updates student psychological profiles. The mental health AI agent runs 24 / 7 on the campus network backend, enabling real-time reasoning, dynamic profiling, and continuous tracking. Reasoning frequency: Each student's real-time psychological risk score is recalculated hourly using a sliding window; Profile update: The individual psychological profile curve is dynamically updated by combining the latest changes in communication behavior (such as a surge in traffic at night for three consecutive days) and changes in psychological data (such as an increase in negative emotions in the student's diary); Tracking mode: The system continuously tracks the trend of psychological state fluctuations, not by scoring once, but by continuously and dynamically updating the score, accurately capturing short-term emotional fluctuations and long-term psychological risks.

[0032] Step S5: The AI ​​agent performs a four-level autonomous decision-making assessment. The AI ​​agent has built-in four fixed thresholds and dual judgment rules, automatically assessing the psychological risk level and matching intervention strategies without human intervention. Risk level classification (0~1 score range): 0~0.3: Normal, stable psychological state, no abnormal fluctuations; 0.3~0.5: Mild mood fluctuations, short-term low mood and anxiety, requiring daily attention; 0.5~0.7: Moderate anxiety risk, persistent negative emotions, psychological stress, requiring close attention and timely guidance; Above 0.7: Severe psychological crisis, with tendencies towards self-harm, depression, and extreme behavior, requiring emergency intervention. Dual judgment: Combining real-time comprehensive score + the magnitude of behavioral changes in the past 7 days (such as abnormal location trajectory, surge in nighttime internet usage) avoids misjudgment based on single data points and improves the accuracy of the classification.

[0033] Step S6: Tiered Automatic Early Warning Push and Personalized Intervention Plan Generation. The AI ​​agent automatically triggers tiered early warnings, pushes information through multiple channels, and generates personalized intervention plans based on risk levels. Mild risk (0.3~0.5): A platform pop-up reminds the homeroom teacher, suggesting daily communication and attention; Moderate risk (0.5~0.7): A dual-terminal early warning is issued via SMS and the campus platform, pushing an analysis of the student's abnormal behavior; Severe risk (above 0.7): An urgent early warning is issued, simultaneously pushing information to school administrators, the psychological center, and the homeroom teacher, initiating an emergency intervention process. Intervention plan generation: Automatically outputs an analysis of the student's abnormal behavior, key points for counseling, psychological guidance scripts, a home-school communication template, and suggested intervention cycles. No manual writing is required throughout the process, directly guiding intervention implementation.

[0034] Step S7: Intervention Result Feedback and Effect Recording. After completing the offline intervention, the school counselor and homeroom teacher quickly feed back the intervention effect on the platform within 1 minute: the improvement of students' status, the effectiveness of the intervention measures, whether continuous follow-up is needed, and the handling results, forming a closed-loop handling record for each warning-intervention-result, which serves as a high-quality training sample for model optimization.

[0035] Step S8: Reinforcement learning closed-loop iterative optimization, automatic model update and upgrade, AI agent regularly (weekly / monthly) extracts intervention feedback data, and iterates backward through reinforcement learning: Optimization direction: adjust cross-modal feature fusion weights, update four-level classification thresholds, and optimize abnormal behavior recognition rules; Optimization effect: the model continuously adapts to the psychological characteristics of students in this school, the recognition accuracy gradually improves, the misjudgment rate continues to decrease, and the problem of solidification and aging of traditional models is completely solved.

[0036] The embodiments of this invention also have the following technical effects: Deep integration of objective behavioral data from telecommunications communication and subjective data from campus mental health education to construct a psychological judgment system that "objectively corroborates subjective evidence"; a unique dual-modal fusion AI model combining communication features and psychological text, with a four-layer fixed structure adapted to campus mental health education scenarios; seamless real-time dynamic continuous monitoring via telecommunications, replacing traditional periodic static questionnaires and achieving early risk warning; a dedicated AI agent for mental health education, enabling automatic grading, automatic early warning, and automatic generation of personalized intervention plans; an intervention feedback reinforcement learning closed-loop iteration mechanism, continuously optimizing the model and making it more accurate with use; and telecommunications-grade national cryptographic desensitization + dedicated network isolation + tiered access control to construct a fully compliant privacy and security system.

[0037] To achieve the above embodiments, such as Figure 3 As shown, this embodiment also provides an AI intelligent agent construction device 10 for a smart campus mental health education platform, including: The multi-source data acquisition and preprocessing module 100 is used to collect students' objective telecommunications communication data, subjective psychological data on mental health education, and campus auxiliary behavior data in real time under the campus private network environment, and to clean, desensitize, and standardize the collected multi-source raw data to generate a structured feature dataset. A multimodal model construction and training module 200 is used to construct a multimodal fusion psychological recognition model. The model includes a communication feature extraction layer, a psychological text feature extraction layer, a cross-modal feature fusion layer, and a psychological state classification output layer. The model is trained using the structured feature dataset to generate a campus-specific psychological recognition model. The psychological risk assessment and profiling module 300 is used to perform real-time reasoning on the structured feature dataset collected in real time using a trained campus-specific psychological recognition model, dynamically generate a psychological risk score and psychological profile for each student, and automatically determine the psychological risk level according to a preset grading threshold and dual judgment rules. The graded early warning iterative optimization module 400 is used to automatically trigger graded early warning and generate personalized intervention plans based on the determined psychological risk level, and backfill intervention effect data after the intervention is completed. The intervention effect data is used to iteratively update the parameters and graded threshold of the model through reinforcement learning.

[0038] This invention discloses an AI intelligent agent construction device for a smart campus mental health education platform. By integrating objective telecommunications behavioral data with subjective mental health education data, a multimodal AI model is constructed to achieve seamless real-time monitoring, automatic hierarchical early warning, and closed-loop iterative optimization. Technical effects: Improves data authenticity and judgment accuracy, enabling proactive risk assessment.

[0039] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 4 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0040] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0041] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0042] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for constructing an AI intelligent agent for a smart campus mental health education platform, characterized in that, include: S10, in the campus private network environment, collects students' objective telecommunications behavior data, subjective psychological data on mental health education, and campus auxiliary behavior data in real time, and cleans, desensitizes, and standardizes the collected multi-source raw data to generate a structured feature dataset. S20, Construct a multimodal fusion psychological recognition model, the model including a communication feature extraction layer, a psychological text feature extraction layer, a cross-modal feature fusion layer and a psychological state classification output layer, and train the model using the structured feature dataset to generate a campus-specific psychological recognition model; S30 utilizes a trained campus-specific psychological recognition model to perform real-time reasoning on the structured feature dataset collected in real time, dynamically generating a psychological risk score and psychological profile for each student, and automatically determining the psychological risk level based on preset grading thresholds and dual judgment rules. S40, based on the determined psychological risk level, automatically trigger a graded early warning and generate a personalized intervention plan, and after the intervention is completed, backfill the intervention effect data, and use the intervention effect data to iteratively update the parameters and graded threshold of the model through reinforcement learning.

2. The method as described in claim 1, characterized in that, S10 includes: After removing abnormal jitter data caused by base station signal drift or brief network outages, duplicate data, and invalid data with a missing rate greater than 30%, cleaned multi-source data is obtained. The sensitive fields of mobile phone number, ID card number, class and name in the cleaned multi-source data were masked and replaced by the national cryptographic SM4 telecom-grade encryption algorithm, and only the anonymous student ID was retained for correlation analysis, thus obtaining the de-identified multi-source data. The numerical features of traffic, duration and frequency in the desensitized multi-source data are normalized by maximum and minimum values, compressed to the range of 0 to 1, and the difference in units is eliminated to obtain the normalized numerical features. The psychological diaries and questionnaire texts in the desensitized multi-source data are segmented, stop words are removed, and semantic errors are corrected to generate Embedding feature vectors, converting the text into computable numerical features, thus obtaining the structured feature dataset.

3. The method as described in claim 2, characterized in that, The process of segmenting, removing stop words, and semantically correcting the psychological diaries and questionnaire texts from the de-identified multi-source data to generate embedding feature vectors includes: Perform word segmentation on the psychological diary and questionnaire text to obtain a word segmentation sequence; Stop words are removed from the segmented sequence to obtain a valid word sequence; Perform semantic error correction on the effective word sequence to obtain the corrected text data; The corrected text data is input into a pre-trained Embedding model to generate the Embedding feature vector.

4. The method as described in claim 1, characterized in that, S20 includes: Construct a dedicated extraction layer for communication features using CNN or GRU networks to automatically extract objective behavioral features such as sudden changes in base station trajectory, surges in nighttime traffic, sharp drops in call frequency, and abnormal online duration; A psychological text feature extraction layer using a BERT pre-trained model is constructed to extract subjective semantic features of emotions such as positive and negative, repression, anxiety, depression, and loneliness. A cross-modal feature fusion layer is constructed by combining feature concatenation and attention weights to deeply fuse the objective behavioral features with the subjective semantic features, highlighting the weights of highly correlated features. Construct a psychological state classification output layer that outputs a continuous psychological risk quantification score from 0 to 1 through a fully connected layer. The higher the score, the higher the psychological risk.

5. The method as described in claim 4, characterized in that, The method further includes: Using a publicly available pre-trained large model as a foundation, historical psychological annotation data of students from this school is input, including questionnaires, consultation records and intervention results. Fine-tuning and training are performed until the model recognition accuracy is greater than 92% and convergence occurs, and then the model is solidified into the campus-specific psychological recognition model.

6. The method as described in claim 1, characterized in that, S30 includes: The real-time psychological risk score for each student is recalculated hourly using a sliding window method. By combining the latest changes in communication behavior and psychological data, the individual psychological profile curve is dynamically updated; It continuously tracks the trend of psychological state fluctuations, updates scores dynamically, and captures short-term emotional fluctuations and long-term psychological risks.

7. The method as described in claim 1, characterized in that, S40 includes: When the risk level is mild emotional fluctuation, the class teacher will be notified via a pop-up on the platform, and it is recommended to pay attention to the situation through daily communication. When the risk level is moderate anxiety risk, an alert will be issued via both SMS and the campus platform, and an analysis of the reasons for the student's abnormal behavior will be pushed to the system. When the risk level is classified as a severe psychological crisis, an urgent warning is issued, and the information is simultaneously sent to school administrators, the psychological counseling center, and homeroom teachers to initiate an emergency intervention process.

8. A smart campus mental health education platform AI intelligent body construction device, characterized in that, include: The multi-source data acquisition and preprocessing module is used to collect students' objective telecommunications behavior data, subjective psychological data on mental health education, and campus auxiliary behavior data in real time under the campus private network environment, and to clean, desensitize, and standardize the collected multi-source raw data to generate a structured feature dataset. A multimodal model construction and training module is used to construct a multimodal fusion psychological recognition model. The model includes a communication feature extraction layer, a psychological text feature extraction layer, a cross-modal feature fusion layer, and a psychological state classification output layer. The model is trained using the structured feature dataset to generate a campus-specific psychological recognition model. The psychological risk assessment and profiling module is used to perform real-time reasoning on the structured feature dataset collected in real time using a trained campus-specific psychological recognition model, dynamically generate a psychological risk score and psychological profile for each student, and automatically determine the psychological risk level according to preset grading thresholds and dual judgment rules. The graded early warning iterative optimization module is used to automatically trigger graded early warnings and generate personalized intervention plans based on the determined psychological risk level. After the intervention is completed, the intervention effect data is backfilled, and the parameters and graded thresholds of the model are updated iteratively through reinforcement learning using the intervention effect data.

9. A computer device, characterized in that, Including processor and memory; The processor runs a program corresponding to the executable program code stored in the memory to implement the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images