An ai psychological resilience real-time intervention method and system for college students

Abstract

The application discloses a kind of AI mental resilience real-time intervention method and system for college student group, the method includes the following steps: acquisition multimodal data;UniStressBERT model is constructed, text data is input into model, and text semantic feature and text emotion score are obtained;Individual baseline standardization is carried out to multimodal data, and standardized data are calculated;Risk monitoring is carried out based on text emotion score and standardized data by SRT algorithm;Scene data is encoded, and scene coding data are obtained;According to text semantic feature, standardized data, scene coding, mental resilience hidden state is calculated;Based on mental resilience hidden state, the PRI score of mental resilience dimension is calculated;Risk level is determined according to PRI score, and intervention strategy is executed according to risk level.The application greatly improves the recognition accuracy of campus stress event through model extraction and PRI dynamic evaluation, and enhances the intervention level.

Description

Technical Field

[0001] This invention relates to the field of mental health, and in particular to an AI-based real-time intervention method and system for psychological resilience in college students. Background Technology

[0002] In the current context of technology, mental health issues among university students are becoming increasingly prominent, but traditional mental health service systems and some solutions that incorporate AI technology still have many shortcomings.

[0003] Traditional university psychological counseling centers generally adopt a "passive waiting" model, meaning students need to actively make appointments and wait in line for counseling. However, due to stigma, lack of awareness, or inconvenient access to resources, the rate of students actively seeking help is generally low, leading to delays in the detection of many psychological crisis events and missing the best intervention opportunity.

[0004] To address this challenge, some universities and technology companies have attempted to introduce artificial intelligence technology and develop AI chatbots. However, existing AI chatbots have significant technical limitations:

[0005] (1) Superficial intervention level: Most products are limited to simple companionship chat or distribution of standardized psychological assessment questionnaires (such as PHQ-9, GAD-7), lacking the ability to intervene in deep psychological problems and unable to provide personalized guidance based on evidence-based psychological intervention theories (such as cognitive behavioral therapy CBT, mindfulness therapy).

[0006] (2) Lack of situational awareness: It is difficult to identify and understand the "stressors unique to college students" in a fine-grained manner, and it is impossible to distinguish the negative emotions caused by different situations such as academic pressure during exam week, job anxiety during job season, breakup of romantic relationship or interpersonal conflict in dormitory, so it is impossible to provide targeted support.

[0007] (3) Data privacy and compliance risks: Students’ mental health data and related campus behavior data (such as library records, consumption habits, social network comments, etc.) are extremely sensitive. Existing solutions often fail to meet strict privacy protection and compliance requirements in data collection, storage and use, making it difficult for the technology to be implemented and promoted on campus.

[0008] (4) Lack of closed loop and explainability: The intervention effect is difficult to quantify and evaluate. The system cannot optimize itself based on the intervention results. At the same time, the intervention process is not transparent to counselors, psychological teachers and other managers. There is a lack of traceable and auditable mechanisms, making it difficult to form a collaborative working closed loop of "AI initial screening intervention - human precise intervention". Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings and deficiencies of the existing technology and provide an AI-based real-time intervention method for psychological resilience in college students.

[0010] Another objective of this invention is to provide an AI-based real-time intervention system for psychological resilience in college students.

[0011] The objective of this invention is achieved through the following technical solution:

[0012] A real-time AI-based intervention method for psychological resilience in university students includes the following steps:

[0013] S1. Collect multimodal data, including text data, physiological data, behavioral data, and scene data;

[0014] S2. Construct the UniStressBERT model: Inject stress term cluster bias into the input layer, insert a domain adapter after each Transformer layer, and add soft ontology attention to the output layer.

[0015] Text data is input into the UniStressBERT model to obtain text semantic features and text sentiment scores;

[0016] S3. Standardize individual baselines based on physiological and behavioral data to obtain a standardization function, and then calculate standardized data using the standardization function.

[0017] S4. Based on text sentiment scores and standardized data, risk monitoring is performed using the SRT algorithm to obtain and determine the risk monitoring results:

[0018] When the risk monitoring result is greater than or equal to the first preset threshold, an assessment is triggered. If the assessment result meets the preset conditions, the process proceeds to step S5; otherwise, it is handled manually.

[0019] When the risk monitoring result is greater than or equal to the second preset threshold and less than the first preset threshold, local light intervention is performed.

[0020] When the risk monitoring result is less than the second preset threshold, the risk will continue to be monitored and the process will jump to S5.

[0021] S5. Encode the scene data to obtain scene-coded data; calculate the latent state of psychological resilience based on text semantic features, standardized data, and scene coding; calculate the PRI score of the psychological resilience dimension based on the latent state of psychological resilience; determine the risk level based on the PRI score and execute corresponding intervention actions.

[0022] In step S2, the pressure term cluster bias is specifically as follows:

[0023] Building a dynamic stress term library The initial vocabulary includes words related to campus stress, categorized by event type into three dimensions: academic, interpersonal, and life. Each dimension contains multiple clusters of stress words. The academic dimension includes clusters of exam stress and academic burnout; the interpersonal dimension includes clusters of dormitory conflict, romantic relationship problems, social anxiety, and club stress; and the life dimension includes clusters of employment anxiety, economic stress, health concerns, family stress, future uncertainty, and other events.

[0024] During the federated learning process, each edge node can report newly discovered stress words, which are then voted on globally and updated. Supports monthly expansion;

[0025] In the input layer, the input data is segmented and then matched. The token is used to embed its original words and add bias, specifically:

[0026] ;

[0027] in, These are pre-trained embedding vectors; To determine the injection intensity, the injection intensity is controlled and dynamically adjusted by federated learning based on the intervention results. The bias vector is shared by each stress word cluster.

[0028] In step S2, the domain adapter... The specific calculation formula is as follows:

[0029] ;

[0030] in, This represents the hidden state of the l-th layer Transformer. ; This is the downprojection matrix, which compresses the 768-dimensional matrix into a low-rank r-dimensional matrix; This is the upward projection matrix, which is used to restore r dimensions to 768 dimensions; The activation function is a Gaussian error linear unit.

[0031] In step S2, the soft ontology attention is specifically as follows:

[0032] Constructing the event ontology vector matrix The specific calculations are as follows:

[0033] ;

[0034] Each column vector Corresponding dynamic pressure term library The stress-related word clusters include: exam stress, academic burnout, dormitory conflict, relationship problems, social anxiety, club stress, employment anxiety, economic stress, health concerns, family stress, future uncertainty, and other events.

[0035] Based on the event ontology vector matrix Calculate the event query vector The specific calculation formula is as follows:

[0036] ;

[0037] in, [CLS] is the output after processing by the stress term cluster embedding and domain adaptation adapter; T is the transpose; The event projection matrix;

[0038] Calculate event attention score The specific calculation is as follows:

[0039] ;

[0040] in, The dot product similarity between the event query vector and the j-th event ontology vector; This is a scaling factor to prevent the softmax gradient from vanishing due to excessively large dot product values; This represents the relevance probability of the current text to the j-th stress word cluster, and the sum of the event attention scores is 1;

[0041] Calculate the weighted event vector The specific calculation is as follows:

[0042] ;

[0043] in, The event attention score; The event entity vector;

[0044] Constructing a dimensional ontology vector matrix The specific calculation is as follows:

[0045] ;

[0046] Each column vector This corresponds to the seven psychological resilience dimensions of PRI, including goal focus, emotional control, positive cognition, job confidence, resilience to setbacks, interpersonal harmony, and campus adaptation.

[0047] Calculate the dimension query vector The specific calculation formula is as follows:

[0048] ;

[0049] in, The dimension projection matrix;

[0050] Calculate the attention score of the dimension The specific calculation is as follows:

[0051] ;

[0052] in, This represents the dot product similarity between the dimension query vector and the i-th dimension ontology vector; The attention weights for the seven dimensions of psychological resilience form a probability distribution, and the sum of the attention scores for the seven dimensions is 1.

[0053] Calculate the weighted dimensional vector based on the dimensional attention score and the dimensional ontology vector. The specific calculation is as follows:

[0054] ;

[0055] The output after processing by the pressure term embedding and domain adaptation adapter Weighted event vector and weighted dimension vector splicing and fusion :

[0056] .

[0057] In step S3, the standardized function is calculated as follows:

[0058] ;

[0059] ;

[0060] ;

[0061] in, The mean baseline; The baseline is the standard deviation; This indicates the sensitivity of the control baseline to recent observations; These are observed values, represented as heart rate (HR), heart rate variability (RMSSD), learning interruption frequency (break), or negative emotion score. The data represents the mean values ​​up to time t-1. The standard deviation data is for time t-1 and earlier.

[0062] In step S4, the SRT algorithm is specifically as follows:

[0063] ;

[0064] in, Weights for multimodal data; Score the sentiment of the text. Indexed by 8 categories of emotions; , and It is obtained from the standardized function, where, This is the standardized value of resting heart rate; This represents the standardized value for heart rate variability. To learn the normalized value of the interrupt frequency; For scene indication functions;

[0065] when When the edge layer depth assessment is triggered, the assessment includes multimodal data fusion to construct observation vectors, and the UniStressBERT model to fully infer and predict the seven-dimensional components of PRI and event categories. If the assessment result meets the preset conditions, proceed to the next step; otherwise, it is processed manually.

[0066] when At the same time, local light interventions are implemented; the content of local light interventions includes sending psychological questionnaires, sending soothing light music, and sending humorous copy;

[0067] when At that time, maintain the monitoring status and proceed to the next step;

[0068] in, and It is a hysteresis dual threshold. The first preset threshold, This is the second preset threshold.

[0069] In step S5, the hidden state of psychological resilience specifically refers to:

[0070] Five types of real-time data are concatenated into an observation vector. The specific calculation formula is as follows:

[0071] ;

[0072] in, This represents the text semantic features output by the UniStressBERT model; Represents scene-coded data;

[0073] Calculate the state transition equation:

[0074] ;

[0075] in, Indicates the intervention effect value; For process noise, Represents a normal distribution;

[0076] The observation model uses a 2-layer MLP, specifically:

[0077] ;

[0078] Where ReLU is the activation function. For the first layer MLP, For the second layer MLP, For observation vectors;

[0079] The latent state of psychological resilience is calculated based on the state transition equation and the observation model. The specific calculation process is as follows:

[0080] ;

[0081] in, Indicates the reliability of the observation; This represents the prior prediction of the state at time t based on all data up to and including time t-1.

[0082] In step S5, the PRI score is specifically as follows:

[0083] PRI is decomposed into seven psychological resilience dimensions. These include vectors for goal focus, emotional control, positive cognition, employment confidence, setback recovery, interpersonal coordination, and campus adaptability.

[0084] The target focus vector is calculated as follows:

[0085] ;

[0086] in, and For feature weights; This function indicates the exam week. This represents the standardized value for late nights;

[0087] The emotion control degree vector is calculated as follows:

[0088] ;

[0089] in, and For feature weights; It is a set of three types of negative sentiment indexes;

[0090] The positive perception vector is calculated as follows:

[0091] ;

[0092] in, and For feature weights, Score the emotional tone of the text as joyful;

[0093] The employment confidence vector is calculated as follows:

[0094] ;

[0095] in, and For feature weights, Score the emotional impact of the anxious text. This is a function indicating the work scenario;

[0096] The frustration recovery vector is calculated as follows:

[0097] ;

[0098] in, and For feature weights, For viewing the window, Let PRI score be the score at time t. for PRI score at any given moment;

[0099] The interpersonal coordination vector is calculated as follows:

[0100] ;

[0101] in, and For feature weights, Score the emotional intensity of the angry text. The total number of interpersonal conflicts;

[0102] The campus fitness vector is calculated as follows:

[0103] ;

[0104] in, and For feature weights, Absence rate;

[0105] The specific calculation of PRI is as follows:

[0106] ;

[0107] Calibrate PRI scores using latent states of psychological resilience:

[0108] ;

[0109] If the equation does not hold true, adjust the weights of each dimension proportionally. Or correct To make it finally aligned When data for a certain dimension is missing, the vector for that dimension is directly assigned a value. Its weight Automatically set to 0.

[0110] The multimodal data uses the national cryptographic SM4-CBC and HMAC algorithms to encrypt and store sensitive data, and the key is rotated within a preset time period; sensitive fields in the multimodal data are customized and replaced using dynamic desensitization technology; a federated learning architecture is adopted, and only the encrypted gradient parameters are uploaded for global updates, without changing the original multimodal data.

[0111] Another objective of this invention is achieved through the following technical solution:

[0112] A real-time AI-based psychological resilience intervention system for university students includes a data acquisition module, an AI decision-making module, a privacy protection module, and a psychological resilience intervention module.

[0113] The data acquisition module collects multimodal data, including text data, physiological data, behavioral data, and scene data.

[0114] AI Decision Module: Constructing the UniStressBERT model: Injecting stress term cluster bias into the input layer, inserting a domain adapter after each Transformer layer, and adding soft ontology attention to the output layer;

[0115] Text data is input into the UniStressBERT model to obtain text semantic features and text sentiment scores;

[0116] Physiological and behavioral data are standardized to individual baselines to obtain a standardization function, and standardized data are calculated using the standardization function.

[0117] Privacy protection module: Based on text sentiment scores and standardized data, risk monitoring is performed using the SRT algorithm to obtain risk monitoring results and make judgments.

[0118] When the risk monitoring result is greater than or equal to the first preset threshold, a deep assessment is triggered;

[0119] When the risk monitoring result is greater than or equal to the second preset threshold and less than the first preset threshold, local light intervention is performed.

[0120] When the risk monitoring result is less than the second preset threshold, the risk will continue to be monitored and the next step will be taken.

[0121] The psychological resilience intervention module encodes the scene data to obtain scene-encoded data; calculates the latent state of psychological resilience based on text semantic features, standardized data, and scene encoding; calculates the PRI score of the psychological resilience dimension based on the latent state of psychological resilience; and determines the risk level based on the PRI score and executes corresponding intervention actions.

[0122] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0123] 1. This invention solves the coarse-grained problem of context perception by constructing a dynamic stress term library and multi-dimensional PRI calculation. It can accurately identify 12 typical campus stress events and associate them with corresponding emotional response patterns, thereby improving the scenario adaptability of intervention suggestions and the proportion of user feedback that "the psychological counseling suggestions provided are more in line with their actual needs".

[0124] 2. This invention constructs the UniStressBERT model to accurately identify campus-specific vocabulary, reducing misjudgments of stress events; and based on the dynamic expansion mechanism of the vocabulary under federated learning, the model has continuous evolution capabilities; by introducing an Adapter, parameters can be trained efficiently and are pluggable; and it achieves interpretable, interventionable, and quantifiable semantic enhancement, thereby enhancing the level of intervention.

[0125] 3. This invention avoids terminal resource consumption and privacy leakage through the SRT algorithm, solving data security and compliance issues; it constructs a triple privacy protection system of "local encrypted storage + dynamic desensitization + federated learning", which meets the data security standards of the education industry, completely eliminates the concerns of universities about data privacy, and clears the way for the implementation of technology on campus.

[0126] 4. This invention, through the combination of "multimodal data acquisition, SRT algorithm and PRI dynamic assessment", completely changes the core limitation of the traditional model of "passively waiting for students to ask for help", and advances the intervention point from "after students ask for help" to "the early stage of risk", thereby improving the early identification rate of potential psychological crises.

[0127] 5. This invention, through a model that prioritizes AI-driven autonomous intervention supplemented by human intervention, greatly improves the efficiency of psychological counseling service coverage and the accessibility of psychological services for students in pilot scenarios. At the same time, AI takes on low-risk counseling needs, eliminating the need for human counselors to handle homogeneous basic issues, thus improving per capita service efficiency. This allows human counselors to focus their efforts on high-risk cases and in-depth interventions, forming an intervention closed loop. Attached Figure Description

[0128] Figure 1 This is a flowchart of the AI-based real-time intervention method for psychological resilience in college students, as described in this invention.

[0129] Figure 2 This is a schematic diagram of the UniStressBERT model described in this invention.

[0130] Figure 3 This is a flowchart of the PRI scoring process described in this invention. Detailed Implementation

[0131] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0132] like Figure 1 A real-time AI-based intervention method for psychological resilience in university students includes the following steps:

[0133] S1. Collect multimodal data, including text data, physiological data, behavioral data, and scene data;

[0134] S2. Construct the UniStressBERT model: Inject stress term cluster bias into the input layer, insert a domain adapter after each Transformer layer, and add soft ontology attention to the output layer.

[0135] Text data is input into the UniStressBERT model to obtain text semantic features and text sentiment scores;

[0136] S3. Standardize individual baselines based on physiological and behavioral data to obtain a standardization function, and then calculate standardized data using the standardization function.

[0137] S4. Based on text sentiment scores and standardized data, risk monitoring is performed using the SRT algorithm to obtain and determine the risk monitoring results:

[0138] When the risk monitoring result is greater than or equal to the first preset threshold, an assessment is triggered. If the assessment result meets the preset conditions, the process proceeds to step S5; otherwise, it is handled manually.

[0139] When the risk monitoring result is greater than or equal to the second preset threshold and less than the first preset threshold, local light intervention is performed.

[0140] When the risk monitoring result is less than the second preset threshold, the risk will continue to be monitored and the process will jump to S5.

[0141] S5. Encode the scene data to obtain scene-encoded data; calculate the latent state of psychological resilience based on text semantic features, standardized data, and scene encoding; calculate the PRI score for the psychological resilience dimension based on the latent state of psychological resilience; determine the risk level based on the PRI score.

[0142] If the risk is low, then AI will intervene;

[0143] If the risk level is medium, then AI intervention and manual review will be used.

[0144] If the risk is high, then manual takeover and intervention will be implemented.

[0145] In step S1, multimodal data acquisition involves real-time collection of text data, behavioral data, physiological data, and scenario data. Text data includes social media posts and online consultation texts; behavioral data includes campus app logs, dormitory access records, library visit durations, and frequency of study interruptions; physiological data includes heart rate, heart rate variability, and sleep quality (duration and number of interruptions); and scenario data includes scenarios such as exam weeks, job hunting season, holidays, and daily routines.

[0146] The multimodal data uses the national cryptographic SM4-CBC and HMAC algorithms to encrypt and store sensitive data, with the key rotating every 24 hours. Sensitive fields in the multimodal data are dynamically de-identified and replaced. A federated learning architecture is adopted, receiving only encrypted gradients. A global model update is performed, with no original data leaving the system; among which, The noise variance for differential privacy is used to protect the model gradient from leaking the original data. N is a normal distribution, and g is the original gradient.

[0147] like Figure 2 This invention proposes the UniStressBERT model, a deep improvement upon the Hugging Face open-source hfl / chinese-roberta-wwm-ext pre-trained language model, namely the RoBERTa model. Its infrastructure consists of a 12-layer Transformer encoder (768 hidden layer dimensions), employing a whole-word masking pre-training strategy and trained on a large-scale Chinese corpus. This invention features three core improvements designed specifically for the psychological resilience of university students: a domain-adaptive adapter, stress-related word cluster embedding injection, and soft ontology attention fusion, enabling it to possess a deep understanding of the semantics of campus stress.

[0148] In terms of basic architecture, this invention does not modify the internal structure of the original 12-layer Transformer. Instead, it inserts a low-rank Adapter module after each layer, injects stress term cluster bias into the input layer, and adds soft ontology attention fusion into the output layer, forming a three-layer enhanced architecture of "Adapter embedding + term cluster enhancement + ontology focusing". The total number of parameters increases by only 4.2%, and the inference latency increases by <15ms.

[0149] The infrastructure includes:

[0150] Input layer: word embedding (768 dimensions) + position embedding (max_len=512) + sentence embedding;

[0151] 12-layer Transformer encoder: Each layer contains multi-head self-attention (12 heads × 64 dimensions) + feedforward network (3072-dimensional hidden layer) + residual connections and layer normalization;

[0152] Output layer: 768-dimensional hidden states labeled [CLS] are represented as sequences.

[0153] The specific pressure term bias is as follows:

[0154] Building a dynamic stress term library Initially containing 200+ words related to campus stress, categorized into 12 groups based on event type, and clearly defined across three dimensions:

[0155] Dimension 1: Academic Dimension

[0156] Exam stress clusters: {"practice questions", "failing a course", "retaking a course", "GPA", "make-up exam", "deferred exam"};

[0157] Academic burnout cluster: {"homework", "paper", "final exam", "deadline", "group", "report", "graduate school recommendation", "graduate school entrance exam", "interview", "adjustment", "score line", "choice", "procrastination", "schedule", "fatigue"};

[0158] Dimension 2: Interpersonal Dimension

[0159] Dormitory conflict clusters: {"Lights out", "Snoring", "Hygiene", "Electricity bill", "Roommates", "Sleep schedule"};

[0160] Relationship troubles cluster: {"relationship", "breakup", "friendship", "betrayal", "social"};

[0161] Social anxiety cluster: {"social", "isolation", "conflict", "help", "bullying"};

[0162] Club stress cluster: {"clubs", "extracurricular activities", "tasks", "division of labor", "stress"};

[0163] Dimension 3: Lifestyle Dimension

[0164] Job anxiety cluster: {"Autumn recruitment", "Spring recruitment", "Offer", "Resume", "Interview", "Tripartite agreement"};

[0165] Economic pressure cluster: {"tuition fees", "living expenses", "part-time jobs", "loans", "consumption", "poverty"};

[0166] Health Concerns Cluster: {"Insomnia", "Anxiety", "Depression", "Illness", "Physical Examination", "Psychological Counseling"};

[0167] Family stress cluster: {"Parents", "Expectations", "Comparisons", "Going home", "Phone calls", "Stress"};

[0168] The group facing uncertainty about the future: {"going abroad", "studying abroad", "career", "confusion", "goals", "uncertainty"};

[0169] Other event clusters: {"sudden", "unexpected", "change", "trouble"}.

[0170] During the federated learning process, each edge node can report newly discovered stress words, which are then voted on in the cloud (confirmed by ≥3 schools) and updated globally. Supports monthly expansion;

[0171] In the input layer, the input data is segmented and then matched. The token is used to embed its original words and add bias, specifically:

[0172] ;

[0173] in, The embedding vectors pre-trained for the RoBERTa model are the original vector representations of the corresponding tokens in the vocabulary lookup matrix provided by the open-source model. To determine the injection intensity, the injection intensity is controlled and dynamically adjusted by federated learning based on the intervention results. The bias vector is shared by each stress word cluster.

[0174] This invention innovatively modifies the hit pressure word cluster library while preserving the pre-trained embedding. For campus-specific vocabulary (such as "guaranteed postgraduate admission" and "autumn recruitment"), additional trainable bias terms are added. This forms an enhanced word embedding. This allows for enhanced perception of campus stress contexts without compromising common semantics.

[0175] This invention uses stress term embedding to inject semantic bias into campus-specific terms such as "failing a course," "autumn recruitment," and "lights out." This reduces the model's misjudgment rate for core stress events such as academics, employment, and interpersonal relationships by about 37%, enabling it to clearly distinguish between "anxiety caused by failing an exam" and ordinary "low mood."

[0176] Thanks to the dynamic expansion mechanism of the lexicon under federated learning, the embedded stress word cluster has the ability to continuously evolve. In the six-month cross-school deployment, the lexicon size automatically expanded from the initial 200+ entries to more than 340 entries, and the recognition coverage of emerging stress topics (such as "online exam anxiety") increased by 65%.

[0177] In traditional techniques, directly fine-tuning all RoBERTa parameters can lead to overfitting (the average number of psychological data for college students is less than 5,000 samples per school), and deploying the full model (approximately 400MB) at the edge consumes too much resources. This invention inserts a low-rank bottleneck layer after the output of each Transformer layer, trains only the Adapter parameters, and freezes the original RoBERTa weights.

[0178] Domain Adaptor The specific calculation formula is as follows:

[0179] ;

[0180] in, This represents the hidden state of the l-th layer Transformer. ; This is the downprojection matrix, which compresses the 768-dimensional matrix into a low-rank r-dimensional matrix; This is the upward projection matrix, which is used to restore r dimensions to 768 dimensions; The activation function is a Gaussian error linear unit.

[0181] The rank r is a configurable parameter that balances performance and efficiency. .when At that time, each Adapter layer only adds With 12 layers, approximately 0.6 million parameters were added, increasing the model size to around 420MB.

[0182] This invention achieves efficient parameter training by introducing a domain adaptation adapter, updating only 0.6M parameters (compared to 102M parameters for full fine-tuning), increasing training speed by 8 times and reducing GPU memory usage by 60%. Different schools can train their adapters independently, and only the adapter parameters are merged during cloud aggregation to avoid catastrophic forgetting. During inference, the adapter can be loaded / unloaded, and low-end terminals can use only the basic RoBERTa (accuracy decrease <15%).

[0183] While the [CLS] representation of the basic RoBERTa model has powerful general semantic representation capabilities, it has two core shortcomings when directly applied to psychological resilience assessment: 1) It lacks explicit alignment with the ontology of stress events for college students, making it impossible to accurately distinguish the semantic differences between campus-specific scenarios such as "exam anxiety" and "employment anxiety"; 2) It lacks explicit mapping of the seven dimensions of psychological resilience, resulting in a mismatch between semantic features and the psychological concept space when calculating the PRI index, affecting the accuracy of the assessment.

[0184] This invention innovatively introduces ontology nodes from a stress terminology library as external semantic anchors and designs a soft ontology attention fusion mechanism, namely a two-layer attention fusion mechanism (event layer + dimension layer). Through a learnable projection matrix, the general language representation space and the psychological professional concept space are forcibly aligned to achieve interpretable, interventionable, and quantifiable semantic enhancement.

[0185] The specific details of the soft ontology attention are as follows:

[0186] Constructing the event ontology vector matrix The specific calculations are as follows:

[0187] ;

[0188] Each column vector Corresponding dynamic pressure term library The stress word clusters include (1) exam stress cluster, (2) academic burnout cluster, (3) dormitory conflict cluster, (4) relationship troubles cluster, (5) social phobia cluster, (6) club stress cluster, (7) employment anxiety cluster, (8) economic stress cluster, (9) health concerns cluster, (10) family stress cluster, (11) future uncertainty cluster, and (12) other events cluster;

[0189] Projection matrix (used to align semantic space):

[0190] ;

[0191] These two matrices map the [CLS] representation and ontology vector to a unified query-key aligned space. They are trainable parameters initialized to a random Gaussian distribution N(0, 0.02). 2 ).

[0192] Based on the event ontology vector matrix Calculate the event query vector (Precise identification of pressure sources), the specific calculation formula is as follows:

[0193] ;

[0194] in, [CLS] is the output after processing by the stress term cluster embedding and domain adaptation adapter; T is the transpose; The event projection matrix maps general semantics to the event query space;

[0195] Calculate event attention score The specific calculation is as follows:

[0196] ;

[0197] in, The dot product similarity between the event query vector and the j-th event ontology vector; This is a scaling factor to prevent the softmax gradient from vanishing due to excessively large dot product values; This represents the probability of relevance between the current text and the j-th stress word cluster, with the sum of the attention scores for the 12 events being 1;

[0198] Calculate the weighted event vector The specific calculation is as follows:

[0199] ;

[0200] in, The event attention score; For the event ontology vector, Convert the generic [CLS] representation into an event-aware semantic representation;

[0201] Constructing a dimensional ontology vector matrix The specific calculation is as follows:

[0202] ;

[0203] Each column vector The seven psychological resilience dimensions of PRI are: (1) goal focus, (2) emotional control, (3) positive cognition, (4) job confidence, (5) setback recovery, (6) interpersonal coordination, and (7) campus adaptability.

[0204] Calculate the dimension query vector The specific calculation formula is as follows:

[0205] ;

[0206] in, The dimension projection matrix;

[0207] Calculate the attention score of the dimension The specific calculation is as follows:

[0208] ;

[0209] in, This represents the dot product similarity between the dimension query vector and the i-th dimension ontology vector; The attention weights for the seven dimensions of psychological resilience form a probability distribution, and the sum of the attention scores for the seven dimensions is 1.

[0210] Calculate the weighted dimensional vector based on the dimensional attention score and the dimensional ontology vector. The specific calculation is as follows:

[0211] ;

[0212] in, The contextual representation of the generated dimension perception is directly used for subsequent PRI seven-dimensional component calculation, realizing an explicit mapping from the language space to the psychometric space.

[0213] The final representation is the fusion and downstream task computation:

[0214] The output after processing by the pressure term embedding and domain adaptation adapter Weighted event vector and weighted dimension vector splicing and fusion :

[0215] ;

[0216] in, It represents 2304 dimensions, consisting of 768 ([CLS]) + 768 (event-aware) + 768 (dimension-aware); Map the general [CLS] to the semantic space of psychological resilience; While preserving the basic semantics, event type and resilience dimension information are explicitly injected, enabling downstream task heads to directly utilize the aligned professional features.

[0217] The three-task downstream design achieves seamless end-to-end integration from text perception to decision input by converting the enhanced semantic representation into standardized psychological indicators required by the system in one go. This UniStressBERT model architecture avoids the cumbersome process of manually designing rules for feature transformation in traditional solutions. It not only reduces semantic loss and cross-module alignment errors, but also allows the three core functions of emotion recognition, event classification, and resilience assessment to share the same interpretable feature space, greatly improving the system's maintainability and iteration efficiency. At the same time, since all outputs originate from a unified ontology-enhanced representation, a complete chain of evidence for intervention decisions is established, meeting the stringent requirements of traceability and audit compliance in psychological service scenarios.

[0218] To address the individual differences in the meaning of "the same heart rate variability has different implications for different students", the system independently maintains an exponential moving baseline for three types of physiological behavioral indicators for each student and calculates standardized functions.

[0219] The standardized function is calculated as follows:

[0220] ;

[0221] ;

[0222] ;

[0223] in, The mean baseline represents the long-term average level of a student's physiological indicator (such as heart rate) and reflects the locational characteristics of that indicator. The standard deviation baseline characterizes the index around... The fluctuation range reflects its discrete characteristics; these two baselines are maintained independently by each student and stored encrypted locally on the terminal, and the data is not uploaded to the cloud;

[0224] This indicates the sensitivity of the control baseline to recent observations. A value of 0.001 indicates a baseline update cycle of approximately 1000 observations (about 42 days), which is suitable for the stabilization period at the beginning of the semester; a value of 0.05 indicates an update cycle of approximately 20 observations (about 1 day), which is suitable for rapid response during exam weeks. These are observed values, represented as heart rate (HR), heart rate variability (RMSSD), learning interruption frequency (break), or negative emotion score. The data represents the mean values ​​up to time t-1. The standard deviation data is for time t-1 and earlier.

[0225] This standardization function ensures that multimodal features participate in subsequent calculations at the same order of magnitude, addressing physiological indicators (such as...) ) and behavioral indicators (such as ) dimensional differences.

[0226] To avoid terminal resource consumption and privacy leakage, this invention designs the SRT (Streaming Risk Trigger) algorithm, which decouples risk identification into rapid end-side screening and deep edge inference.

[0227] The SRT algorithm is specifically as follows:

[0228] ;

[0229] in, These are the weights for multimodal data, and their sum is 1 (default). ;

[0230] Score the sentiment of the text. Eight sentiment indices are generated, calculated using the UniStressBERT model. The specific calculation process is as follows: the terminal loads the quantized UniStressBERT model (retaining the low-rank adapter and stress cluster injection), segments the input text, and then sequentially processes it through an embedding layer, 12 layers of Transformer encoding, and soft ontology attention fusion. The input emotion classification header is processed by softmax normalization, and the output is an 8-dimensional probability distribution, whose k-th dimension is the probability distribution. ;

[0231] , and It is obtained from the standardized function, where, This is a standardized value of resting heart rate, reflecting the level of physiological arousal. This represents the standardized value for heart rate variability. To learn the standardized value of the interruption frequency, which reflects the degree of behavioral disorder; This is a scene indicator function; it defaults to 1 when the calendar is marked as exam week or job hunting season.

[0232] when When the edge layer depth assessment is triggered, the assessment includes multimodal data fusion to construct observation vectors, and the UniStressBERT model to perform complete inference and prediction of the seven-dimensional components of PRI and event categories. Based on the assessment results, it is decided whether to proceed to the next step or transfer to manual processing. For example, if the observation vectors are successfully constructed or the seven-dimensional components of PRI are predicted, proceed to the next step.

[0233] when At the same time, local light interventions are implemented; the content of local light interventions includes sending psychological survey questionnaires, sending soothing light music, and sending humorous copy;

[0234] when At that time, maintain the monitoring status and proceed to the next step;

[0235] in, and It is a hysteresis dual threshold. The first preset threshold is 0.7, with a default value of 0.7. This is the second preset threshold, with a default value of 0.3.

[0236] In step S5, the hidden state of psychological resilience specifically refers to:

[0237] Five types of real-time data are concatenated into an observation vector. The specific calculation formula is as follows:

[0238] ;

[0239] in, This represents the text semantic features output by the UniStressBERT model; Represents scene-coded data (exam week, job-hunting season, holidays, daily life);

[0240] Calculate the state transition equation:

[0241] ;

[0242] in, Indicates the intervention effect value; This represents process noise, with a value of 2, and represents the uncertainty of control state transitions, set by expert experience. Represents a normal distribution;

[0243] The observation model uses a 2-layer MLP, specifically:

[0244] ;

[0245] ReLU is the activation function that accelerates convergence and prevents gradient vanishing. For the first layer MLP, For the second layer MLP, The hidden layer dimension of 512 represents a balance between expressive power and computational cost.

[0246] A fixed-period sampling mechanism is adopted, defining the time index t as the discrete sampling time and the sampling interval as... Minutes; therefore, t represents the current sampling time; t-1 represents the previous sampling time.

[0247] The latent state of psychological resilience is calculated based on the state transition equation and the observation model. The specific calculation process is as follows:

[0248] ;

[0249] in, Indicates the reliability of the observation; This represents the prior prediction of the state at time t based on all data up to and including time t-1, i.e., the estimated state before obtaining new observations at time t. This is the posterior estimate obtained after fusing the observation data at time t, which is the final output PRI hidden state.

[0250] like Figure 3This invention upgrades the Psychological Resilience Index (PRI) from a traditional questionnaire-based static assessment to a dynamic estimation of latent variables, addressing two core issues: individual heterogeneity and the lack of multimodal data. The system utilizes a temporal Bayesian filtering framework to fuse physiological behavioral data from the SRT algorithm with textual semantic features from the UniStressBERT model in real time, generating an interpretable, calibrable, and traceable dynamic resilience profile.

[0251] The specific PRI score is as follows:

[0252] PRI is decomposed into seven psychological resilience dimensions. These include vectors for goal focus, emotional control, positive cognition, employment confidence, setback recovery, interpersonal coordination, and campus adaptability.

[0253] The target focus vector is calculated as follows:

[0254] ;

[0255] in, and These are the feature weights of the exam week indicator function and the standardized value of late nights, respectively. The initial values ​​are set based on the experience of psychology experts and the statistical correlation of large-scale historical samples. , And ensure that the sum of the weights is 1. Subsequently, as the system runs, the weights can be dynamically fine-tuned through an online learning mechanism: after each effective intervention, the system records the actual contribution of the features that played a role in that intervention, and updates the weights using gradient descent to avoid overfitting; This function indicates the exam week. This represents the standardized value for late nights;

[0256] The emotion control degree vector is calculated as follows:

[0257] ;

[0258] in, and These are the feature weights for the sum of negative emotion scores in the text and the standardized heart rate value, respectively, used to quantify the contribution of the two modalities in the assessment of emotion control. Corresponding to the subjective emotional experience dimension, the intensity of students' subjective emotions is reflected by aggregating the text emotion scores of three types of negative emotions (anxiety, depression, and anger); Corresponding to the physiological arousal dimension, through standardized heart rate values To characterize students' physiological stress levels; It is a set of three types of negative sentiment indexes;

[0259] The positive perception vector is calculated as follows:

[0260] ;

[0261] in, and These are the characteristic weights of quantifying the peak of negative emotions and the positive joy emotions on positive cognition, respectively. The fact that negative emotional peaks dominated the assessment reflects the psychological theory of "negativity bias," which states that negative events typically have a more significant impact on an individual's psychology than positive events of equal intensity. This reflects the regulatory effect of positive emotions; Score the emotional tone of the text as joyful;

[0262] The employment confidence vector is calculated as follows:

[0263] ;

[0264] in, and These are the standardized values ​​of the learning interruption frequency. Anxiety score The feature weights are calculated only during the job-hunting season; the initial values ​​are set to... =0.4、 =0.6, to reflect that the impact of anxiety on employment confidence is generally stronger than that of behavioral learning interruption; Score the text's sentiment level as anxious. This is a function indicating the work scenario;

[0265] The frustration recovery vector is calculated as follows:

[0266] ;

[0267] in, and These are the PRI rate of decrease and the standardized value of heart rate variability, respectively. The feature weights, both satisfy the normalization condition. + =1, initial value set to =0.6、 =0.4; The observation window is 72 hours. Let PRI score be the score at time t. for PRI score at any given moment;

[0268] The interpersonal coordination vector is calculated as follows:

[0269] ;

[0270] in, and These are the feature weights for anger score and the feature weights for conflict frequency, respectively. Quantifying the negative impact of emotional intensity on interpersonal harmony. The degree of negative impact of quantified event frequency. Both satisfy the normalization condition. + =1, initial value set to =0.6、 =0.4, to reflect that the intensity of emotion has a slightly higher impact on interpersonal coordination than the frequency of events; Score the emotional intensity of the angry text. The total number of interpersonal conflicts is identified through the text pattern "roommate / classmate / mentor + negative words";

[0271] The campus fitness vector is calculated as follows:

[0272] ;

[0273] in, and These are the feature weights for absenteeism rate and learning interruption frequency, respectively. Quantify the negative impact of absenteeism rate on campus adaptability. The negative impact of the frequency of quantification learning interruptions is assessed. Both satisfy the normalization condition. + =1, initial value set to =0.5、 =0.5, to reflect that academic participation and behavioral stability have equally important effects on campus adjustment; Absence rates are identified through text patterns of "roommate / classmate / tutor + negative words";

[0274] The specific calculation of PRI is as follows:

[0275] ;

[0276] Indicates dynamic weights. (Exam week) , Increased by 30% during job-hunting season Increase by 50%, automatically triggered through scene recognition.

[0277] Calibrate PRI scores using latent states of psychological resilience:

[0278] ;

[0279] If the equation does not hold true, adjust the weights of each dimension proportionally. Or correct To make it finally aligned When data for a certain dimension is missing, the vector for that dimension is directly assigned a value. Its weight Automatically set to 0. The participation of all components in initialization, dynamic calibration and missing compensation is the core constraint variable that ensures the consistency of seven-dimensional decomposition and comprehensive evaluation.

[0280] Risk level is determined based on PRI score:

[0281] If the risk level is low (PRI score ≥ 70), AI intervention will be implemented.

[0282] If the risk level is medium (50 points ≤ PRI < 70 points), then AI intervention and manual review will be used.

[0283] If the risk is high (PRI < 50 points), then manual takeover and intervention will be implemented.

[0284] For example, when AI intervenes, it can generate intervention plans based on the scenario. For instance, based on the matching result of Zhang's "exam stress → decreased focus on goals", three candidate plans can be generated. Each plan includes three parts: "theoretical basis - implementation steps - expected effect".

[0285] Option A: "Pre-exam cognitive restructuring" (based on CBT theory, using 5 rounds of Socratic questioning to correct the catastrophic cognition that "not being able to remember knowledge points = failing the exam", with an expected improvement in focus of 10-15 points);

[0286] Option B: "Pomodoro Technique Time Management" (based on time management psychology, it sets a "25-minute focus + 5-minute rest" cycle, with professional course knowledge point breakdown, and the expected goal is to improve focus by 15-20 minutes).

[0287] Option C: "4-7-8 Breathing Relaxation Training" (based on mindfulness therapy, using voice guidance to control the breathing rhythm, inhale for 4 seconds, hold your breath for 7 seconds, and exhale for 8 seconds, with an expected improvement in emotional control of 12-18 points).

[0288] By calling the user preference model (trained based on Zhang's three past interaction data with an accuracy of 89%), the AI ​​analysis showed that Zhang had a probability of choosing "voice guidance" about 70% of the time, "text interaction" about 30% of the time, and a click conversion rate of 65% for "time management tools" (higher than cognitive reconstruction 42% and breathing training 38%).

[0289] Based on the intervention goal (prioritizing improving focus), the optimal solution was determined to be B (the Pomodoro Technique), and an interactive script integrating "voice + text" was generated (voice was in a female voice, speaking speed of 120 words / minute, text font was Source Han Sans, font size 16, line spacing 1.5).

[0290] The system triggers a voice reminder every 45 minutes (e.g., "Current focus period is over. You can get up and move around for 5 minutes. We suggest you drink some water or look into the distance"). During this period, the effectiveness is monitored every 5 minutes using the following methods:

[0291] Text monitoring: One lightweight feedback question was pushed (e.g., "How focused are you right now? A. Very focused B. Average C. Easily distracted"); Zhang gave "A" twice and "B" once.

[0292] Physiological monitoring: The smart bracelet uploads heart rate data in real time, which has decreased from 85 beats / minute to 72 beats / minute, and HRV has increased from 42ms to 55ms;

[0293] Behavioral monitoring: The library's smart terminal recorded that Zhang's study interruption frequency decreased from 4.2 times / hour to 1.8 times / hour;

[0294] Effectiveness evaluation: PRI scores are calculated every 5 minutes (62 points at the 10th minute, 68 points at the 20th minute, and 73 points at the 25th minute). An improvement of 28.1% ≥ 15% of the effective intervention threshold is recorded. The effectiveness evaluation unit generates an "effective intervention" label and includes the plan and parameters (such as the 25-minute focus cycle and the breakdown of computer science knowledge points) in the AI ​​Agent's "success case library" (category tags: exam week - computer science - target focus). The AI ​​Agent can be iteratively optimized based on the intervention results.

[0295] A real-time AI-based psychological resilience intervention system for university students includes a data acquisition module, an AI decision-making module, a privacy protection module, and a psychological resilience intervention module.

[0296] The data acquisition module collects multimodal data, including text data, physiological data, behavioral data, and scene data.

[0297] AI Decision Module: Constructing the UniStressBERT model: Injecting stress term cluster bias into the input layer, inserting a domain adapter after each Transformer layer, and adding soft ontology attention to the output layer;

[0298] Text data is input into the UniStressBERT model to obtain text semantic features and text sentiment scores;

[0299] Physiological and behavioral data are standardized to individual baselines to obtain a standardization function, and standardized data are calculated using the standardization function.

[0300] Privacy protection module: Based on text sentiment scores and standardized data, risk monitoring is performed using the SRT algorithm to obtain risk monitoring results and make judgments.

[0301] When the risk monitoring result is greater than or equal to the first preset threshold, a deep assessment is triggered;

[0302] When the risk monitoring result is greater than or equal to the second preset threshold and less than the first preset threshold, local light intervention is performed.

[0303] When the risk monitoring result is less than the second preset threshold, the risk will continue to be monitored and the next step will be taken.

[0304] The psychological resilience intervention module encodes scene data to obtain scene-encoded data; it calculates the latent state of psychological resilience based on text semantic features, standardized data, and scene encoding; it calculates the PRI score for the psychological resilience dimension based on the latent state of psychological resilience; and it determines the risk level based on the PRI score.

[0305] If the risk is low, AI will intervene.

[0306] If the risk level is medium, then AI intervention and manual review will be used.

[0307] If the risk is high, then manual takeover and intervention will be implemented.

[0308] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A real-time AI-based intervention method for psychological resilience in university students, characterized in that: Includes the following steps: S1. Collect multimodal data, including text data, physiological data, behavioral data, and scene data; S2. Construct the UniStressBERT model: Inject stress term cluster bias into the input layer, insert a domain adapter after each Transformer layer, and add soft ontology attention to the output layer. Text data is input into the UniStressBERT model to obtain text semantic features and text sentiment scores; The specific details of the soft ontology attention are as follows: Constructing the event ontology vector matrix The specific calculations are as follows: ； Each column vector Corresponding dynamic pressure term library The stress-related word clusters include: exam stress, academic burnout, dormitory conflict, relationship problems, social anxiety, club stress, employment anxiety, economic stress, health concerns, family stress, future uncertainty, and other events. Based on the event ontology vector matrix Calculate the event query vector The specific calculation formula is as follows: ； in, [CLS] is the output after processing by the stress term cluster embedding and domain adaptation adapter; T is the transpose; The event projection matrix; Calculate event attention score The specific calculation is as follows: ； in, The dot product similarity between the event query vector and the j-th event ontology vector; This is a scaling factor to prevent the softmax gradient from vanishing due to excessively large dot product values; This represents the probability of relevance between the current text and the j-th stress word cluster, with the sum of event attention scores being 1; Calculate the weighted event vector The specific calculation is as follows: ； in, The event attention score; The event entity vector; Constructing a dimensional ontology vector matrix The specific calculation is as follows: ； Each column vector This corresponds to the seven psychological resilience dimensions of PRI, including goal focus, emotional control, positive cognition, job confidence, resilience to setbacks, interpersonal harmony, and campus adaptation. Calculate the dimension query vector The specific calculation formula is as follows: ； in, The dimension projection matrix; Calculate the attention score of the dimension The specific calculation is as follows: ； in, This represents the dot product similarity between the dimension query vector and the i-th dimension ontology vector; The attention weights for the seven dimensions of psychological resilience form a probability distribution, and the sum of the attention scores for the seven dimensions is 1. Calculate the weighted dimensional vector based on the dimensional attention score and the dimensional ontology vector. The specific calculation is as follows: ； The output after processing by the pressure term embedding and domain adaptation adapter Weighted event vector and weighted dimension vector splicing and fusion : ； in, This represents the text semantic features output by the UniStressBERT model; S3. Standardize individual baselines based on physiological and behavioral data to obtain a standardization function, and then calculate standardized data using the standardization function. The standardized function is calculated as follows: ； ； ； in, The mean baseline; The baseline is the standard deviation; This indicates the sensitivity of the control baseline to recent observations; These are observed values, represented as heart rate (HR), heart rate variability (RMSSD), learning interruption frequency (break), or negative emotion score. The data represents the mean values ​​up to time t-1. The standard deviation data is for time t-1 and earlier. S4. Based on text sentiment scores and standardized data, risk monitoring is performed using the SRT algorithm to obtain and determine the risk monitoring results: When the risk monitoring result is greater than or equal to the first preset threshold, an assessment is triggered. If the assessment result meets the preset conditions, the process proceeds to step S5; otherwise, it is handled manually. When the risk monitoring result is greater than or equal to the second preset threshold and less than the first preset threshold, local light intervention is performed. When the risk monitoring result is less than the second preset threshold, the risk will continue to be monitored and the process will jump to S5. S5. Encode the scene data to obtain scene-coded data; calculate the latent state of psychological resilience based on text semantic features, standardized data, and scene coding; calculate the PRI score of the psychological resilience dimension based on the latent state of psychological resilience; determine the risk level based on the PRI score and execute corresponding intervention actions.

2. The AI-based real-time intervention method for psychological resilience in university students according to claim 1, characterized in that, In step S2, the pressure term cluster bias is specifically as follows: Building a dynamic stress term library The initial vocabulary includes words related to campus stress, categorized by event type into three dimensions: academic, interpersonal, and life. Each dimension contains multiple clusters of stress words. The academic dimension includes clusters of exam stress and academic burnout; the interpersonal dimension includes clusters of dormitory conflict, romantic relationship problems, social anxiety, and club stress; and the life dimension includes clusters of employment anxiety, economic stress, health concerns, family stress, future uncertainty, and other events. During the federated learning process, each edge node can report newly discovered stress words, which are then voted on globally and updated. Supports monthly expansion; In the input layer, the input data is segmented and then matched. The token is used to embed its original words and add bias, specifically: ； in, These are pre-trained embedding vectors; To determine the injection intensity, the injection intensity is controlled and dynamically adjusted by federated learning based on the intervention results. The bias vector is shared by each stress word cluster.

3. The AI-based real-time intervention method for psychological resilience in university students according to claim 1, characterized in that, In step S2, the domain adapter... The specific calculation formula is as follows: ； in, This represents the hidden state of the l-th layer Transformer. ; This is the downprojection matrix, which compresses the 768-dimensional matrix into a low-rank r-dimensional matrix; This is the upward projection matrix, which is used to restore r dimensions to 768 dimensions; The activation function is a Gaussian error linear unit.

4. The AI-based real-time intervention method for psychological resilience in university students according to claim 1, characterized in that, In step S4, the SRT algorithm is specifically as follows: ； in, For risk monitoring results; Weights for multimodal data; Score the sentiment of the text. Indexed by 8 categories of emotions; , and It is obtained from the standardized function, where, This is the standardized value of resting heart rate; This represents the standardized value for heart rate variability. To learn the normalized value of the interrupt frequency; For scene indication functions; when When the edge layer depth assessment is triggered, the assessment includes multimodal data fusion to construct observation vectors and UniStressBERT model complete inference to predict the seven-dimensional components of PRI and event categories. If the assessment result meets the preset conditions, proceed to the next step; otherwise, it is processed manually. when At the same time, local light interventions are implemented; the content of local light interventions includes sending psychological survey questionnaires, sending soothing light music, and sending humorous copy; when At that time, maintain the monitoring status and proceed to the next step; in, and It is a hysteresis dual threshold. The first preset threshold, This is the second preset threshold.

5. The AI-based real-time intervention method for psychological resilience in university students according to claim 1, characterized in that, In step S5, the hidden state of psychological resilience specifically refers to: Six types of real-time data are concatenated into an observation vector. The specific calculation formula is as follows: ； in, This represents the text semantic features output by the UniStressBERT model; Represents scene-coded data; This is the standardized value of resting heart rate; This represents the standardized value for heart rate variability. To learn the normalized value of the interrupt frequency; Calculate the state transition equation: ； in, Indicates the intervention effect value; For process noise, Represents a normal distribution; The observation model uses a 2-layer MLP, specifically: ； Where ReLU is the activation function. For the first layer MLP, For the second layer MLP, For observation vectors; The latent state of psychological resilience is calculated based on the state transition equation and the observation model. The specific calculation process is as follows: ； in, Indicates the reliability of the observation; This represents the prior prediction of the state at time t based on all data up to and including time t-1.

6. The AI-based real-time intervention method for psychological resilience in university students according to claim 1, characterized in that, In step S5, the PRI score is specifically as follows: PRI is decomposed into seven psychological resilience dimensions. These include vectors for goal focus, emotional control, positive cognition, employment confidence, setback recovery, interpersonal coordination, and campus adaptability. The target focus vector is calculated as follows: ； in, and For feature weights; This function indicates the exam week. This represents the standardized value for late nights; To learn the normalized value of the interrupt frequency; The emotion control degree vector is calculated as follows: ； in, and For feature weights; It is a set of three types of negative sentiment indices; Assign a sentiment score to the text; This is the standardized value of resting heart rate; The positive perception vector is calculated as follows: ； in, and For feature weights, Score the emotional tone of the text as joyful; The employment confidence vector is calculated as follows: ； in, and For feature weights, Score the text's sentiment level as anxious. This is a function indicating the work scenario; The frustration recovery vector is calculated as follows: ； in, and For feature weights, For viewing the window, Let PRI score be the score at time t. for PRI score at any given moment; This represents the standardized value for heart rate variability. The interpersonal coordination vector is calculated as follows: ； in, and For feature weights, Score the emotional intensity of the angry text. The total number of interpersonal conflicts; The campus fitness vector is calculated as follows: ； in, and For feature weights, Absence rate; The specific calculation of PRI is as follows: ； Calibrate PRI scores using latent states of psychological resilience: ； If the equation does not hold true, adjust the weights of each dimension proportionally. Or correct To make it finally aligned When data for a certain dimension is missing, the vector for that dimension is directly assigned a value. Its weight Automatically set to 0.

7. The AI-based real-time intervention method for psychological resilience in university students according to claim 1, characterized in that, The multimodal data uses the national cryptographic SM4-CBC and HMAC algorithms to encrypt and store sensitive data, and the key is rotated within a preset time period; sensitive fields in the multimodal data are customized and replaced using dynamic desensitization technology; a federated learning architecture is adopted, and only the encrypted gradient parameters are uploaded for global updates, without changing the original multimodal data.

8. A real-time AI-based psychological resilience intervention system for university students, characterized in that: The method for implementing the AI-based real-time intervention method for psychological resilience targeting university students as described in any one of claims 1 to 7 includes a data acquisition module, an AI decision-making module, a privacy protection module, and a psychological resilience intervention module. The data acquisition module collects multimodal data, including text data, physiological data, behavioral data, and scene data. AI Decision Module: Constructing the UniStressBERT model: Injecting stress term cluster bias into the input layer, inserting a domain adapter after each Transformer layer, and adding soft ontology attention to the output layer; Text data is input into the UniStressBERT model to obtain text semantic features and text sentiment scores; Physiological and behavioral data are standardized to individual baselines to obtain a standardization function, and standardized data are calculated using the standardization function. Privacy protection module: Based on text sentiment scores and standardized data, risk monitoring is performed using the SRT algorithm to obtain risk monitoring results and make judgments. When the risk monitoring result is greater than or equal to the first preset threshold, a deep assessment is triggered; When the risk monitoring result is greater than or equal to the second preset threshold and less than the first preset threshold, local light intervention is performed. When the risk monitoring result is less than the second preset threshold, the risk will continue to be monitored and the next step will be taken. The psychological resilience intervention module encodes the scene data to obtain scene-encoded data; calculates the latent state of psychological resilience based on text semantic features, standardized data, and scene encoding; calculates the PRI score of the psychological resilience dimension based on the latent state of psychological resilience; and determines the risk level based on the PRI score and executes corresponding intervention actions.

Images