Early warning and intervention system for mental health of teenagers based on multi-modal ai analysis

Abstract

The present application relates to the technical field of computer information processing and artificial intelligence application, and discloses a youth mental health early warning and intervention system based on multi-modal AI analysis, which comprises a campus heart education module, a supervision and overall planning module, an AI operation module and a hierarchical guardianship module, and synchronously collects scale data and audio and video streams, and after time stamp alignment and encryption, submits the data to the AI operation module to extract acoustic, facial light flow displacement and micro-vibration features, and combines with the collection confidence to adaptively allocate weights, and outputs a multi-dimensional risk total index; in the intervention link, the system integrates a large language model to automatically output an interview abstract, realizes closed-loop tracking of the remaining risk value of the work order by introducing an intervention attenuation time penalty model, and the local terminal relies on a graph convolution and a time sequence network to output a regional risk trend prediction tensor. The present application overcomes the subjective limitations of single scale evaluation, and constructs a multi-level collaborative management and control architecture covering precise multi-modal early warning, artificial intervention quantitative evaluation and regional macro prediction.

Description

Technical Field

[0001] This invention relates to the fields of computer information processing and artificial intelligence application technology, specifically to an early warning and intervention system for adolescent mental health based on multimodal AI analysis. Background Technology

[0002] Currently, the mental health early warning systems widely deployed in the education system mainly rely on traditional standardized test scales as the core data source. This single data collection mode heavily depends on the subjective willingness of the test subjects to answer questions. In practical applications, it is often interfered with by the test subjects' defensive mentality or deliberate concealment behavior, causing the risk assessment results to deviate from the true situation. At the same time, the early warning judgment logic of the existing systems is mostly based on static and fixed parameter standards, ignoring the objective differences in psychological baselines between different regional environments and different age groups, making it difficult to adapt to the complex assessment needs in real-world scenarios.

[0003] In the follow-up manual intervention stage for high-risk individuals, the current system design also has obvious limitations. The existing business process usually requires front-line staff to manually enter unstructured interview texts, resulting in low data processing efficiency. More importantly, the existing architecture lacks a mathematical model for dynamically quantifying the effects of manual intervention. The system cannot objectively calculate the continuity of historical intervention behaviors in the present based on the time span, which means that the downgrading and closed-loop operation of early warning work orders lacks continuous numerical tracking support, causing a break in the subsequent psychological monitoring and management process.

[0004] Furthermore, from the perspective of higher-level regional management departments, the data application of existing systems is often limited to the simple aggregation of micro-level individual data. The current architecture has not yet introduced spatial correlation calculation and temporal dependency modeling for the evolution of risk status of various schools under its jurisdiction. As a result, the management side cannot obtain quantitative predictive indicators for the overall development trend. This technical deficiency makes it difficult for education management departments to perceive the aggregation trend of psychological risks in the region in advance, which restricts the early planning and precise allocation of mental health intervention resources at the macro level. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an early warning and intervention system for adolescent mental health based on multimodal AI analysis. This system solves the problems in existing mental health management systems, such as reliance on a single subjective scale leading to distorted assessments, lack of a quantitative closed-loop tracking mechanism for intervention status, and difficulty in predicting regional macro-risks.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multimodal AI analysis-based early warning and intervention system for adolescent mental health, comprising: The campus mental health education module is used to receive standardized psychological scales and deploy test tasks on local terminals, and simultaneously collect audio and video data of the test subjects during the test or interview process; The supervision and coordination module is used to configure scale factors and set risk judgment thresholds at different levels, allocate system access permissions for management terminals at all levels, and receive aggregated data uploaded by nodes in subordinate areas. The AI ​​computing module is used to receive audio and video stream data and scale answer data transmitted from the campus mental health education module. It extracts acoustic emotional features, facial local features, physiological tremor features and scale behavioral features respectively. By assigning different confidence weights, it performs calculations and outputs a multi-dimensional risk index and corresponding risk warning level. The campus monitoring module is used to obtain the warning level calculated by the AI ​​computing module, filter records according to the severity of the warning level, generate alarm commands and push them to the set campus-level management terminals. The local monitoring module is used to receive global early warning data output by the AI ​​computing module in parallel, and to perform cross-level converged data synchronization storage and interface view display.

[0007] Through the modular deployment described above, this system achieves the fusion collection and evaluation of multi-source subjective and objective data at the underlying level, and establishes a structured hierarchical early warning data flow system on this basis.

[0008] Furthermore, the campus mental health education module is equipped with a clock frequency synchronization controller. The clock frequency synchronization controller uses the underlying system clock of the local operating system as the absolute time reference and forcibly writes a unified timestamp tag to the synchronously started continuous audio data stream and continuous video image stream. The campus mental health education module uses memory address bus fusion splicing logic to combine and assemble continuous audio data streams and continuous video image streams within the same sliding time window into standardized multimodal data packets; Before performing cascading splicing, an identifier frame containing data type and length information is pushed into the header of each type of data stream; The encrypted multimodal data packet set is generated by cascading and splicing a set of continuous video image frame sequences with synchronization timestamp tags, a set of continuous audio acoustic sampling data with synchronization timestamp tags, and structured scale response feedback data. It is then generated by applying a symmetric encryption algorithm engine in conjunction with the random symmetric encryption key sequence generated by the system during the handshake phase.

[0009] The above structure ensures strict alignment of multi-source audio and video streams in the time dimension by establishing an absolute time reference, and ensures the security of the transmitted data set by using encrypted fusion processing.

[0010] Furthermore, the AI ​​computing module contains various AI computing sub-modules; each AI computing sub-module includes an AI speech recognition sub-module and a micro-expression analysis sub-module. The AI ​​speech recognition submodule separates the audio stream, runs an endpoint detection algorithm to determine whether there is valid speech activity within the current sliding time window, and runs an acoustic model to perform multilingual speech-to-text transcription if there is valid speech. Simultaneously, it extracts deep acoustic feature sequences such as Mel spectrum and inputs them into the acoustic computing network to output speech emotion feature values. The micro-expression analysis submodule parses video frame sequences, locates key facial feature points, applies optical flow algorithms to calculate the spatiotemporal displacement features of facial muscles, and inputs them into a deep learning network to output abnormal micro-expression probability values.

[0011] Based on this, the system calculates the probability distribution using acoustic feature sequences and facial displacement features, thus forming an objective evaluation index for emotional state.

[0012] Furthermore, each AI computing submodule also includes a video image micro-vibration analysis submodule; the video image micro-vibration analysis submodule reads the video frame sequence, calculates the structural similarity index between adjacent image frames, and when the structural similarity index is higher than the set dynamic threshold to exclude large-scale macroscopic motion interference, it executes a sub-pixel-level spatial frequency domain amplification algorithm to extract the physiological tremor features of the head and neck in a preset frequency band, compares them with individual baseline physiological data, and outputs micro-vibration abnormal feature values; In cases where individual baseline physiological data is lacking, the standard norms of the same age group are dynamically called as the comparison benchmark to calculate the difference. After calculating the difference, the difference is mapped to an anomaly probability value in the range of zero to one hundred through activation function or normalization algorithm.

[0013] The above design utilizes a spatial frequency domain amplification algorithm to obtain small-amplitude physiological tremor characteristics, increases the detection dimension of non-contact physiological indicators, and provides a norm comparison alternative path when individual baselines are missing.

[0014] Furthermore, the AI ​​computing module extracts the audio signal-to-noise ratio and the average pixel brightness gradient of the video as real-time confidence metrics, and adaptively adjusts the weight ratio of each modal feature in the final calculation to output a multi-dimensional risk index. The AI ​​computing module calculates the product of the dynamic confidence of each modality and the corresponding baseline weight, and performs global normalization on the product results of all modalities before performing the dot product operation. The continuous real number of the multidimensional risk index is obtained by multiplying the confidence scores of signal acquisition based on signal-to-noise ratio or brightness characteristics in real time to the range of zero to one hundred for different data modes, and then summing the product of each score together with the baseline feature weight coefficients set in advance for each data mode and the feature dimension risk judgment scalar values ​​extracted and normalized to the range of zero to one hundred by independent network branches.

[0015] By extracting signal-to-noise ratio and brightness gradient adaptively adjusting feature weights, the system can reduce the calculation bias of the multi-dimensional risk index caused by data distortion due to environmental interference.

[0016] Furthermore, the supervision and coordination module supports the customization of multi-dimensional risk assessment benchmark physical lines for users in different jurisdictions and age groups, and dynamically calculates and issues a global risk trigger threshold rule set in combination with regional baseline norms. The corresponding global risk trigger threshold constants generated for different warning levels are obtained through the following mathematical logic calculation: The normalized weight coefficient set by the management terminal for the independent feature dimension is multiplied by the independent basic early warning threshold set for the corresponding feature dimension. The product results of all independent feature dimensions are then linearly weighted and summed. Finally, the baseline compensation deviation parameter is superimposed. The baseline compensation bias parameter is obtained by extracting historical norm data from the past calendar year within the region and calculating the standard deviation mapping of the normal distribution of the historical norm data.

[0017] Based on the historical norm distribution characteristics of the region, the system can dynamically and adaptively adjust the multi-dimensional risk warning threshold.

[0018] Furthermore, the campus monitoring module has a built-in work order flow engine based on a state machine; the work order flow engine performs continuous real-valued evaluations based on the remaining risk assessment value of the current warning work order to determine whether to downgrade or close the loop in the warning status. The remaining risk assessment value is a continuous real number calculated through the following process: extract the preset single human intervention behavior benchmark weighting coefficient and the effectiveness score scalar of the current intervention record, perform multiplication to construct the basic intervention item; extract the risk rebound penalty constant set based on the Ebbinghaus forgetting curve empirical data and the time span quantified value between the time node of the intervention record and the current risk reassessment time node, perform multiplication to construct the time penalty item; The calculation result of a single intervention effect is obtained by subtracting the time penalty from the basic intervention item; the initial multi-dimensional risk index assigned when the work order is initialized is extracted, and the sum of the calculation results of the single intervention effect corresponding to all historical cumulative intervention times is deducted.

[0019] The aforementioned workflow engine transforms the decay process of the effectiveness of manual intervention over time into a time penalty term calculation formula, enabling numerical tracking and closed-loop management of the remaining risk of early warning work orders.

[0020] Furthermore, the system integrates an LLM (Large Language Model) submodule. When performing follow-up interview tasks on the campus terminal, the AI ​​speech recognition submodule continuously transcribes the interview dialogue speech, while the LLM submodule reads the transcribed dialogue text and historical alarm work order data, performs natural language parsing and intent extraction, and guides the model to automatically output structured interview summary records and intervention suggestion forms by injecting preset prompt word templates containing target data format constraints into the LLM submodule.

[0021] This allows for the automatic parsing and transformation of unstructured voice interview content into structured data forms using a large language model, thereby improving the efficiency of interview recording.

[0022] Furthermore, the local monitoring module has a built-in macro-indicator aggregation engine based on a distributed stream processing framework. The macroeconomic indicator aggregation engine performs cross-level indicator fusion calculations based on a preset system and outputs a continuous real number of risk situation assessment index for a specified region. The continuous real number of the risk situation assessment index for a designated area is obtained through the following logic: extract the average multi-dimensional total risk index of the current active status of independent schools and multiply it by the normalized weight ratio of independent schools in the regional system. Then, perform cumulative summation on the total number of registered independent schools under the jurisdiction and multiply it by the risk base weight coefficient initialized by the system backend to form the first summation term. Extract the cumulative number of high-risk intervention work orders with overdue status within the jurisdiction, divide it by the total number of historical intervention work orders, and map it to a percentage system. Then, multiply the number of these numbers by the intervention lag penalty coefficient initialized in the system backend to form the second summation term. Finally, perform an overlay calculation operation on the first and second summation terms.

[0023] The aforementioned macroeconomic indicator aggregation engine performs a mathematical superposition operation on the regional risk benchmark variable and the work order intervention lag penalty coefficient to quantitatively evaluate the overall intervention management execution effectiveness of a specified region.

[0024] Furthermore, the local monitoring module integrates a risk trend prediction engine based on a spatiotemporal graph convolutional neural network; The risk trend prediction engine abstracts each school in the jurisdiction into an independent node of a graph topology and constructs an edge adjacency matrix based on the reciprocal of the geographical travel distance between schools. The risk trend prediction engine receives historical situation assessment indices for each node and associated auxiliary indicators, and generates a three-dimensional input sequence feature matrix with dimensions of total number of nodes multiplied by time step multiplied by number of features. The network architecture, which includes graph convolutional network layers and gated recurrent unit layers, receives a 3D input sequence feature matrix. The graph convolutional network layers use the Laplacian matrix to aggregate the spatial feature distribution information of adjacent nodes. The gated recurrent unit layers perform temporal dependency modeling on the extracted spatial feature sequences to capture the time decay law. At the end of the network, a fully connected layer outputs a prediction tensor that represents the probability distribution of risk outbreaks in each school in the jurisdiction in the next quarter.

[0025] This neural network architecture, based on spatial graph topology and temporal dependency modeling, enables quantitative prediction of the probability of future risk evolution of nodes in the subordinate region.

[0026] This invention provides an early warning and intervention system for adolescent mental health based on multimodal AI analysis. It has the following beneficial effects: 1. This invention improves the accuracy of early warning results and the applicability and security of the system in complex scenarios. The system abandons the subjective limitations of single-scale assessment, integrates behavioral data from scales with multi-dimensional objective features extracted from audio and video streams, such as voice emotion, facial optical flow displacement, and physiological tremors. It also combines real-time signal-to-noise ratio and brightness gradient adaptive allocation of reliability weights to achieve multimodal cross-validation, thereby improving the accuracy of early warning. At the same time, the system supports dynamically issuing risk trigger thresholds based on historical baseline norms of different jurisdictions, which can flexibly adapt to the differentiated assessment needs of primary and secondary schools and vocational schools. In addition, the underlying multimodal data is forcibly written with a unified timestamp and fused using a symmetric encryption algorithm engine during splicing, ensuring the security of sensitive data of adolescents during collection and transmission.

[0027] 2. This invention significantly improves the efficiency of subsequent interventions by simplifying the manual intervention process and realizing dynamic quantitative closed-loop management of intervention work orders. In the actual interview stage, the system's built-in LLM large language model submodule reads the transcribed dialogue text and historical work order data. Through natural language parsing and intent extraction, it automatically outputs structured interview summaries and intervention suggestion forms, reducing the manual writing burden on grassroots staff. At the same time, the work order circulation engine combines the preset intervention benchmark weighting coefficient with the time penalty item set based on the Ebbinghaus forgetting curve, abstracting the decay process of intervention effect over time into a calculable residual risk assessment value. This approach transforms the traditional case closure judgment that relies on human experience into a numerical continuous tracking mechanism, ensuring a complete closed loop in the psychological crisis intervention process.

[0028] 3. This invention enhances the data coordination and macro-prediction decision-making capabilities of regional management departments by constructing a hierarchical and interconnected management architecture. After receiving global early warning data in parallel, the local monitoring module uses a macro-indicator aggregation engine to superimpose the regional average risk index with the intervention lag penalty coefficient of high-risk work orders, objectively quantifying the crisis management execution efficiency of the subordinate areas. At the same time, the system-integrated risk trend prediction engine abstracts each school in the jurisdiction as a topological node, and uses graph convolutional network layers and gated recurrent unit layers to jointly model the spatial feature distribution and time decay law, outputting a prediction tensor representing the future risk outbreak probability of each node. This mechanism enables management departments to intuitively obtain the spatial evolution trend of risks, providing data support for issuing management instructions in advance and accurately allocating educational resources. Attached Figure Description

[0029] Figure 1 This is a diagram of the module architecture of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a diagram illustrating the data acquisition and interaction logic architecture of the present invention. Figure 4 This is a logical architecture diagram of the supervisory and overall authority and configuration of the present invention; Figure 5 This is a diagram showing the internal logic topology and data flow of the AI ​​computing module of the present invention; Figure 6 This is a topology diagram of the interaction and intervention flow logic of the campus monitoring module of the present invention; Figure 7 This is the logical topology diagram of the macro-situation analysis and spatiotemporal prediction of the local monitoring module of the present invention; Figure 8 This is a line graph showing the dynamic allocation of multimodal feature fusion weights according to the present invention. Figure 9 This is a curve showing the decay and convergence of the residual risk assessment value in this invention. Figure 10 This is a three-dimensional heat map showing the spatiotemporal evolution of regional risk situation according to the present invention.

[0030] Among them, 10 is the Campus Mental Health Education Module; 20 is the Supervision and Coordination Module; 30 is the AI ​​Computing Module; 40 is the Campus Monitoring Module; and 50 is the Local Monitoring Module. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0032] See attached document Figure 1 The present invention provides a multimodal AI-based early warning system for adolescent mental health, which may include: a campus mental health education module 10, a supervision and coordination module 20, an AI computing module 30, a campus monitoring module 40, and a local monitoring module 50.

[0033] The Campus Mental Health Education Module 10 is used to receive standardized psychological scales and deploy test tasks on the local terminal, and simultaneously collect audio and video data of the test subjects during the test or interview process.

[0034] The supervision and coordination module 20 is used to configure scale factors and set risk judgment thresholds at different levels, allocate system access permissions to management terminals at all levels, and receive aggregated data uploaded by nodes in subordinate areas.

[0035] AI computing module 30 is used to receive audio and video data and scale answer data transmitted by campus mental health education module 10, extract acoustic emotional features, facial local features, physiological tremor features and scale behavioral features, respectively, and perform calculations by assigning different confidence weights to output a multi-dimensional risk index and corresponding risk warning level.

[0036] The campus monitoring module 40 is used to obtain the warning level calculated by the AI ​​computing module 30, filter records according to the severity of the warning level, generate alarm commands and push them to the set campus-level management terminal.

[0037] The local monitoring module 50 is used to receive global early warning data output by the AI ​​computing module 30 in parallel, and to perform cross-level data synchronization storage and interface view display.

[0038] See attached document Figure 2 This invention provides a method for early warning and intervention of adolescent mental health based on multimodal AI analysis. This method is driven by the system's underlying control logic to coordinate the execution of the aforementioned modules, achieving a closed-loop process from data collection to intervention suggestion generation. The specific implementation steps are as follows: S1, the campus mental health education module 10 receives the assessment instructions issued by the supervision and coordination module 20, activates the standardized psychological scale test interface of the local terminal, and simultaneously turns on the audio and video sensor configured on the local terminal. While acquiring the user's scale answer option data, the sensor synchronously records the continuous video stream and audio stream of the test process. The campus mental health education module 10 encapsulates the heterogeneous data and transmits it to the AI ​​computing module 30.

[0039] S2, after receiving the data packet, the AI ​​computing module 30 starts the various AI computing sub-modules contained within it to process the input data in parallel. The AI ​​speech recognition sub-module separates the audio stream, first runs the endpoint detection algorithm to determine whether there is valid speech activity within the current sliding time window. If there is valid speech, it runs the acoustic model to perform multilingual speech to text transcription operation, and simultaneously extracts deep acoustic feature sequences such as Mel spectrum and inputs them into the acoustic computing network to output speech emotion feature values. The micro-expression analysis sub-module parses the video frame sequence, locates key facial feature points, applies the optical flow algorithm to calculate the spatiotemporal displacement features of facial muscles, inputs them into the deep learning network, and outputs abnormal micro-expression probability values. The video image micro-vibration analysis submodule reads the video frame sequence and calculates the structural similarity index between adjacent image frames. When the structural similarity index is higher than the set dynamic threshold to exclude large-scale macroscopic motion interference, a sub-pixel-level spatial frequency domain amplification algorithm is executed to extract the physiological tremor features of the head and neck in a preset frequency band. The data is compared with the individual baseline physiological data, and the micro-vibration abnormality feature value is output. Finally, the AI ​​calculation module 30 performs structured and normalized conversion on the collected scale answer data and outputs the scale behavioral feature value.

[0040] S3, the risk scoring submodule acquires four sets of independent feature values. The system evaluates the data acquisition completeness of the audio and video streams based on the calculated audio-video signal-to-noise ratio, frame rate stability, and effective data duration ratio, and assigns confidence coefficients for each dimension accordingly. The risk scoring submodule adjusts the preset basic weights according to the confidence coefficients to generate a real-time dynamic weight distribution. The four sets of feature values ​​and their corresponding dynamic weights are multiplied by a dot product to calculate the total risk index of mental health in multiple dimensions. The system compares the total risk index of multiple dimensions with the preset hierarchical classification thresholds and assigns corresponding warning level labels to the total risk index of multiple dimensions.

[0041] S4, the campus monitoring module 40 and the local monitoring module 50 simultaneously obtain the assessment results containing the warning level label. The system retrieves the global risk trigger threshold rule set dynamically calculated and issued by the supervision and coordination module 20 in combination with the regional baseline norm. When the total multi-dimensional risk index reaches the lower limit of the graded alarm trigger set by the rule set, the campus monitoring module 40 locks the tested object and generates an alarm work order containing details of abnormal characteristics in each dimension. The alarm work order is pushed to the corresponding campus-level management terminal according to the preset permission rules. The system instructs the management terminal to establish an intervention follow-up task.

[0042] S5, when performing follow-up interview tasks on the campus terminal, the AI ​​speech recognition submodule continuously transcribes the interview dialogue speech, the LLM large language model submodule reads the transcribed dialogue text and historical alarm work order data, performs natural language parsing and intent extraction, and guides the model to automatically output structured interview summary records and intervention suggestion forms by injecting preset prompt word templates containing target data format constraints such as JSON. The supervision and coordination module 20 receives the structured form data and performs persistent storage, completing the overall system process from risk warning to auxiliary intervention.

[0043] See attached document Figure 3 The campus mental health education module 10, as the front-end data access hub of the system, is configured and runs in the local area network environment. It receives the standardized psychological scale configuration file issued by the supervision and coordination module 20 through the secure socket layer protocol interface. In this embodiment, the configuration file contains the question bank test content and state jump rules encapsulated in the extensible markup language format. The graphics rendering engine built into the campus mental health education module 10 parses the extensible markup language and dynamically renders and generates a visual answer interaction interface on the display device of the local user terminal. The interaction interface continuously listens for click events or screen touch events of peripheral hardware, records the corresponding scale answer options in a structured encoding, and temporarily stores them in the local high-speed cache for synchronous packaging and scheduling.

[0044] At the same moment the questionnaire interaction interface is presented on the local user terminal, the underlying hardware abstraction layer of the campus mental health module 10 concurrently calls the device application programming interface to wake up the inherent digital microphone array and digital camera of the terminal device. Since the subsequent cross-modal fusion calculation depends on strict time dimension constraints, as a preferred method, the campus mental health module 10 is equipped with a clock frequency synchronization controller. The clock frequency synchronization controller uses the underlying system clock of the local operating system as the absolute time reference and forces a unified timestamp label to be written to the synchronously started continuous audio data stream and continuous video image stream.

[0045] In the specific sampling process, the audio sensor channel continuously samples using pulse code modulation, while the video image sensor channel continuously captures a sequence of continuous image frames containing the head and neck region of the subject at a set fixed frame rate. The multi-channel hardware collaborative acquisition process is executed in the background daemon of the operating system.

[0046] The structured questionnaire response data acquired in real time, together with the audio sampling stream data and video frame stream data with unified timestamp labels, constitute the front-end multi-source heterogeneous dataset. The campus mental health education module 10 uses memory address bus fusion splicing logic to combine and assemble heterogeneous data streams within the same sliding time window into standardized multimodal data packets.

[0047] To ensure the identifiability of multi-source data in subsequent parsing, the system pushes an identifier frame containing data type and length information into the header of each type of data stream before performing concatenation. This defines the multimodal data packet logical structure built at the system's underlying level as conforming to a specified combinatorial mathematical parsing expression. ; In the formula, This represents the set of encrypted multimodal data packets generated in real time within a specific sliding time window period; it also represents the structured scale response data generated and extracted within the corresponding time window period. Represents a set of consecutive video image frames with accompanying synchronization timestamp tags; Represents a set of continuous audio acoustic sampling data with accompanying synchronization timestamp tags; symbol This represents the cascading and splicing of heterogeneous media data blocks in the computer's memory address space according to their corresponding identifier frames; symbol This indicates the cryptographic fusion process performed using a symmetric encryption algorithm engine; This indicates that the system generates a random symmetric encryption key sequence during the handshake phase based on the Advanced Encryption Standard (AES) algorithm. The actual value of this key is a 256-bit fixed-length key generated by a pseudo-random number generator according to the AES-256 standard. The system uses this key to encrypt and protect data packets through corresponding encryption operation modes, such as counter mode.

[0048] For the asymmetric key negotiation process required by symmetric encryption algorithms, such as exchanging AES keys via RSA algorithm and the specific symmetric encryption operation execution process, the standard cryptographic dynamic link library of the existing computer operating system can be directly called to implement it. After secure encapsulation and digital signature generated based on hash message authentication code or asymmetric private key and verified by the corresponding public key, the set of verified and legitimate encrypted multimodal data packets is pushed to the data receiving buffer pool of AI computing module 30 through a long connection channel established by broadband network data bus to wait for deep feature parsing.

[0049] See attached document Figure 4 The monitoring and coordination module 20 is deployed on a cloud server or regional core data center. As the control center of the entire architecture, it provides a visual management and configuration backend. Managers can use this module to complete the initialization and writing of the underlying business parameters of the system. In this embodiment, the supervision and coordination module 20 has a built-in scale parsing engine. This engine receives standardized psychological test documents imported by managers and performs structured extraction. Through regular expression matching and natural language segmentation technology, it converts the unstructured question stem text, option scores and corresponding measurement factor mapping relationships into a standard key-value pair data table structure supported by a relational database, thereby realizing the standardized conversion of heterogeneous scale formats to a unified data structure within the system.

[0050] In the scale factor configuration and threshold calibration stage, the supervision and coordination module 20 supports the customization of multi-dimensional risk judgment benchmark physical lines for users in different jurisdictions and age groups. The system provides a graphical sliding adjustment control to receive the risk feature weights and alarm trigger lower limits set by managers for each dimension. The set parameter values ​​are serialized by the underlying logic processing unit into a standard cross-platform threshold rule set. This threshold rule set is asynchronously sent to the rule parsing library of the AI ​​computing module 30 through an internal message queue, serving as an objective benchmark for subsequent calculation of the multi-dimensional risk total index and comparison of risk levels. The global risk trigger threshold configured by the system is defined to follow a specific linear weighted sum mathematical relationship.

[0051] ; In the formula, The monitoring and coordination module 20 generates corresponding global risk trigger threshold constants for different early warning levels, such as low, medium and high levels. When calculating for each early warning level, the system calls the corresponding level configuration parameters and substitutes them into the following formula to solve independently. This represents the total number of independent risk feature dimensions that the system supports for analysis. In this embodiment, this value covers multiple processing channels, including table features and audio / video multimodal features. Indicates that the management terminal is the first The normalized weight coefficients are set for each feature dimension. The value of the coefficients ranges from 0 to 1, and the sum of the coefficients of each dimension is fixed at one. Indicates the corresponding number An independent basic warning threshold is set for each feature dimension, and the value of this threshold is set to a continuous constant from 60 to 95. This represents the baseline compensation deviation parameter. Its actual value is obtained by extracting historical norm data from the past calendar year within the region and calculating the standard deviation of its normal distribution. The actual value after mapping is limited to between -5.0 and +5.0.

[0052] To address the massive amounts of cross-level aggregated data generated during system operation, the Supervision and Coordination Module 20 employs a role-based access control model to perform fine-grained resource authentication and logical isolation. As a preferred approach, administrators configure multi-level role entities corresponding to district / county / bureau and school levels in the background. The system's underlying structure is a tree-like organizational structure directory containing district / county, school, and class nodes. Specific system function menu items are packaged and bound to the corresponding role entities along with the row-level access permissions of different organizational structure directory nodes. When the tested objects at each level initiate a data access request, the authentication interceptor built into the Supervision and Coordination Module 20 extracts the token identifier from the network request header and reverse-parses its mapped organizational structure level. Based on this, it dynamically assembles data retrieval statements with hierarchical filtering conditions in the underlying relational database query engine.

[0053] The database query engine executes the above retrieval statement and returns a set of mental health warning records that have been pruned. This pruning process is enforced by the view isolation mechanism of the database kernel layer, which restricts the terminal interface to only output sensitive business data within the corresponding physical jurisdiction. For the underlying relational data table mapping logic involved in the role-based access control model and the identity token issuance and encryption verification mechanism in the front-end and back-end interaction process, it can be directly implemented using a conventional open-source permission management framework combined with a hash algorithm component.

[0054] See attached document Figure 5 The AI ​​computing module 30 is configured as the core computing engine of the system. It mainly receives the encrypted multimodal data packet set from the data receiving buffer pool and performs decryption and unpacking operations. Under the overall planning of the internal computing power scheduling framework, it routes multi-source heterogeneous data to the corresponding feature extraction branch network.

[0055] As a preferred approach, the AI ​​computing module 30 integrates a speech processing branch based on a deep acoustic model, a facial local feature branch based on a spatiotemporal neural network, and a physiological tremor branch. Finally, the heterogeneous features are mapped to a continuous numerical space of a multi-dimensional risk index with a unified dimension through a dynamic confidence fusion device.

[0056] In the acoustic emotion feature extraction channel, the AI ​​computing module 30 receives a continuous audio sampling stream with a sampling rate of 16 kHz, performs frame-by-frame windowing preprocessing, and extracts 80-dimensional Mel spectrum features as the acoustic input matrix. This matrix is ​​fed into a deep acoustic model with a three-layer bidirectional long short-term memory network and a self-attention mechanism connection structure for sequence modeling. After the model extracts temporal context features, it reduces the dimensionality through a fully connected layer with a modified linear unit activation function. Finally, the hidden state is mapped to a speech emotion feature value representing depression or anxiety tendencies in the range of 0 to 1 through the Softmax output layer.

[0057] For a continuous video image frame sequence, the facial local feature extraction channel uses a residual network with 50 network layers to extract the coordinates of 68 key points on the face. It combines sparse optical flow methods such as the Lucas-Kanade algorithm to calculate the pixel difference in displacement of key points between adjacent image frames to construct sparse optical flow feature sequence data. This sequence data is reconstructed into a three-dimensional feature tensor containing spatial nodes and temporal dimensions, and input into a spatiotemporal three-dimensional convolutional neural network containing three three-dimensional convolutional layers and three-dimensional max pooling layers to extract dynamic spatial geometric change patterns. At the end of the network, the abnormal probability score quantification value representing the stiffness of facial micro-expressions is also output through a fully connected layer.

[0058] To address concealed physiological abnormalities, the AI ​​computing module 30 deploys a sub-pixel-level micro-vibration physiological feature extraction algorithm based on macroscopic motion determination constraints. This algorithm extracts local spatial phase information from video sequences using a high-pass spatial filter, calculates the structural similarity index of adjacent frames, and compares it with a set dynamic threshold of 0.85. If the index is lower than the threshold, it is determined that there is large-scale macroscopic motion interference, and a masking logic is triggered within this sliding window. In the static observation window that excludes macroscopic motion interference, Euler video magnification technology is applied to perform temporal bandpass filtering amplification on the minute spatial phase changes, extracting the amplitude curve reflecting the frequency fluctuation of minute tremors in the head and neck muscles and cross-comparing it with the locally cached individual baseline norm. If it is determined that there is a lack of individual historical baseline data, the standard norm of the same age group is dynamically called as the comparison benchmark. After calculating the difference, it is mapped to an abnormal probability value in the range of 0 to 1 through activation functions such as Sigmoid or Min-Max normalization algorithm, and finally outputs the micro-vibration abnormal feature value.

[0059] After aggregating multi-channel features, the AI ​​computing module 30 employs a dynamic weight allocation strategy based on the confidence level of multimodal data acquisition. It extracts the audio signal-to-noise ratio and the average pixel brightness gradient of the video as real-time confidence metrics, adaptively adjusting the weight ratio of each modality feature in the final calculation to output a multi-dimensional total risk index. The system calculates the product of the dynamic confidence level of each modality and the corresponding baseline weight, and performs global normalization on the product results of all modalities before performing the dot product operation to ensure that the sum of the dynamic weights participating in the final calculation is always equal to one. Before performing the fusion calculation, the system's internal numerical converter performs a unified linear mapping on the voice emotion feature values ​​in the 0 to 1 interval output by each deep network branch, amplifying them to a percentage scale of 0 to 100 to align with the numerical distribution space of the scale's behavioral features. The multi-dimensional total risk index fusion calculation model is defined according to the following mathematical expression.

[0060] ; In the formula, This represents a continuous real number representing the multidimensional total risk index output after cross-modal fusion calculation; This indicates the total number of data modalities participating in the current operation; and These represent the terms for the i-th and j-th respectively. Each data mode is mapped in real time to a signal acquisition confidence score in the 0 to 1 range based on signal-to-noise ratio or brightness characteristics; This represents the data modality index variable in the outer accumulation and summation operation, and its value is a positive integer. This represents the data modality index variable in the inner accumulation and summation operation, and its value is a positive integer. Indicates that the first Each data mode is assigned a baseline feature weight coefficient, and the sum of the baseline weights for each mode is fixed at one. This represents the first element extracted from each independent network branch and normalized to the interval between 0 and 100. The risk assessment scalar value for each feature dimension includes the aforementioned voice emotion feature value, micro-vibration abnormality feature value, etc.

[0061] After obtaining the multi-dimensional risk index and generating an early warning work order, in the subsequent campus follow-up interview, the system inputs the multi-dimensional risk index and the speech-to-text transcript of the interview into the LLM large language model submodule to generate an intervention plan and output it. In addition, to ensure the accuracy of front-end feature extraction, for the construction and offline training process of the aforementioned acoustic and facial deep network models, the system extracts a desensitized multimodal dataset with professional psychologist emotional state annotations and micro-expression abnormality scoring labels as training samples before deployment. Cross-entropy is used as the loss function to measure the difference between the predicted distribution and the true label. The adaptive moment estimation optimization algorithm is used to perform multiple rounds of backpropagation to update the network weight parameters under the set batch size and learning rate parameters. The batch size is set between 32 and 128, and the initial learning rate parameter is set between 0.001 and 0.005. After the loss function value on the validation set converges to a stationary state, the model is packaged and deployed online.

[0062] See attached document Figure 6 The campus monitoring module 40 is configured as the front-end interactive display and business closed-loop management hub of the system. It is deployed on the school's local server or the mobile terminal of authorized teachers. It mainly undertakes the structured intervention plan generated by the multi-dimensional risk index and large language model output by the AI ​​computing module 30.

[0063] In this embodiment, the campus monitoring module 40 has a built-in work order flow engine based on a state machine. This engine parses the received multi-dimensional risk index data packet and compares it with the alarm trigger lower limit execution matrix set by the supervision and coordination module 20 for each level. In this embodiment, the alarm trigger lower limits for low, medium and high priority early warning work orders are preset to 60 points, 75 points and 90 points respectively. Based on the comparison difference, early warning work orders with high, medium and low priority attributes are automatically generated and their initial state is marked as pending processing, thereby constructing the physical closed loop starting point for the flow from risk assessment to physical manual intervention.

[0064] In the automated distribution of early warning work orders, the campus monitoring module 40 performs routing forwarding based on the built-in organizational structure responsibility mapping table. The system extracts the grade and class identification code of the monitored object, combines it with the priority attribute of the current early warning work order, and asynchronously pushes the work order data to the terminal device cache of the corresponding homeroom teacher and school psychological counselor through the internal message queue middleware, triggering the system-level message notification mechanism at the bottom layer of the mobile terminal. As a preferred implementation method, the work order flow engine is configured to follow a specific decay convergence mathematical model for the degradation or closed-loop determination of the early warning status.

[0065] ; In the formula, This represents a continuous real number representing the remaining risk assessment value of the current early warning work order calculated by the campus monitoring module 40 at any given time point. This represents the initial multi-dimensional risk index assigned during work order initialization; This indicates the total number of independent intervention records performed for the subject; it also indicates the pre-configured weighting coefficient for a single manual intervention behavior, with a value between 0 and 1. Indicates the first The effectiveness score scalar of the intervention record is obtained by mapping the relative percentage decrease of the total score of the standardized psychological scale before and after the intervention to the range of 0 to 1. If the total score does not decrease or rebounds, the system determines that the intervention is invalid and forces the effectiveness score scalar to a value of zero. The index variable represents the number of intervention records in the cumulative summation operation, and its value is a positive integer; it represents the number of intervention records in the cumulative summation operation. The quantified value of the time span between the time point of the intervention record and the time point of the current risk reassessment; This represents the risk rebound penalty constant set based on empirical data from the Ebbinghaus forgetting curve, since it is included in the summation term of the formula via a time penalty term ( ) for the current artificial intervention item ( The internal subtraction operation is essentially used to reduce the effectiveness of the current human intervention item. The system reduces the weight of the internal deduction operation, which leads to a compensatory rebound in the remaining risk assessment value. The system is configured with a lower limit truncation function for this internal deduction operation to ensure that the minimum value of the calculation result of a single intervention is truncated to zero, preventing the risk assessment value from abnormally exceeding the initial benchmark due to the infinite extension of time. The function is configured to ensure that the remaining risk assessment value does not become an abnormal negative number.

[0066] The tested subjects who receive the early warning work order enter the intervention process record and the phased feedback information of the tested subjects through the graphical operation interface provided by the campus monitoring module 40. The operation interface adopts a responsive layout to present multimodal feature radar charts and cognitive maps. The terminal device serializes the form data entered by the user into Extensible Markup Language format through a secure connection via Hypertext Transfer Protocol and sends it back to the cloud database. The work order flow engine updates the remaining risk assessment value of the current early warning work order in real time according to the aforementioned decay convergence mathematical model. When the assessment value falls below the set safety baseline value, as a preferred method, the safety baseline value is set between 20 and 30 points, triggering the work order state machine to reverse the current work order state to closed loop and archive it to the historical record database.

[0067] For the long-term accumulated closed-loop work order data and periodic assessment results, the campus monitoring module 40 is equipped with a visualization report rendering engine to perform macro statistics. This rendering engine aggregates all student mental health data under the current school organizational structure, calculates the frequency of risk warnings in each class and the average closed-loop cycle of intervention work orders, and converts the statistical results into a multi-dimensional data aggregation cube structure. Finally, it drives the front-end open-source chart library to dynamically draw a comprehensive monitoring screen covering trend lines and distribution heatmaps. The component rendering mechanism of the underlying graphical operation interface and the asynchronous publish-subscribe mode of the message queue middleware can be directly implemented using conventional front-end frameworks and open-source message middleware.

[0068] See attached document Figure 7 The local monitoring module 50 is configured as the data aggregation center of the regional education management department. It is deployed on the government cloud data center or the independent server cluster inside the education bureau. It mainly receives the summary risk index and historical closed-loop work order records that are periodically reported by all campus monitoring modules 40 in the jurisdiction after differential privacy desensitization processing. The above-mentioned reported content constitutes the global early warning data of the system mentioned above. In this embodiment, the local monitoring module 50 has a built-in macro-indicator aggregation engine based on a distributed stream processing framework. This engine parses global early warning data and performs online analysis and processing based on administrative division codes. It extracts timestamps from structured records to calculate composite monitoring indicators, including risk warning occurrence rate and intervention work order closure rate, forming a multi-dimensional regional psychological data warehouse set.

[0069] For the quantitative assessment of the overall mental health risk situation in the region, the local monitoring module 50 performs cross-level indicator fusion calculation based on the preset system. The system extracts the total number of effective risk samples from each independent school node and combines them with the relapse frequency of previous intervention records. The aforementioned indicators are then input into the risk situation assessment model through the internal calculation unit to output a quantitative benchmark.

[0070] As a preferred implementation, the macro-indicator aggregation engine is configured to calculate the risk situation assessment index for any specified regional node according to a specific composite weighted mathematical model.

[0071] ; In the formula, It represents a continuous real number representing the risk situation assessment index for a designated area, calculated periodically by the local monitoring module 50. This indicates the total number of registered independent schools within the region; Indicates the first The average multidimensional total risk index of the current activation status of the schools; representing the first The absolute number of students currently enrolled in the school; This represents the preset baseline constant for the total number of students in the region. Dividing the two values ​​is used to calculate the normalized weight ratio of the school in the regional system. Configured as a small smoothing constant, it is used in the formula as an anti-jitter offset and placed in the denominator to avoid calculation overflow exceptions caused by division by zero. Its specific value is set to 1×10. -5 ; This indicates the cumulative number of high-risk intervention work orders that are overdue and unprocessed within the jurisdiction; This represents the total number of historical intervention work orders recorded in the regional database. The system extracts the result of dividing these two and multiplies it by one hundred, mapping its dimensions to a percentage space of 0 to 100, so as to align the numerical scale of the multi-dimensional risk index with the index fusion. This represents the risk base weighting coefficient initialized by the system backend, with a value ranging from 0.6 to 0.8. This represents the intervention lag penalty coefficient initialized by the system backend, with a value ranging from 0.2 to 0.4; and and The sum of the two values ​​is fixed at one.

[0072] To address the evolving trends of regional mental health risk characteristics in the spatiotemporal dimension, the local monitoring module 50 integrates a risk trend prediction engine based on a spatiotemporal graph convolutional neural network. The engine abstracts each school in the jurisdiction as an independent node in a graph topology and constructs an edge adjacency matrix using the reciprocal of the geographical distance between schools. The engine receives the historical situation assessment index of each node over the past twelve months and its associated auxiliary indicators, including the node's average multidimensional total risk index for the current month, the number of overdue work orders, and the number of students enrolled. It generates a three-dimensional input sequence feature matrix with dimensions of the total number of nodes multiplied by the time step and the number of features. The sequence feature matrix is ​​fed into a network architecture containing two layers of graph convolutional network and one layer of gated recurrent unit. The graph convolutional network layer uses the Laplacian matrix to aggregate the spatial feature distribution information of adjacent nodes, and the gated recurrent unit layer performs temporal dependency modeling on the extracted spatial feature sequence to capture the time decay law. At the end of the network, a fully connected layer with a Sigmoid or Softmax activation function outputs a prediction tensor representing the probability distribution of risk outbreaks in each school in the jurisdiction for the next quarter.

[0073] In the offline construction and parameter tuning stage of the spatiotemporal graph convolutional neural network model, the system extracts the historical records of multi-source heterogeneous psychological assessments and corresponding clinical diagnosis summary data of the region over the past three years as samples and true distribution labels for supervised learning. The system uses cross-entropy as the loss function to measure the numerical deviation between the model's predicted risk tensor and the actual diagnosis distribution matrix. The adaptive moment estimation optimization algorithm is used to perform iterative backpropagation within the set hyperparameter boundaries to update the weight parameters of the graph network and recurrent network nodes. The model deployment is completed after the validation set evaluation index converges to the set lower limit, which is that the cross-entropy loss value is less than the set convergence threshold, such as 0.05.

[0074] After obtaining the prediction tensor, the local monitoring module 50 is equipped with a geographic information system rendering engine to reconstruct and merge high-risk nodes with digital map layers to output a three-dimensional situational simulation heat map. The distributed computing scheduling mechanism at the bottom layer of the stream processing framework and the tile loading and rendering technology of digital map layers can be directly implemented using open-source big data processing components and mainstream geographic information frameworks.

[0075] Application Examples: Please see the appendix Figure 8 -Appendix Figure 10 To aid in understanding the technical solution of this invention, the following is an application example based on daily mental health screening and intervention in key municipal middle schools.

[0076] In the application embodiment of daily mental health screening and intervention in key municipal middle schools, the test subjects log in to the local terminal in the test room and start the standardized psychological scale test issued by the campus mental health education module 10. During the test subjects' answers, the digital camera and microphone array configured on the terminal are simultaneously awakened and continuously record facial video streams and audio streams. The clock frequency synchronization controller inside the system adds a unified timestamp to the audio stream, video stream and scale answer record. The system splices and encapsulates the above three sets of heterogeneous data and encrypts them locally using the AES-256 algorithm. Finally, the encrypted multimodal data packet is pushed to the data receiving buffer pool of the AI ​​computing module 30.

[0077] The AI ​​computing module 30 receives and decrypts encrypted data packets and initiates multi-branch parallel processing. The scale behavior feature value of the test subject is calculated to be 65 points. The AI ​​speech recognition submodule extracts the Mel spectrum features from the test subject's audio data and outputs the speech emotion feature value representing depressive tendencies. The facial micro-expression analysis submodule calculates the pixel difference value of the displacement of the test subject's facial key points using the sparse optical flow method and outputs the probability value of abnormal micro-expressions. The micro-vibration physiological feature extraction algorithm extracts the physiological tremor features of the test subject's head and neck above the norm baseline using spatial frequency domain amplification technology after excluding macroscopic motion interference and outputs the abnormal micro-vibration feature value. The system reduces the signal acquisition confidence score of the video modality based on the fact that the average pixel brightness gradient of the current video modality is lower than the preset brightness benchmark lower limit. The system simultaneously increases the signal acquisition confidence score of the audio modality with an audio signal-to-noise ratio higher than the preset signal-to-noise ratio benchmark lower limit. The system performs dot product fusion calculation on the aforementioned feature values ​​and the dynamic weights calculated based on the confidence score. Finally, the AI ​​computing module 30 outputs a continuous real number of 82 points for the test subject's multi-dimensional risk total index.

[0078] The campus monitoring module 40 receives a multi-dimensional risk index of 82 points. The campus monitoring module 40 performs a matrix comparison with the intermediate-level warning trigger limit of 75 points preset by the supervision and coordination module 20. When the comparison difference is positive, the system locks the tested object and generates an alarm work order with medium priority attributes. The system pushes the alarm work order to the terminal device cache of the homeroom teacher and school psychological counselor with preset permissions through the internal message queue middleware. In the subsequent follow-up interview tasks, the terminal microphone continuously collects audio, and the AI ​​voice recognition submodule transcribes the audio into text data. The LLM large language model submodule reads the transcribed text data and outputs a structured interview summary record and intervention suggestion form containing core risk elements. Teachers perform two manual intervention actions based on the intervention suggestion form and enter the relative percentage decrease of the total scale score on the terminal to generate an intervention effectiveness score scalar. The work order flow engine inputs the aforementioned data into the time decay convergence mathematical model to calculate the remaining risk assessment value of the current warning work order as 25 points. This remaining risk assessment value is lower than the preset safety baseline value of 30 points. The work order state machine triggers a state reversal instruction to mark the current alarm work order status as closed and archive it to the historical record database.

[0079] Closed-loop alarm work order records and multi-dimensional risk index data, after differential privacy desensitization processing, are periodically reported and aggregated to the local monitoring module 50. The macro-indicator aggregation engine built into the local monitoring module 50 extracts the risk warning occurrence rate and intervention work order closure rate based on a composite weighted mathematical model. The system calculates and updates the continuous real number of the risk status assessment index of the school node in the region by combining the absolute base of the number of students enrolled in the school. The risk trend prediction engine inputs the independent nodes of the graph topology of each school into the spatiotemporal graph convolutional neural network. The network outputs a prediction tensor that represents the distribution of the probability of risk outbreaks in each school in the jurisdiction in the next quarter.

[0080] The testing entity conducted a comparative verification test on 12,000 students from 10 affiliated schools. The test subjects were divided into three groups: control group A, which used a single structured psychological scale for computer-based assessment; control group B, which used a structured psychological scale combined with facial expression recognition technology for feature extraction; and experimental group C, which used a structured psychological scale combined with continuous audio acoustic sampling data and continuous video image frame sequences for feature extraction and used a work order flow engine for state reversal.

[0081] Table 1: Comparison of early warning accuracy and false negative rate for different test groups.

[0082] The test data recorded in Table 1 were extracted. The comprehensive early warning accuracy of experimental group C was 94.7%, which was higher than that of control group A and control group B. The high-risk false negative rate of experimental group C was 3.1%, which was lower than that of control group A and control group B. The scalar values ​​output by the micro-expression feature extraction branch and the physiological tremor feature extraction branch in the multimodal AI fusion computing architecture reduced the false negative probability data.

[0083] Table 2: Comparison of Early Warning Response Time and Intervention Timeliness among Different Test Groups Extract the test data recorded in Table 2. The average response time for early warning recorded in experimental group C was 3 minutes, which was lower than the average response time for early warning recorded in control groups A and B. The average closed-loop cycle of early warning work orders recorded in experimental group C was 5.2 days, which was lower than the average closed-loop cycle for early warning work orders recorded in control groups A and B. The effective improvement rate of stage intervention recorded in experimental group C was 91.8%, which was higher than the effective improvement rate of stage intervention recorded in control groups A and B.

[0084] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for early warning and intervention of adolescent mental health based on multimodal AI analysis, characterized in that, include: The campus mental health education module is used to receive standardized psychological scales and deploy test tasks on local terminals, and simultaneously collect audio and video data of the test subjects during the test or interview process; The supervision and coordination module is used to configure scale factors and set risk judgment thresholds at different levels, allocate system access permissions for management terminals at all levels, and receive aggregated data uploaded by nodes in subordinate areas. The AI ​​computing module is used to receive audio and video data and scale answer data transmitted by the campus mental health education module, extract acoustic emotional features, facial local features, physiological tremor features and scale behavioral features respectively, perform calculations by assigning different confidence weights, and output a multi-dimensional risk index and corresponding risk warning level. The campus monitoring module is used to obtain the warning level calculated and output by the AI ​​computing module, filter records according to the severity corresponding to the warning level, generate alarm commands and push them to the set campus-level management terminal. The local monitoring module is used to receive global early warning data output by the AI ​​computing module in parallel, and perform cross-level data synchronization storage and interface view display.

2. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 1, characterized in that, The campus mental health education module is equipped with a clock frequency synchronization controller. The clock frequency synchronization controller uses the underlying system clock of the local operating system as the absolute time reference and forcibly writes a unified timestamp tag to the synchronously started continuous audio data stream and continuous video image stream. The campus mental health education module uses memory address bus fusion splicing logic to combine and assemble the continuous audio data stream and the continuous video image stream within the same sliding time window into a standardized multimodal data packet; Before performing concatenation, an identifier frame containing data type and length information is pushed into the header of each type of data stream; the encrypted multimodal data packet set is generated by concatenating and splicing a set of continuous video image frame sequences with synchronization timestamp tags, a set of continuous audio acoustic sampling data with synchronization timestamp tags, and structured scale answer feedback data, and is generated by applying a symmetric encryption algorithm engine in combination with the random symmetric encryption key sequence generated by the system during the handshake phase.

3. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 1, characterized in that, The AI ​​computing module contains various AI computing sub-modules; each AI computing sub-module includes an AI speech recognition sub-module and a micro-expression analysis sub-module. The AI ​​speech recognition submodule separates the audio stream, runs an endpoint detection algorithm to determine whether there is valid speech activity within the current sliding time window, and runs an acoustic model to perform multilingual speech-to-text transcription if there is valid speech. Simultaneously, it extracts deep acoustic feature sequences such as Mel spectrum and inputs them into the acoustic computing network to output speech emotion feature values. The micro-expression analysis submodule parses the video frame sequence, locates key facial feature points, applies optical flow algorithm to calculate the spatiotemporal displacement features of facial muscles, and inputs them into the deep learning network to output abnormal micro-expression probability values.

4. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 3, characterized in that, Each AI computing submodule also includes a video image micro-vibration analysis submodule. The video image micro-vibration analysis submodule reads the video frame sequence, calculates the structural similarity index between adjacent image frames, and when the structural similarity index is higher than the set dynamic threshold to exclude large-scale macroscopic motion interference, it executes a sub-pixel-level spatial frequency domain amplification algorithm to extract physiological tremor features of the head and neck in a preset frequency band, compares them with individual baseline physiological data, and outputs micro-vibration abnormality feature values. In cases where baseline physiological data for the individual is lacking, the standard norm of the same age group is dynamically called as a comparison benchmark to calculate the difference. After calculating the difference, the difference is mapped to an abnormal probability value in the range of zero to one hundred through an activation function or normalization algorithm.

5. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 1, characterized in that, The AI ​​computing module extracts the audio signal-to-noise ratio and the average pixel brightness gradient of the video as real-time confidence quantification indicators, and adaptively adjusts the weight ratio of each modal feature in the final calculation to output the multi-dimensional risk index. The AI ​​computing module calculates the product of the dynamic confidence of each modality and the corresponding benchmark weight, and performs global normalization on the product results of all modalities before performing the dot product operation; The continuous real number of the multidimensional risk index is obtained by multiplying the confidence scores of signal acquisition based on signal-to-noise ratio or brightness characteristics in real time to the range of zero to one hundred for different data modes, and then performing an accumulation and summation operation on the product results of each score, combined with the benchmark feature weight coefficients set in advance for each data mode and the feature dimension risk judgment scalar values ​​extracted and normalized to the range of zero to one hundred by independent network branches.

6. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 1, characterized in that, The supervision and coordination module supports the customization of multi-dimensional risk assessment benchmark physical lines for users in different jurisdictions and age groups, and dynamically calculates and issues a global risk trigger threshold rule set in combination with regional baseline norms. The corresponding global risk trigger threshold constants generated for different warning levels are obtained through the following mathematical logic calculation: extract the normalized weight coefficient set by the management terminal for the independent feature dimension, multiply it by the independent basic warning threshold set for the corresponding feature dimension, perform linear weighted summation on the product results of all independent feature dimensions, and finally superimpose the baseline compensation deviation parameter. The baseline compensation bias parameter is obtained by extracting historical norm data from the past calendar year within the region and calculating the standard deviation mapping of the normal distribution of the historical norm data.

7. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 1, characterized in that, The campus monitoring module has a built-in work order flow engine based on a state machine; the work order flow engine performs continuous real-valued evaluations based on the remaining risk assessment value of the current warning work order to determine whether to downgrade or close the loop in the warning status. The remaining risk assessment value is a continuous real number calculated through the following purely textual abstraction process: extract the preset single manual intervention behavior benchmark weighting coefficient and the effectiveness score scalar of the current intervention record, perform multiplication operation to construct the basic intervention item; extract the risk rebound penalty constant set based on the Ebbinghaus forgetting curve empirical data and the time span quantified value between the time node of the intervention record and the current risk reassessment time node, perform multiplication operation to construct the time penalty item; The calculation result of a single intervention effect is obtained by subtracting the time penalty term from the basic intervention term; the initial multi-dimensional risk index assigned when the work order is initialized is extracted, the sum of the calculation results of the single intervention effect corresponding to all historical cumulative intervention times is deducted, and the minimum value of the calculation result of the single intervention effect is set to zero.

8. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 1, characterized in that, The system integrates an LLM (Large Language Model) submodule; When performing follow-up interview tasks on the campus terminal, the AI ​​speech recognition submodule continuously transcribes the interview dialogue speech. The LLM large language model submodule reads the transcribed dialogue text and historical alarm work order data, performs natural language parsing and intent extraction, and guides the model to automatically output structured interview summary records and intervention suggestion forms by injecting preset prompt word templates containing target data format constraints into the LLM large language model submodule.

9. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 1, characterized in that, The local monitoring module has a built-in macro-indicator aggregation engine based on a distributed stream processing framework. The macroeconomic indicator aggregation engine performs cross-level indicator fusion calculations based on a preset system, outputting a continuous real number of the risk situation assessment index for a specified region; the continuous real number of the risk situation assessment index for a specified region is obtained through the following logic: Extract the average multidimensional risk index of the current active status of the independent school and multiply it by the normalized weight ratio of the independent school in the regional system. Then, sum the total number of registered independent schools under its jurisdiction and multiply it by the risk base weight coefficient initialized by the system backend to form the first summation term. Extract the cumulative number of high-risk intervention work orders with overdue status within the jurisdiction, divide it by the total number of historical intervention work orders, and map it to a percentage system. Then, multiply the number of these numbers by the intervention lag penalty coefficient initialized in the system backend to form a second summation term. Finally, perform an overlay calculation operation on the first summation term and the second summation term.

10. The adolescent mental health early warning and intervention system based on multimodal AI analysis according to claim 9, characterized in that, The local monitoring module integrates a risk trend prediction engine based on a spatiotemporal graph convolutional neural network. The risk trend prediction engine abstracts each school in the jurisdiction into an independent node of a graph topology and constructs an edge adjacency matrix based on the reciprocal of the geographical travel distance between schools. The risk trend prediction engine receives the historical situation assessment index of each node and the associated auxiliary indicators, and generates a three-dimensional input sequence feature matrix with the dimension of the total number of nodes multiplied by the time step multiplied by the number of features. The three-dimensional input sequence feature matrix is ​​fed into a network architecture containing a graph convolutional network layer and a gated recurrent unit layer. The graph convolutional network layer uses the Laplacian matrix to aggregate the spatial feature distribution information of adjacent nodes. The gated recurrent unit layer performs temporal dependency modeling on the extracted spatial feature sequence to capture the time decay law. At the end of the network, a fully connected layer outputs a prediction tensor that represents the probability distribution of risk outbreaks in each school in the jurisdiction in the next quarter.

Images