A mental health early warning method and system based on multi-source student data fusion

Abstract

The present application relates to the technical field of complex system modeling and data analysis, and provides a mental health early warning method and system based on multi-source student data fusion, which is used to solve the problem of insufficient early warning accuracy and explainability caused by the lack of description of the time lag causal relationship between behaviors and the dynamic propagation of risks in traditional early warning technology. It constructs a behavior characteristic sequence through time series pattern decoupling, and accordingly constructs and dynamically updates a personal behavior time lag causal network based on a causal discovery algorithm. On this network, a risk diffusion simulation based on a causal propagation model is performed to obtain diffusion paths and predicted arrival times. Then, a dynamic risk diffusion situation map is generated, and based on this, key intervention nodes are identified through intervention effect simulation to generate an intervention path plan, so that the mental health early warning is improved from recognition based on static association to situation awareness and intervention planning based on causal deduction, significantly improving the causal explainability, dynamic foresight and decision support capability of early warning.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of data science and intelligent decision-making, and particularly to complex system modeling, temporal causal discovery and propagation dynamics simulation techniques. Specifically, it relates to a mental health early warning method and system based on the fusion of multi-source student data. Background Technology

[0002] Student mental health early warning refers to the process of using data analysis technology to identify and assess the early signs of mental health risks among students. It is an important task in the field of educational information technology and intelligent decision-making. In smart education management systems such as "Safe Campus," the need for continuous monitoring and precise intervention of the mental state of students is increasingly prominent. The accuracy, timeliness, and interpretability of early warnings directly affect the effectiveness of campus mental health prevention and control.

[0003] Traditional techniques primarily employ statistical analysis or machine learning models for student mental health early warning. One typical approach is threshold judgment based on psychological scale scores, which relies on subjective questionnaires, resulting in limited and outdated data. Another approach utilizes student behavioral data, constructing risk assessment models through feature engineering and classification algorithms. However, these methods typically rely on static or statistical correlations of features for prediction, failing to delve into the dynamic causal relationships between multidimensional behaviors and their temporal delays. They also cannot simulate the propagation and evolution of risk within behavioral networks, leading to early warning results that often remain at a binary "whether there is risk" level. This lack of in-depth insight into the causes, development trends, and optimal intervention points of risks limits the accuracy, foresight, and operability of the early warnings.

[0004] Therefore, how to construct a mental health early warning method that can characterize the time-delayed causal relationship between behaviors, simulate the dynamic transmission process of risks, and output a method that includes prospective intervention path analysis, so as to improve the accuracy, interpretability and decision support capabilities of the early warning, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a mental health early warning method and system based on multi-source student data fusion. This system elevates student mental health early warning from a single-point, static risk identification to a dynamic situational analysis and resilience intervention planning, thereby improving the accuracy, foresight, and decision support effectiveness of the early warning.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a mental health early warning method based on multi-source student data fusion, comprising: responding to a mental health early warning analysis command for a target student, acquiring multi-source heterogeneous behavioral time-series data of the target student, and decomposing the multi-source heterogeneous behavioral time-series data into behavioral feature sequences of N preset behavioral dimensions through time-series pattern decoupling, where N is a natural number; constructing a personal behavioral time-lag causal network based on the behavioral feature sequences, wherein nodes represent behavioral dimensions, edges represent time-lag causal relationships between behavioral dimensions, and the weights of the edges are dynamically learned and updated based on a causal discovery algorithm; determining mental health risk source nodes corresponding to the target mental health risk type, and performing a mental health risk diffusion simulation based on a causal propagation model on the personal behavioral time-lag causal network to obtain state variables characterizing mental health risk along the network edges. The data includes: the diffusion path and estimated arrival time from the source node of mental health risk to other nodes; a dynamic risk diffusion situation map, including a risk heatmap and a diffusion timeline, generated based on the diffusion path and estimated arrival time; key intervention nodes identified through intervention effect simulation based on the dynamic risk diffusion situation map, and intervention path planning generated based on the key intervention nodes, which are nodes that can maximally delay or block the spread of risk to preset key behavioral dimensions of mental health in the simulation; whether preset early warning triggering conditions are met based on the diffusion path and estimated arrival time; if the preset early warning triggering conditions are met, a mental health early warning report generated, including the dynamic risk diffusion situation map, the intervention path planning, and intervention reference prompts for the key intervention nodes.

[0007] The present invention also provides a mental health early warning system based on multi-source student data fusion. This mental health early warning system based on multi-source student data fusion corresponds one-to-one with the above-mentioned mental health early warning method based on multi-source student data fusion, as detailed in the following embodiments.

[0008] Beneficial Effects: This invention decomposes multi-source heterogeneous behavioral time-series data into behavioral feature sequences with clear semantics through temporal pattern decoupling, achieving structured representation and dimensionality reduction of high-dimensional mixed behavioral data. Furthermore, by constructing and dynamically updating a time-delayed causal network of individual behavior and simulating the diffusion of mental health risks based on a causal propagation model, it achieves a leap from behavioral correlation analysis to individualized causal mechanisms and prediction of time-varying risk trajectories. Finally, by generating a dynamic risk diffusion situation map that integrates a risk heatmap and a diffusion time axis, and generating intervention path planning based on intervention effect simulation, it automatically outputs a mental health early warning report containing the above visualization and inference results when preset early warning trigger conditions are met. This transforms the traditional early warning model that relies on static thresholds, single dimensions, or post-event statistics into an intelligent early warning and decision support method based on dynamic causal modeling, prospective risk propagation simulation, and data-driven intervention strategy inference, significantly improving the personalization, prediction accuracy, timeliness, and operability and pertinence of intervention recommendations in mental health early warnings. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the mental health early warning method based on multi-source student data fusion provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the first sub-process of the mental health early warning method based on multi-source student data fusion provided in an embodiment of the present invention; Figure 3 This is a schematic block diagram of a mental health early warning system based on multi-source student data fusion, provided as an embodiment of the present invention. Detailed Implementation

[0010] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0011] It should be noted that the mental health early warning method and system based on multi-source student data fusion involved in the embodiments of the present invention is an auxiliary decision-making tool based on data analysis and complex system modeling. It aims to provide risk situation awareness and intervention decision-making reference for mental health promotion work in educational settings such as schools. The output information of the embodiments of the present invention (including early warning reports, intervention reference prompts, etc.) is only auxiliary reference information based on the data model and does not constitute a medical diagnosis of any psychological or mental illness, nor does it constitute any clinically binding treatment or rehabilitation plan. In practical applications, users should make comprehensive judgments and decisions under the guidance of professionals, taking into account specific circumstances.

[0012] This invention provides a method and system for early warning of mental health based on the fusion of multi-source student data. This system can be an intelligent analysis engine hosted on a server or cloud platform, driving applications including but not limited to: First, within a "Safe Campus" integrated management platform, by integrating this early warning system, school administrators can be provided with dynamic awareness of individual and group mental health risks, assisting in resource allocation and the generation of key student mental health concerns; Second, within a student development guidance and support system, by analyzing intervention path planning and intervention reference prompts, data references can be provided for counselors to conduct personalized communication and support activities.

[0013] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings.

[0014] Example 1, please refer to Figure 1 , Figure 1 This is a flowchart illustrating a mental health early warning method based on multi-source student data fusion, provided in an embodiment of the present invention. Figure 1 As shown, in this embodiment, the method is applied to a server, and the method includes the following steps S11-S18: S11: In response to the mental health early warning analysis instruction for the target student, acquire the multi-source heterogeneous behavioral time series data of the target student, and decompose the multi-source heterogeneous behavioral time series data into behavioral feature sequences of N preset behavioral dimensions through time series mode decoupling, where N is a natural number.

[0015] Multi-source heterogeneous behavioral time-series data refers to a set of data sequences that are continuously collected from multiple independent sensors or data sources (such as campus card access control records, library borrowing system logs, classroom behavior monitoring video streams, dormitory network usage traffic records, wearable device physiological signals, etc.) and are strictly ordered by timestamps.

[0016] Temporal pattern decoupling refers to a data processing method that uses a parameterized algorithm model to separate and map the original multi-source heterogeneous behavioral temporal data to N predefined, semantically clear preset behavioral dimensions (such as the regularity of work and rest, the scope of social activities, the stability of learning input, and the consumption behavior pattern). The goal is to extract independent behavioral patterns from each dimension from the mixed signals and generate a sequence of behavioral features that characterize the temporal evolution of each behavioral dimension.

[0017] A behavioral feature sequence is a sequence of numerical feature vectors arranged in chronological order, generated by further feature calculations performed according to fixed time windows based on the initial signal obtained through temporal pattern decoupling processing on a certain preset behavioral dimension. Each element in the sequence corresponds to a time window, and its value is obtained by statistical aggregation (such as calculating the mean, variance, and entropy) of the decoupled temporal signal within the window or by extraction through a pre-trained feature encoder. This sequence quantitatively and continuously characterizes the pattern, intensity, or regularity of students' evolution over time in the corresponding behavioral dimension.

[0018] The implementation is as follows: First, in response to the received mental health early warning analysis instruction (i.e., the control signal that triggers mental health early warning analysis), the multi-source heterogeneous behavioral time series data of the target students are acquired and preprocessed. The acquisition and preprocessing include: pulling raw records from various data sources, de-identifying them, and uniformly converting them into an internal standard time series format, and aligning and buffering them according to timestamps. Secondly, the pre-loaded "Temporal Pattern Decoupling" module is invoked. Its core is a deep separation model with fixed parameters, based on a fusion architecture of multi-head self-attention and a one-dimensional convolutional neural network. This model performs forward computation on the concatenated multi-source heterogeneous behavioral temporal data: first, it maps each source data to a unified latent space through an embedding layer; then, it uses a multi-head self-attention mechanism to learn the cross-modal dependencies and dynamic temporal relationships between different data sources; finally, it performs convolution and pooling operations on the attention-weighted fusion features through a set of parallel, dimension-specific convolutional filter heads (each head corresponding to a "preset behavioral dimension"), thereby separating the global cluttered signals and focusing them onto each preset behavioral dimension. The deep separation model, pre-trained on a large amount of historical data, has learned to project complex input temporal patterns into N orthogonal subspaces. Thirdly, the output of each convolutional filter head is smoothed and normalized through a fully connected layer, ultimately generating a fixed-length behavioral feature sequence corresponding to the corresponding preset behavioral dimension. Specifically, for a time span of length T, each behavioral dimension will output a T*D feature matrix, where D is the number of dimensions of the corresponding behavioral dimension features. The behavioral feature sequence, as a decoupled structured intermediate representation, is stored in the feature database for subsequent use by the early warning analysis model. The temporal pattern decoupling module adopts an architecture that combines multi-head self-attention with a one-dimensional CNN. This is because multi-head self-attention can effectively capture complex, long-range cross-modal temporal dependencies between multi-source heterogeneous data, while one-dimensional CNN excels at extracting local temporal patterns and achieving dimensionality reduction. This combined design allows the model to simultaneously model global correlations and local details, thereby more accurately decomposing mixed behavioral signals into semantically clear preset behavioral dimensions.

[0019] For example, taking the "Safe Campus" scenario, in response to the weekly mental health early warning analysis instruction for the target student "Zhang San": First, acquire multi-source heterogeneous behavioral time-series data from the past week, which can cover dormitory access control card swipe records (timestamps), campus Wi-Fi access point switching logs (AP ID, timestamps), library seat management system check-in records (location, duration), cafeteria consumption records (time, amount), and anonymized frequency statistics from public area video surveillance, among other multimodal behavioral logs. Second, through the time-series pattern decoupling module, decompose and map the above mixed data streams to preset N=4 behavioral dimensions, generating corresponding behavioral feature sequences. Specifically, the output dimensions include: regularity of daily routine (dimensional A), scope of social activities (dimensional B), stability of learning engagement (dimensional C), and consumption behavior patterns (dimensional D). The behavioral feature sequence for each dimension is a numerical sequence that changes over time, quantifying the intensity or regularity of the behavioral pattern on that dimension.

[0020] S12: Based on the behavioral feature sequence, construct a time-delay causal network for personal behavior, where nodes represent behavioral dimensions, edges represent time-delay causal relationships between behavioral dimensions, and the weights of the edges are dynamically learned and updated based on a causal discovery algorithm.

[0021] The Personal Behavioral Lag Causal Network represents a personalized, dynamically evolving directed weighted graph data structure constructed for a specific target student. Nodes correspond to N predefined behavioral dimensions (such as "regularity of daily routine" and "range of social activities"). Edges represent directed connections from one behavioral dimension node to another. Physically, changes in the source dimension's behavioral pattern, after a corresponding time delay (lag), causally affect changes in the target dimension's behavioral pattern. Each edge is associated with two core attributes: a lag value (a positive integer time unit, such as days) and a weight value (a continuous real number). The weight value quantifies the strength and direction (positive / negative impact) of the causal relationship.

[0022] Lag causality refers to the causal relationship inferred from behavioral feature sequence data. It is defined as follows: for a given time interval τ (lag), if the feature value of behavioral dimension X at time t, after excluding interference from other behavioral dimensions, has the ability to predict or explain the change of feature value of behavioral dimension Y at time t+τ, then there is a causal relationship with a lag of τ from X to Y. It is different from instantaneous correlation, emphasizing the temporal constraint that the cause comes first and the effect comes later, which is more in line with the real physical process of behavioral pattern propagation and evolution. Its inference depends on the analysis of sequence data by the corresponding causal discovery algorithm.

[0023] Causal discovery algorithms represent mathematical models used to automatically infer causal relationships and their time lags between variables from multivariate time series data (i.e., sequences of behavioral features). In this invention, the algorithm is executed periodically in a sliding window manner, dynamically updating the edges (existence, direction, and time lag) and weights of the individual behavioral time-lag causal network based on new data to achieve online adaptive learning of the network.

[0024] The implementation is as follows: First, standardize N behavioral feature sequences (each sequence of length T) and form a T*N multivariate time series matrix. Then, initialize a time-delay causal network for individual behavior: using N preset behavioral dimensions as nodes, initialize a fully directed graph containing all N multiplied by (N-1) possible edges. Each edge e(X→Y) is associated with an initial time delay parameter τ_xy (randomly set within a preset range [τ_min, τ_max]) and an initial weight parameter w_xy (initially 0). Second, in each learning period t (t≥1), based on the sequence data of the current time window, execute the following causal discovery algorithm: 1) Constraint phase (identifying edges and time delays): For each pair of nodes (X, Y), within the predefined time delay search range [τ_min, τ_max], calculate the conditional independence test statistic (such as partial correlation coefficient) of X(t) and Y(t+τ) under the condition that all other node sets Z are at their respective optimal time delays. If at least one time delay τ makes the test significant (p-value less than the threshold), candidate directed edges X→Y are retained, and the time delay with the strongest significance is taken as its current time delay estimate τ_xy. Thus, using time constraints, the possible edges of the individual behavior time delay causal network and the time delays of each edge are initially determined. 2) Score stage (structure optimization and weight learning): Using the network skeleton output from the constraint stage as the initial structure, a scoring function is defined, such as the Bayesian Information Criterion (BIC). The scoring function evaluates both the goodness of fit of the network structure to the data (using a vector autoregression VAR model) and the model complexity. Through heuristic search (such as hill climbing), structural perturbations (addition, deletion, and reverse edges) are performed near the initial network skeleton, and the time delays of the edges are optimized simultaneously to maximize the BIC score. For each candidate network structure evaluated during the search process, the coefficient of each edge e(X→Y) is estimated by fitting a VAR model and minimizing the prediction error. This coefficient is then assigned as the weight w_xy of that edge, with its positive or negative sign indicating a promoting or inhibiting effect, and its absolute value indicating the strength of the effect. This stage outputs the optimal network structure, time delay τ_xy^(t), and weight w_xy^(t) for the current period. Third, a sliding time window with a fixed period (e.g., weekly) is used. When new behavioral feature sequence data is obtained, the network G^(t-1) that converged in the previous period is used as a hot start (i.e., as the initial search starting point for the above-mentioned score stage), and the two-stage algorithm of the constraint stage and score stage is repeated to obtain the updated network G^(t). In this way, the network structure and parameters can continuously evolve, reflecting the dynamic changes in the causal patterns of student behavior.

[0025] For example, following the previous example, based on student Zhang San's behavioral characteristic sequence across four preset behavioral dimensions over the past week, the personal behavioral time-lag causal network structure data is output after processing by a causal discovery algorithm as follows: 1) Node set V: {A (regularity of daily routine), B (range of social activities), C (stability of learning engagement), D (consumption behavior pattern)}. 2) Edge set E and parameters: (A → C, weight: +0.65, time lag: 1 day), (B → A, weight: -0.48, time lag: 0 days), (C → D, weight: +0.30, time lag: 2 days). This time-lag causal network reveals a strong positive feedback connection between nodes A and C, and a negative influence of node B on A.

[0026] S13: Determine the mental health risk source node corresponding to the target mental health risk type, and perform a mental health risk diffusion simulation based on the causal propagation model on the individual behavior time-delay causal network to obtain the diffusion path and expected arrival time of the state variable representing mental health risk from the mental health risk source node to other nodes along the network edge.

[0027] The causal propagation model represents a dynamic computational model used to simulate how state variables representing mental health risks propagate along causal edges on a time-delayed causal network of individual behavior. It employs a hybrid propagation model that combines a linear threshold model based on discrete time steps with the idea of ​​delay differential equations. It treats each network node as a unit with an internal state. The state change of a node is affected by the state of all its upstream neighbor nodes in the previous time step (considering time delay) and the weights of the connecting edges. The current risk state level is determined by an activation function. The causal propagation model mathematically defines the dynamic rules of the spread of risk signals in the network with time and topology.

[0028] Mental health risk diffusion simulation refers to the process of numerical simulation on a time-delayed causal network of individual behavior using a causal propagation model. It takes the initial state of the selected mental health risk source node (e.g., set to a high-risk level) as input, and iteratively calculates how the state variables representing mental health risk gradually propagate through directed edges in the network (considering their respective time delays and weights) and affect other nodes at multiple discrete time steps according to the dynamic equations defined by the causal propagation model. The goal is to predict the future trajectory of risk evolution.

[0029] The implementation is as follows: First, identify the mental health risk source nodes. Based on the target mental health risk type specified by the user or configured by the system (i.e., a predefined risk category, such as academic burnout risk), query the pre-set "risk type-behavioral dimension mapping knowledge base." This knowledge base defines the correlation strength coefficient and anomaly criteria between each risk type and each preset behavioral dimension. Combining this knowledge base with the latest terminal values ​​or short-term trends of the behavioral feature sequences of each dimension, calculate the "current risk exposure" score R_i for each node i. The score R_i is a weighted product of the correlation strength coefficient and the degree of anomaly of the dimension feature (standardized). Select one or a few nodes with the highest scores to determine the set S of mental health risk source nodes for this simulation, and set its initial risk state variable value to a high value (e.g., 1.0), while setting the initial values ​​of other nodes to low values ​​(e.g., 0.0). Second, perform a mental health risk diffusion simulation based on a causal propagation model. Load the time-delay causal network G of the current target student's personal behavior (containing a node set V, an edge set E, and each edge e(u→v) has a weight w_uv and a time delay τ_uv). Initialize the state values ​​x_v(0) of all nodes v at time t=0 (1.0 for the source node, 0.0 for the rest). Then, perform iterative simulations at discrete time steps t=1, 2, ... T_max. At each step t, for each non-source node v∈V (or all nodes in subsequent steps), calculate the total input signal I_v(t) it receives: I_v(t) = Σ_{(u→v)∈E} w_uv * x_u( max(0, t - τ_uv) ); where x_u(k) represents the state value of node u at the k-th discrete time step. When k<0, x_u(k) takes its initial value x_u(0), and max(0, t - τ_uv) ensures that the input signal takes into account the time delay τ_uv on the edges. Then, the state update of node v at time t follows the following rule: x_v(t) = min( 1.0, α * x_v(t-1) + (1-α) * σ( β * I_v(t) + θ_v ) ); where σ is the Sigmoid activation function used to simulate the nonlinear response of the node; α (0<α<1) is the state inertia coefficient, representing the inertia of the node's state change; β is the input gain coefficient; and θ_v is the intrinsic threshold of node v (which can be dynamically adjusted according to the node's historical characteristics). The causal propagation model, by integrating the cumulative characteristics of the linear threshold model with dynamic updates featuring inertia, can more smoothly simulate the propagation and accumulation of risk. Third, the diffusion path and expected arrival time are obtained.During the simulation, the time step t_arrival(v) when the state value x_v(t) of each node v first exceeds the preset "risk concern threshold" ξ (e.g., 0.5) is recorded. For each node v that reaches this threshold within the simulation time window, a "diffusion path" from the source node set S to v is determined by tracing back the main input source of its state change (i.e., the upstream node chain that contributes the most to I_v(t_arrival)). Finally, the set of all "associated and diffused" nodes v, the corresponding "expected arrival time" t_arrival(v) (which can be converted to the actual date), and the main "diffusion path" are output.

[0030] For example, following the previous example, for the target student Zhang San, "academic burnout risk" is set as the target mental health risk type. The system calculates and determines node C, the "stability of learning engagement" node, as the source node of mental health risk in his personal behavioral time-lag causal network. After performing a mental health risk diffusion simulation, the output includes: 1) Diffusion path: C → D (consumption behavior pattern). 2) Estimated arrival time: The estimated arrival time of the risk from node C to node D is 2 days. The simulation also reveals potential feedback paths (such as B→A→C) in the network, reflecting the dynamic evolution of risk in a complex causal structure.

[0031] S14: Based on the diffusion path and the expected arrival time, generate a dynamic risk diffusion situation map, which includes a risk heat map and a diffusion time axis.

[0032] Risk heatmaps are a core component of dynamic risk diffusion situation maps. They are used to dynamically map the risk state prediction intensity of each node in the time-delay causal network of individual behavior at the corresponding simulation time in a visual encoding manner. Through the mapping relationship between node visual attributes (such as color and size) and risk state values, they intuitively present the spatial distribution, clustering pattern and intensity dynamics of risks over time.

[0033] The diffusion timeline is an interactive time control linked to the risk heatmap in the dynamic risk diffusion situation map. Its time scale corresponds to the discrete time steps of the simulation and marks the expected arrival time of each node. By operating this timeline, users can control the heatmap to display the risk distribution at the corresponding moment or play the dynamic process continuously, realizing the visual tracking of the risk evolution sequence.

[0034] The implementation is as follows: First, the diffusion path and expected arrival time (including the sequence of risk state changes of each node over time {x_v(t)}, the expected arrival time of each node t_arrival(v), and the main diffusion path) are input into the graphics rendering engine. Second, a risk heatmap is constructed. The graphics rendering engine is based on the topology of the current individual behavior time-delay causal network (the initial layout can be calculated using algorithms such as force-directed algorithms). For the currently selected time point T_current on the time axis: 1) Query the risk state value x_v(T_current) of each node v at time T_current. 2) Map x_v(T_current) to a specific color value Color_v according to the preset color mapping function. For example, linear interpolation is used: Color_v = interpolate(Color_low, Color_high, x_v(T_current)), where Color_low and Color_high correspond to the start and end colors of the color spectrum, respectively. Optionally, calculate the rendering radius Radius_v of the node based on x_v(T_current) (e.g., Radius_v = R_base + k * x_v(T_current)). 3) On the canvas, draw a graphic (e.g., a circle) representing the node, with Radius_v as the radius and Color_v as the fill color, based on the node's layout coordinates. 4) According to the "diffusion path", highlight the directed edges from the "mental health risk source node" to the currently reached (i.e., x_v(T_current)>= ξ) node, and indicate the propagation direction with arrows. The color or thickness of the edges can be related to their weight or the activity of propagation. Third, integrate the diffusion timeline and achieve dynamic linkage. 1) Draw a horizontal timeline at the bottom or side of the visualization interface, with its starting point at the simulation start time (t=0) and its ending point at the maximum simulation time T_max. 2) On the timeline, for each node v with an "estimated arrival time" t_arrival(v), add a visual marker (such as a labeled vertical line or icon) at the corresponding scale position. 3) Implement interactive linkage between the timeline and the risk heatmap: When the user drags the slider on the timeline, the T_current value changes accordingly, immediately triggering a re-rendering of the risk heatmap to display the risk distribution at time T_current; when the user clicks on the arrival time marker of a node on the timeline, the interface automatically highlights the node's position in the heatmap and its related path.

[0035] S15: Based on the dynamic risk diffusion situation map, key intervention nodes are identified through intervention effect simulation, and intervention path planning is generated based on the key intervention nodes. The key intervention nodes are those nodes that can delay or block the spread of risk to the preset key psychological health behavior dimensions to the greatest extent in the simulation.

[0036] Intervention effect simulation refers to a digital extrapolation method that quantifies the effects of different intervention strategies by controlling variables based on the time-delay causal network and causal propagation model of individual behavior. The process includes: applying virtual interventions to nodes of different preset behavioral dimensions in the network sequentially and individually (i.e., adjusting the state values ​​of the target nodes in the initial conditions of the causal propagation model), rerunning the mental health risk diffusion simulation after each intervention, and calculating the quantitative difference in risk propagation results before and after the intervention, so as to provide a data basis for identifying the optimal intervention.

[0037] Key intervention nodes refer to the nodes of preset behavioral dimensions that are identified through intervention effect simulation and compared with preset quantitative assessment indicators (such as the time delay increment of risk reaching preset key behavioral dimensions of mental health). These nodes are the ones that contribute the most to achieving the goal of "delaying or blocking the spread of risk to preset key behavioral dimensions of mental health to the greatest extent" and are used as priority action targets to guide personalized intervention.

[0038] The implementation is as follows: First, determine the intervention assessment objectives. Based on presets or user specifications, clarify the set K of "preset key behavioral dimensions of mental health" that need to be protected (e.g., K={A: regularity of work and rest, B: scope of social activities}). Furthermore, from the causal network on which the dynamic risk diffusion pattern is based, exclude the set S of "mental health risk source nodes," and use all remaining non-source nodes as the candidate intervention node set C. Set the intensity parameter Δ of the virtual intervention (e.g., reduce the node state value by 0.3). Second, perform intervention effect simulation to identify key intervention nodes: 1) For each node c in the candidate intervention node set C, perform one simulation: on the initial conditions of the original simulation, modify the initial state value x_c(0) of node c to x_c(0) - Δ (ensuring it is not lower than 0), simulating the situation after its state is "improved." 2) Rerun the causal propagation model to obtain the risk diffusion process throughout the network under the new initial conditions. 3) Calculate the effectiveness evaluation index E_c of this intervention. The effectiveness evaluation index is defined as follows: For each preset key psychological health behavior dimension k∈K, calculate the delay increment ΔT_{c,k} = T'k - T_k relative to the original simulation (where T_k is the original arrival time and T'k is the arrival time after intervention; if some prognostic risks do not arrive within the simulation time window, then ΔT{c,k} is set to a large value). Finally, the comprehensive effectiveness index E_c of node c can be the weighted sum or minimum value of ΔT{c,k} for all k∈K. 4) After traversing all c∈C, select the node c* that makes the effectiveness evaluation index E_c optimal (e.g., the largest comprehensive delay increment) and identify it as the key intervention node. Third, generate intervention path planning: 1) Reverse causal analysis: In the individual behavior time-lag causal network G, starting from the key intervention node c, traverse its incoming edges in reverse to identify the set U of upstream nodes that have a direct or indirect significant positive impact on it (i.e., by increasing the state of these nodes, the improvement of c's state can be indirectly promoted), forming a potential "intervention support path". 2) Time-series planning: Combining the expected arrival time of each node in the original simulation, calculate the "last feasible intervention time window" for effective intervention on c before the risk reaches the preset key psychological health behavior dimension K. For example, the end time of the window can be set as min_{k∈K} T_k -τ_{ck} (where τ_{ck} is the shortest path time lag from c to the nearest preset key psychological health behavior dimension k).3) Generate a structured data object containing the following fields for intervention path planning: {primary_target: ID of the key intervention node c* and the suggested state adjustment target (e.g., "increase to value > 0.7"); supporting_nodes: List of upstream supporting nodes to be synchronized or prioritized, sorted by impact weight; time_window: Reference time window for initiating the intervention [t_start, t_end]; expected_effect: Expected technical effect, such as "expected time delay of X days for the risk to reach the preset key psychological health behavior dimension A".}

[0039] S16: Based on the diffusion path and the expected arrival time, determine whether the preset warning triggering conditions are met; S17: If the preset warning triggering conditions are met, generate a mental health warning report, which includes the dynamic risk diffusion situation map, the intervention path planning, and intervention reference prompts for the key intervention nodes; S18: If the preset warning triggering conditions are not met, end the current warning analysis process.

[0040] The process is as follows: Based on the diffusion path and expected arrival time, pre-defined quantitative rules are evaluated sequentially (e.g., determining whether the risk will reach any pre-defined key behavioral dimension of mental health within the next N days). If any rule is triggered, the pre-defined warning trigger condition is met, and a mental health warning report is automatically assembled and generated. The mental health warning report integrates a dynamic risk diffusion trend map (intuitively presenting the risk evolution), intervention path planning (providing structured intervention suggestions), and intervention reference prompts generated based on key intervention nodes (as supplementary reference information). The generated mental health warning report is pushed to the designated management terminal. If the pre-defined warning trigger condition is not met, the analysis process ends and resources are released.

[0041] This invention decomposes multi-source heterogeneous behavioral time-series data into behavioral feature sequences with clear semantics through temporal pattern decoupling, achieving structured representation and dimensionality reduction of high-dimensional mixed behavioral data. Furthermore, by constructing and dynamically updating a time-delayed causal network of individual behavior and simulating the diffusion of mental health risks based on a causal propagation model, a technological leap from behavioral correlation analysis to individualized causal mechanisms and time-varying risk trajectory prediction is achieved. Finally, by generating a dynamic risk diffusion situation map that integrates a risk heatmap and a diffusion time axis, and by generating intervention path planning based on intervention effect simulation, a mental health early warning report containing the above visualization and inference results is automatically output when preset early warning trigger conditions are met. This significantly improves the personalization, prediction accuracy, timeliness, and operability and pertinence of intervention recommendations in mental health early warnings.

[0042] In one embodiment, please refer to Figure 2 , Figure 2 This is a schematic diagram of the first sub-process of the mental health early warning method based on multi-source student data fusion provided in an embodiment of the present invention. Figure 2 As shown, in this embodiment, the multi-source heterogeneous behavioral time-series data is decomposed into behavioral feature sequences of N preset behavioral dimensions through temporal mode decoupling, including: S21: For each data source in the multi-source heterogeneous behavioral time-series data, it is divided into multiple data segments by an adaptive sliding window, and the temporal statistical features and time-frequency domain features of each data segment are extracted to obtain an initial feature matrix; S22: For the initial feature matrix, blind source signal separation is performed using an independent component analysis algorithm to eliminate cross-modal coupling noise between multiple data sources, and decoupled temporal signals of each behavioral dimension are obtained; S23: For each behavioral dimension temporal signal, robust filtering of outliers based on variational autoencoders is applied to generate smooth behavioral feature sequences; S24: The smooth behavioral feature sequences of each behavioral dimension are z-score standardized, and the standardized sequences are concatenated into a behavioral feature sequence matrix with a behavioral dimension of N*T to obtain the behavioral feature sequences of N preset behavioral dimensions corresponding to the multi-source heterogeneous behavioral time-series data, where T is the temporal length.

[0043] Independent Component Analysis (ICA) algorithm: This refers to a statistical computation method for blind source signal separation. Its core assumption is that the observed multi-channel mixed signal (i.e., the initial feature matrix) is composed of a linear mixture of several statistically independent source signals. By finding the unmixing matrix to maximize the statistical independence between the output signals (e.g., minimize mutual information), the independent source signals can be estimated and separated. In this invention, it is applied to separate relatively independent time-series signals corresponding to each preset behavioral dimension from the mixed multi-source features.

[0044] Robust outlier filtering using variational autoencoders (VAEs): This technique utilizes a VAE to reconstruct single-dimensional behavioral time-series signals, identifying and smoothing outliers. A VAE consists of an encoder (mapping the input sequence to latent spatial distribution parameters) and a decoder (for reconstructing the sequence). Training is based on normal behavioral sequence data (such as historical data of target students) to learn the latent distribution of normal patterns. In the filtering phase, after the input sequence is encoded and decoded by the VAE, the reconstruction error is calculated. Data points with errors exceeding a preset threshold are identified as outliers and replaced with reconstructed or interpolated values, thereby generating a smooth behavioral feature sequence that effectively suppresses noise and occasional anomalies.

[0045] The implementation is as follows: First, for each stream of "multi-source heterogeneous behavioral time-series data" (such as access control record streams and consumption flow streams), an adaptive sliding window algorithm is applied to divide it into a series of data segments. For each data segment, a set of "time-series statistical features" (such as the mean and variance of the number of consumptions within that time period) and a set of "time-frequency domain features" (such as the spectral entropy of the frequency of behavioral activities within that time period) are calculated in parallel, and all features are organized by segment to form the initial feature matrix of the data source. Second, the initial feature matrices from all data sources are concatenated after being aligned by time to form a global high-dimensional feature matrix, and an "Independent Component Analysis" (ICA) algorithm is applied to the high-dimensional feature matrix. The ICA algorithm finds the unmixing matrix W through iterative optimization (such as using Infomax or FastICA algorithms) so that the independence between the components of the transformed signal Y = W*X (where X is the input feature matrix) is maximized. These independent components are regarded as the preliminary estimate of the "decoupled behavioral dimension time-series signals". Each component theoretically corresponds to a potential behavioral pattern signal that is relatively less affected by coupling noise from other data sources. Third, for each behavioral dimension time series signal (corresponding to a time series) separated by ICA, "robust outlier filtering based on variational autoencoder" is applied. Specifically, a pre-trained VAE model is used to encode and reconstruct the signal. The error between the original signal value and the VAE reconstructed value at each time point is calculated. Points with errors exceeding a preset threshold are marked as outliers. Then, the VAE reconstructed value of the outlier or the interpolation between it and the normal points before and after it is replaced with the original outlier value, thereby obtaining a smooth behavioral feature sequence, which significantly improves data quality. Fourth, the smooth behavioral feature sequence obtained for each behavioral dimension is subjected to "z-score standardization" (i.e., subtracting the sequence mean and dividing by the sequence standard deviation) to make feature sequences under different dimensions and scales comparable. Finally, all N standardized sequences are concatenated row by row (behavioral dimension) to form an N*T behavioral feature sequence matrix, where T is a uniform time series length.

[0046] In this embodiment of the invention, a refined characterization of multi-source heterogeneous behavioral time-series data is performed through adaptive sliding window and multi-dimensional feature extraction. Then, an independent component analysis algorithm is used to perform blind source separation to eliminate cross-modal coupling noise and obtain preliminary signals corresponding to each preset behavioral dimension. Furthermore, an anomaly robust filtering technique based on variational autoencoder is used to improve sequence quality and robustness, generating a smooth behavioral feature sequence, thereby improving the data quality and analysis accuracy of the early warning chain.

[0047] In one embodiment, constructing a time-delay causal network of individual behavior includes: Initialize the adjacency matrix of the individual behavior time-delay causal network and for each edge (i , j Associate a time delay parameter ; In each learning cycle t Based on the current behavioral feature sequence window data, a temporal causal discovery model based on an attention mechanism is used to calculate the nodes. i For nodes j causal strength With optimal time delay The temporal causal discovery model is optimized by minimizing a prediction loss function with a temporal sparse regularization term; The edges in the adjacency matrix are dynamically updated according to the following formulas: i , j Weights and time delay : ; ; Where λ is the smoothing factor and η is the learning rate of the time delay parameter. This represents the value calculated by the temporal causal discovery model for the edge (in the current learning cycle t). i , j The new time delay estimate, and It is the historical stored value of the network at the beginning of the learning period t. and It is the updated value used for period t+1; calculate and Frobenius norm difference ; Determine whether the norm difference δ is less than a preset convergence threshold. ; If δ < Stop updating. As the final network weight matrix A of the aforementioned individual behavior time-delay causal network, The final time delay matrix of the aforementioned individual behavior time delay causal network and output the network weight matrix. A With the time delay matrix .

[0048] A temporal causal discovery model based on attention mechanisms represents a neural network model used to infer causal relationships and their time lags between variables from a sequence of multivariate behavioral features. Its core is the self-attention mechanism, which calculates the dependency weights between features at any two time points in the sequence to effectively capture long-range temporal dependencies. It typically includes an encoder, a multi-head self-attention layer, and a causal inference head. Through end-to-end training, it learns the mapping from sequence data to causal structures. The temporal causal discovery model is invoked in each learning cycle t to calculate the causal strength between nodes based on the current data. With optimal time delay .

[0049] Temporal sparsity regularization is a penalty term added to the training loss function of a temporal causal discovery model. It is used to encourage the model to learn a sparse causal network structure (e.g., using L1 norm regularization λ * ||W||_1, where W is the causal weight matrix and λ is the regularization coefficient). Based on the prior knowledge that causal effects in behavioral systems are usually limited, temporal sparsity regularization can improve the interpretability of the learned network and prevent overfitting.

[0050] The implementation is as follows: First, based on the number of behavioral dimensions N, initialize the N*N adjacency matrix of the individual behavior time-delay causal network. All its elements (weights) Typically, it is initialized to 0 or a very small random number, indicating that no causal relationship is determined initially. Furthermore, a time delay matrix of the same size is initialized. Each element The initial time delay estimate from node i to node j can be set to a preset default value (e.g., 1) or randomly initialized within a reasonable range. Secondly, in each learning period t (e.g., whenever a new "behavioral feature sequence" window of data has been accumulated over a week): the N-dimensional behavioral feature sequence data within the current time window is input into the attention-based temporal causal discovery model. The temporal causal discovery model (e.g., an architecture combining a temporal convolutional network (TCN) and multi-head self-attention) outputs a causal strength matrix through forward computation. (Element is) and optimal time delay matrix (Element is) The temporal causal discovery model is optimized during training by minimizing a prediction loss function with a temporal sparse regularization term. This loss function typically consists of two parts: the prediction loss (e.g., the mean squared error of predicting the behavior of the next time step using the causal structure inferred by the model) and a sparse regularization term (e.g., an L1 regularization term that penalizes non-zero values). This forces the model to learn a causal structure that is both accurate and sparse. The edges in the adjacency matrix are then dynamically and smoothly updated according to the following formula:i , j Weights and time delay ,: This is a first-order low-pass filter, which makes the weights change smoothly; This is an incremental adjustment using gradient descent.

[0051] Then, calculate the Frobenius norm difference between the weight matrices before and after the update. And determine whether δ is less than the preset convergence threshold. (A small positive number close to 0).

[0052] If δ ≥ The network is not yet stable. and This serves as the initial value for the next cycle (t+1), and the cycle continues. If δ < The network is considered to have converged to a stable state, and updates are stopped. At this point, the obtained... This is the weight matrix of the final individual behavior time-delay causal network. A , This is the final time delay matrix T_au, output. A and This is for use in subsequent risk diffusion simulations.

[0053] In this embodiment of the invention, a temporal causal discovery model based on an attention mechanism is used to periodically estimate causal strength and time delay. A temporal sparse regularization term and a dynamic smoothing update formula are introduced to achieve progressive online learning and optimization of a time-delay causal network for individual behavior. The attention mechanism and sparse constraints ensure the accuracy and interpretability of the network. The smoothing factor λ and the learning rate η enable the network to stably absorb new data and suppress noise. Finally, the Frobenius norm difference δ is used as the convergence criterion to ensure stable output. This significantly improves the network's personalized adaptation ability, the accuracy of temporal evolution tracking, and its robustness as a basic model for risk simulation.

[0054] In one embodiment, performing a mental health risk diffusion simulation based on a causal propagation model includes: according to the network weight matrix A With the time delay matrix A continuous-time, multi-dimensional coupled differential equation system is constructed as a causal propagation model: ; in, Represents a node i In time t Risk state variables, For the node's own dynamic function, g( ) is the coupling function. Noise term; The initial state of the mental health risk source nodes The system is set to an abnormally active state, with the remaining nodes set to the baseline state. The multi-dimensional coupled differential equations are solved using the fourth-order Runge-Kutta numerical integration method to simulate the propagation and evolution of risk states on the time-delay causal network of individual behavior. The time when the risk state variable of each node first exceeds a preset risk threshold is recorded. The expected arrival time is used as the basis for backtracking and generating diffusion paths based on the activation time sequence of each node and edge connectivity.

[0055] In this invention, a continuous-time, multi-dimensional coupled differential equation system refers to a mathematical model used to describe the dynamic propagation and evolution of state variables representing mental health risks on a time-delayed causal network of individual behavior. It contains N differential equations (corresponding to N being the number of behavioral dimensions), with each equation corresponding to a node i in the network. The equation structure is as follows: The core of this is: the rate of change of risk state for each node i ( It is determined by three parts: 1) the dynamics of the node itself (e.g., resilience toward baseline or tendency to deteriorate); 2) Coupling effects from other nodes j ,in It is the causal weight from j to i (taken from weight matrix A). The corresponding time delay (taken from the time delay matrix) ), g( 3) Mapping neighbor node states to influence; (Such as Gaussian white noise), used to simulate unobserved external disturbances, this set of equations characterizes the continuous, smooth dynamic process of risk propagation in the network over the continuous time domain.

[0056] Node self-dynamic function , is the risk state describing node i in the above system of differential equations. A mathematical function describing its own evolution over time in the absence of external node influence. In one example of this invention, it employs a function with a stable attractor, such as... ,in It is the attenuation coefficient (which controls the speed at which the node state returns to the baseline). This represents the node's inherent baseline or equilibrium state (such as the normal value in a state of mental health). The node's own dynamic function ensures that when there is no external risk input, the node's risk state will naturally decay or return to a stable level, simulating the inertia of behavioral habits or the self-regulating tendency of psychological states.

[0057] Coupling function g( ), which represents the risk state of upstream node j. This is a nonlinear mapping function that transforms the state values ​​of neighboring nodes into an applicable driving force. Common forms include linear functions g(x) = x, or nonlinear functions such as the sigmoid function g(x) = 1 / (1 + exp(-β*(x-γ))), where β and γ are parameters, and the nonlinear g(x) = 1 / (1 + exp(-β*(x-γ))). It can simulate the threshold effect, that is, only when the risk of the upstream node exceeds a certain level will it have a significant impact on the downstream node, thus enabling the model to more realistically simulate the nonlinear characteristics in the propagation of behavioral patterns.

[0058] The implementation is as follows: First, model construction: based on the network weight matrix A and time delay matrix and the pre-defined form of the dynamic function (such as...) (g(x) = x or the Sigmoid function), instantiate a continuous-time multidimensional coupled differential equation system, with specific parameters of the equation for each node i (such as... , The initial state of the node can be customized based on the statistical sequence of its historical behavioral characteristics. Secondly, based on the set S of mental health risk source nodes, the initial state of the corresponding node is determined. Set to an abnormal activation state (e.g., uniformly set to 1.0), all other nodes j in the individual behavior time-delay causal network (j The initial state of S) All are set to "baseline state" (e.g., set to 0.1 or its inherent value). Third, the fourth-order Runge-Kutta numerical integration method (RK4) is used to solve the above-initialized differential equation system to simulate the evolution process from t=0 to the preset simulation termination time t=T_max. The specific steps are as follows: 1) Select an appropriate time step h (e.g., 0.1 days). 2) For each time point t_k, calculate the risk state value of all nodes at time t_(k+1)=t_k+h according to the RK4 formula. During the calculation process, when it is necessary to calculate the delay term... At that time, if For non-discrete time points, interpolation methods such as cubic spline interpolation can be used to obtain the values ​​from the calculated historical state sequence to ensure accurate integration of time delays. The above calculation process is then iteratively executed until t_k ≥ T_max. Throughout this process, the state value sequence of all nodes at all discrete time points t_k is recorded. Fourth, 1) Estimated arrival time: Iterate through the state sequence of each node i { Find the time point t_k when its state value first exceeds the preset risk threshold (e.g., ξ=0.5), and record this time. = t_k records the estimated arrival time of the node. If the threshold is not exceeded within the simulation time window, it is marked as "not arrived". 2) Diffusion path: Based on the order of the "estimated arrival times" of each node, combined with the topology of the individual behavior time-delay causal network (weight matrix A), a reverse backtracking is performed. For each non-source node i that is activated in the simulation (i.e., the state exceeds the threshold), find the upstream node j corresponding to all its incoming edges that was activated earlier than i in time and has causal weight Larger nodes are considered the primary sources of risk for node i. Based on this, a causal chain is generated from the mental health risk source node to each activated node through recursive backtracking. The set of causal chains constitutes the diffusion path.

[0059] In this embodiment of the invention, based on the network weight matrix A With time delay matrix A system of multidimensional coupled differential equations in continuous time is constructed as a causal propagation model and solved using numerical integration. This achieves a high-fidelity continuous time-series simulation of the dynamic propagation of risk on a time-delayed causal network of individual behavior. Furthermore, the continuous changes in risk are characterized by differential equations, and the learned time-delay parameters are innovatively incorporated into the model. Embedded Coupling Items In this way, the risk transmission strictly follows the causal time delay, thereby accurately predicting the expected arrival time and tracing the diffusion path, which significantly improves the rationality of the simulation, the accuracy of the time sequence, and the foresight of the early warning method.

[0060] In one embodiment, key intervention nodes are identified through intervention effect simulation, including: defining a set of intervention actions. I ={ }, where each intervention action Preset state controls are applied to one or more behavioral dimension nodes; based on the dynamic risk diffusion situation map, for each candidate intervention node of the individual behavior time-lag causal network. v Simulate the application of each intervention action in sequence. The network state changes after intervention; by solving the modified causal propagation model after the intervention, the process of mental health risk diffusion is re-simulated, and the expected arrival time delay ΔT of the mental health risk diffusion to the preset key mental health behavioral dimensions after intervention is calculated. , ) and the final risk level reduction ΔR of the preset key behavioral dimensions of mental health , Based on the preset cost-benefit weights, calculate the cost-benefit weights for each candidate intervention node. Overall intervention efficacy score across all intervention actions , where U( ) is a utility function used to combine delay and reduction effects; Indicates intervention action The corresponding preset cost weights quantify the implementation of corresponding practical measures (such as...) The estimated resource consumption or operational difficulty of the associated tutoring or activities; the top M nodes with the highest comprehensive intervention effectiveness scores are selected as key intervention nodes. Key intervention nodes are those for which it is recommended to prioritize the implementation of their associated real-world support measures (from the set). I ) behavioral dimension targets.

[0061] Intervention Action Set I ={ } represents a set of predefined virtual operation instructions. Each intervention action Technically defined, this is a mathematical operation used to quantitatively regulate the state parameters (such as initial state values ​​and dynamic parameters) of corresponding nodes in a time-delay causal network of individual behavior. This mathematical operation corresponds to a predefined psychological and behavioral support measure (such as one-on-one counseling or task assignment), and is a quantitative simulation of the expected impact after implementing that measure. For example, actions... The initial value of the "Regularity of Daily Routine" node can be increased from 0.1 to 0.4 to simulate the effect of "Personalized Daily Routine Guidance" measures.

[0062] A utility function (U(·)) is a function that maps two quantitative indicators of the intervention effect—the expected arrival time delay ΔT and the final risk reduction ΔR—to a single scalar value. It is used to comprehensively evaluate the technical benefits of the intervention. For example, a simple linear utility function can be: U(ΔT, ΔR) = a * ΔT + b * ΔR, where a and b are positive coefficients, representing the degree of importance attached to the delay and risk reduction, respectively. The U(·) function is designed to transform multi-objective optimization (longer delay, lower risk) into a single-objective comparison, facilitating selection.

[0063] The implementation is as follows: First, load the predefined set of intervention actions. I , I The measures are stored in the system knowledge base as a "measure-control mapping table," with each record containing: a description of the measure (business meaning), the target behavioral dimension, and specific mathematical control parameters (such as the state adjustment magnitude). Furthermore, based on the risk situation reflected in the dynamic risk diffusion map, a set C of candidate intervention nodes v that need to be evaluated is determined (usually all non-psychological health risk source nodes, or nodes on the critical path selected based on network topology). Next, for each candidate intervention node v, each intervention action is traversed. and according to The defined mathematical regulation modifies the model parameters. Based on the original individual behavioral time-delay causal network and initial conditions for risk diffusion, corresponding preset state regulation is implemented to apply intervention, resulting in a modified causal propagation model after the intervention. This modified causal propagation model is then used to re-simulate the psychological health risk diffusion process, yielding the post-intervention risk propagation trajectory. Furthermore, two core indicators are extracted from the post-intervention simulation results: 1) the expected arrival time delay ΔT(v, ): Calculate the delay in risk arrival time after intervention relative to the original simulated arrival time for each preset key psychological health behavioral dimension. Typically, the minimum or average delay across all preset key psychological health behavioral dimensions is taken as ΔT. , 2) The final reduction in risk level ΔR( , At the end of the simulation, calculate the reduction in the final risk state value of each preset key behavioral dimension of mental health relative to the original simulation final value, and take the average or minimum value as ΔR. , This process, which simulates the effects of interventions, is equivalent to systematically testing various real-world support measures in a digital twin environment. Expected effects for different behavioral problems (node ​​v). Third, calculate the overall intervention efficacy score: after completing all ( , After the combined simulation, for each candidate intervention node v, its comprehensive score for all intervention actions is calculated. Where U(·) is a predefined utility function, Intervention action The pre-defined cost weights and summation operation mean that, for a node v, the effects of all possible interventions applied to it are evaluated and summed in weight to obtain a comprehensive score S(v) representing the overall intervention potential of that node. Fourth, all candidate intervention nodes are sorted from highest to lowest according to their comprehensive intervention efficacy scores S(v), and the top M nodes (M is a pre-defined positive integer, such as 1 or 2) are selected as the identified key intervention nodes. The output will indicate the node and its associated intervention actions. (Specific measures such as one-on-one tutoring, task assignment, and participation in group activities are suggested.) The key intervention point is the point where intervention may yield the greatest overall cost-benefit return under the current risk situation.

[0064] In this embodiment, by elevating intervention optimization from experience-based judgment to computation based on simulation and multi-objective quantitative evaluation, not only is the intervention effect evaluated, but cost considerations are also taken into account. This enables data-driven and accurate identification of the optimal intervention point on an individualized causal network, significantly improving the objectivity and reference value of the intervention path.

[0065] In one embodiment, generating an intervention path plan includes: for each of the key intervention nodes Determine the optimal intervention action. , Among them, the independent variable The value range belongs to the set of intervention actions. I The symbol argmax represents the function that makes the function... The independent variable with the largest value ; Identify the M key intervention nodes and their corresponding optimal intervention actions As the basic nodes of the intervention planning diagram; based on the time-delay causal network of individual behaviors, the causal influence relationship between the key intervention nodes is analyzed, and the intervention synergy effect matrix among the key intervention nodes is constructed; using a Markov decision process model with resource and time constraints, with minimizing the overall risk level and intervention cost as the objective function, and the synergy effect matrix as the auxiliary parameter for state transition, the optimal intervention sequence is solved. According to the optimal intervention sequence It generates an intervention path plan that includes the sequence of intervention nodes, the type of intervention action, and the recommended implementation time window.

[0066] The intervention synergy matrix is ​​an M*M matrix (where M is the number of key intervention nodes), and its elements... Quantitatively representing the prior key intervention nodes Implement its optimal intervention action Then, key intervention nodes. Implement its optimal intervention action At that time, the degree to which the latter's expected effect is enhanced (or weakened). Element Values ​​are typically obtained through additional simulation calculations: comparison pairs The effect of intervention alone versus intervention in combination Further intervention on the basis The differences in effectiveness are illustrated by the intervention synergy matrix, which characterizes the nonlinear interactions between different key intervention nodes / action combinations, providing crucial input for planning ordered intervention sequences.

[0067] The optimal intervention sequence is an ordered action plan output by solving a Markov decision process model. It is a list of triples, where each element... Indicates: During the suggested simulation time (or actual time point), for key intervention nodes Implement the corresponding optimal intervention action . It could be a precise point in time, or it could be the starting point of a time window.

[0068] The implementation is as follows: First, for each key intervention node... Analysis in calculating S( Recorded at the time of intervention, and different intervention actions The corresponding sub-fraction S( , By comparison, determine what makes S( , The largest intervention action, i.e., the optimal intervention action. . M ( , Pairings, serving as fundamental nodes in the intervention planning diagram, represent discrete, optimal intervention options. Secondly, based on the time-lag causal network of individual behaviors, the causal connections (direct or indirect) between the M key intervention nodes are analyzed. For each pair of nodes ( , Perform additional simulations: 1) Simulate separately for... Apply Effect (referred to as basic effect) ); 2) Simulate first Apply Wait for its effects to propagate through the network for a certain period of time (e.g., with a time lag) before proceeding. Apply The combined effect (denoted as the combined effect) Synergistic effect elements It can be defined as ( - ) / ( (Relative enhancement rate) or directly the difference, thereby constructing the intervention synergy matrix E. Third, define the state space (risk state, used resources, current time) and action space (choosing a certain action) of the Markov Decision Process (MDP model). , The search involves a sequence of actions, including a state transition model (based on causal propagation models and synergistic effect matrices to predict the next state), a reward function (e.g., reward = ω1 * risk reduction - ω2 * action cost), and constraints (total resource budget R_max, total time window T_max). Then, using dynamic programming (e.g., value iteration) or sampling-based optimization algorithms (e.g., Monte Carlo tree search), the search seeks the sequence of actions that maximizes the cumulative reward, given the constraints. During the search process, when evaluating an action (intervention)... When querying the synergy effect matrix E, if certain nodes have already been intervened in, then... ,according to Adjust the expected impact of the current action on the state (risk). The strategy obtained from the solution, under the initial state (current risk situation, resources are 0, time is 0), will derive a deterministic sequence of actions. Each action in this sequence (excluding "waiting" actions) is extracted, and based on the execution time point t determined during model solving, an optimal intervention sequence is formed. Fourth, generate intervention path planning: This involves creating the optimal intervention sequence. Converting this into a structured description results in intervention path planning, which explicitly lists: 1) the order of intervention nodes: , , ... (corresponding to) (in order); 2) Intervention action type: the corresponding number of nodes , , ...;3) Recommended implementation time window: based on Based on the expected onset time of the intervention and the expected arrival time of risk transmission, a launch time window is recommended for each intervention. , ].

[0069] For example, continuing from the previous example, suppose two key intervention points are identified: A (lifestyle) and D (consumption), and their optimal intervention actions are respectively (Moderately raising the baseline of daily routine) and (Slightly enhancing consumption regularity). Simulations show that improving daily routines (A) may indirectly make subsequent interventions in consumption (D) more effective by improving overall well-being (E_{D|A}>0), while the synergistic effect is weak. Furthermore, under the constraint of limited "attentional energy" resources, the MDP model, through simulation, finds that Option 1: First... back Due to positive synergy, the overall risk reduction is greater and more lasting; Option 2, the reverse order, is less effective. Therefore, the solution... Finally, the following intervention path planning recommendations are generated: First (in the next few days) focus on supporting their regularity of daily routines (node ​​A), and then (about a week later) supplement this with attention to consumption behavior (node ​​D).

[0070] In this embodiment of the invention, by upgrading the intervention strategy from static node selection to dynamic and collaborative sequence optimization, and by using the intervention synergy effect matrix and constraint optimization model, the overall intervention effectiveness is maximized under the premise of considering the coordination between measures, resource constraints and timing. This significantly improves the systematicness, operability and expected effect of the intervention recommendations.

[0071] In one embodiment, the preset early warning triggering conditions include at least one of the following: there exists at least one diffusion path, the terminal node of the diffusion path is a preset key psychological health behavior dimension; for any preset key psychological health behavior dimension, its expected arrival time is less than a preset first time threshold; the topological importance weighted sum of the set of nodes covered by all diffusion paths is calculated, the topological importance weighted sum exceeds a preset second threshold; based on the dynamic risk diffusion situation map, the growth rate of the overall network risk entropy within a preset time period Δt in the future is calculated, the growth rate exceeds a preset third threshold.

[0072] The overall network risk entropy, constructed based on information entropy theory, is a scalar indicator used to characterize the uncertainty or disorder of the risk state distribution at a certain moment in a time-delayed causal network of individual behavior. It is calculated as follows: at a given simulation time t, the risk state values ​​{X_i(t)} of all nodes are normalized so that their sum is 1, resulting in a probability distribution p_i(t) = X_i(t) / Σ_j X_j(t). Then, the information entropy of this distribution is calculated: H(t) = -Σ_i [p_i(t) *log(p_i(t))]. When the risk is highly concentrated in a few nodes, the entropy value is low (ordered and deterministic state); when the risk is evenly distributed among many nodes, the entropy value is high (disordered and uncertain state). Changes in entropy can reflect the evolution pattern of risk distribution.

[0073] The growth rate of the overall network risk entropy refers to the rate of change of the overall network risk entropy H(t) within a predetermined time period Δt. It can be calculated using a difference approximation: Growth_Rate = (H(t+Δt) - H(t)) / Δt, where H(t+Δt) and H(t) are obtained through risk diffusion simulation. A positive growth rate indicates that the risk is becoming more dispersed, the system's disorder or uncertainty is increasing, and it foreshadows a worsening of the risk; a negative growth rate indicates that the risk is tending to concentrate. A growth rate exceeding a predetermined third threshold indicates that the risk distribution is undergoing a rapid and unfavorable evolution.

[0074] The implementation is as follows: The preset early warning triggering conditions are a logical set consisting of multiple quantitative judgment rules that can be used independently or in combination. The following conditions are evaluated sequentially or in parallel: Condition 1) Existence check of key dimension paths. Traverse all diffusion paths and check if there is at least one path whose terminal node (i.e., the last node of the path) belongs to the predefined preset set of key psychological health behavioral dimensions K. If it exists, Condition 1 is triggered, indicating that the risk prediction will directly affect the core behavioral dimension that needs the most protection. Condition 2) Urgency check of key dimension arrival time. For each node k ∈ K of the preset key psychological health behavioral dimension, read its expected arrival time T_k (set to infinity if it has not arrived), calculate the difference between T_k and the current time. If there exists any k such that (T_k - current time) ≤ the preset first time threshold (e.g., 3 days), Condition 2 is triggered, indicating that the risk will soon affect the key dimension in the short term, and time is urgent. Condition 3) Scope and severity check of risk impact. Determine the set of nodes V_affected covered by all diffusion paths. Based on the current time-delay causal network structure of individual behavior, pre-calculate or calculate in real-time the topological importance weight w_i^topo of each node i (e.g., calculate eigenvector centrality using the NetworkX library), and calculate the topological importance weighted sum: S_top = Σ_{i∈V_affected} w_i^topo. If S_top > the preset second threshold, condition three is triggered, indicating that the risk has not only spread but has also affected many structurally important nodes in the network, indicating systemic dysfunction. Condition four) Dynamic deterioration check of risk distribution structure. Based on the simulation data of the dynamic risk diffusion situation map, extract the risk states {X_i(t)} and {X_i(t+Δt)} of each node in the network at the current time t and at the future time t+Δt (e.g., one day later). Calculate the overall network risk entropy H(t) and H(t+Δt) at time t and t+Δt respectively. The entropy growth rate is calculated as: G = (H(t+Δt) - H(t)) / Δt. If G > the preset third threshold (a positive number), condition four is triggered, indicating that the risk is rapidly spreading from the local to the global level in the short term, and the uncertainty and disorder of the system state are rapidly increasing, signaling an accelerated deterioration of the risk. The above conditions are combined using a logical "OR" relationship. If any one of the conditions is evaluated as true (triggered), the preset warning trigger condition is met, and a warning report is generated; otherwise, the process ends.

[0075] In this embodiment of the invention, by constructing a comprehensive automatic early warning decision logic, it not only assesses whether and when a risk affects key targets, but also evaluates the breadth and structural evolution trend of its impact from a system perspective. In particular, it captures early signals of risk spread by the overall network risk entropy growth rate, and achieves intelligent judgment from multiple dimensions such as target, time, space and structure, which significantly improves the accuracy, foresight and reliability of early warning.

[0076] In one embodiment, after generating a mental health early warning report, the method further includes: associating and storing the generated dynamic risk diffusion trend map, the intervention path planning, the network snapshot at the time of early warning triggering, and the intervention implementation record to establish a personal mental health early warning file for the target student; based on the target student's historical personal mental health early warning file, by comparing the network structure evolution and the migration pattern of mental health risk source nodes in previous early warnings, using a time-series pattern mining algorithm, generating an evolutionary trend analysis and periodic report corresponding to the target student's mental health risk; and using the evolutionary trend analysis and periodic report as a supplementary attachment to the mental health early warning report, or pushing it to the corresponding management terminal according to preset rules.

[0077] The implementation is as follows: First, after each mental health early warning report is generated, the dynamic risk diffusion trend map, intervention path planning, and network snapshot of the triggering time (including weight matrix A and time delay matrix) of this early warning are automatically extracted. After being linked to student identifiers, these data are stored in the student's personal mental health early warning file. Authorized users can supplement this file with actual intervention implementation records. Secondly, the personal mental health early warning files of target students are periodically analyzed. Time-series pattern mining algorithms (such as trend decomposition and sequence pattern mining) are applied to mine the network structure indicator sequences and mental health risk source node sequences of previous early warnings, automatically generating evolutionary trend analyses and periodic reports describing long-term changes and cyclical patterns. Finally, the generated reports can be used as historical background to supplement new early warning reports or pushed independently to the management terminal.

[0078] Furthermore, using a time-series pattern mining algorithm, an evolutionary trend analysis and periodic report corresponding to the mental health risks of the target students are generated, including: extracting the network topology feature vector of the individual behavior time-lag causal network at each warning time to construct a network feature evolution time-series sequence; applying singular spectrum analysis to the network feature evolution time-series sequence to extract long-term trend components and periodic components; performing Fourier transform on the periodic components to identify the main period; and generating an evolutionary trend analysis and periodic report based on the long-term trend components and the main period.

[0079] The implementation is as follows: First, network snapshots at each warning time are extracted from the individual mental health early warning files of target students in chronological order. For each network snapshot, feature vectors reflecting its topological structure (such as network density, average clustering coefficient, etc.) are calculated to construct a time-series sequence of network feature evolution. Second, singular spectral analysis (SSA) is applied to the time-series sequence of network feature evolution to decompose it into long-term trend components and periodic components. Furthermore, frequency domain analysis (such as Fourier transform) is performed on the periodic components to identify the main period with the strongest energy. Third, the macroscopic evolution direction of network structural features is described based on the long-term trend components, and its regular fluctuations are described based on the main period and periodic components, integrating these to generate an evolutionary trend analysis and a periodic report.

[0080] In this embodiment of the invention, by establishing and maintaining a personal mental health early warning file after an early warning, the response to a single event is upgraded to continuous process management. Based on this file, a time-series pattern mining algorithm is applied to automatically discover the evolution trend and cycle of individual risk patterns and generate an analysis report with long-term insights. The system is built into an intelligent platform with memory and learning capabilities, which significantly enhances the continuity and foresight of early warning.

[0081] The mental health early warning method based on multi-source student data fusion described in the above embodiments can be recombined as needed to obtain a combined implementation scheme, but all are within the protection scope claimed by this invention.

[0082] In one embodiment, a mental health early warning system based on multi-source student data fusion is provided. This system corresponds one-to-one with the mental health early warning method based on multi-source student data fusion described in the above embodiments. Please refer to [link / reference]. Figure 3 , Figure 3 This is a schematic block diagram of a mental health early warning system based on multi-source student data fusion, provided as an embodiment of the present invention. Figure 3 As shown, the mental health early warning system 30 based on multi-source student data fusion is applied to a server and includes: a first decomposition module 31, a first construction module 32, a first simulation module 33, a first generation module 34, a second generation module 35, a first judgment module 36, and a third generation module 37. The detailed descriptions of the above functional modules are as follows: The first decomposition module 31 is used to respond to the mental health early warning analysis command for the target student, acquire the multi-source heterogeneous behavioral time-series data of the target student, and decompose the multi-source heterogeneous behavioral time-series data into behavioral feature sequences of N preset behavioral dimensions through time-series pattern decoupling, where N is a natural number; the first construction module 32 is used to construct a personal behavioral time-lag causal network based on the behavioral feature sequences, wherein nodes represent behavioral dimensions, edges represent time-lag causal relationships between behavioral dimensions, and the weights of the edges are dynamically learned and updated based on a causal discovery algorithm; the first simulation module 33 is used to determine the mental health risk source node corresponding to the target mental health risk type, and perform a mental health risk diffusion simulation based on a causal propagation model on the personal behavioral time-lag causal network to obtain the diffusion path and expected arrival time of the state variable representing mental health risk from the mental health risk source node to other nodes along the network edges. The first generation module 34 is used to generate a dynamic risk diffusion situation map based on the diffusion path and the expected arrival time. The dynamic risk diffusion situation map includes a risk heat map and a diffusion time axis. The second generation module 35 is used to identify key intervention nodes based on the dynamic risk diffusion situation map through intervention effect simulation, and generate an intervention path plan based on the key intervention nodes. The key intervention nodes are nodes that can delay or block the spread of risk to preset key psychological health behaviors to the greatest extent in the simulation. The first judgment module 36 is used to determine whether the preset early warning triggering conditions are met based on the diffusion path and the expected arrival time. The third generation module 37 is used to generate a psychological health early warning report if the preset early warning triggering conditions are met. The psychological health early warning report includes the dynamic risk diffusion situation map, the intervention path plan, and intervention reference prompts for the key intervention nodes.

[0083] In one embodiment, the first decomposition module 31 includes: a first extraction submodule, used to divide each data source in the multi-source heterogeneous behavioral time-series data into multiple data segments using an adaptive sliding window, and extract the temporal statistical features and time-frequency domain features of each data segment to obtain an initial feature matrix; a first separation submodule, used to perform blind source signal separation on the initial feature matrix using an independent component analysis algorithm to eliminate cross-modal coupling noise between multiple data sources, and obtain decoupled temporal signals for each behavioral dimension; a first filtering submodule, used to apply robust outlier filtering based on variational autoencoder to each of the behavioral dimension temporal signals to generate smooth behavioral feature sequences; and a first splicing submodule, used to perform z-score normalization on the smooth behavioral feature sequences of each behavioral dimension, and splice the normalized sequences into a behavioral feature sequence matrix with a behavioral dimension of N*T, to obtain N preset behavioral feature sequences corresponding to the multi-source heterogeneous behavioral time-series data, where T is the temporal length.

[0084] In one embodiment, the first construction module 32 includes: a first initialization submodule, used to initialize the adjacency matrix of the individual behavior time-delay causal network. and for each edge ( i , j Associate a time delay parameter The first calculation submodule is used in each learning cycle. t Based on the current behavioral feature sequence window data, a temporal causal discovery model based on an attention mechanism is used to calculate the nodes. i For nodes j causal strength With optimal time delay The temporal causal discovery model is optimized by minimizing a prediction loss function with a temporal sparse regularization term; the first update submodule is used to dynamically update the edges in the adjacency matrix according to the following formula ( i , j Weights and time delay : ; ; Where λ is the smoothing factor and η is the learning rate of the time delay parameter. This represents the value calculated by the temporal causal discovery model for the edge (in the current learning cycle t). i , j The new time delay estimate, and It is the historical stored value of the network at the beginning of the learning period t. and This is the updated value used for period t+1; the second calculation submodule is used for calculation. and Frobenius norm difference The first judgment submodule is used to determine whether the norm difference δ is less than a preset convergence threshold. The first output submodule is used if δ < Stop updating. As the final network weight matrix A of the aforementioned individual behavior time-delay causal network, The final time delay matrix of the aforementioned individual behavior time delay causal network and output the network weight matrix. A With the time delay matrix .

[0085] In one embodiment, the first simulation module 33 includes: a first construction submodule, configured to, based on the network weight matrix... AWith the time delay matrix A continuous-time, multi-dimensional coupled differential equation system is constructed as a causal propagation model: ;in, Represents a node i In time t Risk state variables, For the node's own dynamic function, g( ) is the coupling function. For noise terms; the first setting submodule is used to set the initial state of the mental health risk source node. The system is set to an abnormal activation state, with the remaining nodes set to the baseline state; the first solution submodule is used to solve the multidimensional coupled differential equations using the fourth-order Runge-Kutta numerical integration method to simulate the propagation and evolution of risk states on the time-delay causal network of individual behavior; the first acquisition submodule is used to record the time when the risk state variable of each node first exceeds a preset risk threshold. The expected arrival time is used as the basis for backtracking and generating diffusion paths based on the activation time sequence of each node and edge connectivity.

[0086] In one embodiment, the second generation module 35 includes: a first definition submodule, used to define a set of intervention actions. I ={ }, where each intervention action The system can apply preset state controls to one or more behavioral dimension nodes; the first simulation submodule is used to apply preset state controls to each candidate intervention node of the individual behavior time-lag causal network based on the dynamic risk diffusion situation map. v Simulate the application of each intervention action in sequence. The network state changes after intervention; the third calculation submodule is used to re-simulate the mental health risk diffusion process by solving the modified causal propagation model after the intervention, and calculate the expected arrival time delay ΔT of the mental health risk diffusion to the preset key mental health behavioral dimensions after intervention. , ) and the final risk level reduction ΔR of the preset key behavioral dimensions of mental health , The fourth calculation submodule is used to calculate the cost-benefit weights for each candidate intervention node. Overall intervention efficacy score across all intervention actions , where U( Let ) be the utility function. Indicates intervention action The corresponding preset cost weights; the first selection submodule is used to select the top M nodes with the highest comprehensive intervention effectiveness scores as key intervention nodes.

[0087] In one embodiment, the second generation module 35 includes: a first determination submodule, used for determining each of the key intervention nodes. Determine the optimal intervention action. , Among them, the independent variable The value range belongs to the set of intervention actions. I The symbol argmax represents the function that makes the function... The independent variable with the largest value The second determining submodule is used to determine the M key intervention nodes and their corresponding optimal intervention actions. The first submodule serves as the basic node of the intervention planning diagram; the second submodule is used to analyze the causal relationship between the key intervention nodes based on the time-delay causal network of individual behaviors, and construct the intervention synergy effect matrix between the key intervention nodes; the second submodule is used to solve for the optimal intervention sequence using a Markov decision process model with resource and time constraints, with the objective function of minimizing the overall risk level and intervention cost, and with the synergy effect matrix as the auxiliary parameter for state transition. The first generation submodule is used to generate the optimal intervention sequence. It generates an intervention path plan that includes the sequence of intervention nodes, the type of intervention action, and the recommended implementation time window.

[0088] In one embodiment, the preset early warning triggering conditions include at least one of the following: there exists at least one diffusion path, and the terminal node of the diffusion path is a preset key psychological health behavior dimension; for any preset key psychological health behavior dimension, its expected arrival time is less than a preset first time threshold; a fifth calculation submodule is used to calculate the topological importance weighted sum of the set of nodes covered by all diffusion paths, and the topological importance weighted sum exceeds a preset second threshold; a sixth calculation submodule is used to calculate the growth rate of the overall network risk entropy within a preset time period Δt in the future based on the dynamic risk diffusion situation map, and the growth rate exceeds a preset third threshold.

[0089] In one embodiment, the mental health early warning system 30 further includes: a first establishment module, used to associate and store the dynamically generated risk diffusion situation map, the intervention path planning, the network snapshot at the time of early warning triggering, and the intervention implementation record to establish a personal mental health early warning file for the target student; a fourth generation module, used to generate an evolution trend analysis and periodic report corresponding to the mental health risk of the target student based on the target student's historical personal mental health early warning file, by comparing the network structure evolution and the migration pattern of mental health risk source nodes in previous early warnings, and using a time-series pattern mining algorithm; and a first processing module, used to use the evolution trend analysis and periodic report as a supplementary attachment to the mental health early warning report, or push it to the corresponding management terminal according to preset rules.

[0090] In one embodiment, the fourth generation module includes: a first composition submodule, used to extract the network topology feature vector of the individual behavior time-lag causal network at each warning time, to form a network feature evolution time series sequence; a second extraction submodule, used to apply singular spectrum analysis to the network feature evolution time series sequence to extract long-term trend components and periodic components; a first identification submodule, used to perform Fourier transform on the periodic components to identify the main period; and a second generation submodule, used to generate an evolution trend analysis and a periodic report based on the long-term trend components and the main period.

[0091] Specific limitations regarding the mental health early warning system based on multi-source student data fusion can be found in the limitations of the mental health early warning method based on multi-source student data fusion mentioned above, and will not be repeated here. Each module in the aforementioned mental health early warning system based on multi-source student data fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0092] Those skilled in the art will understand that the methods and systems provided in the embodiments of the present invention can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. The methods can also be implemented as a computer program product stored in one or more computer-readable storage media, including but not limited to: disks, optical disks, read-only memory (ROM), random access memory (RAM), flash memory, etc. When the computer program product is executed by one or more data processing devices (such as computers), the devices perform the steps as described in any of the preceding method embodiments.

[0093] Software tools, components, or models not belonging to this company that appear in the embodiments of this invention are merely illustrative examples and do not represent actual use. The data collection methods used in the embodiments of this invention comply with relevant laws and regulations, such as the "Data Security Law of the People's Republic of China," the "Personal Information Protection Law of the People's Republic of China," GDPR (General Data Protection Regulation of the European Union), or information security standards of other countries and regions.

[0094] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A mental health early warning method based on multi-source student data fusion, characterized in that, include: In response to the mental health early warning analysis command for the target student, the multi-source heterogeneous behavior time series data of the target student is acquired, and the multi-source heterogeneous behavior time series data is decomposed into behavior feature sequences of N preset behavior dimensions through time series pattern decoupling, where N is a natural number. Based on the behavioral feature sequence, a time-delay causal network for personal behavior is constructed, where nodes represent behavioral dimensions, edges represent time-delay causal relationships between behavioral dimensions, and the weights of the edges are dynamically learned and updated based on a causal discovery algorithm. Identify the mental health risk source node corresponding to the target mental health risk type, and perform a mental health risk diffusion simulation based on the causal propagation model on the individual behavior time-delay causal network to obtain the diffusion path and expected arrival time of the state variable representing mental health risk from the mental health risk source node to other nodes along the network edge. Based on the diffusion path and the expected arrival time, a dynamic risk diffusion situation map is generated, which includes a risk heat map and a diffusion time axis. Based on the dynamic risk diffusion situation map, key intervention nodes are identified through intervention effect simulation, and intervention path planning is generated based on the key intervention nodes. The key intervention nodes are those nodes that can delay or block the spread of risk to preset key psychological health behaviors to the greatest extent in the simulation. Based on the diffusion path and the expected arrival time, determine whether the preset early warning triggering conditions are met; If the preset early warning triggering conditions are met, a mental health early warning report is generated. The mental health early warning report includes the dynamic risk diffusion situation map, the intervention path planning, and intervention reference prompts for the key intervention nodes.

2. The mental health early warning method based on multi-source student data fusion as described in claim 1, characterized in that, By decoupling through temporal patterns, the multi-source heterogeneous behavioral temporal data is decomposed into behavioral feature sequences of N preset behavioral dimensions, including: For each data source in the multi-source heterogeneous behavioral time series data, it is divided into multiple data segments by an adaptive sliding window, and the time series statistical features and time-frequency domain features of each data segment are extracted to obtain an initial feature matrix; For the initial feature matrix, the independent component analysis algorithm is used to separate blind source signals to eliminate cross-modal coupling noise between multiple data sources and obtain decoupled time-series signals of each behavioral dimension; For each of the aforementioned behavioral dimension time-series signals, anomaly robust filtering based on variational autoencoder is applied to generate smooth behavioral feature sequences. The smoothed behavioral feature sequence of each behavioral dimension is z-score normalized, and the normalized sequences are concatenated into a behavioral feature sequence matrix of N*T behavioral dimension to obtain the behavioral feature sequences of N preset behavioral dimensions corresponding to the multi-source heterogeneous behavioral time series data, where T is the time series length.

3. The mental health early warning method based on multi-source student data fusion as described in claim 1 or 2, characterized in that, Constructing a time-delayed causal network of individual behaviors includes: Initialize the adjacency matrix of the individual behavior time-delay causal network and for each edge ( i , j Associated with a time delay parameter ; In each learning cycle t Based on the current behavioral feature sequence window data, a temporal causal discovery model based on an attention mechanism is used to calculate the nodes. i For nodes j causal strength With optimal time delay The temporal causal discovery model is optimized by minimizing a prediction loss function with a temporal sparse regularization term; The edges in the adjacency matrix are dynamically updated according to the following formulas: i , j Weights and time delay : ; ; Where λ is the smoothing factor and η is the learning rate of the time delay parameter. This represents the value calculated by the temporal causal discovery model for the edge (in the current learning cycle t). i , j The new time delay estimate, and It is the historical stored value of the network at the beginning of the learning period t. and It is the updated value used for period t+1; calculate and Frobenius norm difference ; Determine whether the norm difference δ is less than a preset convergence threshold. ; If δ < Stop updating. As the final network weight matrix A of the aforementioned individual behavior time-delay causal network, The final time delay matrix of the aforementioned individual behavior time delay causal network and output the network weight matrix. A With the time delay matrix .

4. The mental health early warning method based on multi-source student data fusion as described in claim 3, characterized in that, Perform a simulation of the diffusion of mental health risks based on a causal propagation model, including: According to the network weight matrix A With the time delay matrix A continuous-time, multi-dimensional coupled differential equation system is constructed as a causal propagation model: ; in, Represents a node i In time t Risk state variables, For the node's own dynamic function, g( ) is the coupling function. Noise term; The initial state of the mental health risk source nodes Set the node to an abnormal activation state, and set the remaining nodes to the baseline state; The fourth-order Runge-Kutta numerical integration method is used to solve the multidimensional coupled differential equations to simulate the propagation and evolution of risk state on the time-delay causal network of individual behavior. Record the time when the risk status variable of each node first exceeds the preset risk threshold. The expected arrival time is used as the basis for backtracking and generating diffusion paths based on the activation time sequence of each node and edge connectivity.

5. The mental health early warning method based on multi-source student data fusion as described in claim 1, characterized in that, Through intervention effect simulation, key intervention nodes are identified, including: Define the set of intervention actions I ={ }, where each intervention action Apply preset state control to one or more behavioral dimension nodes; Based on the dynamic risk diffusion situation map, for each candidate intervention node in the time-delay causal network of individual behavior v Simulate the application of each intervention action in sequence. Subsequent changes in network state; By solving the modified causal propagation model after the intervention, the diffusion process of mental health risks is re-simulated, and the expected arrival time delay ΔT of the mental health risks after the intervention to the preset key mental health behavioral dimensions is calculated. , ) and the final risk level reduction ΔR of the preset key behavioral dimensions of mental health , ); Calculate the cost-benefit weights for each candidate intervention node. Overall intervention efficacy score across all intervention actions , where U( Let ) be the utility function. Indicates intervention action The corresponding preset cost weights; The top M nodes with the highest comprehensive intervention effectiveness scores were selected as key intervention nodes.

6. The mental health early warning method based on multi-source student data fusion as described in claim 5, characterized in that, Generate intervention pathway planning, including: For each of the aforementioned key intervention nodes Determine the optimal intervention action. , Among them, the independent variable The value range belongs to the set of intervention actions. I The symbol argmax represents the function that makes the function... The independent variable with the largest value ; M key intervention nodes and their corresponding optimal intervention actions As a basic node in the intervention planning map; Based on the time-delayed causal network of individual behavior, the causal influence relationship between the key intervention nodes is analyzed, and the intervention synergy effect matrix between the key intervention nodes is constructed. A Markov decision process model with resource and time constraints is adopted, with the objective function of minimizing the overall risk level and intervention cost, and the synergy matrix used as an auxiliary parameter for state transition, to solve for the optimal intervention sequence. ; According to the optimal intervention sequence It generates an intervention path plan that includes the sequence of intervention nodes, the type of intervention action, and the recommended implementation time window.

7. The mental health early warning method based on multi-source student data fusion as described in claim 1, characterized in that, Preset warning trigger conditions include at least one of the following: There exists at least one of the aforementioned diffusion paths, and the terminal node of the diffusion path is a preset key behavioral dimension of mental health. For any preset key behavioral dimension of mental health, the estimated arrival time is less than a preset first time threshold; Calculate the topological importance weighted sum of the set of nodes covered by all diffusion paths, where the topological importance weighted sum exceeds a preset second threshold; Based on the dynamic risk diffusion situation map, the growth rate of the overall network risk entropy within a future preset time period Δt is calculated, and the growth rate exceeds a preset third threshold.

8. The mental health early warning method based on multi-source student data fusion as described in claim 1, characterized in that, After generating the mental health early warning report, it also includes: The dynamic risk diffusion situation map, the intervention path plan, the network snapshot when the early warning is triggered, and the intervention implementation record generated this time will be stored together to establish a personal mental health early warning file for the target student. Based on the historical personal mental health early warning files of the target students, by comparing the evolution of network structure and the migration patterns of mental health risk source nodes in previous early warnings, and using a time-series pattern mining algorithm, an evolutionary trend analysis and periodic report corresponding to the mental health risks of the target students are generated. The evolutionary trend analysis and periodic reports can be used as supplementary attachments to the mental health early warning report, or pushed to the corresponding management terminal according to preset rules.

9. The mental health early warning method based on multi-source student data fusion as described in claim 8, characterized in that, Using a time-series pattern mining algorithm, an evolutionary trend analysis and periodic report on the mental health risks of the target students are generated, including: Extract the network topology feature vector of the individual behavior time-delay causal network at each warning time to construct a network feature evolution time series sequence; Singular spectral analysis was applied to the time series sequence of network feature evolution to extract long-term trend components and periodic components; Perform a Fourier transform on the periodic components to identify the main period; Based on the long-term trend components and the main period, an evolutionary trend analysis and periodic report are generated.

10. A mental health early warning system based on multi-source student data fusion, characterized in that, include: The first decomposition module is used to respond to the mental health early warning analysis command for the target student, acquire the multi-source heterogeneous behavior time series data of the target student, and decompose the multi-source heterogeneous behavior time series data into N preset behavior dimension behavior feature sequences through time series mode decoupling, where N is a natural number. The first construction module is used to construct a time-delay causal network of personal behavior based on the behavioral feature sequence, wherein nodes represent behavioral dimensions, edges represent time-delay causal relationships between behavioral dimensions, and the weights of the edges are dynamically learned and updated based on a causal discovery algorithm. The first simulation module is used to determine the mental health risk source node corresponding to the target mental health risk type, and to perform a mental health risk diffusion simulation based on the causal propagation model on the individual behavior time-delay causal network to obtain the diffusion path and expected arrival time of the state variable representing mental health risk from the mental health risk source node to other nodes along the network edge. The first generation module is used to generate a dynamic risk diffusion situation map based on the diffusion path and the expected arrival time. The dynamic risk diffusion situation map includes a risk heat map and a diffusion time axis. The second generation module is used to identify key intervention nodes based on the dynamic risk diffusion situation map and through intervention effect simulation, and to generate intervention path planning based on the key intervention nodes. The key intervention nodes are nodes that can delay or block the spread of risk to preset key psychological health behaviors to the greatest extent in the simulation. The first judgment module is used to determine whether the preset early warning triggering conditions are met based on the diffusion path and the expected arrival time. The third generation module is used to generate a mental health early warning report if the preset early warning triggering conditions are met. The mental health early warning report includes the dynamic risk diffusion situation map, the intervention path planning, and intervention reference prompts for the key intervention nodes.

Images