A method for dynamically monitoring psychological health state of college students by side writing data

Abstract

The present application relates to the technical field of mental health information monitoring, and discloses a method for dynamically monitoring the mental health state of college students based on side-writing data, which comprises the following steps: collecting multi-source side-writing data of college students within discrete time slices; obtaining calendar events, course arrangements and time period information and constructing a context vector; generating expected side-writing features through a context alignment counterfactual digital twin model based on historical summary states and the context vector; calculating counterfactual residuals according to actual side-writing features and expected side-writing features; dividing the feature domain and accumulating the time window to construct a drift evidence chain; combining uncertainty indicators to obtain risk intensity and risk level through risk inference; calculating the remaining stable period according to the risk intensity change trend; and executing active micro-measurement and updating the stable period sample set or risk inference parameters when the trigger condition is met. Through the joint modeling mechanism of constructing side-writing feature vectors and context vectors, semantic bias caused by independent analysis of single behavior features is avoided.

Description

Technical Field

[0001] This invention relates to the field of information-based monitoring technology for mental health, specifically a method for dynamically monitoring the mental health status of college students using profile data. Background Technology

[0002] With the advancement of information technology construction in higher education institutions, the monitoring of college students' mental health is gradually shifting from traditional manual assessment methods to data-driven analysis. Current technologies typically rely on psychological scales, interview records, and counselor observations for mental health identification. Some solutions incorporate student behavior data for auxiliary analysis, such as course attendance records, daily routine statistics, campus card usage, and mobile device usage. These solutions then classify or assess student mental health based on statistical analysis or machine learning models. This type of technology achieves quantitative analysis of mental health by collecting students' daily behavioral data and constructing behavioral characteristics.

[0003] However, the aforementioned technical solutions mostly rely on the independent statistical characteristics of student behavior data as the basis for analysis, without structurally modeling the external contextual factors that generate the behavior. In practical applications, student behavior patterns are significantly influenced by contextual conditions such as course schedules, exam cycles, semester rhythms, and environmental factors; the same behavioral changes have different meanings in different contexts. Existing technologies do not introduce a unified representation and joint modeling mechanism for contextual information during behavioral feature analysis, leading to unstable semantic correspondences between behavioral features and actual states, which can easily cause biases and affect the consistency of monitoring results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for dynamically monitoring the mental health status of college students using profile data. This method solves the problem that existing technologies do not introduce a unified representation and joint modeling mechanism for contextual information during behavioral feature analysis, which leads to unstable semantic correspondence between behavioral features and actual states, easily causing biases and affecting the consistency of monitoring results.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for dynamically monitoring the mental health status of college students using profile data, comprising the following steps: S1. Collect multi-source profile data of college students in discrete time slices, and perform statistical feature extraction and desensitization aggregation on the multi-source profile data to generate profile feature vectors for the corresponding time slices. S2. Obtain the context information corresponding to the time slice, and encode the context information to generate the context vector corresponding to the time slice; S3. Based on the profile feature vector and context vector of historical time slices, update the historical summary status of individual students through the state update function; S4. Align the historical summary state with the context vector input context to a counterfactual digital twin model to obtain the expected profile features under the current context conditions, and form the expected profile feature vector. S5. Calculate the counterfactual residual based on the profile feature vector and the expected profile feature vector to obtain the counterfactual residual vector of the current time slice; S6. The counterfactual residual vector is standardized, and the standardized counterfactual residuals are aggregated according to the preset feature domain partitioning method to obtain the domain residual strength of multiple feature domains. S7. Perform time-accumulated calculation on the domain residual strength within a preset time window to construct a drift evidence chain, and determine the consistency of the evidence chain based on the domain residual strength of multiple feature domains. S8. Calculate the risk intensity based on the drift evidence chain, evidence chain consistency, situation vector, and uncertainty index, and generate a mental health risk level according to a preset grading threshold. S9. Calculate the remaining stable period based on the changing trend of the risk intensity within the time window; S10. When the risk intensity meets the active micro-test triggering condition, perform active micro-test and obtain subjective state information, and update the stable period sample set or risk inference parameters according to the subjective state information. S11. Output includes monitoring results of mental health risk level, uncertainty index, drift evidence chain information, and remaining stable period.

[0006] Preferably, the profiling feature vector in step S1 is generated through the following steps: Align multi-source profiled data according to time slices; Execute the feature mapping function for each type of data source to obtain the corresponding data source feature vector; The feature vectors from various data sources are concatenated to form a profile feature vector of uniform dimension, and the profile feature vector is then normalized.

[0007] Preferably, the context vector in step S2 is generated by a context encoding function, and the context information includes at least one of the following: School calendar event information, course schedule information, time period information, and environmental information; The context encoding function converts the context information into a numerical vector and forms the context vector.

[0008] Preferably, the context-aligned counterfactual digital twin model in step S4 is used to predict the desired profiling features based on the historical summary state and context vector, and its calculation process includes: The historical summary state of the current time slice is updated using a state update function based on the historical summary state of the previous time slice, the profile feature vector, and the context vector. The updated historical summary state and the current context vector are input into the conditional prediction function to obtain the desired profile feature vector.

[0009] Preferably, the counterfactual residual in step S5 is calculated by the difference between the profiled feature vector and the expected profiled feature vector, and then standardized by the residual scaling vector to obtain the standardized counterfactual residual vector.

[0010] Preferably, the drift evidence chain in step S7 is constructed through the following steps: The standardized counterfactual residuals are divided into multiple feature domains according to a preset feature domain; The absolute value of the residual is aggregated within each feature domain to obtain the domain residual strength. The domain residual strength is weighted and accumulated within a preset time window to form the domain chain strength; The consistency of the evidence chain is determined based on the domain chain strength of multiple feature domains, and a drift evidence chain is formed.

[0011] Preferably, the uncertainty index in step S8 is calculated based on at least one of the following factors: Data completeness of profiled feature vectors; Prediction variance of context-aligned counterfactual digital twin models; Missing input data.

[0012] Preferably, the risk intensity in step S8 is calculated by a risk inference function, which takes drift evidence chain, evidence chain consistency, situation vector and uncertainty index as inputs and outputs a continuous risk intensity value.

[0013] Preferably, the remaining stable period in step S9 is calculated based on the trend of risk intensity changes within a time window, and the calculation process includes: The slope of risk change is obtained by trend estimation of the risk intensity sequence within a time window. The remaining stable period is calculated based on the slope of the risk change and the high-risk threshold.

[0014] Preferably, the proactive micro-test in step S10 is achieved by sending a brief psychological state inquiry to the student's terminal and obtaining subjective state information, and performing at least one of the following update operations based on the subjective state information: Update the stable period sample set; Update the parameters of the risk inference function; Update the risk classification threshold.

[0015] This invention provides a method for dynamically monitoring the mental health status of college students using profiling data. It has the following beneficial effects: 1. This invention achieves the structured fusion of college students' daily behavior data and environmental context information by constructing a joint modeling mechanism of profile feature vectors and context vectors. This enables the monitoring process to consider both behavioral performance and external context factors simultaneously, avoiding semantic bias caused by independent analysis of single behavioral features.

[0016] 2. This invention introduces a context-aligned counterfactual digital twin model to generate a desired profile feature vector based on the historical summary state and the current context vector, enabling the monitoring system to obtain a behavioral reference baseline under the current context conditions, thereby providing a calculable control quantity for subsequent deviation identification.

[0017] 3. This invention constructs counterfactual residuals, using the difference between the actual profile feature vector and the expected profile feature vector as a deviation representation, thereby stripping away the influence of contextual factors and transforming the anomaly detection object from its original behavioral changes into a deviation after context alignment, thus improving the consistency of deviation identification.

[0018] 4. This invention constructs a drift evidence chain by dividing the counterfactual residuals into feature domains and accumulating them over time windows, thereby achieving continuous aggregation of deviation information from multiple time slices and feature domains, and making risk judgment based on continuous changes.

[0019] 5. This invention uses an evidence chain consistency determination mechanism to compare the domain chain strength of multiple feature domains by threshold, forming a structured representation of cross-domain collaborative deviation, thereby improving the overall judgment capability of multi-dimensional behavioral changes. Attached Figure Description

[0020] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart illustrating the overall process of the method of the present invention. Figure 3 This is a flowchart of the profile data acquisition and feature generation process of the present invention; Figure 4 This is a flowchart of the context information processing and context vector construction process of the present invention; Figure 5 The flowchart for the context-aligned counterfactual digital twin model processing of this invention is shown below; Figure 6 This is a flowchart of the risk inference and remaining stable period calculation of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see the appendix Figure 1 -Appendix Figure 6 This invention provides a method for dynamically monitoring the mental health status of college students using profile data, comprising the following steps: S1. Collect multi-source profile data of college students in discrete time slices, and perform statistical feature extraction and desensitization aggregation on the multi-source profile data to generate profile feature vectors for the corresponding time slices. Furthermore, the profiling feature vector in step S1 is generated through the following steps: Align multi-source profiled data according to time slices; Execute the feature mapping function for each type of data source to obtain the corresponding data source feature vector; The feature vectors from various data sources are concatenated to form a profile feature vector with a unified dimension, and then the profile feature vector is normalized.

[0023] Specifically, in step S1, multi-source profile data of university students within discrete time slices is collected, and statistical feature extraction and de-identification aggregation are performed on the multi-source profile data to generate profile feature vectors for the corresponding time slices. .

[0024] In this embodiment, the multi-source profiling data includes at least learning profiling data, lifestyle profiling data, and device profiling data. Learning profiling data is generated by the teaching platform and includes login events, resource access events, assignment submission events, and course attendance events. Lifestyle profiling data is generated by the campus card system and access control system and includes consumption events and entry / exit events. Device profiling data is generated by student terminals and wearable devices and includes screen unlock events, application category usage statistics, sensor statistics, sleep statistics, and activity statistics. All types of data are generated in the form of event streams, and each event includes at least the event type, event timestamp, and an anonymized student identifier.

[0025] In this embodiment, the feature extraction and desensitization aggregation module merges events into corresponding time slices according to time slice alignment rules; it performs fixed statistical feature extraction operations for each data source, including at least: event count statistics, cumulative duration statistics, intraday segment count statistics, event interval statistics, and historical difference statistics; then, it concatenates the features of each data source in a preset order to form a profile feature vector. The de-identification aggregation follows fixed rules: categorical fields are mapped to preset category numbers; time information only retains the time slice index and segment number; specific location and application identifiers are not stored, only the statistical results are retained; the monitoring server only stores... It generates a log identifier but does not store the original event details.

[0026] S2. Obtain the context information corresponding to the time slice, encode the context information, and generate the context vector corresponding to the time slice; Furthermore, the context vector in step S2 is generated through a context encoding function, and the context information includes at least one of the following: School calendar event information, course schedule information, time period information, and environmental information; The context encoding function converts context information into a numerical vector and forms a context vector.

[0027] Specifically, in step S2, the context information corresponding to the time slice is obtained, and the context information is encoded to generate a context vector for the corresponding time slice. .

[0028] In this implementation, the context information includes academic calendar event information, course schedule information, time period information, and environmental information. Academic calendar event information includes semester week order, exam schedule, and statutory holiday identifiers; course schedule information includes the daily course count and morning / evening course density; time period information includes weekday number and intraday time slot number; and environmental information is provided by a defined data source and bound to a time slice index. The context vector construction module performs multi-hot encoding on discrete fields, fixed-boundary scaling encoding on continuous fields, and fixed-mapping encoding on periodic fields, and concatenates the encoding results in a preset order to form a context vector. The context vector and the profiling feature vector are stored aligned using the same time slice index t.

[0029] S3. Based on the profile feature vector and context vector of historical time slices, update the historical summary status of individual students through the state update function; Specifically, in step S3, based on the profiled feature vector and context vector of the historical time slice, the historical summary state of each student is updated through a state update function. .

[0030] In this embodiment, the historical summary status update module maintains the historical summary status for each student. The historical summary state is updated using a state update function. The inputs are the historical summary state of the previous time slice, the profile feature vector of the previous time slice, and the context vector of the previous time slice. The output is the historical summary state of the current time slice. The initial value of the historical summary state is determined by the system initialization configuration and is stored in conjunction with the student identifier.

[0031] S4. Align the historical summary state and the context vector into the context-aligned counterfactual digital twin model to obtain the expected profile features under the current context conditions and form the expected profile feature vector. Furthermore, the context-aligned counterfactual digital twin model in step S4 is used to predict desired profiling features based on historical summary states and context vectors. Its calculation process includes: The historical summary state of the current time slice is updated using a state update function based on the historical summary state of the previous time slice, the profile feature vector, and the context vector. The updated historical summary state and the current context vector are input into the conditional prediction function to obtain the desired profile feature vector.

[0032] Specifically, in step S4, the historical summary state is aligned with the context vector input context using a counterfactual digital twin model to obtain the expected profile features under the current context conditions, and an expected profile feature vector is formed. .

[0033] In this embodiment, the context-aligned counterfactual digital twin model module maintains a conditional prediction function for each student, and the conditional prediction function is based on historical summary states. With context vector As input, output the desired profile feature vector. The parameters of the conditional prediction function are updated driven by the stable-period sample set, which is indexed by time slice. This indicates that model updates are triggered by the monitoring server and record the model parameter version number and the index of the effective time slice.

[0034] The core calculation relationships described in this step are illustrated by the following formulas: ; in: Students In time slice The expected profile feature vector; Students The corresponding conditional prediction function, i.e., the prediction function of the context-aligned counterfactual digital twin model: Students In time slice Historical summary status: Students In time slice Context vector; Represents an individual student index; This represents the discrete time slice index.

[0035] Among them, historical summary status Used to represent a summary of a student's continuous behavior prior to the current time slice; context vector Used to characterize the calendar event information, course schedule information, time period information, and environmental information corresponding to the current time slice; conditional prediction function. Output the expected profile feature vector under the given context conditions.

[0036] S5. Calculate the counterfactual residuals based on the profiled feature vector and the expected profiled feature vector to obtain the counterfactual residual vector of the current time slice; Furthermore, the counterfactual residual in step S5 is calculated by the difference between the profiled feature vector and the expected profiled feature vector, and then standardized by the residual scaling vector to obtain the standardized counterfactual residual vector.

[0037] Specifically, in step S5, the counterfactual residual is calculated based on the profiled feature vector and the expected profiled feature vector to obtain the counterfactual residual vector for the current time slice. In this embodiment, the counterfactual residual calculation module performs a difference operation on the profile feature vector and the expected profile feature vector at the same time slice index t to obtain the counterfactual residual vector. and will It is stored in association with student identifiers and time slice indexes for subsequent drift evidence chain construction and risk inference.

[0038] The core calculation relationships described in this step are illustrated by the following formulas: ; in: Students In time slice counterfactual residual vector Students In time slice The profiled feature vector; Students In time slice The expected profile feature vector; Index representing individual students: This represents the discrete time slice index.

[0039] Among them, profile feature vector The counterfactual residual vector is formed by statistical feature extraction and de-identification aggregation of learning profile data, life profile data, and device profile data. Used for constructing subsequent drift evidence chains.

[0040] S6. Standardize the counterfactual residual vector and aggregate the standardized counterfactual residuals according to the preset feature domain partitioning method to obtain the domain residual strength of multiple feature domains. Specifically, in step S6, the counterfactual residual vector is standardized, and the standardized counterfactual residuals are aggregated according to a preset feature domain partitioning method to obtain the domain residual strengths of multiple feature domains.

[0041] In this embodiment, the drift evidence chain construction module performs standardization processing on the counterfactual residual vectors. The standardization process uses a student-level residual scaling parameter, which is statistically obtained from the counterfactual residual vectors within the stable-period sample set and updated as the stable-period sample set is updated. The standardized counterfactual residuals are allocated to G feature domains according to a preset feature domain partitioning method, with the feature domain partitioning indexed by dimension. It is stated that each They are mutually exclusive and cover all feature dimensions. For each feature domain g, a fixed intra-domain aggregation operator is performed on the absolute value of the standardized residuals within the domain to obtain the domain residual strength for that time slice. The type of the intra-domain aggregation operator is fixed and its version number is recorded during system initialization.

[0042] S7. Perform time-accumulated calculation of the domain residual strength within a preset time window, construct a drift evidence chain, and determine the consistency of the evidence chain based on the domain residual strength of multiple feature domains. Furthermore, the drift evidence chain in step S7 is constructed through the following steps: The standardized counterfactual residuals are divided into multiple feature domains according to a preset feature domain; The absolute value of the residual is aggregated within each feature domain to obtain the domain residual strength. The domain residual strength is calculated by weighted accumulation within a preset time window to form the domain chain strength. The consistency of the evidence chain is determined based on the domain chain strength of multiple feature domains, and a drift evidence chain is formed.

[0043] Specifically, in step S7, the residual strength of the domain is calculated over time within a preset time window to construct a drift evidence chain, and the consistency of the evidence chain is determined based on the domain chain strength of multiple feature domains.

[0044] In this embodiment, a domain residual strength sequence with a preset time window length W is maintained for each feature domain. The drift evidence chain construction module performs a time accumulation operation on the domain residual strengths within the window. The time accumulation operation uses a fixed time weight sequence, which is stored together with the window length W. The results of each feature domain after window accumulation constitute the drift evidence chain vector. Consistency of the chain of evidence

[0045] The domain threshold is obtained by comparing the domain chain strength in the drift evidence chain vector with the corresponding domain threshold. The domain threshold is fixed according to the feature domain configuration and is consistent with the feature domain division version.

[0046] The core calculation relationships described in this step are illustrated by the following formulas: ; ; ; in: Students In time slice The Domain chain strength of each feature domain; Students In time slice The Domain residual strength of each feature domain; Indicates time slice Relative to the current time slice Time weighting; Indicates the preset time window length; Students In time slice The drifting multi-evidence chain vector; Indicates the total number of feature domains; Students In time slice Consistency of the chain of evidence; Indicates the first The threshold values ​​corresponding to each feature domain; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This represents the index of a historical time slice located within a time window; Represents an individual student index; Indicates the current discrete time slice index; Indicates the feature domain index.

[0047] Among them, domain residual strength The time weights are calculated from the standardized counterfactual residuals within the corresponding feature domain using the in-domain aggregation operator. Used to characterize the contribution of different historical time slices to the current drift evidence chain; drift evidence chain vector Composed of the domain chain strength of all feature domains; consistency of the chain of evidence. Used to characterize the number of feature fields that reach the domain threshold.

[0048] S8. Calculate the risk intensity based on the drift evidence chain, evidence chain consistency, situation vector, and uncertainty index, and generate a mental health risk level according to the preset grading threshold. Furthermore, the uncertainty index in step S8 is calculated based on at least one of the following factors: Data completeness of profiled feature vectors; Prediction variance of context-aligned counterfactual digital twin models; Missing input data.

[0049] The risk intensity in step S8 is calculated using a risk inference function, which takes drift evidence chain, evidence chain consistency, scenario vector, and uncertainty index as inputs and outputs a continuous risk intensity value.

[0050] Specifically, in step S8, the risk intensity is calculated based on the drift evidence chain, evidence chain consistency, situation vector, and uncertainty index, and a mental health risk level is generated according to a preset grading threshold.

[0051] In this embodiment, the uncertainty index calculation module generates an uncertainty index based on data integrity. Data completeness is determined by the ratio of the number of valid feature dimensions in the current time slice to the total number of feature dimensions. The number of valid feature dimensions is obtained from missing label statistics. Uncertainty index Stored in conjunction with the time-slice index.

[0052] In this embodiment, the risk inference module receives the drift evidence chain vector. Consistency of the chain of evidence Context vector With uncertainty index The risk intensity is calculated according to the fixed risk inference function. Risk level Risk intensity With preset grading threshold The comparison revealed that the grading thresholds are fixed in the configuration and their version numbers are recorded. Both risk intensity and risk level are stored in a time-slice index.

[0053] S9. Calculate the remaining stable period based on the trend of risk intensity changes within the time window; Furthermore, the remaining stable period in step S9 is calculated based on the trend of risk intensity changes within the time window, and the calculation process includes: The slope of risk change is obtained by trend estimation of the risk intensity sequence within a time window. The remaining stable period is calculated based on the slope of risk change and the high-risk threshold.

[0054] Specifically, in step S9, the remaining stable period is calculated based on the changing trend of the risk intensity within the time window. .

[0055] In this embodiment, the remaining stable period calculation module calculates the remaining stable period when the remaining stable period output gating condition is met. The remaining stable period output gating condition is determined by comparing the uncertainty index with a gating threshold, where the gating threshold... Fixed storage. When the output gating condition is met, a trend estimation operation is performed on the risk intensity sequence within the time window. The trend estimation operation uses a fixed linear fitting process to obtain the slope of risk change. When the risk change slope is greater than zero, the remaining stable period is obtained by dividing the difference between the risk intensity and the high-risk threshold by the risk change slope; when the risk change slope is not greater than zero, the remaining stable period is recorded as unavailable (NA). The remaining stable period is stored as a number of time slices and bound to the time slice length Δ for conversion output.

[0056] The core calculation relationships described in this step are illustrated by the following formulas: ; in: Students In time slice The remaining stable period; Indicates a high-risk threshold; Students In time slice The intensity of the risk; Students Slope of risk changes within the time window; This represents the preset zero-reduction constant, and ; Represents an individual student index; This represents the index of the current discrete time slice.

[0057] Among them, the slope of risk change The risk intensity sequence within the time window is obtained through trend estimation; when When the value is greater than zero, the remaining stable period is calculated according to the above formula; when... When the value is not greater than zero, the remaining stable period is marked as unavailable. The remaining stable period is used to characterize the number of time slices remaining before reaching the high-risk threshold under the current risk trend.

[0058] S10. When the risk intensity meets the active micro-testing triggering condition, perform active micro-testing and obtain subjective state information, and update the stable period sample set or risk inference parameters based on the subjective state information. Furthermore, the proactive micro-test in step S10 is achieved by sending a brief psychological state inquiry to the student's terminal and obtaining subjective state information, and then performing at least one of the following update operations based on the subjective state information: Update the stable period sample set; Update the parameters of the risk inference function; Update the risk classification threshold.

[0059] Specifically, in step S10, when the risk intensity meets the active micro-test triggering condition, active micro-testing is performed and subjective state information is obtained, and the stable period sample set or risk inference parameters are updated according to the subjective state information.

[0060] In this embodiment, the proactive quiz module sends a proactive quiz request to the student terminal based on the proactive quiz trigger conditions. The proactive quiz request includes a question identifier, a set of options, and a timeout parameter; the student terminal returns subjective status information. The monitoring server maps according to a fixed threshold rule. Mapped to discretized subjective state levels The set of discretized subjective state level values ​​is as follows The stable period sample set is indexed by time slice. It means that when Time slice t will be included when the preset inclusion criteria are met. ,when When the preset exclusion conditions are met, time slice t is changed from... Exclude and record the exclusion flag. After the sample set is updated during the stabilization period, the monitoring server updates the parameters of the scenario-aligned counterfactual digital twin model according to the fixed training trigger rules and records the model parameter version number.

[0061] In this embodiment, the active micro-testing trigger condition is determined by the risk intensity and uncertainty index, and the trigger criterion is fixed and stored in the configuration.

[0062] The core calculation relationships described in this step are illustrated by the following formulas: ; in: Students In time slice Active micro-test trigger identifier; This represents the lower threshold of medium risk in risk classification; This represents the high-risk threshold in the risk classification; Students In time slice The intensity of the risk; Students In time slice Uncertainty index; Indicates the active micro-test trigger threshold; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Represents an individual student index; This represents the index of the current discrete time slice.

[0063] Among them, when When the active quiz module initiates an active quiz request to the student's terminal; when At that time, active micro-testing is not performed.

[0064] In this embodiment, updating the risk inference parameters includes at least one of the following: updating the risk inference function parameters and updating the risk classification threshold. , Or update the feature domain threshold θ g Parameter updates are performed using a structured configuration versioning approach. After the update, a new version number is generated and bound to the effective time slice index. The monitoring server retains the old and new version numbers and records the effective range of the version.

[0065] S11. Output includes monitoring results of mental health risk level, uncertainty index, drift evidence chain information, and remaining stable period.

[0066] Specifically, step S11 outputs monitoring results including mental health risk level, uncertainty index, drift evidence chain information, and remaining stable period.

[0067] In this embodiment, the monitoring result output module generates a structured output record, which includes at least: student identifier. Time slice index Risk intensity Risk level Uncertainty index Drift evidence chain vector Consistency of the chain of evidence Remaining stable period Or the NA marker and active microprobe trigger marker may not be available. Model parameter version number and configuration version number. Output records are written to the database and output through the management interface.

[0068] In this embodiment, the monitoring server associates and stores the profiled feature vector for each output record. Context vector Expected profile feature vector With counterfactual residual vector Log identifiers are assigned to records, and these log identifiers, along with model parameter version numbers and configuration version numbers, are used for audit queries and consistency checks.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically monitoring the mental health status of college students using profiling data, characterized in that, Includes the following steps: S1. Collect multi-source profile data of college students in discrete time slices, and perform statistical feature extraction and desensitization aggregation on the multi-source profile data to generate profile feature vectors for the corresponding time slices. S2. Obtain the context information corresponding to the time slice, and encode the context information to generate the context vector corresponding to the time slice; S3. Based on the profile feature vector and context vector of historical time slices, update the historical summary status of individual students through the state update function; S4. Align the historical summary state with the context vector input context to a counterfactual digital twin model to obtain the expected profile features under the current context conditions, and form the expected profile feature vector. S5. Calculate the counterfactual residual based on the profile feature vector and the expected profile feature vector to obtain the counterfactual residual vector of the current time slice; S6. The counterfactual residual vector is standardized, and the standardized counterfactual residuals are aggregated according to the preset feature domain partitioning method to obtain the domain residual strength of multiple feature domains. S7. Perform time-accumulated calculation on the domain residual strength within a preset time window to construct a drift evidence chain, and determine the consistency of the evidence chain based on the domain residual strength of multiple feature domains. S8. Calculate the risk intensity based on the drift evidence chain, evidence chain consistency, situation vector, and uncertainty index, and generate a mental health risk level according to a preset grading threshold. S9. Calculate the remaining stable period based on the changing trend of the risk intensity within the time window; S10. When the risk intensity meets the active micro-test triggering condition, perform active micro-test and obtain subjective state information, and update the stable period sample set or risk inference parameters according to the subjective state information. S11. Output includes monitoring results of mental health risk level, uncertainty index, drift evidence chain information, and remaining stable period.

2. The method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The profiling feature vector in step S1 is generated through the following steps: Align multi-source profiled data according to time slices; Execute the feature mapping function for each type of data source to obtain the corresponding data source feature vector; The feature vectors from various data sources are concatenated to form a profile feature vector of uniform dimension, and the profile feature vector is then normalized.

3. The method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The context vector in step S2 is generated through a context encoding function, and the context information includes at least one of the following: School calendar event information, course schedule information, time period information, and environmental information; The context encoding function converts the context information into a numerical vector and forms the context vector.

4. The method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The context-aligned counterfactual digital twin model in step S4 is used to predict desired profiling features based on historical summary states and context vectors. Its calculation process includes: The historical summary state of the current time slice is updated using a state update function based on the historical summary state of the previous time slice, the profile feature vector, and the context vector. The updated historical summary state and the current context vector are input into the conditional prediction function to obtain the desired profile feature vector.

5. The method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The counterfactual residual in step S5 is calculated by the difference between the profiled feature vector and the expected profiled feature vector, and then standardized by the residual scaling vector to obtain the standardized counterfactual residual vector.

6. The method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The drift evidence chain in step S7 is constructed through the following steps: The standardized counterfactual residuals are divided into multiple feature domains according to a preset feature domain; The absolute value of the residual is aggregated within each feature domain to obtain the domain residual strength. The domain residual strength is weighted and accumulated within a preset time window to form the domain chain strength; The consistency of the evidence chain is determined based on the domain chain strength of multiple feature domains, and a drift evidence chain is formed.

7. The method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The uncertainty index in step S8 is calculated based on at least one of the following factors: Data completeness of profiled feature vectors; Prediction variance of context-aligned counterfactual digital twin models; Missing input data.

8. The method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The risk intensity in step S8 is calculated by a risk inference function, which takes drift evidence chain, evidence chain consistency, scenario vector and uncertainty index as inputs and outputs a continuous risk intensity value.

9. A method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The remaining stable period in step S9 is calculated based on the trend of risk intensity changes within a time window, and the calculation process includes: The slope of risk change is obtained by trend estimation of the risk intensity sequence within a time window. The remaining stable period is calculated based on the slope of the risk change and the high-risk threshold.

10. A method for dynamically monitoring the mental health status of college students using profiling data according to claim 1, characterized in that, The active micro-test in step S10 is achieved by sending a brief psychological state inquiry to the student's terminal and obtaining subjective state information, and then performing at least one of the following update operations based on the subjective state information: Update the stable period sample set; Update the parameters of the risk inference function; Update the risk classification threshold.

Images