Mental health group screening system based on behavior data feature fusion

Abstract

The invention relates to the technical field of mental health monitoring and intelligent information processing, in particular to a mental health group screening system based on behavior data feature fusion, which comprises a data acquisition terminal, a feature fusion server and a monitoring and early warning terminal, collecting multi-dimensional behavior sequence data of the target object including a first type of feature dimensions and a second type of feature dimensions; the server calculates a probability distribution distance in combination with historical baseline behavior data, and generates a feature drift index, a disguise probability value and a data deviation entropy value; when the entropy value is lower than a safety threshold value, a first feature fusion strategy is adopted to generate a first state evaluation result, and when the entropy value is not lower than the safety threshold value, a suboptimal degradation fusion strategy is triggered to generate a second state evaluation result; the monitoring and early warning terminal generates an intervention instruction based on the evaluation result and feeds back and displays the instruction; the present invention can maintain screening availability during deliberate disguise, privacy avoidance, or noise interference.

Description

Technical Field

[0001] This invention relates to the field of mental health monitoring and intelligent information processing technology, specifically a mental health group screening system based on behavioral data feature fusion. Background Technology

[0002] Mental health screening technology is a common method for collecting, analyzing, and assessing the behavioral data of target groups. It can combine multi-dimensional behavioral characteristics such as social interaction text and device usage rhythm to continuously identify the psychological risks of groups such as students and employees. It has the technical characteristics of large-scale concurrent processing and remote automated data monitoring. With the development of digital health management, mental health screening has gradually evolved from single questionnaire assessment to intelligent screening based on multi-source behavioral data, and improved the state recognition ability through machine learning models.

[0003] However, when conducting continuous mental health screenings on large groups, existing technologies are still susceptible to the decline in data credibility, resulting in insufficient stability of screening results. On the one hand, target subjects may deliberately disguise their behavioral patterns, disable authorization for highly sensitive data, or actively generate noise, causing fine-grained features to be distorted but still included in the fusion calculation, leading to misjudgments. On the other hand, existing solutions usually lack the ability to automatically switch fusion strategies and provide system failure warnings when data credibility changes, making it difficult to maintain basic screening availability when privacy features are unavailable or input data is continuously abnormal. Summary of the Invention

[0004] The purpose of this invention is to provide a mental health group screening system based on behavioral data feature fusion, and to solve the following technical problems:

[0005] The key focus is on addressing the usability issues of the screening system when the screened population exhibits non-authentic behavioral trajectory manipulation, disables authorization for highly sensitive data, or actively generates noise. By transforming feature distribution drift into an executable data deviation entropy value determination, and dynamically switching between the full-dimensional first feature fusion strategy and the suboptimal degradation fusion strategy accordingly, the system can effectively prevent the loss of decision confidence when data reliability deviates. This achieves a dynamic balance between short-term early warning capabilities and the maintenance of a long-term data confidence pool, and enables more robust identification and continuous early warning of mental health risks.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A mental health screening system based on behavioral data feature fusion includes: a data acquisition terminal, a feature fusion server, and a monitoring and early warning terminal;

[0008] The data acquisition terminal is used to acquire multidimensional behavior sequence data of the target object and send the multidimensional behavior sequence data to the feature fusion server. The multidimensional behavior sequence data includes a first type of feature dimension pre-labeled as involving privacy and a second type of feature dimension pre-labeled as public.

[0009] The feature fusion server is used to obtain the historical baseline behavior data of the target object from a pre-configured database, generate a feature drift index by calculating the probability distribution distance between the historical baseline behavior data and the multidimensional behavior sequence data, input the feature drift index into a preset probability evaluation function for normalization mapping, calculate the spoofing probability value of the multidimensional behavior sequence data in the interval of 0 to 1, and use a preset information entropy calculation rule to multiply the spoofing probability value and the opposite probability corresponding to the spoofing probability value with their corresponding logarithms, and then perform a summation and inversion operation to convert it into the data deviation entropy value of the target object.

[0010] The feature fusion server is further configured to process the multidimensional behavior sequence data using a first feature fusion strategy when the data deviation entropy value is lower than a preset safety threshold, so as to generate a first state evaluation result. The first feature fusion strategy includes: obtaining a preset first feature weight set, using the first feature weight set to perform weighted concatenation of the multidimensional behavior sequence data to generate a full-dimensional fusion feature matrix, and inputting the full-dimensional fusion feature matrix into a pre-trained machine learning state classification model to output the first state evaluation result.

[0011] The feature fusion server is further configured to trigger a suboptimal degradation fusion strategy to process the multidimensional behavior sequence data when the data deviation entropy value is not lower than the preset safety threshold, so as to generate a second state evaluation result. The triggering of the suboptimal degradation fusion strategy includes: resetting the weights corresponding to the first type of feature dimension in the first feature weight set to zero to generate a second feature weight set; using the second feature weight set to extract the second type of feature dimension by weighting to generate a dimension-reduced macro feature matrix; and inputting the dimension-reduced macro feature matrix into the pre-trained machine learning state classification model to output the second state evaluation result.

[0012] The feature fusion server is also used to send the first state evaluation result or the second state evaluation result to the monitoring and early warning terminal;

[0013] The monitoring and early warning terminal is used to generate intervention instructions based on the first state assessment result or the second state assessment result, and send them to the data acquisition terminal for feedback display.

[0014] Optionally, the feature fusion server is also used for:

[0015] Obtain the historical data deviation entropy value sequence of the target object within a preset time window;

[0016] When the number of effective entropy points in the historical data deviation entropy value sequence reaches a preset number, based on the values ​​of each time node in the historical data deviation entropy value sequence, a linear fitting algorithm is used to calculate the slope of the historical data deviation entropy value sequence as a function of time, and the absolute value of the slope is used as the system confidence decay rate.

[0017] In response to the system confidence decay rate exceeding a preset risk threshold, a system failure warning signal is sent to the monitoring and early warning terminal; or,

[0018] If the system confidence decay rate is not greater than the preset risk threshold, the current feature fusion strategy is maintained.

[0019] Optionally, the system may also include: a policy configuration terminal;

[0020] The monitoring and early warning terminal is also used to send a policy evaluation request to the policy configuration terminal;

[0021] The policy configuration terminal is used to update the first feature fusion policy and the suboptimal degradation fusion policy stored in the feature fusion server based on the policy evaluation request.

[0022] Optionally, the policy configuration terminal is further configured to: send an access request carrying the account information of the policy configuration terminal to the feature fusion server;

[0023] The feature fusion server is further configured to: determine whether the policy configuration terminal has access to the feature fusion server, and if the policy configuration terminal has access to the feature fusion server, process the access request of the policy configuration terminal;

[0024] Alternatively, if the policy configuration terminal does not have access to the feature fusion server, the access request of the policy configuration terminal may be denied.

[0025] Optionally, the feature fusion server is also used for:

[0026] Extract the account information of the policy configuration terminal from the access request of the policy configuration terminal;

[0027] Obtain pre-stored reference account information from the feature fusion server;

[0028] The reference account information is compared with the account information of the policy configuration terminal;

[0029] In response to the inconsistency between the reference account information and the account information of the policy configuration terminal, it is determined that the policy configuration terminal does not have access to the feature fusion server; or,

[0030] In response to the fact that the reference account information matches the account information of the policy configuration terminal, it is determined that the policy configuration terminal has access to the feature fusion server.

[0031] Optionally, the data acquisition terminal includes:

[0032] Interactive log acquisition module and sensor status monitoring module;

[0033] The interaction log collection module is used to collect social network interaction text data of the target object;

[0034] The sensor status monitoring module is used to collect the device usage rhythm data of the target object;

[0035] The social network interaction text data corresponds to the first type of feature dimension, and the device usage rhythm data corresponds to the second type of feature dimension. Together, they constitute the multidimensional behavioral sequence data.

[0036] Optionally, the monitoring and early warning terminal is also used for:

[0037] The intervention commands and system failure warning signals are displayed visually and broadcast via voice.

[0038] The beneficial effects of this invention are:

[0039] 1. This invention calculates the feature drift index of multidimensional behavioral data and converts it into data deviation entropy value. When the data confidence decreases, it automatically triggers a suboptimal degradation fusion strategy. The weights of features involving privacy dimensions are reset to zero, and only public features are used for dimensionality reduction evaluation. This effectively avoids system misjudgment caused by the disguised trajectory or deliberate modification of the target object, and maintains the basic availability of the screening system when sensitive data is unavailable.

[0040] 2. This invention acquires historical data deviation entropy value sequences and calculates the system confidence decay rate, enabling real-time monitoring of the decay trend of the system confidence environment. When the decay rate exceeds the risk threshold, it promptly triggers a system failure warning signal and broadcasts it through multiple channels, effectively reminding the intervention network to pay attention to group data distortion and preventing the screening system from continuously outputting erroneous conclusions when it loses the confidence basis to support decision-making.

[0041] 3. This invention introduces a policy configuration terminal and access control mechanism. The monitoring terminal can initiate policy evaluation requests to dynamically update the first feature fusion policy and the downgrade policy. Combined with strict reference account information comparison and authentication, it not only realizes the adaptive evolution of feature weight rules following the real feedback of the group, but also ensures the controllability of high-risk configuration operations and prevents external interference from maliciously modifying the warning threshold.

[0042] 4. This invention acquires social interaction text data through an interaction log collection module and combines it with device usage rhythm data acquired through a sensor status monitoring module to build a complementary data foundation. This not only provides detailed expression layers for full-dimensional fusion but also preserves objective behavioral rhythm traces that are difficult to deliberately modify in the downgraded fusion mode, thereby enhancing the anti-modification ability of group psychological risk screening from the data source. Attached Figure Description

[0043] The invention will now be further described with reference to the accompanying drawings.

[0044] Figure 1 A schematic diagram of a group mental health screening system based on behavioral data feature fusion provided in this application embodiment. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Please see Figure 1 A mental health group screening system based on behavioral data feature fusion includes: a data acquisition terminal, a feature fusion server, and a monitoring and early warning terminal; wherein, the data acquisition terminal is used to acquire multidimensional behavioral sequence data of the target object and send the multidimensional behavioral sequence data to the feature fusion server, the multidimensional behavioral sequence data including a first type of feature dimension pre-labeled as involving privacy and a second type of feature dimension pre-labeled as public;

[0047] The feature fusion server is used to obtain the historical baseline behavior data of the target object from a pre-configured database, generate a feature drift index by calculating the probability distribution distance between the historical baseline behavior data and the multidimensional behavior sequence data, input the feature drift index into a preset probability evaluation function for normalization mapping, calculate the spoofing probability value of the multidimensional behavior sequence data in the interval of 0 to 1, and use a preset information entropy calculation rule to multiply the spoofing probability value and the opposite probability corresponding to the spoofing probability value with their corresponding logarithms, and then perform a summation and inversion operation to convert it into the data deviation entropy value of the target object.

[0048] The feature fusion server is further configured to process the multidimensional behavior sequence data using a first feature fusion strategy when the data deviation entropy value is lower than a preset safety threshold, so as to generate a first state evaluation result. The first feature fusion strategy includes: obtaining a preset first feature weight set, using the first feature weight set to perform weighted concatenation of the multidimensional behavior sequence data to generate a full-dimensional fusion feature matrix, and inputting the full-dimensional fusion feature matrix into a pre-trained machine learning state classification model to output the first state evaluation result.

[0049] The feature fusion server is further configured to trigger a suboptimal degradation fusion strategy to process the multidimensional behavior sequence data when the data deviation entropy value is not lower than the preset safety threshold, so as to generate a second state evaluation result. The triggering of the suboptimal degradation fusion strategy includes: resetting the weights corresponding to the first type of feature dimension in the first feature weight set to zero to generate a second feature weight set; using the second feature weight set to extract the second type of feature dimension by weighting to generate a dimension-reduced macro feature matrix; and inputting the dimension-reduced macro feature matrix into the pre-trained machine learning state classification model to output the second state evaluation result.

[0050] The feature fusion server is also used to send the first state assessment result or the second state assessment result to the monitoring and early warning terminal; the monitoring and early warning terminal is used to generate an intervention command based on the first state assessment result or the second state assessment result, and send it to the data acquisition terminal for feedback display.

[0051] This embodiment provides a psychological health group screening mechanism based on behavioral data feature fusion. Specifically, this embodiment takes a regional public psychological health crisis intervention network as the main scenario. The network covers a large university town and its surrounding industrial parks. During the preset peak period of group psychological load, the preset target group is continuously assessed for psychological health risk status.

[0052] The entire system consists of data collection terminals deployed on personal mobile devices, feature fusion servers deployed in medical information centers, and monitoring and early warning terminals deployed at regional psychological crisis intervention command posts. It is designed to address the issue of maintaining the availability of the screening system even when groups exhibit non-authentic behavioral trajectory manipulation, close authorization for highly sensitive data, or actively generate noise.

[0053] The details are as follows: The data acquisition terminal continuously acquires multidimensional behavioral sequence data of the target object, and pre-labels the feature dimensions before acquisition; the first type of feature dimensions involving privacy may include fine-grained interactive text at the individual level, details of device usage at night, and high-frequency behavioral trajectories within a continuous time period; the second type of publicly disclosed feature dimensions may include device activity rhythms with global statistical features, frequency of public social interactions, and distribution of application access categories.

[0054] This pre-labeling is not only used for classification and storage, but also to reserve controllable boundaries for subsequent fusion strategy switching; because in the state assessment scenario, the granularity of the feature dimension and the correlation of the implicit state representation are positively correlated with the sensitivity of short-term state recognition and the probability of the target object triggering privacy avoidance strategies.

[0055] The feature fusion server retrieves the historical baseline behavior data of the target object from a pre-configured database; the historical baseline behavior data refers to the distribution of behavior features that have passed quality verification during a relatively stable period, and is used to describe the long-term habits of the object rather than occasional fluctuations at a certain point in time.

[0056] For example, if a target user's nighttime device usage frequency is lower than a preset activity threshold, the number of words in their social media texts is lower than a preset length range, and their weekday routine is regular, then this set of long-term stable patterns can constitute their individual baseline; the server compares the distribution differences between the currently collected multidimensional behavioral sequence data and this baseline to generate feature drift indicators;

[0057] The calculation of distribution differences and probability distribution distance can be carried out by using algorithms such as KL divergence, JS divergence or Wasserstein distance. By quantifying the degree of difference between the current feature distribution and the baseline feature distribution, a specific distance value is calculated as a feature drift index.

[0058] The drift here is not simply equivalent to an anomaly. Its engineering meaning is to determine whether the current behavioral change is a natural deviation that conforms to the real stress response, or whether there are signs of artificial normalization, unrealistic modification, or noise injection.

[0059] For example, when the actual psychological load increases, multiple dimensions often show consistent changes, while non-realistic modifications often show a consistent and abnormal preference distribution in some highly sensitive dimensions, but other macro rhythms do not recover synchronously.

[0060] The server inputs the drift index into a preset probability evaluation function to obtain the spoofing probability value, and then converts it into a data deviation entropy value according to the information entropy calculation rules; the data deviation entropy value here can be understood as a quantitative indicator of whether the current input data still has a high system confidence level.

[0061] Specifically, the preset probability evaluation function can be a Sigmoid function or an exponential normalization function, which smoothly maps the unbounded feature drift index to the interval between 0 and 1, and uses this as the spoofing probability value; the information entropy calculation rule specifically adopts the Shannon information entropy formula, which multiplies the spoofing probability value and the opposite probability corresponding to the spoofing probability value by their corresponding logarithms, and then performs summation and inversion operations to finally obtain the quantified data deviation entropy value, thereby objectively reflecting the uncertainty of data distribution;

[0062] If the value is low, it indicates that although the current collected data has changed from the historical baseline, the changes are interpretable and consistent with each other, making it suitable to continue using full-dimensional fusion; if the value is high, it indicates that the internal consistency of the data has decreased, which may have been affected by structurally abnormal inputs, privacy avoidance, or group-based unrealistic modifications.

[0063] The aforementioned preset safety thresholds and the preset risk thresholds mentioned later are all pre-configured in the system by statistically analyzing the data deviation entropy and decay rate of a group of the same size in the region during a historical normal and stable period, extracting the upper boundary value or empirical critical value of its confidence interval.

[0064] When the data deviation entropy value is lower than the preset safety threshold, the server adopts the first feature fusion strategy; specifically, the server reads the first feature weight set, performs weighted splicing on the multi-dimensional behavioral sequence data, and generates a full-dimensional fusion feature matrix.

[0065] The preset first feature weight set can be pre-calculated and allocated by performing feature correlation analysis on historical group sample data, or initialized by clinical psychology experts based on the importance of psychological crisis manifestations of each feature dimension.

[0066] For ease of explanation, a simplified schematic can be used to represent this: Assume the current object has four feature slices F1, F2, F3, and F4, where F1 and F2 belong to the first type of feature dimension, and F3 and F4 belong to the second type of feature dimension; if the first feature weight set is {0.35, 0.25, 0.20, 0.20}, it means that the system allows fine-grained privacy features to participate in the judgment under a high confidence state;

[0067] The fused feature matrix obtained after splicing retains the relative contributions of different dimensions. It is then input into a pre-trained machine learning state classification model and outputs the first state evaluation result, such as low risk requiring tracking and high risk requiring manual review. At this time, the model pays more attention to individual-level details, which is conducive to improving the sensitivity of short-term early warning.

[0068] The pre-trained machine learning state classification model can specifically adopt classification algorithm models such as random forest, support vector machine or multilayer perceptron; its pre-training process includes: acquiring a sample dataset containing a large amount of multidimensional behavioral sequence data of target objects and corresponding expert-annotated psychological state results, inputting the sample dataset into the initial classification model for forward propagation and backpropagation parameter optimization to minimize the classification loss between the predicted output and the real annotation, until the model converges to reach usable accuracy.

[0069] When the data deviation entropy value is not lower than the preset safety threshold, the system no longer adheres to the high-sensitivity full-feature scheme, but instead triggers a suboptimal degradation fusion strategy; specifically, the server resets the weights corresponding to the first type of feature dimension in the first feature weight set to zero, and generates the second feature weight set;

[0070] Taking the above four feature slices as an example, if the weights become {0, 0, 0.20, 0.20} after resetting, it means that the server actively abandons the fine-grained privacy features and only retains the public and more macroscopic second type of feature dimension, forming a dimensionality-reduced macroscopic feature matrix;

[0071] The matrix is ​​then input into the same machine learning state classification model, and the output is the second state evaluation result. The suboptimal degradation fusion strategy still retains the preset basic classification ability, but maintains the basic screening ability of the system at the cost of reducing the local recognition precision when the confidence decreases, and avoids continuing to use highly sensitive features that may have already undergone distribution shift, which would aggravate misjudgment.

[0072] After receiving the first or second state assessment results, the monitoring and early warning terminal generates an intervention instruction. This intervention instruction may include: sending a self-service psychological assessment portal to the data collection terminal, suggesting an appointment for psychological counseling, initiating a class counselor review prompt, or triggering a regional hotline follow-up task.

[0073] The feedback displayed here can be directed to the target individual or to authorized intervention personnel. Its configuration logic avoids directly outputting absolute status judgment classifications. Instead, it dynamically matches the corresponding intervention execution strategy based on the data confidence level of the status assessment results. When the confidence level is high, fine-grained individual-level interventions can be carried out. When the confidence level is low, low-invasive and low-stimulation universal support measures are prioritized.

[0074] In boundary cases, if the data acquisition terminal only acquires the first type of feature dimension and lacks the second type of feature dimension in a certain sampling period, the server will not execute the downgraded fusion result output for the time being. Instead, it will mark the period as having incomplete information structure and switch to waiting for supplementary acquisition or use the second type of feature dimension of the most recent complete period for short-term replacement. If only the second type of feature dimension is acquired, the downgraded fusion process can be directly entered.

[0075] If historical baseline behavioral data is missing, for example, if newly included monitoring subjects do not have sufficient historical records, the server will temporarily use the group reference baseline of the same age group and the same management scenario as a substitute, and gradually switch to the individual baseline after subsequent continuous collection reaches the set window; if the current data deviation entropy value is exactly in the vicinity of the threshold, in order to avoid frequent switching of fusion strategy, a buffer can be set, and the strategy can be switched after maintaining the same direction of change for several consecutive cycles.

[0076] On the third day after the start of the exam week in the university town, the frequency of public social interactions of a graduate student target suddenly increased. The emotional polarity score of the nighttime text expression was higher than the baseline average, but the device usage rhythm still showed continuous screen-on in the early morning and increased fluctuations in sleep start and end times during the week. After comparing the current behavior with the baseline of the previous month, the server identified an inconsistency between the trend of positive social network interaction text data and the failure of device usage rhythm data to reflect the true physiological state, forming a high probability of spoofing, and calculated a high data deviation entropy value.

[0077] The system therefore no longer relies on the social network interaction text data, but only retains the second type of feature dimension such as the device's rhythm data for downgraded fusion, and outputs the second state evaluation result that needs to be tracked; the monitoring and early warning terminal generates low-intrusion intervention instructions accordingly, such as showing the campus psychological service entrance and anonymous consultation channel to its terminal, without triggering the push of the warning information of the preset highest intervention level, thereby reducing the possibility of it further avoiding system collection;

[0078] Furthermore, to ensure that both the full-dimensional fusion feature matrix and the dimensionality-reduced macroscopic feature matrix can be input into the same pre-trained machine learning state classification model, the server adopts a unified input position template during the model training phase.

[0079] Specifically, each feature dimension corresponds to a fixed input slot in the template; under the first feature fusion strategy, both the first and second type feature dimensions are written into their corresponding slots; under the suboptimal degradation fusion strategy, the first type feature dimension, which is reset to zero weight, retains its original slot, but its slot value is set to zero or to the numerical representation corresponding to the preset masking mark, while the second type feature dimension continues to be written into its original slot.

[0080] Thus, although a dimension-reduced macroscopic feature matrix is ​​formed semantically, the matrix length, field order, and positional encoding relationship remain consistent with the full-dimensional fusion feature matrix at the model input interface level, thereby avoiding the problem that the model cannot handle due to changes in the input dimension; accordingly, the aforementioned weighted concatenation and weighted extraction are performed according to the rules under the unified template, where extraction emphasizes that only the second type of feature dimension participates in the effective judgment, rather than changing the model interface format;

[0081] Furthermore, the generation of the spoofing probability value and the data deviation entropy value follows a deterministic processing order: first, the feature drift index is obtained based on the distribution difference between historical baseline behavior data and current multidimensional behavior sequence data; then, the feature drift index is mapped to the spoofing probability value; and finally, the spoofing probability value is converted into the data deviation entropy value according to the preset information entropy calculation rules.

[0082] The output of the previous stage serves as the input of the next stage, and there is no situation where the state evaluation result is used in reverse to participate in the calculation of the current cycle entropy value. This ensures that the strategy switching decision is based on the reliability of the collected data itself, rather than on the cyclical inference based on the posterior intervention result.

[0083] Furthermore, to avoid name drift of the same technical object in the whole text due to contextual description, the descriptions of device activity rhythm, device usage rhythm, public interaction rhythm, public social interaction frequency, and macro rhythm characteristics in this paper are all illustrative sub-items or derived representations of the second type of feature dimension. When performing weight allocation, weighted extraction and downgrade fusion, they are uniformly processed according to the second type of feature dimension.

[0084] Correspondingly, descriptions such as fine-grained interactive text, nighttime text sentiment, text expression features, and highly sensitive text content are all illustrative sub-items or derived representations of the first type of feature dimension. They can participate in full-dimensional fusion under high confidence conditions, and their corresponding weights are reset to zero when the suboptimal degradation fusion strategy is triggered. Therefore, the different terms mentioned above are used to describe the data source or manifestation and do not indicate the introduction of new feature categories.

[0085] Furthermore, the aforementioned full-dimensional fusion feature matrix and dimensionality-reduced macroscopic feature matrix correspond to the effective participation range under the first feature fusion strategy and the effective participation range under the suboptimal degradation fusion strategy, respectively. The semantic distinction between the two is whether the feature set participating in the judgment includes the first type of feature dimension, rather than indicating that the model input interface adopts two different formats.

[0086] Specifically, dimensionality reduction means that the effective information dimension is shrunk to the second type of feature dimension. Macroscopically, it means that the feature semantics involved in the judgment are more inclined to public and rhythmic behavioral representations. Under the unified input position template, the first type of feature dimension that is blocked still retains a fixed slot but does not participate in the effective judgment.

[0087] This approach maintains consistency with weighted extraction to generate a reduced-dimensional macroscopic feature matrix while avoiding the misinterpretation that a completely new model input object has been constructed that is entirely different from the full-dimensional fused feature matrix.

[0088] The purpose of this step is to transform the question of whether to continue to trust fine-grained, highly sensitive features when data reliability deviates into an executable entropy value determination and fusion strategy switching process, thereby achieving a dynamic balance between short-term early warning capabilities and long-term data confidence pool maintenance.

[0089] In a preferred embodiment of the present invention, the feature fusion server is further configured to: obtain a historical data deviation entropy value sequence of the target object within a preset time window; when the number of effective entropy value points in the historical data deviation entropy value sequence reaches a preset number, calculate the slope of the historical data deviation entropy value sequence as a function of time based on the values ​​of each time node in the historical data deviation entropy value sequence, and use the absolute value of the slope as the system confidence decay rate;

[0090] If the system confidence decay rate is greater than a preset risk threshold, a system failure warning signal is sent to the monitoring and early warning terminal; or if the system confidence decay rate is not greater than the preset risk threshold, the current feature fusion strategy is maintained.

[0091] This embodiment provides a system failure early warning mechanism for scenarios where data confidence is continuously declining. Specifically, in the aforementioned scheme, the data deviation entropy value of a single sampling period can only reflect the data confidence at a certain point in time. However, in regional psychological screening, the real danger is not a single instance of data distortion, but rather the continuous spread and accelerated accumulation of distortion.

[0092] If the fusion strategy is switched based solely on the single-point entropy value, two problems may arise: First, the system may frequently oscillate between full-dimensional fusion and downgraded fusion; second, the system cannot determine whether the current confidence deviation is approaching the overall failure boundary. Therefore, this embodiment further introduces the historical data deviation entropy value sequence and the system confidence decay rate to evaluate the decay rate of data confidence.

[0093] The details are as follows: The feature fusion server continuously records the data deviation entropy value of the target object within a preset time window, forming a historical data deviation entropy value sequence; this time window can be set at the hourly, daily, or shift management cycle; in engineering, this sequence does not reflect the psychological state itself, but rather the trajectory of the system's confidence in the input data of the object; if the trajectory is stable, it means that although there are fluctuations in individual cycles, the overall input can still be explained.

[0094] If the trajectory shows a monotonically increasing or monotonically decreasing trend, and the absolute value of the slope of change is greater than the preset mutation threshold, it indicates that the data source environment has undergone structural changes, such as the emergence of monitoring avoidance, the presentation of highly consistent abnormal feature distribution, or the centralized closure of feature authorization within the group.

[0095] The server calculates the slope of the sequence over time using a linear fitting method and takes its absolute value as the system confidence decay rate. The reason for using the absolute value of the slope is that the system focuses on whether the confidence structure changes rapidly, rather than being limited to changing in a fixed direction.

[0096] For example, if the entropy values ​​of an object at four consecutive time points are 0.20, 0.23, 0.25, and 0.27, it indicates that the deviation is gradually increasing.

[0097] If the values ​​are 0.52, 0.40, 0.29, and 0.18, it indicates that the system has suddenly dropped from a low confidence state to a high confidence state. This may also mean that the terminal has been changed, the sampling method has changed, or discontinuous data has appeared in the short term. It does not necessarily mean that the risk has been eliminated. Therefore, the rate of change still needs to be monitored. Linear fitting mainly plays the role of suppressing single-point noise and extracting the trend direction.

[0098] When the system confidence decay rate exceeds a preset risk threshold, the server sends a system failure warning signal to the monitoring and early warning terminal. This warning signal does not mean that a certain object is necessarily in an extreme psychological crisis, but rather it prompts the intervention network to indicate that the data confidence environment for this object or this type of object is rapidly declining, and the stability of the original screening conclusions is insufficient.

[0099] The proportion of manual review should be increased, the reliance on automatic decision-making should be reduced, and if necessary, a switch to universal intervention at the group level should be made. Conversely, when the decay rate is not greater than the risk threshold, it indicates that the existing fusion strategy is still applicable and the system can maintain the current strategy.

[0100] As a supplementary implementation, if the number of effective entropy points within the preset time window is insufficient, for example, due to the access of new objects, interruption of data upload, or long-term offline status of the terminal resulting in a short time series, the decay rate is not directly calculated. Instead, the object is marked as having insufficient trend information, the most recently determined fusion strategy is maintained, and a prompt is given to continue collecting data.

[0101] If the number of consecutive missing data periods within the time window exceeds the preset span, such as multiple consecutive missing sampling points, the missing segments are marked intermittently before fitting to avoid misjudging interruptions as trend transitions; if the decay rate is close to the risk threshold but has not been stably exceeded, a dual-period confirmation mechanism can be set to reduce false system failure warnings caused by occasional events.

[0102] In the aforementioned scenario of joint screening between university towns and industrial parks, after the start of exam week, a group of students from a certain department collectively spread a counter-strategy on social media platforms, using the unified release of positive content to circumvent system monitoring; the data deviation entropy value of a certain target gradually increased from a stable state over a continuous 6-hour period, and the upward trend continued.

[0103] After the server extracts the entropy value sequence within the time window, it identifies that the confidence decay rate has exceeded the risk threshold. Therefore, in addition to retaining the downgraded fusion result, it also sends a system failure warning signal to the monitoring and early warning terminal. After receiving the signal, the command center no longer focuses solely on individual screening results, but judges that the department may have a group of structural abnormal inputs. It then arranges for counselors to provide non-labeled psychological support and appropriately extends the manual review period for the group.

[0104] Furthermore, the linear fitting algorithm takes the historical data deviation entropy value sequence arranged in chronological order as input, and establishes a fitting relationship according to the sampling time and entropy value corresponding to each sampling point. When the sampling period is fixed, the sampling sequence number can be directly used to replace the actual timestamp for fitting. When the sampling period is not fixed, the actual time interval is used for fitting to avoid treating long gaps and short-term fluctuations as the same.

[0105] The slope obtained thus represents the rate of change of the data deviation entropy value per unit time, and its absolute value positively reflects the magnitude of the change in the system's confidence state.

[0106] Furthermore, the aforementioned system confidence decay rate is an engineering term for system stability, and its monitoring focuses on the magnitude of changes in the system confidence structure, rather than being limited to a unidirectional decrease in confidence. Therefore, when the data deviation entropy value continues to rise, it can be understood that the system's confidence in the input data is weakening.

[0107] When the data deviation entropy value continues to decrease rapidly, it can be understood that the system confidence structure has undergone rapid reconstruction, and manual review should be prompted by an early warning. This is consistent with the calculation method of the absolute value of the slope and avoids mistakenly assuming that the rapid decline has been naturally eliminated.

[0108] The purpose of this mechanism is to move beyond the reliability of a single cycle to the level of whether the system is losing its continuous confidence, thereby achieving early warning of the 72-hour failure boundary and avoiding the situation where the screening system has actually lost the confidence basis to support decision-making even though it still has output results on the surface.

[0109] In a preferred embodiment of the present invention, the system further includes: a policy configuration terminal; the monitoring and early warning terminal is further configured to send a policy evaluation request to the policy configuration terminal; the policy configuration terminal is configured to update the first feature fusion policy and the suboptimal degradation fusion policy stored in the feature fusion server based on the policy evaluation request.

[0110] This embodiment provides a configuration update mechanism for policy adaptive evolution; specifically, although the aforementioned scheme can switch the fusion strategy between trusted and untrusted data states, in a continuously high-pressure social environment, the fixed first feature fusion strategy and the fixed suboptimal degradation fusion strategy may still gradually become ineffective.

[0111] The reason is that structurally abnormal behavior patterns themselves can evolve: initially they may only manifest as disabling text authorization, but they may evolve into uniformly creating traces of normal routines, and eventually even show a highly consistent distribution of low-risk behaviors across groups; therefore, it is necessary to introduce a strategy configuration terminal so that human experts and the system can form a closed loop and reorganize the fusion rules in stages.

[0112] The details are as follows: After receiving the status assessment results, intervention feedback results, and system failure warning signals, the monitoring and early warning terminal can initiate a policy assessment request to the policy configuration terminal. This request can carry information such as the false alarm situation, signs of missed alarms, data deviation entropy value distribution, verification results after manual intervention, and authorization withdrawal ratio for the current period.

[0113] The strategy configuration terminal is generally deployed in regional mental health centers or medical information management departments and is used by algorithm maintenance personnel, clinical psychologists or data security managers with corresponding responsibilities. Its update content includes, but is not limited to: adjusting the relative weights of different dimensions in the first feature weight set, adjusting the security threshold, updating the second type of feature set that can be retained in the degradation fusion, and replacing or retraining the state classification model.

[0114] For ease of explanation, a simplified schematic can be used to represent the logical flow of the strategy update; assume that the first feature weight set in a certain stage is {F1:0.35, F2:0.25, F3:0.20, F4:0.20}, and the second feature weight set is {F1:0, F2:0, F3:0.50, F4:0.50}.

[0115] After continuous operation, manual review revealed that the probability of the nighttime text sentiment dimension corresponding to F2 being affected by the target object's active modification behavior was higher than the preset deviation threshold, causing high-risk objects to show normalized results in this dimension for a long time, while the error variance of the device's use of rhythm data to reflect the objective load status was less than the preset threshold.

[0116] At this point, the strategy configuration terminal can update the first feature weight to {0.20, 0.10, 0.25, 0.45}, and further shrink the degradation strategy to retain the macro feature structure with F4 as the main component; this is not only to pursue model evaluation indicators, but also to feed the on-site verification conclusions into the feature confidence structure.

[0117] In boundary situations, if the information contained in the policy evaluation request is incomplete, such as only having false alarms but lacking human review conclusions, the policy configuration terminal may temporarily not execute the formal update, but instead return an instruction to maintain the current policy and collect supplementary feedback data.

[0118] If communication between the policy configuration terminal and the feature fusion server is abnormal, the server will continue to use the latest version of the policy that has been implemented locally to avoid screening interruption due to configuration unreachability; if a significant increase in error is found in a short period of time after the new policy is issued, the previous version of the policy can be retained as a rollback version and automatically restored after the anomaly is confirmed.

[0119] In the aforementioned university town and industrial park joint network, during the second preset time period after triggering the preset peak psychological load condition, the monitoring and early warning terminal found that the automatic screening conclusions of several departments deviated from the results of manual follow-up visits by the psychological hotline: some people who really needed attention were highly positive in their public texts, but their device usage rhythms were continuously disordered.

[0120] The monitoring and early warning terminal then sends a policy evaluation request to the policy configuration terminal. After evaluation, the policy configuration terminal lowers the weight of text-related fine-grained features in the trusted state, while increasing the weight of device rhythm and public interaction stability features, and simultaneously tightens the feature set available in the degraded mode. After the updated policy is written to the feature fusion server, the system's adaptability to similar distribution offset scenarios is enhanced.

[0121] Furthermore, when the strategy configuration terminal updates the first feature fusion strategy and the suboptimal degradation fusion strategy, it can use a version management method to record the strategy before the update, the strategy after the update, the effective time, and the evaluation basis for triggering the update.

[0122] This allows subsequent manual review results to correspond to specific strategy versions, facilitating rollbacks when strategy effectiveness is unstable and ensuring traceability of screening results at different stages. This additional explanation clarifies that updates include not only parameter replacements but also controlled writes, effect switching, and version retention.

[0123] The purpose of this mechanism is to ensure that the screening system does not statically execute preset rules, but rather makes phased adjustments to the fusion strategy based on the real feedback from the intervention network, thereby achieving the evolution from usability to continuous usability.

[0124] In a preferred embodiment of the present invention, the policy configuration terminal is further configured to: send an access request carrying the account information of the policy configuration terminal to the feature fusion server; the feature fusion server is further configured to: determine whether the policy configuration terminal has access rights to the feature fusion server, and if the policy configuration terminal has access rights to the feature fusion server, process the access request of the policy configuration terminal; or if the policy configuration terminal does not have access rights to the feature fusion server, reject the access request of the policy configuration terminal.

[0125] This embodiment provides an access control mechanism during the policy update process. Specifically, in the aforementioned scheme, the policy configuration terminal can modify the fusion policy, and this capability itself has a high impact. If access permission verification is lacking, once an unauthorized external terminal accesses the feature fusion server, it may unauthorizedly tamper with the policy to reduce the sensitivity of the warning, relax the security threshold, or forcibly close the downgrade policy, thereby directly disrupting the stable operation of the regional psychological screening network. Therefore, in addition to the policy update closed loop, an access request confirmation process needs to be established before the access request is processed.

[0126] The details are as follows: After the policy configuration terminal sends an access request to the feature fusion server, the server does not immediately perform configuration reading, policy modification, or model replacement, but instead enters the permission determination stage; the core of this stage is to distinguish between legitimate terminals with policy management responsibilities and other terminals that do not have the right to modify system behavior.

[0127] Only after the server confirms that the terminal has access rights will subsequent actions such as uploading policy evaluation results, adjusting weights, and changing thresholds be processed; otherwise, the server will reject the access request and may log the rejection for security auditing.

[0128] Access control here has dual engineering significance; on the one hand, it protects the screening strategy from being arbitrarily modified, preventing on-site personnel from adjusting the system trigger threshold to a state deviating from the preset safe range due to short-term pressure and non-standard procedures; on the other hand, it prevents external interference from directly affecting the system output through the policy entry point during periods of heightened group emotional confrontation.

[0129] Especially in medical informatics scenarios, policy changes are equivalent to redefining screening rules, and their security level should be higher than that of general query access.

[0130] As an alternative implementation, if the server is unable to complete the verification during the permission determination process, for example, due to the authentication service being temporarily unavailable, network link interruption, or security log module malfunction, the server can place the access request in a pending state and temporarily refrain from executing the policy update.

[0131] If a terminal sends multiple rejected requests in a short period of time, the server can also send a security alert to the monitoring and warning terminal to remind the operation and maintenance personnel to check for any abnormal attempts; if the permissions are granted but the request content lacks necessary fields, the server will only allow read-only access and will not allow write or update.

[0132] After the aforementioned regional network entered its third week of operation, due to the system issuing downgraded screening results multiple times, some non-policy management personnel attempted to directly modify feature weights through self-built terminals in order to improve the screening sensitivity of individual colleges. After the terminal initiated an access request to the feature fusion server, the server first performed permission verification, and those who failed were directly rejected, and their requests did not enter the policy writing process. This avoided policy changes without unified evaluation from interfering with the screening order of the entire region.

[0133] Furthermore, to avoid confusion in the text regarding terms such as access request, access permission, query permission, and read-only access, the access permission described in this article specifically refers to the processing permission for policy configuration resources in the feature fusion server, which includes at least the qualification judgment for controlled reading of the current policy version and submission of policy update requests.

[0134] The judgment object is whether the policy configuration terminal has entered the policy management link, rather than referring to all information query permissions within the system; therefore, even if a general query account has the permission to view the screening results in other business scenarios, as long as it does not have policy management responsibilities, it is regarded as not having the access permission of the feature fusion server in this embodiment.

[0135] The aforementioned read-only access refers to the situation where, when the policy management layer's permission verification has passed but the request message lacks the required fields for writing, the server only allows viewing the currently effective policy version and does not allow policy writing. This processing still belongs to the controlled processing of access requests and does not change the logical order of first determining permissions and then processing requests in this embodiment.

[0136] The purpose of this mechanism is to control the evolution of the strategy within the authorized boundaries, so as to ensure that the screening capability can be adjusted but not arbitrarily tampered with, and to ensure that the system still has basic governance order under high pressure.

[0137] In a preferred embodiment of the present invention, the feature fusion server is further configured to: extract the account information of the policy configuration terminal from the access request of the policy configuration terminal; obtain pre-stored reference account information from the feature fusion server; and compare the reference account information with the account information of the policy configuration terminal.

[0138] If the reference account information is inconsistent with the account information of the policy configuration terminal, it is determined that the policy configuration terminal does not have access to the feature fusion server; or if the reference account information is consistent with the account information of the policy configuration terminal, it is determined that the policy configuration terminal has access to the feature fusion server.

[0139] This embodiment provides an access permission confirmation step based on account information comparison. Specifically, the aforementioned access control mechanism clearly defines the principle of authentication before processing, but it is still not conducive to implementation if the authentication basis is not further disclosed. This embodiment uses account information comparison to realize permission confirmation, so that the server can determine whether the policy configuration terminal belongs to the pre-authorized scope in a clear and auditable manner.

[0140] The details are as follows: After receiving an access request, the feature fusion server extracts the terminal account information from the request message; the account information may include account identifier, affiliated organization identifier, job number or pre-bound terminal identity field; the server obtains pre-stored reference account information from the local secure storage area and compares the two;

[0141] When the two match, the terminal configured by the policy is deemed to have access rights; when the two do not match, it is deemed to have no access rights. The match here can be a match of all fields, or a match of key fields according to preset rules. For example, access is only granted when the account identifier and the organization identifier match at the same time.

[0142] As a specific example of data comparison logic; assuming the reference account information pre-stored in the server is account A-Regional Psychological Center-Strategy Position, and a certain access request carries the account information of account A-Regional Psychological Center-Strategy Position, then it is judged to pass; if the requesting account is account A-Department Counselor-General Position, or account B-Regional Psychological Center-Strategy Position, it is judged to fail due to inconsistencies in key information; this can avoid mistakenly allowing access based solely on matching a single field;

[0143] This account information comparison-based solution is suitable for high-impact configuration operations in medical information management scenarios because the frequency of policy modification is lower than the lower limit of the frequency of regular business queries, and the operating subject is clear. Using auditable account consistency verification is beneficial for tracing the chain of responsibility after the fact. Compared with fuzzy authorization methods, this solution is more suitable for clearly restricting the policy adjustment entry point in a regional psychological crisis screening system.

[0144] As a supplementary implementation, if the access request lacks an account information field, the server will default to determining that it does not have access rights and will reject the request; if the reference account information on the server is being updated or there are multiple authorization lists, the currently effective version can be locked for comparison to avoid authorization ambiguity caused by concurrent updates.

[0145] If the same account information initiates policy write requests from different terminal locations within a short period of time, the server may temporarily not process the write, only allow querying and prompt manual verification;

[0146] During the operation of the aforementioned regional network, the official policy management terminal of the regional mental health center plans to update the downgrade fusion rules. The account information carried in its access request is consistent with the server's reference account information. Therefore, the server allows it to read the current policy version and submit the update.

[0147] Conversely, although a college's information administrator has system query privileges, their account identifier does not match the pre-stored policy management account. Therefore, the server rejects their modification request and only retains the log record; this ensures that the policy adjustments that actually take effect all come from the predetermined management chain.

[0148] The purpose of this step is to complete permission confirmation through a clear and executable account comparison method, thereby achieving controlled opening of policy entry points and preventing high-risk configuration operations from escaping the authorization system.

[0149] In a preferred embodiment of the present invention, the data acquisition terminal includes: an interaction log acquisition module and a sensor status monitoring module; wherein, the interaction log acquisition module is used to acquire social network interaction text data of the target object; the sensor status monitoring module is used to acquire device usage rhythm data of the target object; the social network interaction text data corresponds to the first type of feature dimension, and the device usage rhythm data corresponds to the second type of feature dimension, the two together constituting the multidimensional behavioral sequence data.

[0150] This embodiment provides a terminal acquisition mechanism oriented towards specific data sources. Specifically, in the aforementioned scheme, multidimensional behavioral sequence data has been used for fusion and degradation decision-making, but if it is not specified which module actually collects the data, the data source is easily abstracted. This embodiment further limits the data acquisition terminal to include at least an interactive log acquisition module and a sensor status monitoring module, which respectively correspond to the two complementary information chains of explicit expression and objective rhythm in psychological screening.

[0151] The following is a detailed description: The interaction log collection module is used to collect text data of social network interactions; this data can come from authorized campus social platforms, public post replies, text summaries in instant messages, or interaction frequency records; its technology is to reflect changes in an individual's language expression style, degree of emotional expression, social avoidance, or sudden activity;

[0152] However, the probability that the social network interaction text data is actively modified is higher than the preset baseline probability, because the target object can change the text feature distribution by changing the wording, reply delay or applying preset templates; therefore, this part of the data is usually more suitable as a highly sensitive feature, participating in fine screening when the system confidence is sufficient, and being preferentially downweighted or removed when the confidence decreases.

[0153] The sensor status monitoring module is used to collect device usage rhythm data; this data may include screen-on time distribution, unlock frequency, continuous nighttime usage duration, charging time pattern, and switching rhythm of frequently used application categories, etc.

[0154] Compared to text data from social network interactions, the device uses rhythmic data that possesses objective physical attributes independent of subjective intent. Even with non-authentic modifications, its continuous distribution is highly correlated with the actual physiological workload, and the probability of synchronous non-authentic modifications is lower than the system's preset threshold. For example, when actual sleep is impaired, the terminal is more likely to leave rhythmic imprints when woken up late at night and used repeatedly for short periods. Therefore, this part of the data is often used as an important component of the second type of feature dimension and plays a core role in the downgrade fusion stage.

[0155] The two types of data together constitute multidimensional behavioral sequence data. For ease of explanation, a simplified schematic can be used to represent this: Within a sampling day, the interactive log acquisition module outputs text segments T1 and T2, and the sensor status monitoring module outputs rhythm segments R1, R2, and R3. Then, the data sequence uploaded by the terminal can be combined into {T1, T2, R1, R2, R3}. Among them, the T-type segments can be pre-labeled as the first type of feature dimension, and the R-type segments can be pre-labeled as the second type of feature dimension. The server can then perform the aforementioned entropy calculation and fusion strategy switching accordingly.

[0156] In boundary cases, if the interaction log collection module does not output data due to the target object withdrawing text authorization, the terminal can still only upload device usage rhythm data, and the system will directly enter the processing flow dominated by the second type of feature.

[0157] If the sensor status monitoring module fails to function properly due to system permission restrictions or terminal hardware failure, the terminal should include a missing data flag in the upload packet so that the server can avoid mistaking text data as complete input. If neither module has valid data, the sampling period will only be recorded as a data missing state, and no new status assessment results will be output. Instead, the low-intensity intervention state of the previous valid period will continue.

[0158] During the aforementioned high-pressure screening in the university town, a target individual continuously posted positive and frequently interactive text messages on social media platforms, and the interaction log collection module collected more positive text fragments than the preset baseline threshold; however, the device experienced high-frequency wake-ups and prolonged screen-on times during three consecutive nighttime periods, and the sensor status monitoring module collected significantly disordered usage rhythms.

[0159] After the system uploads the two sets of data together, the server identifies the difference between the positive text and the rhythm imbalance, thereby increasing its vigilance on the probability of spoofing and switching to the degraded fusion mode when necessary.

[0160] Furthermore, the multidimensional behavioral sequence data composed of social network interactive text data and device rhythm data refers to the terminal first completing the structured organization of the raw data into feature segments, and then encapsulating these feature segments into a unified data packet for uploading in chronological order, rather than requiring the server to directly mix and process the unorganized raw full-text text or raw sensor bottom-level signals.

[0161] Specifically, the interaction log collection module can output text-derived feature segments such as text length, sentiment summary, interaction frequency, and response delay;

[0162] The sensor status monitoring module can output rhythm-derived feature slices such as screen-on duration, number of nighttime wake-ups, unlock frequency, and application category switching rhythm. The former can be classified into fine-grained expression features in the first feature dimension, and the latter can be classified into macro-rhythmic features in the second feature dimension, thus maintaining consistency with the feature pre-labeling, weight allocation, and fusion switching process.

[0163] Furthermore, to maintain consistency with the terminology used above, in this paper, device usage rhythm data, device active rhythm, device active rhythm data, and macroscopic rhythm characteristics all refer to the same type of rhythmic behavior data or its derived characteristics collected and processed by the sensor status monitoring module at the data source level.

[0164] At the data source level, the fine-grained interactive text expression features of social network interactive text data all refer to text-based behavioral data or their derived features collected and organized by the interactive log collection module; the former is uniformly used as the source basis for the second type of feature dimension, and the latter is uniformly used as the source basis for the first type of feature dimension.

[0165] Therefore, the different names used in the preceding text are only to distinguish the granularity of data presentation, authorization status, or example content, and do not mean that the data acquisition terminal contains additional new acquisition modules that are not recorded.

[0166] Furthermore, when the terminal uploads the multidimensional behavior sequence data, it can simultaneously attach a sampling period identifier, a feature missing identifier, and an authorization status identifier so that the feature fusion server can distinguish between three different situations: the object does not have this type of behavior, the module has not collected this type of data, and this type of data is unavailable due to permission reasons.

[0167] This can prevent the server from misjudging missing data as stable behavior, and also provide a clearer data source basis for subsequent downgrade and fusion strategy switching.

[0168] Furthermore, the sampling results of the interactive log collection module and the sensor status monitoring module can be first aligned with timestamps, archived periodically, and marked with missing items according to a unified field format on the terminal side, and then merged into the multidimensional behavior sequence data for uploading;

[0169] This allows the server to read text-type feature slices and rhythm-type feature slices within the same sampling period, reducing artifacts caused by inconsistent sampling times, thus making subsequent feature drift comparisons and entropy value determinations more consistent.

[0170] The purpose of this mechanism is to provide the screening system with an input foundation that includes both expressive information and behavioral rhythm information through the parallel collection of two types of heterogeneous data, thereby achieving a more robust identification of psychological risks.

[0171] In a preferred embodiment of the present invention, the monitoring and early warning terminal is further used to: visualize and broadcast the intervention command and the system failure early warning signal.

[0172] This embodiment provides a multi-channel early warning display mechanism for regional command and manual duty scenarios. Specifically, in the aforementioned solution, the monitoring and early warning terminal is already able to receive status assessment results, generate intervention instructions, and receive system failure early warning signals. However, if it is only displayed in the form of a static list, the duty personnel may not be able to distinguish between the two different types of events, namely object risk and system confidence risk, in a timely manner under high concurrency and high pressure scenarios.

[0173] Therefore, this embodiment further improves the perceptibility and responsiveness of the intervention chain through visual display and voice broadcast;

[0174] The details are as follows: When the monitoring and early warning terminal visualizes intervention instructions, it can differentiate and present information such as object-level risk, the group to which the instruction belongs, the currently adopted fusion strategy, and the suggested intervention method on the interface. For example, intervention instructions generated from the first-state assessment results can be labeled as fine-tuning screening results, while intervention instructions generated from the second-state assessment results can be labeled as downgraded screening results, to indicate to the on-duty personnel the different levels of subsequent review intensity. The display of system failure early warning signals emphasizes trend information, such as showing the system confidence decay within a certain department, a certain park, or a certain time window, to help managers identify whether it is an individual problem or a group-wide decline in data confidence.

[0175] Voice broadcasts are used to ensure that personnel on duty can be informed of critical events in a timely manner even when they are not in a visual interface interactive state; especially when on duty at night, handling multiple tasks concurrently, or when the command center is monitoring multiple units at the same time, voice reminders can shorten the time difference between system judgment and human attention.

[0176] For intervention instructions, the broadcast content can be simplified to: a certain group has identified individuals to be tracked; please review the intervention recommendations. For system failure warning signals, the broadcast content can be simplified to: the confidence level of data in a certain area is rapidly declining; please check the screening strategy and manual review queue. The design focuses on conveying the event type and priority, rather than disclosing detailed individual privacy information.

[0177] If only a visual interface is relied upon, staff on duty may miss system failure warnings while handling other urgent matters; if only voice broadcasts are relied upon, it is difficult to retain the complete context for review.

[0178] Therefore, the combination of the two can complement each other; the former is suitable for presenting trends, hierarchy and object relationships, while the latter is suitable for providing instant reminders.

[0179] As an alternative implementation, if the terminal is in a preset physical environment with low noise restrictions, the system can automatically turn off the external speaker and switch to the headphone channel or vibration alert; if the visualization module is temporarily unavailable, the voice broadcast will still retain the key event alert; conversely, if the voice module malfunctions, the event will be presented on the interface in the form of a highlight, pop-up window or priority queue.

[0180] If a large number of warnings occur at the same time, in order to avoid information congestion, they can be broadcast in order of priority, that is, system failure warnings take precedence over individual intervention instructions.

[0181] When the aforementioned regional network enters the critical 72-hour monitoring period, the monitoring and early warning terminal displays on the large screen that the screening results of a certain college have changed from full-dimensional fusion to downgraded fusion. On the other hand, it announces with voice that the confidence level of the college's data is rapidly declining and requests manual review to be initiated.

[0182] Based on this, the staff on duty immediately determined that there were not only individuals requiring intervention, but also a possible spread of abnormal structural inputs within the group. Therefore, they simultaneously adjusted the shifts for the psychological hotline and the queue for manual review tasks.

[0183] The purpose of this mechanism is to clearly communicate object-level intervention information and system-level failure risks through different channels, thereby enabling rapid response and tiered handling in high-pressure duty scenarios.

[0184] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A mental health group screening system based on behavioral data feature fusion, characterized in that, include: Data acquisition terminal, feature fusion server, and monitoring and early warning terminal; The data acquisition terminal is used to acquire multidimensional behavior sequence data of the target object and send the multidimensional behavior sequence data to the feature fusion server. The multidimensional behavior sequence data includes a first type of feature dimension pre-labeled as involving privacy and a second type of feature dimension pre-labeled as public. The feature fusion server is used to obtain the historical baseline behavior data of the target object from a pre-configured database, generate a feature drift index by calculating the probability distribution distance between the historical baseline behavior data and the multidimensional behavior sequence data, input the feature drift index into a preset probability evaluation function for normalization mapping, calculate the spoofing probability value of the multidimensional behavior sequence data in the interval of 0 to 1, and use a preset information entropy calculation rule to multiply the spoofing probability value and the opposite probability corresponding to the spoofing probability value with their corresponding logarithms, and then perform a summation and inversion operation to convert it into the data deviation entropy value of the target object. The feature fusion server is further configured to process the multidimensional behavior sequence data using a first feature fusion strategy when the data deviation entropy value is lower than a preset safety threshold, so as to generate a first state evaluation result. The first feature fusion strategy includes: obtaining a preset first feature weight set, using the first feature weight set to perform weighted concatenation of the multidimensional behavior sequence data to generate a full-dimensional fusion feature matrix, and inputting the full-dimensional fusion feature matrix into a pre-trained machine learning state classification model to output the first state evaluation result. The feature fusion server is further configured to trigger a suboptimal degradation fusion strategy to process the multidimensional behavior sequence data when the data deviation entropy value is not lower than the preset safety threshold, so as to generate a second state evaluation result. The triggering of the suboptimal degradation fusion strategy includes: resetting the weights corresponding to the first type of feature dimension in the first feature weight set to zero to generate a second feature weight set; using the second feature weight set to extract the second type of feature dimension by weighting to generate a dimension-reduced macro feature matrix; and inputting the dimension-reduced macro feature matrix into the pre-trained machine learning state classification model to output the second state evaluation result. The feature fusion server is also used to send the first state evaluation result or the second state evaluation result to the monitoring and early warning terminal; The monitoring and early warning terminal is used to generate intervention instructions based on the first state assessment result or the second state assessment result, and send them to the data acquisition terminal for feedback display.

2. The mental health group screening system based on behavioral data feature fusion according to claim 1, characterized in that, The feature fusion server is also used for: Obtain the historical data deviation entropy value sequence of the target object within a preset time window; When the number of effective entropy points in the historical data deviation entropy value sequence reaches a preset number, based on the values ​​of each time node in the historical data deviation entropy value sequence, a linear fitting algorithm is used to calculate the slope of the historical data deviation entropy value sequence as a function of time, and the absolute value of the slope is used as the system confidence decay rate. In response to the system confidence decay rate being greater than a preset risk threshold, a system failure warning signal is sent to the monitoring and early warning terminal; or, If the system confidence decay rate is not greater than the preset risk threshold, the current feature fusion strategy is maintained.

3. The mental health group screening system based on behavioral data feature fusion according to claim 1, characterized in that, The system also includes: a policy configuration terminal; The monitoring and early warning terminal is also used to send a policy evaluation request to the policy configuration terminal; The policy configuration terminal is used to update the first feature fusion policy and the suboptimal degradation fusion policy stored in the feature fusion server based on the policy evaluation request.

4. The mental health group screening system based on behavioral data feature fusion according to claim 3, characterized in that, The policy configuration terminal is also used to: send an access request carrying the account information of the policy configuration terminal to the feature fusion server; The feature fusion server is further configured to: determine whether the policy configuration terminal has access to the feature fusion server, and if the policy configuration terminal has access to the feature fusion server, process the access request of the policy configuration terminal; Alternatively, if the policy configuration terminal does not have access to the feature fusion server, the access request of the policy configuration terminal may be denied.

5. The mental health group screening system based on behavioral data feature fusion according to claim 4, characterized in that, The feature fusion server is also used for: Extract the account information of the policy configuration terminal from the access request of the policy configuration terminal; Obtain pre-stored reference account information from the feature fusion server; The reference account information is compared with the account information of the policy configuration terminal; In response to the inconsistency between the reference account information and the account information of the policy configuration terminal, it is determined that the policy configuration terminal does not have access to the feature fusion server; or, In response to the fact that the reference account information matches the account information of the policy configuration terminal, it is determined that the policy configuration terminal has access to the feature fusion server.

6. The mental health group screening system based on behavioral data feature fusion according to any one of claims 1-5, characterized in that, The data acquisition terminal includes: Interactive log acquisition module and sensor status monitoring module; The interaction log collection module is used to collect social network interaction text data of the target object; The sensor status monitoring module is used to collect the device usage rhythm data of the target object; The social network interaction text data corresponds to the first type of feature dimension, and the device usage rhythm data corresponds to the second type of feature dimension. Together, they constitute the multidimensional behavioral sequence data.

7. The mental health group screening system based on behavioral data feature fusion according to claim 2, characterized in that, The monitoring and early warning terminal is also used for: The intervention commands and system failure warning signals are displayed visually and broadcast via voice.

Images