Dredging auxiliary platform based on psychological health education

Abstract

The invention relates to the technical field of psychological health education, and discloses a psychological health education-based dredging auxiliary platform, which comprises a psychological evaluation module, a data analysis module, an intervention execution module, a user archive library, a real-time monitoring module and the like. And the data analysis module is connected with the user archive library, generates a psychological state trend and intervention demand curve according to historical psychological data and behavior record data, and interacts data with other modules. And the intervention execution module receives related data, generates a real-time grooming instruction, and stores and records the instruction. The real-time monitoring module and the health rating module collect data to generate a corresponding matrix. The cognitive model optimization module optimizes parameters through a semantic association analysis network and the like. The state prediction model adopts a sliding time pane analysis method, and the health rating module adopts a fuzzy analytic hierarchy process. The platform realizes dynamic evaluation and accurate intervention, and improves the psychological health dredging effect.

Description

Technical Field

[0001] The present invention relates to the technical field of mental health education, and particularly to a counseling assistance platform based on mental health education. Background Art

[0002] Existing mental health assessment methods are relatively single, mostly using fixed questionnaire forms and lacking a dynamic adjustment mechanism. Fixed questionnaires cannot be personalized according to the user's real-time mental state and historical data, resulting in insufficient accuracy and pertinence of assessment results. For example, for users with different personalities and different life experiences, using the same questionnaire may not comprehensively and deeply reflect their true mental conditions, thus affecting the effectiveness of subsequent counseling plans.

[0003] In terms of psychological intervention, traditional methods lack systematicness and dynamic tracking mechanisms. Once a counseling plan is formulated, it is often difficult to adjust strategies in a timely manner according to the user's real-time feedback during the counseling process. For example, during the counseling process, due to changes in the external environment or the user's own emotional fluctuations, the original counseling plan may no longer be applicable, but traditional methods are difficult to detect and adjust in a timely manner, thus affecting the counseling effect.

[0004] In addition, the ability to integrate and analyze users' psychological data in existing technologies is insufficient. The user's mental state is affected by various factors, including physiological data, behavior records, environmental factors, etc., while traditional technologies often cannot effectively integrate and deeply analyze these multi-dimensional data, making it difficult to discover the potential laws and trends behind the data and unable to provide a more scientific and comprehensive basis for mental health assessment and intervention.

[0005] With the development of information technology, although some Internet-based mental health assistance platforms have emerged, most of these platforms have single functions and lack coordinated linkage between modules. For example, the data transmission and interaction between the assessment module, analysis module, and intervention module are not smooth enough to form a complete closed-loop system, resulting in the overall efficiency of the platform not being fully exerted. Summary of the Invention

[0006] The purpose of the present invention is to provide a counseling assistance platform based on mental health education to solve the problems raised in the above background art.

[0007] To achieve the above purpose, the present invention provides the following technical solution: A counseling assistance platform based on mental health education, the platform includes:

[0008] A mental assessment module, a data analysis module, an intervention execution module, and a user profile database;

[0009] The data analysis module is connected to the user profile database, where historical psychological data and behavioral record data are stored. The data analysis module obtains the historical psychological data in the user profile database, generates the trend of the user's psychological state through the state prediction model, and generates an intervention demand curve in combination with the behavioral record data. The data analysis module is connected to the psychological assessment module and sends the intervention demand curve to the psychological assessment module. The psychological assessment module adjusts the dynamic generation strategy of the assessment questionnaire according to the intervention demand curve transmitted by the data analysis module. The data analysis module is connected to the intervention execution module, sends the trend of the user's psychological state to the intervention execution module, and receives the real-time counseling instructions generated by the intervention execution module. The intervention execution module is connected to the user profile database, stores the counseling records in the user profile database for the data analysis module to call.

[0010] Preferably, the platform further includes a real-time monitoring module, a health rating module, and a cognitive model optimization module. The data analysis module is connected to the real-time monitoring module and the health rating module, and is used to send acquisition instructions to the real-time monitoring module and the health rating module. The acquisition instructions include obtaining the user's real-time physiological data and interaction behavior characteristics, uploading the collected real-time data to the user profile database. The real-time monitoring module and the health rating module send the psychological dynamic matrix and the health rating matrix to the intervention execution module. The data analysis module is connected to the cognitive model optimization module. When the health rating module determines that the psychological fluctuation amplitude exceeds the threshold, the data analysis module requests a correction parameter set from the cognitive model optimization module and sends it to the intervention execution module. The data analysis module generates an index of the real-time data during the operation process and the counseling result of the intervention execution module and transmits it to the cognitive model optimization module. The cognitive model optimization module includes a semantic association analysis network. The cognitive model optimization module receives the correction instruction from the data analysis module, generates a correction parameter set matching the current user state, and sends it to the data analysis module. After obtaining the correction parameter set, the data analysis module transmits it to the intervention execution module. The intervention execution module adjusts the push frequency of the counseling content according to the correction parameter set to stabilize the user's emotional fluctuation. After each counseling cycle ends, the cognitive model optimization module receives the real-time data and counseling result of the data analysis module and updates the parameters of the semantic association analysis network.

[0011] Preferably, the intervention demand curve includes a short-term emotional fluctuation threshold and a long-term psychological improvement trend function.

[0012] Preferably, according to the instruction of the data analysis module, the real-time monitoring module collects the skin conductivity and heart rate variability data of the user according to the preset event trigger mechanism, stores them in the user profile database after superimposing the environmental influence factors, and converts them into a psychological dynamic matrix and sends it to the intervention execution module through an encrypted channel.

[0013] Preferably, the health rating module periodically obtains the user's language logic and social activity parameters, adds the stress accumulation coefficient, stores them in the user profile library, and converts the collected parameters into a health rating matrix and sends it to the intervention execution module through an encrypted channel.

[0014] Preferably, at the beginning of the counseling cycle, the intervention execution module receives the user's psychological state trend transmitted by the data analysis module; during the counseling process, it receives the psychological dynamic matrix and health rating matrix transmitted by the real-time monitoring module and the health rating module, and uses the behavior chain matching method to perform content adaptation calculation on the psychological state and the counseling plan, and generates a real-time counseling instruction and feeds it back to the data analysis module, and at the same time stores it in the user profile library according to the timestamp.

[0015] Preferably, the semantic association analysis network in the cognitive model optimization module is a multi-level attention architecture. In the multi-level attention architecture, each layer of semantic attention head corresponds to emotional features of different dimensions, and each layer of associated memory unit binds the scenario dependence relationship of historical psychological data. The semantic tensors output by all levels together constitute a correction parameter group; the cognitive model optimization module reads the behavior data and psychological records in the current user profile library, inputs them into the multi-level attention architecture, and initializes each layer of the network to the context matching the current user state; based on the emotional stability constraint condition, adjusts the associated weights of each layer of the network to generate an optimal correction parameter group and sends it to the data analysis module.

[0016] Preferably, the platform further includes a resource adaptation module, and the resource adaptation module includes a knowledge base unit, a scenario simulation unit, and a feedback adjustment unit. In the multi-level attention architecture, the bottom layer network corresponds to the information retrieval delay parameter of the knowledge base unit, the middle layer network corresponds to the response matching degree of the scenario simulation unit, the upper layer network corresponds to the priority determination parameter of the feedback adjustment unit, and the final output layer corresponds to the counseling strategy deviation amount.

[0017] Preferably, in each layer network of the multi-level attention architecture, the feature fusion order is set from front to back as the environmental interference factor, the behavior pattern characteristic, and the cognitive deviation parameter. Among them, the environmental interference factor is a normalized vector of the noise level and social density; the behavior pattern characteristic is the periodic clustering part of the user behavior log; the cognitive deviation parameter is a normalized combination of the logical confusion index and the emotion regulation ability; in the multi-level attention architecture, according to the classification of the environmental interference factor, the behavior pattern characteristic, and the cognitive deviation parameter, a multi-dimensional feature cross-validation mechanism is configured in the cross-layer attention mechanism and the memory association layer of the network.

[0018] Preferably, the state prediction model is constructed by using the sliding time window analysis method, taking the weekly and monthly cycle parts of the psychological data as input. After outlier removal processing, the user's psychological state trend is generated through a multi-modal feature fusion module. The user's psychological state trend includes a basic emotion baseline part, a stress response interval, and a random stress range. The health rating module uses the fuzzy analytic hierarchy process to grade the user's comprehensive health status, maps the language coherence, facial expression recognition data, and interaction response duration into a health index matrix through a membership function, and then generates a health rating matrix through a feature dimensionality reduction module.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] Through the connection between the data analysis module and the user profile library, the platform can make full use of the stored historical psychological data and behavior record data. By using the state prediction model to deeply analyze the historical data, the user's psychological state trend is generated, and at the same time, the intervention demand curve is generated in combination with the behavior record data. This process realizes the forward-looking prediction of the user's psychological state, changes the limitation of the traditional method that only evaluates based on the current state, and enables psychological assessment and intervention to better meet the actual needs and development trends of users. For example, by analyzing the user's weekly and monthly cycle psychological data, potential psychological problem tendencies can be discovered in advance, and the guidance strategy can be adjusted in time to achieve early detection and early intervention of psychological problems.

[0021] According to the intervention demand curve transmitted by the data analysis module, the psychological assessment module adjusts the dynamic generation strategy of the assessment questionnaire. Compared with the traditional fixed questionnaire, this dynamic adjustment mechanism can generate a more targeted questionnaire according to the individual differences and real-time psychological state of the user, improving the accuracy and effectiveness of the assessment results. For example, for users with large emotional fluctuations, the system can automatically increase the proportion of relevant questions to more comprehensively understand the reasons and degrees of their emotional changes, providing a more accurate basis for subsequent guidance.

[0022] The close cooperation of the intervention execution module with multiple modules realizes the dynamic and personalized guidance process. At the beginning of the guidance cycle, it receives the user's psychological state trend. During the guidance process, it receives the psychological dynamic matrix and the health rating matrix in real time, and uses the behavior chain matching method to perform content adaptation calculation to generate real-time guidance instructions. This real-time feedback and dynamic adjustment mechanism enables the guidance plan to be optimized in time according to the user's real-time reaction during the guidance process, improving the guidance effect. For example, when the user's emotional fluctuations intensify during the guidance process, the system can quickly adjust the push frequency and method of the guidance content to stabilize the user's emotions and avoid the rigidity and lag of the guidance plan.

[0023] The real-time monitoring module and the health rating module collect multi-dimensional information such as users' physiological data and behavioral characteristics, providing rich real-time data support for the platform. The real-time monitoring module collects skin conductivity and heart rate variability data according to a preset event trigger mechanism, superimposes environmental influencing factors, and converts them into a psychological dynamic matrix; the health rating module periodically obtains parameters such as language logic and social activity, superimposes the stress accumulation coefficient, and generates a health rating matrix. The collection and analysis of these data enable the platform to understand users' psychological states more comprehensively and deeply, providing a more scientific basis for psychological assessment and intervention, and making up for the deficiency of traditional methods relying only on subjective reports.

[0024] The cognitive model optimization module realizes the intelligent optimization of the counseling strategy through a semantic association analysis network and a multi-level attention architecture. When the health rating module determines that the psychological fluctuation amplitude exceeds the threshold, it can timely generate a set of correction parameters matching the current user state and adjust the push frequency of the counseling content. At the same time, after each counseling cycle, the parameters of the semantic association analysis network are updated according to the real-time data and counseling results, enabling the platform to continuously learn and adapt to user changes, and improving the effectiveness and adaptability of the counseling strategy. This intelligent optimization mechanism enables the platform to have the ability of self-evolution and can continuously improve the service quality in long-term use.

[0025] The introduction of the resource adaptation module further improves the platform's collaborative operation ability and resource utilization efficiency. By corresponding the knowledge base unit, the scenario simulation unit, the feedback regulation unit to each layer of the multi-level attention architecture network, the multi-dimensional optimization of the counseling strategy is realized. For example, the bottom layer network corresponds to the information retrieval delay parameter of the knowledge base unit, ensuring the rapid acquisition of counseling resources; the top layer network corresponds to the priority determination parameter of the feedback regulation unit, enabling the counseling strategy to be reasonably sorted and adjusted according to the actual needs and feedback of users, improving the adaptability and utilization rate of resources.

[0026] Through data transmission and interaction among various modules, the platform forms a complete closed-loop system. From data collection, analysis and evaluation, intervention execution to model optimization, each link is closely connected, realizing the whole process management of mental health counseling. This systematic design not only improves the overall efficiency of the platform, but also provides users with coherent and continuous psychological support services, enhancing users' experience and trust. Brief Description of the Drawings

[0027] Figure 1 It is the working principle diagram of the counseling assistance platform based on mental health education described in the present invention;

[0028] Figure 2 It is the design diagram of real-time monitoring and cognitive optimization;

[0029] Figure 3 Design drawings generated for the psychological dynamic matrix;

[0030] Figure 4 Design drawings for the multi-level attention architecture;

[0031] Figure 5 Design drawings for the construction of the state prediction model. Specific implementation manners

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0033] Please refer to Figures 1-5 , a counseling assistance platform based on mental health education according to the present invention, the system includes:

[0034] A psychological assessment module, a data analysis module, an intervention execution module, and a user profile database.

[0035] The data analysis module is connected to the user profile database, and the user profile database stores historical psychological data and behavior record data. The data analysis module obtains the historical psychological data in the user profile database, generates the trend of the user's psychological state through the state prediction model, and generates an intervention demand curve in combination with the behavior record data.

[0036] The data analysis module is connected to the psychological assessment module, and sends the intervention demand curve to the psychological assessment module. The psychological assessment module adjusts the dynamic generation strategy of the assessment questionnaire according to the curve.

[0037] The data analysis module is connected to the intervention execution module, sends the trend of the user's psychological state to the intervention execution module, and at the same time receives the real-time counseling instruction generated by the intervention execution module.

[0038] The intervention execution module is connected to the user profile database, stores the counseling record in the user profile database for the data analysis module to call.

[0039] Embodiment 1:

[0040] The platform further includes a real-time monitoring module, a health rating module, and a cognitive model optimization module.

[0041] As the core processing unit, the data analysis module establishes two-way data transmission links with the real-time monitoring module and the health rating module. Specifically, the data analysis module sends acquisition instructions to the real-time monitoring module and the health rating module through a preset instruction protocol. The acquisition instructions contain clear parameter acquisition types and cycle rules. Among them, the parameter acquisition types cover two categories: users' real-time physiological data and interaction behavior characteristics. The real-time physiological data includes, but is not limited to, indicators reflecting the activities of the autonomic nervous system such as skin conductivity, heart rate variability, and respiratory rate. The interaction behavior characteristics involve digital behavior trajectories of users within the platform, such as click frequency, page stay duration, input text length, and social interaction frequency. The acquisition cycle is adaptively adjusted according to the dynamic change degree of the user's mental state. When the data analysis module finds through historical data modeling that the user's mental state fluctuates more severely, it automatically shortens the acquisition cycle to improve data real-time performance; otherwise, it lengthens the cycle to reduce system resource consumption.

[0042] After receiving the acquisition instructions, the real-time monitoring module and the health rating module respectively execute data acquisition and processing processes. The real-time monitoring module is equipped with a multi-modal physiological sensor interface, which is compatible with mainstream wearable devices (such as smart bracelets, heart rate monitoring chest straps, etc.) and platform-built-in sensors (such as cameras, microphones, etc.). It acquires users' physiological data in real time according to a preset event trigger mechanism. The event trigger mechanism is set based on double conditions of a time threshold and a behavior threshold. The time threshold can be set as a fixed time interval (such as every minute, every hour), and the behavior threshold is automatically triggered for acquisition according to specific operations of the user within the platform (such as starting a psychological assessment, receiving counseling content). The original data such as skin conductivity and heart rate variability collected are first preprocessed, including standard operations such as denoising filtering and baseline correction. Subsequently, environmental impact factors (such as the current environmental noise level, social scene density, etc., obtained through platform-built-in environmental sensors or manual input by users) are superimposed to form physiological data records containing spatio-temporal context. After being processed by an encryption algorithm, they are stored in the corresponding real-time data partition of the user profile library. At the same time, the real-time monitoring module maps the preprocessed physiological data into a psychological dynamic matrix. This matrix represents different dimensions of the user's real-time mental state (such as anxiety level, arousal degree, emotional stability, etc.) in the form of multi-dimensional vectors and is transmitted to the intervention execution module through an encrypted channel to provide physiological index basis for the dynamic adjustment of real-time counseling strategies.

[0043] The health rating module focuses on the behavioral and linguistic representations of the user's mental state. Through natural language processing technology and social behavior analysis algorithms, it periodically obtains parameters such as the user's language logic and social activity. The language logic analysis is based on the text content input by the user within the platform (such as the answers to the assessment questionnaire, feedback messages during the counseling process, etc.). Through algorithms such as syntactic analysis, semantic coherence measurement, and logical contradiction detection, a language logic index is generated. The social activity parameter is comprehensively obtained by calculating indicators such as the number of social interactions initiated or participated in by the user within a specified time window, the diversity of interaction objects, and the duration of interaction. In addition, the health rating module introduces a stress accumulation coefficient model. Based on parameters such as stress event records, emotional fluctuation frequency, and duration in the user's historical mental data, the current stress accumulation level is dynamically calculated through the exponential smoothing algorithm. After normalization, the language logic index, social activity parameter, and stress accumulation coefficient are linearly combined according to a preset weight matrix to be transformed into a health rating matrix. The health rating matrix presents the health level (such as low risk, medium risk, high risk) of the user's current mental state and the distribution of risk factors in each dimension in a hierarchical structure. It is also transmitted to the intervention execution module through an encrypted channel to assist in judging the urgency and pertinence of the counseling plan.

[0044] A closed-loop feedback mechanism is constructed between the data analysis module and the cognitive model optimization module. When the health rating module determines through the preset threshold detection algorithm that the user's mental fluctuation amplitude exceeds the safety threshold (such as the high-risk level in the health rating matrix continuously exceeds a specific duration), an abnormal state response process is triggered: the data analysis module sends a request for a correction parameter group to the cognitive model optimization module. This request includes context information such as the trend of the user's current mental state, the intervention demand curve, and real-time physiological and behavioral data. After receiving the request, the cognitive model optimization module starts the semantic association analysis network for parameter optimization calculation. The semantic association analysis network adopts a deep learning architecture, and its core components include a multi-level attention mechanism and a scenario memory unit: the multi-level attention mechanism allows the model to capture the semantic features of the user's mental state at different abstraction levels. Each layer of attention head corresponds to the emotional features of a specific dimension (such as anger, anxiety, depression, etc.), and the association weights between the features are automatically learned through the self-attention mechanism; the scenario memory unit binds the scenario-dependent relationships in the user's historical mental data, such as the mapping relationship between specific life events, social scenarios, and emotional responses, and realizes the temporal modeling of historical scenarios through the long short-term memory network (LSTM) or the Transformer architecture.

[0045] The cognitive model optimization module first reads the behavior data and psychological records in the current user profile library, encodes them into high-dimensional feature vectors, and then inputs them into a multi-level attention architecture. In the network initialization stage, through transfer learning techniques, the parameters of each layer are adjusted to the context space matching the current user state. For example, according to prior information such as the user's age, gender, and type of psychological problems, the weight distribution of the attention heads is initialized. Subsequently, based on the emotional stability constraint conditions (such as limiting the fluctuation range of the psychological state within a preset safe interval), the associated weights of each layer of the network are adjusted through the backpropagation algorithm, and the optimization goal is to minimize the deviation between the counseling strategy and the user's real-time psychological state. During the optimization process, the data analysis module continuously transmits real-time data during the operation process (such as physiological data collected by the real-time monitoring module, behavior parameters generated by the health rating module, etc.) and the counseling results of the intervention execution module (such as the types of counseling content already pushed, the push frequency, user interaction feedback, etc.) to the cognitive model optimization module. These data are processed through feature engineering to generate index tags, which are used to supervise the parameter update of the semantic association analysis network.

[0046] The correction parameter group generated by the cognitive model optimization module is fused by the semantic tensors output by each level, and specifically includes parameters such as the theme weight of the counseling content, the push time window, and the preference for interaction forms. The correction parameter group is transmitted to the data analysis module through the data interface. After the latter performs format verification and compatibility verification on the parameter group, it forwards it to the intervention execution module. The intervention execution module adjusts the push strategy of the counseling content according to the correction parameter group. For example, for users with a high anxiety level, the push frequency of mindfulness meditation content is increased, or for users with a significant decline in social activity, social skills training courses are preferentially recommended. After each counseling cycle ends (the counseling cycle can be set to a fixed duration or dynamically adjusted according to the complexity of the user's psychological state), the cognitive model optimization module receives the real-time data and counseling results summarized by the data analysis module, and updates the parameters of the semantic association analysis network through an online learning algorithm to achieve the adaptive evolution of the model to changes in the user's psychological state.

[0047] The data collection, processing, and transmission processes of the real-time monitoring module and the health rating module all follow strict data security specifications. All original data and derivative matrices (such as the psychological dynamics matrix, health rating matrix) are protected by national secret-level encryption algorithms (such as SM4) during storage and transmission to prevent data leakage or tampering. At the same time, the platform establishes a data access control mechanism to hierarchically manage the database access permissions of components such as the real-time monitoring module, the health rating module, and the cognitive model optimization module, ensuring that only authorized modules can call specific types of data resources.

[0048] In terms of system resource scheduling, the data analysis module dynamically allocates computing tasks to the real-time monitoring module, the health rating module, and the cognitive model optimization module through a load balancing algorithm to avoid computational resource overload in a single module. For example, when multiple users enter a high-risk psychological state simultaneously, the data analysis module preferentially allocates more computing power resources to the cognitive model optimization module to ensure the generation efficiency of the corrected parameter set; in normal states, resources are evenly allocated according to a preset ratio to maintain the stability of the overall system operation efficiency.

[0049] Embodiment 2:

[0050] The intervention demand curve includes a short-term emotional fluctuation threshold and a long-term psychological improvement trend function.

[0051] When generating the intervention demand curve, the data analysis module first performs time series decomposition on the historical psychological data in the user profile database. The historical psychological data covers the scores of psychological assessment questionnaires completed by the user in the past, emotional log records, feedback data during the counseling process, etc. The data is divided into two time scales, short-term (e.g., the most recent week) and long-term (e.g., the most recent three months), through the sliding time window analysis method. For short-term data, the data analysis module uses statistical analysis methods to calculate features such as the mean, standard deviation, and extreme values of emotional indicators, and combines a preset emotional fluctuation threshold algorithm (such as the dynamic percentile method based on historical data) to determine the short-term emotional fluctuation threshold. This threshold represents the critical value at which the user may experience emotional abnormalities in the short term. When the real-time monitored emotional indicator breaks through this threshold, an immediate intervention mechanism is triggered. The long-term psychological improvement trend function is constructed through a regression model in machine learning. Key features that can reflect changes in psychological state (such as the scores of the self-rating anxiety scale, the frequency of positive emotions, etc.) are selected as independent variables, and the timestamp is used as the dependent variable. The trend function curve is fitted through the least squares method or the gradient descent algorithm to predict the overall evolution direction of the user's psychological state in the future period of time.

[0052] The real-time monitoring module executes the physiological data collection task according to the instructions of the data analysis module, and its core mechanism is based on a preset event-triggering mechanism and a multi-source data fusion strategy. The event-triggering mechanism includes two types of triggering conditions: one is time-driven conditions, such as automatically starting physiological data collection at fixed times of the day (e.g., 10 am, 3 pm); the other is behavior-driven conditions, when the user performs specific operations within the platform (such as starting a new psychological assessment, clicking on a counseling content link), the collection process is automatically triggered through the event listening interface. The physiological data collected has skin conductivity and heart rate variability as the core indicators: the skin conductivity sensor obtains the sweat gland activity signal of the user's fingertip or palm through a contact electrode patch, and this signal is positively correlated with the level of emotional arousal; heart rate variability (HRV) is collected through photoplethysmography (PPG) technology, and by analyzing the variability of adjacent inter-beat intervals, it reflects the sympathetic-parasympathetic nerve balance state of the autonomic nervous system, and further evaluates the stress level and emotional stability.

[0053] After the original physiological data is collected, it enters the environmental impact factor superposition processing link. The environmental impact factors are obtained through the environmental perception module built into the platform, specifically including: (1) noise level, by collecting environmental sound signals through a microphone, analyzing the energy distribution of each frequency band through Fourier transform, and calculating the equivalent continuous A-weighted sound level (Leq) to characterize the environmental noise intensity; (2) social density, recorded through the location data authorized by the user or the social function of the platform, and counting the number of interaction objects in the physical space or virtual social scene where the user is currently located. The environmental impact factors are spatially and temporally aligned with the physiological data in the form of a standardized vector (value range [0,1]), for example, associating the noise level value at the same time point with the skin conductivity data to form a composite data unit containing environmental context. After the composite data unit undergoes data cleaning (removing outliers) and feature engineering (extracting time-domain and frequency-domain features), it is stored in the real-time physiological data partition of the user profile library and transmitted to the matrix generation unit of the real-time monitoring module through a data interface.

[0054] The matrix generation unit of the real-time monitoring module converts the processed physiological data and environmental impact factor composite data into a psychological dynamics matrix. The psychological dynamics matrix adopts a multi-dimensional tensor structure, and its dimensions are defined as follows: the first dimension is the time axis, with the minute as the minimum time unit; the second dimension is the physiological index dimension, including sub-dimensions such as skin conductivity, heart rate variability, and respiratory rate; the third dimension is the environmental impact dimension, including sub-dimensions such as noise level and social density. Each element value in the matrix represents the standardized data value under the corresponding time point, corresponding index, and environmental factor. The psychological dynamics matrix is sent to the intervention execution module through an AES-256 encrypted channel to provide a quantitative basis at the physiological level for the real-time adjustment of the counseling strategy.

[0055] The parameter collection and processing process of the health rating module focuses on the user's language behavior and social behavior characteristics. The acquisition of language logic parameters is based on natural language processing (NLP) technology, and the specific steps are as follows: First, all text data input by the user within the platform is obtained through the text collection interface, including the answers to psychological assessment questionnaires, conversation records with counseling robots, post content in social communities, etc.; Second, the text is segmented using a segmentation tool (such as jieba), and a text syntax tree is constructed by combining techniques such as part-of-speech tagging and named entity recognition; Then, the logical coherence between text paragraphs is evaluated through semantic similarity calculation (such as a sentence embedding model based on BERT), and semantic contradictions or topic jumps are detected; Finally, the grammar correctness score, semantic coherence score, and logical consistency score are weighted and summed to generate a language logic index (the value range is 0-100, and the higher the value, the stronger the language logic).

[0056] The calculation of social activity parameters is based on the user's social interaction records within the platform, specifically including: (1) The number of interaction initiations, which counts the number of times the user actively initiates chats, creates discussion groups, participates in activities, etc. within a unit of time; (2) The diversity of interaction objects, which calculates the distribution uniformity of the user's interaction objects through the Shannon entropy algorithm, and the higher the entropy value, the more diverse the interaction objects; (3) The duration of interaction, which calculates the average duration of a single interaction and the proportion of the total duration. The above indicators are obtained through the analysis of social behavior logs and are comprehensively integrated into a social activity index (the value range is 0-1).

[0057] The stress accumulation coefficient model uses the exponentially weighted moving average (EWMA) algorithm to dynamically model the user's historical stress events. Stress events include life events actively reported by the user in psychological assessments (such as exam failures, interpersonal conflicts, etc.), continuous high emotional fluctuation periods detected by the health rating module, etc. The model formula is:

[0058]

[0059] Among them, is the stress accumulation coefficient at the current moment, is the cumulative coefficient at the previous moment, is the forgetting factor (the value range is 0.8-0.95), is an indicator variable indicating whether a stress event occurs at the current moment (1 if it occurs, otherwise 0). The stress accumulation coefficient decays exponentially over time, ensuring that the impact weight of recent stress events on the current state is higher.

[0060] After being standardized, the language logic index, social activity index, and stress accumulation coefficient are fused through the health rating matrix generation model. This model uses the Analytic Hierarchy Process (AHP) to determine the weight vectors of each parameter. The weights are set based on the results of psychological research. For example, the contribution weight of language logic to the assessment of mental state is set to 0.4, social activity to 0.3, and stress accumulation coefficient to 0.3. The fused health rating matrix is presented in the form of a two-dimensional heat map. The horizontal axis is the time axis, and the vertical axis is the health risk level (low, medium, high). The depth of the color block at each time point reflects the health degree of the mental state at the corresponding moment. The health rating matrix is also transmitted to the intervention execution module through an encrypted channel, forming multi-modal data fusion with the mental dynamics matrix to jointly support the dynamic generation of the guidance plan.

[0061] In terms of data transmission and storage security, the mental dynamics matrix and health rating matrix generated by the real-time monitoring module and the health rating module are both stored using blockchain technology. A unique hash value is generated for each data matrix before transmission and written into the blockchain distributed ledger through a smart contract to ensure the immutability and traceability of the data. At the same time, the platform establishes a data life cycle management mechanism to anonymize or securely delete physiological data and behavior data that exceed the preset storage period (such as three years), meeting the requirements of privacy protection regulations.

[0062] At the system architecture level, the real-time monitoring module and the health rating module adopt a microservices architecture design and are each deployed as independent containerized services, communicating with the data analysis module and the intervention execution module through an API gateway. This architecture design realizes the decoupling between modules, facilitating the independent expansion of computing resources (such as adding sensor interface nodes and enhancing the computing power of the NLP model), and at the same time reducing the impact of a single module failure on the overall system. For example, when the health rating module experiences processing delays due to a surge in language data, the number of container instances can be automatically increased through the dynamic scaling mechanism to ensure the real-time collection of parameters and matrix generation.

[0063] Example 3:

[0064] The guidance process of the intervention execution module is divided into two stages: periodic initialization and real-time dynamic adjustment. At the beginning of the guidance cycle, the intervention execution module receives the user's psychological state trend transmitted by the data analysis module through the data interface. This trend includes the basic emotion baseline, stress response interval, and random stress range generated based on the state prediction model. The basic emotion baseline reflects the user's normal psychological state without significant external stimuli. The stress response interval represents the range of emotional fluctuations when the user faces stress events. The random stress range covers short-term emotional changes caused by occasional environmental disturbances. Based on the above trend data, the intervention execution module filters the initial guidance content from the preset guidance plan library, which is indexed multidimensionally according to psychological problem types (such as anxiety, depression, social phobia, etc.), intervention forms (such as cognitive behavioral therapy, mindfulness training, music guidance, etc.), and user characteristics (such as age, gender, education level).

[0065] During the guidance process, the intervention execution module continuously receives the psychological dynamics matrix transmitted by the real-time monitoring module and the health rating matrix transmitted by the health rating module. The psychological dynamics matrix reflects the user's physiological arousal level in real time in the form of a three-dimensional tensor (time × physiological indicators × environmental factors). The health rating matrix constructs a risk level map of the psychological state through behavioral parameters such as language logic and social activity. The intervention execution module uses the behavioral chain matching method to calculate the content adaptation between the psychological state and the guidance plan. This method constructs the probability transfer relationship between the user's behavior sequence and the guidance effect based on the hidden Markov model (HMM). Specifically, the behavioral chain matching method encodes the real-time behavior data of the user during the guidance process (such as click preference for guidance content, stay duration, feedback text sentiment tendency, etc.) as the observation sequence and defines the core intervention elements of the guidance plan (such as content theme, interaction form, push frequency, etc.) as the state sequence to achieve the dynamic matching of the guidance content and the user's real-time state through maximizing the posterior probability . The calculation of the posterior probability is based on Bayes' theorem:

[0066]

[0067] where: is the likelihood probability of the observation sequence under the state sequence, characterized by the emission probability matrix trained by historical data; is the prior probability of the state sequence, jointly determined by the preset logic of the guidance plan and the user's characteristics; is the marginal probability of the observation sequence, which is used as a normalization constant to ensure that the probability value is within the legal range.

[0068] By solving the above model through the Viterbi algorithm, the optimal state at each time step can be obtained , that is, the parameters of the current most suitable psychological state guidance plan (such as the weight of recommending a certain type of guidance content, adjusting the visual style of the interaction interface, etc.). The intervention execution module generates real-time guidance instructions according to the optimal state, and the instruction content includes but is not limited to: (1) pushing the link of the guidance content of a specific theme; (2) triggering an interactive psychological training task (such as a breathing regulation guidance animation); (3) sending an emotion recording reminder notice to the user. The real-time guidance instructions are synchronously fed back to the data analysis module through the message middleware for updating the input parameters of the user psychological state trend model, and at the same time, they are stored in the guidance record partition of the user profile library according to the timestamp (accurate to the millisecond level) to form a complete intervention process log.

[0069] The semantic association analysis network of the cognitive model optimization module adopts a multi-level attention architecture, and each layer of the network corresponds to the processing of emotion features at different abstract levels. Specifically, the semantic attention head of each layer is configured to focus on emotion features in specific dimensions. For example, the first layer of attention head focuses on the "anxiety - calm" dimension, the second layer focuses on the "depression - pleasure" dimension, and the third layer focuses on the "social avoidance - social approach" dimension. The associated memory unit of each layer adopts a long short-term memory (LSTM) network structure to capture the context-dependent relationships in historical psychological data, such as the emotion response patterns of users in specific social scenarios (such as workplace meetings, family gatherings). The semantic tensors output by all layers are fused through a fully connected layer to form a set of correction parameters containing the weights of emotion features in each dimension.

[0070] When the cognitive model optimization module receives the correction request from the data analysis module, it first reads the behavior data (such as the type of guidance content clicked, the distribution of interaction duration) and psychological records (such as real-time psychological assessment scores, emotion self-assessment logs) within the current period from the user profile library, and converts them into a mixed input of word embedding vectors and numerical feature vectors. The initial context parameters of the multi-level attention architecture are personalized initialized according to the user's demographic characteristics (such as age, occupation) and historical psychological intervention response patterns. For example, the attention weight of gamified guidance content is increased for the young user group. Subsequently, the model adjusts the associated weights of each layer based on the emotion stability constraint conditions (such as limiting the amplitude of skin conductance fluctuations in the psychological dynamic matrix not to exceed 1.5 times the baseline value), and the goal is to minimize the guidance strategy deviation amount (that is, the sum of the absolute values of the parameter differences between the preset guidance plan and the real-time adapted plan).

[0071] In the parameter update stage, the cognitive model optimization module receives the real-time data index and the guidance result label transmitted by the data analysis module. The real-time data index includes the feature summary of the psychological dynamics matrix (such as the mean and standard deviation of each physiological index) and the risk level change trajectory of the health rating matrix. The guidance result label extracts the emotional polarity (such as positive, neutral, negative) and the content satisfaction score from the user feedback text through natural language processing technology. These data serve as supervision signals to drive the weight update of the semantic association analysis network, enabling the model to adaptively learn the individual differences of users. After each guidance cycle ends, the cognitive model optimization module serially stores the parameters of the multi-level attention architecture for loading during the initialization of the next cycle, avoiding the computational resource consumption caused by repeated training.

[0072] At the data transmission level, asynchronous communication is implemented between the intervention execution module and the data analysis module using a message queue (such as RabbitMQ) to ensure the low-latency transmission and high-concurrency processing of real-time guidance instructions. Large-capacity data such as the psychological dynamics matrix and the health rating matrix are stored and transmitted in chunks through a distributed file system (such as HDFS) to improve data throughput efficiency. All sensitive data related to user privacy (such as the original values of physiological indicators and the specific content of social interactions) are processed through homomorphic encryption technology to ensure that model calculations and parameter updates are completed in the ciphertext state, meeting the requirements of privacy protection.

[0073] Example 4:

[0074] The resource adaptation module added to the platform includes a knowledge base unit, a scenario simulation unit, and a feedback adjustment unit. The three form a structured mapping with the multi-level attention architecture of the cognitive model optimization module to achieve refined management and dynamic adaptation of guidance resources.

[0075] The knowledge base unit stores a large amount of mental health education resources, including text materials (such as explanations of the principles of cognitive behavioral therapy), audio and video content (such as mindfulness meditation guidance audio, social skills training videos), interactive training tools (such as emotional diary templates, stress coping strategy tests), etc. These resources are multi-dimensionally labeled according to knowledge categories (such as psychotherapy, emotional management, interpersonal relationships), applicable groups (such as teenagers, working people, the elderly), and intervention goals (such as relieving anxiety, improving self-esteem, and improving sleep). In the multi-level attention architecture, the underlying network corresponds to the information retrieval delay parameter of the knowledge base unit, which reflects the time efficiency from resource request to content presentation. For example, when a user triggers the need for "instant anxiety relief", the underlying network prioritizes retrieving resources labeled "anxiety intervention-immediate effect-audio", and reduces the retrieval delay by optimizing the retrieval algorithm (such as combining inverted index with vector space model) to ensure that the adapted content is pushed within the critical time window of the user's emotional fluctuations. In actual applications, the underlying network can dynamically adjust the resource loading strategy according to the network environment in the user's area, such as giving priority to pushing text resources to users with low network bandwidth to avoid audio and video loading jams affecting the counseling experience.

[0076] The scenario simulation unit builds immersive psychological training scenarios based on virtual reality (VR) and augmented reality (AR) technologies, such as simulated workplace speeches, social gatherings, examinations, and other typical scenarios that are prone to psychological stress. The middle-level network corresponds to the response matching degree of the scenario simulation unit, that is, the degree of fit between the simulation scenario parameters and the user's real psychological state. Taking a user with social phobia as an example, the system determines that his social avoidance is at a high risk level by analyzing the social activity parameters in his health rating matrix (such as 0 active social times in the past week and only family members interacting with him) and the physiological indicators in the psychological dynamic matrix (such as a 30% decrease in heart rate variability when mentioning social scenarios). The middle-level network adjusts the scenario simulation parameters accordingly: low-intensity social scenarios (such as virtual one-on-one coffee conversations) are used in the initial stage, and the expressions and conversation content of the virtual characters in the scenes are pre-set to be gentle and friendly; as the user's anxiety level in the simulated scenario (judged by real-time monitoring of skin conductivity) gradually decreases, the middle-level network automatically increases the complexity of the scenario (such as increasing the number of virtual characters to 3-5 people and introducing slightly conflicting conversations) to achieve a dynamic match between the difficulty of the simulated scenario and the user's psychological tolerance. The response matching degree of the scenario simulation unit is also reflected in the authenticity of the sensory details. For example, based on the historical data in the user archive, the main color tone of the virtual scene is adjusted for users who are sensitive to specific colors, thereby enhancing the psychological intervention effect of the immersive experience.

[0077] The feedback adjustment unit is responsible for collecting real-time feedback data from users on the counseling content, including click behaviors (such as skipping a certain video segment, repeatedly watching a certain part of the content), text feedback (such as leaving a message in the comment area saying "the content is too abstract"), physiological feedback (such as a sudden increase in heart rate while watching the counseling video), etc., and converting the feedback data into priority determination parameters. The high-level network adjusts the priority of the counseling strategy according to these parameters. For example, when the completion rate of a certain type of mindfulness training video is lower than 30% for most users and the feedback is "boring", the high-level network will lower the push priority of this type of content and increase the priority of the "interactive emotional painting" type of content with positive user feedback; if a user's skin conductance continuously exceeds the baseline value by 20% and clicks the "pause" button multiple times when receiving workplace stress counseling content, the system determines that the current content does not match the user's needs and immediately triggers the feedback adjustment mechanism, and retrieves short-term content labeled "workplace stress - quick relaxation - breathing training" from the knowledge base unit for replacement. The feedback adjustment unit also has the ability to analyze trends across user groups. For example, by analyzing the high participation data of the adolescent group in gamified counseling tools, the high-level network automatically generates exclusive priority rules for this group and increases the push weight of psychological counseling games.

[0078] In the feature fusion link of each layer of the network, the environmental interference factor, behavior pattern characteristics, and cognitive bias parameters are processed in a fixed order. The environmental interference factor includes the standardized vectors of the noise level (such as the decibel value of the environment where the user is currently located) and the social density (such as the number of people in the physical space). For example, when a user uses the platform in a noisy public place, the noise level value in the environmental interference factor vector is close to 1, and the system accordingly reduces the push probability of text-based counseling content that requires deep concentration and preferentially recommends audio-guided content that does not require visual concentration. The behavior pattern characteristics are obtained through periodic clustering analysis of the user's behavior log. For example, the behavior log of a certain user shows that they frequently access the anxiety counseling section after 22:00 every day. The system determines that this period is the user's high-anxiety period and enhances the weight of the relevant data during this period during feature fusion to generate a targeted counseling plan in advance. The cognitive bias parameter is composed of the normalized combination of the logical confusion index (obtained through language coherence analysis) and the emotion regulation ability (evaluated through the emotion recovery speed in the historical counseling records). For example, users with a higher logical confusion index may have cognitive distortions, and the system focuses on invoking cognitive-behavioral therapy-related resources during feature fusion to improve their thinking patterns through structured cognitive restructuring training.

[0079] The cross-layer attention mechanism and the memory association layer configure a multi-dimensional feature cross-validation mechanism to ensure the accuracy and reliability of feature fusion. For example, when the underlying network detects that the user's current network environment is poor, resulting in a high retrieval delay of the knowledge base, the cross-layer attention mechanism will send a signal to the middle layer network to temporarily reduce the demand of the scenario simulation unit for real-time data interaction, avoiding affecting the fluency of the simulation scenario due to data transmission delay; the memory association layer verifies whether the current feature fusion result conforms to the individual's psychological change law by comparing the user's current behavior pattern with similar scenarios in historical data (such as the anxious performance before a certain exam half a year ago). If there is a significant deviation (such as the current anxiety level far exceeding the same period in history), the abnormal data marking process will be triggered to prompt the data analysis module to review the real-time monitoring data.

[0080] Taking a 25-year-old female workplace user as an example, her records in the user profile database show that she has recently had emotional problems such as difficulty falling asleep and irritability due to project pressure. The health rating matrix indicates that her stress accumulation coefficient has reached 0.8 (full value 1.0), the language logic index has dropped from 85 points to 72 points, and the social activity index is 0.3 (lower than the normal threshold of 0.5). The real-time monitoring module collects that her skin conductivity has increased by an average of 15% compared to the baseline value during the period from 18:00 to 20:00 on weekdays, and the heart rate variability has decreased by 22%. Adding that the noise level in her office environment reaches 65 decibels (higher than the comfortable threshold of 55 decibels), the psychological dynamic matrix shows a significant increase in the anxiety dimension value.

[0081] The processing flow of the resource adaptation module is as follows:

[0082] Underlying network: According to the user's current network environment (4G mobile network) and the knowledge base retrieval delay parameter, lightweight text + audio combination resources are preferentially retrieved, such as the audio file of "5-minute workplace stress breathing method" and the supporting text illustrations, to avoid video content affecting the timeliness of intervention due to slow loading.

[0083] Middle layer network: Based on its low social activity and cognitive deviation parameters (the increase in the logic confusion index), the scenario simulation unit generates a low-intensity workplace communication simulation scenario - the virtual colleague requests cooperation in a gentle tone, and the user responds through text input or voice commands. The system analyzes the logic and emotional tendency of the response content in real time, dynamically adjusts the reaction mode of the virtual colleague, and gradually increases the communication difficulty.

[0084] Upper layer network: When the user first uses the simulation scenario, pauses many times and mentions in the feedback that "the virtual character's expression is stiff", the feedback adjustment unit raises the priority of the "scene realism" parameter. Based on this, the middle layer network calls more delicate facial expression animation resources and adds environmental details (such as the green plants on the desk and background light music) in subsequent simulations. The user's subsequent feedback shows that "the sense of immersion is enhanced and the anxiety is relieved to some extent".

[0085] Feature Fusion and Verification: The cross-layer attention mechanism detects that the amplitude of the user's skin conductivity fluctuation in the simulation scenario is relatively large (more than 1.8 times the baseline value). The memory association layer compares its historical data and finds that a similar reaction occurred after a real workplace conflict, verifying that the current intervention plan reaches its core stress source. Continue to promote the dredging according to the established strategy. At the same time, increase the push frequency of relaxation audio through the underlying network to form a cyclic intervention mode of "stress exposure - relaxation training".

[0086] Through the collaborative operation of the resource adaptation module and the multi-level attention architecture, the platform realizes the dynamic allocation of all-chain resources from environmental perception, behavior analysis to cognitive intervention, ensuring that the dredging content reaches the optimal combination in terms of timeliness, matching degree, user experience, etc., and providing precise mental health education services for users in different mental states and different usage scenarios.

[0087] Example 5:

[0088] The state prediction model and the health rating module constitute the core components for the platform to conduct long-term analysis and comprehensive evaluation of the user's mental state. The two realize the quantitative modeling and grade division of the mental state through a data-driven algorithm framework.

[0089] Construction and Application of the State Prediction Model:

[0090] The state prediction model adopts the sliding time window analysis method to hierarchically analyze the time series characteristics of the user's mental data. Taking the weekly cycle and monthly cycle as the main analysis scales, for example, collect the user's mental assessment data in the past 12 weeks (such as the scores of the Self-Rating Anxiety Scale SAS and the Self-Rating Depression Scale SDS), daily emotion log records (the user's self-rating of the emotion of the day from 1 to 10), dredging intervention records (such as the number of times of receiving cognitive behavioral therapy, the emotion feedback after each intervention), etc. In the data preprocessing stage, outliers are removed through the interquartile range (IQR) method. For example, if the SAS score of a certain week suddenly increases to 90 points (far exceeding the user's historical mean ± 3 times the standard deviation) and there is no clear record of stress events, it is determined as an outlier and filtered.

[0091] The cleaned data is input into the multi-modal feature fusion module, which converts numerical features (such as scale scores, self-assessed emotion scores), text features (such as keyword frequencies in emotion logs), and behavioral features (such as the number of clicks on counseling content) into feature vectors of a unified dimension. For example, if a user clicks on the "social anxiety counseling" content multiple times within a monthly cycle, the "social anxiety intervention preference" dimension in the corresponding behavioral feature vector is assigned a high value; if words such as "nervous" and "irritable" frequently appear in the emotion log, the weight of the "negative emotion word frequency" dimension in the text feature vector is increased. After the multi-modal features are dimension-reduced by principal component analysis (PCA), they are input into a time series prediction model (such as a long short-term memory network LSTM) to generate a psychological state trend that includes the basic emotion baseline, stress response interval, and random stress range.

[0092] Taking a 28-year-old male user as an example, his weekly cycle psychological data shows that in the past 6 weeks, his SAS score fluctuated between 55 and 65 (the normal threshold is below 50), the average self-assessed emotion score was 6.8 (the highest anxiety level is 10), and he received 2 online counseling sessions per week. Through the analysis of a sliding time window (window length 4 weeks, step size 1 week) by the state prediction model, it is found that his basic emotion baseline corresponds to an SAS score of 58, the stress response interval is 60 - 65 (corresponding to predictable stress events such as work project nodes), and the random stress range is 55 - 58 (triggered by daily trivial matters). The model predicts that there is a 70% probability that his SAS score will remain in the range of 55 - 65 next month, indicating that continuous anxiety intervention is required and the counseling frequency should be increased one week before the project node.

[0093] The grading assessment process of the health rating module:

[0094] The health rating module uses the fuzzy analytic hierarchy process (FAHP) to comprehensively evaluate the user's psychological state from multiple dimensions such as language coherence, facial expression recognition data, and interaction response duration. The language coherence analysis is based on the open-ended response text of the user in the psychological assessment questionnaire. For example, the user is required to describe "the scene of feeling anxious last time". The logical coherence is evaluated by calculating the semantic similarity between sentences in the text (using the cosine similarity algorithm). If in the user's answer, "speaking at a meeting" is mentioned first and then jumps to "crowded subway", and there is no logical connection between the two parts, and the semantic similarity is less than 0.3, then the language coherence is judged to be poor.

[0095] Facial expression recognition data is collected through the built-in camera of the platform, and computer vision algorithms (such as OpenFace) are used to detect micro-expression features, including the frequency of frowning, the amplitude of muscle movement at the corners of the eyes, and the fixation duration. For example, when the user is watching a simulated social scenario video, if the frowning frequency is higher than 5 times per minute and lasts for more than 30 seconds, combined with the decrease in heart rate variability in the real-time physiological data, it is determined as a negative emotional reaction related to social anxiety. The interactive response duration refers to the operation delay of the user to the counseling content, such as the time interval for clicking the "Next" button. If the average response duration exceeds 90 seconds and is accompanied by repeated reading of the same content multiple times, it may indicate that the content difficulty does not match the user's cognitive level.

[0096] The above parameters are mapped into a health index matrix through a membership function. Taking language coherence as an example, five fuzzy levels of "excellent", "good", "medium", "poor", and "very poor" are set, and the corresponding membership function is a trapezoidal distribution: when the semantic similarity ≥ 0.8, the membership degree is 1 (excellent), linearly decreasing between 0.6 - 0.8, and when ≤ 0.4, the membership degree is 0 (poor and below). The frowning frequency in the facial expression recognition data uses a Gaussian membership function, and the frequency threshold intervals corresponding to different anxiety levels are fitted according to historical data. The membership function of the interactive response duration is set as an S-shaped curve. When the response duration exceeds 60 seconds, the probability of belonging to the "attention dispersion" level gradually increases with time.

[0097] The health index matrix is compressed to a two-dimensional space through a feature dimensionality reduction module (such as the t-SNE algorithm) to generate a health rating matrix. Each point in the matrix represents the user's mental state at a certain time point. The horizontal axis is the emotional stability dimension, and the vertical axis is the cognitive function dimension. Different color regions correspond to low, medium, and high risk levels. For example, a user's language coherence membership degree is 0.5 (medium), the facial expression recognition shows that the frequency of anxiety-related micro-expressions is 4 times per minute (corresponding to the medium risk threshold), and the average interactive response duration is 75 seconds (belonging to the medium risk of "attention dispersion"). After dimensionality reduction, the coordinates of its health rating matrix fall into the yellow area (medium risk), indicating that the counseling plan needs to be adjusted to increase cognitive training and attention guidance content.

[0098] Taking a 16-year-old high school student user as an example, the records in their user profile show that in the past three months, due to academic pressure, they have had problems such as difficulty falling asleep and inattentiveness in class, and have actively used the platform for psychological counseling. The state prediction model analyzed their monthly cycle psychological data and found that the self-evaluation scores of their emotions on Wednesdays and Thursdays were significantly lower than other periods (with an average score of 4.2 points on a 10-point scale). Corresponding to the time of math exams and mock tests, the SAS score in the stress response interval reached 68 points (moderate anxiety), and the random stress range was mainly triggered by a sharp increase in the amount of homework. The health rating module collected their language coherence data: in the answer to "describing the source of stress", the semantic similarity between sentences was 0.45, belonging to the "lower-middle" level; facial expression recognition showed that the frowning frequency reached 6 times per minute when mentioning exams, and the viewing screen time decreased by 30%; the average interactive response time when receiving academic stress counseling content was 120 seconds, significantly higher than the platform average (80 seconds).

[0099] Based on the above data, the platform executes the following intervention process:

[0100] State prediction application: According to the stress response pattern in the middle and later part of the week, "pre-exam relaxation training" audio content is pushed in advance on Tuesday evenings, including progressive muscle relaxation guidance. The duration of 7 days is a cycle, covering the key psychological adjustment window before the exam.

[0101] Health rating application: For the problem of poor language coherence, the "thinking visualization" tool is added to the counseling plan, such as filling in the mind map in the emotion diary template, to help users express stress feelings in a structured way; the facial expression and interactive response data indicate that they have cognitive avoidance of academic-related content, so the content presentation form is adjusted to short videos + dynamic charts, transforming the abstract stress management theory into visual cases (such as "application demonstration of the Pomodoro Technique in homework"), and shortening the duration of each piece of content to within 5 minutes to match their attention maintenance ability.

[0102] Data closed-loop verification: New data collected two weeks later showed that the self-evaluation scores of the user's emotions on Wednesdays and Thursdays increased to 5.8 points, and the SAS score decreased to 62 points; the semantic similarity of language coherence increased to 0.61, and the membership level rose to "medium"; the interactive response time for watching academic counseling videos shortened to 90 seconds, and the health rating matrix coordinates shifted towards the low-risk area. Based on this, the system adjusted the counseling strategy, reduced the passive push frequency, increased the proportion of counseling content that users can choose independently, and gradually cultivated their subjective initiative in psychological adjustment.

[0103] All data processed by the state prediction model and the health rating module follow the principle of minimum necessity. Only information directly related to psychological state assessment is collected, and user identities are hidden through de-identification techniques (such as replacing real names with hash values). The model parameters are updated using a federated learning framework. The local data of each user is only used for local model training, and the central server optimizes the global model by aggregating the model update gradients of each node, avoiding the leakage of private data. For example, the weight matrix of the fuzzy analytic hierarchy process in the health rating module is updated every two weeks based on the statistical results of anonymized data of all users to ensure that the model can adapt to the changing trends of the psychological characteristics of different groups.

[0104] It should be noted that in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0105] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A counseling assistance platform based on mental health education, characterized in that: It includes a psychological assessment module, a data analysis module, an intervention execution module, and a user profile database; the data analysis module is connected to the user profile database, and historical psychological data and behavior record data are stored in the user profile database. The data analysis module obtains the historical psychological data in the user profile database, generates the trend of the user's psychological state through a state prediction model, and generates an intervention demand curve in combination with the behavior record data; the data analysis module is connected to the psychological assessment module and sends the intervention demand curve to the psychological assessment module; the psychological assessment module adjusts the dynamic generation strategy of the assessment questionnaire according to the intervention demand curve transmitted by the data analysis module; the data analysis module is connected to the intervention execution module, sends the trend of the user's psychological state to the intervention execution module, and receives the real-time counseling instruction generated by the intervention execution module; the intervention execution module is connected to the user profile database, stores the counseling record in the user profile database for the data analysis module to call.

2. The counseling assistance platform based on mental health education according to claim 1, characterized in that It also includes a real-time monitoring module, a health rating module, and a cognitive model optimization module; the data analysis module is connected to the real-time monitoring module and the health rating module, and is used to send acquisition instructions to the real-time monitoring module and the health rating module. The acquisition instructions include obtaining the user's real-time physiological data and interactive behavior characteristics, uploading the collected real-time data to the user profile database, and the real-time monitoring module and the health rating module send the psychological dynamic matrix and the health rating matrix to the intervention execution module; the data analysis module is connected to the cognitive model optimization module. When the health rating module determines that the psychological fluctuation amplitude exceeds the threshold, the data analysis module requests a correction parameter group from the cognitive model optimization module and sends it to the intervention execution module; the data analysis module generates an index of the real-time data and the counseling result during the operation process and transmits it to the cognitive model optimization module; The cognitive model optimization module includes a semantic association analysis network. The cognitive model optimization module receives the correction instruction from the data analysis module, generates a correction parameter group that matches the current user state, and sends it to the data analysis module. After obtaining the correction parameter group, the data analysis module transmits it to the intervention execution module. The intervention execution module adjusts the push frequency of the counseling content according to the correction parameter group to stabilize the user's emotional fluctuation; after each counseling cycle ends, the cognitive model optimization module receives the real-time data and the counseling result from the data analysis module and updates the parameters of the semantic association analysis network.

3. The counseling assistance platform based on mental health education according to claim 1, characterized in that: The intervention demand curve includes a short-term emotional fluctuation threshold and a long-term psychological improvement trend function.

4. The counseling and assistance platform based on mental health education according to claim 2, characterized in that: According to the instruction of the data analysis module, the real-time monitoring module collects the user's skin conductivity and heart rate variability data according to a preset event trigger mechanism, stores them in the user profile database after superimposing environmental influencing factors, and converts them into a psychological dynamic matrix and sends it to the intervention execution module through an encrypted channel.

5. The counseling assistance platform based on mental health education according to claim 2, characterized in that: The health rating module periodically obtains the user's language logic and social activity parameters, stores them in the user profile database after superimposing the stress accumulation coefficient, and converts the collected parameters into a health rating matrix and sends it to the intervention execution module through an encrypted channel.

6. The counseling and assistance platform based on mental health education according to claim 5, characterized in that: At the beginning of the counseling cycle, the intervention execution module receives the user's psychological state trend transmitted by the data analysis module; during the counseling process, it receives the psychological dynamics matrix and the health rating matrix transmitted by the real-time monitoring module and the health rating module, uses the behavior chain matching method to perform content adaptation calculation on the psychological state and the counseling plan, generates a real-time counseling instruction and feeds it back to the data analysis module, and stores it in the user profile database according to the time stamp.

7. The counseling assistance platform based on mental health education according to claim 2, characterized in that: The semantic association analysis network in the cognitive model optimization module is a multi-level attention architecture. In the multi-level attention architecture, the semantic attention heads of each layer correspond to emotional features of different dimensions, and the associated memory units of each layer bind the context dependence relationship of historical psychological data. The semantic tensors output by all levels together constitute the correction parameter group; The cognitive model optimization module reads the behavior data and psychological records in the current user profile database, inputs them into the multi-level attention architecture, initializes each layer of the network to the context matching the current user state; based on the emotional stability constraint condition, adjusts the association weights of each layer of the network, generates the optimal correction parameter group, and sends it to the data analysis module.

8. The counseling assistance platform based on mental health education according to claim 7, characterized in that, It further includes a resource adaptation module. The resource adaptation module includes a knowledge base unit, a scenario simulation unit, and a feedback adjustment unit. In the multi-level attention architecture, the bottom layer network corresponds to the information retrieval delay parameter of the knowledge base unit, the middle layer network corresponds to the response matching degree of the scenario simulation unit, the upper layer network corresponds to the priority determination parameter of the feedback adjustment unit, and the final output layer corresponds to the counseling strategy deviation amount.

9. The counseling assistance platform based on mental health education according to claim 8, characterized in that: In each layer of the network of the multi-level attention architecture, the feature fusion order is set from front to back as the environmental interference factor, the behavior pattern characteristic, and the cognitive bias parameter. Among them, the environmental interference factor is the normalized vector of the noise level and the social density; the behavior pattern characteristic is the periodic clustering part of the user behavior log; the cognitive bias parameter is the normalized combination of the logical confusion index and the emotion regulation ability. According to the classification of the environmental interference factor, the behavior pattern characteristic, and the cognitive bias parameter in the multi-level attention architecture, a multi-dimensional feature cross-validation mechanism is configured in the cross-layer attention mechanism and the memory association layer of the network.

10. The counseling assistance platform based on mental health education according to claim 6, characterized in that: The state prediction model is constructed using the sliding time window analysis method. Taking the weekly and monthly cycle parts of the psychological data as inputs, after outlier removal processing, the user's psychological state trend is generated through the multi-modal feature fusion module. The user's psychological state trend includes the basic emotion baseline part, the stress response interval, and the random stress range; the health rating module uses the fuzzy analytic hierarchy process to grade the user's comprehensive health status, maps the language coherence, facial expression recognition data, and interaction response duration to the health index matrix through the membership function, and then generates the health rating matrix through the feature dimensionality reduction module.

Images