A multi-dimensional psychological evaluation system based on big data analysis

Abstract

The present application relates to the technical field of psychological assessment, in particular to a multi-dimensional psychological assessment system based on big data analysis, comprising the following modules: a multi-source data acquisition module; a data management and fusion module; a multi-dimensional analysis and calculation module connected with the data management and fusion module and configured to extract features based on the time series state vector; an intelligent evaluation and intervention module; an application and feedback module connected with the intelligent evaluation and intervention module; a privacy security and federated learning framework connected with all the above modules; the present application synchronously fuses active assessment, behavior logs, physiological signals and environmental context data through the multi-source data acquisition module, and adopts multi-modal feature dynamic weighted fusion based on attention mechanism, which significantly overcomes the problems that single self-report questionnaire is easily affected by social approval, self-cognitive bias and random answering, and improves the objectivity and reliability of the assessment results.

Description

Technical Field

[0001] This invention relates to the field of psychological assessment technology, specifically a multidimensional psychological assessment system based on big data analysis. Background Technology

[0002] Current mainstream psychological assessment methods mainly rely on three major categories of techniques: traditional self-report questionnaires, behavioral task assessments, and physiological signal collection. While questionnaires are widely used, they are severely limited by subjective factors such as the social desirability effect, self-perception bias, and random responses, resulting in low reliability and validity. Behavioral and physiological data are typically used as independent research tools, failing to be deeply integrated with subjective reports, leading to a single assessment dimension and an inability to comprehensively reflect the user's true psychological state. Furthermore, existing psychological assessments mostly involve static cross-sectional data collection, reflecting only the user's psychological state at a "certain moment" or "over the past week," failing to capture the dynamic evolution of emotions, stress, and cognitive function over time. Furthermore, the data cannot accurately predict future psychological risks; in addition, most systems use fixed, universal norms for comparison, ignoring individual differences in users' age, cultural background, occupation, and daily rhythms, leading to a large number of false positives or false negatives; assessment results are often presented as simple scores and descriptive reports, lacking in-depth analysis of potential causes and failing to provide personalized, dynamically adjustable intervention suggestions; assessment data from different institutions are isolated, forming "data silos," and highly sensitive mental health data is at risk of leakage and misuse during collection, transmission, storage, and third-party analysis, seriously restricting the development of accurate, continuous, and inclusive psychological assessments.

[0003] Therefore, we have made improvements to this by proposing a multidimensional psychological assessment system based on big data analysis. Summary of the Invention

[0004] The purpose of this invention is to provide a multidimensional psychological assessment system based on big data analysis to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution, comprising the following modules: The multi-source data acquisition module is configured to simultaneously collect users' active assessment data, behavioral log data, physiological signal data, and environmental and social context data. The data governance and fusion module is connected to the multi-source data acquisition module and is configured to perform time alignment, noise reduction and cleaning, and multimodal feature fusion on multi-source data to generate a unified time series state vector. The multidimensional analysis and calculation module is connected to the data governance and fusion module and is configured to extract features based on the time series state vector, construct personalized dynamic norms for users using federated learning and dynamic clustering algorithms, and learn the causal graph structure between features using causal inference algorithms. The intelligent assessment and intervention module is connected to the multidimensional analysis and calculation module and is configured to generate a multidimensional psychological profile based on the personalized dynamic norm and causal graph structure, use a time-series prediction model for risk warning, and output personalized intervention actions through a reinforcement learning recommendation engine. The application and feedback module is connected to the intelligent assessment and intervention module and is configured to provide users or institutions with visual reports and intervention interaction interfaces, and to collect user feedback data on the intervention effect and send it back to the multidimensional analysis and calculation module to update the model parameters. The privacy and security and federated learning framework connects with all the above modules and is configured to use cross-island federated learning, differential privacy and homomorphic encryption technologies to collaboratively train a global model among multiple data holders, and the original data does not leave the local machine.

[0006] As a preferred technical solution of this application, the multi-source data acquisition module further includes: The adaptive questionnaire submodule dynamically adjusts the difficulty and dimensions of the questions based on item response theory, and incorporates attention verification questions and random response detection logic. The behavior recording submodule is configured to collect context-independent behavior metadata from user terminals, including application usage duration classification, keystroke dynamics, mouse movement trajectory, and behavioral performance of gamified cognitive microtasks. The physiological sensing submodule is configured to interface with wearable devices or non-contact sensors to collect heart rate, heart rate variability, skin conductance, skin temperature and acceleration data in real time. The Environment and Social submodule is configured to acquire fuzzy geolocation entropy, Wi-Fi signal characteristics, and social interaction frequency statistics.

[0007] As a preferred technical solution of this application, the data governance and fusion module includes: The time alignment and cleaning unit is configured to resample multi-source data at a fixed time granularity, use Kalman filtering or linear interpolation to denoise physiological signals, and remove outliers that exceed physiological thresholds. The multimodal fusion unit is configured to dynamically weight and fuse features from each modality based on an attention mechanism to generate a unified time-series state vector; The time-series database storage unit is configured to use a time-series database to store raw data and feature vectors in a hierarchical manner, with hot data stored on high-speed media and cold data compressed and stored in object storage.

[0008] As a preferred technical solution of this application, the specific method for constructing personalized dynamic norms in the multidimensional analysis and calculation module is as follows: A collaborative training method based on federated averaging and distributed K-Means clustering is adopted. The original samples are not exchanged on the local data of multiple participants. Only the cluster center parameters or model gradients are exchanged to jointly construct a global user latent feature space. For any user, project their time-series state vector onto the feature space, and dynamically match the Top-K most similar psychological peer groups by calculating the distance to the existing cluster centers.

[0009] As a preferred technical solution of this application, the causal inference algorithm in the multidimensional analysis and calculation module is specifically as follows: The Peter-Clark (PC) algorithm or the Fast Causal Inference (FCI) algorithm is used to learn the Bayesian network structure from the time series state vector and determine the directed causal edges between each feature variable; It further includes a counterfactual reasoning unit, configured to simulate the expected change in a target psychological indicator when a certain antecedent variable is changed, based on the learned causal graph.

[0010] As a preferred technical solution of this application, the multi-dimensional psychological profile generation method in the intelligent assessment and intervention module includes: Based on the time series state vector and causal graph structure, quantitative evaluation values ​​for five dimensions are calculated: emotional stability, cognitive resilience, social approach / avoidance, stress load, and sleep recovery. The evaluation value for each dimension is obtained by weighted fusion of the corresponding physiological, behavioral, and questionnaire sub-features, and the weights are adaptively adjusted according to the population distribution in the personalized dynamic norm.

[0011] As a preferred technical solution of this application, the risk warning in the intelligent assessment and intervention module is implemented using a Long Short-Term Memory (LSTM) network model: Using the time series state vectors of N consecutive days as the input sequence, predict the probability of a preset psychological risk event occurring on the Tth day in the future; The warning threshold is dynamically set to add twice the standard deviation to each user's personalized baseline. When the predicted probability exceeds this threshold, a warning notification is pushed through the application and feedback module.

[0012] As a preferred technical solution of this application, the reinforcement learning recommendation engine in the intelligent evaluation and intervention module adopts a deep Q-network (DQN) architecture: The user's current time-series state vector and causal graph state are used as the state space, and the candidate intervention actions are used as the action space. The improvement degree of the user's state vector collected the next day, together with the explicit score collected by the application and feedback module, constitutes the reward signal; By continuously updating the DQN network parameters through online interaction, the personalized intervention actions output by the recommendation engine can maximize the accumulation of expected rewards.

[0013] As a preferred technical solution in this application, the privacy and security and federated learning framework is specifically implemented as follows: The horizontal federated learning unit is configured such that the central server only coordinates the model gradients or cluster center parameters uploaded by each participant, and uses a secure aggregation protocol to encrypt and sum the gradients, so that no party can know the original data of the other party. The differential privacy unit is configured to add Laplace or Gaussian noise that satisfies ε-differential privacy to the output population statistics results. The homomorphic encryption unit is configured to perform ciphertext computation using partial homomorphic encryption on sensitive statistics for cross-party computation. The zero-knowledge proof unit is configured to verify to a third party that a user meets a certain mental health status conclusion without disclosing any specific data.

[0014] As a preferred technical solution in this application, the specific process is as follows: During initial access, subjective and behavioral baselines are obtained through adaptive questionnaires and gamified cognitive tasks; During continuous operation, the multi-source data acquisition module records data at the minute level and transmits it to the data governance and fusion module to generate a time-series state vector; The multidimensional analysis and computation module uses federated learning dynamic norms and causal inference to perform real-time analysis of the current state vector, and the intelligent assessment and intervention module triggers personalized intervention pushes based on the analysis results. Users provide feedback on the intervention effect through the application and feedback module. This feedback data serves as a reward signal input to the reinforcement learning recommendation engine, and at the same time as a new sample to update the LSTM prediction model and causal graph structure. The updated model parameters are securely transmitted back to the central server through the federated learning framework for global model aggregation in the next iteration, thus forming a closed-loop evolution system of "perception-analysis-intervention-feedback-model self-optimization".

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By synchronously integrating active assessment, behavioral logs, physiological signals and environmental context data through a multi-source data acquisition module, and adopting dynamic weighted fusion of multimodal features based on attention mechanism, the problem of single self-report questionnaires being susceptible to social desirability, self-perception bias and random responses is significantly overcome, thereby improving the objectivity and reliability of the assessment results.

[0016] 2. By utilizing hierarchical storage of time-series databases and continuous data acquisition at the minute level, combined with Long Short-Term Memory (LSTM) networks, the evolution sequence of psychological states is modeled and future risks are predicted. This breaks through the limitations of traditional static cross-sectional assessments and enables dynamic tracking and early warning of emotions, stress, and cognitive functions. Personalized dynamic norms are constructed through federated learning and distributed clustering. The distribution of the user's psychological peer group is used as a comparison benchmark and is automatically updated weekly. This solves the problem of misjudgment caused by general fixed norms ignoring individual differences, making psychological assessments more relevant to the user's age, culture, occupation, and lifestyle.

[0017] 3. By introducing the Peter-Clark (PC) or Fast Causal Inference (FCI) algorithm to learn the Bayesian network structure between features, and combining it with counterfactual reasoning and Deep Q-Network (DQN) reinforcement learning recommendation engine, we can not only output psychological profile scores, but also reveal the causal path of "why it is abnormal" and provide dynamically optimized personalized intervention actions. At the same time, through user feedback, we form a closed loop of "perception-analysis-intervention-feedback-model self-optimization", which significantly improves the interpretability of the assessment results and the effectiveness of the intervention.

[0018] 4. By adopting cross-island federated learning, differential privacy, homomorphic encryption and zero-knowledge proof technologies, the original data of each participant does not leave their local machine. They only exchange encrypted gradients or clustering parameters. While breaking down traditional data silos and realizing cross-domain collaborative modeling, it strictly protects the security and compliance of users' sensitive mental health data. Attached Figure Description

[0019] Figure 1 This is a block diagram of the overall structure of the present invention. Detailed Implementation

[0020] 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.

[0021] This invention provides a technical solution: such as Figure 1 The multidimensional psychological assessment system based on big data analysis shown includes the following modules: The multi-source data acquisition module is configured to simultaneously collect users' active assessment data, behavior log data, physiological signal data, and environmental and social context data; the multi-source data acquisition module continuously collects data at a minute-level sampling period and pushes the timestamped raw data to the data governance and fusion module; The data governance and fusion module is connected to the multi-source data acquisition module and is configured to perform time alignment, noise reduction and cleaning, and multimodal feature fusion on the multi-source data to generate a unified time series state vector. After completing the fusion, the data governance and fusion module simultaneously sends the time series state vector to the multidimensional analysis and calculation module and the time series database storage unit. The multidimensional analysis and computation module, connected to the data governance and fusion module, is configured to extract features based on the time series state vector, construct personalized dynamic norms for users using federated learning and dynamic clustering algorithms, and learn the causal graph structure between features using causal inference algorithms; the multidimensional analysis and computation module also receives user feedback data from the application and feedback module and sends the updated norm parameters and causal graph structure to the intelligent evaluation and intervention module. The intelligent assessment and intervention module is connected to the multidimensional analysis and calculation module. It is configured to generate a multidimensional psychological profile based on the personalized dynamic norm and causal graph structure, use a time-series prediction model for risk warning, and output personalized intervention actions through a reinforcement learning recommendation engine. The intelligent assessment and intervention module sends the output personalized intervention actions to the application and feedback module, and simultaneously receives the intervention execution status returned by the application and feedback module, forming an intervention closed loop. The application and feedback module, connected to the intelligent assessment and intervention module, is configured to provide users or institutions with visual reports and an interactive intervention interface, and to collect user feedback data on the intervention effect and send it back to the multidimensional analysis and calculation module to update model parameters; the application and feedback module also sends the feedback data to the reinforcement learning recommendation engine in the intelligent assessment and intervention module for real-time updates of reward signals. The privacy and security and federated learning framework connects to all the above modules and is configured to use cross-island federated learning, differential privacy and homomorphic encryption technology to collaboratively train a global model among multiple data holders, and the original data does not leave the local machine; the privacy and security and federated learning framework protects the data transmission between all modules with encrypted channels and periodically injects differential privacy noise into the intermediate calculation results stored locally.

[0022] Furthermore, the multi-source data acquisition module further includes: The adaptive questionnaire submodule dynamically adjusts the difficulty and dimensions of the questions based on item response theory, and embeds attention verification questions and random answer detection logic; after each assessment, the adaptive questionnaire submodule sends the answer time and answer sequence of each question as behavioral features to the data governance and fusion module. The behavior recording submodule is configured to collect context-independent behavior metadata from user terminals, including application usage duration classification, keystroke dynamics, mouse movement trajectory, and behavior performance of gamified cognitive micro-tasks. The behavior recording submodule obtains explicit authorization from the user before collection and performs local anonymization processing on the raw data, uploading only statistical features. The physiological sensing submodule is configured to interface with wearable devices or non-contact sensors to collect heart rate, heart rate variability, skin conductance, skin temperature, and acceleration data in real time. The physiological sensing submodule has a built-in signal quality evaluation unit that automatically marks the data for that period as unusable when the signal-to-noise ratio is lower than a preset threshold. The Environment and Social submodule is configured to acquire fuzzy geolocation entropy, Wi-Fi signal characteristics, and social interaction frequency statistics. The Environment and Social submodule collects data every 30 minutes and temporarily stores it in a local cache after differential privacy processing.

[0023] Furthermore, the data governance and fusion module includes: The time alignment and cleaning unit is configured to resample multi-source data at a fixed time granularity, use Kalman filtering or linear interpolation to reduce noise in physiological signals, and remove outliers that exceed physiological thresholds; the time alignment and cleaning unit uses 1 minute as the basic granularity to unify all data onto the same time axis, and marks and skips missing data segments. The multimodal fusion unit is configured to dynamically weight and fuse features of each modality based on an attention mechanism to generate a unified time-series state vector; the feature vector output by the multimodal fusion unit at each time point has a fixed dimension of 128, which includes 32-dimensional physiological features, 32-dimensional behavioral features, 32-dimensional questionnaire features, and 32-dimensional environmental features. The time-series database storage unit is configured to use a time-series database to store raw data and feature vectors in a hierarchical manner. Hot data is stored on a high-speed medium, and cold data is compressed and stored in object storage. The time-series database storage unit retains data for each user for 5 years and applies AES-256 symmetric encryption before storage.

[0024] Furthermore, the specific method for constructing personalized dynamic norms in the multidimensional analysis and calculation module is as follows: A collaborative training method based on federated averaging and distributed K-Means clustering is adopted. The original samples are not exchanged on the local data of multiple participants. Only the cluster center parameters or model gradients are exchanged to jointly construct a global user latent feature space. The collaborative training is triggered once a week and only participates in the current round of updates when the amount of new local data added by a participant exceeds 1000 person-days. For any user, their time-series state vector is projected onto this feature space, and the most similar Top-K psychological peer groups are dynamically matched by calculating the distance to existing cluster centers; the default value of K is 50, and it can be automatically adjusted according to the historical matching stability of individual users; The characteristic distribution of this group is used as the personalized dynamic norm of the user, and the norm parameters are automatically updated according to a preset period (e.g., weekly). The updated norm parameters are distributed to the local servers of each participant through the federated learning framework for the next evaluation.

[0025] Furthermore, the causal inference algorithm in the multidimensional analysis and calculation module is specifically as follows: The Peter-Clark (PC) algorithm or the Fast Causal Inference (FCI) algorithm is used to learn the Bayesian network structure from the time series state vector and determine the directed causal edges between each feature variable; the causal inference algorithm runs once a day at midnight, based only on user data from the past 30 days, to avoid excessive computation time; It further includes a counterfactual reasoning unit, configured to simulate the expected change in the target psychological indicator when a certain antecedent variable is changed, based on the learned causal graph; the counterfactual reasoning unit outputs a readable statement "If X is changed to the target value Y, the target indicator Z is expected to change by ΔZ", with a 95% confidence interval.

[0026] Furthermore, the multi-dimensional psychological profile generation method in the intelligent assessment and intervention module includes: Based on the time series state vector and causal graph structure, quantitative evaluation values ​​for five dimensions are calculated: emotional stability, cognitive resilience, social approach / avoidance, stress load, and sleep recovery. The evaluation value for each dimension ranges from 0 to 100, and is accompanied by a percentile ranking compared to the user's personalized dynamic norm. The evaluation value of each dimension is obtained by weighted fusion of the corresponding physiological, behavioral and questionnaire sub-features, and the weights are adaptively adjusted according to the population distribution in the personalized dynamic norm; the adaptive adjustment adopts a soft maximization function to ensure that the sum of the weights of each dimension is 1, and is recalculated with each daily norm update.

[0027] Furthermore, the risk warning in the intelligent assessment and intervention module is implemented using a Long Short-Term Memory (LSTM) network model: Using the time series state vectors of N consecutive days as the input sequence, predict the probability of a preset psychological risk event occurring on the Tth day (e.g., the 7th day); where N is 14 days, T is 7 days, and the LSTM model contains two hidden layers, each with 128 units. The warning threshold is dynamically set to add twice the standard deviation to each user's personalized baseline. When the predicted probability exceeds this threshold, a warning notification is pushed through the application and feedback module. The warning notification includes the top three contributing features that led to the increased probability and their causal paths.

[0028] Furthermore, the reinforcement learning recommendation engine in the intelligent evaluation and intervention module adopts a deep Q-network (DQN) architecture: The user's current time-series state vector and causal graph state are used as the state space, and the candidate intervention actions (including recommending mindfulness practice, adjusting task priority, suggesting social activities, and pushing cognitive behavioral therapy exercises) are used as the action space; the action space also includes the empty action of "no intervention" in order to learn the optimal intervention frequency. The improvement degree of the user's state vector collected the next day, together with the explicit score collected by the application and feedback module, constitutes the reward signal; the reward signal calculation formula is: ,in This is a comprehensive reward signal; used in reinforcement learning recommendation engines to measure the overall benefit a user gains after receiving personalized intervention; the larger the value, the better the intervention effect, and it is used to update the Q network parameters to optimize subsequent intervention strategies. The weighted positive change in the multidimensional psychological profile of the next day compared to today; specifically, The objective improvement is calculated based on a five-dimensional psychological profile (emotional stability, cognitive resilience, social approach / avoidance, stress load, and sleep recovery). The change value of each dimension is first weighted according to its importance in the dynamic norm, and then summed to obtain the total positive change score. The range is usually normalized to [-1,1] or [0,1], with positive values ​​indicating improved state. The user's explicit subjective rating of the intervention effect; the rating is given directly by the user in the application and feedback module, with the original score range of 1-5 points, which is converted to the interval [-1,1] through linear mapping; the mapping rule is usually: 1 point → -1, 3 points → 0, 5 points → 1, and the intermediate value is linearly interpolated. By continuously updating the DQN network parameters through online interaction, the personalized intervention actions output by the recommendation engine can maximize the accumulation of expected rewards. The DQN network uses an experience replay pool with a capacity of 10,000 transformations, and updates the target network every 200 steps.

[0029] Furthermore, the privacy and security and federated learning framework is specifically implemented as follows: The horizontal federated learning unit is configured such that the central server only coordinates the model gradients or cluster center parameters uploaded by each participant. The gradients are encrypted and summed using a secure aggregation protocol, so that no party can know the original data of the other party. The secure aggregation protocol is based on threshold homomorphic encryption and requires at least two-thirds of the participants to be online to complete a round of aggregation. The differential privacy unit is configured to add Laplace or Gaussian noise that satisfies ε-differential privacy to the output population statistics results; the ε value is 0.5, and the noise amplitude is adaptively adjusted according to the sensitivity of the results; The homomorphic encryption unit is configured to perform ciphertext computation using partial homomorphic encryption on sensitive statistics for cross-square computation; the homomorphic encryption is only used for statistics such as median and variance, and not for large-scale matrix operations; The zero-knowledge proof unit is configured to verify to a third party that a user meets a certain mental health status conclusion without disclosing any specific data; the zero-knowledge proof is implemented using the Schnorr protocol, and the verification conclusion is a Boolean value.

[0030] Furthermore, the specific process is as follows: During initial access, subjective and behavioral baselines are obtained through adaptive questionnaires and gamified cognitive tasks; if users already have historical data in other participating systems, their anonymized norm indexes are synchronized through the federated learning framework, skipping the initial baseline collection step. During continuous operation, the multi-source data acquisition module records data at the minute level and transmits it to the data governance and fusion module to generate a time-series state vector; The multidimensional analysis and computation module uses federated learning dynamic norms and causal inference to perform real-time analysis of the current state vector, and the intelligent assessment and intervention module triggers personalized intervention pushes based on the analysis results. Users provide feedback on the intervention effect through the application and feedback module. This feedback data is used as a reward signal to input the reinforcement learning recommendation engine, and at the same time as new samples to update the LSTM prediction model and the causal graph structure. The LSTM prediction model is incrementally trained weekly using the data of the most recent 90 days, and the causal graph structure is completely relearned monthly. The updated model parameters are securely transmitted back to the central server through the federated learning framework for global model aggregation in the next iteration, thus forming a closed-loop evolution system of "perception-analysis-intervention-feedback-model self-optimization". When a user has no feedback data for 30 consecutive days, the system automatically reduces the data collection frequency to once per hour and suspends personalized intervention push until the user reactivates.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended technical solutions and their equivalents.

Claims

1. A multidimensional psychological assessment system based on big data analysis, characterized in that: Includes the following modules: The multi-source data acquisition module is configured to simultaneously collect users' active assessment data, behavioral log data, physiological signal data, and environmental and social context data. The data governance and fusion module is connected to the multi-source data acquisition module and is configured to perform time alignment, noise reduction and cleaning, and multimodal feature fusion on multi-source data to generate a unified time series state vector. The multidimensional analysis and calculation module is connected to the data governance and fusion module and is configured to extract features based on the time series state vector, construct personalized dynamic norms for users using federated learning and dynamic clustering algorithms, and learn the causal graph structure between features using causal inference algorithms. The intelligent assessment and intervention module is connected to the multidimensional analysis and calculation module and is configured to generate a multidimensional psychological profile based on the personalized dynamic norm and causal graph structure. The application and feedback module is connected to the intelligent assessment and intervention module and is configured to provide users or institutions with visual reports and an interactive intervention interface. The privacy and security and federated learning framework connects to all the modules mentioned above and is configured to employ cross-island federated learning, differential privacy, and homomorphic encryption technologies.

2. The multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The multi-source data acquisition module further includes: The adaptive questionnaire submodule dynamically adjusts the difficulty and dimensions of the questions based on item response theory, and incorporates attention verification questions and random response detection logic. The behavior recording submodule is configured to collect context-independent behavior metadata from user terminals, including application usage duration classification, keystroke dynamics, mouse movement trajectory, and behavioral performance of gamified cognitive microtasks. The physiological sensing submodule is configured to interface with wearable devices or non-contact sensors to collect heart rate, heart rate variability, skin conductance, skin temperature and acceleration data in real time. The Environment and Social submodule is configured to acquire fuzzy geolocation entropy, Wi-Fi signal characteristics, and social interaction frequency statistics.

3. The multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The data governance and fusion module includes: The time alignment and cleaning unit is configured to resample multi-source data at a fixed time granularity, use Kalman filtering or linear interpolation to denoise physiological signals, and remove outliers that exceed physiological thresholds. The multimodal fusion unit is configured to dynamically weight and fuse features from each modality based on an attention mechanism to generate a unified time-series state vector; The time-series database storage unit is configured to use a time-series database to store raw data and feature vectors in a hierarchical manner, with hot data stored on high-speed media and cold data compressed and stored in object storage.

4. The multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The specific method for constructing personalized dynamic norms in the multidimensional analysis and calculation module is as follows: A collaborative training method based on federated averaging and distributed K-Means clustering is adopted. The original samples are not exchanged on the local data of multiple participants. Only the cluster center parameters or model gradients are exchanged to jointly construct a global user latent feature space. For any user, project their time-series state vector onto the feature space, and dynamically match the Top-K most similar psychological peer groups by calculating the distance to the existing cluster centers.

5. The multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The causal inference algorithm in the multidimensional analysis and calculation module is specifically as follows: The Peter-Clark algorithm or the fast causal inference algorithm is used to learn the Bayesian network structure from the time series state vector and determine the directed causal edges between each feature variable; It further includes a counterfactual reasoning unit, configured to simulate the expected change in a target psychological indicator when a certain antecedent variable is changed, based on the learned causal graph.

6. The multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The multi-dimensional psychological profile generation methods in the intelligent assessment and intervention module include: Based on the time series state vector and causal graph structure, quantitative evaluation values ​​for five dimensions are calculated: emotional stability, cognitive resilience, social approach / avoidance, stress load, and sleep recovery. The evaluation value for each dimension is obtained by weighted fusion of the corresponding physiological, behavioral, and questionnaire sub-features, and the weights are adaptively adjusted according to the population distribution in the personalized dynamic norm.

7. The multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The risk warning in the intelligent assessment and intervention module is implemented using a long short-term memory network model. Using the time series state vectors of N consecutive days as the input sequence, predict the probability of a preset psychological risk event occurring on the Tth day in the future; The warning threshold is dynamically set to add twice the standard deviation to each user's personalized baseline. When the predicted probability exceeds this threshold, a warning notification is pushed through the application and feedback module.

8. The multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The reinforcement learning recommendation engine in the intelligent evaluation and intervention module adopts a deep Q-network architecture. The user's current time-series state vector and causal graph state are used as the state space, and the candidate intervention actions are used as the action space. The improvement degree of the user's state vector collected the next day, together with the explicit score collected by the application and feedback module, constitutes the reward signal; By continuously updating the DQN network parameters through online interaction, the personalized intervention actions output by the recommendation engine can maximize the accumulation of expected rewards.

9. The multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The privacy and federated learning framework is specifically implemented as follows: The horizontal federated learning unit is configured such that the central server only coordinates the model gradients or cluster center parameters uploaded by each participant, and uses a secure aggregation protocol to encrypt and sum the gradients, so that no party can know the original data of the other party. The differential privacy unit is configured to add Laplace or Gaussian noise that satisfies ε-differential privacy to the output population statistics results. The homomorphic encryption unit is configured to perform ciphertext computation using partial homomorphic encryption on sensitive statistics for cross-party computation. The zero-knowledge proof unit is configured to verify to a third party that a user meets a certain mental health status conclusion without disclosing any specific data.

10. A multidimensional psychological assessment system based on big data analysis according to claim 1, characterized in that: The specific process is as follows: During initial access, subjective and behavioral baselines are obtained through adaptive questionnaires and gamified cognitive tasks; During continuous operation, the multi-source data acquisition module records data at the minute level and transmits it to the data governance and fusion module to generate a time-series state vector; The multidimensional analysis and computation module uses federated learning dynamic norms and causal inference to perform real-time analysis of the current state vector, and the intelligent assessment and intervention module triggers personalized intervention pushes based on the analysis results. Users provide feedback on the intervention effect through the application and feedback module. This feedback data serves as a reward signal input to the reinforcement learning recommendation engine, and at the same time as a new sample to update the LSTM prediction model and causal graph structure. The updated model parameters are securely transmitted back to the central server through the federated learning framework for global model aggregation in the next iteration, thus forming a closed-loop evolution system of "perception-analysis-intervention-feedback-model self-optimization".

Images