Safety monitoring and evaluation method and system based on multi-dimensional data

Abstract

The invention relates to the technical field of multi-modal data processing and intelligent health management, in particular to a safety monitoring and evaluation method and system based on multi-dimensional data, and the method comprises the steps: mapping multi-modal feature data into health state vectors, and generating health state deviation and change trend by combining with time sequence analysis; identifying a health risk and generating a behavior intervention path; the system expands scarce scene data through a generative adversarial network, verifies the validity of a path in a virtual simulation environment, and optimizes parameters and association weights of key nodes in the path; dynamic collection of user feedback data further supports real-time adjustment of paths and models; finally, dynamic monitoring, accurate evaluation and personalized behavior intervention of the health state are achieved, and the method has high adaptability and perspectiveness.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal data processing and intelligent health management, and in particular to a safety monitoring and evaluation method and system based on multidimensional data. Background Art

[0002] The health status of the elderly is complex and is often affected by physiological changes, behavioral patterns, and environmental factors. Traditional safety monitoring technology is difficult to fully capture the complex relationship between these multimodal factors, and it is even more difficult to dynamically evaluate the changing trend of health status. The existing technology (Chinese invention patent, publication number: CN118379173B, name: A comprehensive safety monitoring method based on digital elderly care) mainly relies on dynamic Bayesian networks (DBN) to analyze the relationship between behavioral data and physiological data, and optimize behavioral recommendations by mapping weights. However, the existing technology has the following defects: Although existing technologies combine behavioral data and physiological data for analysis, they lack comprehensive consideration of environmental factors and long-term cumulative effects, which limits the accuracy and comprehensiveness of health assessments and behavioral interventions. Existing technologies tend to plan behavioral data based on impact weights, but are unable to dynamically capture the complex causal relationships of changes in health status in multiple dimensions, resulting in insufficient predictability and foresight. The model parameters of existing technologies mainly rely on static training results, lack the ability to dynamically adjust based on real-time feedback, and are difficult to adapt to individual differences and scenario changes. Summary of the invention

[0003] In view of the many problems existing in the above-mentioned prior art, the present invention provides a safety monitoring and evaluation method and system based on multidimensional data. The present invention maps multimodal feature data into a health state vector by enhancing the causal chain model, and dynamically captures the causal relationship between physiological, behavioral and environmental factors. Combining generative adversarial networks and virtual simulation technology, the system generates personalized behavioral intervention paths, and dynamically adjusts model weights and path parameters through real-time user feedback data. The present invention realizes accurate monitoring and forward-looking intervention of health status, and provides a scientific and personalized health management solution.

[0004] A safety monitoring and evaluation method based on multidimensional data comprises the following steps: Collect multimodal data through distributed sensor networks, including physiological data, behavioral data, environmental data, and social interaction data, pre-process and prioritize the collected data, and generate optimized sampling data and event cluster data; Based on optimized sampling data and event cluster data, multimodal feature data is extracted and fused to generate cross-modal feature data; a causal chain model is constructed by combining the causal inference network to generate short-term causal chain data and long-term causal chain data, and an enhanced causal chain model is generated through an optimization algorithm; Using enhanced causal chain models and cross-modal feature data, health status is evaluated and health status vectors and health status deviation data are generated; through path optimization and time series modeling, health risk path data, health trend prediction data and trend risk assessment data are generated; Based on health risk path data and health trend prediction data, a behavioral intervention path is generated, and the effectiveness of the intervention path is verified by combining simulation and optimization techniques; user feedback data is collected, and the key node weights of the behavioral intervention path and the associated parameters of the causal chain model are dynamically adjusted to generate optimized behavioral intervention path data and an updated causal chain model.

[0005] Preferably, the step of extracting and fusing multimodal feature data includes: extracting heart rate fluctuation characteristics and blood pressure change characteristics in physiological feature data, gait stability characteristics and activity frequency characteristics in behavioral feature data, and temperature and humidity fluctuation characteristics and light intensity change characteristics in environmental feature data based on optimized sampling data and event cluster data; and fusing the above feature data using a tensor decomposition method to generate cross-modal feature data.

[0006] Preferably, the tensor decomposition method includes: dynamically adjusting the correlation weights between feature data by decomposing the feature matrix; specifically including constructing a third-order tensor for physiological feature data, behavioral feature data, and environmental feature data, extracting key feature dimensions through singular value decomposition, and dynamically adjusting weight parameters to optimize the fusion effect and generate cross-modal feature data.

[0007] Preferably, the step of constructing a causal chain model includes: mapping cross-modal feature data into causal nodes using a causal inference network; generating short-term causal chain data based on time series analysis, and generating long-term causal chain data by combining the cumulative effects of multiple events; optimizing the correlation relationship and global weight of nodes in the causal chain through a nonlinear collaborative inference algorithm to generate an enhanced causal chain model.

[0008] Preferably, the nonlinear collaborative inference algorithm includes: applying a dynamic weight allocation strategy to adjust the node connection strength of short-term causal chain data according to the characteristic distribution of causal nodes; correcting the cumulative effect of long-term causal chain data through a recursive optimization mechanism, and using a multi-objective optimization method to calculate the global influence of nodes in the causal chain to generate a dynamically enhanced causal chain model.

[0009] Preferably, the step of evaluating the health status includes: constructing a health status mapping model based on the enhanced causal chain model and cross-modal feature data to generate a health status vector; generating health status deviation data by measuring the difference between the health status vector and the ideal health state, and generating health status change trend data in combination with historical health status change data.

[0010] Preferably, the step of generating health status deviation data includes: combining the current health status vector with the historical health status vector, analyzing the change trend using time series modeling, and generating a health status deviation curve; calculating the state deviation change rate of key nodes to generate health status deviation data.

[0011] Preferably, the step of generating a behavioral intervention path includes: generating preliminary behavioral intervention path data based on health risk path data and health trend prediction data in combination with a generative adversarial game optimization algorithm; verifying the improvement effect of each node in the preliminary behavioral intervention path on the health status through virtual simulation technology, adjusting the node parameters and connection weights in the path, and generating a behavioral intervention path.

[0012] Preferably, the step of collecting user feedback data includes: collecting feedback data after the user executes the behavior intervention path, including behavior completion rate, subjective feelings and health status change data; combining the generative adversarial network to expand scarce scenario data, dynamically adjusting the key node parameters of the behavior intervention path and the associated weights of the causal chain model, and generating optimized behavior intervention path data and an updated causal chain model.

[0013] A system for implementing the multi-dimensional data-based safety monitoring and assessment method, comprising: A data collection module for collecting physiological data, behavioral data, environmental data, and social interaction data through a distributed sensor network; The data processing module is used to pre-process and prioritize the collected multi-modal data to generate optimized sampling data and event cluster data; A feature extraction and fusion module is used to extract physiological feature data, behavioral feature data and environmental feature data based on optimized sampling data and event cluster data, and fuse the above feature data through a tensor decomposition method to generate cross-modal feature data; The causal inference and chain building module is used to combine the causal inference network, map the cross-modal feature data into causal nodes, generate short-term causal chain data and long-term causal chain data, and generate an enhanced causal chain model through an optimization algorithm; The health status assessment module is used to assess the health status using the enhanced causal chain model and cross-modal feature data, generate health status vectors and health status deviation data; generate health risk path data, health trend prediction data and trend risk assessment data through path optimization and time series modeling; The behavior intervention path generation module is used to generate preliminary behavior intervention path data based on health risk path data and health trend prediction data by generating adversarial game optimization algorithms, and verify the effectiveness of the path in combination with virtual simulation technology to generate behavior intervention paths; The feedback and optimization module is used to collect user feedback data, dynamically adjust the key node weights of the behavior intervention path and the associated parameters of the causal chain model, and generate optimized behavior intervention path data and updated causal chain model.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are: The present invention achieves accurate modeling and dynamic optimization of multi-dimensional causal relationships of health status by enhancing the causal chain model technology; The present invention achieves the improvement of model robustness and adaptability under data scarcity conditions by using the generative adversarial network to expand scarce scene data. The present invention realizes the scientific optimization and verification of the behavior intervention path through dynamic programming and virtual simulation technology, ensuring the effectiveness and personalization of the intervention measures; The present invention realizes real-time optimization and enhances the adaptability to complex changes in health status by dynamically adjusting the behavioral intervention path and the causal chain weight technical means. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 It is a schematic diagram of the construction process of the enhanced causal chain model in the present invention; Figure 3 It is a schematic diagram of the generation and optimization process of the behavior intervention path in the present invention; Figure 4 A schematic diagram of path optimization driven by dynamic feedback in the present invention; Figure 5 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0016] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0017] like Figure 1 As shown, a safety monitoring and evaluation method based on multidimensional data includes the following steps: Collect multimodal data through distributed sensor networks, including physiological data, behavioral data, environmental data, and social interaction data, pre-process and prioritize the collected data, and generate optimized sampling data and event cluster data; The present invention uses a distributed sensor network to collect multimodal data to provide data support for monitoring and evaluating the health status of the elderly. The distributed sensor network is composed of various types of sensors, including physiological sensors, behavioral sensors, environmental sensors, and social interaction sensors, which are installed on the devices worn by the elderly, the living environment, and the interactive terminal. Through network collaborative collection, it can cover the main data sources of the elderly's daily life and lay the foundation for multi-dimensional health analysis.

[0018] The collection of multimodal data is achieved through the following equipment and principles: Physiological data collection devices: such as smart bracelets or wearable blood pressure monitors, which record heart rate, blood oxygen, blood pressure and other indicators in real time. These data are realized through the optical detection principle of sensors (such as PPG and ECG). For example, heart rate monitoring is based on changes in light absorption caused by blood flow.

[0019] Behavioral data collection devices: such as gait monitors, which use inertial sensors (such as accelerometers and gyroscopes) to monitor gait smoothness, activity frequency and other behavioral characteristics. Such sensors can identify abnormal behaviors, such as falls or long periods of inactivity.

[0020] Environmental data collection equipment: such as temperature and humidity sensors and light detectors, which are used to monitor the temperature, humidity and light intensity of the living environment. By analyzing the environmental data, it can be inferred whether there are potential risks in the environment where the elderly live.

[0021] Social interaction data collection equipment: such as voice assistants or home interaction terminals, used to record data such as call frequency and visitor records to assist in assessing the social activity of the elderly.

[0022] The collected data is often affected by noise, equipment errors, etc. To ensure the reliability and consistency of the data, the raw data needs to be preprocessed. Align data from different sources through a unified timestamp to ensure temporal consistency between data. Use noise reduction algorithms suitable for the characteristics of different modal data, such as: using the Kalman filter algorithm for heart rate and blood pressure data to smooth random fluctuations in physiological data. Use the median filter algorithm for gait data to eliminate error points in abnormal behavior data.

[0023] In order to improve the efficiency and accuracy of data analysis, it is necessary to prioritize the preprocessed data.

[0024] The data is marked as an event using a threshold-based method. For example, a heart rate exceeding 120 bpm or a sudden gait stop is set as an abnormal event and the relevant data is marked. The DBSCAN algorithm is used to perform density clustering on the features of the marked event data. The algorithm calculates the local density of data points, identifies highly correlated abnormal data clusters, and generates event cluster data.

[0025] Through the above-mentioned process of data collection, preprocessing and priority screening, the accuracy and analysis efficiency of the health data of the elderly can be significantly improved. Through noise reduction and time synchronization, environmental interference and data deviation can be eliminated to provide high-quality data for subsequent analysis. Through priority screening, only high-priority data with strong relevance to health monitoring is retained to avoid interference of non-critical data on subsequent calculations. Abnormal event data is marked in real time and event clusters are generated to provide support for rapid assessment of health status. For example, potential risk events such as abnormal heart rate and falls can be discovered in a timely manner.

[0026] Example: A 75-year-old man wears a smart bracelet and a gait monitor, and a temperature and humidity sensor is installed in his living environment. Through the distributed sensor network, the following process is achieved: Data collection: The smart bracelet collects heart rate and blood pressure data, recording the current heart rate as 135 bpm (high) and blood pressure as 160 / 100 mmHg (high). The gait monitor records the elderly person's sudden gait stop and detects a large degree of posture tilt. The environmental sensor records the indoor temperature as 28°C and the humidity as 70%, with no abnormal environmental risks.

[0027] Data preprocessing: After the heart rate data was processed by Kalman filtering, random jitter was removed and the abnormal heart rate duration was confirmed to be more than 2 minutes. After the gait data was processed by median filtering, the posture tilt trend was identified to be consistent with the gait stop time.

[0028] Priority screening and event cluster generation: Based on the threshold method, heart rate higher than 120 bpm is marked as an abnormal event, and gait cessation for more than 10 seconds is marked as a behavioral abnormality event. The DBSCAN algorithm is used to cluster the marked events, confirming that heart rate abnormality and gait cessation are in the same event cluster, and generate event cluster data.

[0029] The system analyzes and determines that the elderly may have fallen. It combines event cluster data to send early warning information to caregivers in real time, including timestamps and key feature values ​​of heart rate and gait abnormalities. At the same time, the data is recorded for subsequent health assessment and behavior analysis.

[0030] like Figure 2 As shown, based on the optimized sampling data and event cluster data, multimodal feature data is extracted and fused to generate cross-modal feature data; a causal chain model is constructed in combination with a causal inference network to generate short-term causal chain data and long-term causal chain data, and an enhanced causal chain model is generated through an optimization algorithm; In the elderly care scenario of the present invention, multimodal data (such as physiological, behavioral, environmental and social data) can provide comprehensive information support for health monitoring and risk assessment. However, due to the heterogeneity of these data in feature space and differences in time scale, it is necessary to capture their inherent correlations and potential laws through feature extraction, fusion and causal chain modeling.

[0031] Based on the optimized sampling data and event cluster data, feature extraction is performed on data of different modalities, including heart rate fluctuation features and blood pressure change features in physiological data, gait stability features and activity frequency features in behavioral data, and temperature and humidity fluctuation features and light intensity change features in environmental data. After extraction, these multimodal feature data are fused through the tensor decomposition method to construct a high-dimensional feature space and generate cross-modal feature data. The tensor decomposition method captures the correlation between each modality data by decomposing the feature matrix.

[0032] Based on cross-modal feature data, a causal inference network (CIN) is used to build a causal chain model. The causal inference network generates short-term causal chain data through time series analysis, and generates long-term causal chain data by combining the long-term cumulative effects of multiple events. The specific steps are as follows: Generation of short-term causal chain data: Dynamically model the time series in cross-modal feature data and analyze the causal relationship between key events. Quantify the causal impact of event A on event B through Granger causality analysis method; Generation of long-term causal chain data: Construct a long-term causal chain for the time cumulative effects of multiple events, and calculate the total causal strength of the long-term effect through the cumulative weight function; Enhanced optimization of causal chain model: The association and weight of nodes in the causal chain are optimized through nonlinear collaborative inference algorithm. Nonlinear collaborative inference dynamically adjusts the weight balance between short-term causal chain and long-term causal chain through multi-objective optimization method to ensure that the causal chain model can reflect both immediate effect and cumulative effect.

[0033] Preferably, the step of extracting and fusing multimodal feature data includes: Based on the optimized sampling data and event cluster data, the heart rate fluctuation characteristics and blood pressure change characteristics in the physiological characteristic data, the gait stability characteristics and activity frequency characteristics in the behavioral characteristic data, and the temperature and humidity fluctuation characteristics and light intensity change characteristics in the environmental characteristic data are extracted; the above feature data are fused using the tensor decomposition method to generate cross-modal feature data.

[0034] In the elderly care scenario, the multimodal characteristics of physiological, behavioral and environmental data lead to differences and complex interactions between different modal data. For example, fluctuations in heart rate may be directly related to gait stability, while changes in ambient temperature and humidity may indirectly affect behavioral patterns through long-term cumulative effects. Therefore, the purpose of extracting and fusing multimodal feature data is to convert these heterogeneous data into a unified high-dimensional feature representation to provide support for subsequent causal analysis and health status assessment. The present invention achieves structured unification of multidimensional data by extracting features based on optimized sampling data and event cluster data, and fusion of multimodal data using tensor decomposition methods.

[0035] Based on the optimized sampling data and event cluster data, feature extraction is performed on physiological data, behavioral data and environmental data respectively.

[0036] Extraction of physiological characteristic data: Extract heart rate fluctuation characteristics and blood pressure change characteristics from heart rate and blood pressure data through time series analysis. Heart rate fluctuation characteristics reflect short-term abnormalities in the cardiovascular state of the elderly. For example, the standard deviation of the fluctuation amplitude is used to quantify the degree of fluctuation: ,

[0037] in, Indicates the characteristic value of heart rate fluctuation; Indicates Heart rate value at the moment; Indicates the mean heart rate; Indicates the time series length of the heart rate data. The blood pressure change characteristics identify the key points of long-term and short-term blood pressure changes through trend analysis, such as identifying peak points and downward trends.

[0038] Behavioral feature data extraction is based on gait stability and activity frequency analysis behavior data. Gait stability is obtained by analyzing the acceleration rate and angular velocity. For example, the standard deviation acceleration rate can be used to quantify the degree of instability. Activity frequency features are calculated by counting the number of movements in a fixed time period to generate feature data related to the daily activity patterns of the elderly.

[0039] Environmental feature data extraction: extract temperature and humidity fluctuation features and light intensity change features from environmental data. Temperature and humidity fluctuation features are calculated based on the change amplitude of the time series, such as trend segmentation of data through the sliding window method. Light intensity change features are generated by identifying the rising and falling cycles of light, and are used to infer the correlation between activity time and ambient light.

[0040] The core of multimodal data fusion is to map the above extracted features into a unified high-dimensional feature space to construct cross-modal feature data. The present invention adopts tensor decomposition method for data fusion, and the specific steps are as follows: Constructing feature tensors ,in: : Physiological characteristic dimension (such as heart rate fluctuation characteristics, blood pressure change characteristics); : Behavioral characteristic dimensions (such as gait stability characteristics, activity frequency characteristics); : Environmental characteristic dimensions (such as temperature and humidity fluctuation characteristics, light intensity change characteristics).

[0041] Decompose the tensor and extract the key feature dimensions: ,

[0042] in, Represents a multimodal feature tensor; represents the rank of tensor decomposition; Indicates The weight of the feature components; , , Represents the component vectors of physiological, behavioral, and environmental characteristics respectively. By dynamically adjusting the rank of the tensor decomposition , which can control the complexity of fusion data and balance the correlation between different modalities. The decomposed features are reconstructed into high-dimensional feature vectors to generate cross-modal feature data, ensuring that the data of different modalities can reflect their relevance and importance under a unified representation.

[0043] Converting heterogeneous physiological, behavioral, and environmental feature data into a unified high-dimensional feature representation facilitates subsequent causal analysis and health status assessment. By fusing multimodal data through tensor decomposition methods, not only the feature associations within a single modality are retained, but also the potential interactions between modalities are captured, such as the indirect effect of insufficient ambient light on gait stability. Highly correlated cross-modal feature data can support more accurate health risk assessments, such as identifying fall risks earlier by analyzing the combined features of heart rate fluctuations and activity frequency.

[0044] In the example, an elderly person was doing outdoor activities in winter, and the sensor recorded the following data: Data collection, heart rate records showed large fluctuations in heart rate over a short period of time, with a standard deviation of 12 bpm; gait monitoring records showed a decrease in activity frequency and an unstable gait was detected; environmental data recorded a temperature of 10°C, a humidity of 80%, and light intensity below the standard threshold.

[0045] Feature extraction, physiological feature data: the heart rate fluctuation feature extraction value is 12, and the blood pressure changes show a downward trend in systolic blood pressure; behavioral feature data: the gait stability feature value is 0.6 (based on the angular velocity change rate), and the activity frequency feature is 5 times per minute (lower than the daily activity level); environmental feature data: the temperature and humidity fluctuation feature value is medium to high, and the light intensity change feature shows a continuous low-light environment during the activity period.

[0046] Data fusion, tensor construction : Contains heart rate fluctuation characteristics and blood pressure change characteristics; Contains gait stability characteristics and activity frequency characteristics; Contains temperature and humidity fluctuation characteristics and light intensity change characteristics.

[0047] Tensor decomposition extracts key dimensions, and cross-modal feature data is generated after weight adjustment. The fused cross-modal feature data shows that there is a strong correlation between low-light environment and gait instability. Combined with the heart rate fluctuation characteristics, it further shows that activities in low light for the elderly may increase the risk of falling. The system uses this risk for subsequent health status assessment and behavioral intervention path generation.

[0048] Preferably, the tensor decomposition method includes: dynamically adjusting the correlation weights between feature data by decomposing the feature matrix; specifically including constructing a third-order tensor for physiological feature data, behavioral feature data, and environmental feature data, extracting key feature dimensions through singular value decomposition, and dynamically adjusting weight parameters to optimize the fusion effect and generate cross-modal feature data.

[0049] In the elderly care scenario, the monitoring and evaluation of health status requires the integration of multimodal data (such as physiological characteristic data, behavioral characteristic data, and environmental characteristic data). Due to the different sources of these data, their characteristics and dimensions are significantly different. For example, heart rate fluctuations and gait stability reflect short-term changes, while environmental temperature and humidity have long-term effects. Therefore, by using the tensor decomposition method to fuse multimodal feature data, the correlation weights of different modal features can be dynamically adjusted to ensure the scientificity and accuracy of feature fusion.

[0050] Based on the extracted multimodal feature data, a third-order tensor is constructed , to represent the multidimensional relationship between physiological, behavioral and environmental characteristics. Each element of Indicates Physiological characteristics, Behavioral characteristics and The interaction value between the environmental features.

[0051] The third-order tensor constructed Perform singular value decomposition (SVD) to extract key feature dimensions. The decomposition formula is: By selecting The component with the largest singular value can capture the main cross-modal feature associations, while reducing the feature dimension and avoiding redundant calculations.

[0052] In order to further optimize the feature fusion effect, the decomposed weight parameters Make dynamic adjustments. The basis for adjustment is as follows: The weights of physiological, behavioral, and environmental characteristics are dynamically balanced according to their weight ratio in a specific scenario. For example, in fall risk assessment, behavioral characteristics may be more important than environmental characteristics.

[0053] According to the temporal correlation strength of the feature data, the feature weight decay rate is adjusted. The dynamic adjustment formula is: ,

[0054] in, represents the adjusted weight parameter; represents the original weight parameter; Represents the adjustment factor, which is calculated based on the importance and time correlation of the modal features.

[0055] Through the adjusted tensor decomposition results, the decomposed weights and component vectors are reconstructed into high-dimensional feature vectors to generate cross-modal feature data. This data uniformly represents the multidimensional association of physiological, behavioral and environmental characteristics, providing support for subsequent causal analysis and health assessment.

[0056] The tensor decomposition method can capture the deep interactions between multimodal data, such as the combined effects of heart rate fluctuations, gait instability, and changes in light intensity, providing support for a comprehensive assessment of health status. Dynamic adjustment of weight parameters optimizes the feature fusion effect according to different scenarios. For example, in fall risk assessment, highlighting the weight of behavioral features helps to more accurately identify risk events. By extracting key feature dimensions through singular value decomposition, the interference of redundant features is reduced, the computational efficiency is improved, and the main feature information is retained.

[0057] In the example, an elderly person wears a health monitoring device during nighttime activities and collects the following data: Characteristic data: Physiological characteristic data, heart rate fluctuation characteristic value is 15 bpm, blood pressure change trend shows systolic blood pressure decreases. Behavioral characteristic data, gait stability characteristic value is 0.4 (low), activity frequency is 3 times per minute (significantly decreased). Environmental characteristic data, temperature and humidity fluctuation characteristics show high humidity (80%), light intensity is dim (less than 50 lux).

[0058] Constructing tensors: physiological feature dimensions : Heart rate fluctuation characteristics, blood pressure change characteristics; behavioral characteristics dimension :Gait stability characteristics, activity frequency characteristics; environmental characteristics dimension :Temperature and humidity fluctuation characteristics, light intensity change characteristics. Construct tensor , where the element values ​​represent the interaction effects of different feature combinations.

[0059] Extract the first two main eigencomponents by SVD , .

[0060] Dynamically adjust the weight parameters to give behavioral features higher priority. The adjustment factor is , The adjusted weight is , .

[0061] The cross-modal feature vector is reconstructed through the adjusted tensor decomposition results, and finally unified feature data is generated for causal analysis of fall risk and health status assessment. The system identifies the strong correlation between excessive humidity, insufficient light and unstable gait, and further assesses the fall risk of the elderly by combining the heart rate fluctuation characteristics, prompting caregivers to adjust the ambient light and optimize activity arrangements.

[0062] Preferably, the step of constructing a causal chain model includes: mapping cross-modal feature data into causal nodes using a causal inference network; generating short-term causal chain data based on time series analysis, and generating long-term causal chain data by combining the cumulative effects of multiple events; optimizing the correlation relationship and global weight of nodes in the causal chain through a nonlinear collaborative inference algorithm to generate an enhanced causal chain model.

[0063] In the elderly care scenario, the health status is affected by multimodal factors such as physiology, behavior, and environment, and there are complex causal relationships between these factors. In order to identify and evaluate these causal relationships, it is necessary to capture the interaction effects and cumulative effects of various factors through a causal chain model. The present invention maps cross-modal feature data into causal nodes through a causal inference network, generates short-term causal chain data based on time series analysis, and generates long-term causal chain data based on the cumulative effects of multiple events. Subsequently, the causal chain is optimized through a nonlinear collaborative inference algorithm to generate an enhanced causal chain model, which has both immediate response capabilities and long-term trend analysis capabilities.

[0064] Causal nodes are the basic units of the causal chain model, representing key events with causal relationships in feature data. Cross-modal feature data are mapped to causal nodes through the Causal Inference Network (CIN). The mapping rules are based on the conditional probability and correlation strength between feature data. For example, for the heart rate fluctuation feature (event A) and the gait stability feature (event B), by calculating the conditional probability and the total probability The difference in causal strength is determined by: ,

[0065] If the causal strength exceeds the set threshold , then connect event A and event B as causal nodes. In the elderly care scenario, causal nodes include: heart rate fluctuation (physiological node), unstable gait (behavioral node), and abnormal ambient temperature and humidity (environmental node).

[0066] Short-term causal chains are used to capture the immediate associations between health events based on time series analysis in cross-modal feature data. The specific process is as follows: Time series modeling uses the timestamps of events to construct time series data.

[0067] Causal strength calculation, through the Granger causality analysis method, quantifies the causal influence of event A on event B. The calculation formula is: ,

[0068] in, It represents the Granger causal strength of event A on event B; represents the variance of event B under the condition of event A; represents the total variance of event B.

[0069] Short-term causal chain data records the causal nodes and their strengths within the current time window, for example: Heart rate fluctuation → unstable gait, causal strength is 0.85.

[0070] Changes in temperature and humidity → heart rate fluctuations, the causal strength is 0.6.

[0071] Long-term causal chains capture the cumulative effects of multiple events and are particularly suitable for assessing the long-term impact of environmental factors. For example, persistent humidity abnormalities may lead to respiratory health problems. The specific process includes: Cumulative causal calculation, through the time-weighted cumulative model, calculates the long-term causal strength: , in, Indicates time the long-term causal strength of Indicates The weight of each event node; Represents a decay function, which is used to represent the effect of time interval on causal strength. Dynamically adjust the node relationship in the long-term causal chain and optimize the weight according to the frequency of event occurrence and cumulative strength.

[0072] There may be contradictions in the weight distribution and correlation between short-term and long-term causal chains, so the causal chain needs to be optimized through nonlinear collaborative inference algorithms. The optimization process includes: Error calculation, calculate the correlation error of the short-term causal chain and the long-term causal chain respectively: , in, Errors that represent short-term causal chains; Errors that represent long-term causal chains; , Represents the weight parameter, which is used to balance the correlation between short-term and long-term.

[0073] Optimize the target, reduce the overall error through recursive iteration, update the causal node weights and chain associations, and generate an enhanced causal chain model.

[0074] Through the short-term causal chain, the dynamic causal relationship between health events of the elderly can be captured instantly, such as the immediate correlation between heart rate fluctuations and gait abnormalities. The long-term causal chain effectively captures the long-term impact of environmental factors, such as the potential cumulative risk of humidity changes on the respiratory system. Nonlinear collaborative inference optimizes the causal chain model so that it can reflect both immediate relationships and analyze long-term trends, providing a scientific basis for health assessment and behavioral intervention.

[0075] In the embodiment, an elderly person's heart rate fluctuates more and his gait becomes unstable during nighttime activities, and the environment is low light and high humidity. The system analyzes the causal relationship between events through a causal chain model.

[0076] Causal node mapping: Heart rate fluctuation (physiological node): event A; Unstable gait (behavior node): event B; Insufficient ambient light (environment node): event C; Abnormal ambient humidity (environment node): event D.

[0077] Short-term causal chain generation: Heart rate fluctuation → unstable gait, causal strength is 0.8; Low light → unstable gait, causal strength is 0.6.

[0078] Long-term causal chain generation: Abnormal humidity → unstable gait, with a cumulative causal strength of 0.7.

[0079] Nonlinear collaborative optimization: After optimization, the weight of heart rate fluctuation on gait instability was adjusted to 0.85, and the weight of insufficient light was adjusted to 0.65.

[0080] The analysis results show that the risk of falling is caused by the combined action of multiple factors, among which heart rate fluctuations are immediate triggering factors and abnormal humidity is a cumulative risk factor. The optimized causal chain model supports the formulation of intervention strategies for lighting enhancement and humidity control.

[0081] Preferably, the nonlinear collaborative inference algorithm includes: applying a dynamic weight allocation strategy to adjust the node connection strength of short-term causal chain data according to the characteristic distribution of causal nodes; correcting the cumulative effect of long-term causal chain data through a recursive optimization mechanism, and using a multi-objective optimization method to calculate the global influence of nodes in the causal chain to generate a dynamically enhanced causal chain model.

[0082] In the elderly health monitoring, the causal chain model needs to analyze the immediate relationship of short-term events and the cumulative effect of long-term events at the same time. The dynamic changes and complex interactions of these causal relationships need to be optimized through nonlinear collaborative inference algorithms to improve the accuracy and adaptability of the causal chain model. The algorithm optimizes the connection strength of short-term causal chain nodes through a dynamic weight allocation strategy, corrects the cumulative effect of long-term causal chains using a recursive optimization mechanism, and calculates the global influence of causal nodes through a multi-objective optimization method, thereby generating a dynamically enhanced causal chain model.

[0083] The goal of the dynamic weight allocation strategy is to optimize the connection strength between nodes in the short-term causal chain according to the characteristic distribution of causal nodes, so that the model can more accurately reflect the immediate causal relationship.

[0084] Each node in the short-term causal chain data (such as heart rate fluctuation feature node A, gait instability feature node B) is assigned an initial weight according to the distribution of its feature value. The weight adjustment follows the following rules: high-correlation feature nodes (i.e. conditional probability is significantly higher than background probability) are assigned higher weights; the weights of low-correlation feature nodes are suppressed to avoid noise amplification.

[0085] Dynamic weight adjustment is achieved through the following formula: , in, Represents the connection strength between node A and node B; The conditional probability represents the immediate causal strength of node A on node B; It represents the correlation measure between the feature distribution of node A and node B. Through the above weight adjustment, the short-term causal chain can more accurately quantify the immediate causal relationship and ensure that the weights of key nodes are fully reflected in the model.

[0086] Long-term causal chains are used to analyze the cumulative effects of events, but these cumulative effects are uneven between different nodes. For example, the cumulative effect of abnormal environmental humidity on gait instability may be much greater than the change in light intensity. Therefore, the recursive optimization mechanism corrects the deviation of the cumulative effect through multiple rounds of iterations to ensure the balance and global accuracy of the model. According to historical event data, the cumulative effects of the nodes in the long-term causal chain are corrected through the weight decay function: , in, represents the corrected cumulative causal strength between node A and node B; represents the raw cumulative causal strength; Represents the time decay function, which is used to represent the decreasing relationship between the cumulative effect of nodes over time. In each iteration, the system updates the cumulative effect of the nodes and verifies the new weight distribution. The optimization termination condition is that the overall error is lower than the set threshold .

[0087] The overall optimization of the causal chain model requires finding a balance between short-term immediate response and long-term trend analysis. Through the multi-objective optimization method, the global influence of each causal node is calculated and the global weight of the node is adjusted. The optimization objective expression is: , in, Representation Node Short-term causal errors; Representation Node long-term causal error; , The equilibrium parameters representing the short-term and long-term causal chains; Represents the total number of causal nodes.

[0088] Global influence calculation: , in, Representation Node global influence; , Represents the weight of the node in the short-term and long-term causal chain. Through multi-objective optimization, the global influence of the node is dynamically adjusted to generate a more balanced and accurate causal chain model.

[0089] The dynamic weight allocation strategy enables the short-term causal chain to accurately capture the immediate causal relationship between key events, such as the direct impact of heart rate fluctuations on gait instability. The recursive optimization mechanism corrects the deviations in the long-term causal chain, making the analysis of cumulative effects more accurate, such as the long-term impact of abnormal humidity on the respiratory health of the elderly. Through multi-objective optimization, the causal chain model achieves a dynamic balance between short-term immediate response and long-term trend prediction, providing more comprehensive data support for health status assessment and behavioral intervention.

[0090] Example: An elderly person's gait becomes unstable and his heart rate fluctuates more during nighttime activities. At the same time, the ambient light is weak and the humidity is high. The system optimizes the causal chain model through a nonlinear collaborative inference algorithm. The specific process is as follows: In the short-term causal chain, the immediate causal strength of heart rate fluctuation (node ​​A) on unstable gait (node ​​B) is 0.85, and the connection strength is adjusted to 0.9 through the dynamic weight allocation strategy. The weight of insufficient ambient light (node ​​C) on unstable gait is adjusted to 0.65.

[0091] In the long-term causal chain, the original value of the cumulative causal strength of humidity anomaly (node ​​D) on gait instability is 0.7, which is corrected to 0.75 after recursive optimization.

[0092] The equilibrium parameters of the short-term causal chain and the long-term causal chain are set as , The global influence of the unstable gait node is calculated as: .

[0093] The optimized causal chain model shows that abnormal humidity has a significant long-term correlation with gait instability, and heart rate fluctuations are the immediate factors that directly trigger falls. The system recommends optimizing the night activity environment (increasing light brightness, reducing humidity) and monitoring heart rate fluctuations.

[0094] Using enhanced causal chain models and cross-modal feature data, health status is evaluated and health status vectors and health status deviation data are generated; through path optimization and time series modeling, health risk path data, health trend prediction data and trend risk assessment data are generated; The health status of the elderly is the result of the comprehensive effect of multimodal data, such as physiological data (such as heart rate fluctuations), behavioral data (such as gait stability) and environmental data (such as light intensity and humidity changes). The present invention evaluates health status by enhancing the causal chain model and cross-modal feature data, and combines path optimization and time series modeling to generate health risk path data, health trend prediction data and trend risk assessment data, providing a scientific basis for dynamic health management and risk intervention.

[0095] The core of health status assessment is to combine the enhanced causal chain model with cross-modal feature data, build a health status mapping model, and generate health status vectors and health status deviation data. The health status mapping model uses the weights and associations of the nodes in the enhanced causal chain model to convert cross-modal feature data into a multidimensional health status vector. Each component of the health status vector corresponds to a health indicator (such as heart rate fluctuations, gait stability). The health status deviation is used to measure the difference between the current health status and the ideal health status. The health status deviation data is generated by calculating the Euclidean distance between the health status vector and the ideal health status vector.

[0096] The health risk path is generated between the nodes of the enhanced causal chain model through the path optimization algorithm, describing the best intervention path from the current health state to the ideal health state. The path optimization is based on the dynamic programming method, with the goal of minimizing the health state deviation, and generates the optimal path from the current node to the ideal state node. The total weight of the path represents the required intervention intensity.

[0097] Health trend prediction is based on time series modeling, capturing the changing trend of health status and combining trend characteristics for risk assessment. Time series modeling generates health trend prediction data by predicting the future value of the health status vector. Common methods include temporal convolutional networks (TCNs), the core of which is to learn time characteristics through historical status data and predict future changes. Health trend prediction data is combined with the weights in the enhanced causal chain model to generate trend risk assessment data, which is used to quantify health risks in the future.

[0098] Preferably, the step of evaluating the health status includes: constructing a health status mapping model based on the enhanced causal chain model and cross-modal feature data to generate a health status vector; generating health status deviation data by measuring the difference between the health status vector and the ideal health state, and generating health status change trend data in combination with historical health status change data.

[0099] In the elderly care scenario, health status is affected by the comprehensive influence of physiological, behavioral and environmental multimodal factors. The key to assessing health status is to build a model that can integrate multimodal data and compare it with the ideal health status. The present invention builds a health status mapping model by enhancing the causal chain model and cross-modal feature data, generates a health status vector to quantify the current health status, and combines health status deviation data and historical health status change data to generate health status change trend data. This method can capture changes in health status in real time and provide accurate risk assessment and intervention basis.

[0100] The health state mapping model uses the node weights and associations in the enhanced causal chain model to map cross-modal feature data into a multidimensional health state vector. Each component of represents a health indicator (such as heart rate fluctuation characteristics, gait stability characteristics).

[0101] Input data to enhance the causal chain model to provide the weights of causal nodes , cross-modal feature data provides feature values . Mapping formula, the calculation formula of the health state vector is: , in, Represents the health status vector Component, indicating the health indicators; Indicates The causal node The weight of each health indicator; Indicates The characteristic value of each causal node; Represents the total number of causal nodes. Through the above formula, the health status mapping model integrates multimodal data into a unified health status vector, reflecting the current overall health status of the elderly.

[0102] Health status deviation is used to quantify the difference between the current health status and the ideal health status, and is a basic indicator for health risk assessment. Definition of ideal health status: Construct an ideal health status vector based on health standards or health benchmark data of elderly individuals. The health state deviation is calculated by the Euclidean distance between the health state vector and the ideal health state vector: , in, Indicates health status deviation; Represents the current health status vector Quantity; The ideal health state vector Quantity; Represents the dimension of the health status vector. The greater the health status deviation, the greater the gap between the current health status and the ideal health status, and more timely intervention measures are needed.

[0103] Health status change trend data is used to reflect the historical change trajectory and future trend of health status. It is generated by combining historical health status data and time series modeling. Through time series analysis, the change trend (such as increase, decrease or fluctuation) of the health status vector is extracted. Linear fitting or polynomial fitting is performed on the historical health status vector to generate trend characteristic values, for example: Rise rate: indicates the speed of improvement of health status; Fluctuation amplitude: reflects the instability of health status. Compare the current health status vector with the historical mean, and the calculation expression is: , in, Indicates Trend change values ​​of health indicators; Represents the current health status vector Quantity; Represents the historical average of health status. Trend change data can provide a basis for subsequent health intervention strategies, such as recommending that the elderly improve their exercise frequency or adjust their living environment.

[0104] The health status vector and health status deviation data can comprehensively and quantitatively describe the health status of the elderly and the gap between them and their ideal state. The health status change trend data captures the dynamic changes of health indicators and provides reliable support for early warning and intervention. Combining health status deviation data and trend change data, the system can formulate personalized health management plans, such as strengthening the monitoring of abnormal heart rate fluctuations or optimizing gait stability.

[0105] In the embodiment, an elderly person shows abnormal heart rate fluctuations and unstable gait during the monitoring period, and the ambient light is low. The system generates the following results by evaluating the health status: Health state mapping: cross-modal feature data, heart rate fluctuation feature value is 15 bpm, gait stability feature value is 0.6, and light intensity feature value is 40 lux; enhanced causal chain model weight, heart rate fluctuation weight is 0.7, gait stability weight is 0.6, and light intensity weight is 0.5.

[0106] Health state vector calculation: ; Ideal health state vector: ; Health status deviation: ; Health status change trend, historical average: heart rate fluctuation average is 12 bpm, gait stability average is 0.7; trend change value: It shows that the heart rate fluctuation is abnormally increased and the gait stability is reduced.

[0107] The system identifies that abnormal heart rate fluctuations are the main factor in the deterioration of health status. Unstable gait and insufficient ambient light may further increase the risk. It is recommended that caregivers increase the frequency of heart rate fluctuation monitoring and improve indoor lighting conditions.

[0108] Preferably, the step of generating health status deviation data includes: combining the current health status vector with the historical health status vector, analyzing the change trend using time series modeling, and generating a health status deviation curve; calculating the state deviation change rate of key nodes to generate health status deviation data.

[0109] In the elderly care scenario, health status deviation data is used to quantify the difference between the current health status of the elderly and the historical health trends, providing in-depth analysis and early warning of health changes. By combining the current health status vector with the historical health status vector, the health status deviation curve is generated using time series modeling, and the state deviation change rate of key nodes is further calculated to provide support for health risk assessment and intervention strategies.

[0110] The health status deviation curve reflects the trend of the health status of the elderly over time and can capture the dynamic characteristics of multidimensional health status. and the current health status vector As the input of time series modeling, the trend of health status over time is extracted. The time series model uses a sliding window method to fit and predict continuous health status data.

[0111] , in, Represents the predicted health state vector at the next time point; Represents the historical health status vector at the current time point and before; A mapping function that represents a time series model, such as a Temporal Convolutional Network (TCN) or an LSTM.

[0112] By current health status vector and historical predicted health status vector The Euclidean distance of , generates the health status deviation curve: , in, Indicates time point Health status deviation value; represents the Euclidean distance.

[0113] The health status deviation change rate is used to measure the extent of changes in health status at key nodes (such as heart rate fluctuations and gait instability) and is a key indicator of health risk. By analyzing the gradient of the health status deviation curve, key nodes with significant changes are determined, such as the time points when heart rate fluctuations increase significantly or gait stability decreases significantly. Perform the first-order difference of the health status deviation value at the key node and calculate the state deviation change rate: , in, Indicates key nodes The rate of change of state deviation; Indicates time point Health status deviation value; Represents the time interval between adjacent key nodes. By calculating the state deviation change rate, the fluctuation intensity of the health state can be quantified, further supporting the quantitative analysis of health risks.

[0114] The health status deviation curve can reflect the dynamic changes in the health status of the elderly in real time, such as increased heart rate fluctuations or decreased gait stability, providing a basis for early risk warning.

[0115] The state deviation change rate quantitatively describes the speed of health state changes and can quickly identify high-risk events, such as acceleration that increases heart rate fluctuations. By analyzing health state deviation data, the system can develop highly targeted intervention measures, such as optimizing ambient light intensity when insufficient light causes unstable gait.

[0116] In the example, an elderly person showed increased heart rate fluctuations and unstable gait during the monitoring period, and the ambient humidity was high. The system generates health status deviation data to complete the following analysis: Health state deviation curve generation: Historical health state vector: , corresponding to heart rate fluctuation, gait stability, and light intensity respectively; current health status vector: ; Predicting health status through time series modeling: ; Health status deviation curve: .

[0117] State deviation change rate calculation: Key node 1 (heart rate fluctuation): The state deviation increases from 2.0 to 4.69, and the change rate is: Key node 2 (gait stability): The state deviation decreases from 0.2 to 0.1, and the rate of change is: .

[0118] The system identified a significant increase in heart rate fluctuations and a decrease in gait stability. Analysis of the rate of change showed that heart rate fluctuations were high-risk change points, while gait stability changes were more moderate. The system recommended strengthening real-time monitoring of heart rate fluctuations and gradually optimizing gait stability.

[0119] Based on health risk path data and health trend prediction data, a behavioral intervention path is generated, and the effectiveness of the intervention path is verified by combining simulation and optimization techniques; user feedback data is collected, and the key node weights of the behavioral intervention path and the associated parameters of the causal chain model are dynamically adjusted to generate optimized behavioral intervention path data and an updated causal chain model.

[0120] The behavioral intervention path is generated based on the health risk path data and health trend prediction data through the generative adversarial game optimization algorithm. Input data: health risk path data, which represents the optimal adjustment path and its cost from the current health state to the ideal health state; health trend prediction data, which describes the trend of health status over time, indicating risk points and optimization directions.

[0121] The behavioral intervention path generation process uses dynamic programming methods to find the least costly adjustment path on the health risk path. Output results: The generated preliminary behavioral intervention path includes multiple key nodes, each of which corresponds to an intervention action (such as increasing walking time, optimizing indoor lighting).

[0122] The generated preliminary behavioral intervention path needs to be verified for effectiveness through simulation technology, and the key nodes in the path need to be optimized using reinforcement learning algorithms. In a virtual simulation environment, each node of the intervention path is simulated and tested. For example, the effect of increasing the light brightness indoors on gait stability and heart rate fluctuations is simulated. The node intervention effect value is calculated based on the simulation results. The reinforcement learning algorithm optimizes the key node parameters in the path based on the simulation results, including adjusting the connection weights and action priorities between nodes. The optimized path is more adaptable and has better execution effect.

[0123] The optimization of the behavior intervention path requires the dynamic adjustment of the key node weights and the associated parameters of the causal chain model in combination with the user's execution feedback data. The system collects the user's behavior completion rate, subjective feelings (such as comfort) and health status change data after executing the intervention path in real time as the basis for dynamic adjustment. Update the node weights and the associated parameters of the causal chain model according to the feedback data. Output results: The dynamically adjusted behavior intervention path data and the updated causal chain model can better adapt to the individual needs of the elderly and improve the intervention effect.

[0124] Preferably, Figure 3 As shown, the steps of generating a behavioral intervention path include: generating preliminary behavioral intervention path data based on health risk path data and health trend prediction data in combination with a generative adversarial game optimization algorithm; verifying the improvement effect of each node in the preliminary behavioral intervention path on the health status through virtual simulation technology, adjusting the node parameters and connection weights in the path, and generating a behavioral intervention path.

[0125] In the elderly care scenario, the generation of behavioral intervention paths aims to plan effective intervention measures to improve the health status of the elderly by analyzing health risk path data and health trend prediction data. The present invention uses a generative adversarial game optimization algorithm to generate preliminary behavioral intervention path data, and verifies the health improvement effect of each key node in the path through virtual simulation technology. Subsequently, the optimized behavioral intervention path is generated by adjusting the node parameters and connection weights in the path.

[0126] The generative adversarial game optimization algorithm dynamically plans the initial version of the intervention path through interactive analysis of health risk path data and health trend prediction data.

[0127] Input data: health risk path data, indicating the adjustment path from the current health status to the ideal health status, including risk nodes and adjustment costs; health trend prediction data, describing the future change trend of health status and identifying nodes that require priority intervention.

[0128] The generative adversarial game optimization algorithm is based on the principle of multi-objective optimization, and considers both health improvement effects and implementation costs in intervention path planning. The generation of the preliminary behavioral intervention path follows the following optimization objectives: , in, Representation Node To Node The weight of represents the intervention effect; Represents a slave node To Node implementation costs; Represents the node in the intervention path of health benefits; Indicates the total number of nodes in the path.

[0129] Output data: The preliminary behavioral intervention pathway consists of several key nodes, each of which corresponds to a specific intervention measure (such as increasing activity time, improving light intensity, etc.).

[0130] The initially generated behavioral intervention path needs to be verified in a virtual simulation environment to evaluate the improvement effect of each node in the path on the health status and optimize the parameters and weights of the path. The simulation environment simulates the actual situation of the elderly executing the intervention path, including the dynamic changes in execution effect and health status. For example, in the node where the light intensity is increased, the simulation environment calculates the impact of light changes on gait stability and heart rate fluctuations. The health improvement effect of each node is quantitatively evaluated and its comprehensive effect value is calculated: , in, Representation Node The health improvement effect value; Representation Node Health characteristics The weight of Representation characteristics of improvement value.

[0131] According to the virtual simulation results, the parameters and connection weights of the nodes in the path are adjusted to optimize the adaptability and implementation effect of the intervention path. Node parameter optimization includes: adjusting the intervention intensity in the node, such as increasing the daily activity time from 5 minutes to 8 minutes; modifying the node priority according to the simulation effect to ensure that efficient intervention is executed first. The connection weights between nodes are updated using reinforcement learning algorithms to balance the health improvement benefits and implementation costs of the intervention path.

[0132] The preliminary behavioral intervention path is dynamically generated through the game optimization algorithm, which can fully consider the health improvement effect and implementation cost, and formulate a scientific intervention plan for the elderly. Through virtual simulation technology, the health improvement effect of each node in the path is evaluated to ensure the feasibility and practical adaptability of the intervention path. Combined with the simulation verification results, the node parameters and weights are adjusted to make the intervention path more in line with the health needs of the elderly and improve the intervention effect.

[0133] In this example, an elderly person has unstable gait and abnormal heart rate fluctuations due to reduced indoor activity time and insufficient ambient light. The system generates a behavioral intervention path and performs simulation and optimization to complete the following steps: Behavioral intervention path generation: Input data: The health risk path shows that it is necessary to prioritize increasing activity time and optimizing light intensity; the health trend forecast shows that heart rate fluctuations may further increase. Output path: Node 1: Increase daily activity time to 5 minutes; Node 2: Increase indoor light brightness to 300 lux.

[0134] Virtual simulation verification: Simulation test: Node 1 increases the activity time to improve gait stability by 0.4 and heart rate fluctuation by 0.3; Node 2 increases the light brightness to improve gait stability by 0.2 and heart rate fluctuation by 0.1. Comprehensive effect value: Node 1 effect value: E1=0.7; Node 2 effect value: E2=0.3.

[0135] Path optimization adjustment: adjust the activity time of node 1 to 8 minutes per day; increase the priority of node 2 to improve the light environment earlier. The optimized behavioral intervention path shows that heart rate fluctuations and gait stability have been significantly improved, the indoor activity time of the elderly has been significantly increased, and health risks have been effectively reduced.

[0136] Preferably, Figure 4 As shown, the step of collecting user feedback data includes: collecting feedback data after the user executes the behavior intervention path, including behavior completion rate, subjective feelings and health status change data; combining the generative adversarial network to expand scarce scenario data, dynamically adjusting the key node parameters of the behavior intervention path and the associated weights of the causal chain model, and generating optimized behavior intervention path data and an updated causal chain model.

[0137] The purpose of collecting user feedback data is to evaluate the effectiveness of the user's implementation of the intervention path and its actual relevance to health status. The collection content includes: Behavior completion rate: records the ratio of the intervention behavior actually completed by the user to the expected plan. For example, if the plan is to walk for 10 minutes and the actual completion is 7 minutes, the behavior completion rate is 70%. Subjective feelings: record the comfort, fatigue or other subjective evaluation after the behavior is executed through user feedback. Health status change data: changes in multimodal data recorded by sensors, such as the improvement of gait stability characteristic values ​​and the reduction of heart rate fluctuations.

[0138] After standardizing the collected feedback data, calculate the specific impact value of the feedback on each node in the intervention path. For example: , in, Representation Node The feedback impact value of It represents the impact value of behavior completion rate; Indicates the subjective impact value; Indicates the impact value of health status change; , , Represents the weight parameter, which is dynamically adjusted according to the scenario.

[0139] Since some scenarios (such as extreme environments and low-frequency events) may have scarce data in actual feedback, the system uses generative adversarial networks (GANs) to expand data to improve the robustness and adaptability of the model. GAN generates high-fidelity simulation data through an adversarial generation mechanism. The generator generates new possible feedback scenarios based on known feedback data, and the discriminator verifies the authenticity of the generated data and optimizes the output effect of the generator. For example, the expanded scenario may include the impact of different humidity levels on gait stability. The expanded simulation data is fused with the actual feedback data to enrich the diversity of feedback data and make the dynamic adjustment process more widely applicable.

[0140] Combined with actual feedback and expanded data, the system dynamically adjusts the key node parameters in the behavior intervention path and the associated weights of the causal chain model. For nodes with low completion rates, reduce the intensity of intervention or adjust the intervention method, such as reducing daily walking time or switching to low-intensity exercise; for nodes with high feedback satisfaction, increase priority and increase intervention intensity. For example, optimizing indoor lighting brightness has a significant intervention effect on gait stability, and the brightness can be increased to a better level. Redistribute the associated weights of nodes in the causal chain based on feedback data: ,

[0141] in, represents the updated node association weight; represents the original node association weight; Indicates the node feedback impact value; Represents the adjustment coefficient, which is used to control the impact of feedback data on weight update.

[0142] By collecting and analyzing user feedback, the behavioral intervention path can be dynamically adjusted according to the actual implementation effect to ensure the personalization and adaptability of the intervention plan. Generative adversarial networks expand the feedback data of low-frequency scenarios and enhance the system's response ability and robustness in scarce scenarios. Dynamically adjusting the association weights of the causal chain model improves the model's adaptability to complex health status changes and provides more accurate support for subsequent health assessment and intervention strategy formulation.

[0143] In the example, an elderly person actually completed part of the intervention behavior while executing the behavior intervention path, and the feedback data showed that his health status had improved. The system collects user feedback data and dynamically adjusts and optimizes the intervention path: Planned path: walk for 10 minutes every day and increase the indoor light brightness to 300 lux. Feedback data: Behavior completion rate: walking completion rate is 60%, light brightness adjustment completion rate is 90%; Subjective feeling: The subjective fatigue score of walking behavior was 3 (out of 5, high fatigue); Changes in health status: The improvement value of gait stability characteristics is 0.3, and the improvement value of heart rate fluctuation characteristics is 0.4.

[0144] Combined with the feedback, the feedback influence value of the walking node is calculated to be 0.25, and the feedback influence value of the light node is 0.4.

[0145] Generate adversarial network expansion data: expand the scene, generate data on the impact of different light levels (250 lux, 350 lux) on gait stability; generate data on the improvement of heart rate fluctuations for different walking times (5 minutes, 8 minutes). Output result: The expanded data shows that the improvement value of walking for 8 minutes on heart rate fluctuations reaches 0.45, which is higher than the current behavior completion.

[0146] Dynamic adjustment and optimization: Path adjustment, adjusting walking time to 8 minutes per day, increasing priority; light nodes remain at 300 lux, but increase monitoring of ambient brightness stability. Causal chain model update: walking node weight increased to 0.8, light node weight updated to 0.7. The optimized behavioral intervention path is more in line with the actual health needs of users, and the system predicts that heart rate fluctuations and gait stability will be further improved in subsequent monitoring cycles.

[0147] like Figure 5As shown, a system for implementing the multi-dimensional data-based safety monitoring and assessment method comprises: A data collection module for collecting physiological data, behavioral data, environmental data, and social interaction data through a distributed sensor network; The data processing module is used to pre-process and prioritize the collected multi-modal data to generate optimized sampling data and event cluster data; A feature extraction and fusion module is used to extract physiological feature data, behavioral feature data and environmental feature data based on optimized sampling data and event cluster data, and fuse the above feature data through a tensor decomposition method to generate cross-modal feature data; The causal inference and chain building module is used to combine the causal inference network, map the cross-modal feature data into causal nodes, generate short-term causal chain data and long-term causal chain data, and generate an enhanced causal chain model through an optimization algorithm; The health status assessment module is used to assess the health status using the enhanced causal chain model and cross-modal feature data, generate health status vectors and health status deviation data; generate health risk path data, health trend prediction data and trend risk assessment data through path optimization and time series modeling; The behavior intervention path generation module is used to generate preliminary behavior intervention path data based on health risk path data and health trend prediction data by generating adversarial game optimization algorithms, and verify the effectiveness of the path in combination with virtual simulation technology to generate behavior intervention paths; The feedback and optimization module is used to collect user feedback data, dynamically adjust the key node weights of the behavior intervention path and the associated parameters of the causal chain model, and generate optimized behavior intervention path data and updated causal chain model.

[0148] The system of the present invention collects physiological data (such as heart rate, blood pressure), behavioral data (such as gait stability, activity frequency), environmental data (such as temperature and humidity, light intensity) and social interaction data in the daily life of the elderly through a distributed sensor network. The collected multimodal data is preprocessed and prioritized to generate optimized sampling data and event cluster data, providing high-quality data input for subsequent analysis.

[0149] The feature extraction and fusion module uses tensor decomposition methods to unify the key features in multimodal data and generate cross-modal feature data. These data are mapped into causal nodes by the causal inference network, and through the construction of short-term causal chains and long-term causal chains, an enhanced causal chain model is generated to capture the complex causal relationship between physiological, behavioral and environmental factors.

[0150] Through the health status assessment module, the system generates health status vectors and health status deviation data, and combines path optimization and time series modeling to generate health risk paths and trend prediction data, providing support for real-time monitoring and forward-looking analysis of health status.

[0151] In the behavioral intervention phase, the system generates behavioral intervention paths based on health risk path data and trend prediction data, and verifies the effectiveness of the paths through virtual simulation technology. At the same time, the system collects user feedback data, and continuously optimizes the paths and models by dynamically adjusting the key node weights of the behavioral intervention paths and the associated parameters of the causal chain model.

[0152] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A safety monitoring and evaluation method based on multidimensional data, characterized in that: The following steps are involved: Collect multimodal data through distributed sensor networks, including physiological data, behavioral data, environmental data, and social interaction data, pre-process and prioritize the collected data, and generate optimized sampling data and event cluster data; Based on optimized sampling data and event cluster data, multimodal feature data is extracted and fused to generate cross-modal feature data; Combine the causal inference network to build a causal chain model, generate short-term causal chain data and long-term causal chain data, and generate an enhanced causal chain model through an optimization algorithm; Using enhanced causal chain models and cross-modal feature data, health status is evaluated and health status vectors and health status deviation data are generated; Generate health risk pathway data, health trend prediction data, and trend risk assessment data through pathway optimization and time series modeling; Generate behavioral intervention paths based on health risk path data and health trend prediction data, and verify the effectiveness of the intervention paths by combining simulation and optimization techniques; Collect user feedback data, dynamically adjust the key node weights of the behavior intervention path and the associated parameters of the causal chain model, and generate optimized behavior intervention path data and updated causal chain model.

2. The safety monitoring and evaluation method based on multidimensional data according to claim 1 is characterized in that: The step of extracting and fusing multimodal feature data comprises: Based on the optimized sampling data and event cluster data, the heart rate fluctuation characteristics and blood pressure change characteristics in the physiological characteristic data, the gait stability characteristics and activity frequency characteristics in the behavioral characteristic data, and the temperature and humidity fluctuation characteristics and light intensity change characteristics in the environmental characteristic data are extracted; the above feature data are fused using the tensor decomposition method to generate cross-modal feature data.

3. The safety monitoring and evaluation method based on multidimensional data according to claim 2 is characterized in that: The tensor decomposition method comprises: The correlation weights between feature data are dynamically adjusted by decomposing the feature matrix; specifically, a third-order tensor is constructed for physiological feature data, behavioral feature data, and environmental feature data, key feature dimensions are extracted through singular value decomposition, and weight parameters are dynamically adjusted to optimize the fusion effect and generate cross-modal feature data.

4. The safety monitoring and evaluation method based on multidimensional data according to claim 1 is characterized in that: The steps of constructing the causal chain model include: The cross-modal feature data is mapped into causal nodes using a causal inference network. Short-term causal chain data is generated based on time series analysis, and long-term causal chain data is generated by combining the cumulative effects of multiple events. The correlation and global weight of the nodes in the causal chain are optimized through a nonlinear collaborative inference algorithm to generate an enhanced causal chain model.

5. The multi-dimensional data-based safety monitoring and assessment method according to claim 4, characterized in that: The nonlinear collaborative inference algorithm includes: A dynamic weight allocation strategy is applied according to the characteristic distribution of causal nodes to adjust the node connection strength of short-term causal chain data; a recursive optimization mechanism is used to correct the cumulative effect of long-term causal chain data, and a multi-objective optimization method is used to calculate the global influence of nodes in the causal chain to generate a dynamically enhanced causal chain model.

6. The safety monitoring and evaluation method based on multidimensional data according to claim 1 is characterized in that: The steps of assessing health status include: A health status mapping model is constructed based on the enhanced causal chain model and cross-modal feature data to generate a health status vector. Health status deviation data is generated by measuring the difference between the health status vector and the ideal health state, and health status change trend data is generated by combining historical health status change data.

7. The safety monitoring and evaluation method based on multidimensional data according to claim 6 is characterized in that: The step of generating health status deviation data comprises: By combining the current health state vector with the historical health state vector, time series modeling is used to analyze the change trend and generate a health state deviation curve; the state deviation change rate of key nodes is calculated to generate health state deviation data.

8. The safety monitoring and evaluation method based on multidimensional data according to claim 1 is characterized in that: The steps of generating a behavioral intervention path include: Based on health risk path data and health trend prediction data, preliminary behavioral intervention path data are generated in combination with the generative adversarial game optimization algorithm. The improvement effect of each node in the preliminary behavioral intervention path on the health status is verified through virtual simulation technology, and the node parameters and connection weights in the path are adjusted to generate a behavioral intervention path.

9. The safety monitoring and evaluation method based on multidimensional data according to claim 1 is characterized in that: The step of collecting user feedback data comprises: Collect feedback data from users after they execute the behavioral intervention path, including behavior completion rate, subjective feelings, and health status change data; combine the generative adversarial network to expand scarce scenario data, dynamically adjust the key node parameters of the behavioral intervention path and the associated weights of the causal chain model, and generate optimized behavioral intervention path data and updated causal chain model.

10. A system for implementing the multi-dimensional data-based safety monitoring and assessment method according to any one of claims 1 to 9, characterized in that: include: A data collection module for collecting physiological data, behavioral data, environmental data, and social interaction data through a distributed sensor network; The data processing module is used to pre-process and prioritize the collected multi-modal data to generate optimized sampling data and event cluster data; A feature extraction and fusion module is used to extract physiological feature data, behavioral feature data and environmental feature data based on optimized sampling data and event cluster data, and fuse the above feature data through a tensor decomposition method to generate cross-modal feature data; The causal inference and chain building module is used to combine the causal inference network, map the cross-modal feature data into causal nodes, generate short-term causal chain data and long-term causal chain data, and generate an enhanced causal chain model through an optimization algorithm; A health status assessment module is used to assess the health status using an enhanced causal chain model and cross-modal feature data, and to generate a health status vector and health status deviation data; Generate health risk pathway data, health trend prediction data, and trend risk assessment data through pathway optimization and time series modeling; The behavior intervention path generation module is used to generate preliminary behavior intervention path data based on health risk path data and health trend prediction data by generating adversarial game optimization algorithms, and verify the effectiveness of the path in combination with virtual simulation technology to generate behavior intervention paths; The feedback and optimization module is used to collect user feedback data, dynamically adjust the key node weights of the behavior intervention path and the associated parameters of the causal chain model, and generate optimized behavior intervention path data and updated causal chain model.

Images