An unmanned aerial vehicle assisted artificial intelligence wisdom health data collection and analysis system

Abstract

The application discloses an unmanned aerial vehicle (UAV) assisted artificial intelligence intelligent health care health data collection and analysis system, relates to the technical field of data processing, and solves the following technical problems: firstly, it is difficult to realize indoor and outdoor centimeter-level precision positioning and dynamic monitoring coverage; secondly, it is difficult to fuse and calibrate multi-dimensional health data in real time; thirdly, it is difficult to construct a dynamic health assessment and risk prediction system, lacks spatiotemporal correlation analysis of health states and environmental factors, cannot update health baselines in real time, cannot accurately calculate health risk indexes and the probability of environment-induced risks, and cannot effectively predict future risks; and finally, it is difficult to optimize health data collection strategies.The application integrates a multi-modal sensor array on a UAV cluster, combines a UWB positioning technology to construct a centimeter-level precision positioning system, and integrates health data collection and processing, analysis, management and other modules to realize intelligent processing of health data in the whole process.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically a drone-assisted artificial intelligence-based smart health and wellness data collection and analysis system. Background Technology

[0002] With the accelerating aging of the global population and the increasing demand for health management, smart elderly care, as an emerging field integrating technologies such as artificial intelligence, the Internet of Things, and mobile healthcare, is gradually becoming an important direction for addressing the shortage of elderly care resources and improving the quality of health services. In smart elderly care scenarios, comprehensive, real-time, and accurate collection and analysis of health data for the elderly and special populations is a core prerequisite for achieving personalized health monitoring, risk warning, and precise intervention. Currently, existing health data collection and analysis technologies mostly rely on fixed monitoring devices such as smart wearable devices, fixed home sensors, or periodic manual collection. While these can obtain some basic health information, their adaptability and level of intelligence in complex scenarios remain significantly limited.

[0003] The existing technologies have the following problems: First, it is difficult to achieve centimeter-level accuracy in indoor and outdoor positioning and dynamic monitoring coverage; second, it is difficult to fuse and calibrate multidimensional health data in real time; third, it is difficult to build a dynamic health assessment and risk prediction system, lacking spatiotemporal correlation analysis of health status and environmental factors, unable to update the health baseline in real time, accurately calculate the health risk index and the probability of environmentally induced risks, and also difficult to effectively predict future risks; finally, it is difficult to optimize health data collection strategies. Summary of the Invention

[0004] Therefore, this application provides an AI-assisted drone-based smart healthcare data collection and analysis system to address the problems existing in the prior art, including:

[0005] First, it is difficult to achieve centimeter-level accuracy in indoor and outdoor positioning and dynamic monitoring coverage; second, it is difficult to integrate and calibrate multidimensional health data in real time; third, it is difficult to build a dynamic health assessment and risk prediction system, lacking spatiotemporal correlation analysis of health status and environmental factors, unable to update health baselines in real time, accurately calculate health risk indices and the probability of environmentally induced risks, and also difficult to effectively predict future risks; finally, it is difficult to optimize health data collection strategies.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] Firstly, a drone-assisted AI-powered smart healthcare data collection and analysis system includes the following modules:

[0008] Multimodal deployment module: The drone swarm is equipped with a multimodal sensor array to build a centimeter-level precision indoor and outdoor positioning system based on UWB positioning technology;

[0009] Health data collection and processing module: dynamically adjusts the monitoring range based on user activity patterns using an adaptive cruise path planning algorithm; utilizes a multi-sensor data fusion algorithm, combined with Kalman filtering and attention mechanisms, to process and calibrate the collected multi-dimensional health data in real time;

[0010] Health data analysis module: Constructs a spatiotemporal correlation model based on multidimensional health data; establishes a dynamic baseline model and calculates the real-time health risk index and health monitoring coverage; utilizes drones for environmental monitoring, calculates the comprehensive environmental stress index and the probability of environmentally induced health risks; constructs a health knowledge graph based on the OWL ontology; and uses a time series prediction model based on the health knowledge graph to predict future risks.

[0011] Health data management module: Introduces reinforcement learning mechanism to optimize drone health data collection strategy.

[0012] Optionally, the monitoring range can be dynamically adjusted based on user activity patterns using an adaptive cruise path planning algorithm, including the following steps:

[0013] Based on historical and real-time UWB positioning data, a dynamically updated heatmap of user spatiotemporal probability distribution is generated.

[0014] With the goal of covering the user's activity area, and taking into account the drone's endurance constraints, an objective function is defined, and an intelligent optimization algorithm is used to calculate the optimal observation point sequence in real time.

[0015] Dynamically adjust the path planning based on real-time heat maps;

[0016] The drone is controlled to fly to the target location according to a preset path or real-time command, collect user health data in different areas, and then return to the designated location after the data collection is completed.

[0017] Optionally, a spatiotemporal correlation model can be constructed based on multidimensional health data, including the following steps:

[0018] Drones acquire centimeter-level location information of users through UWB positioning technology and combine it with timestamps to add spatiotemporal tags to various health data; at the same time, they integrate environmental data and user behavior data to form a multi-dimensional spatiotemporal health dataset.

[0019] Utilize graph neural networks or spatiotemporal convolutional networks to construct a spatiotemporal correlation model of user health status;

[0020] By inputting a multi-dimensional spatiotemporal health dataset, a spatiotemporal correlation model is trained, and a user's health status representation vector is output.

[0021] Optionally, a dynamic baseline model is established to update the health baseline in real time and calculate the real-time health risk index and health monitoring coverage, including the following steps:

[0022] A dynamic baseline model is established based on the historical health status representation vector sequence output by the spatiotemporal correlation model.

[0023] Using a sliding time window to extract N consecutive state vectors as input, the number of clusters K is optimized by the silhouette coefficient method, and the initial health baseline is determined by K-means clustering.

[0024] The Mahalanobis distance D between the current health status representation vector and the health baseline is calculated in real time. When D is greater than or equal to 1.5 times the initial standard deviation, a density-based spatiotemporal clustering algorithm is used to cluster the location coordinates of abnormal events to generate a heat map of abnormal areas.

[0025] The formula for calculating the real-time health risk index is:

[0026] in, Let be the dynamic baseline value of the i-th health parameter at time t. Let be the i-th health parameter value at time t. It is the historical standard deviation of the i-th health parameter. is the weight of the i-th health parameter, n is the total number of health parameters; y is the area of ​​the abnormal region in the heat map; ;

[0027] Formula for calculating health surveillance coverage:

[0028] in, It is the total monitoring duration. It is the total area of ​​the user activity heatmap. For real-time health risk index; For the real-time flight speed of the drone, That is the maximum flight speed of the drone. It is the signal-to-noise ratio. It is the maximum signal-to-noise ratio.

[0029] Alternatively, drones can be used for environmental monitoring to calculate a comprehensive environmental stress index and the probability of environmentally induced health risks, including the following steps:

[0030] Based on the environmental data monitored by drones, the spatiotemporal environment tensor is generated by organizing the data according to time series and spatial location.

[0031] The spatiotemporal environmental tensor and user health data are analyzed using the sparse group Lasso regression model to automatically identify environmental factors that affect health status.

[0032] Calculate the comprehensive environmental stress index

[0033] in, This represents the measured value of the j-th environmental factor at time t. and These are the historical mean and standard deviation of environmental factor j, respectively. Let be the weight of the j-th environmental factor; m is the number of environmental factors.

[0034] The Sigmoid function is used to perform a nonlinear transformation on the environmental stress index to calculate the probability of environmentally induced health risks. The formula is as follows:

[0035] in, It is a comprehensive environmental stress index. It is a real-time health risk index. and These are the adjustment coefficients.

[0036] Further, optionally, the health knowledge graph constructed based on the OWL ontology includes the following steps:

[0037] Based on health data collected by drones, health-related entities, attributes, and relationships are extracted using data processing technology.

[0038] Classes, attributes, and relationships are defined using the OWL ontology language to construct the ontology model and form the schema layer;

[0039] The extracted content is instantiated based on the model and stored in a graph database to generate an initial knowledge graph.

[0040] A streaming data incremental update algorithm is adopted. When the drone collects new health data, it inserts the map through SPARQL Update and uses timestamps or credibility rules to handle conflicts.

[0041] The updated graph deployment graph attention network learns the embedded representations of nodes and edges, and an incremental embedding algorithm is used to optimize the dynamic update.

[0042] Optionally, based on a health knowledge graph, a time series prediction model is used to predict future risks, including the following steps:

[0043] Extract user's historical health status representation vector sequence, associated environmental factor time series data, health risk index change records and entity relationships from the health knowledge graph to construct a feature set for time series prediction.

[0044] Long Short-Term Memory Networks or Temporal Graph Neural Networks are used as time series prediction models. The feature set is input into the model step by step, and the entity embedding and relation embedding of the health knowledge graph serve as prior knowledge constraints for the model.

[0045] The trained model outputs a predicted sequence of health status representation vectors for the user over the next T time units, and calculates the corresponding future health risk index trend curve.

[0046] By combining the causal relationship between environmental factors and health risks in the health knowledge graph, we can identify the top three key environmental factors that contribute to future health risks and their corresponding impact weights.

[0047] Based on the predicted future health risk index and the probability of environmentally induced health risks, the probability of potential abnormal health events is calculated, and the spatial region and time window of the abnormal event are predicted by reasoning the spatiotemporal relationship of the health knowledge graph.

[0048] Optionally, the health data management module includes the following steps: introducing the Q-learning algorithm from reinforcement learning to optimize the data collection strategy, using health monitoring coverage as the reward function, and dynamically adjusting the drone sampling frequency; using a federated learning framework to store distributed data, and combining blockchain technology to support data traceability and access control.

[0049] Compared with the prior art, this application has at least the following beneficial effects:

[0050] This invention achieves centimeter-level indoor and outdoor positioning through UWB positioning technology. Combined with an adaptive cruise path planning algorithm, it can dynamically adjust the monitoring range according to the user's activity patterns, solving the problems of fixed monitoring range and many blind spots in existing technologies, and improving the targeting and coverage of health data collection.

[0051] This invention employs a multi-sensor data fusion algorithm to process and calibrate multidimensional health data in real time. It constructs a spatiotemporal correlation model to mine the spatiotemporal patterns of health data, establishes a dynamic baseline model to update the health baseline in real time and calculate the risk index, combines environmental monitoring to assess environmentally induced risks, and constructs a health knowledge graph based on the OWL ontology and uses a time series prediction model. This invention can not only accurately assess the current health status but also predict future health risks.

[0052] This invention optimizes the data collection strategy of unmanned aerial vehicles by introducing a reinforcement learning mechanism, making the collection process more intelligent and efficient. At the same time, through the dynamic updating and management of the knowledge graph, it realizes the accumulation and reuse of health knowledge, thereby improving the overall intelligence level of the smart health and wellness system. Attached Figure Description

[0053] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0054] Figure 1 This is a schematic diagram of a module structure provided in one embodiment of this application. Detailed Implementation

[0055] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] Please see Figure 1 The first aspect of this invention provides an unmanned aerial vehicle (UAV)-assisted artificial intelligence-based smart health and wellness data collection and analysis system, comprising the following modules:

[0057] Multimodal deployment module: The drone swarm is equipped with a multimodal sensor array to build a centimeter-level precision indoor and outdoor positioning system based on UWB positioning technology;

[0058] Health data collection and processing module: dynamically adjusts the monitoring range based on user activity patterns using an adaptive cruise path planning algorithm; utilizes a multi-sensor data fusion algorithm, combined with Kalman filtering and attention mechanisms, to process and calibrate the collected multi-dimensional health data in real time;

[0059] Health data analysis module: Constructs a spatiotemporal correlation model based on multidimensional health data; establishes a dynamic baseline model and calculates the real-time health risk index and health monitoring coverage; utilizes drones for environmental monitoring, calculates the comprehensive environmental stress index and the probability of environmentally induced health risks; constructs a health knowledge graph based on the OWL ontology; and uses a time series prediction model based on the health knowledge graph to predict future risks.

[0060] Health data management module: Introduces reinforcement learning mechanism to optimize drone health data collection strategy.

[0061] Specifically, a cluster system of 5-10 small drones is employed, each equipped with a multimodal sensor array including: UWB positioning tags, infrared physiological sensors, environmental sensors (temperature, humidity, PM2.5, noise, etc.), and high-definition cameras. Based on UWB positioning technology, 4-6 UWB base stations are deployed indoors and outdoors in the health and wellness community. The distances between the drones and base stations, and between the drones and the UWB tags worn by users, are calculated using the Time-of-Flight (TOF) ranging principle. Combined with triangulation, centimeter-level positioning accuracy (±3cm) is achieved, providing a spatiotemporal reference for subsequent data collection. Based on the collected data and user activity patterns, the monitoring range is dynamically adjusted using an adaptive cruise path planning algorithm. A multi-sensor data fusion algorithm, combined with Kalman filtering and attention mechanisms, is used to process and calibrate the collected multidimensional health data in real time. Kalman filtering is used to process the respiratory signals from millimeter-wave radar, suppressing motion artifacts. An attention mechanism is used, dynamically allocating sensor weights through a Transformer architecture; for example, when infrared thermal imaging detects localized high temperatures, the weight of that channel's data is automatically increased to reduce environmental noise interference. Based on the processed data, a spatiotemporal correlation model is constructed using multidimensional health data; a dynamic baseline model is established and real-time health risk indices and health monitoring coverage are calculated; environmental monitoring is conducted using drones to calculate the comprehensive environmental stress index and the probability of environmentally induced health risks; a health knowledge graph is constructed based on the OWL ontology; and a time series prediction model is used based on the health knowledge graph to predict future risks. Through the synergy of these four modules, the system achieves a closed-loop process from centimeter-level positioning, dynamic data acquisition, multi-dimensional analysis to intelligent management, solving problems such as blind spots in existing technologies and insufficient data fusion, and improving the accuracy and intelligence level of smart healthcare.

[0062] Example 1: Based on user activity patterns, the monitoring range is dynamically adjusted using an adaptive cruise path planning algorithm, including the following steps:

[0063] Based on historical and real-time UWB positioning data, a dynamically updated heatmap of user spatiotemporal probability distribution is generated.

[0064] With the goal of covering the user's activity area, and taking into account the drone's endurance constraints, an objective function is defined, and an intelligent optimization algorithm is used to calculate the optimal observation point sequence in real time.

[0065] Dynamically adjust the path planning based on real-time heat maps;

[0066] The drone is controlled to fly to the target location according to a preset path or real-time command, collect user health data in different areas, and then return to the designated location after the data collection is completed.

[0067] Specifically, the system uses UWB positioning modules mounted on drone swarms to receive signals from UWB tags worn by users in real time, with a positioning accuracy of ±3cm. It also retrieves user location data from the past 7 days (including location coordinates and timestamps from 6:00 AM to 10:00 PM daily). Employing a kernel density estimation algorithm, the system maps the user's indoor and outdoor activity locations onto a two-dimensional plane at 5-minute intervals, generating a dynamically updated spatiotemporal probability distribution heatmap.

[0068] For example, in the heat map, the "garden area (8:00-10:00)" is a red high-probability area (users concentrate on exercising here in the morning), the "bedroom area (22:00-6:00)" is a yellow medium-probability area (users rest at night), and the inactive area is a blue low-probability area.

[0069] The calculation of the optimal observation point sequence involves: taking the coverage of 90% of the high-probability areas in the user activity heatmap as the core objective, and combining the drone's endurance constraint (a single drone has a flight time of 30 minutes, and 5 minutes of battery power needs to be reserved for round trips to the base station), defining the objective function as: F = α × coverage rate + β × endurance utilization rate - γ × path overlap, where α, β, and γ are weighting coefficients (set to 0.6, 0.3, and 0.1 respectively). The coverage rate reflects the proportion of high-probability areas covered, the endurance utilization rate is the ratio of actual flight time to maximum endurance time, and the path overlap rate avoids duplicate flights between drones. A particle swarm optimization algorithm (population size 50, iterations 30) is used to calculate the optimal observation point sequence in real time.

[0070] For example, for the morning garden area, three observation points are planned (garden entrance, central square, and rest pavilion) to ensure that the drone completes data collection and returns to the base station within 25 minutes. The system updates the user's spatiotemporal probability distribution heatmap every 10 minutes, and triggers a real-time path adjustment mechanism when a change in the user's activity area is detected.

[0071] For example, if the heat value of the "Garden Center Plaza" observation point drops from 0.8 to 0.3, the system automatically removes this point and adds "Living Room Center" as a new observation point. The system also recalculates the flight path (shortening the flight distance to the garden and extending the flight path indoors) to ensure that the coverage of high-probability areas (≥0.7) in the new heat map remains above 90%. The UAV ground control center sends flight commands (including target coordinates and flight speed ≤5m / s) to the designated UAV based on the adjusted path plan. After arriving at the "Living Room Center" observation point, the UAV collects user heart rate and body surface temperature data using its onboard infrared physiological sensor, while simultaneously recording the UWB positioning coordinates (e.g., living room (x=4.5m, y=3.2m)) and timestamp (10:15:30). After a single UAV completes data collection at three observation points, the system automatically triggers a return-to-home command, controlling the UAV to return to the charging base station and uploading the encrypted health data to the health data processing module via the 5G network.

[0072] Example 2: Constructing a spatiotemporal correlation model based on multidimensional health data, including the following steps:

[0073] Drones acquire centimeter-level location information of users through UWB positioning technology and combine it with timestamps to add spatiotemporal tags to various health data; at the same time, they integrate environmental data and user behavior data to form a multi-dimensional spatiotemporal health dataset.

[0074] Utilize graph neural networks or spatiotemporal convolutional networks to construct a spatiotemporal correlation model of user health status;

[0075] By inputting a multi-dimensional spatiotemporal health dataset, a spatiotemporal correlation model is trained, and a user's health status representation vector is output.

[0076] Specifically, the drone swarm receives signals from the UWB tags worn by users in real time via the UWB positioning module, obtaining centimeter-level location information and recording timestamps simultaneously. For the various health data collected, the system automatically adds spatiotemporal tags (example: {heart rate: 82, location: (3.2, 2.5), time: 08:30:22.156}).

[0077] Meanwhile, the environmental sensors on the drone collect environmental data such as temperature and humidity, PM2.5 concentration, and noise levels. This data is then combined with user behavior data identified by the computer vision module and fused with the aforementioned spatiotemporally labeled health data to form a multi-dimensional spatiotemporal health dataset.

[0078] The system employs a spatiotemporal convolutional network to construct a spatiotemporal correlation model of user health status. The model's input layer consists of 30-minute continuous time-series data segments. The feature dimensions for each time step include: location features: the user's average location coordinates, distance traveled, and duration of stay within that time period; physiological features: the mean and standard deviation of indicators such as heart rate, blood oxygen, and blood pressure; environmental features: the rate of change of the mean values ​​of environmental factors such as temperature, humidity, and PM2.5; and behavioral features: the type of behavior and its duration. The model's intermediate layers contain three spatiotemporal convolutional blocks, each consisting of one temporal convolutional layer and one spatial convolutional layer. Temporal convolution captures the changing trends of health data over time, while spatial convolution mines the correlation between location and health status (e.g., blood oxygen levels are generally higher in garden areas than indoors). The output layer is a fully connected layer that compresses the multidimensional features into a 128-dimensional user health status representation vector.

[0079] Using a multi-dimensional spatiotemporal health dataset of 50 users from a senior living community over 30 consecutive days as training samples, the Adam optimizer was employed to train the spatiotemporal correlation model. During training, the cross-entropy loss between the model's output representation vector and manually labeled health states (stable, slightly fluctuating, and abnormal) was used as the optimization objective. Training stopped when the loss function converged. After model deployment, the latest 30 minutes of multi-dimensional spatiotemporal health data from users was input in real time, outputting a 128-dimensional health state representation vector. Each dimension value in this vector, after normalization (range 0-1), corresponds to heart rate stability, the correlation between location movement and health, and environmental adaptability, comprehensively reflecting the user's health state within a specific spatiotemporal range.

[0080] Example 3: Establishing a dynamic baseline model to update the health baseline in real time and calculate the real-time health risk index and health monitoring coverage, including the following steps:

[0081] A dynamic baseline model is established based on the historical health status representation vector sequence output by the spatiotemporal correlation model.

[0082] Using a sliding time window to extract N consecutive state vectors as input, the number of clusters K is optimized by the silhouette coefficient method, and the initial health baseline is determined by K-means clustering.

[0083] The Mahalanobis distance D between the current health status representation vector and the health baseline is calculated in real time. When D is greater than or equal to 1.5 times the initial standard deviation, a density-based spatiotemporal clustering algorithm is used to cluster the location coordinates of abnormal events to generate a heat map of abnormal areas.

[0084] The formula for calculating the real-time health risk index is:

[0085] in, Let be the dynamic baseline value of the i-th health parameter at time t. Let be the i-th health parameter value at time t. It is the historical standard deviation of the i-th health parameter. is the weight of the i-th health parameter, n is the total number of health parameters; y is the area of ​​the abnormal region in the heat map; ;

[0086] Formula for calculating health surveillance coverage:

[0087] in, It is the total monitoring duration. It is the total area of ​​the user activity heatmap. For real-time health risk index; For the real-time flight speed of the drone, That is the maximum flight speed of the drone. It is the signal-to-noise ratio. It is the maximum signal-to-noise ratio.

[0088] Specifically, based on the 128-dimensional health status representation vector sequence output by the spatiotemporal correlation model in Example 2, historical data of a user for 14 consecutive days (1344 vectors generated every 30 minutes for each day, 24 hours a day) was selected as the sample. A sliding time window mechanism was used to extract the data, with a window size of N=24 (i.e., a vector sequence of 12 consecutive hours, corresponding to 48 vectors) and a window sliding step of 4 (i.e., the input was updated every 2 hours). The number of clusters K was optimized using the silhouette coefficient method: the silhouette coefficient was calculated by traversing the range of K=2 to K=8. When K=3, the silhouette coefficient was the highest, and the optimal number of clusters was determined to be 3. The K-means clustering algorithm was used to cluster the historical vector sequence, resulting in 3 cluster centers, corresponding to "stable state", "mildly fluctuating state", and "potentially risky state", respectively. These 3 center vectors were used as the initial health baseline. After model deployment, the system receives the user's current health status representation vector in real time and calculates the Mahalanobis distance D between this vector and the three types of baselines in the dynamic baseline model (the covariance matrix is ​​used to correct the correlation of features in different dimensions, avoiding the influence of single-dimensional bias on the judgment). The initial standard deviation is set to the overall standard deviation of the historical vector sequence. When the real-time D ≥ 1.5 * initial standard deviation, an anomaly detection mechanism is triggered. The density-based spatiotemporal clustering algorithm DBSCAN is used, with a neighborhood radius of 2 meters and a minimum sample size of 5. Clustering is performed on the UWB positioning coordinates and living room of all abnormal events within the past 10 minutes, generating an anomaly area heatmap: the area with the highest cluster density is marked in red, and the area with the second highest density is marked in yellow, visually presenting the concentrated areas of anomalies. Health parameters i include heart rate, blood pressure, blood oxygen, etc., and their weights wᵢ are determined using the analytic hierarchy process (e.g., heart rate w1 = 0.3, blood pressure w2 = 0.25, and the weights of other parameters decrease sequentially), satisfying Σwᵢ = 1; or the weights are determined using the entropy weight method. The dynamic baseline value bᵢ(t) of the i-th health parameter at time t is extracted from the features of the corresponding cluster center in the dynamic baseline model; the real-time health parameter Kᵢ(t) is the current data collected by the drone; the historical standard deviation σᵢ is the standard deviation of the fluctuation of this parameter for the user over the past 30 days; the area y of the abnormal region heatmap is calculated by counting pixels and then substituted into the formula to finally obtain τ(t). The signal-to-noise ratio is usually defined as the ratio of signal power to noise power. The health monitoring coverage rate is finally obtained by substituting various parameters into the formula.

[0089] Example 4: Using drones for environmental monitoring to calculate the comprehensive environmental stress index and the probability of environmentally induced health risks, including the following steps:

[0090] Based on the environmental data monitored by drones, the spatiotemporal environment tensor is generated by organizing the data according to time series and spatial location.

[0091] The spatiotemporal environmental tensor and user health data are analyzed using the sparse group Lasso regression model to automatically identify environmental factors that affect health status.

[0092] Calculate the comprehensive environmental stress index

[0093] in, This represents the measured value of the j-th environmental factor at time t. and These are the historical mean and standard deviation of environmental factor j, respectively. Let be the weight of the j-th environmental factor; m is the number of environmental factors.

[0094] The Sigmoid function is used to perform a nonlinear transformation on the environmental stress index to calculate the probability of environmentally induced health risks. The formula is as follows:

[0095] in, It is a comprehensive environmental stress index. It is a real-time health risk index. and These are the adjustment coefficients.

[0096] Specifically, the drone swarm is equipped with temperature and humidity sensors, PM2.5 detectors, noise meters, and light sensors to monitor the environment, collecting data every 5 minutes. The collected environmental data includes temperature, humidity, PM2.5 concentration, noise levels, and light intensity, and simultaneously records the collection timestamp and UWB positioning coordinates. The system organizes the data according to a three-dimensional structure of "time-space-environmental factors": the time dimension is divided into 1-hour units, with 24 time slices per day; the spatial dimension corresponds to the monitoring area; and the environmental factor dimension includes environmental parameters, forming a spatiotemporal environmental tensor.

[0097] For example, the tensor slice data for "time slice 8:00-9:00, garden area" is [26℃, 55%, 30μg / m 3 [45dB, 800lux], intuitively presenting environmental characteristics under different times and spaces.

[0098] User health data, which is time-matched to environmental data from the same period, is retrieved. This data includes health status representation vectors output by the spatiotemporal correlation model and real-time health risk indices. These data, along with the spatiotemporal environmental tensor, are input into a sparse cluster Lasso regression model. The model divides environmental factors into three "factor groups" (e.g., temperature + humidity as the "microclimate group," PM2.5 as the "air quality group," and noise + light as the "sensory stimulus group"). Through a combination of L1 regularization controlling overall sparsity and L2 regularization controlling intra-group sparsity, the model automatically filters out environmental factors that significantly affect health status.

[0099] Using one month of monitoring data from a community, with 720 time slices × 10 areas × 5 environmental factors as training samples, the regression coefficients output by the model show that PM2.5 (absolute value of coefficient 0.82), temperature (0.65), and humidity (0.43) are the key influencing factors (absolute value of coefficient > 0.4), while the influence of noise and light is relatively weak (absolute value of coefficient < 0.2). Therefore, these three factors are included in the subsequent calculations.

[0100] Based on the formula for the comprehensive environmental stress index, taking environmental monitoring data at a certain time t (e.g., 10:00) as an example, the following calculation is performed: The environmental factor measurement value Eⱼ(t) is calculated using historical mean μⱼ and standard deviation σⱼ obtained from data collected over the past 90 days. The factor weights wⱼ are determined based on the coefficient proportions of the sparse group Lasso regression results: w1=0.3 (temperature), w2=0.2 (humidity), w3=0.5 (PM2.5), and satisfy Σwⱼ=1. Substituting these values ​​into the formula, Y(t) is calculated, which is the comprehensive environmental stress index Y(t). The higher the value, the greater the potential health pressure from the environment.

[0101] Based on the aforementioned Sigmoid function, and combined with the real-time health risk index τ(t) for the same period, the time derivatives of the environmental stress index and the health risk index are first calculated: dY(t) / dt is the rate of change of Y(t) in the past hour; dτ(t) / dt is the rate of change of the health risk index in the past hour; the adjustment coefficients α and β are determined by calibration with historical data, α=0.8 and β=1.2, with the weights biased towards the changing trend of health risk itself, and can be adjusted according to historical data; the probability of environmentally induced health risk is then calculated.

[0102] Example 5: A health knowledge graph constructed based on OWL ontology, including the following steps:

[0103] Based on health data collected by drones, health-related entities, attributes, and relationships are extracted using data processing technology.

[0104] Classes, attributes, and relationships are defined using the OWL ontology language to construct the ontology model and form the schema layer;

[0105] The extracted content is instantiated based on the model and stored in a graph database to generate an initial knowledge graph.

[0106] A streaming data incremental update algorithm is adopted. When the drone collects new health data, it inserts the map through SPARQL Update and uses timestamps or credibility rules to handle conflicts.

[0107] The updated graph deployment graph attention network learns the embedded representations of nodes and edges, and an incremental embedding algorithm is used to optimize the dynamic update.

[0108] Specifically, health-related entities, attributes, and relationships are extracted based on multi-dimensional data collected by drones, including user physiological data (heart rate, blood pressure, blood oxygen, etc.); environmental data (temperature, PM2.5, etc.); behavioral data (sitting, walking, etc.); and spatiotemporal data (location coordinates, timestamps, etc.). Information is extracted through data processing techniques: named entity recognition models such as BERT fine-tuning models are used to extract entities from textual data, including users, health parameters, environmental factors, behavioral states, spatiotemporal locations, etc.

[0109] Entity attributes are extracted using rule matching and entity linking techniques;

[0110] For example, the attributes of "User A" are "Age: 75 years old" and "Underlying disease: Hypertension", and the attributes of "Heart rate" are "Measurement value: 75 beats / min" and "Unit: beats / min";

[0111] Based on dependency parsing, identify the relationships between entities;

[0112] Examples include "User A - has - heart rate", "heart rate - affected by - temperature", "User A - performs - meditation", "meditation - occurs in - bedroom", etc.

[0113] The OWL DL is used to describe logical subsets, defining classes, attributes, and relationships to form the schema layer of the knowledge graph. Specifically, this includes: defining core classes such as "User," "HealthParameter," "EnvironmentalFactor," "Behavior," and "SpatialPosition," and setting class hierarchies using "rdfs:subClassOf" (e.g., "VitalSign" is a subclass of "HealthParameter," containing heart rate, blood pressure, etc.); connecting individuals with basic data types, such as "hasAge" with domain "User" and value domain "xsd:integer"; "hasValue" with domain "HealthParameter" and value domain "xsd:float"; and object attributes connecting two individuals, such as "hasHealthParameter" with domain "User" and value domain "HealthParameter"; and "isAffectedBy." The domain is "HealthParameter" and the value domain is "EnvironmentalFactor"; the domain of "occursIn" is "Behavior" and the value domain is "SpatialPosition".

[0114] The extracted entities, attributes, and relationships are instantiated based on the ontology model:

[0115] For example: Individual instance: "UserA", "HeartRate1 (heart rate instance)", "PM2.5_30 (PM2.5 instance, value 30μg / m³)" 3 "Sitting (sit-down instance)" "Bedroom (bedroom instance)"; Attribute assignment: Bind data attributes to individuals, for example: "UserA hasAge 75" "HeartRate1 hasValue 75";

[0116] Relationship binding: Associating individuals through object attributes, for example: "UserA hasHealthParameterHeartRate1", "HeartRate1 isAffectedBy PM2.5_30", "Sitting occursIn Bedroom".

[0117] Instantiated data is stored in the Neo4j graph database to form an initial health knowledge graph, which is visualized as a network of relationships between "user - health parameters - environment - behavior - location".

[0118] A streaming data incremental update algorithm is adopted. When the drone collects new health data, it inserts the map through SPARQL Update and uses timestamps or credibility rules to handle conflicts.

[0119] Perform incremental inserts using SPARQL statements, for example: INSERT DATA { <usera> <hashealthparameter> <heartrate2> . <heartrate2> <hasvalue> 85 . <heartrate2> <occursin> <livingroom>};

[0120] If new data conflicts with historical data, the latest data is retained using timestamp rules, or sensor data is prioritized based on reliability rules such as a drone sensor accuracy of 95% being higher than the 80% of manually recorded data. Based on the updated knowledge graph, a graph attention network is deployed to learn the embedded representations of nodes and edges.

[0121] Taking all nodes in the graph as input, and using the attention mechanism of GAT, the output is a 128-dimensional node embedding vector, which reflects the semantic and structural features of the nodes in the graph.

[0122] When the graph is incrementally updated, an incremental embedding algorithm is used. Only the embeddings of newly added nodes and affected neighboring nodes are recalculated, while the embeddings of other nodes remain unchanged. Compared with full retraining, the computational efficiency is improved by more than 60%.

[0123] Example 6: Based on a health knowledge graph, a time series prediction model is used to predict future risks, including the following steps:

[0124] Extract user's historical health status representation vector sequence, associated environmental factor time series data, health risk index change records and entity relationships from the health knowledge graph to construct a feature set for time series prediction.

[0125] Long Short-Term Memory Networks or Temporal Graph Neural Networks are used as time series prediction models. The feature set is input into the model step by step, and the entity embedding and relation embedding of the health knowledge graph serve as prior knowledge constraints for the model.

[0126] The trained model outputs a predicted sequence of health status representation vectors for the user over the next T time units, and calculates the corresponding future health risk index trend curve.

[0127] By combining the causal relationship between environmental factors and health risks in the health knowledge graph, we can identify the top three key environmental factors that contribute to future health risks and their corresponding impact weights.

[0128] Based on the predicted future health risk index and the probability of environmentally induced health risks, the probability of potential abnormal health events is calculated, and the spatial region and time window of the abnormal event are predicted by reasoning the spatiotemporal relationship of the health knowledge graph.

[0129] Specifically, from the health knowledge graph stored in the Neo4j graph database, multi-dimensional time-series data and entity relationships related to user A are extracted: 128-dimensional health status representation vectors output by the spatiotemporal correlation model over the past 30 days are retrieved, covering features such as heart rate stability and blood pressure fluctuation; time-series sequences of PM2.5 concentration, temperature, and humidity during the same period are extracted and associated with entity relationships such as "PM2.5 - affects - blood pressure" and "temperature - affects - heart rate" in the knowledge graph; the real-time health risk index τ(t) sequence over the past 30 days is obtained; key relationships such as "user A - has - hypertension", "hypertension - susceptible to - PM2.5", and "user A - nighttime activities in - bedroom" are extracted from the knowledge graph and converted into binary feature vectors. The above data are aligned in 1-hour time steps to form a multi-dimensional feature set containing "health representation vector + environmental factors + risk index + relationship features". Each sample corresponds to a comprehensive feature set for 1 hour, used for model training and prediction.

[0130] A prediction model is constructed using a Long Short-Term Memory (LSTM) network combined with knowledge graph embedding features. The specific settings are as follows: the input layer receives a feature set from the previous 12 hours; entity embeddings output by a graph attention network are used as prior knowledge, fused with the hidden states of an LSTM through a fully connected layer to constrain the model to focus on key associations; the feature set from the past 30 days is used as training data, with the first 25 days as the training set and the last 5 days as the validation set. The Adam optimizer is used, with the mean squared error between the predicted health status representation vector and the actual vector as the loss function. The model converges after 50 training rounds. After deployment, the model takes the latest 12-hour feature set from user A as input and outputs a predicted sequence of health status representation vectors for the next 24 hours. Using the dynamic baseline model in Example 3, the predicted vectors are converted into a real-time health risk index, generating a trend curve.

[0131] Based on the causal relationship of "environmental factors - influence - health status" in the health knowledge graph, the contribution of each environmental factor to future risk is calculated using the attention weights of the graph attention network. Combining the predicted future health risk index with the environmental induced risk probability calculated in Example 4, the probability of an abnormal event is calculated using a weighted formula: Abnormal probability = 0.6 × τ(t) + 0.4 × P(t);

[0132] Based on the spatiotemporal relationships in the knowledge graph, SPARQL query reasoning is used to deduce the spatial region and time window where anomalies occur. For example, anomalies are most likely to occur in the bedroom area, with a time window of 22:00-23:00.

[0133] Example 7: The health data management module includes the following steps: introducing the Q-learning algorithm from reinforcement learning to optimize the data collection strategy, using health monitoring coverage as the reward function, and dynamically adjusting the sampling frequency of the drone; using a federated learning framework to store distributed data, and combining blockchain technology to support data traceability and access control.

[0134] Specifically, the Q-learning algorithm is introduced, with a reward function based on the monitoring coverage rate, such as a target of ≥90%. When the coverage rate is below 85%, the drone sampling frequency is increased from 5 minutes / time to 2 minutes / time; a positive reward is given when the coverage rate is ≥90%, and a negative reward is given otherwise, dynamically adjusting the drone sampling frequency. A federated learning framework is adopted, where user health data is processed locally on the terminal and only model parameters are uploaded. Blockchain technology is used to record the drone ID, time, and hash value of the data collector, enabling traceability and access control.

[0135] The technical features of the above embodiments can be combined in any way, as long as there is no contradiction in the combination of these technical features. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.< / livingroom> < / occursin> < / heartrate2> < / hasvalue> < / heartrate2> < / heartrate2> < / hashealthparameter> < / usera>

Claims

1. An unmanned aerial vehicle-assisted artificial intelligence health data collection and analysis system, characterized in that, Includes the following modules: Multimodal deployment module: The drone swarm is equipped with a multimodal sensor array to build a centimeter-level precision indoor and outdoor positioning system based on UWB positioning technology; Health data collection and processing module: dynamically adjusts the monitoring range based on user activity patterns using an adaptive cruise path planning algorithm; utilizes a multi-sensor data fusion algorithm, combined with Kalman filtering and attention mechanisms, to process and calibrate the collected multi-dimensional health data in real time; Health data analysis module: Constructs a spatiotemporal correlation model based on multidimensional health data; establishes a dynamic baseline model and calculates the real-time health risk index and health monitoring coverage; utilizes drones for environmental monitoring and calculates the comprehensive environmental stress index and the probability of environmentally induced health risks; A health knowledge graph built on OWL ontology; Based on health knowledge graphs, time series prediction models are used to predict future risks; Health data management module: Introduces reinforcement learning mechanism to optimize drone health data collection strategy; Based on user activity patterns, the monitoring range is dynamically adjusted using an adaptive cruise path planning algorithm, including the following steps: Based on historical and real-time UWB positioning data, a dynamically updated heatmap of user spatiotemporal probability distribution is generated. With the goal of covering the user's activity area, and taking into account the drone's endurance constraints, an objective function is defined, and an intelligent optimization algorithm is used to calculate the optimal observation point sequence in real time. Dynamically adjust the path planning based on real-time heat maps; The drone is controlled to fly to the target location according to a preset path or real-time command, collect user health data in different areas, and return to the designated location after the collection is completed. Constructing a spatiotemporal correlation model based on multidimensional health data includes the following steps: Drones acquire centimeter-level location information of users through UWB positioning technology and combine it with timestamps to add spatiotemporal tags to various health data; at the same time, they integrate environmental data and user behavior data to form a multi-dimensional spatiotemporal health dataset. Utilize graph neural networks or spatiotemporal convolutional networks to construct a spatiotemporal correlation model of user health status; By inputting a multi-dimensional spatiotemporal health dataset, a spatiotemporal correlation model is trained, and a user's health status representation vector is output. Establish a dynamic baseline model to update the health baseline in real time and calculate the real-time health risk index and health monitoring coverage, including the following steps: A dynamic baseline model is established based on the historical health status representation vector sequence output by the spatiotemporal correlation model. Using a sliding time window to extract N consecutive state vectors as input, the number of clusters K is optimized by the silhouette coefficient method, and the initial health baseline is determined by K-means clustering. The Mahalanobis distance D between the current health status representation vector and the health baseline is calculated in real time. When D is greater than or equal to 1.5 times the initial standard deviation, a density-based spatiotemporal clustering algorithm is used to cluster the location coordinates of abnormal events to generate a heat map of abnormal areas. The real-time health risk index formula is calculated as: in, Let be the dynamic baseline value of the i-th health parameter at time t. Let be the i-th health parameter value at time t. It is the historical standard deviation of the i-th health parameter. y is the weight of the i-th health parameter, n is the total number of health parameters; y is the area of ​​the abnormal region in the heat map. ; Computing the health monitoring coverage formula: in, It is the total monitoring duration. It is the total area of ​​the user activity heatmap. For real-time health risk index; For the real-time flight speed of the drone, That is the maximum flight speed of the drone. It is the signal-to-noise ratio. It is the maximum signal-to-noise ratio. 2.The unmanned aerial vehicle (UAV) -assisted artificial intelligence (AI) health data collection and analysis system of claim 1, wherein Using drones for environmental monitoring to calculate the comprehensive environmental stress index and the probability of environmentally induced health risks includes the following steps: Based on the environmental data monitored by drones, the spatiotemporal environment tensor is generated by organizing the data according to time series and spatial location. The spatiotemporal environmental tensor and user health data are analyzed using the sparse group Lasso regression model to automatically identify environmental factors that affect health status. Computing a comprehensive environmental stress index in, This represents the measured value of the j-th environmental factor at time t. and These are the historical mean and standard deviation of environmental factor j, respectively. Let be the weight of the j-th environmental factor; m is the number of environmental factors. The environmental pressure index is nonlinearly converted by using a sigmoid function to calculate the probability of environmental-induced health risks, and the formula is: wherein, is the comprehensive environmental stress index, is the real-time health risk index, and are the adjustment coefficients, respectively. 3.The unmanned aerial vehicle (UAV) -assisted artificial intelligence (AI) health data collection and analysis system of claim 1, wherein The health knowledge graph built on the OWL ontology includes the following steps: Based on health data collected by drones, health-related entities, attributes, and relationships are extracted using data processing technology. Classes, attributes, and relationships are defined using the OWL ontology language to construct the ontology model and form the schema layer; The extracted content is instantiated based on the model and stored in a graph database to generate an initial knowledge graph. The streaming data incremental update algorithm is adopted. When the drone collects new health data, it inserts it into the map through SPARQL Update and uses timestamps or credibility rules to handle conflicts. The updated graph deployment graph attention network learns the embedded representations of nodes and edges, and an incremental embedding algorithm is used to optimize the dynamic update.

4. The unmanned aerial vehicle-assisted artificial intelligence health data collection and analysis system according to claim 3, wherein Based on a health knowledge graph, a time series forecasting model is used to predict future risks, including the following steps: Extract user's historical health status representation vector sequence, associated environmental factor time series data, health risk index change records and entity relationships from the health knowledge graph to construct a feature set for time series prediction. Long Short-Term Memory Networks or Temporal Graph Neural Networks are used as time series prediction models. The feature set is input into the model step by step, and the entity embedding and relation embedding of the health knowledge graph serve as prior knowledge constraints for the model. The trained model outputs a predicted sequence of health status representation vectors for the user over the next T time units, and calculates the corresponding future health risk index trend curve. By combining the causal relationship between environmental factors and health risks in the health knowledge graph, we can identify the top three key environmental factors that contribute to future health risks and their corresponding impact weights. Based on the predicted future health risk index and the probability of environmentally induced health risks, the probability of potential abnormal health events is calculated, and the spatial region and time window of the abnormal event are predicted by reasoning the spatiotemporal relationship of the health knowledge graph.

5. The drone-assisted AI-powered smart healthcare data collection and analysis system according to claim 1, characterized in that, The health data management module includes the following steps: introducing the Q-learning algorithm from reinforcement learning to optimize the data collection strategy, using health monitoring coverage as the reward function, and dynamically adjusting the sampling frequency of drones; using a federated learning framework to store distributed data, and combining blockchain technology to support data traceability and access control.

Images