Whole-house health monitoring AI analysis system based on virtual reality

Abstract

This invention discloses a whole-house smart health monitoring and analysis system based on virtual reality, comprising: collecting room boundary data, room connectivity data, and room coordinate data to establish a three-dimensional spatial digital model, generating a graph node sequence and three-dimensional spatial features; mapping physiological data, activity data, and indoor environmental data to the graph node sequence according to human body location data to obtain room health features; calculating the spatial accessibility cost matrix and generating a node transmission cost matrix by combining risk intensity values; using the Sinkhorn algorithm to generate a node transmission weight matrix and a weighted neighbor node set; and performing dual-branch aggregation and risk gating correction through an improved GraphSAGE model to obtain a corrected health risk representation, and outputting room risk level, prompt instructions, and environmental control instructions. This invention enables the three-dimensional spatial digital model to participate in whole-house health risk analysis and the linkage of smart home environment devices.

Description

Technical Field

[0001] This invention relates to the field of spatial computing technology, and in particular to a whole-house smart health monitoring and analysis system based on virtual reality. Background Technology

[0002] With the development of home health monitoring, virtual reality modeling, and smart home control technologies, data on human status, activity behavior, and indoor environment throughout the home are gradually shifting from single-point collection to multi-source collaborative collection. Related technologies largely revolve around establishing data collection links using wearable devices, environmental sensors, indoor positioning devices, and home display terminals. This allows for the acquisition of physiological states, activity levels, location of individuals, and environmental information such as temperature, humidity, and air quality. These data are then displayed through visual interfaces to show room status or user health status, providing a data foundation for home health management.

[0003] Existing virtual reality home monitoring solutions mostly construct a three-dimensional display scene of the house, overlaying various monitoring data on corresponding room locations, allowing users to view the health or environmental status of different spatial areas. Some solutions integrate with smart home devices, issuing alerts or triggering environmental device actions when monitoring data exceeds thresholds. The technical focus of these solutions is concentrated on spatial visualization, data display, and rule triggering. The spatial model mainly serves as an interface and location labeling function, lacking in-depth modeling of the connectivity between rooms, spatial distance relationships, and the correlation changes of health risks in adjacent spaces.

[0004] In real-world home settings, there is a coupling relationship between people's activity locations, physiological states, changes in the indoor environment, and room connectivity. The same health risk may be influenced by the accessibility of adjacent rooms, spatial distribution, and environmental propagation conditions. Relying solely on monitoring data from a single room can easily disconnect the overall spatial structure of the house from the health risk analysis process, preventing virtual reality home scenarios from participating in health risk identification and coordinated decision-making. Therefore, in whole-house intelligent health monitoring scenarios, a technical solution is still needed that can combine virtual reality spatial models, multi-source health data, and spatial accessibility for risk analysis.

[0005] Therefore, how to provide a whole-house smart health monitoring and analysis system based on virtual reality is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a whole-house smart health monitoring and analysis system based on virtual reality. This invention uses a three-dimensional spatial digital model, Sinkhorn edge weight sampling, and improved GraphSAGE aggregation to enable the spatial relationships of the whole house to participate in health risk analysis and device linkage.

[0007] According to an embodiment of the present invention, a whole-house smart health monitoring and analysis system based on virtual reality includes: The spatial model construction module collects room boundary data, room connectivity data, and room coordinate data to establish a three-dimensional spatial digital model, determine graph nodes, and generate graph node sequences and three-dimensional spatial features. The health data mapping module collects physiological data, activity data, human location data, and indoor environmental data, and maps them to a graph node sequence to obtain the room's health characteristics. The access cost calculation module calculates the spatial access cost between graph nodes based on the room connectivity data and room coordinate data in the three-dimensional spatial features, and obtains the spatial access cost matrix. The transmission cost generation module correlates and fuses the spatial access cost in the spatial access cost matrix with the risk intensity value in the room health characteristics to obtain the node transmission cost matrix. The edge weight sampling generation module performs row and column normalization iterative processing on the node transmission cost matrix to generate the node transmission weight matrix, and filters the neighbor nodes in the graph node sequence to obtain a weighted neighbor node set. The dual-branch aggregation module inputs 3D spatial features and room health features into the dual-branch aggregation layer of the improved GraphSAGE model, performs spatial feature aggregation and health feature aggregation on the weighted neighbor node set, and generates a health risk representation. The risk gating correction module calculates the neighbor aggregation weight correction coefficient based on the health risk characterization and corrects the node transmission weight of the weighted neighbor node set to obtain the corrected health risk characterization. The device linkage output module determines the room risk level based on the corrected health risk profile and sends it to the home monitoring terminal. When the room risk level exceeds the preset risk level, it generates a prompt command and an environmental control command, outputs a prompt through the home monitoring terminal, and simultaneously sends the environmental control command to the smart home environment devices.

[0008] Optionally, the spatial model construction module specifically includes: The boundary points in the room boundary data are transformed to a unified three-dimensional coordinate system, and the room ownership is marked according to the room coordinate data to obtain coordinate-based ownership data. Candidate room regions are established based on the coordinate-based attribution data, and coordinate region mapping results are generated according to the candidate room regions where the room coordinate data is located. The connectivity order between candidate room areas is determined based on the room connectivity data, and the connectivity order is mapped to the coordinate area mapping results to obtain the candidate graph node arrangement results; Based on the candidate graph node arrangement results, the boundary connection relationship between candidate room areas is read, and the boundary connection relationship is compared with the room connectivity data to obtain the node consistency test result. When the node consistency test result represents a boundary connection relationship that is inconsistent with the room connectivity data, the inconsistent candidate room area is taken as the backflow room area, and the room boundary data is selected to obtain the backflow boundary data. The candidate room regions are updated based on the reflux boundary data, and a three-dimensional spatial digital model is established. The room regions in the 3D spatial digital model are identified as graph nodes. The graph nodes are sorted according to the room connectivity data to obtain a graph node sequence. The room boundary data, room connectivity data and room coordinate data are extracted according to the graph node sequence to obtain the 3D spatial features.

[0009] Optionally, the health data mapping module specifically includes: The room area in the three-dimensional spatial digital model is extracted according to the graph node sequence, and the human body position data is converted to a unified three-dimensional coordinate system to obtain human body coordinate data. Based on the coordinate attribution relationship between human body coordinate data and room area, the human body coordinate data is mapped to the graph node sequence to obtain the human body mapping result; Physiological and activity data are mapped to graph node sequences according to the human body mapping results to obtain health candidate data; The indoor environmental data is mapped to a sequence of graph nodes according to room areas to obtain the environmental mapping results; The health candidate data and the environment mapping results are matched according to the same graph nodes to obtain the room candidate data; When the human body mapping result in the room candidate data is inconsistent with the environment mapping result, the candidate adjacent nodes are determined based on the environment mapping result and the room connectivity data. The distance between the human body coordinate data and the room boundary data corresponding to the candidate adjacent nodes is compared. The candidate adjacent node with the smallest distance is selected to update the human body mapping result and obtain the human body correction result. When the human body mapping result in the room candidate data is consistent with the environment mapping result, the human body mapping result is determined as the human body correction result. The room candidate data is updated based on the human body correction results, and the risk intensity values ​​corresponding to the physiological data, activity data and indoor environmental data are extracted. The risk intensity values ​​are arranged according to the graph node sequence to obtain the room health characteristics.

[0010] Optionally, the access cost calculation module specifically includes: Based on the direct connectivity relationships between nodes in the room connectivity data marker graph, a set of directly connected node pairs is obtained; Based on the room coordinate data corresponding to two graph nodes in the set of directly connected node pairs, calculate the absolute value of the coordinate difference and sum them to obtain the direct connection cost; Based on the room connectivity data, identify the graph node combinations that do not form direct connectivity relationships in the graph node sequence to obtain the set of indirect node pairs; Based on the room connectivity data, the set of indirect node pairs is expanded layer by layer to obtain a set of candidate paths, and indirect node pairs that do not generate a set of candidate paths are marked as isolation costs. Based on the direct connection cost, the set of directly connected node pairs contained in each extended path in the candidate path set is accumulated, and the number of coordinate direction changes between adjacent sets of directly connected node pairs in each extended path is counted to obtain the path candidate value. When the same indirect node corresponds to multiple path candidate values, sequential filtering is performed to obtain the target path cost; When the same indirect node pair corresponds to only one path candidate value, the path candidate value is determined as the target path cost; The direct connection cost, target path cost, and isolation cost are respectively used as the spatial accessibility cost between the corresponding graph nodes, and sorted according to the graph node sequence to obtain the spatial accessibility cost matrix.

[0011] Optionally, the transmission cost generation module specifically comprises: Extract the starting and ending graph nodes with spatial accessibility costs from the spatial accessibility cost matrix to obtain a set of candidate node pairs; Based on the starting and ending graph nodes in the candidate node pair set, risk intensity values ​​are extracted from the room health features to obtain risk data. Subtracting the risk intensity value of the terminal node from the risk intensity value of the starting node in the risk data yields the risk difference data. When the risk intensity value of the starting graph node represented by the risk difference data is greater than the risk intensity value of the ending graph node, the spatial accessibility cost is subtracted from the risk difference data to obtain the forward return flow data; When the risk intensity value of the starting graph node represented by the risk difference data is less than the risk intensity value of the ending graph node, the spatial accessibility cost is added to the absolute value of the risk difference data to obtain the reverse correction data. When the risk intensity value of the starting graph node is equal to the risk intensity value of the ending graph node, the spatial accessibility cost is determined as the horizontal cost data. The node transmission cost corresponding to the candidate node pair set is determined based on the forward backflow data, reverse correction data, and horizontal cost data, and written into the matrix row and column positions corresponding to the starting graph node and the ending graph node to obtain the node transmission cost matrix.

[0012] Optionally, the edge weight sampling generation module specifically comprises: Extract the matrix row and column positions and node transmission costs corresponding to the starting and ending graph nodes from the node transmission cost matrix to obtain cost row and column data. Based on the cost row and column data, the node transmission cost that is less than the preset minimum positive value is replaced with the preset minimum positive value, and the matrix row and column positions that are the same for the starting graph node and the ending graph node are marked as node mask positions to obtain positive cost data. The Sinkhorn algorithm is used to perform row normalization on the positive cost data to obtain row-normalized data, and the Sinkhorn algorithm is used to perform column normalization on the row-normalized data to obtain column-normalized data. The column-normalized data is fed back to the row-normalized processing, and the row-normalized and column-normalized processing are performed repeatedly until the preset normalization condition is met, thus obtaining the node transmission weight matrix. Based on the node masking bit, delete the weights of the starting graph nodes that point to the same ending graph nodes in the node transmission weight matrix, and filter a preset number of ending graph nodes to obtain candidate neighbor nodes. The candidate neighbor nodes are bound to their corresponding node transmission weights and written into the corresponding starting graph node according to the graph node sequence to obtain the weighted neighbor node set.

[0013] Optionally, the dual-branch aggregation module specifically comprises: Extract the starting graph node, candidate neighbor node, and corresponding node transmission weight from the weighted neighbor node set to obtain neighbor weight data; Based on the neighbor weight data, spatial features corresponding to the starting graph node and candidate neighbor nodes are extracted from the three-dimensional spatial features to obtain spatial neighborhood data. Based on the neighbor weight data, the risk intensity values ​​corresponding to the starting graph node and candidate neighbor node are extracted from the room health features to obtain the health neighborhood data; In the dual-branch aggregation layer of the improved GraphSAGE model, spatial features in the spatial neighborhood data are weighted and aggregated according to the node transmission weights to obtain spatial aggregated data. In the dual-branch aggregation layer of the improved GraphSAGE model, the risk intensity values ​​in the health neighborhood data are weighted and aggregated according to the node transmission weights to obtain health aggregated data. The values ​​in the spatial aggregated data and the health aggregated data are summed separately to obtain the spatial aggregated value and the health aggregated value. The spatial aggregated value is then subtracted from the health aggregated value to obtain the aggregated differential data. Construct an initial zero-value matrix containing spatial correction positions and health correction positions. When the aggregated differential data indicates that the health aggregate value is greater than the spatial aggregate value, write the aggregated differential data into the health correction position and keep the spatial correction position at zero. When the aggregated differential data indicates that the health aggregate value is not greater than the spatial aggregate value, write the absolute value of the aggregated differential data into the spatial correction position and keep the health correction position at zero to obtain the branch correction data. Health risk characterization is generated by splicing spatial aggregated data, health aggregated data, and branch correction data in a fixed order.

[0014] Optionally, the risk gating correction module specifically includes: Extract the representation data corresponding to each starting graph node and candidate neighbor node from the health risk representation, and sum them respectively to obtain the gated sum data; Subtract the gated summation value corresponding to the starting graph node from the gated summation value corresponding to the candidate neighbor node in the gated summation data to obtain the gated differential data; When the gated summation value corresponding to the candidate neighbor node in the gated differential data is greater than the gated summation value corresponding to the node in the initial graph, the gated differential data is incremented by 1 to obtain the enhanced candidate coefficient. When the enhancement candidate coefficient is greater than the preset correction upper limit, the preset correction upper limit is determined as the enhancement correction coefficient; when the enhancement candidate coefficient is not greater than the preset correction upper limit, the enhancement candidate coefficient is determined as the enhancement correction coefficient. When the gated summation value corresponding to the candidate neighbor node in the gated differential data representation is not greater than the gated summation value corresponding to the node in the initial graph, the value 1 is determined as the maintenance correction coefficient. The neighbor aggregation weight correction factor is determined based on the enhancement correction factor and the preservation correction factor, and the node transmission weight in the weighted neighbor node set is multiplied by the neighbor aggregation weight correction factor to obtain the corrected node weight; The modified health risk representation is obtained by weighting and summing the representation data corresponding to the candidate neighbor nodes according to the modified node weights, and then concatenating it with the representation data corresponding to the nodes in the initial graph.

[0015] Optionally, the device linkage output module specifically includes: Extract the representation data corresponding to each graph node from the modified health risk representation, and obtain the room representation data according to the room area in the three-dimensional spatial digital model corresponding to the graph node sequence; The room risk level is obtained by summing the representative values ​​in the room representative data and matching them with the preset risk level. The room risk level is linked to the corresponding room area to obtain room level data; When the room risk level in the room risk level data does not exceed the preset risk level, the room risk level data is sent to the home monitoring terminal to obtain the data displayed on the terminal. When the room risk level in the room level data exceeds the preset risk level, the risk intensity value corresponding to the indoor environmental data of the corresponding room area is extracted from the room health characteristics, and the spatial accessibility cost between the corresponding room area and the adjacent room area is extracted according to the spatial accessibility cost matrix to obtain the environmental accessibility data. When the risk intensity value in the environmental accessibility data exceeds the preset environmental threshold and the space access cost is lower than the preset access threshold, a first linkage instruction containing fresh air control instructions and prompt instructions is generated. When the risk intensity value in the environmental accessibility data exceeds the preset environmental threshold and the spatial access cost is not lower than the preset access threshold, a second linkage instruction containing environmental control instructions and prompt instructions is generated. When the risk intensity value in the environmental accessibility data does not exceed the preset environmental threshold, a third linkage instruction containing a prompt instruction is generated; The prompts from the first, second, and third linkage commands are sent to smart home environment devices and home monitoring terminals to obtain device linkage results.

[0016] Optionally, the improved GraphSAGE model includes a feature input layer, a weighted neighbor access layer, a dual-branch aggregation layer, a branch correction layer, and a representation output layer, specifically: The feature input layer is connected to the weighted neighbor access layer, receiving 3D spatial features, room health features, and a set of weighted neighbor nodes, and aligning the nodes according to the graph node sequence to obtain the model input data; The weighted neighbor access layer is connected to the dual-branch aggregation layer. It extracts the starting graph node, candidate neighbor node and corresponding node transmission weight from the model input data to obtain neighbor weight data. The dual-branch aggregation layer includes a spatial aggregation branch and a health aggregation branch. The spatial aggregation branch performs weighted aggregation of three-dimensional spatial features based on neighbor weight data to obtain spatial aggregation data. The health aggregation branch performs weighted aggregation of risk intensity values ​​in room health features based on neighbor weight data to obtain health aggregation data. The branch correction layer is connected to the dual-branch aggregation layer, which sums the spatial aggregation data and the health aggregation data respectively to obtain the spatial aggregation value and the health aggregation value. The branch correction layer then subtracts the spatial aggregation value from the health aggregation value to obtain the aggregated differential data. The branch correction layer constructs an initial zero-value matrix containing spatial correction positions and health correction positions, and writes corresponding values ​​to the spatial correction positions and health correction positions according to the numerical direction of the aggregation difference data to obtain the branch correction data; The representation output layer is connected to the branch correction layer, and the data is spliced ​​together in a fixed order of spatial aggregated data, health aggregated data and branch correction data. The health risk representation is then output according to the graph node sequence.

[0017] The beneficial effects of this invention are: (1) This invention establishes a three-dimensional spatial digital model by collecting room boundary data, room connectivity data and room coordinate data, and determines the room area as a graph node, generating a graph node sequence and three-dimensional spatial features, thereby realizing the conversion of home space structure into graph node data, so that the whole house space model can participate in the health risk analysis process.

[0018] (2) This invention maps physiological data, activity data and indoor environmental data to a graph node sequence according to human body location data, and combines spatial accessibility cost matrix, risk intensity value, Sinkhorn algorithm to generate node transmission weight matrix and weighted neighbor node set, thereby realizing the correlation processing of health data, spatial accessibility relationship and neighbor node weight, thus avoiding the disconnect between single room monitoring data and the whole room spatial relationship.

[0019] (3) The improved GraphSAGE model in this invention processes the three-dimensional spatial features and room health features through the weighted neighbor access layer, the dual branch aggregation layer and the branch correction layer in the whole-house home health risk characterization generation process, and obtains the corrected health risk characterization by combining risk gating correction, thereby providing a data basis for room risk level judgment, prompt instructions and environmental control instructions generation. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the system structure of the whole-house smart health monitoring and analysis system based on virtual reality proposed in this invention; Figure 2 This is a schematic diagram of the improved GraphSAGE model structure of the whole-house smart health monitoring AI analysis system based on virtual reality proposed in this invention. Detailed Implementation

[0021] Example 1: Reference Figures 1-2 A whole-house smart health monitoring and analysis system based on virtual reality, including: The spatial model construction module collects room boundary data, room connectivity data, and room coordinate data. It establishes a three-dimensional spatial digital model based on a unified three-dimensional coordinate system and identifies the room regions in the three-dimensional spatial digital model as graph nodes, generating graph node sequences and three-dimensional spatial features. The health data mapping module collects physiological data, activity data, human body location data, and indoor environment data, and maps the physiological data, activity data, and indoor environment data to a graph node sequence based on the human body location data to obtain the room's health characteristics. The access cost calculation module calculates the spatial access cost between graph nodes based on the room connectivity data and room coordinate data in the three-dimensional spatial features, and obtains the spatial access cost matrix. The transmission cost generation module correlates and fuses the spatial access cost in the spatial access cost matrix with the risk intensity value in the room health characteristics to obtain the node transmission cost matrix. The edge weight sampling generation module uses the Sinkhorn algorithm to perform row and column normalization iterative processing on the node transmission cost matrix to generate the node transmission weight matrix, and then filters the neighbor nodes in the graph node sequence according to the node transmission weight matrix to obtain the weighted neighbor node set. The dual-branch aggregation module inputs 3D spatial features and room health features into the dual-branch aggregation layer of the improved GraphSAGE model, performs spatial feature aggregation and health feature aggregation on the weighted neighbor node set, and generates a health risk representation. The risk gating correction module calculates the neighbor aggregation weight correction coefficient based on the health risk characterization and corrects the node transmission weight of the weighted neighbor node set to obtain the corrected health risk characterization. The device linkage output module determines the room risk level based on the corrected health risk profile and sends it to the home monitoring terminal. When the room risk level exceeds the preset risk level, it generates a prompt command and an environmental control command, outputs a prompt through the home monitoring terminal, and simultaneously sends the environmental control command to the smart home environment devices.

[0022] In this embodiment, the spatial model construction module specifically comprises: The boundary points in the room boundary data are transformed to a unified three-dimensional coordinate system, and the room ownership is marked according to the room coordinate data to obtain coordinate-based ownership data. Candidate room regions are established based on the coordinate-based attribution data, and coordinate region mapping results are generated according to the candidate room regions where the room coordinate data is located. The connectivity order between candidate room areas is determined based on the room connectivity data, and the connectivity order is mapped to the coordinate area mapping results to obtain the candidate graph node arrangement results; Based on the candidate graph node arrangement results, the boundary connection relationship between candidate room areas is read, and the boundary connection relationship is compared with the room connectivity data to obtain the node consistency test result. When the node consistency test result represents a boundary connection relationship that is inconsistent with the room connectivity data, the inconsistent candidate room area is taken as the return room area, and the room boundary data is reselected based on the room coordinate data corresponding to the return room area to obtain the return boundary data. The candidate room regions are updated based on the reflux boundary data, and a three-dimensional spatial digital model is established. In Example 1, the node consistency test result is used to distinguish the processing paths of candidate room regions entering the 3D spatial digital model. When the node consistency test result indicates that the boundary connection relationship is consistent with the room connectivity data, the candidate room region is directly determined as the room region data to be modeled. When the node consistency test result indicates that the boundary connection relationship is inconsistent with the room connectivity data, the inconsistent candidate room region is used as the return room region, and the room boundary data is reselected based on the room coordinate data corresponding to the return room region to obtain the return boundary data. The candidate room region is updated according to the return boundary data, and the updated candidate room region is determined as the room region data to be modeled. A 3D spatial digital model is established based on the room region data to be modeled. When reselecting room boundary data, a coordinate retrieval center is established for the room coordinate data corresponding to the return room area, and the set of boundary points containing the coordinate retrieval center is retrieved from the collected room boundary data to obtain the first candidate boundary data. The first candidate boundary data is transformed to a unified three-dimensional coordinate system, and room affiliation is marked according to the room coordinate data corresponding to the return room area to obtain the second candidate boundary data. According to the room connectivity data, the candidate room area formed by the second candidate boundary data is checked for boundary adjacency with adjacent candidate room areas, and the second candidate boundary data that does not form a boundary adjacency relationship is deleted to obtain the third candidate boundary data. When there are multiple data items in the third candidate boundary data, the number of consistent rooms between the candidate room area corresponding to each data item and the room connectivity data is counted, and the third candidate boundary data with the largest number of consistent rooms is selected as the return boundary data. When there is only one data item in the third candidate boundary data, the third candidate boundary data is used as the return boundary data. The room regions in the 3D spatial digital model are identified as graph nodes. The graph nodes are sorted according to the room connectivity data to obtain a graph node sequence. The room boundary data, room connectivity data and room coordinate data are extracted according to the graph node sequence to obtain the 3D spatial features.

[0023] In Example 1, the room regions in the 3D spatial digital model originate from candidate room regions. When the node consistency test result indicates that the boundary connection relationship is consistent with the room connectivity data, the candidate room region is not updated backflow, and the candidate room region enters the 3D spatial digital model establishment process and is used as the room region in the 3D spatial digital model. When the node consistency test result indicates that the boundary connection relationship is inconsistent with the room connectivity data, the backflow boundary data replaces the room boundary data corresponding to the inconsistent candidate room region, and the replaced candidate room region enters the 3D spatial digital model establishment process and is used as the room region in the 3D spatial digital model. After the 3D spatial digital model is established, the room regions in the 3D spatial digital model and the candidate room regions entering the 3D spatial digital model establishment process maintain consistency in room ownership labels, room coordinate data, and room connectivity data. Each room region in the 3D spatial digital model is converted into one graph node according to the room ownership label, and the graph nodes are sorted according to the room connectivity data to obtain a graph node sequence. The room boundary data, room connectivity data, and room coordinate data of the corresponding graph nodes are extracted according to the graph node sequence to obtain the 3D spatial features.

[0024] In this embodiment, the health data mapping module specifically comprises: The room area in the three-dimensional spatial digital model is extracted according to the graph node sequence, and the human body position data is converted to a unified three-dimensional coordinate system to obtain human body coordinate data. Based on the coordinate attribution relationship between human body coordinate data and room area, the human body coordinate data is mapped to the graph node sequence to obtain the human body mapping result; Physiological and activity data are mapped to graph node sequences according to the human body mapping results to obtain health candidate data; The indoor environmental data is mapped to a sequence of graph nodes according to room areas to obtain the environmental mapping results; The health candidate data and the environment mapping results are matched according to the same graph nodes to obtain the room candidate data; When the human body mapping result in the room candidate data is inconsistent with the environment mapping result, the candidate adjacent nodes are determined based on the environment mapping result and the room connectivity data. The distance between the human body coordinate data and the room boundary data corresponding to the candidate adjacent nodes is compared. The candidate adjacent node with the smallest distance is selected to update the human body mapping result and obtain the human body correction result. In Example 1, the room boundary data corresponding to the candidate adjacent nodes are converted into a set of boundary line segments according to the boundary point connection relationship. For each candidate adjacent node, the human coordinate points in the human coordinate data are projected onto each boundary line segment in the boundary line segment set. When the projected point is located between the two endpoints of the boundary line segment, the absolute value of the three-dimensional coordinate difference between the human coordinate point and the projected point is calculated and summed to obtain the line segment distance value. When the projected point exceeds the range between the two endpoints of the boundary line segment, the absolute value of the three-dimensional coordinate difference between the human coordinate point and the two endpoints is calculated and summed, and the smaller value is selected as the line segment distance value. The minimum value of the line segment distance values ​​corresponding to the same candidate adjacent node is selected to obtain the node boundary distance. The node boundary distances corresponding to each candidate adjacent node are compared, and the candidate adjacent node with the smallest node boundary distance is selected to update the human mapping result. When the human body mapping result in the room candidate data is consistent with the environment mapping result, the human body mapping result is determined as the human body correction result. The room candidate data is updated based on the human body correction results, and the risk intensity values ​​corresponding to the physiological data, activity data and indoor environmental data are extracted. The risk intensity values ​​are arranged according to the graph node sequence to obtain the room health characteristics.

[0025] In Example 1, the risk intensity value is obtained by converting physiological data, activity data, and indoor environmental data from the room candidate data. For each piece of physiological data, activity data, and indoor environmental data, a safety lower limit, a safety upper limit, a risk lower limit, and a risk upper limit are set. When the data value is between the safety lower limit and the safety upper limit, the risk intensity value is set to 0. When the data value is lower than the safety lower limit, the difference between the safety lower limit and the data value is calculated, and then the difference is divided by the difference between the safety lower limit and the risk lower limit to obtain the risk intensity value. When the data value is higher than the safety upper limit, the difference between the data value and the safety upper limit is calculated, and then the difference is divided by the difference between the risk upper limit and the safety upper limit to obtain the risk intensity value. The risk intensity values ​​corresponding to the physiological data, activity data, and indoor environmental data are arranged according to the graph node sequence to obtain the room health characteristics.

[0026] In this embodiment, the access cost calculation module specifically comprises: Extract the room coordinate data corresponding to each graph node according to the graph node sequence, and mark the direct connectivity between graph nodes according to the room connectivity data to obtain a set of directly connected node pairs; Based on the room coordinate data corresponding to two graph nodes in the set of directly connected node pairs, calculate the absolute value of the coordinate difference and sum them to obtain the direct connection cost; Based on the room connectivity data, identify the graph node combinations that do not form direct connectivity relationships in the graph node sequence to obtain the set of indirect node pairs; Based on the room connectivity data, the indirect node pair set is expanded layer by layer, and the same graph node is prohibited from appearing repeatedly in a single expansion path to obtain a candidate path set. Indirect node pairs that do not generate a candidate path set are marked as isolation costs. Based on the direct connection cost, the set of directly connected node pairs contained in each extended path in the candidate path set is accumulated, and the number of coordinate direction changes between adjacent sets of directly connected node pairs in each extended path is counted to obtain the path candidate value. When the same indirect node corresponds to multiple path candidate values, the target path cost is obtained by filtering according to the execution order of the cumulative direct connection cost from small to large, the number of coordinate direction changes from few to many, and the number of graph nodes from few to many. In Example 1, when counting the number of coordinate direction changes, each extended path is divided into sets of adjacent directly connected node pairs according to the order in which the graph nodes pass. For each set of directly connected node pairs, the room coordinate data corresponding to the two graph nodes are extracted, and the absolute values ​​of the differences on the X-axis, Y-axis, and Z-axis are calculated respectively. The coordinate axis with the largest value is selected as the dominant coordinate direction of the set of directly connected node pairs. When two coordinate axes have the same value, the coordinate axis with the first value in the fixed order of X-axis, Y-axis, and Z-axis is selected as the dominant coordinate direction. The dominant coordinate directions of adjacent sets of directly connected node pairs are compared sequentially along the extended path. When the current dominant coordinate direction is inconsistent with the previous dominant coordinate direction, the number of coordinate direction changes is increased by 1. After comparing all sets of adjacent directly connected node pairs in the extended path, the number of coordinate direction changes corresponding to the extended path is obtained. When the same indirect node pair corresponds to only one path candidate value, the path candidate value is determined as the target path cost; The direct connection cost, target path cost, and isolation cost are respectively used as the spatial accessibility cost between the corresponding graph nodes, and sorted according to the graph node sequence to obtain the spatial accessibility cost matrix.

[0027] In this embodiment, the transmission cost generation module specifically comprises: Extract the starting and ending graph nodes with spatial accessibility costs from the spatial accessibility cost matrix to obtain a set of candidate node pairs; Based on the starting and ending graph nodes in the candidate node pair set, risk intensity values ​​are extracted from the room health features to obtain risk data. Subtracting the risk intensity value of the terminal node from the risk intensity value of the starting node in the risk data yields the risk difference data. When the risk intensity value of the starting graph node represented by the risk difference data is greater than the risk intensity value of the ending graph node, the spatial accessibility cost is subtracted from the risk difference data to obtain the forward return flow data; In Example 1, when calculating forward backflow data, the spatial accessibility cost is subtracted from the risk difference data to obtain forward candidate data; when the forward candidate data is not greater than a preset minimum positive value, the forward candidate data is replaced with the preset minimum positive value, and the preset minimum positive value is used as the forward backflow data; when the forward candidate data is greater than the preset minimum positive value, the forward candidate data is used as the forward backflow data; the preset minimum positive value is set to 0.01; the node transmission cost is determined based on the forward backflow data, reverse correction data, and horizontal cost data, and the node transmission cost is written into the matrix row and column positions corresponding to the starting graph node and the ending graph node to obtain the node transmission cost matrix; When the risk intensity value of the starting graph node represented by the risk difference data is less than the risk intensity value of the ending graph node, the spatial accessibility cost is added to the absolute value of the risk difference data to obtain the reverse correction data. When the risk intensity value of the starting graph node is equal to the risk intensity value of the ending graph node, the spatial accessibility cost is determined as the horizontal cost data. The node transmission cost corresponding to the candidate node pair set is determined based on the forward backflow data, reverse correction data, and horizontal cost data, and written into the matrix row and column positions corresponding to the starting graph node and the ending graph node to obtain the node transmission cost matrix.

[0028] In Example 1, for each candidate node pair in the candidate node pair set, the risk difference data is used as a branch identifier; when the risk difference data indicates that the risk intensity value of the starting graph node is greater than the risk intensity value of the ending graph node, the forward backflow data is determined as the node transmission cost; when the risk difference data indicates that the risk intensity value of the starting graph node is less than the risk intensity value of the ending graph node, the reverse correction data is determined as the node transmission cost; when the risk difference data indicates that the risk intensity value of the starting graph node is equal to the risk intensity value of the ending graph node, the horizontal cost data is determined as the node transmission cost; the node transmission cost corresponding to each candidate node pair is written into the matrix row and column positions corresponding to the starting graph node and the ending graph node to obtain the node transmission cost matrix.

[0029] In this embodiment, the edge weight sampling generation module specifically comprises: Extract the matrix row and column positions and node transmission costs corresponding to the starting and ending graph nodes from the node transmission cost matrix to obtain cost row and column data. Based on the cost row and column data, the node transmission cost that is less than the preset minimum positive value is replaced with the preset minimum positive value, and the matrix row and column positions that are the same for the starting graph node and the ending graph node are marked as node mask positions to obtain positive cost data. The Sinkhorn algorithm is used to perform row normalization on the positive cost data to obtain row-normalized data, and the Sinkhorn algorithm is used to perform column normalization on the row-normalized data to obtain column-normalized data. In Example 1, when the Sinkhorn algorithm performs row normalization, for each row of positive cost data, it retains the masking status of the node mask bits and summarizes the transmission costs of unmasked nodes in the same row to obtain the row data sum; it divides the transmission cost of each unmasked node in the same row by the row data sum to obtain the row weight value; it writes the row weight value back to the corresponding matrix row and column position to obtain the row-normalized data. When the Sinkhorn algorithm performs column normalization, for each column of the row-normalized data, it retains the masking status of the node mask bits and summarizes the unmasked row weight values ​​in the same column to obtain the column data sum; it divides the unmasked row weight value in the same column by the column data sum to obtain the column weight value; it writes the column weight value back to the corresponding matrix row and column position to obtain the column-normalized data. The column-normalized data is fed back to the row-normalized processing, and the row-normalized processing and column-normalized processing are executed in a loop until the sum of the weights of each row and the sum of the weights of each column in the column-normalized data reach the preset normalization condition, thus obtaining the node transmission weight matrix. In Example 1, the preset normalization condition is set as follows: the absolute value of the difference between the sum of weights in each row (excluding node masking bits) and 1 in the column-normalized data is not greater than 0.001, and the absolute value of the difference between the sum of weights in each column (excluding node masking bits) and 1 is not greater than 0.001. When the sum of weights in each row and the sum of weights in each column both satisfy the preset normalization condition, the loop of row normalization and column normalization is stopped, and the current column-normalized data is used as the node transmission weight matrix. When there are rows where the sum of weights does not satisfy the preset normalization condition, or columns where the sum of weights does not satisfy the preset normalization condition, the current column-normalized data is returned to the row normalization process. When the loop reaches 100 times, the loop is stopped, and the column-normalized data obtained in the 100th iteration is used as the node transmission weight matrix. Based on the node masking bit, delete the weights of the starting graph nodes that point to the same ending graph nodes in the node transmission weight matrix, and filter a preset number of ending graph nodes in each row according to the node transmission weight from largest to smallest to obtain candidate neighbor nodes. In Example 1, the preset quantity is set to 3. After deleting the weights of the starting graph nodes pointing to the same ending graph nodes in the node transmission weight matrix according to the node masking bit, the undeleted ending graph nodes and their corresponding node transmission weights are retained for each row. When the number of undeleted ending graph nodes in each row is greater than 3, the top 3 ending graph nodes are selected in descending order of node transmission weights to obtain candidate neighbor nodes. When the number of undeleted ending graph nodes in each row is not greater than 3, all undeleted ending graph nodes are used as candidate neighbor nodes. After the candidate neighbor nodes are bound with their corresponding node transmission weights, they are written into the corresponding starting graph nodes to obtain a weighted neighbor node set. The candidate neighbor nodes are bound to their corresponding node transmission weights and written into the corresponding starting graph node according to the graph node sequence to obtain the weighted neighbor node set.

[0030] In this embodiment, the dual-branch aggregation module specifically refers to: Extract the starting graph node, candidate neighbor node, and corresponding node transmission weight from the weighted neighbor node set to obtain neighbor weight data; Based on the neighbor weight data, spatial features corresponding to the starting graph node and candidate neighbor nodes are extracted from the three-dimensional spatial features to obtain spatial neighborhood data. Based on the neighbor weight data, the risk intensity values ​​corresponding to the starting graph node and candidate neighbor node are extracted from the room health features to obtain the health neighborhood data; In the dual-branch aggregation layer of the improved GraphSAGE model, spatial features in the spatial neighborhood data are weighted and aggregated according to the node transmission weights to obtain spatial aggregated data. In Example 1, when generating spatial aggregation data, for each starting graph node in the neighbor weight data, the spatial features and node transmission weights corresponding to candidate neighbor nodes are obtained in the spatial neighborhood data; the spatial features corresponding to the candidate neighbor nodes are decomposed into room boundary data components, room connectivity data components, and room coordinate data components; the room boundary data components, room connectivity data components, and room coordinate data components of each candidate neighbor node are multiplied by the corresponding node transmission weights to obtain neighbor weighted features; for the same starting graph node, the values ​​of the same component positions in all neighbor weighted features are accumulated to obtain the spatial neighborhood aggregation value; the spatial neighborhood aggregation value is written into the spatial aggregation position corresponding to the starting graph node to obtain spatial aggregation data. In the dual-branch aggregation layer of the improved GraphSAGE model, the risk intensity values ​​in the health neighborhood data are weighted and aggregated according to the node transmission weights to obtain health aggregated data. In Example 1, when generating health aggregated data, for each starting graph node in the neighbor weight data, the risk intensity value and node transmission weight corresponding to the candidate neighbor node are obtained in the health neighborhood data; the risk intensity value corresponding to each candidate neighbor node is multiplied by the corresponding node transmission weight to obtain the neighbor weighted risk; for the same starting graph node, the weighted risks of all neighbors are accumulated to obtain the health neighborhood aggregated value; the risk intensity value of the starting graph node in the room health features is concatenated with the health neighborhood aggregated value in a fixed order to obtain the node health aggregated value; the node health aggregated value is written into the health aggregated position corresponding to the starting graph node to obtain the health aggregated data. The values ​​in the spatial aggregated data and the health aggregated data are summed separately to obtain the spatial aggregated value and the health aggregated value. The spatial aggregated value is then subtracted from the health aggregated value to obtain the aggregated differential data. Construct an initial zero-value matrix containing spatial correction positions and health correction positions. When the aggregated differential data indicates that the health aggregate value is greater than the spatial aggregate value, write the aggregated differential data into the health correction position and keep the spatial correction position at zero. When the aggregated differential data indicates that the health aggregate value is not greater than the spatial aggregate value, write the absolute value of the aggregated differential data into the spatial correction position and keep the health correction position at zero to obtain the branch correction data. In Example 1, when constructing the initial zero-value matrix, the number of matrix rows is determined based on the number of starting graph nodes corresponding to the spatial aggregation data and the health aggregation data. Two column positions, spatial correction position and health correction position, are set in each row. The spatial correction position and health correction position in each row are set to 0, resulting in the initial zero-value matrix. The aggregation differential data is written into the corresponding rows of the initial zero-value matrix according to the starting graph nodes. When the aggregation differential data indicates that the health aggregation value is greater than the spatial aggregation value, the aggregation differential data is written into the health correction position, while keeping the spatial correction position in the same row at 0. When the aggregation differential data indicates that the health aggregation value is not greater than the spatial aggregation value, the absolute value of the aggregation differential data is written into the spatial correction position, while keeping the health correction position in the same row at 0. After writing all the rows corresponding to the starting graph nodes, the branch correction data is obtained. Health risk characterization is generated by splicing spatial aggregated data, health aggregated data, and branch correction data in a fixed order.

[0031] In this embodiment, the risk gating correction module specifically comprises: Extract the representation data corresponding to each starting graph node and candidate neighbor node from the health risk representation, and sum them respectively to obtain the gated sum data; Subtract the gated summation value corresponding to the starting graph node from the gated summation value corresponding to the candidate neighbor node in the gated summation data to obtain the gated differential data; When the gated summation value corresponding to the candidate neighbor node in the gated differential data is greater than the gated summation value corresponding to the node in the initial graph, the gated differential data is incremented by 1 to obtain the enhanced candidate coefficient. When the enhancement candidate coefficient is greater than the preset correction upper limit, the preset correction upper limit is determined as the enhancement correction coefficient; when the enhancement candidate coefficient is not greater than the preset correction upper limit, the enhancement candidate coefficient is determined as the enhancement correction coefficient. When the gated summation value corresponding to the candidate neighbor node in the gated differential data representation is not greater than the gated summation value corresponding to the node in the initial graph, the value 1 is determined as the maintenance correction coefficient. In Example 1, the preset correction upper limit is set to 3. After the risk gating correction module obtains the enhancement candidate coefficient, it compares the enhancement candidate coefficient with the preset correction upper limit. When the enhancement candidate coefficient is greater than 3, the value 3 is determined as the enhancement correction coefficient. When the enhancement candidate coefficient is not greater than 3, the enhancement candidate coefficient is determined as the enhancement correction coefficient. When the gated summation value corresponding to the candidate neighbor node in the gated differential data representation is not greater than the gated summation value corresponding to the starting graph node, the value 1 is determined as the maintenance correction coefficient. The neighbor aggregation weight correction factor is determined based on the enhancement correction factor and the preservation correction factor, and the node transmission weight in the weighted neighbor node set is multiplied by the neighbor aggregation weight correction factor to obtain the corrected node weight; In Example 1, for each starting graph node and candidate neighbor node, the risk gating correction module selects a corresponding coefficient based on the gating differential data. When the gating differential data indicates that the gating summation value corresponding to the candidate neighbor node is greater than the gating summation value corresponding to the starting graph node, the enhancement correction coefficient is determined as the neighbor aggregation weight correction coefficient. When the gating differential data indicates that the gating summation value corresponding to the candidate neighbor node is not greater than the gating summation value corresponding to the starting graph node, the maintenance correction coefficient is determined as the neighbor aggregation weight correction coefficient. The neighbor aggregation weight correction coefficient is multiplied by the node transmission weight in the weighted neighbor node set to obtain the corrected node weight. The modified health risk representation is obtained by weighting and summing the representation data corresponding to the candidate neighbor nodes according to the modified node weights, and then concatenating it with the representation data corresponding to the nodes in the initial graph.

[0032] In this embodiment, the device linkage output module specifically comprises: Extract the representation data corresponding to each graph node from the modified health risk representation, and obtain the room representation data according to the room area in the three-dimensional spatial digital model corresponding to the graph node sequence; The room risk level is obtained by summing the representative values ​​in the room representative data and matching them with the preset risk level. The room risk level is linked to the corresponding room area to obtain room level data; In Example 1, the preset risk level is set to Level 3, and the room risk levels are divided into Level 1, Level 2, Level 3, and Level 4. For the room representation data corresponding to each graph node, all representation values ​​in the room representation data are added together to obtain the room representation sum value. When the room representation sum value is not greater than 1, it matches Level 1; when the room representation sum value is greater than 1 but not greater than 2, it matches Level 2; when the room representation sum value is greater than 2 but not greater than 3, it matches Level 3; when the room representation sum value is greater than 3, it matches Level 4, thus obtaining the room risk level. According to the graph node sequence, the room risk level and the corresponding room area in the three-dimensional spatial digital model are written into the same data item to obtain the room level data. When the room risk level in the room risk level data does not exceed the preset risk level, the room risk level data is sent to the home monitoring terminal to obtain the data displayed on the terminal. When the room risk level in the room level data exceeds the preset risk level, the risk intensity value corresponding to the indoor environmental data of the corresponding room area is extracted from the room health characteristics, and the spatial accessibility cost between the corresponding room area and the adjacent room area is extracted according to the spatial accessibility cost matrix to obtain the environmental accessibility data. When the risk intensity value in the environmental accessibility data exceeds the preset environmental threshold and the space access cost is lower than the preset access threshold, a first linkage instruction containing fresh air control instructions and prompt instructions is generated. In Example 1, the preset environmental threshold is set to 0.6, and the preset accessibility threshold is set to 2. When the risk intensity value in the environmental accessibility data is greater than 0.6, it is determined that the risk intensity value exceeds the preset environmental threshold. When the risk intensity value in the environmental accessibility data is not greater than 0.6, it is determined that the risk intensity value does not exceed the preset environmental threshold. When the spatial accessibility cost in the environmental accessibility data is less than 2, it is determined that the spatial accessibility cost is lower than the preset accessibility threshold. When the spatial accessibility cost in the environmental accessibility data is not less than 2, it is determined that the spatial accessibility cost is not lower than the preset accessibility threshold. The determination result enters the generation process of the first linkage instruction, the second linkage instruction, and the third linkage instruction. When the risk intensity value in the environmental accessibility data exceeds the preset environmental threshold and the spatial access cost is not lower than the preset access threshold, a second linkage instruction containing environmental control instructions and prompt instructions is generated. When the risk intensity value in the environmental accessibility data does not exceed the preset environmental threshold, a third linkage instruction containing a prompt instruction is generated; The prompts from the first, second, and third linkage commands are sent to smart home environment devices and home monitoring terminals to obtain device linkage results.

[0033] In this embodiment, the improved GraphSAGE model includes a feature input layer, a weighted neighbor access layer, a dual-branch aggregation layer, a branch correction layer, and a representation output layer, specifically: The feature input layer is connected to the weighted neighbor access layer, receiving 3D spatial features, room health features, and a set of weighted neighbor nodes, and aligning the nodes according to the graph node sequence to obtain the model input data; The weighted neighbor access layer is connected to the dual-branch aggregation layer. It extracts the starting graph node, candidate neighbor node and corresponding node transmission weight from the model input data to obtain neighbor weight data. The dual-branch aggregation layer includes a spatial aggregation branch and a health aggregation branch. The spatial aggregation branch performs weighted aggregation of three-dimensional spatial features based on neighbor weight data to obtain spatial aggregation data. The health aggregation branch performs weighted aggregation of risk intensity values ​​in room health features based on neighbor weight data to obtain health aggregation data. The branch correction layer is connected to the dual-branch aggregation layer, which sums the spatial aggregation data and the health aggregation data respectively to obtain the spatial aggregation value and the health aggregation value. The branch correction layer then subtracts the spatial aggregation value from the health aggregation value to obtain the aggregated differential data. The branch correction layer constructs an initial zero-value matrix containing spatial correction positions and health correction positions, and writes corresponding values ​​to the spatial correction positions and health correction positions according to the numerical direction of the aggregation difference data to obtain the branch correction data; The representation output layer is connected to the branch correction layer, and the data is spliced ​​together in a fixed order of spatial aggregated data, health aggregated data and branch correction data. The health risk representation is then output according to the graph node sequence.

[0034] In Example 1, the improved GraphSAGE model includes a feature input layer, a weighted neighbor access layer, a dual-branch aggregation layer, a branch correction layer, and a representation output layer. The network backbone has a 5-layer structure, and the dual-branch aggregation layer includes spatial aggregation branches and health aggregation branches; the number of graph nodes is assumed to be... The maximum number of candidate neighbor nodes corresponding to each starting graph node in the weighted neighbor set is . , If there are fewer than 3 candidate neighbor nodes, the transmission weight of the node at the vacant position is reset to 0; 3D spatial feature dimension. Take 7, where the room boundary data component is 3-dimensional, corresponding to the room boundary data in a unified 3D coordinate system. axis, shaft and The boundary span along the axis; the room connectivity data component is 1-dimensional, corresponding to the number of directly connected nodes in the graph; the room coordinate data component is 3-dimensional, corresponding to the room coordinate data in a unified 3D coordinate system. axis coordinate values axis coordinate values ​​and Axis coordinate values; Room health characteristic dimensions The value is set to 3, corresponding to the risk intensity values ​​obtained from the conversion of physiological data, activity data, and indoor environmental data, respectively; the feature input layer receives data in dimensionality 3. The three-dimensional spatial features multiplied by 7, with dimensions of 7 Room health characteristics and dimensions multiplied by 3 Candidate neighbor node index data multiplied by 3 and dimension is The node transmission weight data is multiplied by 3; the weighted neighbor access layer outputs the candidate neighbor nodes and their transmission weights for each starting graph node based on the candidate neighbor node index data and node transmission weight data; the spatial aggregation branch performs weighted aggregation of the three-dimensional spatial features corresponding to the candidate neighbor nodes based on the node transmission weights, and uses a 7x7 spatial mapping weight matrix to linearly map the weighted aggregation results to obtain a dimension of... Spatial aggregation data multiplied by 7; the health aggregation branch performs weighted aggregation of room health features corresponding to candidate neighbor nodes based on node transmission weights, and uses a 3x3 health mapping weight matrix to linearly map the weighted aggregation results, resulting in a dimension of... The health aggregate data is multiplied by 3; the branch correction layer sums the spatial aggregate data and the health aggregate data respectively to obtain the spatial aggregate value and the health aggregate value. The spatial aggregate value is subtracted from the health aggregate value to obtain the aggregated differential data, and a dimension is generated according to the numerical direction of the aggregated differential data. The branch correction data is multiplied by 2; the representation output layer concatenates the spatial aggregation data, health aggregation data, and branch correction data in a fixed order, and the output dimension is... health risk characterization multiplied by 12; the trainable parameters of the model consist of a 7-row by 7-column spatial mapping weight matrix and a 3-row by 3-column health mapping weight matrix, neither the spatial mapping weight matrix nor the health mapping weight matrix is provided with a bias term, and the parameter scale is 58 weight parameters; the weighted neighbor access layer, branch correction layer and characterization output layer are not provided with additional trainable parameters; the training data is derived from historically collected room boundary data, room connectivity data, room coordinate data, physiological data, activity data, human body position data and indoor environment data, as well as the room risk levels in historical device linkage output records; during training, the historically collected data is processed by the spatial model construction module, the health data mapping module, the accessibility cost calculation module, the transmission cost generation module and the edge weight sampling generation module to form training samples, the improved GraphSAGE model outputs a health risk characterization, the risk gating correction module generates a corrected health risk characterization according to the health risk characterization, and the device linkage output module sums the room characterization data corresponding to the corrected health risk characterization to obtain a training output value; the first-level, second-level, third-level and fourth-level room risk levels in the historical device linkage output records are converted into supervised values of 0.5, 1.5, 2.5 and 3.5 respectively, and the spatial mapping weight matrix and the health mapping weight matrix are trained by using the mean square error between the training output value and the supervised value, with the training batch size set to 32, the learning rate set to 0.001, and the number of iteration rounds set to 100 rounds; compared with the existing GraphSAGE model, the improvement of the improved GraphSAGE model lies in that the weighted neighbor node set determined by the node transmission weight is accessed through the weighted neighbor access layer, the three-dimensional spatial features and the room health features are respectively processed through the dual-branch aggregation layer, and the difference relationship between the spatial aggregation value and the health aggregation value is written into the health risk characterization through the branch correction layer, so that the three-dimensional spatial digital model, the spatial accessibility cost and the room health features jointly participate in the generation of graph node characterization.

[0035] Embodiment 2:

[0036] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a whole-house home health monitoring scenario, where the application object is a continuous living space composed of a living room, a bedroom, a kitchen, a bathroom and a study, and the implementation scenario is to perform linkage analysis on the physiological data, activity data, human body position data and indoor environment data of occupants in a three-dimensional spatial digital model, and output prompt instructions and environment regulation instructions according to the room risk level. In this scenario, most of the existing methods perform threshold judgment based on the sensing data of a single room, and the connectivity relationship between rooms, spatial accessibility cost and risk correlation of adjacent rooms are not involved in the risk analysis process, resulting in that the virtual reality home scenario mainly undertakes the display function and is difficult to support whole-house health risk analysis and device linkage output.

[0037] During implementation, room boundary data, room connectivity data, and room coordinate data are collected. Boundary points in the room boundary data are converted to a unified 3D coordinate system, and room assignments are marked according to the room coordinate data to obtain coordinate-based assignment data. Candidate room regions are established based on the coordinate-based assignment data, and coordinate region mapping results are generated. The connectivity order between candidate room regions is determined based on the room connectivity data and correlated with the coordinate region mapping results to obtain candidate graph node arrangement results. The boundary connection relationships between candidate room regions are read based on the candidate graph node arrangement results and compared with the room connectivity data to obtain node consistency verification results. When the node consistency verification results indicate that the boundary connection relationships are inconsistent with the room connectivity data, the inconsistent candidate room regions are designated as backflow room regions. Room boundary data is reselected based on the room coordinate data corresponding to the backflow room regions to obtain backflow boundary data. The candidate room regions are updated based on the backflow boundary data, and a 3D spatial digital model is established. The room regions in the 3D spatial digital model are identified as graph nodes. The graph nodes are sorted according to the room connectivity data to obtain a graph node sequence, and room boundary data, room connectivity data, and room coordinate data are extracted to obtain 3D spatial features.

[0038] In the health data mapping process, human position data is transformed to a unified three-dimensional coordinate system to obtain human coordinate data. The human mapping result is then obtained based on the coordinate attribution relationship between the human coordinate data and room areas. Physiological and activity data are mapped to a graph node sequence according to the human mapping result to obtain health candidate data. Indoor environmental data is mapped to a graph node sequence according to room areas to obtain environmental mapping results. Health candidate data and environmental mapping results are matched using the same graph nodes to obtain room candidate data. When the human mapping result in the room candidate data is inconsistent with the environmental mapping result, candidate adjacent nodes are determined based on the environmental mapping result and room connectivity data. The distance between the human coordinate data and the room boundary data corresponding to each candidate adjacent node is compared, and the candidate adjacent node with the smallest distance is selected to update the human mapping result, obtaining the human correction result. When the human mapping result in the room candidate data is consistent with the environmental mapping result, the human mapping result is determined as the human correction result. The room candidate data is updated based on the human correction result, and the risk intensity values ​​corresponding to physiological, activity, and indoor environmental data are extracted. The risk intensity values ​​are arranged according to the graph node sequence to obtain the room health characteristics. The risk intensity value range is set to 0 to 1.

[0039] In the spatial accessibility processing, the direct connectivity relationships between graph nodes are marked based on room connectivity data to obtain a set of directly connected node pairs; the direct connection cost is calculated based on the room coordinate data corresponding to two graph nodes in the set of directly connected node pairs; the graph node combinations that do not form a direct connectivity relationship are determined based on the room connectivity data to obtain a set of indirect node pairs; the set of indirect node pairs is expanded layer by layer, and the same graph node is prohibited from appearing repeatedly in a single expansion path to obtain a set of candidate paths; the set of directly connected node pairs contained in each expansion path in the candidate path set is accumulated based on the direct connection cost, and the number of coordinate direction changes is counted to obtain path candidate values; the target path cost is filtered according to the execution order of the accumulated direct connection cost from small to large, the number of coordinate direction changes from few to many, and the number of graph nodes from few to many; the direct connection cost, the target path cost, and the isolation cost are respectively used as the spatial accessibility cost between the corresponding graph nodes to obtain the spatial accessibility cost matrix. Subsequently, candidate node pairs are extracted from the spatial accessibility cost matrix. Risk intensity values ​​of the starting and ending graph nodes are extracted from the room health features to obtain risk data. The risk intensity values ​​of the starting and ending graph nodes are subtracted from the risk intensity values ​​of the ending graph nodes to obtain risk difference data. Forward backflow data, reverse correction data, and horizontal cost data are generated based on the risk difference data and written into the matrix row and column positions to obtain the node transmission cost matrix. The Sinkhorn algorithm is used to perform row and column normalization iterative processing on the node transmission cost matrix. The preset normalization condition is that the absolute value of the difference between the sum of weights in each row and 1 is no greater than 0.001, and the absolute value of the difference between the sum of weights in each column and 1 is no greater than 0.001. The maximum number of iterations is set to 100. Weights of starting graph nodes pointing to the same ending graph node are deleted based on the node masking bit. In each row, a preset number of ending graph nodes (3) are selected according to their transmission weights from largest to smallest to obtain a weighted neighbor node set.

[0040] The improved GraphSAGE model comprises a feature input layer, a weighted neighbor access layer, a dual-branch aggregation layer, a branch correction layer, and a representation output layer, with a 5-layer backbone. In implementation, the number of graph nodes is 5, and the maximum number of candidate neighbor nodes is 3. If there are fewer than 3 candidate neighbor nodes, the transmission weight of the node at the vacant position is reset to 0. The 3D spatial feature dimension is 7, the room health feature dimension is 3, and the representation output layer outputs a 5x12 health risk representation. The model's trainable parameters consist of a 7x7 spatial mapping weight matrix and a 3x3 health mapping weight matrix. Neither the spatial mapping weight matrix nor the health mapping weight matrix has bias terms, and the parameter size is 58 weight parameters. The training data came from historically collected room boundary data, room connectivity data, room coordinate data, physiological data, activity data, human position data, and indoor environmental data, as well as room risk levels from historical device linkage output records. The training batch size was set to 32, the learning rate to 0.001, and the number of iterations to 100. The level 1, level 2, level 3, and level 4 room risk levels were converted into supervised values ​​of 0.5, 1.5, 2.5, and 3.5, respectively. The mean squared error between the training output values ​​and the supervised values ​​was used to train the spatial mapping weight matrix and the health mapping weight matrix. After training, the mean squared error was verified to be 0.084.

[0041] In the risk gating correction process, the representation data corresponding to each starting graph node and candidate neighbor node is extracted from the health risk representation, and summed separately to obtain the gating sum data. The gating sum value corresponding to the candidate neighbor node is subtracted from the gating sum value corresponding to the starting graph node to obtain the gating differential data. When the gating differential data represents the gating sum value corresponding to the candidate neighbor node as greater than the gating sum value corresponding to the starting graph node, the gating differential data is incremented by 1 to obtain the enhancement candidate coefficient. The preset correction upper limit is set to 3. When the enhancement candidate coefficient is greater than 3, 3 is determined as the enhancement correction coefficient. When the coefficient is no greater than 3, the enhancement candidate coefficient is determined as the enhancement correction coefficient; when the gated summation value corresponding to the candidate neighbor node in the gated differential data representation is no greater than the gated summation value corresponding to the starting graph node, the value 1 is determined as the maintenance correction coefficient; the neighbor aggregation weight correction coefficient is determined based on the enhancement correction coefficient and the maintenance correction coefficient, and the node transmission weight is multiplied by the neighbor aggregation weight correction coefficient to obtain the corrected node weight; the representation data corresponding to the candidate neighbor node is weighted and summed according to the corrected node weight, and then concatenated with the representation data corresponding to the starting graph node to obtain the corrected health risk representation. The device linkage output module obtains room representation data based on the corrected health risk representation, sums the representation values ​​in the room representation data, and matches them with the preset risk level to obtain the room risk level; the preset risk level is set to level three, the preset environmental threshold is set to 0.6, the preset access threshold is set to 2, and the first linkage command, the second linkage command, or the third linkage command is generated based on the environmental access data to obtain the device linkage result.

[0042] Table 1 Results of Whole-House Health Monitoring

[0043] As shown in Table 1, the risk levels of the living room and bedroom do not exceed the preset risk levels. The device linkage output module sends the room level data to the home monitoring terminal to form the terminal display data. The sum of the room characteristics of the kitchen is 3.48, the room risk level is level four, the environmental risk value is 0.82 and the space access cost is 1.40, which meets the conditions that the risk intensity value exceeds the preset environmental threshold and the space access cost is lower than the preset access threshold. Therefore, the first linkage instruction containing fresh air control instructions and prompt instructions is generated. The sum of the room characteristics of the bathroom is 3.15, the environmental risk value is 0.76 and the space access cost is 2.30, which meets the conditions that the risk intensity value exceeds the preset environmental threshold and the space access cost is not lower than the preset access threshold. Therefore, the second linkage instruction containing environmental control instructions and prompt instructions is generated.

[0044] The maximum node transmission weight reaches 0.57 in the kitchen, indicating that the node transmission weight matrix can form a weighted set of neighbor nodes under the combined effect of spatial accessibility cost and risk intensity value, and provide input for the dual-branch aggregation of the improved GraphSAGE model. The above implementation results show that this invention can complete the construction of a three-dimensional spatial digital model, room health feature mapping, spatial accessibility cost calculation, Sinkhorn edge weight sampling, improved GraphSAGE model aggregation, risk gating correction, and device linkage output in the same process, enabling the spatial relationships throughout the house to participate in the health risk analysis process and forming corresponding linkages with home monitoring terminals and smart home environment devices.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A whole-house smart health monitoring and analysis system based on virtual reality, characterized in that: include: The spatial model construction module collects room boundary data, room connectivity data, and room coordinate data to establish a three-dimensional spatial digital model, determine graph nodes, and generate graph node sequences and three-dimensional spatial features. The health data mapping module collects physiological data, activity data, human location data, and indoor environmental data, and maps them to a graph node sequence to obtain the room's health characteristics. The access cost calculation module calculates the spatial access cost between graph nodes based on the room connectivity data and room coordinate data in the three-dimensional spatial features, and obtains the spatial access cost matrix. The transmission cost generation module correlates and fuses the spatial access cost in the spatial access cost matrix with the risk intensity value in the room health characteristics to obtain the node transmission cost matrix. The edge weight sampling generation module performs row and column normalization iterative processing on the node transmission cost matrix to generate the node transmission weight matrix, and filters the neighbor nodes in the graph node sequence to obtain a weighted neighbor node set. The dual-branch aggregation module inputs 3D spatial features and room health features into the dual-branch aggregation layer of the improved GraphSAGE model, performs spatial feature aggregation and health feature aggregation on the weighted neighbor node set, and generates a health risk representation. The risk gating correction module calculates the neighbor aggregation weight correction coefficient based on the health risk characterization and corrects the node transmission weight of the weighted neighbor node set to obtain the corrected health risk characterization. The device linkage output module determines the room risk level based on the corrected health risk profile and sends it to the home monitoring terminal. When the room risk level exceeds the preset risk level, it generates a prompt command and an environmental control command, outputs a prompt through the home monitoring terminal, and simultaneously sends the environmental control command to the smart home environment devices.

2. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 1, characterized in that, The spatial model construction module is specifically as follows: The boundary points in the room boundary data are transformed to a unified three-dimensional coordinate system, and the room ownership is marked according to the room coordinate data to obtain coordinate-based ownership data. Candidate room regions are established based on the coordinate-based attribution data, and coordinate region mapping results are generated according to the candidate room regions where the room coordinate data is located. The connectivity order between candidate room areas is determined based on the room connectivity data, and the connectivity order is mapped to the coordinate area mapping results to obtain the candidate graph node arrangement results; Based on the candidate graph node arrangement results, the boundary connection relationship between candidate room areas is read, and the boundary connection relationship is compared with the room connectivity data to obtain the node consistency test result. When the node consistency test result represents a boundary connection relationship that is inconsistent with the room connectivity data, the inconsistent candidate room area is taken as the backflow room area, and the room boundary data is selected to obtain the backflow boundary data. The candidate room regions are updated based on the reflux boundary data, and a three-dimensional spatial digital model is established. The room regions in the 3D spatial digital model are identified as graph nodes. The graph nodes are sorted according to the room connectivity data to obtain a graph node sequence. The room boundary data, room connectivity data and room coordinate data are extracted according to the graph node sequence to obtain the 3D spatial features.

3. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 1, characterized in that, The health data mapping module specifically comprises: The room area in the three-dimensional spatial digital model is extracted according to the graph node sequence, and the human body position data is converted to a unified three-dimensional coordinate system to obtain human body coordinate data. Based on the coordinate attribution relationship between human body coordinate data and room area, the human body coordinate data is mapped to the graph node sequence to obtain the human body mapping result; Physiological and activity data are mapped to graph node sequences according to the human body mapping results to obtain health candidate data; The indoor environmental data is mapped to a sequence of graph nodes according to room areas to obtain the environmental mapping results; The health candidate data and the environment mapping results are matched according to the same graph nodes to obtain the room candidate data; When the human body mapping result in the room candidate data is inconsistent with the environment mapping result, the candidate adjacent nodes are determined based on the environment mapping result and the room connectivity data. The distance between the human body coordinate data and the room boundary data corresponding to the candidate adjacent nodes is compared. The candidate adjacent node with the smallest distance is selected to update the human body mapping result and obtain the human body correction result. When the human body mapping result in the room candidate data is consistent with the environment mapping result, the human body mapping result is determined as the human body correction result. The room candidate data is updated based on the human body correction results, and the risk intensity values ​​corresponding to the physiological data, activity data and indoor environmental data are extracted. The risk intensity values ​​are arranged according to the graph node sequence to obtain the room health characteristics.

4. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 1, characterized in that, The access cost calculation module is specifically as follows: Based on the direct connectivity relationships between nodes in the room connectivity data marker graph, a set of directly connected node pairs is obtained; Based on the room coordinate data corresponding to two graph nodes in the set of directly connected node pairs, calculate the absolute value of the coordinate difference and sum them to obtain the direct connection cost; Based on the room connectivity data, identify the graph node combinations that do not form direct connectivity relationships in the graph node sequence to obtain the set of indirect node pairs; Based on the room connectivity data, the set of indirect node pairs is expanded layer by layer to obtain a set of candidate paths, and indirect node pairs that do not generate a set of candidate paths are marked as isolation costs. Based on the direct connection cost, the set of directly connected node pairs contained in each extended path in the candidate path set is accumulated, and the number of coordinate direction changes between adjacent sets of directly connected node pairs in each extended path is counted to obtain the path candidate value. When the same indirect node corresponds to multiple path candidate values, sequential filtering is performed to obtain the target path cost; When the same indirect node pair corresponds to only one path candidate value, the path candidate value is determined as the target path cost; The direct connection cost, target path cost, and isolation cost are respectively used as the spatial accessibility cost between the corresponding graph nodes, and sorted according to the graph node sequence to obtain the spatial accessibility cost matrix.

5. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 1, characterized in that, The transmission cost generation module specifically comprises: Extract the starting and ending graph nodes with spatial accessibility costs from the spatial accessibility cost matrix to obtain a set of candidate node pairs; Based on the starting and ending graph nodes in the candidate node pair set, risk intensity values ​​are extracted from the room health features to obtain risk data. Subtracting the risk intensity value of the terminal node from the risk intensity value of the starting node in the risk data yields the risk difference data. When the risk intensity value of the starting graph node represented by the risk difference data is greater than the risk intensity value of the ending graph node, the spatial accessibility cost is subtracted from the risk difference data to obtain the forward return flow data; When the risk intensity value of the starting graph node represented by the risk difference data is less than the risk intensity value of the ending graph node, the spatial accessibility cost is added to the absolute value of the risk difference data to obtain the reverse correction data. When the risk intensity value of the starting graph node is equal to the risk intensity value of the ending graph node, the spatial accessibility cost is determined as the horizontal cost data. The node transmission cost corresponding to the candidate node pair set is determined based on the forward backflow data, reverse correction data, and horizontal cost data, and written into the matrix row and column positions corresponding to the starting graph node and the ending graph node to obtain the node transmission cost matrix.

6. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 1, characterized in that, The edge weight sampling and generation module specifically comprises: Extract the matrix row and column positions and node transmission costs corresponding to the starting and ending graph nodes from the node transmission cost matrix to obtain cost row and column data. Based on the cost row and column data, the node transmission cost that is less than the preset minimum positive value is replaced with the preset minimum positive value, and the matrix row and column positions that are the same for the starting graph node and the ending graph node are marked as node mask positions to obtain positive cost data. The Sinkhorn algorithm is used to perform row normalization on the positive cost data to obtain row-normalized data, and the Sinkhorn algorithm is used to perform column normalization on the row-normalized data to obtain column-normalized data. The column-normalized data is fed back to the row-normalized processing, and the row-normalized and column-normalized processing are performed repeatedly until the preset normalization condition is met, thus obtaining the node transmission weight matrix. Based on the node masking bit, delete the weights of the starting graph nodes that point to the same ending graph nodes in the node transmission weight matrix, and filter a preset number of ending graph nodes to obtain candidate neighbor nodes. The candidate neighbor nodes are bound to their corresponding node transmission weights and written into the corresponding starting graph node according to the graph node sequence to obtain the weighted neighbor node set.

7. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 1, characterized in that, The dual-branch aggregation module is specifically as follows: Extract the starting graph node, candidate neighbor node, and corresponding node transmission weight from the weighted neighbor node set to obtain neighbor weight data; Based on the neighbor weight data, spatial features corresponding to the starting graph node and candidate neighbor nodes are extracted from the three-dimensional spatial features to obtain spatial neighborhood data. Based on the neighbor weight data, the risk intensity values ​​corresponding to the starting graph node and candidate neighbor node are extracted from the room health features to obtain the health neighborhood data; In the dual-branch aggregation layer of the improved GraphSAGE model, spatial features in the spatial neighborhood data are weighted and aggregated according to the node transmission weights to obtain spatial aggregated data. In the dual-branch aggregation layer of the improved GraphSAGE model, the risk intensity values ​​in the health neighborhood data are weighted and aggregated according to the node transmission weights to obtain health aggregated data. The values ​​in the spatial aggregated data and the health aggregated data are summed separately to obtain the spatial aggregated value and the health aggregated value. The spatial aggregated value is then subtracted from the health aggregated value to obtain the aggregated differential data. Construct an initial zero-value matrix containing spatial correction positions and health correction positions. When the aggregated differential data indicates that the health aggregate value is greater than the spatial aggregate value, write the aggregated differential data into the health correction position and keep the spatial correction position at zero. When the aggregated differential data indicates that the health aggregate value is not greater than the spatial aggregate value, write the absolute value of the aggregated differential data into the spatial correction position and keep the health correction position at zero to obtain the branch correction data. Health risk characterization is generated by splicing spatial aggregated data, health aggregated data, and branch correction data in a fixed order.

8. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 1, characterized in that, The risk gating correction module is specifically as follows: Extract the representation data corresponding to each starting graph node and candidate neighbor node from the health risk representation, and sum them respectively to obtain the gated sum data; Subtract the gated summation value corresponding to the starting graph node from the gated summation value corresponding to the candidate neighbor node in the gated summation data to obtain the gated differential data; When the gated summation value corresponding to the candidate neighbor node in the gated differential data is greater than the gated summation value corresponding to the node in the initial graph, the gated differential data is incremented by 1 to obtain the enhanced candidate coefficient. When the enhancement candidate coefficient is greater than the preset correction upper limit, the preset correction upper limit is determined as the enhancement correction coefficient; when the enhancement candidate coefficient is not greater than the preset correction upper limit, the enhancement candidate coefficient is determined as the enhancement correction coefficient. When the gated summation value corresponding to the candidate neighbor node in the gated differential data representation is not greater than the gated summation value corresponding to the node in the initial graph, the value 1 is determined as the maintenance correction coefficient. The neighbor aggregation weight correction factor is determined based on the enhancement correction factor and the preservation correction factor, and the node transmission weight in the weighted neighbor node set is multiplied by the neighbor aggregation weight correction factor to obtain the corrected node weight; The modified health risk representation is obtained by weighting and summing the representation data corresponding to the candidate neighbor nodes according to the modified node weights, and then concatenating it with the representation data corresponding to the nodes in the initial graph.

9. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 1, characterized in that, The device linkage output module is specifically: Extract the representation data corresponding to each graph node from the modified health risk representation, and obtain the room representation data according to the room area in the three-dimensional spatial digital model corresponding to the graph node sequence; The room risk level is obtained by summing the representative values ​​in the room representative data and matching them with the preset risk level. The room risk level is linked to the corresponding room area to obtain room level data; When the room risk level in the room risk level data does not exceed the preset risk level, the room risk level data is sent to the home monitoring terminal to obtain the data displayed on the terminal. When the room risk level in the room level data exceeds the preset risk level, the risk intensity value corresponding to the indoor environmental data of the corresponding room area is extracted from the room health characteristics, and the spatial accessibility cost between the corresponding room area and the adjacent room area is extracted according to the spatial accessibility cost matrix to obtain the environmental accessibility data. When the risk intensity value in the environmental accessibility data exceeds the preset environmental threshold and the space access cost is lower than the preset access threshold, a first linkage instruction containing fresh air control instructions and prompt instructions is generated. When the risk intensity value in the environmental accessibility data exceeds the preset environmental threshold and the spatial access cost is not lower than the preset access threshold, a second linkage instruction containing environmental control instructions and prompt instructions is generated. When the risk intensity value in the environmental accessibility data does not exceed the preset environmental threshold, a third linkage instruction containing a prompt instruction is generated; The prompts from the first, second, and third linkage commands are sent to smart home environment devices and home monitoring terminals to obtain device linkage results.

10. The whole-house smart health monitoring and analysis system based on virtual reality according to claim 7, characterized in that, The improved GraphSAGE model includes a feature input layer, a weighted neighbor access layer, a dual-branch aggregation layer, a branch correction layer, and a representation output layer, specifically: The feature input layer is connected to the weighted neighbor access layer, receiving 3D spatial features, room health features, and a set of weighted neighbor nodes, and aligning the nodes according to the graph node sequence to obtain the model input data; The weighted neighbor access layer is connected to the dual-branch aggregation layer. It extracts the starting graph node, candidate neighbor node and corresponding node transmission weight from the model input data to obtain neighbor weight data. The dual-branch aggregation layer includes a spatial aggregation branch and a health aggregation branch. The spatial aggregation branch performs weighted aggregation of three-dimensional spatial features based on neighbor weight data to obtain spatial aggregation data. The health aggregation branch performs weighted aggregation of risk intensity values ​​in room health features based on neighbor weight data to obtain health aggregation data. The branch correction layer is connected to the dual-branch aggregation layer, which sums the spatial aggregation data and the health aggregation data respectively to obtain the spatial aggregation value and the health aggregation value. The branch correction layer then subtracts the spatial aggregation value from the health aggregation value to obtain the aggregated differential data. The branch correction layer constructs an initial zero-value matrix containing spatial correction positions and health correction positions, and writes corresponding values ​​to the spatial correction positions and health correction positions according to the numerical direction of the aggregation difference data to obtain the branch correction data; The representation output layer is connected to the branch correction layer, and the data is spliced ​​together in a fixed order of spatial aggregated data, health aggregated data and branch correction data. The health risk representation is then output according to the graph node sequence.

Images