An old person multi-parameter monitoring intelligent monitoring method and system based on an internet of things

Abstract

The application provides an old person multi-parameter monitoring intelligent monitoring method and system based on Internet of Things, and relates to the technical field of intelligent monitoring. The method comprises the following steps: acquiring different types of monitoring data of a target object and determining real-time change quantity thereof; updating the weight of each monitoring data according to the change; dividing a computing node cluster based on the state of an edge computing node and distributing data processing tasks according to the weight; extracting a high-risk feature set from the cluster processing result and uploading the high-risk feature set to a cloud end for verification; and finally executing an early warning operation. The application can realize dynamic weight adjustment of multi-source monitoring data and edge-cloud collaborative processing, effectively balance the real-time performance and accuracy of data processing, and improve the accuracy and response speed of home safety monitoring of the old people.

Description

Technical Field

[0001] This disclosure generally relates to the field of intelligent monitoring technology, and in particular to an intelligent monitoring method and system for multi-parameter monitoring of the elderly based on the Internet of Things. Background Technology

[0002] As the aging population continues to grow, the health and safety monitoring of the elderly has become an important social concern. In particular, in home-based and community-based elderly care settings, how to effectively protect the lives of the elderly and improve their quality of life has become an urgent social need to be addressed.

[0003] Currently, monitoring technologies for the elderly mainly involve deploying various sensor devices in the home environment to collect data on the elderly's physical condition and environmental parameters, forming a multi-dimensional monitoring data stream. This data stream is then transmitted to a central processing unit for unified analysis, triggering corresponding early warning mechanisms based on the analysis results.

[0004] However, in existing technologies, the integration and processing of multi-source monitoring data faces significant challenges, and the process has inherent limitations when processing data. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in related technologies, this application provides an intelligent monitoring method and system for multi-parameter monitoring of the elderly based on the Internet of Things, which aims to solve the problems of insufficient multi-source information fusion capability, unreasonable division of labor between local and remote processing, and difficulty in balancing the real-time performance and accuracy of data processing in existing elderly monitoring methods.

[0006] Firstly, an intelligent monitoring method for multi-parameter monitoring of the elderly based on the Internet of Things is provided, including: Acquire monitoring data of different types of target objects and determine the real-time changes of each monitoring data; The weights of each monitoring data are updated based on the real-time changes in each monitoring data. Each edge computing node is divided into multiple computing node clusters based on its load and communication status. The monitoring data is then distributed to the corresponding computing node cluster for processing based on the updated weights. Extract the feature components that meet the high-risk conditions from the processing results returned by each computing node cluster, and use them as a high-risk feature set; The high-risk feature set is transmitted to the cloud for verification, and an early warning operation is performed based on the verification results.

[0007] This application provides an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly. By constructing a layered processing architecture with end-to-end collaboration, it achieves an integrated closed loop from data acquisition to early warning and intervention. Specifically, the method first acquires different types of monitoring data of the target object and determines the dynamic changes of each monitoring data in real time. Due to the adoption of a multi-source data fusion sensing strategy, it can comprehensively capture the elderly's movement status and environmental interaction information, laying a solid data foundation for subsequent accurate judgment. On this basis, the weights of each monitoring data are dynamically updated according to the real-time changes, enabling the system to adaptively identify the most influential risk factors in the current scenario, thereby avoiding misjudgments or omissions caused by a single data source dominating the judgment. Subsequently, the method divides the nodes into multiple computing clusters according to the load and communication status of each edge computing node, and allocates the monitoring data to the corresponding clusters for processing based on the updated weights. This edge collaboration mechanism effectively solves the contradiction between limited local computing power and remote transmission latency in traditional solutions. It utilizes the proximity processing capabilities of edge nodes to ensure response speed, and avoids resource idleness or overload through load balancing between clusters, achieving a balance between real-time performance and accuracy. Furthermore, features meeting high-risk criteria are extracted from the processing results returned by each computing node cluster to form a high-risk feature set. This refines massive amounts of raw data into key risk information, significantly reducing the burden of transmitting invalid data. Finally, the high-risk feature set is transmitted to the cloud for verification. The powerful computing resources of the cloud are used to perform secondary confirmation of high-risk events, and precise intervention instructions are generated based on preset warning conditions. Through the aforementioned layered processing mechanism of "local real-time response, edge collaborative sharing, and precise cloud verification," this method ensures both the timeliness of abnormal event triggering and the high reliability of the final warning, significantly improving the intelligence level and comprehensive protection capabilities of home safety monitoring for the elderly.

[0008] Secondly, an intelligent monitoring system for multi-parameter monitoring of the elderly based on the Internet of Things is provided, the system comprising: The acquisition and determination module is used to acquire monitoring data of different types of target objects and determine the real-time changes of each monitoring data. The update module is used to update the weight of each monitoring data according to the real-time changes of each monitoring data. The partitioning and allocation module is used to divide each edge computing node into multiple computing node clusters according to the load status and communication status of each edge computing node, and to allocate each monitoring data to the corresponding computing node cluster for processing based on the updated weights. The extraction module is used to extract the feature parts that meet the high-risk conditions from the processing results returned by each computing node cluster, and use them as a high-risk feature set. The verification and early warning module is used to transmit the high-risk feature set to the cloud for verification and to issue an early warning based on the verification results. Attached Figure Description

[0009] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, as provided in this application embodiment; Figure 2 A flowchart illustrating an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, provided in an embodiment of this application; Figure 3 A flowchart illustrating an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, provided in an embodiment of this application; Figure 4 A flowchart illustrating an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, provided in an embodiment of this application; Figure 5 A flowchart illustrating an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, provided in an embodiment of this application; Figure 6 A flowchart illustrating an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, provided in an embodiment of this application; Figure 7 A flowchart illustrating an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, provided in an embodiment of this application; Figure 8 A flowchart illustrating an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, provided in an embodiment of this application; Figure 9 A flowchart illustrating an IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly, provided in an embodiment of this application; Figure 10 This is a structural block diagram of an IoT-based intelligent monitoring system for multi-parameter monitoring of the elderly, provided in an embodiment of this application. Detailed Implementation

[0010] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0011] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0012] The following is combined with Figure 1 For an explanation of the IoT-based intelligent monitoring method for multi-parameter monitoring of the elderly provided in this application, please refer to [link / reference]. Figure 1 , Figure 1 A flowchart illustrating the steps of an exemplary embodiment of this application provides a method for intelligent monitoring of multiple parameters of the elderly based on the Internet of Things. The flowchart includes the following steps: Step S20: Obtain monitoring data of different types of target objects and determine the real-time changes of each monitoring data. The target group in this application may refer to elderly people who need to be monitored. These elderly people may be those living at home or those living in elderly care facilities.

[0013] Since the purpose of this application is to ensure the safety of the elderly and improve their quality of life, all monitoring data must be acquired with the elderly person as the core focus. Therefore, we can acquire data reflecting the elderly person's movement status, environmental data reflecting the elderly person's activity environment, and data reflecting the frequency of contact between the elderly person and obstacles. This data can include, for example, foot movement trajectory data, foot movement speed data, the frequency and intensity of contact between the elderly person and tables placed in their living environment, the density of obstacles in the elderly person's living environment, ground humidity, and ground flatness, among other types of data.

[0014] These data can be collected, for example, by using accelerometers, infrared rangefinders, pressure sensors, gyroscopes, lidar, anemometers, humidity sensors, and levels installed on the target object's feet. For instance, this application could collect the target object's movement speed at a frequency of 50 times per second using an ankle bracelet deployed on the target object's feet, and collect the obstacle density in the target object's living environment at a frequency of 50 times per second using lidar deployed in the target object's living environment.

[0015] After the data from the aforementioned multiple sources is collected by the corresponding acquisition devices, it can be transmitted to the processing devices via local area networks, wireless networks, wired networks, etc., for subsequent analysis.

[0016] The real-time change in monitoring data can refer to the numerical offset or statistical characteristic change of the corresponding monitoring data within a continuous sampling period. For example, the real-time change in foot movement trajectory data can be the rate of change of the standard deviation of the displacement vector magnitude within the current sampling window relative to the previous window; the real-time change in obstacle contact data can be the increment of the contact event count per unit time. This application can, for example, calculate the rate of change of the displacement vector magnitude of the foot movement trajectory data within the sliding time window to characterize the increasing trend of gait instability; this application can also, for example, determine the surge in contact behavior based on the increment of the event count of obstacle contact data per unit time.

[0017] For example, during the target object's nighttime activity, the processing device reads a set of six-axis raw data from the accelerometer and gyroscope every 200 milliseconds, and generates a three-dimensional trajectory point sequence of the feet after attitude calculation; at the same time, the infrared distance sensor detects the distance between the feet and fixed obstacles such as the bed edge and door frame at a frequency of 50Hz, and records a contact event when the distance is less than 15cm and the duration is more than 300ms; the processing device aligns the trajectory point sequence and the contact event sequence within the most recent second in time, calculates the rate of change of the standard deviation of the trajectory curvature and the frequency increment of the contact event, and outputs the two as the real-time changes of the foot movement trajectory data and the obstacle contact data, respectively.

[0018] Step S30: Update the weight of each monitoring data according to the real-time changes of each monitoring data. In this application, after obtaining the real-time changes of each monitoring data through the above steps, the original weights are not used for risk assessment. Instead, the weights of each monitoring data are redefined so that the risk assessment can be carried out again based on the updated weights.

[0019] It should be noted here that the weight of the monitoring data is a dimensionless coefficient used to characterize the relative importance of the corresponding monitoring data in the subsequent fusion judgment; this weight can be dynamically adjusted according to the real-time changes of each monitoring data, and is not fixed in advance; its value range is, for example, [0, 1], and the sum of the weights of all monitoring data is, for example, 1. For example, when the real-time change of the foot movement trajectory data increases, its weight increases accordingly; when the real-time change of the obstacle contact data increases significantly, its weight also increases synchronously.

[0020] For example, this application may determine the magnitude of weight adjustment based on the ratio of the real-time change of each monitoring data to its historical average; this application may also normalize the real-time change of each monitoring data to the [0, 1] interval and then allocate it to the corresponding weight increment in a linear proportion, this application does not limit this.

[0021] For example, during the system initialization phase, this application may assign initial weights of 0.4, 0.35, and 0.25 to foot movement trajectory data, obstacle contact data, and environmental monitoring data, respectively. When the real-time change in foot movement trajectory data exceeds twice the standard deviation of its historical mean, its weight is increased by 0.15. If the real-time change in obstacle contact data increases by 300% at the same time, its weight is increased by 0.2. After adjustment, the weights of the three become 0.55, 0.55, and 0.25.

[0022] Step S40: Divide each edge computing node into multiple computing node clusters according to the load status and communication status of each edge computing node, and allocate each monitoring data to the corresponding computing node cluster for processing based on the updated weights. One issue with existing technologies is the inconsistency in the division of labor between local and remote processing, as well as the latency in information processing. Therefore, this application addresses this by dividing dispersed edge computing nodes into multiple collaborative computing node clusters, providing a computing resource pool for subsequent monitoring data allocation.

[0023] The cluster partitioning is not static, but dynamically adjusted based on the real-time load and communication status of each edge computing node to adapt to changes in the network environment and computing load. This application can partition edge computing nodes that meet the same load balancing and communication connectivity constraints into the same computing node cluster. For example, this application can use a weighted composite load score based on the CPU utilization and remaining memory of each edge computing node, combined with inter-node communication latency to construct a communication quality matrix, and then use a K-means clustering algorithm to divide all nodes into k computing node clusters. Alternatively, this application can use a graph partitioning algorithm, constructing a communication topology graph with nodes as vertices and communication latency as edge weights, and partitioning into k computing node clusters according to the minimum cut principle. Furthermore, this application can also use a greedy strategy, sequentially adding the node with the lowest load to the current cluster until the communication latency of nodes within the cluster meets a threshold and the load variance is below a set limit to obtain k computing node clusters. This application does not limit the specific implementation of these methods.

[0024] It should be explained that the load status of an edge computing node can refer to the resource usage of a single edge computing node at the current moment, including CPU utilization, remaining memory, disk I / O throughput, etc.; the communication status of an edge computing node can refer to the communication quality indicators between the node and other edge nodes, including average communication latency between nodes, packet loss rate, bandwidth availability, etc.

[0025] After dividing the computing node clusters according to the above process, this application can allocate the monitoring data to computing node clusters with corresponding processing capabilities based on the weight of each monitoring data. The monitoring data with higher weights are preferentially allocated to clusters with larger load margins and lower communication latency. This can reflect the core idea of ​​this application to allocate execution resources on demand and maximize the overall processing efficiency.

[0026] For example, this application may connect a total of 8 edge computing nodes, numbered N1 to N8; real-time monitoring shows that the CPU utilization of N1 to N4 is 32%, 28%, 35%, and 30%, respectively, the remaining memory is all above 60%, and the communication latency between them is all below 45ms; the CPU utilization of N5 to N8 is all above 70%, and the communication latency fluctuates greatly; based on this, the system divides N1 to N4 into the first computing node cluster, and N5 to N8 are not included in the active cluster for the time being; when the weight of foot movement trajectory data is updated to 0.55, the system assigns it to the first computing node cluster for trajectory anomaly detection; when the weight of obstacle contact data is 0.55, it is also assigned to the same cluster for contact pattern analysis; environmental monitoring data, due to its low weight (0.25), is assigned to the lightweight sub-cluster N1 for separate processing.

[0027] Step S50: Extract the feature parts that meet the high-risk conditions from the processing results returned by each computing node cluster, and use them as the high-risk feature set; This step enables intelligent screening and fusion of distributed processing results into condensed risk information. After completing their assigned computing tasks, each computing node cluster returns the processing results to the processing device. These results are not raw monitoring data, but rather structured information after local analysis. Optionally, these results can also be encapsulated in structured data formats such as JSON or Protobuf, facilitating subsequent parsing and feature extraction.

[0028] For example, the first computing node cluster returns a processing result indicating that the obstacle density exceeds 2 per square meter and the wind speed remains above 12 m / s. This example demonstrates that the processing result includes at least: Quantitative indicators: obstacle density (e.g., 2 obstacles / square meter), wind speed (e.g., 12 meters / second); Status determination: Whether the preset threshold is exceeded ("exceeds", "continuously higher than").

[0029] High-risk conditions are pre-defined judgment rules used to define what situations constitute a high-risk state that requires cloud verification. Examples include obstacle density exceeding 2 per square meter, foot movement speed exceeding 2.5 meters per second, and obstacle contact frequency exceeding 5 times per minute.

[0030] The feature that meets the high-risk criteria can refer to the data segment that meets the preset high-risk criteria in the output results after each computing node cluster has completed local processing.

[0031] A high-risk feature set can refer to a structured data set formed by aggregating feature components that meet the high-risk criteria. Its elements may include feature identifiers, original data indexes, anomaly type labels, confidence scores, etc. This set may not contain the full content of the original monitoring data, but only retain the key feature dimensions that support the high-risk determination. There is no limitation on this.

[0032] For example, this application can filter entries with scores higher than 0.85 based on the anomaly score field in the results returned by each cluster; this application can also combine timestamp sequences to determine whether there are feature segments that trigger anomaly marking for more than 3 consecutive sampling periods; further, this application can also perform geofencing matching between the spatial coordinates output by each cluster and preset high-incidence areas of falls (such as toilet thresholds, stair corners) to extract feature entries that fall within the fence. This application can obtain a simplified set of high-risk clues that can be used for cloud-based deep verification based on any of the above methods, without limitation.

[0033] For example, in this application, the processing result returned by the first computing node cluster may include a feature record labeled as a sudden change in gait curvature, with an anomaly score of 0.92, a timestamp falling within the most recent 500ms window, and the corresponding foot spatial coordinates located within 1m of the bathroom doorway; this record meets the high-risk condition, is extracted, and written into the high-risk feature set; at the same time, another feature with a sudden increase in contact frequency, although scoring 0.88, is not included because its timestamp is isolated and has no continuity; the final high-risk feature set contains one feature entry, including feature identifier GaitCurv_001, original data index TS_20240512_192345_001, anomaly type label gait instability, and confidence score of 0.92.

[0034] Step S60: Transmit the high-risk feature set to the cloud for verification, and issue an early warning based on the verification results.

[0035] This step is the core link of the cloud layer in the layered processing mechanism of this application. It realizes the secondary confirmation of the high-risk feature set generated at the edge, ensuring the accuracy and reliability of the early warning decision. It can solve the technical problems of unreasonable division of labor between local and remote processing and difficulty in balancing the real-time performance and accuracy of data processing.

[0036] Specifically, since the computing node cluster is located at the edge, its high-risk judgment may be limited by computing resources and model accuracy. Therefore, this application further verifies it through the cloud, which has more powerful computing capabilities and more complex models, in order to reduce the false alarm rate.

[0037] Meanwhile, this application combines rapid response at the edge with cloud-based verification, with the two working together to balance response speed and accuracy.

[0038] In practical implementation, this application may, for example, first compress and encrypt the high-risk feature set, and then upload it to the cloud based on the optimal transmission path selected by the RTT measurement results.

[0039] Then, the high-risk feature set is divided into multiple feature subsets, which are then scheduled to different cloud nodes for parallel execution of lightweight model detection.

[0040] Finally, this application can also perform spatial consistency verification between the bounding box coordinates, obstacle categories, and confidence scores output from the cloud and the coarse localization results output by the computing node cluster, and only adopt the cloud results when the deviation is less than the tolerance threshold, thereby realizing cloud verification.

[0041] Furthermore, this application executes corresponding early warning operations based on the verification results in the cloud. These early warning operations may include, for example, further logically comparing the verification results with early warning conditions, and then, upon determining that the early warning conditions are met, generating an early warning signal, generating an intervention command, and sending it to the terminal device. The early warning conditions may include at least one of the following: the obstacle category belongs to a preset danger category (e.g., sharp furniture edges, unsecured carpet edges), the confidence level exceeds a preset confidence threshold (e.g., ≥0.9), or the estimated distance is less than a preset distance threshold (e.g., <0.8m).

[0042] For example, in this application, after the high-risk feature set 001 is uploaded to the cloud, it is assigned to the cloud node C3 with the lowest load; C3 calls the YOLOv8 model to perform joint analysis on its associated LiDAR point cloud and RGB image, and outputs the obstacle category as tilted slipper with a confidence level of 0.94 and an estimated distance of 0.62m from the center point to the foot; the result meets the warning condition that the distance estimate is less than the preset distance threshold (0.8m); the system then generates a warning signal to detect the obstacle with slipping risk near the foot, and generates an intervention command to start a voice reminder 'Please lift your foot to avoid the slipper in front', which is pushed to the smart bracelet terminal worn by the target object through the MQTT protocol.

[0043] This application achieves dynamic perception of risk evolution trends by acquiring different types of monitoring data of the target object and determining its real-time changes; it updates the weights of each monitoring data according to the real-time changes, enabling the system to adapt to the shift of risk focus; it divides the computing node cluster by the load and communication status of edge computing nodes and allocates processing tasks according to weights, improving local response efficiency and resource utilization rationality; it accurately extracts high-risk feature sets from the cluster return results, ensuring the representativeness and effectiveness of cloud data; and finally, through a closed loop of cloud verification and early warning operations, it balances judgment accuracy and response timeliness, thereby effectively solving the problems of insufficient multi-source information fusion capability, unreasonable division of labor between local and remote processing, and difficulty in balancing real-time data processing and accuracy in existing technologies.

[0044] In an optional embodiment, such as Figure 2 As shown, Figure 2 Another optional method embodiment of the IoT-based multi-parameter intelligent monitoring method for the elderly provided in this application includes: Step S201: Identify whether there is any abnormality in the movement state of the target object based on the foot movement trajectory data and obstacle contact data; The foot movement trajectory data can originate from accelerometers and gyroscopes deployed on smart shoes or ankle bracelets worn by the target individual. This data reflects the trajectory, speed, and acceleration of foot movement. The obstacle contact data can originate from infrared distance sensors and pressure sensors. This data reflects the frequency and force of contact between the foot and the obstacle. Foot movement trajectory data is used to characterize the gait dynamics of the target individual, while obstacle contact data is used to characterize the physical constraints of the target individual's interaction with the environment. The two are strongly coupled in the temporal dimension, and their joint analysis can reveal abnormal movement states such as pre-fall warning signs, tripping tendencies, or transitional imbalances.

[0045] This application may identify whether there is an anomaly in the motion state by means of whether the standard deviation of the displacement vector of the motion trajectory sequence exceeds the limit, and by means of whether the frequency of events per unit time of the contact event sequence exceeds the limit; this application may also identify whether there is an anomaly in the motion state by means of whether the rate of change of velocity curvature of the motion trajectory sequence changes abruptly, and by means of means of whether the distribution of contact duration of the contact event sequence is significantly offset, without limitation.

[0046] For example, in a home-based elderly care scenario, Mr. Zhang is walking along a corridor. An accelerometer and gyroscope collect three-axis acceleration and angular velocity data of his feet 50 times per second. After coordinate transformation and integration, a continuous sequence of foot movement trajectories is generated. At the same time, a pressure sensor deployed on the sole of the shoe and an infrared distance sensor 1.2 meters in front of him synchronously capture contact events and record the time and force of each foot approaching or touching an obstacle. When the system detects that the standard deviation of the foot displacement vector within a 1-second sliding window reaches 0.18m (exceeding the threshold of 0.12m), and the frequency of contact events within the window reaches 4 times (exceeding the threshold of 2 times), it determines that the current movement state is abnormal.

[0047] Step S202: When an abnormality in the motion state is detected, a preliminary risk warning signal is generated; The preliminary risk warning signal can be a lightweight asynchronous notification signal generated locally at the edge to initiate the subsequent risk feature extraction process. Generating the preliminary risk warning signal at the edge eliminates the need to wait for a response from the cloud, ensuring rapid response. Furthermore, the preliminary risk warning signal acts as a logical switch for transmitting abnormal events from the local perception layer to the intermediate processing layer. Its generation signifies that the system has confirmed suspicious behavior requiring further analysis from the original sensor stream, thus avoiding redundant processing of all monitoring data and reducing edge computing resource consumption.

[0048] For example, in this application, when the above-mentioned motion state is determined to be abnormal, the terminal device immediately generates a binary warning signal with a length of 64 bytes, wherein the first 8 bytes are a timestamp (Unix millisecond level), the 9th to 16th bytes are an abnormal identification code (0x02 indicates abnormal foot-obstacle interaction), the 17th to 24th bytes are a data source index (0x0103 indicates left foot trajectory + right foot obstacle contact), and the remaining bytes are reserved; the signal is written to the shared memory area for the risk feature extraction module to poll and read.

[0049] Step S203: In response to the generation of the preliminary risk warning signal, extract the risk characteristic information of each monitoring data; Risk characteristic information refers to a highly discriminative subset of data that has been screened, extracted, standardized, and structured from all monitoring data. This risk characteristic information is used to transform raw multi-source sensor data into semantic representations for risk modeling. Its extraction process is triggered by preliminary risk warning signals, ensuring that the high-overhead feature engineering process is only initiated when there are confirmed abnormal clues, thus improving system efficiency. This information forms the technical basis for subsequent judgment of potential risk events.

[0050] For example, after receiving the aforementioned preliminary risk warning signal, the system may call the local model loader to load a pre-trained anomaly detector and perform joint inference on the foot movement trajectory data and obstacle contact data within 5 seconds before and after the signal triggering time; the output anomaly probability is 0.93, and based on this, all acceleration peak points, contact force mutation points and corresponding timestamps within the 5-second window are extracted to form an initial anomaly dataset; then, the Z-score is used to normalize the data of each dimension, and then the labeling is completed according to the risk level rules (contact force >20N and duration >300ms is marked as high risk, 10–20N is marked as medium risk); finally, K-means (k=3) is used to cluster the anomaly points, and low-dimensional components with a PCA contribution rate of less than 5% are removed to form the final risk feature information.

[0051] Step S204: When a potential risk event is identified based on the risk characteristic information, determine the real-time change of each monitoring data.

[0052] Among them, a potential risk event can refer to an intermediate state that has not yet developed into an actual fall or collision, but has the potential to evolve into a high-risk event. The criteria for its determination are, for example, that the abnormal score in the risk feature information exceeds the score threshold and the time series analysis confirms the existence of an environmental obstacle event. This event indicates that the processing equipment needs to move from the local anomaly identification stage to the global dynamic evaluation stage.

[0053] Potential risk events serve as key state transition points connecting local perception and edge collaborative processing. Their identification results directly trigger the quantitative confirmation of the real-time changes in each monitoring data, providing an input premise for updating the weights of each monitoring data based on the real-time changes in each monitoring data in this application.

[0054] For example, this application may determine whether the z-score of each abnormal data in the risk feature information exceeds a score threshold (e.g., 2.5). If so, a sliding window model is invoked to perform temporal pattern recognition on the risk feature information. Alternatively, this application may determine whether the temporal aggregation density of abnormal data in the risk feature information exceeds a unit time threshold (e.g., ≥3 high-risk markers appear every 2 seconds). If so, a local time series analysis module is activated to determine whether an environmental obstacle event exists. Furthermore, this application may determine whether an environmental obstacle event exists based on whether the cross-correlation coefficient between the contact event sequence and the motion trajectory sequence in the risk feature information is continuously lower than 0.3 within multiple consecutive windows. This application does not limit the scope of this application.

[0055] This application achieves early identification of abnormal movement states in the elderly through joint modeling of foot movement trajectory data and obstacle contact data; it uses preliminary risk warning signals as a lightweight triggering mechanism to accurately initiate the subsequent risk feature extraction process; the extracted risk feature information further supports the reliable identification of potential risk events, thereby driving the confirmation of real-time changes in monitoring data in a closed loop; this mechanism completes the key leap from raw perception to risk semantic transformation without relying on cloud intervention, significantly improving the risk sensitivity and response efficiency of the local perception layer, and providing solid and reliable front-end anomaly detection capabilities to support the multi-parameter monitoring intelligent monitoring method of the above embodiments.

[0056] In an optional embodiment, such as Figure 3 As shown, Figure 3 An optional method embodiment provided in this application for identifying whether the movement state of a target object is abnormal based on foot movement trajectory data and obstacle contact data includes: Step S301: Construct a motion trajectory sequence based on foot motion trajectory data, and construct a contact event sequence based on obstacle contact data; The purpose of this step is to transform the raw sensor data into a structured time series, providing a data foundation for subsequent correlation analysis.

[0057] A motion trajectory sequence is a temporal organization of foot motion trajectory data, which can be represented in the following form: M(t)={(x1, y1, z1, v1, a1, t1), (x2, y2, z2, v2, a2, t2),..., (x n y n , z n v n a n , t n )} Where: (x, y, z) are the three-dimensional spatial coordinates; v is the instantaneous velocity; a is the acceleration vector; t is the timestamp.

[0058] This application may construct motion trajectory sequences by means of timestamp alignment and sliding window aggregation; this application may also construct motion trajectory sequences by means of Kalman filtering smoothing and trajectory segment resampling, and this application does not limit it in this way.

[0059] A contact event sequence is a temporal organization of obstacle contact data. A contact event sequence can be represented in the following form: C(t) = {(d1, f1, t1), (d2, f2, t2), ..., (d...} n f n , t n)}, where: d: contact force (measured by a pressure sensor); f: contact frequency (number of contacts per unit time); t: timestamp.

[0060] This application may construct a contact event sequence based on a criterion of exceeding the limit of the contact signal amplitude and duration; this application may also construct a contact event sequence based on a criterion of multi-sensor signal fusion (infrared distance abrupt change + pressure signal rising edge), and this application does not limit it in this way.

[0061] Step S302: Analyze the temporal correlation between the motion trajectory sequence and the contact event sequence; In this application, the motion trajectory sequence and the contact event sequence are aligned and correlated on the time axis through this step to explore the temporal relationship between the two types of events.

[0062] In this embodiment, the analysis process does not rely on a pre-trained deep learning model, but is based on joint statistical modeling within a sliding time window. For example, it includes: calculating motion stability indicators such as displacement standard deviation, velocity kurtosis, and acceleration zero-crossing rate within a unit window for motion trajectory sequences; calculating interaction activity indicators such as event count, average contact intensity, and event interval variance within a unit window for contact event sequences; and then quantifying the collaborative change trend of the two under different time lags through methods such as cross-correlation functions, Granger causality tests, or simple time-series offset matching.

[0063] For example, this application may set the sliding time window length to 2 seconds and the step size to 0.5 seconds; calculate the standard deviation of acceleration of the motion trajectory sequence (reflecting the intensity of motion) and the event count of the contact event sequence (reflecting the frequency of interaction) in each window; then calculate the cross-correlation coefficient between the two in the range of 0 to 1 second lag; if the absolute value of the cross-correlation coefficient in a certain window is greater than 0.6 and the lag is concentrated around 0.3 seconds, it is considered that there is a strong leader-response correlation between the motion change and the contact behavior in that window; this result serves as one of the intermediate pieces of evidence for judging whether the motion state is abnormal.

[0064] Step S303: If the motion trajectory sequence shows abnormal fluctuations and the frequency of the contact event sequence exceeds a preset threshold within a preset time window, then it is determined that the motion state of the target object is abnormal.

[0065] The preset time window can refer to the time interval used for local observation and joint judgment, and its time span is a fixed duration or an adaptively adjusted dynamic interval. The window is used to limit the data range on which the analysis depends, so as to ensure the real-time performance and local stability of the judgment. In this embodiment, the length of the window is set to 1.5 seconds for example.

[0066] Abnormal fluctuations can refer to dynamic characteristics of a motion trajectory sequence that significantly deviate from the normal gait pattern within a preset time window, including but not limited to abrupt changes in displacement, sudden drops in velocity, abnormal increases in acceleration kurtosis, and non-periodic oscillations in trajectory curvature. Such fluctuations need to be confirmed by comparison with local historical gait templates or population statistical benchmarks, rather than solely based on instantaneous extreme values. For example, this application can extract multiple feature indicators from the motion trajectory sequence, and determine the existence of abnormal fluctuations when any indicator exceeds a corresponding threshold. Alternatively, a comprehensive scoring method can be used, determining an anomaly when the weighted comprehensive score exceeds a threshold.

[0067] The frequency exceeding a preset threshold can refer to the number of events occurring in a contact event sequence within a preset time window exceeding an upper limit set empirically or obtained through dynamic learning. This threshold is used to distinguish between normal touches (such as lightly wiping a wall) and high-risk touches (such as repeatedly kicking or bumping into obstacles), and its value can be personalized according to the target object's age, activity level, and environmental complexity. This application may, for example, use Poisson process fitting and overdispersion tests to determine whether the frequency of the contact event sequence exceeds the preset threshold; alternatively, this application may use a method comparing the cumulative distribution of event counts within a sliding window with a quantile threshold to determine whether the frequency of the contact event sequence exceeds the preset threshold, and this application does not limit this method.

[0068] For example, in this application, the system can maintain a sliding window with a length of 1.5 seconds during continuous monitoring; when the standard deviation of the acceleration of the motion trajectory sequence within the window exceeds the local 7-day average plus twice the standard deviation, and the event count of the contact event sequence is greater than 5 times, it is determined that there is an abnormality in the motion state within the current window.

[0069] This application achieves structured modeling of foot movement behavior and environmental interaction behavior by constructing movement trajectory sequences and contact event sequences. By analyzing the correlation between the two in the time dimension, a mapping bridge between movement intention and actual interaction is established. Furthermore, through a dual constraint judgment mechanism of abnormal fluctuations and frequency exceeding thresholds within a preset time window, highly robust identification of abnormal movement states is completed on the edge side without relying on cloud computing power. This effectively reduces the false alarm rate caused by single-modal noise or individual behavioral differences, and improves the ability to capture real fall precursor behaviors (such as staggering, dragging, and kicking obstacles), thus providing a reliable premise for subsequent risk classification and early warning response.

[0070] In an optional embodiment, such as Figure 4 As shown, Figure 4 An optional method embodiment for extracting risk characteristic information from various monitoring data provided in this application includes the following steps: Step S401: Sort the monitoring data by importance and select the target risk dataset based on the sorting results; This step, as a specific implementation of risk feature information extraction, provides a high-quality data foundation for subsequent judgment of potential risk events.

[0071] This application assesses and ranks the importance of monitoring data from different sources and of different types, selecting the most valuable key data for current risk assessment to form a target risk dataset. Importance ranking can be based on data source accuracy, preset weights, priority scores, etc. This application, by selecting the target risk dataset through importance ranking, focuses on key data sources, avoids redundant processing of low-relevance and low-stability data, thereby reducing subsequent computational load and improving the targeting and efficiency of risk feature extraction.

[0072] This application may, for example, rank all monitoring data from high to low based on a priority score calculated by comprehensively considering the accuracy of each monitoring data source in identifying historical risk events, data update frequency, signal-to-noise ratio, and the strength of causal correlation with abnormal motion states.

[0073] For example, foot movement trajectory data, obstacle distribution data, and subjective risk perception data can be collected from three data sources: smart shoe sensors, environmental cameras, and user feedback, respectively. Assuming the accuracy rate of the smart shoe sensor data is 95%, the environmental camera is 80%, and the user feedback is 60%, a weighted average algorithm is used to calculate the priority score. The formula is: Priority Score = Accuracy × Weight, where the weights are 0.5, 0.3, and 0.2, respectively. The final score is 47.5 for the smart shoe sensor, 24 for the environmental camera, and 12 for the user feedback. Therefore, the priority ranking is that of the smart shoe sensor is the highest, and the foot movement trajectory and obstacle contact frequency are selected as the target risk dataset.

[0074] Step S402: Extract abnormal data from the target risk dataset and perform standardization processing to obtain an abnormal dataset; Here, abnormal data can refer to data segments in the target risk dataset whose values ​​deviate from the statistical distribution range of the data type under normal movement conditions by more than a preset threshold; for example, this application can set an abnormal threshold (such as foot movement speed exceeding 2.5 m / s or obstacle contact frequency exceeding 5 times / minute) and combine it with a fast hash index algorithm to quickly extract abnormal data segments.

[0075] Standardization can eliminate the influence of units, unify data scale, and quantify the degree of anomalies. This application, for example, may use a standardization formula to standardize outlier data, which is: Z = (X - μ) / σ; where: Z: standardized data value; X: original data value; μ: data mean; σ: data standard deviation.

[0076] For example, this application may calculate the mean μ1 and standard deviation σ1 of foot movement trajectory data in the target risk dataset in the most recent 60 seconds of normal walking samples, and set the anomaly judgment threshold as |X−μ1| / σ1>3; when a certain segment of 5 consecutive sampling points (corresponding to 100ms) all meet this condition, the 10 sampling points before and after it (a total of 30 sampling points, corresponding to 600ms) are extracted as an abnormal data segment; z-score standardization is performed on the segment to obtain standardized abnormal data; similarly, the same process is performed on obstacle contact data, and finally all standardized abnormal segments are aligned by timestamp and merged into an abnormal dataset.

[0077] Step S403: Perform risk level labeling, cluster reduction and noise filtering on each abnormal data in the abnormal dataset to obtain risk feature information.

[0078] This step is used to deeply refine the standardized anomaly dataset, generating high-quality risk feature information through three layers of processing: risk level labeling, cluster reduction, and noise filtering.

[0079] Risk level labeling can refer to assigning risk level labels (levels one to three) to anomalous data based on its deviation, duration, frequency, and matching degree with known high-risk event patterns. Level one might represent potential interference, level two might represent moderate risk, and level three might represent high risk. Risk level labeling is used to establish a risk stratification management mechanism. The risk level can be mapped to the corresponding risk level by looking up a table, for example, based on the absolute value of the anomalous data's z-score, the duration of the anomalous event, and its frequency of occurrence within a preset time window. Alternatively, this application can use a pre-trained lightweight decision tree model, inputting the temporal statistical characteristics (such as kurtosis, skewness, and zero-crossing rate) and frequency energy ratio of the anomalous data, and outputting a risk level prediction result; this application does not limit the specific approach.

[0080] For example, this application may determine the risk level of an anomaly as level three if the z-score of an anomaly is greater than 0.8 and the z-score of another anomaly is greater than 0.7.

[0081] Clustering reduction employs the K-means clustering algorithm to divide anomalous segments in the dataset with similar fluctuation patterns, temporal distribution characteristics, and amplitude response properties into several clusters. Each cluster uses its centroid as a representative feature vector, achieving data dimensionality compression and semantic aggregation. This application uses clustering reduction to eliminate redundant similar anomalies and retain typical patterns.

[0082] For example, this application uses the K-means clustering algorithm, with Euclidean distance as the similarity measure, to divide all standardized abnormal fragment vectors in the abnormal dataset into K clusters (K is a positive integer, ranging from 2 to 5), and the center vector of each cluster is used as the representative feature of this type of abnormality; this application may also use a hierarchical clustering algorithm, which constructs a clustering tree based on the dynamic time-normalized distance between abnormal fragments, and cuts from top to bottom to obtain the optimal cluster division, and this application does not limit this.

[0083] The noise filtering process employs principal component analysis to reduce the dimensionality of the clustered feature vectors and removes minor components with a cumulative contribution rate below a preset threshold (85%). Simultaneously, it combines sliding window midpoint filtering to remove impulse interference. This application uses noise filtering to suppress unstructured interference and improve feature robustness. The noise filtering process, along with the aforementioned risk level labeling and cluster reduction, works synergistically to output risk feature information that is structurally clear, semantically explicit, and highly resistant to interference.

[0084] For example, this application calculates the mean absolute value of the z-score, the maximum duration (based on the number of consecutive abnormal sampling points), and the number of occurrences within the last 10 seconds for all standardized abnormal segments in the abnormal dataset. These values ​​are then input into a preset rule table. If the mean z-score is >4.5, the duration is >8 sampling points, and the number of occurrences is ≥3, the segment is labeled as a Level 3 risk. Subsequently, all Level 3 risk abnormal segments are input into a K-means clustering module (K=3) to obtain three cluster centers, representing three typical patterns: emergency stop foot braking, high-frequency obstacle probing contact, and large-scale posture imbalance. PCA dimensionality reduction is performed on each cluster center vector (retaining the first two principal components), and then mean filtering is applied to the dimensionality-reduced vectors. Finally, three two-dimensional risk feature vectors are output, which, together with their corresponding risk level labels, constitute the risk feature information.

[0085] This application focuses on key data sources by prioritizing importance, unifies the scale of multi-source heterogeneous signals by extracting and standardizing abnormal data, and then achieves hierarchical discrimination, cluster reduction to preserve typical patterns, and noise filtering to enhance feature robustness through risk level labeling. Finally, it collaboratively generates risk feature information with clear structure, explicit semantics, and strong anti-interference ability, providing a high-quality input foundation for the accurate identification of potential risk events in this application.

[0086] In an optional embodiment, such as Figure 5 As shown, Figure 5 An optional method embodiment provided in this application for determining the real-time changes of various monitoring data when a potential risk event is identified based on risk characteristic information includes the following steps: Step S501: Determine the anomaly score of each abnormal data in the risk feature information; This step is used to quantify and score each abnormal data component in the risk feature information, converting the degree of abnormality into a comparable numerical indicator, and providing a quantitative basis for subsequent threshold judgment.

[0087] Anomaly score can refer to a numerical indicator used to quantify the degree to which a single anomalous data point deviates from the normal distribution in risk feature information. Its calculation does not depend on a specific algorithm model and can be achieved based on common evaluation dimensions such as statistical outlier, density bias, or gradient change rate. For example, this application may calculate the anomaly score by normalizing the distance between each anomalous data point and its cluster center in the risk feature information using the z-score normalization method; it may also quantify the local density bias of each anomalous data point using the Local Anomaly Factor (LOF) algorithm; furthermore, this application may also calculate the anomaly score based on the ratio of the gradient magnitude change rate of each anomalous data point in the time dimension to the historical mean, without limitation.

[0088] For example, this application may extract the third type of abnormal data subset after K-means clustering from the risk feature information output by this application, calculate the Euclidean distance between each data point in the subset and the center of the cluster, and standardize it using the z-score formula. The result is the abnormal score of each abnormal data. When an abnormal score is greater than 2.5, it is determined that it exceeds the normal fluctuation range and proceeds to the next judgment process.

[0089] Step S502: If there is abnormal data with abnormal scores exceeding the score threshold, perform time series analysis on the risk feature information to determine whether there is an environmental obstacle event. The scoring threshold can be a preset numerical boundary used to distinguish between general anomalies and high-risk anomalies requiring special attention. Its setting is not bound to a specific value and can be dynamically configured based on historical false alarm rate statistics, equipment accuracy calibration, or simulation results of typical fall scenarios. The scoring threshold serves as a decision threshold for whether to initiate time-series modeling analysis. Only when at least one anomaly score exceeds this threshold will the time-series analysis capability of the local edge node be invoked, avoiding redundant modeling of low-risk disturbances. This threshold is loaded during system initialization or periodically updated by the cloud based on recent warning accuracy feedback and then distributed to the edge side.

[0090] Time series analysis refers to the modeling and reasoning process of the evolution of risk characteristic information over continuous time segments. It does not rely on a specific neural network structure and can be implemented using methods with time-dependent modeling capabilities, such as sliding window statistics, autoregressive modeling, or recurrent neural networks. Time series analysis is used to identify whether abnormal patterns exhibit persistence, periodicity, or accelerating deterioration trends, thereby distinguishing isolated anomalies from real environmental obstacle events. This analysis process takes the risk characteristic information output by this application as input and the existence of environmental obstacle events as the binary output objective. It runs locally on edge computing nodes and does not rely on cloud intervention.

[0091] An environmental obstacle event can refer to a real risk situation in which an elderly person's movement is hindered or their balance is lost due to physical obstacles in the home environment (such as thresholds, scattered items, or slippery areas on the ground). It does not refer to a specific type of obstacle, but is a general term for a type of environmental state that is spatially intrusive, frequently encountered, and disruptive to movement. It is a key condition for determining whether it is necessary to reconfirm the real-time changes in monitoring data. Its existence conclusion is output by the time series analysis module and serves as a logical switch to trigger the subsequent weight adjustment mechanism. The determination result of this event is directly related to the timing of the action of determining the real-time changes of each monitoring data in this application.

[0092] For example, this application may use a sliding window model to model the time series of risk feature information, with a window length of 10 seconds and a step size of 1 second. The model outputs the probability value of whether an environmental obstacle event exists in each window. Alternatively, this application may construct a dual threshold criterion based on the combination of the mean and variance of the abnormal scores in the sliding window. When the mean is higher than the first threshold for three consecutive windows and the variance is lower than the second threshold, it is determined that an environmental obstacle event exists. This application does not limit this to any particular method.

[0093] For example, this application may select risk feature information from the time period before and after the abnormal data with the highest abnormal score, totaling 15 seconds, to construct a temporal tensor with a length of 150 frames (sampling frequency of 10Hz); input it into a lightweight model deployed on edge nodes, the model having 64 hidden layer units and an output layer that is a classifier; if the model output probability is higher than 0.85 for 5 consecutive frames, it is determined that there is an environmental obstacle event; the determination result is marked as True and passed to the next step.

[0094] Step S503: If an environmental obstacle event exists, determine the real-time change of each monitoring data.

[0095] The real-time changes of each monitoring data point have already been described above and will not be repeated here. These changes can refer to the dynamic shift of foot movement trajectory data, obstacle contact data, and environmental monitoring data relative to the previous stable period at the current moment.

[0096] In this embodiment, the real-time changes of each monitoring data are the direct input basis for the weight update operation in this application. The determination process is strictly limited to after the environmental obstacle event is confirmed, ensuring that the weight adjustment only responds to the real and continuous risk evolution process, rather than instantaneous noise or individual behavioral fluctuations.

[0097] This application quantifies the intensity of risk characteristic information through anomaly scoring and uses a scoring threshold to filter out high-risk anomalies requiring in-depth analysis. Furthermore, it identifies whether these high-risk anomalies exhibit continuous evolutionary characteristics through time-series analysis, thereby accurately determining whether an environmental obstacle event exists. Based on this, the real-time change confirmation process for foot movement trajectory data, obstacle contact data, and environmental monitoring data is initiated only when an environmental obstacle event is confirmed. This three-level progressive mechanism effectively avoids false triggers caused by instantaneous disturbances, sensor jitter, or individual sporadic behavior, ensuring that the data change signals relied upon by subsequent dynamic weight adjustment, edge task allocation, and cloud verification have real risk orientation and spatiotemporal consistency, ultimately improving the robustness and clinical usability of the entire intelligent monitoring method for multi-parameter monitoring of the elderly.

[0098] In an optional embodiment, such as Figure 6 As shown, Figure 6 An optional method embodiment for updating the weights of each monitoring data point based on the real-time changes of each monitoring data point, provided in this application, includes the following steps: Step S601: Determine the degree of influence of each monitoring data based on the real-time changes of each monitoring data. The degree of influence can refer to the strength of the driving role of the monitoring data in the evolution of the current risk situation. Its technical focus is on the sensitivity contribution of the data change to the overall risk assessment result. In this embodiment, the degree of influence is used to characterize whether the change of a certain type of monitoring data is dominating the formation or aggravation of the risk situation. For example, a sudden change in the acceleration of foot movement trajectory data reflects a premonition of a fall, a sudden increase in the contact frequency of obstacle contact data indicates an abnormal increase in the density of spatial obstacles, and a rapid increase in ground humidity in environmental monitoring data indicates an increased risk of slipping. The degree of influence is positively correlated with the real-time change, that is, the more drastic the change, the more significant its influence on the current risk assessment.

[0099] For example, in a home-based elderly care scenario, when the obstacle contact data collected by the infrared distance sensor jumps from an average of 0.3 times / second to 2.8 times / second within a 1-second time window, the change reaches 2.5 times / second, exceeding the historical average by ±1.5 times standard deviation, and the influence of this data is determined to be significantly increased. At the same time, if the foot movement trajectory data collected by the accelerometer shows a pulse signal with an absolute acceleration value exceeding 5g and a duration of ≥200ms within the same window, its influence is also synchronously increased. Based on this, the system identifies that the coordinated changes of the two constitute a high-risk movement pattern, thus providing an input basis for subsequent weight adjustments.

[0100] Step S602: Update the weight of each monitoring data according to the degree of influence of each monitoring data and the adjustment rules. The adjustment rules include: the higher the degree of influence of the monitoring data, the greater the weight.

[0101] The adjustment rule can refer to a functional relationship constraint that maps the degree of influence to weight coefficients; it can be a linear proportional relationship, a piecewise threshold mapping relationship, or a monotonically increasing nonlinear mapping relationship, etc., without limitation. In this embodiment, the adjustment rule is used to ensure that the system automatically enhances the voice of the most active and discriminative data channel during the dynamic risk evolution process, thereby improving the targeting and response efficiency of task allocation in the edge node cluster; for example, when the degree of influence of obstacle contact data increases, its weight increases accordingly, so that when monitoring data is subsequently allocated to the computing node cluster, this type of data is preferentially scheduled to the cluster with stronger computing power and more stable communication for processing.

[0102] The weight of monitoring data can refer to the weighted contribution coefficient of this type of data to the final risk assessment result during the multi-source data fusion decision-making process. This application, for example, updates the weight of each monitoring data point using a linear scaling method based on the magnitude of the impact; or it updates the weight of each monitoring data point using a step-by-step assignment method based on whether the impact exceeds a preset grading threshold.

[0103] For example, during initialization, the initial weights for foot motion trajectory data, obstacle contact data, and environmental monitoring data are 0.4, 0.35, and 0.25, respectively. When the influence of obstacle contact data is detected to rise from a low level to a high level (corresponding to the real-time change exceeding the threshold), its weight is dynamically increased from 0.35 to 0.55. At the same time, the influence of foot motion trajectory data remains at a medium level, and its weight is slightly adjusted to 0.38. The influence of environmental monitoring data remains unchanged, and its weight remains unchanged at 0.25. The updated weight combination is immediately applied to the subsequent steps of allocating each monitoring data to the corresponding computing node cluster to guide the task scheduling strategy.

[0104] This application uses the real-time changes of each monitoring data as a quantitative basis for the degree of influence, and dynamically updates the weights according to the adjustment rule that the higher the degree of influence, the greater the weight. This enables the system to adaptively focus on the monitoring dimension with the most discriminative value under different home environments and behavioral states, avoiding the dilution of key risk factors due to fixed weights. This supports the edge computing node cluster to prioritize and efficiently coordinate high-risk data, and ultimately improves the real-time response accuracy and scenario robustness of the multi-parameter monitoring system.

[0105] In an optional embodiment, such as Figure 7 As shown, Figure 7 An optional method embodiment provided in this application for allocating monitoring data to corresponding computing node clusters for processing using updated weights includes the following steps: Step S701: Determine the priority of each monitoring data according to the updated weights; The updated weights are quantitative indicators representing the relative importance of the monitoring data in the current risk assessment task. They are used to map the processing priority of the monitoring data, and the mapping relationship can be linear, piecewise threshold, or normalized ranking. Priorities are used to guide the edge-side task scheduler in processing different monitoring data and allocating resources accordingly, ensuring higher execution guarantees for risk identification tasks corresponding to high-weight data. For example, this application can normalize and sort the data according to the updated weights, arranging each monitoring data from high to low weight values ​​and assigning priority numbers 1, 2, 3, etc.; this application can also compare the updated weights with a preset priority threshold group, with weights falling into different threshold ranges corresponding to different priority levels (e.g., weight ∈ [0.8, 1.0] → high priority; [0.5, 0.8) → medium priority; [0, 0.5) → low priority). For example, when the updated weight of foot movement trajectory data is 0.92, obstacle contact data is 0.85, and environmental humidity data is 0.31, after normalization and sorting, they are assigned priorities 1, 2, and 3 respectively.

[0106] Step S702: Based on the priority of each monitoring data and the load and communication status of each computing node cluster, the monitoring data is allocated to the corresponding computing node cluster for processing.

[0107] The priority of the monitoring data is used to characterize the level of computing resources to which the monitoring data should be scheduled in the current risk assessment task. In this embodiment, the allocation of monitoring data is completed based on the joint matching relationship between priority and cluster status. Its function is to avoid high-priority data being allocated to high-load or high-latency clusters, thereby ensuring the timely extraction and feedback of key risk features.

[0108] For example, when foot motion trajectory data is determined to be priority 1, and cluster A has a load status score of 0.82 (out of 1) and a communication status score of 0.91, while cluster B has a load status score of 0.45 and a communication status score of 0.63, the scheduler selects cluster A to perform trajectory sequence modeling and fluctuation analysis of the data.

[0109] This application achieves targeted scheduling of high-risk related data by converting the updated weights into monitoring data priorities and combining them with the load and communication status of each computing node cluster for joint matching. On this basis, by leveraging a priority-driven resource selection mechanism, it ensures that high-weight monitoring data such as foot movement trajectory data and obstacle contact data are allocated to clusters with strong computing power and high communication quality for processing, thereby improving the execution efficiency and reliability of key risk identification tasks. Ultimately, without increasing investment in edge hardware, it significantly enhances the system's real-time response capability to high-risk events such as sudden falls and trips.

[0110] In an optional embodiment, such as Figure 8 As shown, Figure 8 An optional method embodiment for transmitting a high-risk feature set to the cloud for verification, provided in this application, includes the following steps: Step S801: Divide the high-risk feature set into multiple feature subsets; The high-risk feature set can refer to the data set consisting of features that meet the high-risk criteria extracted from the processing results returned by each computing node cluster. This set may include multi-dimensional temporal feature vectors, spatial coordinate information, and event identifiers. The feature subset can refer to several non-overlapping sub-data sets that cover the entire set, formed by logically segmenting the high-risk feature set according to data dimensions, time segments, or semantic categories. This feature subset can be a subset divided according to the sensor type to which the feature belongs (e.g., a foot motion trajectory subset, an obstacle contact subset, an environmental monitoring subset), a subset divided according to a time sliding window (e.g., feature sequences within the last 1, 2, and 3 seconds), or a subset divided according to the requirements of the target detection task (e.g., a geometric subset containing only spatial coordinates, an appearance subset containing only intensity and texture). No specific limitations are imposed on the comparison.

[0111] In this embodiment, dividing the high-risk feature set into multiple feature subsets serves to: enable subsequent target detection tasks to be deployed in parallel on multiple cloud nodes, reducing the computational pressure on a single node; support different lightweight strategies for different subsets; and facilitate dynamic matching of the optimal detection capability node based on the data characteristics of each subset.

[0112] Step S802: Assign each feature subset to multiple target cloud environments that meet preset load conditions for target detection and analysis, and obtain verification results. The verification results include at least obstacle category, confidence level, and location information.

[0113] Among them, the load status meeting the preset conditions can refer to at least one of the following indicators: the current CPU utilization of the target cloud node is lower than the first threshold (e.g., 70%), the remaining memory is higher than the second threshold (e.g., 30%), and the network latency is lower than the third threshold (e.g., 50ms).

[0114] The target cloud can refer to the cloud computing nodes that participate in this distributed verification task. These nodes have the ability to run lightweight target detection models and are connected to a unified task scheduling and result aggregation service.

[0115] Target detection analysis can refer to the technical process of performing obstacle recognition, localization, and confidence assessment on a subset of input features, which is a general task in the field of computer vision; the verification result can refer to the structured judgment conclusion output after target detection analysis, which is used to reverse verify whether the high-risk features identified by the local edge side truly reflect the real risks.

[0116] In an optional embodiment, the verification result of this application may include the obstacle category identified by the target cloud, the confidence level of the identified obstacle, and the location information of the identified obstacle. The obstacle category may be the semantic category of the obstacle obtained by the detection model set on the target cloud, such as steps, wires, furniture edges, or slippery areas on the ground; the confidence level may be a probabilistic evaluation value of the identified obstacle category by the detection model set on the target cloud, with a value ranging from 0 to 1; the identified obstacle may be represented by its relative coordinates in three-dimensional space or its bounding box coordinates in a two-dimensional image plane, used to characterize its spatial distribution.

[0117] This application enables elastic scheduling and fault isolation of computing resources by allocating each feature subset to the target cloud; improves the consistency and robustness of detection results through multi-node collaboration; and avoids overall verification failure due to the anomaly of a single node.

[0118] This application can, for example, allocate each feature subset to the node with the lowest overall load score by using a weighted round-robin strategy based on the real-time load metrics reported by each target cloud node, combined with the estimated data size and computational complexity of the feature subset; this application can also, for example, match target cloud nodes with corresponding acceleration capabilities (such as GPUs, NPU inference units) based on the data modality of the feature subset (such as time series, image, point cloud), and prioritize those whose load status meets preset conditions; furthermore, this application can also construct a node capability profile based on historical task response latency and accuracy feedback, and dynamically select the node most suitable for the current feature subset type to perform target detection analysis, etc.

[0119] For example, there are 5 available target cloud instances (A to E), each node periodically reporting its own load status. In a certain verification task, the processing device determines that A (CPU utilization 62%, memory remaining 38%, latency 42ms), C (CPU utilization 55%, memory remaining 41%, latency 47ms), and D (CPU utilization 68%, memory remaining 33%, latency 49ms) meet preset conditions. The processing device assigns the aforementioned trajectory subset to A to run the LSTM enhanced temporal classifier, and assigns the contact subset to C to run the event graph neural network. The network model assigns a spatial subset to D to run the YOLOv8s lightweight object detection model. After the three nodes complete the analysis synchronously, they respectively send back the verification results: A outputs a step-type obstacle with a confidence score of 0.83, located 1.2m to the left front; C outputs a high-frequency contact event with a confidence score of 0.91, corresponding to touching the edge of the step; D outputs the obstacle category: step, with a confidence score of 0.87, and bounding box coordinates [320, 180, 410, 260] (pixels). The processing device performs spatiotemporal alignment and logical fusion on the three results to generate the final verification result.

[0120] This application achieves parallel decomposition and elastic resource scheduling of verification tasks by dividing the high-risk feature set into multiple feature subsets and distributing them to multiple target cloud environments with load conditions that meet preset conditions for target detection and analysis. By leveraging the proprietary processing capabilities of each target cloud environment for different modal feature subsets, the accuracy and granularity of obstacle recognition are improved. Furthermore, based on the cross-comparison and fusion of multi-source verification results, the credibility and anti-interference capability of high-risk judgment are enhanced, thereby effectively supporting the accurate triggering and timely response of early warning operations.

[0121] In an optional embodiment, such as Figure 9 As shown, Figure 9 An optional method embodiment for performing early warning operations based on verification results provided in this application includes the following steps: Step S901: Compare the verification results with the preset warning conditions; The verification result can refer to the structured data set output by the cloud-based target detection and analysis, such as the obstacle categories identified by the cloud, the confidence level of the identified obstacles, and the location information of the identified obstacles.

[0122] Preset warning conditions can refer to a set of logical judgment rules used to determine whether to trigger a warning response. In essence, they are the union of multi-dimensional risk criteria; they are a set of configurable business rules set according to the home safety standards for the elderly.

[0123] For example, if the verification result is the obstacle category identified by the target cloud, the preset warning can be that the obstacle category belongs to a preset danger category, such as the edge of a step, exposed wires, or unsecured carpet edges.

[0124] If the verification result is a confidence level of obstacle detection, the preset warning can be that the confidence level exceeds a preset confidence threshold, such as 0.75.

[0125] Optionally, this application may further invert the distance estimate between the target object and the obstacle when the verification result includes the location information of the identified obstacle. The distance estimate may be represented by Euclidean distance, Manhattan distance, or an approximate distance calculated based on a depth map. Then, the distance estimate is compared with a preset distance threshold to determine whether the preset warning condition is met. It is understood that this application may set the example estimate being less than or equal to the preset distance threshold as the preset warning condition. The preset distance threshold may be, for example, 0.8 meters.

[0126] For example, the verification result returned by the cloud includes an obstacle detection record, whose obstacle category is step edge, which belongs to the preset danger category set; at the same time, its confidence level is 0.82, which is greater than the preset confidence threshold of 0.75; and its distance estimate is 0.63 meters, which is less than the preset distance threshold of 0.8 meters; at least one of the three conditions is met, so it is determined that the preset warning conditions are met.

[0127] Step S902: When the verification result meets the preset warning conditions, a warning signal and an intervention command corresponding to the warning signal are generated. Among them, the early warning signal can refer to a structured notification message used to characterize the existence of a risk event that requires immediate attention, such as a data packet including a timestamp, risk level, obstacle type, location of occurrence, and suggested response action; its role is to serve as a cross-module communication carrier to transmit unified semantic decision output to subsequent intervention execution units.

[0128] Intervention commands can refer to operation commands generated based on warning signals that can be directly parsed and executed by terminal devices, such as text prompt commands, audio playback commands, vibration reminder commands, path replanning commands, or device control commands. In this embodiment, its role is to transform abstract risks into specific behavioral guidance, thereby achieving a closed loop from risk awareness to risk avoidance.

[0129] This application may generate intervention instructions by looking up a table and matching a predefined intervention template based on the combination of risk level and obstacle type carried in the warning signal; this application may also generate intervention instructions by dynamically adjusting the intensity of the intervention instructions based on the distance estimate, for example, the closer the distance, the louder the voice prompt and the higher the vibration frequency.

[0130] For example, when the warning signal indicates that the obstacle is the edge of a step and the estimated distance is 0.63 meters, the system calls the near-step avoidance intervention template to generate an intervention instruction that includes a voice message saying "There is a step ahead, please lift your heels and step slowly" along with a synchronized vibration rhythm.

[0131] Step S903: Send the intervention command to the terminal device to which the target object belongs.

[0132] The terminal device can refer to a hardware device deployed around the target object that has the ability to receive and execute intervention commands, such as a smart bracelet, a walking robot controller, an embedded voice broadcast module, or a smartphone. In this embodiment, its role is to serve as the physical execution port for the final risk response, completing the last link in human-machine collaborative intervention.

[0133] For example, when the processing device recognizes that the target object is currently wearing a smart bracelet and the Bluetooth connection is normal, it encrypts the above-generated voice intervention command and pushes it to the bracelet via the wireless network. After the bracelet decrypts, it triggers the local engine to play the corresponding voice prompt and simultaneously starts the wrist dual-frequency vibration feedback.

[0134] This application constructs a robust and highly configurable early warning triggering logic by comparing the verification results with multi-dimensional preset early warning conditions in parallel and by using three complementary mechanisms: obstacle category matching, confidence threshold judgment, and distance estimation comparison. On this basis, it generates structured early warning signals and semantically aligned intervention instructions, and delivers the instructions accurately to the adapted terminal devices through multiple communication paths. This enables millisecond-level response and multimodal intervention after risk confirmation, effectively improving the ability of the elderly to avoid accidents such as falls.

[0135] It should be noted that although the operations of the method of the present invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart may be performed in a different order.

[0136] Furthermore, such as Figure 10 As shown, this application also provides an intelligent monitoring system for multi-parameter monitoring of the elderly based on the Internet of Things. The system includes: an acquisition and determination module 1001, an update module 1002, a division and allocation module 1003, an extraction module 1004, and a verification and early warning module 1005.

[0137] The acquisition and determination module 1001 is used to acquire monitoring data of different types of target objects and determine the real-time changes of each monitoring data. The update module 1002 is used to update the weight of each monitoring data according to the real-time changes of each monitoring data. The partitioning and allocation module 1003 is used to divide each edge computing node into multiple computing node clusters according to the load status and communication status of each edge computing node, and to allocate each monitoring data to the corresponding computing node cluster for processing based on the updated weights. The extraction module 1004 is used to extract the feature parts that meet the high-risk conditions from the processing results returned by each computing node cluster, and use them as a high-risk feature set. The verification and early warning module 1005 is used to transmit high-risk feature sets to the cloud for verification and to perform early warning operations based on the verification results.

[0138] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A multi-parameter intelligent monitoring method for the elderly based on the Internet of Things, characterized in that, include: Acquire monitoring data of different types of target objects and determine the real-time changes of each monitoring data; The weights of each monitoring data are updated based on the real-time changes in each monitoring data. Each edge computing node is divided into multiple computing node clusters based on its load and communication status. The monitoring data is then distributed to the corresponding computing node cluster for processing based on the updated weights. Extract the feature components that meet the high-risk conditions from the processing results returned by each computing node cluster, and use them as a high-risk feature set; The high-risk feature set is transmitted to the cloud for verification, and an early warning operation is performed based on the verification results.

2. The intelligent monitoring method for multi-parameter monitoring of the elderly based on the Internet of Things according to claim 1, characterized in that, The monitoring data includes foot movement trajectory data and obstacle contact data, and the method further includes: The system identifies whether the target object's movement state is abnormal based on the foot movement trajectory data and the obstacle contact data. When an abnormality is detected in the motion state, a preliminary risk warning signal is generated; In response to the generation of the preliminary risk warning signal, risk characteristic information of each monitoring data is extracted; When a potential risk event is identified based on the risk characteristic information, the real-time change of each monitoring data point is determined.

3. The intelligent monitoring method for multi-parameter monitoring of the elderly based on the Internet of Things according to claim 2, characterized in that, The step of identifying whether the target object's movement state is abnormal based on the foot movement trajectory data and the obstacle contact data includes: A motion trajectory sequence is constructed based on the foot motion trajectory data, and a contact event sequence is constructed based on the obstacle contact data; Analyze the temporal correlation between the motion trajectory sequence and the contact event sequence; If the motion trajectory sequence exhibits abnormal fluctuations and the frequency of the contact event sequence exceeds a preset threshold within a preset time window, then the motion state of the target object is determined to be abnormal.

4. The method according to claim 2, characterized in that, The extraction of risk characteristic information from each of the monitoring data includes: The monitoring data is ranked by importance, and the target risk dataset is selected based on the ranking results; Extract abnormal data from the target risk dataset and perform standardization processing to obtain an abnormal dataset; The risk feature information is obtained by performing risk level labeling, cluster reduction, and noise filtering on each of the abnormal data in the abnormal dataset.

5. The method according to claim 4, characterized in that, When a potential risk event is identified based on the risk characteristic information, determining the real-time change of each monitoring data item includes: Determine the anomaly score for each of the abnormal data in the risk feature information; If there is abnormal data with abnormal scores exceeding the score threshold, then the risk feature information is analyzed over time to determine whether an environmental obstacle event exists. If the aforementioned environmental obstacle event exists, the real-time change in each of the monitoring data is determined.

6. The method according to claim 1, characterized in that, The step of updating the weights of each monitoring data according to the real-time changes of each monitoring data includes: The degree of influence of each monitoring data is determined based on the real-time changes in each monitoring data. The weights of each monitoring data point are updated based on the degree of influence of each monitoring data point and in accordance with adjustment rules, wherein the adjustment rules include: the higher the degree of influence of the monitoring data point, the greater its weight.

7. The method according to claim 1, characterized in that, The step of allocating each monitoring data point to the corresponding computing node cluster for processing using the updated weights includes: The priority of each monitoring data is determined based on the updated weights; Based on the priority of each monitoring data and the load and communication status of each computing node cluster, the monitoring data is allocated to the corresponding computing node cluster for processing.

8. The method according to claim 1, characterized in that, The step of transmitting the high-risk feature set to the cloud for verification includes: The high-risk feature set is divided into multiple feature subsets; Each of the aforementioned feature subsets is assigned to multiple target cloud environments that meet preset load conditions for target detection and analysis, thereby obtaining the verification results.

9. The method according to claim 8, characterized in that, The alert operation based on the verification results includes: The verification results are compared with preset warning conditions; When the verification result meets the preset warning condition, a warning signal and an intervention command corresponding to the warning signal are generated; The intervention command is sent to the terminal device to which the target object belongs.

10. An intelligent monitoring system for multi-parameter monitoring of the elderly based on the Internet of Things, characterized in that, The system includes: The acquisition and determination module is used to acquire monitoring data of different types of target objects and determine the real-time changes of each monitoring data. The update module is used to update the weight of each monitoring data according to the real-time changes of each monitoring data. The partitioning and allocation module is used to divide each edge computing node into multiple computing node clusters according to the load status and communication status of each edge computing node, and to allocate each monitoring data to the corresponding computing node cluster for processing based on the updated weights. The extraction module is used to extract the feature parts that meet the high-risk conditions from the processing results returned by each computing node cluster, and use them as a high-risk feature set. The verification and early warning module is used to transmit the high-risk feature set to the cloud for verification and to issue an early warning based on the verification results.

Images