A safety monitoring system for a solitary elderly person

Abstract

The application relates to a lonely old people safety monitoring system, belonging to the technical field of wisdom endowment, the system comprises a millimeter wave radar sensing module for generating three-dimensional point cloud data of a human body; a posture reconstruction unit reconstructs a skeleton and identifies a posture category through a PointNet++ network; a micro-motion feature extraction unit extracts a breathing frequency and a body motion micro-motion amplitude from radar echo; an environment semantic sensing module collects multi-source time sequence signals through sensors arranged in a household pipeline and household appliance nodes, and generates a behavior semantic vector through a Transform model; a multi-modal mutual verification unit, when an abnormal posture is detected, synchronously judges whether the breathing frequency is lower than a first preset threshold value, whether the body motion micro-motion amplitude is lower than the product of the average body motion amplitude in a normal activity state and a second preset threshold value, and whether the behavior semantic vector before the occurrence of the abnormality matches a preset rule library, and outputs a high-risk alarm when the three conditions are simultaneously satisfied.

Description

Technical Field

[0001] This invention relates to the field of smart elderly care technology, specifically a safety monitoring system for elderly people living alone. Background Technology

[0002] Currently, safety monitoring technologies for elderly people living alone are mainly divided into three categories: The first category is monitoring solutions based on environmental sensors. These solutions deploy sensors at water, electricity, and gas pipelines, as well as doors and windows, analyzing data on water, electricity, and gas usage and door magnetic switches to indirectly infer the elderly person's home activity status. While this type of solution does not require the elderly person to wear a device, it only provides indirect information at the environmental level and cannot directly perceive the elderly person's posture and physiological state. When an elderly person experiences a fall or other emergency, the environmental data often shows no obvious abnormalities, leading to a high false negative rate. The second category is monitoring solutions based on wearable devices. These devices, such as smart bracelets, smart insoles, and beacon tags, collect data on the elderly person's heart rate, blood oxygen, gait, and posture. This type of solution can directly obtain the elderly person's physiological and movement information, but it heavily relies on the user's active wearing and maintenance. Practical problems exist, such as forgetting to wear the device, unwillingness to wear it, insufficient device battery, and beacon signal obstruction, resulting in low user compliance in actual implementation. The third type is vision-based monitoring solutions, which use cameras to capture video images and computer vision technology to identify the elderly’s posture. Although this type of solution can achieve contactless monitoring, it has technical bottlenecks such as privacy concerns, dependence on lighting conditions, occlusion problems, and algorithm processing delays, making it difficult to apply widely in home environments.

[0003] Existing technologies include dual-modal monitoring schemes that combine environmental and attitude perception. For example, an environmental perception layer is constructed using water flow sensors, power sensors, gas sensors, and door magnetic sensors, while an attitude perception layer is formed using UWB and BLE dual-mode beacons, including head, hand, torso, and foot beacons. Attitude analysis is performed by scanning the relative distances between beacons using beacon base stations, and a rescan is triggered for cross-validation when environmental data is abnormal. While this type of scheme reduces the false alarm rate to some extent through dual-modal collaboration, its attitude perception relies entirely on the user wearing multiple beacon devices. This not only increases the user's burden but also requires manual surveying and installation at multiple sites based on the building layout, resulting in high deployment costs. Furthermore, its cross-validation mechanism is based solely on a secondary scan of the beacon spacing threshold, making the verification dimension relatively simple. When beacon signals are obstructed or interfered with, the verification reliability decreases. Furthermore, such solutions lack the ability to perceive physiological micro-movement characteristics of the elderly, such as respiratory rate and amplitude of body movements, and cannot distinguish between abnormal stillness caused by falls and normal sleep or rest states, leaving room for improvement in terms of false alarms and missed alarms. To address these issues, this invention proposes a safety monitoring system for elderly people living alone that does not require wearable devices and integrates millimeter-wave radar point cloud attitude reconstruction with multi-source environmental behavioral semantic mutual verification. Summary of the Invention

[0004] In order to solve the problems of the prior art, the present invention provides a safety monitoring system for elderly people living alone.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Firstly, a safety monitoring system for elderly people living alone, comprising: The millimeter-wave radar sensing module is deployed within the monitoring area to transmit frequency-modulated continuous wave signals and receive echo signals, generating three-dimensional point cloud data of the target human body. The attitude reconstruction unit is connected to the millimeter-wave radar sensing module and is used to input the three-dimensional point cloud data into the attitude recognition network to reconstruct the three-dimensional skeleton structure of the target human body and identify the attitude category. The micro-motion feature extraction unit, connected to the millimeter-wave radar sensing module, is used to extract the respiratory rate and body micro-motion amplitude of the target human body from the echo signal; wherein, the body micro-motion amplitude is obtained by energy integration of the envelope of the chest cavity displacement time-domain signal within a preset time window, and the calculation formula is as follows: In the formula The amplitude of the body's micro-motion. This represents the envelope of the time-domain signal of thoracic cavity displacement. Preset time window; The environmental semantic perception module includes a sensor group deployed at the water, electricity, and gas pipelines and key home appliance nodes to collect multi-source environmental time-series signals and input the multi-source environmental time-series signals into the behavioral semantic model to generate a behavioral semantic vector within a preset time window. The behavioral semantic vector contains the probability value of each preset behavioral semantic category. The multimodal mutual verification unit is connected to the posture reconstruction unit, the micro-motion feature extraction unit, and the environmental semantic perception module, respectively. When the posture recognition result is an abnormal posture, it is used to simultaneously acquire the current breathing rate, body micro-motion amplitude, and the behavioral semantic vector sequence within a preset time window before the abnormality occurred. When the following three conditions are met simultaneously, a high-risk alarm signal is output: First condition: The respiratory rate is lower than a first preset threshold, where the first preset threshold is the lower limit of the normal resting respiratory rate of a human body; The second condition is that the amplitude of the micro-movement is lower than the product of the second preset threshold and the average amplitude of the micro-movement in the normal activity state. The second preset threshold is used to characterize the degree of decline in human mobility after a fall. The third condition is that within a preset time window before the anomaly occurs, there is at least one moment in which the sum of the probabilities of the preceding behavioral semantic categories associated with the anomaly posture category exceeds the third preset threshold.

[0006] In one specific embodiment of the first aspect, the center frequency of the millimeter-wave radar sensing module is 60 GHz, the transmission bandwidth is 4 GHz to 6 GHz, each frame of three-dimensional point cloud data contains no less than 200 scattering points, and the frame rate is no less than 10 frames per second; the three-dimensional point cloud data includes the three-dimensional spatial coordinates and radial velocity of each scattering point.

[0007] In one specific embodiment of the first aspect, the pose recognition network is a PointNet++ network or a graph convolutional network, the input is the three-dimensional point cloud data, and the output is the coordinate vector of each major joint of the human body in three-dimensional space; the pose categories include at least one of standing, sitting, lying down, falling, leaning forward, and falling sideways; when the recognition results of multiple consecutive frames are falling, leaning forward, or falling sideways, it is determined to be an abnormal pose.

[0008] In one specific embodiment of the first aspect, the micro-motion feature extraction unit includes: The phase demodulation module is used to select the distance cell with the highest energy from the echo signal, and perform phase unwrapping on the slow time dimension signal of the distance cell to obtain the time domain signal of the thoracic cavity displacement. A bandpass filter with a passband of 0.1 Hz to 0.8 Hz is used to extract the respiratory frequency component from the time-domain signal of chest displacement. The respiratory frequency is obtained by peak detection of the extracted respiratory signal. The respiratory frequency is defined as the number of breaths per minute. The body motion amplitude calculation module is used to perform Hilbert transform on the time-domain signal of thoracic cavity displacement to obtain an envelope signal, and calculate the energy integral of the envelope signal as the body motion micro-amplitude within a preset time window, wherein the preset time window is 5 seconds to 30 seconds.

[0009] In one specific embodiment of the first aspect, the sensor group of the environmental semantic perception module includes: a vibration sensor or flow sensor deployed on the water pipe entering the house, a non-invasive current sensor deployed on the electricity meter entering the house, a dual-mode gas sensor deployed on the gas pipeline entering the house, and a vibration sensor or current characteristic sensor deployed at the home appliance node, wherein the home appliance node includes at least one or more of a refrigerator, microwave oven, washing machine, and water heater. The behavioral semantic model is a Transformer model or a Long Short-Term Memory network. The input is a data matrix of the multi-source environmental time-series signal within a time window, which is 30 minutes to 120 minutes. The output is a behavioral semantic probability vector. The behavioral semantic categories include at least one of the following: using the toilet, cooking, eating, taking medication, sleeping, going out, and getting up at night.

[0010] In one specific implementation of the first aspect, the mean amplitude of body movement during normal activity is defined as the median value of body movement amplitude calculated every 10 minutes from 8:00 to 20:00 each day for 7 consecutive days. The first preset threshold is 10 times / minute to 18 times / minute; The second preset threshold is 30% to 50%; The third preset threshold is 0.6 to 0.8; The preset time window before the anomaly occurs is 10 to 60 minutes. The semantic category of the preceding behavior associated with the abnormal posture category is determined by an association rule mining algorithm. Association rules with a confidence of not less than 70% and a support of not less than 5% are selected to construct a rule base. The confidence threshold and support threshold are determined by ROC curve analysis. The threshold corresponding to the lowest false positive rate is selected on the training set under the condition that the true positive rate is not less than 95%.

[0011] In one specific implementation of the first aspect, an edge computing node is further included, deployed within the monitoring area, for integrating the millimeter-wave radar sensing module, the attitude reconstruction unit, the micro-motion feature extraction unit, the environmental semantic perception module, and the multimodal mutual verification unit, to complete attitude recognition, micro-motion feature extraction, behavioral semantic generation, and mutual verification decision locally; the edge computing node only uploads high-risk alarm signals and desensitized feature data to the cloud or community server, the desensitized feature data includes attitude category labels, respiratory rate values, body micro-motion amplitude values, and behavioral semantic category labels, but does not include original point cloud data and original environmental time-series signals.

[0012] Secondly, a method for monitoring the safety of elderly people living alone includes the following steps: S1. The millimeter-wave radar sensing module transmits frequency-modulated continuous wave signals and receives echo signals to generate three-dimensional point cloud data of the target human body. S2. The three-dimensional point cloud data is input into the attitude recognition network through the attitude reconstruction unit to reconstruct the three-dimensional skeleton structure and identify the attitude category; S3. The respiratory rate and body movement amplitude of the target human body are extracted from the echo signal by the micro-motion feature extraction unit. The body movement amplitude is obtained by energy integration of the envelope of the chest cavity displacement time-domain signal within a preset time window. The calculation formula is as follows: ; S4. Collect multi-source environmental time-series signals through the environmental semantic perception module and generate behavioral semantic vectors within a preset time window; S5. When the identified posture category is an abnormal posture, the respiratory rate, body movement amplitude, and behavioral semantic vector sequence within the preset time window before the abnormality occurs are obtained simultaneously. If the respiratory rate is lower than the first preset threshold, the body movement amplitude is lower than the product of the second preset threshold and the average body movement amplitude in the normal activity state, and there is at least one moment within the preset time window before the abnormality occurs where the sum of probabilities of the preceding behavioral semantic categories associated with the abnormal posture category exceeds the third preset threshold, then a high-risk alarm signal is output.

[0013] In one specific embodiment of the second aspect, step S3, extracting the respiratory rate, specifically includes: Perform a distance-dimensional fast Fourier transform on the echo signal and select the distance cell with the highest energy as the distance cell where the human target is located. Phase unwrapping is performed on the slow-time dimension signal of the selected distance cell to obtain the time-domain signal of the thoracic cavity displacement; The time-domain signal of the thoracic cavity displacement was filtered using a bandpass filter ranging from 0.1 Hz to 0.8 Hz to obtain the respiratory signal; Peak detection is performed on the respiratory signal, and the time interval between adjacent peaks is calculated. The respiratory rate is defined as the number of breaths per minute, which is obtained by taking the median and then taking the reciprocal of the time interval.

[0014] In one specific implementation of the second aspect, step S4, generating the behavioral semantic vector, specifically includes: Data cleaning and time alignment are performed on the multi-source environmental time-series signals collected by the sensor array; The aligned timing signal is sliced ​​according to a preset time window, with each time window being 60 minutes long and a sliding step of 15 minutes. Multi-source time-series data within each time window are input into a trained Transformer model, which outputs a feature vector. This feature vector is then mapped to a behavioral semantic probability vector via a linear layer and a softmax function. The behavioral semantic probability vector contains probability values ​​for each preset behavioral semantic category, which includes at least one of the following: using the toilet, cooking, eating, taking medication, sleeping, going out, and getting up at night.

[0015] The beneficial effects of this invention are as follows: 1. By constructing a dual-modal heterogeneous architecture of millimeter-wave radar point cloud perception and environmental semantic perception, non-contact synchronous acquisition of human posture and physiological micro-motion features is achieved without relying on wearable devices. Specifically, the posture reconstruction unit uses the PointNet++ network to perform joint reconstruction and temporal posture classification on 3D point cloud data of no less than 200 scattering points per frame, enabling the identification of abnormal postures such as falls, forward leaning, and side falls without any physical contact. The micro-motion feature extraction unit extracts the respiratory frequency component in the 0.1Hz to 0.8Hz frequency band by performing phase demodulation and bandpass filtering on the radar echo signal, and calculates the amplitude of body micro-motion by combining Hilbert transform and energy integration, thereby realizing the quantitative characterization of the intensity of the target human physiological activity. Building upon this foundation, the multimodal mutual verification unit establishes a three-condition joint decision mechanism: when a posture abnormality is triggered, it simultaneously determines whether the respiratory rate is below a first preset threshold of 10 to 18 breaths per minute, whether the amplitude of body micro-movements is below 30% to 50% of the average amplitude of body micro-movements during normal activity, and whether the behavioral semantic vector within 10 to 60 minutes prior to the abnormality matches a preset rule base (confidence not less than 70%, support not less than 5%). This architecture effectively eliminates false alarms caused by non-realistic risk factors such as pet activity, clothing slippage, and equipment noise by triple-associating and verifying posture abnormalities with physiological micro-movement attenuation and behavioral context. While maintaining a true positive rate of fall detection of not less than 95%, it reduces the false alarm rate to below 5%, and completely solves the compliance problem caused by users actively wearing multiple types of beacons in existing technologies.

[0016] 2. This invention achieves synergistic optimization of technical effects at the system deployment level through localized processing of edge computing nodes and a self-organizing network topology generation mechanism. Each edge computing node measures signal attenuation and angle of arrival using channel state information, and automatically generates an indoor spatial topology map using a multi-dimensional scaling analysis algorithm. This eliminates the need for manual surveying and professional configuration, enabling multi-base station coverage optimization and significantly reducing deployment costs and implementation barriers. The environmental semantic perception module collects multi-source time-series signals through sensor groups deployed in the in-home pipelines and appliance nodes. These signals are then transformed into high-level behavioral semantic vectors such as those for toileting, cooking, and getting up at night using a Transformer model. This provides behavioral contextual association for abnormal postures, expanding the decision logic from instantaneous detection at a single moment to semantic verification based on time-series associations. The average amplitude of body movements during normal activity is dynamically updated using the median of seven consecutive days of daytime data, achieving adaptive calibration for individual user differences and behavioral habits. Edge computing nodes only upload anonymized feature data (posture category labels, respiratory rate values, body micro-motion amplitude values, and behavioral semantic category labels) to the community server. The original point cloud data and environmental time-series signals remain within the user's premises, thus achieving data privacy protection while fulfilling monitoring functions. These technologies collectively constitute a fully closed-loop intelligent monitoring system from perception and processing to response. While improving monitoring accuracy and system robustness, this significantly reduces the burden of user cooperation and deployment and maintenance costs, providing reliable technical support for the large-scale application of safety monitoring for elderly people living alone. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0018] Figure 2 This is a schematic diagram of the multimodal mutual verification decision process of the present invention.

[0019] Figure 3 This is a schematic diagram of the attitude reconstruction and micro-motion feature extraction algorithm of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figures 1 to 3 The image shows a safety monitoring system for elderly people living alone.

[0022] I. System Architecture Overview This invention provides a safety monitoring system for elderly people living alone, employing a layered architecture design, including a perception layer, a transmission layer, a processing layer, and an application layer. The perception layer comprises a millimeter-wave radar perception module and an environmental semantic perception module, used to collect human posture micro-motion data and home environment behavior data, respectively; the transmission layer achieves local data processing and uploading through edge computing nodes; the processing layer includes a posture reconstruction unit, a micro-motion feature extraction unit, an environmental semantic perception module, and a multimodal mutual verification unit, completing multi-source data fusion and anomaly detection; the application layer is responsible for alarm push notifications and visualization display.

[0023] II. Implementation Methods of Core Modules 2.1 Millimeter-wave radar sensing module The millimeter-wave radar sensing module is deployed in the monitoring area, and in practice, it can be installed in high-activity areas such as the living room, bedroom, and bathroom. This module uses a frequency-modulated continuous wave radar sensor with a center frequency of 60 GHz, a transmission bandwidth of 4 GHz to 6 GHz, and a modulation period of 100 microseconds. The radar achieves angular resolution through a multi-input multi-output antenna array, with at least 16 virtual channels.

[0024] Data flow description: Input: The echo signal of the radar signal after being reflected by the human body.

[0025] Processing procedure: Perform a range-dimensional fast Fourier transform on the echo signal to obtain the range-Doppler image; A constant false alarm rate (CFAR) detection algorithm is used to extract target scattering points, with the detection threshold set to a false alarm rate of no more than 1%. The azimuth and elevation angles of each scattering point are calculated using the angle of arrival estimation algorithm to obtain the three-dimensional spatial coordinates; Calculate the radial velocity of each scattering point based on the Doppler frequency.

[0026] Output: Each frame contains 3D point cloud data with no fewer than 200 scattering points, at a frame rate of no less than 10 frames per second. Each point cloud data includes the 3D spatial coordinates of that point. and radial velocity .

[0027] 2.2 Attitude Reconstruction Unit The posture reconstruction unit receives 3D point cloud data output by the millimeter-wave radar sensing module, reconstructs the 3D human skeleton and identifies the posture through a deep learning network.

[0028] Data flow description: Input: Multiple consecutive frames of 3D point cloud data, each frame in the following format. The coordinate matrix, N≥200, .

[0029] Processing procedure: Point cloud data is input into a pre-trained PointNet++ network, the network structure of which includes 4 layers of ensemble abstraction layers and 4 layers of feature propagation layers; The network outputs the three-dimensional coordinates of 15 to 25 major joints of the human body, including the head, shoulders, elbows, wrists, hips, knees, and ankles. Input a sequence of joint coordinates from 10 to 30 consecutive frames into a temporal convolutional network to identify pose categories.

[0030] Output: Posture category label, including at least one of standing, sitting, lying down, falling, leaning forward, and falling sideways. When the recognition result is falling, leaning forward, or falling sideways for 5 to 10 consecutive frames, it is judged as an abnormal posture, triggering the subsequent mutual verification process.

[0031] 2.3 Micro-motion Feature Extraction Unit The micro-motion feature extraction unit extracts respiratory frequency and body micro-motion amplitude from radar echo signals to determine whether the human body is in a normal physiological state.

[0032] Data flow description: Input: Slow-time dimension signal of radar echo.

[0033] Processing procedure: Target distance cell selection: Perform a distance-dimensional fast Fourier transform on the echo signal and select the distance cell with the highest energy as the distance cell where the human target is located; Phase demodulation: The slow-time dimension signal of this distance cell is unwrapped in phase to obtain the time-domain signal of the thoracic displacement. ; Respiratory rate extraction: A bandpass filter from 0.1 Hz to 0.8 Hz was used. Filtering is performed to obtain the respiratory signal; peak detection is performed on the respiratory signal, and the time interval between adjacent peaks is calculated. respiratory rate Unit: times / minute; Calculation of body movement amplitude: For The envelope signal is obtained by performing a Hilbert transform. within the preset time window The amplitude of the body's micromotion is obtained by internal calculation of energy integration. ,in Take 5 to 30 seconds.

[0034] Output: Current breathing rate (times / minute) and amplitude of micro-movements . 2.4 Environmental Semantic Awareness Module The environmental semantic perception module collects multi-source environmental time-series signals through sensor groups deployed in the inlet pipeline and home appliance nodes, and converts them into high-level behavioral semantics.

[0035] Sensor deployment solutions: Data flow description: Input: Multi-source environmental timing signals ,in This represents the number of sensors.

[0036] Processing procedure: Data cleaning and alignment: Outlier removal and timestamp alignment of raw sensor data; Time window slicing: Slice the data into 60-minute time windows with a sliding step of 15 minutes; Semantic feature extraction: Multi-source time series data within each time window are input into the trained Transformer model. The model contains a 4-layer encoder, each layer of which contains a multi-head self-attention mechanism and a feedforward network, and outputs a 128-dimensional feature vector. Probability mapping: A linear layer is used to map the probability vectors to behavioral semantics using a softmax function. ,in For the number of semantic categories, Indicates the first The probability value of class semantics, and .

[0037] Output: Behavioral semantic probability vectors, with semantic categories including toileting, cooking, dining, taking medication, sleeping, going out, getting up at night, and sitting for long periods of time.

[0038] 2.5 Multimodal Mutual Verification Unit The multimodal mutual verification unit is the core of this invention, responsible for integrating pose recognition, micro-motion features and behavioral semantics to make a comprehensive decision.

[0039] Data flow description: enter: Abnormal pose categories output by the pose reconstruction unit and the time of occurrence of the anomaly ; Current respiratory rate output by the micro-motion feature extraction unit amplitude of body micro-motion ; The sequence of behavioral semantic vectors output by the environmental semantic perception module for the period from 10 to 60 minutes prior to the occurrence of the anomaly. .

[0040] Judgment Logic: A high-risk alarm signal will be output when the following three conditions are met: Condition 1 (Abnormal respiratory rate): The first preset threshold is set between 10 and 18 breaths per minute, corresponding to the lower limit of the normal resting respiratory rate. When the respiratory rate falls below this threshold, it indicates that the person may be experiencing impaired consciousness or serious physiological abnormalities.

[0041] Condition 2 (Abnormal body movement amplitude): The mean amplitude of body movement during normal activity is defined as the median value of body movement amplitude calculated every 10 minutes from 8:00 to 20:00 each day for 7 consecutive days. The second preset threshold, ranging from 30% to 50%, is used to characterize the degree of decline in human mobility after a fall. When the amplitude of body movement is lower than 30% to 50% of the normal average, it indicates that the human body is in an abnormally still state.

[0042] Condition 3 (Behavioral semantic association matching): To be related to abnormal posture categories The set of semantic categories of preceding behaviors associated with a relationship is determined through association rule mining. A preset time window is set before the anomaly occurs, with a value range of 10 minutes to 60 minutes; The third preset threshold, ranging from 0.6 to 0.8, represents the minimum probability requirement for the behavior semantic vector to match the rule base.

[0043] Output: High-risk alarm signal, including information such as anomaly type, occurrence time, and confidence level.

[0044] 2.6 Construction of the Rule Base for Anomaly Precursor Behaviors The abnormal preceding behavior rule base is used to establish the association between abnormal poses and preceding behaviors. The specific construction method is as follows: Data Acquisition: Collect historical abnormal event data, with each event including an abnormal posture category. The sequence of behavioral semantic vectors within 60 minutes prior to the occurrence of the anomaly, along with manually labeled real risk tags.

[0045] Association rule mining: The Apriori algorithm or FP-Growth algorithm is used for association rule mining. For each anomalous pose category... Define association rules in Given the set of semantic categories for preceding actions, calculate the confidence score for each rule. With support .

[0046] Threshold selection: Thresholds were determined through ROC curve analysis. On the training set, ensuring a true positive rate of at least 95%, the confidence and support thresholds corresponding to the lowest false positive rate were selected. Finally, rules with a confidence level of at least 70% and a support level of at least 5% were added to the rule base.

[0047] Example of a rule: If the abnormal posture is "falling down", the associated preceding behaviors include "getting up at night" or "taking a bath"; If the abnormal posture is "sitting for a long time", the associated preceding behaviors include "using the toilet" or "cooking".

[0048] III. System Deployment Architecture This invention adopts an edge computing architecture, and the specific deployment method is as follows: Edge computing nodes: Deployed in key areas such as living rooms, bedrooms, and bathrooms, each node integrates a millimeter-wave radar sensing module, attitude reconstruction unit, micro-motion feature extraction unit, and some mutual verification logic. The nodes use ARM architecture processors with a computing power of no less than 1 TOPS, supporting local real-time processing.

[0049] Self-organizing network and topology generation: The indoor spatial topology map is automatically generated between edge computing nodes through wireless signal interference. The nodes measure signal attenuation and angle of arrival using channel state information to construct a distance matrix between nodes. A multidimensional scaling analysis algorithm is used to map the distance matrix into two-dimensional spatial coordinates to generate the indoor spatial topology map. Based on this map, the nodes autonomously decide on the master-slave relationship and coverage area.

[0050] Data upload strategy: Edge computing nodes only upload high-risk alarm signals and desensitized feature data to the community server. The desensitized data includes posture category labels, respiratory rate values, body micro-motion amplitude values, and behavioral semantic category labels, but does not include raw point cloud data and raw environmental time-series signals, to ensure user privacy and security.

[0051] IV. Examples Example 1: Nighttime Fall Monitoring Scenario Description: Mr. Zhang, a 78-year-old senior citizen, lives in a two-bedroom apartment. The system has been installed and trained, and the parameters of each module have been personalized based on Mr. Zhang's daily behavior data.

[0052] Time of incident: 2:35 AM Event sequence: Preceding behavior perception: Between 2:20 and 2:30 AM, the environmental semantic perception module detected a continuous water flow signal through a vibration sensor on the household water pipe. Combined with the fluctuation characteristics of the toilet tank water level, the module output a probability value of 0.92 for the "getting up at night" category in the behavioral semantic vector, exceeding the recognition threshold of 0.8. The system records the semantic label and timestamp of this behavior.

[0053] Anomaly detection: At 2:35 AM, the millimeter-wave radar sensing module detected a drastic change in the human point cloud morphology within the restroom area. The attitude reconstruction unit analyzed the continuous point cloud data and output the attitude recognition results: the first 5 frames showed "standing," the keypoint coordinate vectors changed drastically from the 6th to the 10th frames, and the 11th to 15th frames were identified as "falling," with the probability of the "falling" category consistently higher than 0.9. The attitude reconstruction unit determined this to be an abnormal attitude and triggered the mutual verification process.

[0054] Micro-motion feature extraction: The multimodal mutual verification unit synchronously acquires the output data of the micro-motion feature extraction unit: respiratory rate ; Body movement micro-motion amplitude ; The system retrieves the average amplitude of Grandpa Zhang's body movements during normal activity. (Calculated based on daytime data from the past 7 days) Current body movement amplitude The second condition is met.

[0055] Respiratory rate 14 breaths / minute, first preset threshold The system is set to 12 times / minute. Since 14 > 12, the first condition is not met. The system then enters continuous monitoring mode.

[0056] Continuous monitoring and secondary verification: The system continues monitoring after the abnormality is triggered, and acquires the micro-motion data again after 30 seconds: Respiratory rate decreased to =10 times / minute; Body movement micro-motion amplitude 0.28.

[0057] at this time The first condition is met; The second condition is met.

[0058] Behavioral semantic association verification: The system retrieves the behavioral semantic vector sequence within the 30 minutes prior to the anomaly and queries the rule base of preceding behaviors associated with "fall". The rule base lists "getting up at night" as a preceding behavior associated with "fall". Within the 2:20 AM to 2:30 AM time window, the semantic probability value of the "getting up at night" behavior is highest at 0.92, exceeding the third preset threshold. The third condition is met.

[0059] Alarm output: When all three conditions are met, the multimodal mutual verification unit outputs a high-risk alarm signal, which includes the following information: anomaly type "fall", occurrence time "2:35", confidence level "0.96", and location "toilet".

[0060] Response and handling: The edge computing node pushed the alarm signal to the community server, which then notified the grid worker and emergency contact via both telephone voice and app push notifications. The grid worker arrived at the scene at 2:45 AM and found that Mr. Zhang had fallen due to dizziness after getting up at night. He was conscious but unable to get up on his own, so the worker immediately contacted 120 for an ambulance. He was diagnosed with dizziness caused by low blood pressure; timely medical attention prevented further injury.

[0061] This system accurately identified real fall events by verifying the correlation between the semantics of "getting up at night" behavior and the posture of "falling," combined with the micro-motion characteristics of breathing frequency and body movement amplitude, without generating false alarms, and realizing a closed loop of the entire process from perception to handling.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A safety monitoring system for elderly people living alone, characterized in that, include: The millimeter-wave radar sensing module is deployed within the monitoring area to transmit frequency-modulated continuous wave signals and receive echo signals, generating three-dimensional point cloud data of the target human body. The attitude reconstruction unit is connected to the millimeter-wave radar sensing module and is used to input the three-dimensional point cloud data into the attitude recognition network to reconstruct the three-dimensional skeleton structure of the target human body and identify the attitude category. The micro-motion feature extraction unit, connected to the millimeter-wave radar sensing module, is used to extract the respiratory rate and body micro-motion amplitude of the target human body from the echo signal; wherein, the body micro-motion amplitude is obtained by energy integration of the envelope of the chest cavity displacement time-domain signal within a preset time window, and the calculation formula is as follows: In the formula The amplitude of the body's micro-motion. This represents the envelope of the time-domain signal of thoracic cavity displacement. Preset time window; The environmental semantic perception module includes a sensor group deployed at the water, electricity, and gas pipelines and key home appliance nodes to collect multi-source environmental time-series signals and input the multi-source environmental time-series signals into the behavioral semantic model to generate a behavioral semantic vector within a preset time window. The behavioral semantic vector contains the probability value of each preset behavioral semantic category. The multimodal mutual verification unit is connected to the posture reconstruction unit, the micro-motion feature extraction unit, and the environmental semantic perception module, respectively. When the posture recognition result is an abnormal posture, it is used to simultaneously acquire the current breathing rate, body micro-motion amplitude, and the behavioral semantic vector sequence within a preset time window before the abnormality occurred. When the following three conditions are met simultaneously, a high-risk alarm signal is output: First condition: The respiratory rate is lower than a first preset threshold, where the first preset threshold is the lower limit of the normal resting respiratory rate of a human body; The second condition is that the amplitude of the micro-movement is lower than the product of the second preset threshold and the average amplitude of the micro-movement in the normal activity state. The second preset threshold is used to characterize the degree of decline in human mobility after a fall. The third condition is that within a preset time window before the anomaly occurs, there is at least one moment in which the sum of the probabilities of the preceding behavioral semantic categories associated with the anomaly posture category exceeds the third preset threshold.

2. The safety monitoring system for elderly people living alone according to claim 1, characterized in that, The center frequency of the millimeter-wave radar sensing module is 60 GHz, the transmission bandwidth is 4 GHz to 6 GHz, each frame of three-dimensional point cloud data contains no less than 200 scattering points, and the frame rate is no less than 10 frames per second; the three-dimensional point cloud data includes the three-dimensional spatial coordinates and radial velocity of each scattering point.

3. The safety monitoring system for elderly people living alone according to claim 1, characterized in that, The posture recognition network is a PointNet++ network or a graph convolutional network. The input is the three-dimensional point cloud data, and the output is the coordinate vector of each major joint of the human body in three-dimensional space. The posture categories include at least one of the following: standing, sitting, lying down, falling, leaning forward, and falling to the side. When the recognition results of multiple consecutive frames are falling, leaning forward, or falling to the side, it is determined to be an abnormal posture.

4. The safety monitoring system for elderly people living alone according to claim 1, characterized in that, The micro-motion feature extraction unit includes: The phase demodulation module is used to select the distance cell with the highest energy from the echo signal, and perform phase unwrapping on the slow time dimension signal of the distance cell to obtain the time domain signal of the thoracic cavity displacement. A bandpass filter with a passband of 0.1 Hz to 0.8 Hz is used to extract the respiratory frequency component from the time-domain signal of chest displacement. The respiratory frequency is obtained by peak detection of the extracted respiratory signal. The respiratory frequency is defined as the number of breaths per minute. The body motion amplitude calculation module is used to perform Hilbert transform on the time-domain signal of thoracic cavity displacement to obtain an envelope signal, and calculate the energy integral of the envelope signal as the body motion micro-amplitude within a preset time window, wherein the preset time window is 5 seconds to 30 seconds.

5. The safety monitoring system for elderly people living alone according to claim 1, characterized in that, The sensor group of the environmental semantic perception module includes: a vibration sensor or flow sensor deployed on the water pipe at the entrance of the house, a non-intrusive current sensor deployed at the electricity meter at the entrance of the house, a dual-mode gas sensor deployed on the gas pipeline at the entrance of the house, and a vibration sensor or current characteristic sensor deployed at the home appliance node, wherein the home appliance node includes at least one or more of the following: refrigerator, microwave oven, washing machine, and water heater. The behavioral semantic model is a Transformer model or a Long Short-Term Memory network. The input is a data matrix of the multi-source environmental time-series signal within a time window, which is 30 minutes to 120 minutes. The output is a behavioral semantic probability vector. The behavioral semantic categories include at least one of the following: using the toilet, cooking, eating, taking medication, sleeping, going out, and getting up at night.

6. The safety monitoring system for elderly people living alone according to claim 1, characterized in that, The mean amplitude of body movement during normal activity is defined as the median value of body movement amplitude calculated every 10 minutes from 8:00 to 20:00 each day for 7 consecutive days. The first preset threshold is 10 times / minute to 18 times / minute; The second preset threshold is 30% to 50%; The third preset threshold is 0.6 to 0.8; The preset time window before the anomaly occurs is 10 to 60 minutes. The semantic category of the preceding behavior associated with the abnormal posture category is determined by an association rule mining algorithm. Association rules with a confidence of not less than 70% and a support of not less than 5% are selected to construct a rule base. The confidence threshold and support threshold are determined by ROC curve analysis. The threshold corresponding to the lowest false positive rate is selected on the training set under the condition that the true positive rate is not less than 95%.

7. The safety monitoring system for elderly people living alone according to claim 1, characterized in that, It also includes edge computing nodes, deployed within the monitoring area, used to integrate the millimeter-wave radar sensing module, the attitude reconstruction unit, the micro-motion feature extraction unit, the environmental semantic perception module, and the multimodal mutual verification unit, to complete attitude recognition, micro-motion feature extraction, behavioral semantic generation, and mutual verification decision locally; the edge computing nodes only upload high-risk alarm signals and desensitized feature data to the cloud or community server, the desensitized feature data includes attitude category labels, respiratory rate values, body micro-motion amplitude values, and behavioral semantic category labels, but does not include original point cloud data and original environmental time-series signals.

8. A method for monitoring the safety of elderly people living alone, characterized in that, Includes the following steps: S1. The millimeter-wave radar sensing module transmits frequency-modulated continuous wave signals and receives echo signals to generate three-dimensional point cloud data of the target human body. S2. The three-dimensional point cloud data is input into the attitude recognition network through the attitude reconstruction unit to reconstruct the three-dimensional skeleton structure and identify the attitude category; S3. The respiratory rate and body movement amplitude of the target human body are extracted from the echo signal by the micro-motion feature extraction unit. The body movement amplitude is obtained by energy integration of the envelope of the chest cavity displacement time-domain signal within a preset time window. The calculation formula is as follows: ; S4. Collect multi-source environmental time-series signals through the environmental semantic perception module and generate behavioral semantic vectors within a preset time window; S5. When the identified posture category is an abnormal posture, the respiratory rate, body movement amplitude, and behavioral semantic vector sequence within the preset time window before the abnormality occurs are obtained simultaneously. If the respiratory rate is lower than the first preset threshold, the body movement amplitude is lower than the product of the second preset threshold and the average body movement amplitude in the normal activity state, and there is at least one moment within the preset time window before the abnormality occurs where the sum of probabilities of the preceding behavioral semantic categories associated with the abnormal posture category exceeds the third preset threshold, then a high-risk alarm signal is output.

9. A method for monitoring the safety of elderly people living alone according to claim 8, characterized in that, In step S3, extracting the respiratory rate specifically includes: Perform a distance-dimensional fast Fourier transform on the echo signal and select the distance cell with the highest energy as the distance cell where the human target is located. Phase unwrapping is performed on the slow-time dimension signal of the selected distance cell to obtain the time-domain signal of the thoracic cavity displacement; The time-domain signal of the thoracic cavity displacement was filtered using a bandpass filter ranging from 0.1 Hz to 0.8 Hz to obtain the respiratory signal; Peak detection is performed on the respiratory signal, and the time interval between adjacent peaks is calculated. The respiratory rate is defined as the number of breaths per minute, which is obtained by taking the median and then taking the reciprocal of the time interval.

10. A method for monitoring the safety of elderly people living alone according to claim 8, characterized in that, In step S4, generating the behavior semantic vector specifically includes: Data cleaning and time alignment are performed on the multi-source environmental time-series signals collected by the sensor array; The aligned timing signal is sliced ​​according to a preset time window, with each time window being 60 minutes long and a sliding step of 15 minutes. Multi-source time-series data within each time window are input into a trained Transformer model, which outputs a feature vector. This feature vector is then mapped to a behavioral semantic probability vector via a linear layer and a softmax function. The behavioral semantic probability vector contains probability values ​​for each preset behavioral semantic category, which includes at least one of the following: using the toilet, cooking, eating, taking medication, sleeping, going out, and getting up at night.

Images