A long-term care home-based old-age service monitoring method and system based on millimeter wave radar

Abstract

The application discloses a kind of long care risk home-based old-age service monitoring method and system based on millimeter wave radar, through continuously tracking the position information of the person being cared and the caregiver, relative distance and multiple parameters such as body activity, whether the nursing service is truly occurred and its quality is comprehensively judged.The application uses millimeter wave radar as core sensor, through signal processing, target tracking and behavior recognition algorithm, realizes the non-contact, full privacy protection monitoring to the nursing service process.The algorithm core includes multi-target trajectory tracking model, relative motion relationship analysis model and body activity quantification model, accurately identifies the nursing service state by multidimensional data fusion and threshold judgment.The application solves the problems existing in the service process of long care risk, such as difficult to verify the service authenticity, difficult to quantify the service quality, etc., and provides an effective supervision means for insurance handling departments.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, and in particular to a method and system for monitoring long-term care home-based elderly care services based on millimeter-wave radar. Background Technology

[0002] With the accelerating aging of my country's population, the care needs of disabled and semi-disabled elderly people continue to rise. Long-term care insurance (hereinafter referred to as CLC), as an important social security system to guarantee the basic living care and medical care needs of disabled people, has become a key link in improving the elderly care service system. Home-based CLC services, with their advantages of aligning with the elderly's living habits, reducing care costs, and preserving the family atmosphere, have become one of the mainstream service models. However, the service process often takes place in the private setting of the home, lacking effective offline supervision throughout. Insurance agencies find it difficult to accurately verify whether caregivers arrive on time, provide services according to standards, and complete the required quantities as agreed. The authenticity and quality of services have become the core pain point in the implementation of the CLC system, urgently requiring targeted technical solutions.

[0003] While various monitoring methods have been attempted for long-term care insurance-based home-based elderly care services, they all suffer from significant technical flaws and application limitations, failing to meet actual regulatory needs. Firstly, while video surveillance can visually capture the service process, it directly collects sensitive images of the elderly and caregivers' faces and bodies, severely infringing on personal privacy. It is completely unsuitable for use in core private spaces such as bedrooms and bathrooms, and is prone to causing psychological resistance among the elderly, making widespread adoption difficult. Secondly, wearable device monitoring requires the elderly or caregivers to actively wear the relevant terminals, leading to inconvenience, poor daily comfort, and easy loss or forgetting. For disabled elderly with cognitive impairments or severely limited physical movement, the device's suitability and practicality are significantly reduced, and it can only collect the movement data of a single wearer, failing to reflect the service interaction between the two parties. Thirdly, simple check-in records only record the caregiver's arrival and departure times, unable to effectively verify the actual service content, duration, and standardization of service actions. This easily leads to service fraud such as "empty card swiping" and "proxy check-ins," resulting in the loss of long-term care insurance funds and failing to protect the elderly's actual care rights.

[0004] Millimeter-wave radar technology, with its advantages of non-contact measurement, complete privacy protection without generating visual images, unaffected by light or darkness, and ability to penetrate common household obstructions such as clothing and furniture, is gradually becoming an ideal technology choice for monitoring human status in home environments. This technology can accurately acquire information on the distance, speed, angle, and vital signs such as limb movements, breathing, and heartbeat of the monitored target by analyzing radar echo signals. It achieves unobtrusive and continuous monitoring of a person's location and movement status, technically solving the challenges of privacy protection and scenario adaptation in home-based elderly care service monitoring. However, directly applying millimeter-wave radar technology to the monitoring of home-based elderly care services under long-term care insurance still faces many unresolved technical challenges: In complex home environments, static objects such as furniture and walls can cause interference, resulting in insufficient accuracy in multi-target identification and tracking, making it difficult to quickly and accurately distinguish between elderly individuals and caregivers; specific actions in nursing services, such as turning over, massage, and blood glucose monitoring, have small differences in micro-motion characteristics and complex movement patterns, limiting the accuracy of existing behavior recognition algorithms and making it impossible to accurately quantify the caregiver's limb activity; currently, there is a lack of dedicated evaluation models for the 36 aspects of daily living care and medical care services under long-term care insurance, making it difficult to extract effective features from the raw radar monitoring data to achieve service authenticity determination and quantitative evaluation of service quality. Therefore, developing a millimeter-wave radar monitoring system and method specifically adapted to the home-based elderly care service scenario under long-term care insurance, and overcoming the above technical bottlenecks, has become an urgent technical problem to be solved in this field, possessing significant practical and social value. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring long-term care insurance home-based elderly care services based on millimeter-wave radar, which solves the problem of balancing privacy protection and service quality supervision in the existing technology, and realizes effective monitoring and quality assessment of the nursing service process.

[0006] This invention provides a method for monitoring long-term care insurance home-based elderly care services based on millimeter-wave radar, comprising: Acquire raw radar echo data, extract point cloud data of the human target, and obtain the centroid position of the human target; The real-time state vector of the target is obtained based on the centroid position of the human target, and the two targets, the elderly and the caregiver, are distinguished. Based on the target's real-time state vector, determine whether there is a service-oriented interaction between the elderly and caregivers; The system extracts caregiver limb movement characteristics from raw radar echo data, quantifies their limb activity index, and identifies specific service actions of caregivers; and Establish a comprehensive evaluation model to determine whether nursing services actually occur, quantify and evaluate the service quality level, and detect abnormal service behaviors.

[0007] In one embodiment of the present invention, the steps of acquiring raw radar echo data, extracting point cloud data of the human target, and obtaining the centroid position of the human target include: The FMCW millimeter-wave radar transmits a continuous wave signal whose frequency varies linearly with time, and receives the echo signal reflected from the target. By calculating the frequency difference between the transmitted and received signals, it outputs a range-Doppler heat map and point cloud data. The static clutter filtering algorithm eliminates the reflection interference of static objects, and the constant false alarm rate detection eliminates false target points while retaining the echo signal of dynamic human targets. An adaptive clustering algorithm is used to cluster the filtered point cloud data to identify the centroid position of the human target and eliminate the influence of scattered noise points.

[0008] In one embodiment of the present invention, the target distance is calculated using the following formula: ; Where d is the straight-line distance between the target and the radar; C is the speed of light; Tc is the radar frequency modulation period, which is an inherent parameter of the FMCW radar hardware; f b The frequency difference between the transmitted signal and the received echo is obtained by the radar in real time. B represents the radar frequency modulation bandwidth, an inherent parameter of the FMCW radar hardware.

[0009] In one embodiment of the present invention, obtaining the real-time state vector of the target based on the centroid position of the human target and distinguishing between the two targets, the elderly and the caregiver, includes: Based on the centroid of the human target, initialize a motion state vector for each target: ; Where X is the two-dimensional motion state vector of the target; x, y are the coordinates of the target in the two-dimensional plane, determined by the centroid of the radar point cloud; v x ,v y The velocity components of the target in the x and y directions are calculated from the rate of change of coordinates at adjacent time points; a x ,a y The acceleration components of the target in the x and y directions are calculated from the rate of change of velocity between adjacent time points; The MODT algorithm is used to estimate the real-time motion state of each target, establish a preliminary model of the target motion trajectory, and realize the basic trajectory tracking of multiple targets. The initial trajectory is optimized based on the CNN-GRU correction algorithm, and the real-time update of the target's real-time state vector is completed by combining Kalman filtering. By continuously analyzing trajectory features, we can distinguish between elderly people and caregivers.

[0010] In one embodiment of the present invention, determining whether there is service-oriented interaction between the elderly person and the caregiver based on the target real-time state vector includes: Based on the target's real-time state vector, the relative motion parameters are calculated, including: The straight-line distance between the two people in a two-dimensional plane; The difference in velocity between the two individuals in the x and y directions; The angle between the directions of movement of the two people; If two people remain within 2 meters of each other, in the same direction, or relatively stationary, it is determined that there is potential service interaction. The dynamic time warping algorithm is used to match the motion trajectory sequences of the two people, quantify the trajectory similarity, and further verify the authenticity of the service interaction.

[0011] In one embodiment of the present invention, the step of extracting the caregiver's limb movement characteristics from the raw radar echo data, quantifying its limb activity index, and identifying the caregiver's specific service actions includes: Micro-Doppler features of the caregiver's limbs are extracted from the original radar echo signal. The time-domain signal is converted into a time-frequency distribution through short-time Fourier transform to obtain the frequency and amplitude features of limb movement. Based on the frequency and amplitude characteristics of limb movements, and combined with a frequency weighting function, the limb activity index of caregivers is calculated. Based on the micro-Doppler features, the point cloud data of the caregiver's limbs is clustered, service-specific feature vectors are extracted, and matched with a pre-trained standard service action library to identify the caregiver's specific service actions.

[0012] In one embodiment of the present invention, the limb activity index is: ; LAI stands for Limb Activity Index, which is dimensionless and ranges from 0 to 1. A higher value indicates greater limb activity. T represents the statistical time window, which is set according to the service type; STFT(t,f) is the short-time Fourier transform value of the radar echo at time t and frequency f, reflecting the time-frequency characteristics of limb micro-movements; w(f) is a frequency weighting function that assigns weights only to the typical frequency range of limb movements, while other frequencies are weighted at 0. Micro-Doppler features are extracted using the following formula: ; Where S(t,f) is the time-frequency distribution value, and f is the square of the magnitude of the STFT; x(τ) is the original echo signal received by the radar; w(τ−t) is a window function used for frame processing of time-domain signals; j is the imaginary unit; τ is the integration variable, representing time; The service-specific feature vector is: ; Among them, F service This is a service-specific feature vector used to distinguish different nursing service actions; f centroid The centroid of the spectrum reflects the dominant frequency of limb movement; f bandwidth The spectral bandwidth reflects the frequency distribution range of limb movements; f entropy Spectral entropy reflects the uniformity of the frequency distribution of limb movements; f peak The peak value represents the strongest frequency component of limb movement.

[0013] In one embodiment of the present invention, the step of establishing a comprehensive evaluation model to determine whether nursing services actually occurred, quantifying and evaluating the service quality level, and detecting abnormal service behaviors includes: A multi-dimensional feature baseline database was established for 36 long-term care insurance services. Actual monitoring data was initially matched with baseline parameters to determine the types of services to be evaluated. Analyze the caregiver's action sequence in real time to determine the timing of service switching and ensure that the assessment is consistent with the actual service process; By introducing service item specificity weights and time consistency tests, a comprehensive scoring function is constructed to quantify service quality into a comprehensive score, with a higher score indicating better service quality. If all indicators meet the minimum valid threshold and there are no obvious anomalies, the service is judged to have actually occurred; otherwise, it is judged to have been a fake service. Based on the comprehensive score, the service quality is divided into four levels: excellent, good, satisfactory, and unsatisfactory.

[0014] In one embodiment of the present invention, a service status sequence is calculated in real time using the following formula to identify service switching: ; in Let be the forward probability of being in service state j at time t; λ is the global parameter; N represents the total number of service states; Let be the forward probability of being in service state i at time t-1; The critical moments of service switching are identified using a mutation point detection algorithm, and the criteria for judgment are as follows: ; in Let be the rate of change of position at time t. For motion pattern feature vectors, The preset switching threshold; The comprehensive scoring function is: ; in The service is rated as a whole, dimensionless value ranging from 0 to 100, with higher quality being better. α, β, γ, δ, ϵ are the weight coefficients of each indicator, satisfying α+β+γ+δ+ϵ=1, and are dynamically adjusted according to the service type; Distance scoring is calculated based on the match between the actual average interaction distance and the baseline distance; The duration score is calculated based on the degree of matching between the actual service duration and the baseline duration. Activity level is scored based on the degree of match between the actual physical activity index and the baseline physical activity index. The service is scored based on its continuity over time, assessing whether the service process was interrupted; no interruption results in a perfect score. The service switching naturalness score is determined by evaluating whether the order of service switching is logical; if it is, it is a perfect score.

[0015] This invention also provides a long-term care insurance home-based elderly care service monitoring system based on millimeter-wave radar, comprising: The millimeter-wave radar acquisition module is configured to transmit and receive echo signals, acquire point cloud data and range-Doppler heatmaps of human targets, and provide raw radar data for all analysis stages. The signal preprocessing module is configured to reduce noise and extract effective targets from the raw radar data. It eliminates reflection interference from the static environment through a static clutter filtering algorithm, and combines an adaptive clustering algorithm and constant false alarm rate detection to remove false noise points, identify the centroid position of human targets, and output effective human target point cloud data. The multi-target tracking module is configured to distinguish and continuously track the trajectories of elderly people and caregivers in a scene. The behavior recognition module is configured to extract limb micro-Doppler features from radar echoes, combine them with short-time Fourier transform to analyze the frequency and amplitude features of limb movements, identify the caregiver's specific nursing service actions, and calculate a quantified limb activity index to determine the caregiver's limb activity level. The service assessment module is configured to determine whether nursing services actually occur, and to quantify and grade the service quality.

[0016] The present invention has the following beneficial effects: (1) The non-imaging feature of millimeter-wave radar avoids the risk of privacy leakage in video surveillance. Deep data processing ensures that no image information that can identify an individual is generated, making it particularly suitable for monitoring private spaces such as bedrooms and bathrooms.

[0017] (2) Through multi-target tracking and behavior recognition algorithms, it is possible to accurately distinguish between the elderly and caregivers, identify specific service actions, and reduce misjudgments and omissions.

[0018] (3) By adopting static clutter filtering and multipath interference elimination technology, the false detection problem caused by static objects such as furniture and decorations and wall reflections in the home environment can be effectively overcome.

[0019] (4) Through multi-dimensional data analysis, service quality is quantified into comparable numerical indicators, providing objective and unified evaluation standards for long-term care insurance management departments and reducing subjective judgment bias. Attached Figure Description

[0020] Figure 1 A flowchart of a long-term care insurance home-based elderly care service monitoring method based on millimeter-wave radar is shown in one embodiment of the present invention. Detailed Implementation

[0021] In the following description, the invention is described with reference to various embodiments. However, those skilled in the art will recognize that the embodiments may be practiced without one or more specific details or with other alternatives and / or additional methods, materials, or components. In other instances, well-known structures, materials, or operations are not shown or described in detail so as not to obscure the inventive points of the invention. Similarly, for illustrative purposes, specific quantities, materials, and configurations are set forth to provide a comprehensive understanding of embodiments of the invention. However, the invention is not limited to these specific details.

[0022] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.

[0023] In this specification, references to "an embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention. The phrase "in one embodiment" appearing throughout this specification does not necessarily refer to the same embodiment in all instances.

[0024] Furthermore, the numbering of the steps in the methods of the present invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps may be executed in different orders.

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0026] Figure 1 A flowchart of a long-term care insurance home-based elderly care service monitoring method based on millimeter-wave radar is shown in one embodiment of the present invention.

[0027] like Figure 1 As shown in this embodiment, the long-term care insurance home-based elderly care service monitoring method based on millimeter-wave radar includes: S100, Radar Data Acquisition and Preprocessing: Acquire raw radar echo data, eliminate environmental clutter interference, and extract high-precision effective point cloud data of human targets, laying the foundation for subsequent analysis, including: The FMCW millimeter-wave radar transmits a continuous wave signal whose frequency varies linearly with time, and receives the echo signal reflected by the target (elderly person, caregiver). By calculating the frequency difference between the transmitted and received signals, it outputs a range-Doppler heat map (RDM) and point cloud data as the raw data for subsequent processing. The static clutter filtering algorithm eliminates the reflection interference from static objects such as furniture and walls, and the constant false alarm rate (CFAR) detection eliminates false target points while retaining the echo signal of dynamic human targets. An adaptive clustering algorithm is used to cluster the filtered point cloud data, accurately identifying the centroid positions of human targets such as the elderly and caregivers, and eliminating the influence of scattered noise points.

[0028] The core formula is as follows: The target distance is calculated using the following formula: ; Where d is the straight-line distance between the target and the radar; C is the speed of light; Tc is the radar frequency modulation period, which is an inherent parameter of the FMCW radar hardware; f b The frequency difference between the transmitted signal and the received echo is obtained by the radar in real time. B represents the radar frequency modulation bandwidth, an inherent parameter of the FMCW radar hardware.

[0029] The S200 multi-target recognition and tracking system accurately distinguishes between elderly individuals and caregivers in complex home environments, continuously tracking their movement trajectories to ensure a position detection error RMSE ≤ 0.15 meters. The steps are as follows: Based on the centroid of the human target extracted in step one, a motion state vector is initialized for each target, including core parameters such as position, velocity, and acceleration. The MODT algorithm is used to estimate the real-time motion state of each target, establish a preliminary model of the target motion trajectory, and realize the basic trajectory tracking of multiple targets. To address the issue of sparse point clouds in millimeter-wave radar, a CNN-GRU correction (CGC) algorithm is introduced to optimize the initial trajectory. Combined with Kalman filtering, the target state is updated in real time, thereby improving tracking accuracy. By continuously analyzing trajectory features (such as movement speed and range of motion), we can accurately distinguish between elderly people (who usually have a small range of motion and slow speed) and caregivers (who usually have a large range of motion and fast speed), thus preparing for subsequent relative motion analysis.

[0030] The multi-target tracking state vector is updated using the Kalman filter algorithm, and the multi-target tracking state vector is: ; Where X is the two-dimensional motion state vector of the target; x, y are the coordinates of the target in the two-dimensional plane, determined by the centroid of the radar point cloud; v x ,v y The velocity components of the target in the x and y directions are calculated from the rate of change of coordinates at adjacent time points; a x ,a y The acceleration components of the target in the x and y directions are calculated from the rate of change of velocity at adjacent time points.

[0031] S300, Relative Motion Analysis, analyzes the spatial position, movement speed, and direction correlation between the elderly and caregivers to determine whether there is a service-oriented interaction (such as close care provided by the caregiver to the elderly). The specific process is as follows: Basic relative parameter calculation: Based on the real-time state vectors of the two individuals obtained in step two, calculate the core relative motion parameters: Euclidean distance: the straight-line distance between two people in a two-dimensional plane; Relative velocity: The difference in velocity between two people in the x and y directions; Angle of movement direction: The angle between the two people's movement directions (same direction / opposite direction / perpendicular).

[0032] Interaction pattern determination: The typical interaction characteristics of nursing services are close distance (<2 meters), same direction or relatively stationary. If two people continuously meet these characteristics, it is determined that there is potential service interaction.

[0033] The Dynamic Time Warping (DTW) algorithm is used to match the motion trajectory sequences of the two individuals, quantify the trajectory similarity, and further verify the authenticity of the service interaction (the trajectories of the two individuals are usually highly correlated during the service process).

[0034] The reference standards include: The effective service interaction distance threshold is ≤3 meters (minimum effective standard), and the core service interaction distance is ≤2 meters; The smaller the DTW distance, the higher the trajectory similarity, and the stronger the authenticity of the service interaction.

[0035] S400, Limb Activity Measurement and Behavior Recognition, extracts the limb movement characteristics of caregivers, quantifies their Limb Activity Index (LAI), and identifies specific service actions of caregivers (such as turning over, massage, blood glucose monitoring, etc.). The specific process is as follows: Micro-Doppler features of the caregiver's limbs (changes in echo frequency caused by limb micro-movements) are extracted from the radar echo signal. The time-domain signal is converted into a time-frequency distribution through short-time Fourier transform (STFT) to obtain the frequency and amplitude features of limb movement.

[0036] Based on time-frequency distribution data and combined with a frequency weighting function, the caregiver's LAI is calculated to quantify the activity level of limb movement.

[0037] Micro-Doppler features are input into a CNN-LSTM deep learning network, and the improved DBSCAN algorithm is used to cluster the point cloud of the caregiver's limbs to extract service-specific feature vectors. These vectors are then matched with a pre-trained standard service action library (such as warm water sponging and pressure ulcer care) to accurately identify the caregiver's specific service actions.

[0038] In this step, the Physical Activity Index (LAI) is: ; LAI stands for Limb Activity Index, which is dimensionless and ranges from 0 to 1. A higher value indicates greater limb activity. T represents the statistical time window, which is set according to the service type; STFT(t,f) is the short-time Fourier transform value of the radar echo at time t and frequency f, reflecting the time-frequency characteristics of limb micro-movements; w(f) is a frequency weighting function that assigns weights only to the typical frequency range of limb movements, while other frequencies are weighted at 0. Micro-Doppler features are extracted using the following formula: ; Where S(t,f) is the time-frequency distribution value, and f is the square of the magnitude of the STFT; x(τ) is the original echo signal received by the radar; w(τ−t) is a window function used for frame processing of time-domain signals; j is the imaginary unit; τ is the integration variable, representing time; The service-specific feature vector is: ; Among them, F service This is a service-specific feature vector used to distinguish different nursing service actions; f centroid The centroid of the spectrum reflects the dominant frequency of limb movement; f bandwidth The spectral bandwidth reflects the frequency distribution range of limb movements; f entropy Spectral entropy reflects the uniformity of the frequency distribution of limb movements; f peak The peak value represents the strongest frequency component of limb movement.

[0039] S500 service compliance assessment integrates multi-dimensional data such as distance, duration, activity level, action type, and service sequence to establish a comprehensive assessment model. This model determines whether nursing services actually occur, quantifies the service quality level, and detects abnormal service behaviors. The specific process is as follows: This step is the core output of the entire method, and it is divided into four sub-steps: baseline parameter matching, service sequence identification, multi-dimensional scoring, and anomaly detection. Each step is equipped with its own specific algorithm and formula. S510, Baseline Parameter Matching for Tax Collection: A multidimensional feature baseline database was established for the 36 long-term care insurance services (20 daily living care services + 16 medical care services). The actual monitoring data was initially matched with the baseline parameters (expected duration, LAI baseline, interaction distance, point cloud motion characteristics) to determine the service types to be evaluated.

[0040] Examples of baseline parameters are shown in the table below: S520, HMM-based service sequence identification: By using a Hidden Markov Model (HMM) to analyze caregiver action sequences in real time, the timing of service switching can be determined to ensure that the assessment is consistent with the actual service process.

[0041] State sequence: , representing the service status at time T (e.g., "assisting with eating", "blood glucose monitoring"); Observation sequence: , representing radar monitoring feature data (LAI, range, point cloud features, etc.) at time T.

[0042] State transition probability matrix: ,in , representing the probability of switching from service i to service j (based on service logic settings); Observation probability matrix: ,in The radar observation data is correlated with the service status through calculation using a Gaussian mixture model (GMM).

[0043] By using a forward algorithm to calculate the most likely sequence of service states in real time, service switching can be accurately identified.

[0044] The formula for the forward algorithm is: ; in Let be the forward probability of being in service state j at time t; λ is the global parameter; N represents the total number of service states; Let be the forward probability of being in service state i at time t-1; The critical moments of service switching are identified using a mutation point detection algorithm, and the criteria for judgment are as follows: ; in Let be the rate of change of position at time t. For motion pattern feature vectors, This is the preset switching threshold.

[0045] S530, a multi-dimensional comprehensive scoring system, introduces service item specific weights and time consistency checks to construct a comprehensive scoring function that quantifies service quality into specific values. The higher the score, the better the service quality.

[0046] The comprehensive scoring function is: ; in The service is rated as a whole, dimensionless value ranging from 0 to 100, with higher quality being better. α, β, γ, δ, ϵ are the weight coefficients of each indicator, satisfying α+β+γ+δ+ϵ=1, and are dynamically adjusted according to the service type; Distance scoring is calculated based on the match between the actual average interaction distance and the baseline distance; The duration score is calculated based on the degree of matching between the actual service duration and the baseline duration. Activity level is scored based on the degree of match between the actual physical activity index and the baseline physical activity index. The service is scored based on its continuity over time, assessing whether the service process was interrupted; no interruption results in a perfect score. The service switching naturalness score is determined by evaluating whether the order of service switching is logical; if it is, it is a perfect score.

[0047] Compare the actual service with the baseline model using Dynamic Time Warping (DTW) distance: ; Where Q represents the time characteristic sequence of the actual service, and C represents the baseline pattern.

[0048] Duration compliance score: ; Where T actual为 Actual service duration (unit: minutes); T baseline Baseline expected duration for the service (in minutes); T min The minimum valid duration (in minutes) for a service item, and the lower limit of the baseline duration; This is an indicator function. If the condition in parentheses is true, it takes the value 1; otherwise, it takes the value 0 (if the actual duration is less than the minimum requirement, the duration score is 0). It is a natural exponential function that realizes a non-linear mapping between time deviation and score.

[0049] S540, anomaly service detection and adaptive learning, defines four types of service anomaly judgment rules, and detects behaviors such as fake service and perfunctory service in real time: Duration abnormality: (The actual duration was less than 60% of the baseline). Abnormal activity level: or ; Both excessively low and excessively high activity levels are considered abnormal. Distance anomaly: The actual average interaction distance exceeds the baseline range by 2 standard deviations; Switching error: The service item switching frequency is too high / too low, or the switching order is illogical.

[0050] The system continuously collects real service data, optimizes baseline parameters, and adapts to the service styles of different caregivers. The formula is: ; For updated baseline parameters (such as LAI baseline, duration baseline); λ is the forgetting factor (0 < λ < 1), which controls the weight of historical data (usually taken as 0.7-0.9). Baseline parameters accumulated over time; These are the new baseline parameters obtained from current monitoring.

[0051] S550, Final Evaluation Result Output: If the three core indicators of distance, duration, and LAI all meet the minimum effective threshold (e.g., distance ≤ 3 meters, duration ≥ 15 minutes, LAI ≥ 50% of baseline) and there are no obvious abnormalities, the service is judged to have actually occurred; otherwise, it is judged to be a fake service / invalid service.

[0052] Based on the Sservice comprehensive score, service quality is divided into four levels: excellent (90-100), good (80-89), satisfactory (60-79), and unsatisfactory (<60), providing a regulatory basis for long-term care insurance management departments.

[0053] In another embodiment of the present invention, a long-term care insurance home-based elderly care service monitoring system based on millimeter-wave radar is also provided, comprising: The millimeter-wave radar acquisition module is the core of the system's data source. It adopts a 76-81GHz frequency modulated continuous wave (FMCW) radar, which has the ability to measure distance, velocity, and angle. It can penetrate common furniture materials, transmit and receive echo signals, and acquire high-precision point cloud data and range-Doppler heatmaps (RDM) of human targets, providing raw radar data for all subsequent analysis stages.

[0054] The signal preprocessing module performs noise reduction and effective target extraction on the raw radar data. It eliminates reflection interference from static environments such as furniture and walls through a static clutter filtering algorithm. Combined with an adaptive clustering algorithm and constant false alarm rate (CFAR) detection, it removes false noise points and accurately identifies the centroid positions of human targets such as the elderly and caregivers. It outputs clean and effective human target point cloud data to ensure the accuracy of subsequent module analysis.

[0055] The multi-target tracking module accurately distinguishes and continuously tracks the trajectories of elderly people and caregivers in a scene. It adopts the Multi-Target Discrete Tracking (MODT) algorithm combined with the CNN-GRU correction (CGC) algorithm to overcome the sparse point cloud problem of millimeter-wave radar, estimate and update the target motion state in real time, accurately track the position changes of both, and the position detection error RMSE does not exceed 0.15 meters. At the same time, it effectively distinguishes between elderly people and caregivers by the difference in motion characteristics.

[0056] The behavior recognition module focuses on the analysis and quantification of caregivers' limb movements. It extracts micro-Doppler features of the limbs from radar echoes, combines short-time Fourier transform (STFT) to analyze the frequency and amplitude features of limb movements, and uses a CNN-LSTM deep learning model and an improved DBSCAN algorithm to identify caregivers' specific nursing service actions (such as turning over, massage, blood glucose monitoring, etc.) and calculates the quantified limb activity index (LAI) to determine the caregiver's limb activity level.

[0057] The service evaluation module, the core decision-making output module of the system, establishes a multi-dimensional service compliance evaluation model. It comprehensively analyzes key indicators such as the relative distance between the elderly and caregivers, relative movement patterns, caregiver's physical activity level, service action type, and service duration output by the preceding modules. Combined with the exclusive baseline parameters of 36 long-term care insurance services, it determines whether the nursing service actually occurred. At the same time, through multi-threshold judgment and Hidden Markov Model (HMM) service sequence identification, it quantifies and scores the service quality and classifies it into levels. It can also detect abnormal service behaviors and continuously optimize the evaluation baseline parameters through adaptive learning to adapt to the service styles of different caregivers.

[0058] Although various embodiments of the invention have been described above, it should be understood that they are presented by way of example only and not as limitations. It will be apparent to those skilled in the art that various combinations, modifications, and alterations can be made without departing from the spirit and scope of the invention. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely by the appended claims and their equivalents.

Claims

1. A method for monitoring long-term care home-based elderly care services based on millimeter-wave radar, characterized in that, include: Acquire raw radar echo data, extract point cloud data of the human target, and obtain the centroid position of the human target; The real-time state vector of the target is obtained based on the centroid position of the human target, and the two targets, the elderly and the caregiver, are distinguished. Based on the target's real-time state vector, determine whether there is a service-oriented interaction between the elderly and caregivers; The limb movement characteristics of caregivers are extracted from raw radar echo data, their limb activity index is quantified, and their specific service actions are identified. as well as Establish a comprehensive evaluation model to determine whether nursing services actually occur, quantify and evaluate the service quality level, and detect abnormal service behaviors.

2. The method according to claim 1, characterized in that, The steps of acquiring raw radar echo data, extracting point cloud data of the human target, and obtaining the centroid position of the human target include: The FMCW millimeter-wave radar transmits a continuous wave signal whose frequency varies linearly with time, and receives the echo signal reflected from the target. By calculating the frequency difference between the transmitted and received signals, it outputs a range-Doppler heat map and point cloud data. The static clutter filtering algorithm eliminates the reflection interference of static objects, and the constant false alarm rate detection eliminates false target points while retaining the echo signal of dynamic human targets. An adaptive clustering algorithm is used to cluster the filtered point cloud data to identify the centroid position of the human target and eliminate the influence of scattered noise points.

3. The method according to claim 2, characterized in that, The target distance is calculated using the following formula: ； Where d is the straight-line distance between the target and the radar; C is the speed of light; Tc is the radar frequency modulation period, which is an inherent parameter of the FMCW radar hardware; f b The frequency difference between the transmitted signal and the received echo is obtained by the radar in real time. B represents the radar frequency modulation bandwidth, an inherent parameter of the FMCW radar hardware.

4. The method according to claim 1, characterized in that, The process of obtaining the real-time state vector of the target based on the centroid position of the human target, and distinguishing between the elderly and caregiver targets, includes: Based on the centroid of the human target, initialize a motion state vector for each target: ； Where X is the two-dimensional motion state vector of the target; x, y are the coordinates of the target in the two-dimensional plane, determined by the centroid of the radar point cloud; v x ,v y The velocity components of the target in the x and y directions are calculated from the rate of change of coordinates at adjacent time points; a x ,a y The acceleration components of the target in the x and y directions are calculated from the rate of change of velocity between adjacent time points; The MODT algorithm is used to estimate the real-time motion state of each target, establish a preliminary model of the target motion trajectory, and realize the basic trajectory tracking of multiple targets. The initial trajectory is optimized based on the CNN-GRU correction algorithm, and the real-time update of the target's real-time state vector is completed by combining Kalman filtering. By continuously analyzing trajectory features, we can distinguish between elderly people and caregivers.

5. The method according to claim 1, characterized in that, The determination of whether there is service-oriented interaction between the elderly and caregivers based on the target's real-time state vector includes: Based on the target's real-time state vector, the relative motion parameters are calculated, including: The straight-line distance between the two people in a two-dimensional plane; The difference in velocity between the two individuals in the x and y directions; The angle between the directions of movement of the two people; If two people remain within 2 meters of each other, in the same direction, or relatively stationary, it is determined that there is potential service interaction. The dynamic time warping algorithm is used to match the motion trajectory sequences of the two people, quantify the trajectory similarity, and further verify the authenticity of the service interaction.

6. The method according to claim 1, characterized in that, The process of extracting caregiver limb movement characteristics from raw radar echo data, quantifying their limb activity index, and identifying caregiver-specific service actions includes: Micro-Doppler features of the caregiver's limbs are extracted from the original radar echo signal. The time-domain signal is converted into a time-frequency distribution through short-time Fourier transform to obtain the frequency and amplitude features of limb movement. Based on the frequency and amplitude characteristics of limb movements, and combined with a frequency weighting function, the limb activity index of caregivers is calculated. Based on the micro-Doppler features, the point cloud data of the caregiver's limbs is clustered, service-specific feature vectors are extracted, and matched with a pre-trained standard service action library to identify the caregiver's specific service actions.

7. The method according to claim 6, characterized in that, The limb activity index is: ； LAI stands for Limb Activity Index, which is dimensionless and ranges from 0 to 1. A higher value indicates greater limb activity. T represents the statistical time window, which is set according to the service type; STFT(t,f) is the short-time Fourier transform value of the radar echo at time t and frequency f, reflecting the time-frequency characteristics of limb micro-movements; w(f) is a frequency weighting function that assigns weights only to the typical frequency range of limb movements, while other frequencies are weighted at 0. Micro-Doppler features are extracted using the following formula: ； Where S(t,f) is the time-frequency distribution value, and f is the square of the magnitude of the STFT; x(τ) is the original echo signal received by the radar; w(τ−t) is a window function used for frame processing of time-domain signals; j is the imaginary unit; τ is the integration variable, representing time; The service-specific feature vector is: ； Among them, F service This is a service-specific feature vector used to distinguish different nursing service actions; f centroid The centroid of the spectrum reflects the dominant frequency of limb movement; f bandwidth The spectral bandwidth reflects the frequency distribution range of limb movements; f entropy Spectral entropy reflects the uniformity of the frequency distribution of limb movements; f peak The peak value represents the strongest frequency component of limb movement.

8. The method according to claim 1, characterized in that, The establishment of a comprehensive evaluation model to determine whether nursing services actually occur, to quantify and evaluate the service quality level, and to detect abnormal service behaviors includes: A multi-dimensional feature baseline database was established for 36 long-term care insurance services. Actual monitoring data was initially matched with baseline parameters to determine the types of services to be evaluated. Analyze the caregiver's action sequence in real time to determine the timing of service switching and ensure that the assessment is consistent with the actual service process; By introducing service item specificity weights and time consistency tests, a comprehensive scoring function is constructed to quantify service quality into a comprehensive score, with a higher score indicating better service quality. If all indicators meet the minimum valid threshold and there are no obvious anomalies, the service is judged to have actually occurred; otherwise, it is judged to have been a fake service. Based on the comprehensive score, the service quality is divided into four levels: excellent, good, satisfactory, and unsatisfactory.

9. The method according to claim 1, characterized in that, The following formula is used to calculate the service status sequence in real time and identify service switching: ； in Let be the forward probability of being in service state j at time t; λ is the global parameter; N represents the total number of service states; Let be the forward probability of being in service state i at time t-1; The critical moments of service switching are identified using a mutation point detection algorithm, and the criteria for judgment are as follows: ； in Let be the rate of change of position at time t. For motion pattern feature vectors, The preset switching threshold; The comprehensive scoring function is as follows: ； in The service is rated as a whole, dimensionless value ranging from 0 to 100, with higher quality being better. α, β, γ, δ, ϵ are the weight coefficients of each indicator, satisfying α+β+γ+δ+ϵ=1, and are dynamically adjusted according to the service type; Distance scoring is calculated based on the match between the actual average interaction distance and the baseline distance; The duration score is calculated based on the degree of matching between the actual service duration and the baseline duration. Activity level is scored based on the degree of match between the actual physical activity index and the baseline physical activity index. The service is scored based on its continuity over time, assessing whether the service process was interrupted; no interruption results in a perfect score. The service switching naturalness score is determined by evaluating whether the order of service switching is logical; if it is, it is a perfect score.

10. A long-term care home-based elderly care service monitoring system based on millimeter-wave radar, characterized in that, include: The millimeter-wave radar acquisition module is configured to transmit and receive echo signals, acquire point cloud data and range-Doppler heatmaps of human targets, and provide raw radar data for all analysis stages. The signal preprocessing module is configured to reduce noise and extract effective targets from the raw radar data. It eliminates reflection interference from the static environment through a static clutter filtering algorithm, and combines an adaptive clustering algorithm and constant false alarm rate detection to remove false noise points, identify the centroid position of human targets, and output effective human target point cloud data. The multi-target tracking module is configured to distinguish and continuously track the trajectories of elderly people and caregivers in a scene. The behavior recognition module is configured to extract limb micro-Doppler features from radar echoes, combine them with short-time Fourier transform to analyze the frequency and amplitude features of limb movements, identify the caregiver's specific nursing service actions, and calculate a quantified limb activity index to determine the caregiver's limb activity level. The service assessment module is configured to determine whether nursing services actually occur, and to quantify and grade the service quality.

Images