A long-term care insurance service quality control method and system based on millimeter wave point cloud

Abstract

The application discloses a long-term care insurance service quality control method and system based on a millimeter wave point cloud, relates to the technical field of pattern recognition, and comprises the following steps: collecting long-term care insurance service orders and quality control specifications, writing service places and nursing roles into a service process, converting quality control requirements extracted from the quality control specifications into nursing action nodes and connection relationships, simultaneously configuring time length requirements and action state requirements, and constructing a long-term care insurance service quality control task graph; based on the service places and the nursing roles in the long-term care insurance service quality control task graph, starting a millimeter wave radar, pre-processing continuously collected millimeter wave point clouds, identifying nursing staff and nursing objects, simultaneously calculating point cloud observability according to a preset observation window, and forming a nursing role observable point cloud time sequence. The application classifies nursing action nodes that are not synchronized and aligned into service action loss and point cloud evidence deficiency, so that nursing behavior space-time recognition representation has a clear point cloud evidence basis.

Description

Technical Field

[0001] This invention relates to the field of pattern recognition technology, and in particular to a long-term care insurance service quality control method and system based on millimeter-wave point clouds. Background Technology

[0002] As the scale of long-term care insurance services continues to expand, service quality verification is gradually developing towards digitalization, continuity, and objectivity. Millimeter-wave radar can acquire human point clouds in low-light, occluded, and privacy-sensitive scenarios, providing a non-visual perception foundation for the identification of caregivers and care recipients, as well as the recognition of care behavior patterns. By combining the time series of observable point clouds of caregivers, the PointNet++ and LSTM fusion model, and the long-term care insurance service quality control task graph, continuous representation of caregiving actions, caregiver interactions, service duration, and changes in the state of care recipients can be achieved, providing technical support for the standardized verification, quality assessment, and record retention of long-term care insurance service processes.

[0003] Current long-term care insurance service quality control relies heavily on manual inspections, video verification, or single nursing action recognition. Nursing actions obtained through pattern recognition, object recognition, and subject recognition are directly used as the basis for determining whether a service action has occurred. When millimeter-wave point clouds are affected by the sitting or lying posture of the patient, obstruction by caregivers, the installation location of millimeter-wave radar, and the layout of the service area, the failure to recognize nursing actions may stem from either missing service actions or insufficient point cloud evidence. Existing technologies lack a mechanism to integrate and verify nursing action nodes, the patient's state of action, and the observability of point clouds, making it difficult to distinguish the reasons for unrecognized nursing actions and affecting the determination of nursing operation standardization levels and the location of abnormal sources. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a long-term care insurance service quality control method based on millimeter-wave point clouds to solve the problem of difficulty in distinguishing between missing service actions and insufficient point cloud evidence in long-term care insurance service quality control.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, this invention provides a long-term care insurance service quality control method based on millimeter-wave point clouds, comprising: collecting long-term care insurance service orders and quality control specifications; writing service locations and nursing roles into the service process; converting quality control requirements extracted from the quality control specifications into nursing action nodes and connection relationships; configuring duration requirements and function status requirements; and constructing a long-term care insurance service quality control task map; activating millimeter-wave radar based on the service locations and nursing roles in the long-term care insurance service quality control task map; preprocessing continuously collected millimeter-wave point clouds and identifying nursing staff and nursing subjects; calculating the observability of point clouds according to a preset observation window to form a time sequence of observable point clouds for nursing roles; and constructing a PointNet++ and LSTM fusion model, inputting the time sequence of observable point clouds for nursing roles into the PointNet++ and LSTM fusion model to extract nursing action nodes and connection relationships. The spatial posture characteristics of caregivers and patients, along with the temporal characteristics of nursing actions, are used to identify changes in the patient's state, forming a spatiotemporal representation of nursing behavior. This representation is then matched with the long-term care insurance service quality control task map. The compliance of actions and the effective service duration are verified according to the nursing action nodes and their connections, generating a nursing action compliance verification record. Changes in the patient's state are also verified according to the requirements of the function state, generating a long-term care insurance service quality control judgment. Based on the non-synchronized alignment of nursing action nodes, the observability of the point cloud, and changes in the patient's state in the long-term care insurance service quality control judgment, assumptions of node execution and non-execution are constructed. The synchronized alignment time period is reconstructed, and the consistency of the function state is reviewed. This distinguishes between nursing action omissions, missing service actions, and insufficient point cloud evidence, determining the source of anomalies and the level of nursing operation standardization, ultimately forming a long-term care insurance service quality control file.

[0007] As a preferred embodiment of the long-term care insurance service quality control method based on millimeter-wave point clouds described in this invention, the steps for constructing the long-term care insurance service quality control task graph are as follows: Collect long-term care insurance service orders and quality control standards, analyze the service item execution relationship and arrange the service process, write the corresponding service location and nursing role, and form a long-term care insurance service process; According to the service items in the long-term care insurance service process, the corresponding quality control standards are retrieved, and the requirements for the execution of nursing actions, duration requirements, and the status requirements of the nursing subjects are extracted to form the service item quality control requirements; The nursing action execution requirements in the service project quality control requirements are converted into nursing action nodes one by one. The sequential constraints of the actions are extracted and the connection relationship between the nursing action nodes is established. The corresponding duration requirements and function status requirements are configured to the nursing action nodes to construct a long-term care insurance service quality control task map.

[0008] As a preferred embodiment of the long-term care insurance service quality control method based on millimeter-wave point clouds described in this invention, the steps for forming the observable point cloud sequence for nursing roles are as follows: According to the long-term care insurance service quality control task map, the service sites are matched with the corresponding millimeter-wave radars, the sensing areas are set up and a unified time and spatial coordinate reference is established. The corresponding millimeter-wave radars are started to continuously collect point cloud frames carrying the collection time and spatial attributes, and the original millimeter-wave point cloud sequence is formed by arranging them according to the collection time. Background reflection suppression, outlier removal and density filtering are performed on the original millimeter-wave point cloud sequence. Nursing roles are matched by target clustering and cross-frame tracking to identify nursing staff and nursing objects and form nursing role point cloud trajectories. The point cloud trajectory of the nursing role is divided into continuous point cloud segments according to the preset observation window. The observability of each continuous point cloud segment is calculated and written into the corresponding continuous point cloud segment to form the observable point cloud time sequence of the nursing role.

[0009] As a preferred embodiment of the long-term care insurance service quality control method based on millimeter-wave point clouds described in this invention, the steps for constructing the PointNet++ and LSTM fusion model are as follows: Collect the time sequence of observable point cloud of nursing roles that has been manually verified, and label the nursing action categories and nursing object status change categories according to the nursing action nodes and nursing object status requirements in the long-term care insurance service quality control task map to form a nursing behavior recognition training sample set; A local coordinate system is established using the centroid and body axis of the nursing object point cloud. The nursing staff point cloud and nursing object point cloud in the nursing behavior recognition training sample set are transformed into the local coordinate system. Spatial pose features are extracted by establishing the nursing staff PointNet++ branch and the nursing object PointNet++ branch respectively. The nursing role interaction features formed by the relative position and relative motion relationship between the nursing staff point cloud and the nursing object point cloud are fused with the observability of the corresponding point cloud to form a dual-role spatial training feature sequence. The dual-role space training feature sequence is input into the LSTM temporal feature extraction network. Nursing action recognition branch and nursing object state change recognition branch are set up. The model is trained according to the connection relationship between nursing action nodes, and a PointNet++ and LSTM fusion model is constructed.

[0010] As a preferred embodiment of the long-term care insurance service quality control method based on millimeter-wave point clouds described in this invention, the steps for forming a spatiotemporal recognition representation of nursing behavior are as follows: The observable point cloud of nursing roles is input into the PointNet++ and LSTM fusion model. Spatial pose features of nursing staff and nursing objects are extracted through the PointNet++ branches of nursing staff and nursing objects. The timing features of nursing actions are extracted through LSTM to form a joint recognition sequence of nursing actions and nursing object states. The joint recognition sequence of nursing actions and nursing object status is merged into continuous action intervals, the nursing object status changes corresponding to each nursing action are identified, and the corresponding point cloud observability is written into the form of spatiotemporal recognition representation of nursing behavior.

[0011] As a preferred embodiment of the long-term care insurance service quality control method based on millimeter-wave point clouds described in this invention, the steps for forming a compliance verification record of nursing actions are as follows: Organize the spatiotemporal recognition representation of nursing behaviors according to the collection time, merge adjacent spatiotemporal recognition representations of nursing behaviors with the same nursing action category and continuous time into continuous nursing events, and re-divide continuous nursing events when either the nursing action category changes or the time interval between adjacent collection times exceeds the window duration of the preset observation window, forming a continuous nursing event flow; The continuous nursing event flow is matched with the long-term care insurance service quality control task map. The nursing action category, start and end time and point cloud observability are read from the continuous nursing event flow and compared with the nursing action nodes, duration requirements and node connection relationship. The alignment cost is calculated and candidate alignment paths are generated by dynamic time warping. The candidate alignment path with the lowest cumulative alignment cost is determined as the target alignment path. The time interval of nursing events in the target alignment path that are consistent with the nursing action category and nursing action node and whose node transfer conforms to the node connection relationship is determined as the synchronous alignment period. The order of nursing actions is verified according to the node connection relationship, and the synchronous alignment period whose duration meets the corresponding time requirement is accumulated as the effective service time, forming a nursing action compliance verification record.

[0012] As a preferred embodiment of the long-term care insurance service quality control method based on millimeter-wave point clouds described in this invention, the steps for generating long-term care insurance service quality control judgments are as follows: For each nursing action node in the nursing action compliance verification record, a status response window is set after the synchronization alignment period ends, according to the status achievement time limit in the nursing object's function status requirement. The nursing object status change in the status response window is matched with the nursing object's function status requirement to form a nursing action status verification record. Based on the nursing action compliance verification record and nursing action status verification record, nursing action nodes with deviations in action sequence, insufficient effective service time, and invalid patient status are identified and corresponding verification identifiers are written. The location of nursing action nodes that are not synchronized and the observability of point cloud for the corresponding time period are recorded to generate a long-term care insurance service quality control judgment.

[0013] As a preferred embodiment of the long-term care insurance service quality control method based on millimeter-wave point clouds described in this invention, the steps for forming nursing service anomaly classification records are as follows: From the quality control judgment of long-term care insurance services, we extract the nursing action nodes that are not synchronized, the adjacent synchronized time periods, the observability of point cloud and the changes in the status of the nursing object. We construct the node execution hypothesis and the node non-execution hypothesis respectively. We fill the node execution hypothesis into the corresponding node position and reconstruct the synchronized time period. We leave the node non-execution hypothesis with node vacancy. We verify the subsequent node connection relationship, effective service duration and function status requirements respectively to form dual hypothesis verification information. Based on the dual hypothesis verification information, when the observability of the point cloud does not reach the preset point cloud observability threshold, it is marked as insufficient point cloud evidence. When it reaches the preset point cloud observability threshold and the node has executed the hypothesis to restore the connection relationship and the consistency of the function state, it is marked as nursing action recognition omission. When the node has not executed the hypothesis to maintain subsequent alignment and the function state is not established, it is marked as service action missing, thus forming a nursing service abnormality classification record.

[0014] As a preferred embodiment of the long-term care insurance service quality control method based on millimeter-wave point clouds described in this invention, the steps for forming a long-term care insurance service quality control file are as follows: Write the nursing service anomaly classification record into the long-term care insurance service quality control task map, use the node corresponding to the anomaly classification identifier as the candidate anomaly node, copy it to form a verification task map, replace the state change of the candidate anomaly node with the function state requirement along the node connection relationship, review the subsequent nodes, distinguish between independent anomaly sources and related anomalies, and form an anomaly source location record. Based on the independent abnormal sources and associated abnormal ranges in the abnormal source location record, and in conjunction with the long-term care insurance service quality control judgment, the nursing operation standard level is determined according to the level judgment rules in the quality control specification. The nursing operation standard level is written into the abnormal source location record and archived in conjunction with the long-term care insurance service quality control judgment to form a long-term care insurance service quality control file.

[0015] Secondly, the present invention provides a long-term care insurance service quality control system based on millimeter-wave point clouds, including a service task module, a point cloud perception module, a spatiotemporal recognition module, a quality control verification module, and a quality control archiving module; The service task module is used to collect long-term care insurance service orders and quality control specifications, write service locations and nursing roles into the service process, and convert the quality control requirements extracted from the quality control specifications into nursing action nodes and connection relationships. At the same time, it configures duration requirements and function status requirements to construct a long-term care insurance service quality control task diagram. The point cloud perception module is used to activate millimeter-wave radar based on the service location and nursing role in the long-term care insurance service quality control task map, preprocess the continuously collected millimeter-wave point cloud and identify nursing staff and nursing objects, and calculate the observability of the point cloud according to the preset observation window to form the observable point cloud time sequence of nursing roles. The spatiotemporal recognition module is used to construct a PointNet++ and LSTM fusion model. The time sequence of the observable point cloud of nursing roles is input into the PointNet++ and LSTM fusion model to extract the spatial posture features and time sequence features of nursing actions of nursing staff and nursing objects, and to identify changes in the state of nursing objects, forming a spatiotemporal recognition representation of nursing behavior. The quality control verification module is used to match the spatiotemporal identification representation of nursing behavior with the long-term care insurance service quality control task map, verify the compliance of actions and effective service duration according to nursing action nodes and connection relationships, and verify the changes in the state of the nursing object according to the requirements of the function state, and generate a long-term care insurance service quality control judgment. The quality control archiving module is used to extract nursing action nodes, observability of point clouds, and changes in the status of nursing subjects from the quality control judgment of long-term care insurance services. Based on the observability of point clouds and the requirements of the function status, it distinguishes between nursing action omissions, service action deficiencies, and insufficient point cloud evidence, forming a nursing service anomaly classification record. Combined with the node position, it determines the nursing operation standard level and the source of the anomaly, forming a long-term care insurance service quality control archive.

[0016] The beneficial effects of this invention are as follows: by extracting the location and observability of nursing action nodes from the quality control judgment of long-term care insurance services, and classifying non-synchronously aligned nursing action nodes as missing service actions and insufficient point cloud evidence, the spatiotemporal recognition and representation of nursing behavior has a clear point cloud evidence basis; based on the observability of the point cloud within the period to be verified, it can be confirmed whether the nursing role point cloud and nursing role point cloud trajectory can support nursing action recognition, and the nursing operation standard level, the first abnormal node, the period to be verified and the abnormal classification label can be recorded, realizing the node location, time period recording and evidence archiving of nursing abnormalities. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a long-term care insurance service quality control method based on millimeter-wave point clouds.

[0019] Figure 2 A flowchart for creating a quality control file for long-term care insurance services.

[0020] Figure 3 A comparative chart showing the accuracy of three types of nursing action recognition under different point cloud observability conditions.

[0021] Figure 4A comparative data graph showing the components of observability of point clouds under different point cloud degradation intensities. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a long-term care insurance service quality control method based on millimeter-wave point clouds, including the following steps: S1. Collect long-term care insurance service orders and quality control specifications, write service locations and nursing roles into the service process, and convert the quality control requirements extracted from the quality control specifications into nursing action nodes and connection relationships. At the same time, configure duration requirements and function status requirements to construct a long-term care insurance service quality control task map.

[0026] S1.1: Collect long-term care insurance service orders and quality control specifications, analyze the service item execution relationship and arrange the service process, write the corresponding service location and nursing role, and form a long-term care insurance service process; Specifically, long-term care insurance service orders and quality control standards are collected, service items and their order are read from the long-term care insurance service orders, the execution relationship of service items is determined according to the order of service items and the service process is arranged, and the service location and caregiver role recorded in the long-term care insurance service order are written into the corresponding service items to form the long-term care insurance service process.

[0027] It should be noted that long-term care insurance service orders originate from business records generated by long-term care insurance agencies or designated nursing service providers after confirming the disability level of the care recipient and the service plan. Long-term care insurance service orders record service items, service periods, service locations, and the caregivers and care recipients, where the caregivers and care recipients constitute the caregiver roles. Quality control standards originate from service quality standards issued by the long-term care insurance authorities or agencies. Quality control standards record the execution requirements, sequence requirements, duration requirements, and status requirements of care recipients for each service item. When collecting quality control standards, the source, applicable region, version, and effective period of the standards are linked to determine the scope of application of the corresponding quality control requirements.

[0028] S1.2: Retrieve the relevant quality control standards according to the service items in the long-term care insurance service process, and extract the requirements for the execution of nursing actions, duration requirements, and the requirements for the status of the nursing subjects to form the service item quality control requirements; Specifically, the corresponding quality control specifications are retrieved according to the service item names in the long-term care insurance service process. The nursing action content and the sequence of actions corresponding to the service item in the quality control specifications are read as nursing action execution requirements. The duration of the nursing action is read as the duration requirement. The state that the nursing patient should reach after the nursing action is read as the nursing patient's functional state requirement. The nursing action execution requirements, duration requirements and nursing patient's functional state requirements are written into the corresponding service item to form the service item quality control requirements.

[0029] It should be noted that nursing action content refers to the operational behaviors of nursing staff and the cooperation process between nursing staff and patients as specified in the quality control standards for specific service items. This includes the approach position, body posture, limb movement, contact points, direction of movement, sequence of movement, and duration of the process. For example, assisting with turning over includes the nursing staff approaching the bedside, supporting the patient's shoulders and hips, guiding the patient to turn to a side-lying position and adjusting the support position; assisting with getting up includes the nursing staff supporting the patient's upper body, guiding the patient to bend their legs and turn from a supine to a sitting position; assisting with transfer includes the nursing staff supporting the patient away from the bed area and moving them to the wheelchair area. The requirements for performing nursing actions can also correspond to bathing care, postural care, and rehabilitation training. Among these, bathing care can be divided into nursing actions based on the nursing staff's support in positioning, body part cleaning, maintaining the patient's posture, and postural recovery after cleaning; postural care can be divided into nursing actions based on the patient's postural changes, adjustment of support position, and maintenance of the target posture; and rehabilitation training can be divided into nursing actions based on the nursing staff's support position, the direction of the patient's limb movement, the reciprocating movement process, and the stable state after training. The state that the patient should reach after a nursing action refers to the target state specified by the quality control standard for the specific nursing action and that can be identified by millimeter-wave point cloud. This includes the body posture, spatial position, support relationship, limb movement, and stability. For example, after assisting the patient to turn over, the patient should change from supine to lateral and maintain this position. After assisting the patient to get up, the patient's torso should change from supine to sitting and remain stable. After assisting the patient to transfer, the patient's center of gravity should move from the bed area to the wheelchair area and the support position should change accordingly.

[0030] S1.3: Convert the nursing action execution requirements in the service project quality control requirements into nursing action nodes one by one, extract the action sequence constraints and establish the connection relationship between nursing action nodes, configure the corresponding duration requirements and function status requirements to the nursing action nodes, and construct the long-term care insurance service quality control task diagram.

[0031] Specifically, following the rule that one nursing action execution requirement corresponds to one nursing action node, the action name is written as the node name and the action content is written into the corresponding node; the action sequence constraints are extracted from the preceding and subsequent actions recorded in the nursing action execution requirements, and the corresponding nursing action nodes are connected according to the action sequence constraints to form node connection relationships; according to the correspondence between nursing action execution requirements, duration requirements, and function status requirements, the duration requirements and function status requirements are written into the same nursing action node; with the nursing action node as the vertex, the node connection relationship formed by the action sequence constraints in the nursing action execution requirements as the directed edge, and the duration requirements and function status requirements as the node attributes of the corresponding nursing action node, a long-term care insurance service quality control task graph is constructed.

[0032] S2. Based on the service locations and nursing roles in the long-term care insurance service quality control task map, the millimeter-wave radar is activated to preprocess the continuously collected millimeter-wave point clouds and identify nursing staff and nursing subjects. At the same time, the observability of the point clouds is calculated according to the preset observation window to form the observable point cloud sequence of nursing roles.

[0033] S2.1: According to the long-term care insurance service quality control task map, match the service site to the corresponding millimeter-wave radar, set the sensing area and establish a unified time reference and spatial coordinate reference, start the corresponding millimeter-wave radar to continuously collect point cloud frames carrying the collection time and spatial attributes, and arrange them according to the collection time to form the original millimeter-wave point cloud sequence. Specifically, the service locations recorded in the long-term care insurance service quality control task map are read, and the service locations are compared with the installation locations registered by the millimeter-wave radar to determine the millimeter-wave radar installed at the corresponding service location. The coverage area of ​​the millimeter-wave radar is defined according to the scope of nursing services implemented within the service location, and the space where the scope of nursing services implemented and the coverage area of ​​the millimeter-wave radar overlap is set as the sensing area. The continuous point cloud frames collected using the same clock source are written with the collection time, and the spatial coordinate reference is established with the installation location of the millimeter-wave radar as the coordinate origin and the ground direction and vertical direction of the service location. The grassroots edge agent starts the corresponding millimeter-wave radar to continuously collect point cloud frames, and receives the corresponding point cloud frames through local communication connection. The three-dimensional position of the point cloud points under the spatial coordinate reference is written as a spatial attribute into the corresponding point cloud frame, and arranged according to the collection time to form the original millimeter-wave point cloud sequence.

[0034] It should be noted that the grassroots edge intelligence agent is set up in the service location or designated nursing service institution and establishes a local communication connection with the corresponding millimeter-wave radar. It includes a processor, a memory, and a point cloud processing program and a model inference program stored in the memory. It is used to perform preprocessing of the original millimeter-wave point cloud sequence, nursing role recognition, nursing role point cloud trajectory formation, point cloud observability calculation, nursing behavior recognition, and long-term care insurance service quality control verification locally. The scope of nursing service implementation refers to the actual spatial range in which the corresponding service item in the long-term care insurance service order is carried out. It is determined by the nursing location and nursing actions recorded in the service item. For example, bedside care can be limited to the area one to two meters around the bed, assisted transfer can be limited to the movement area between the bed and the wheelchair parking position, and indoor activity care can be limited to the room where the patient is located. The millimeter-wave radar coverage area refers to the spatial range in which the millimeter-wave radar can stably collect human point clouds after installation. It is determined by the detection distance, horizontal field of view, and vertical field of view of the millimeter-wave radar, and the effective boundary is confirmed by the movement of nursing staff within the service area. For example, the detection distance can be set to 0.5 meters to 6 meters, the horizontal field of view can be set to 45° to 60° left and right, and the vertical field of view can be set to 15° to 30° up and down. When setting the sensing area, the space within the scope of nursing service implementation that is within the millimeter-wave radar coverage area and is not continuously blocked by fixed obstacles is defined as the sensing area.

[0035] S2.2: Background reflection suppression, outlier removal and density filtering are performed on the original millimeter-wave point cloud sequence. Nursing roles are matched by target clustering and cross-frame tracking to identify nursing staff and nursing objects and form nursing role point cloud trajectories. Specifically, a background point cloud is formed using empty field point cloud frames collected before the service begins. Points in the original millimeter-wave point cloud sequence that overlap with the corresponding spatial position of the background point cloud and have a radial velocity of zero are deleted. The spatial positions of the remaining points and their adjacent spatial positions are checked, and isolated points without other points around them are deleted. Points with fewer than a preset density threshold in their neighborhood are also deleted. Points located in adjacent spatial positions with the same radial velocity are connected to form a target point group. Target point groups in adjacent point cloud frames that overlap in spatial range and have the same radial velocity are connected according to the time of collection to form a target trajectory. Combining the nursing roles in the long-term care insurance service quality control task map, target trajectories located within the nursing service implementation area at the service start time and not entering from the boundary of the sensing area are marked as nursing object trajectories. Target trajectories that enter from the boundary of the sensing area and move into the nursing service implementation area are marked as nursing staff trajectories, thus forming a nursing role point cloud trajectory.

[0036] It should be noted that when multiple people exist simultaneously or their target trajectories intersect, the role identification is maintained based on the previous position, movement direction, and activity area of ​​the bound role; when a nursing role leaves and re-enters, the role binding is restored based on the departure boundary, entry boundary, and time continuity; if it cannot be uniquely identified, it is marked as an undetermined nursing role. An empty field point cloud frame is a point cloud frame that contains only reflections from fixed objects such as walls or furniture, collected by millimeter-wave radar when no one is in the service area. The density threshold is the minimum number of points required to determine whether a point cloud has effective point cloud support in adjacent spatial locations. When setting the density threshold, under the condition that the millimeter-wave radar installation location, sensing area, and acquisition configuration remain consistent, point cloud frames of the air field and point cloud frames of caregivers and care recipients in standing, sitting, lying, and assisted transfer states are collected respectively. The number of points contained in adjacent spatial locations for discrete points in the air field and the main body point group is counted. The number of points that can eliminate discrete points in the air field, retain the main body point group, and maintain the continuity of the target trajectory is determined as the density threshold. For example, the density threshold can be set to 3 to 8 neighboring points. The upper limit interval is used when the distance between the millimeter-wave radar and the caregiver is short and the obstruction range is small, and the lower limit interval is used when the distance between the millimeter-wave radar and the caregiver is long, the care recipient is in a sitting or lying position, or there is partial obstruction.

[0037] S2.3: Divide the nursing role point cloud trajectory into continuous point cloud segments according to the preset observation window, calculate the observability of each continuous point cloud segment and write it into the corresponding continuous point cloud segment to form the observable point cloud time sequence of the nursing role.

[0038] Specifically, the point cloud trajectories of nursing roles are captured according to the collection time from morning to night, based on the start and end times of the preset observation window. The trajectory points of nursing staff and nursing objects within the same preset observation window are grouped into the same continuous point cloud segment. Each point cloud frame within a continuous point cloud segment is examined individually. Point cloud frames with the number of nursing staff points reaching a preset threshold for the number of human target point groups are recorded as valid nursing staff point cloud frames, and point cloud frames with the number of nursing object points reaching a preset threshold for the number of human target point groups are recorded as valid nursing object point cloud frames. The proportions of the number of valid nursing staff point cloud frames and the number of valid nursing object point cloud frames to the total number of point cloud frames in the continuous point cloud segment are determined as the completeness of the nursing staff point cloud and the completeness of the nursing object point cloud, respectively. The nursing staff target point groups and nursing object target point groups in adjacent point cloud frames are checked frame by frame. Adjacent target point groups with overlapping spatial ranges and consistent radial velocities are marked as continuous trajectory segments of the same target trajectory. Adjacent target point groups with non-overlapping spatial ranges or inconsistent radial velocities are marked as trajectory interruptions. The trajectory connection status of nursing staff and nursing objects are recorded separately. The proportion of the number of successfully connected adjacent point cloud frame groups to the total number of adjacent point cloud frame groups is determined as the continuity of nursing staff trajectory and nursing object trajectory. The proportion of the number of point cloud frames with valid point cloud frames for both nursing staff and nursing objects to the total number of point cloud frames in a continuous point cloud segment is determined as the nursing role coverage. The observability of point cloud in a continuous point cloud segment is calculated. The observability of point cloud and the corresponding acquisition time period are written into the continuous point cloud segment and arranged according to the acquisition time to form the observable point cloud time sequence of nursing roles.

[0039] The expression for calculating the observability of a point cloud is: ; in, Indicates the first The observability of a continuous point cloud segment, with a value ranging from zero to one; Indicates the first The number of point cloud frames in a continuous point cloud segment whose number of nursing staff points reaches the density threshold. Indicates the first The number of point cloud frames in a continuous point cloud segment whose number of nursing object points reaches the density threshold. Indicates the first The number of point cloud frame groups in which the target point group of nursing staff is successfully connected in adjacent point cloud frames in a continuous point cloud segment; Indicates the first The number of point cloud frame groups in which the target point group of the nursing object is successfully connected in adjacent point cloud frames of a continuous point cloud segment; Indicates the first The number of point cloud frames in a continuous point cloud segment whose number of nursing staff points and nursing object points simultaneously reach the density threshold. Indicates the first The total number of point cloud frames contained in a continuous point cloud segment; This indicates the numbering of consecutive point cloud segments arranged according to the time of acquisition; This indicates that the minimum value is selected from the proportions within the parentheses.

[0040] It should be noted that the point cloud observability expression is only applicable when a continuous point cloud segment contains at least two point cloud frames and the roles of both the nursing staff's trajectory and the nursing object's trajectory have been confirmed. If a continuous point cloud segment contains only one point cloud frame, any nursing role's trajectory is missing, or the nursing role cannot be uniquely confirmed, the point cloud observability expression is not executed, and the point cloud observability is recorded as zero. If the target point group of the nursing staff or the target point group of the nursing object does not reach the threshold for the number of human target point groups, the corresponding point cloud completeness and nursing role coverage are recorded as zero. All components are statistically analyzed according to the same continuous point cloud segment, and the value range is limited to zero to one. The threshold for the number of human target points is determined based on the spatial resolution of the millimeter-wave radar, installation distance, sensing area, acquisition parameters, and the posture of the nursing role. Under the condition that the millimeter-wave radar installation location and acquisition configuration remain consistent, point cloud frames are acquired in an empty field and in standing, sitting, lying, moving, and partially occluded states for both the nursing staff and the nursing subject. Target point groups that can stably represent the human body contour and maintain the target trajectory are manually verified and marked. The distribution of the number of valid human target points and discrete empty field points is statistically analyzed. The number of points that can eliminate non-human point groups, retain human point groups in sitting, lying, and occluded states, and maintain continuous trajectory is determined as the threshold for the number of human target points. For example, the threshold for the number of human target points can be set to 15 to 40 points. The observation window is set based on the acquisition interval of millimeter-wave radar point cloud frames, the duration requirements corresponding to nursing action nodes in the long-term care insurance service quality control task diagram, and the continuous time periods of posture and position changes in the nursing role point cloud trajectory. When setting it, the acquisition interval of continuous point cloud frames is read, and the continuous time periods corresponding to turning over, sitting up, transferring, and maintaining body position are selected from the nursing role point cloud trajectory that has been manually reviewed. The continuous point cloud segments are divided in turn using different time lengths. The time length that can cover the continuous changes of the same nursing action, retain the trajectory of the nursing staff and the trajectory of the nursing object, and avoid classifying the previous and subsequent nursing actions into the same continuous point cloud segment is determined as the observation window. For example, the observation window can be set to one to three seconds, one to two seconds for turning over, sitting up, and transferring, and two to three seconds for maintaining body position.

[0041] Figure 3 This illustrates the changes in the recognition accuracy of three types of nursing actions—assisted turning over, assisted getting up, and assisted transfer—under different observability levels of point clouds during the nursing behavior recognition process. The horizontal axis represents the observability level of the point cloud, the vertical axis represents the recognition accuracy of the nursing action, the three curves represent the recognition accuracy of assisted turning over, assisted getting up, and assisted transfer, respectively, and the error bars represent the standard deviation of the recognition accuracy of each nursing service record within the same observability level range of the point cloud. Figure 3It is known that when the observability of the point cloud is in the low range, the point clouds of nursing staff, nursing subjects, and nursing roles are difficult to stably represent the nursing action process, and the recognition accuracy of the three types of nursing actions is at a low level. As the observability of the point cloud increases, the recognition accuracy of the three types of nursing actions generally increases, but the change in the recognition accuracy of each type of nursing action varies due to the influence of the degree of human body occlusion, the interaction range of nursing roles, and the continuity of nursing role point cloud trajectories. When the observability of the point cloud enters the high range, the recognition accuracy of the three types of nursing actions remains at a high level, indicating that the nursing role point cloud and nursing role point cloud trajectory can provide sufficient point cloud evidence for the PointNet++ and LSTM fusion model to extract spatial pose features and nursing action temporal features. Therefore, under the data conditions used in this embodiment, the observability of the point cloud is generally positively correlated with the recognition accuracy of the three types of nursing actions. The observability of the point cloud can quantitatively represent the sufficiency of nursing role point cloud evidence and is used to confirm whether the point cloud evidence in the period to be verified can support the recognition of nursing actions, providing a quantitative basis for the subsequent abnormal classification of unsynchronized nursing action nodes.

[0042] S3. Construct a PointNet++ and LSTM fusion model. Input the time sequence of the observable point cloud of nursing roles into the PointNet++ and LSTM fusion model to extract the spatial posture features and time sequence features of nursing actions of nursing staff and nursing objects, and identify the state changes of nursing objects to form a spatiotemporal recognition representation of nursing behavior.

[0043] S3.1: Collect the time sequence of observable point cloud of nursing roles that has been manually reviewed, and label the nursing action categories and nursing object status change categories according to the nursing action nodes and nursing object status requirements in the long-term care insurance service quality control task map to form a nursing behavior recognition training sample set; Specifically, the observable point cloud time sequence of nursing roles, which has been manually verified, is collected. The point cloud frame corresponding to the start and end of the nursing action is determined according to the collection time. The continuous point cloud frame interval between the two point cloud frames is taken as the corresponding time period of the nursing action. The nursing action name recorded in the corresponding nursing action node in the long-term care insurance service quality control task map is labeled as the nursing action category. The point cloud frame corresponding to the start of changes in the body posture, spatial position or support relationship of the nursing object and the point cloud frame corresponding to the nursing object reaching the required state of the nursing object are determined. The continuous point cloud frame interval between the two point cloud frames is taken as the corresponding time period of the nursing object state change. The change type from the state before the change to the state recorded in the required state of the nursing object is labeled as the nursing object state change category. This forms a nursing behavior recognition training sample set that includes nursing staff point clouds, nursing object point clouds, nursing action categories and nursing object state change categories.

[0044] It should be noted that joint grouping is established based on the nursing object identifier and the service location identifier. The time series of observable point clouds of nursing roles formed by the same nursing object in different service locations, as well as the time series of observable point clouds of nursing roles corresponding to different nursing objects in the same service location, are all classified into the same joint group to avoid the repeated appearance of nursing objects or service locations in different datasets. According to the sample distribution of nursing action category, nursing object state change category, and point cloud observability interval, each joint group is exemplarily allocated as training data, validation data, and test data in proportions of 70%, 15%, and 15%, respectively. Among them, the training data is used to update the parameters of the PointNet++ and LSTM fusion model, the validation data is used to determine the recognition confidence threshold and training stopping condition, and the test data is only used to evaluate the recognition performance after training.

[0045] S3.2: Establish a local coordinate system based on the centroid and body axis of the nursing object point cloud, transform the nursing staff point cloud and nursing object point cloud in the nursing behavior recognition training sample set to the local coordinate system, establish the nursing staff PointNet++ branch and the nursing object PointNet++ branch respectively to extract spatial pose features, and fuse the nursing role interaction features formed by the relative position and relative motion relationship between the nursing staff point cloud and the nursing object point cloud with the observability of the corresponding point cloud to form a dual-role spatial training feature sequence; Specifically, the point cloud of the nursing object is read from the training sample set for nursing behavior recognition. The centroid of the nursing object point cloud is set as the origin of the local coordinate system. The main axis of the body is determined based on the maximum extension direction of the nursing object point cloud in three-dimensional space, without relying on the segmentation of the head point cloud and the torso point cloud. The positive direction of the main axis of the body is determined based on the pre-calibrated head direction of the bed or the initial orientation of the nursing object in the service location. Subsequent valid point cloud frames select the main axis direction of the body that is consistent with the direction of the previous valid point cloud frame to avoid orientation flipping between adjacent point cloud frames. When the nursing object is in a supine position, the longitudinal direction of the bed is used for correction. When the body axis is partially occluded or the head point cloud is missing, the maximum extension direction of the main body of the nursing object point cloud is retained. When the nursing object point cloud is too sparse or the body axis cannot be stably determined, the body axis of the previous effective point cloud frame is used, and two other mutually perpendicular coordinate axes are determined based on the spatial coordinate reference of the service location. According to the position of the nursing staff point cloud and the nursing object point cloud relative to the origin of the local coordinate system and each coordinate axis, the nursing staff point cloud and the nursing object point cloud are transformed to the local coordinate system, and the nursing staff PointNet++ branch and the nursing object PointNet++ branch are established respectively. The point clouds of nursing staff and nursing subjects in the nursing behavior recognition training sample set are read. At each acquisition time, the farthest point sampling is performed by the PointNet++ branch of nursing staff and the PointNet++ branch of nursing subjects respectively, and the center point is selected from the corresponding point cloud. The spherical neighborhood query is performed with the center point as the reference, and the points within the preset neighborhood range are assigned to the corresponding neighborhood point set. Neighborhood point features are formed according to the spatial position and radial velocity of the neighborhood points relative to the center point. The neighborhood point features are fed into a shared multilayer perceptron for point-by-point feature mapping. Max pooling is used to retain the features in the neighborhood point set that can represent body contour, limb position and posture direction. The point cloud coverage is expanded according to the PointNet++ step-by-step sampling, neighborhood set-up and feature aggregation method to obtain the spatial posture features of nursing staff and nursing subjects respectively. Based on the positional relationship between the nurse point cloud and the nursing object point cloud at the same acquisition time, and the continuous change of the positional relationship at adjacent acquisition times, nursing role interaction features are formed. The spatial posture features of nurses, spatial posture features of nursing objects, nursing role interaction features and the observability of corresponding point clouds are fused and arranged according to the acquisition time to form a dual-role spatial training feature sequence.

[0046] It should be noted that the neighborhood range is set based on the spatial resolution capability of the millimeter-wave radar, the density between points in the nurse point cloud and the nurse object point cloud, the body scale of the nurse role, the installation distance of the millimeter-wave radar, and the occlusion of the service location. Nursing staff point clouds and nurse object point clouds corresponding to different nursing action categories and nurse object state change categories are selected from the nursing behavior recognition training sample set. PointNet++ spherical neighborhood queries are then performed sequentially using different candidate neighborhood ranges to check whether the neighborhood point set can preserve the local contours of limbs, trunk posture, and overall body shape, while avoiding grouping the nurse point cloud and the nurse object point cloud into the same neighborhood point set. The neighborhood range used for each layer of PointNet++ is determined accordingly. For example, the layer for extracting local limb contours can be 0.1 meters to 0.2 meters, the layer for extracting trunk posture can be 0.2 meters to 0.4 meters, and the layer for extracting overall body shape can be 0.4 meters to 0.8 meters.

[0047] S3.3: Input the dual-role space training feature sequence into the LSTM temporal feature extraction network, set up nursing action recognition branch and nursing object state change recognition branch, and train the model according to the connection relationship between nursing action nodes to build a PointNet++ and LSTM fusion model.

[0048] Specifically, the dual-role spatial training feature sequence is sequentially fed into an LSTM temporal feature extraction network according to the acquisition time. The LSTM temporal feature extraction network forms temporal features corresponding to each acquisition time. Nursing action recognition and nursing object state change recognition branches are set after the LSTM temporal feature extraction network. The temporal features formed by the LSTM temporal feature extraction network at each acquisition time are fed into the nursing action recognition and nursing object state change recognition branches respectively. Fully connected layers are configured with corresponding neurons according to the number of nursing action categories and the number of nursing object state change categories. Each neuron maps the temporal feature to the corresponding category score through the weights obtained during training. The Softmax function then converts each category score into a category probability that sums to one. These category probabilities are used as the nursing action category determination value and the nursing object state change determination value. The object state change category judgment value; the nursing action categories and nursing object state change categories labeled in the nursing behavior recognition training sample set are used as supervision targets, and the cross-entropy loss function is used respectively; the connection relationship between nursing action nodes in the long-term care insurance service quality control task graph is read. When the nursing action category changes at adjacent collection times, the nursing action category that does not conform to the allowed connection order is recorded in the sequence constraint loss. The two cross-entropy losses and the sequence constraint loss are used together to update the parameters of the nursing staff PointNet++ branch, the nursing object PointNet++ branch, the LSTM temporal feature extraction network, the nursing action recognition branch and the nursing object state change recognition branch. The LSTM temporal feature extraction network training is completed to obtain the PointNet++ and LSTM fusion model.

[0049] It should be noted that the allowed connection order is the node order from the preceding nursing action node to the following nursing action node in the long-term care insurance service quality control task diagram. During the duration of the same nursing action, the nursing action category is allowed to remain unchanged. The sequence constraint loss sets a penalty weight for switching nursing action categories that do not conform to the allowed connection order, but does not delete the corresponding nursing action category output, so that the actual abnormal nursing action sequence can be retained in the subsequent nursing action sequence verification process.

[0050] S3.4: Input the time sequence of the observable point cloud of nursing roles into the PointNet++ and LSTM fusion model, extract the spatial pose features of nursing staff and nursing objects through the PointNet++ branch of nursing staff and the PointNet++ branch of nursing objects, and extract the time sequence features of nursing actions through LSTM to form a joint recognition sequence of nursing actions and nursing object status; Specifically, the PointNet++ and LSTM fusion model is deployed in the grassroots edge agent. The spatial pose features of nurses, the spatial pose features of nursing objects, the interaction features of nursing roles, and the corresponding observability of the point cloud data are sequentially fed into the PointNet++ and LSTM fusion model in the grassroots edge agent according to the acquisition time. The nurse PointNet++ branch performs layer-by-layer neighborhood grouping of the nurse point cloud at each acquisition time and extracts the spatial positional relationships, body orientation, and limb postures of each body part of the nurse. The nursing object PointNet++ branch performs layer-by-layer neighborhood grouping of the nursing object point cloud at each acquisition time and extracts the spatial positional relationships, body orientation, and limb postures of each body part of the nursing object. The spatial positional relationships, body orientation, and limb postures of the nurses and the patients are used to form the spatial posture features of the nurses and the patients. The spatial posture features of the nurses and the patients, along with the observability of the corresponding point cloud, are sequentially fed into the LSTM temporal feature extraction network according to the acquisition time. The LSTM temporal feature extraction network preserves the continuous changes in the spatial postures of the nurses and the patients, as well as the interaction relationships of the nursing roles, between adjacent acquisition times, forming the temporal features of nursing actions. The nursing action recognition branch determines the category of nursing action based on the temporal features of nursing actions, and the nursing patient state change recognition branch determines the category of nursing patient state change based on the temporal features of nursing actions, forming a joint recognition sequence of nursing actions and nursing patient state.

[0051] Figure 4 This diagram illustrates the changes in the completeness of the nurse point cloud, the completeness of the nurse object point cloud, the continuity of the nurse trajectory, the continuity of the nurse object trajectory, the coverage of the nursing role, and the observability of the point cloud under different point cloud degradation intensities during the formation of the observable point cloud for nursing roles. The horizontal axis represents the point cloud degradation intensity, and the vertical axis represents the normalization value corresponding to each indicator. The six curves represent the completeness of the nurse point cloud, the completeness of the nurse object point cloud, the continuity of the nurse trajectory, the continuity of the nurse object trajectory, the coverage of the nursing role, and the observability of the point cloud, respectively. Figure 4 It can be seen that as the point cloud degradation intensity increases, the number of point cloud frames that meet the effective observation conditions for nursing staff and nursing subjects gradually decreases. The continuity of the nursing staff and nursing subject trajectories formed by cross-frame tracking decreases, and the coverage of effective point clouds shared by nursing staff and nursing subjects within the same continuous point cloud segment decreases synchronously. The observability of the point cloud is determined by the lowest value among the components, and therefore decreases synchronously with the component with the largest decrease in point cloud evidence. The different variation amplitudes of the curves of different components indicate that point cloud sparsity, missing continuous point cloud frames, and occlusion of the nursing role interaction area have different effects on the trajectories of nursing staff point clouds, nursing subject point clouds, and nursing role point clouds. Therefore, Figure 4This indicates that the observability of point clouds is not a confidence level for nursing action recognition, but rather a point cloud evidence index jointly formed by the completeness of the point cloud of nursing staff, the completeness of the point cloud of nursing subjects, the continuity of the trajectory of nursing staff, the continuity of the trajectory of nursing subjects, and the coverage of nursing roles. It can quantitatively characterize the sufficiency of point cloud evidence related to nursing roles and provide a recordable and traceable point cloud evidence basis for the spatiotemporal recognition and representation of nursing behaviors and subsequent anomaly classification.

[0052] S3.5: Merge continuous action intervals in the joint recognition sequence of nursing actions and nursing object status, identify the changes in nursing object status corresponding to each nursing action, and write them into the corresponding point cloud observability level to form a spatiotemporal recognition representation of nursing behavior.

[0053] Specifically, the joint recognition sequence of nursing actions and patient states is arranged from earliest to latest according to the collection time. Adjacent entries with the same nursing action category and consecutive collection times are grouped into the same continuous action interval. When the nursing action category changes or the interval between adjacent collection times exceeds the window duration of the preset observation window, the continuous action interval is redefined. For each continuous action interval, the patient state change category, category judgment value, and collection time output by the PointNet++ and LSTM fusion model are read. Entries whose category judgment value reaches the recognition confidence threshold and whose duration reaches the state stability duration are identified as valid state change entries. The state stability duration is read from the state achievement time limit recorded in the quality control specification. If the quality control specification does not record it, it is determined based on the state stability duration in the manually reviewed samples. When there is more than one valid state change category, the patient state change category that last stabilized and continued until the end of the continuous action interval is selected as the associated state change. When there are no valid state change entries, the patient state change is marked as undetermined. The lowest observable point cloud level within the continuous action interval is read, and the associated state change and point cloud observability level are written into the continuous action interval to form a spatiotemporal recognition representation of nursing behavior.

[0054] It should be noted that the identification confidence threshold is set based on the consistency between the maximum class probability output by the PointNet++ and LSTM fusion model in the validation data and the manually reviewed labels. The maximum class probabilities in the validation data are arranged from low to high, and the median value of adjacent different maximum class probabilities is used as the threshold test point. The number of correct identifications, the number of incorrect identifications, and the number of unidentified items under each threshold test point are counted respectively. The threshold test point with the most correct identifications is selected as the identification confidence threshold. When the number of correct identifications is the same, the threshold test point with fewer incorrect identifications is selected. If they are still the same, the threshold test point with the higher value is selected. For example, the identification confidence threshold can be set to 0.60 to 0.85.

[0055] S4. Match the spatiotemporal identification representation of nursing behavior with the long-term care insurance service quality control task map, verify the compliance of actions and effective service duration according to nursing action nodes and connection relationships, form a nursing action compliance verification record, and verify the changes in the status of the nursing object according to the function status requirements, and generate a long-term care insurance service quality control judgment.

[0056] S4.1: Organize the spatiotemporal recognition representation of nursing behavior according to the collection time, merge the spatiotemporal recognition representation of adjacent nursing behaviors with the same nursing action category and continuous time into continuous nursing events, and re-divide the continuous nursing events when either the nursing action category changes or the time interval between adjacent collection times exceeds the window duration of the preset observation window, forming a continuous nursing event flow. Specifically, the continuous action intervals in the spatiotemporal recognition representation of nursing behavior are read and arranged from earliest to latest according to the start time of the continuous action intervals. When the nursing action categories of adjacent continuous action intervals are the same, and the time interval between the start time of the later continuous action interval and the end time of the earlier continuous action interval does not exceed the window duration of the preset observation window, the adjacent continuous action intervals are classified into the same continuous nursing event. The earliest start time and the latest end time in the continuous nursing event are respectively used as the start time and end time of the continuous nursing event, and the lowest observability of the point cloud in the contained continuous action intervals is written as the observability of the point cloud of the continuous nursing event. When the nursing action category changes or the time interval exceeds the window duration of the preset observation window, the later continuous action interval is classified into a new continuous nursing event and arranged according to the start time to form a continuous nursing event flow.

[0057] S4.2: Match the continuous nursing event flow with the long-term care insurance service quality control task graph, read the nursing action category, start and end time and point cloud observability from the continuous nursing event flow, and compare it with nursing action nodes, duration requirements and node connection relationships to calculate the alignment cost. Then, use dynamic time warping to generate candidate alignment paths and determine the candidate alignment path with the lowest cumulative alignment cost as the target alignment path. Specifically, the nursing action nodes corresponding to the current service item are arranged into a first sequence according to their node connection relationships. When there are branches in the node connection relationships, the first sequence is formed according to each allowed branch. Nursing action categories, start times, end times, and point cloud observability are read from the continuous nursing event stream and arranged into a second sequence according to the start time from earliest to latest. Nursing action nodes in the first sequence are compared one by one with continuous nursing events in the second sequence. A category consistency cost is set when the nursing action categories are consistent, and a category inconsistency cost is set when the nursing action categories are inconsistent. A duration consistency cost is set when the duration of the continuous nursing event meets the duration requirement, and a duration non-consistency cost is set when it does not. A corresponding point cloud evidence cost is also set based on the point cloud observability, forming an alignment cost between each nursing action node and each continuous nursing event. During the candidate alignment path extension process, priority is given to selecting... Select a pair that matches the nursing action category and the nursing action node. If no extendable pair with the same nursing action category exists, retain the pair with inconsistent nursing action categories and write an inconsistent category identifier. Establish a cumulative alignment cost matrix according to dynamic time warping. Select the position with the lowest cumulative alignment cost from the allowed positions on the left, top, and upper left to extend the current pair. Restrict the nursing action node to either maintain the current nursing action node or move to the subsequent nursing action node recorded in the node connection relationship. Backtrack from the end position of the cumulative alignment cost matrix along the direction of the lowest cumulative alignment cost to form a candidate alignment path. Compare the number of inconsistent category identifiers from all candidate alignment paths and determine the candidate alignment path with the fewest inconsistent category identifiers as the priority path. If there are multiple priority paths, select the candidate alignment path with the lowest cumulative alignment cost from the multiple priority paths as the target alignment path.

[0058] The expression for calculating the alignment cost is: ; ; ; in, Indicates the first The nursing action node and the first The alignment cost between consecutive care events ranges from zero to one. Indicates the first The nursing action node and the first The cost of nursing action categories between consecutive nursing events is zero when the nursing action categories are the same and one when they are different. Indicates the first The duration of the first consecutive nursing event relative to the first The time deviation cost generated by the time requirement configured for each nursing action node ranges from zero to one. Indicates the first The cost of point cloud evidence corresponding to a series of nursing events; Indicates the first The minimum duration requirement for configuring each nursing action node; Indicates the first The duration of a series of nursing events is determined by the termination time and the start time; Indicates the first The maximum duration requirement for each nursing action node; Indicates the first The observability of the point cloud corresponding to a continuous nursing event ranges from zero to one. This indicates the number of nursing action nodes arranged according to their connection relationships; This indicates the sequence number of consecutive nursing events arranged according to their start time.

[0059] It should be noted that when the continuous nursing event flow is empty, a cumulative alignment cost matrix is ​​not established. All nursing action nodes corresponding to the current service item are marked as not synchronized and the service start time to service end time recorded in the long-term care insurance service order are determined as the corresponding time period.

[0060] S4.3: The time interval of nursing events in the target alignment path that is consistent with the nursing action category and nursing action node and whose node transfer conforms to the node connection relationship is determined as the synchronous alignment period. The nursing action sequence is verified according to the node connection relationship, and the synchronous alignment period whose duration meets the corresponding duration requirement is accumulated as the effective service time, forming a nursing action compliance verification record.

[0061] Specifically, pairs of nursing action categories matching nursing action nodes are extracted from the target alignment path and arranged from earliest to latest according to the start time of the corresponding nursing events. If the nursing action node corresponding to the first pair is a starting nursing action node recorded in the long-term care insurance service quality control task map, an action sequence compliance flag is written; otherwise, an action sequence deviation flag is written. For subsequent pairs, it is verified whether the transfer from the nursing action node corresponding to the previous pair to the nursing action node corresponding to the current pair constitutes a node connection relationship. If it does, an action sequence compliance flag is written, and the nursing event time interval is determined as the synchronous alignment period; otherwise, an action sequence deviation flag is written. Spatial posture features of nursing staff and nursing personnel are extracted from the spatiotemporal recognition representation of nursing behavior corresponding to the synchronous alignment period. The system analyzes the spatial posture characteristics of the nursing object and the interaction characteristics of the nursing role, and extracts the requirements for the nurse's body orientation, limb posture, relative position of the nursing role, contact point, and direction of movement from the nursing action execution requirements recorded in the corresponding nursing action nodes. The extracted spatial posture characteristics and nursing role interaction characteristics are compared with the corresponding requirements item by item. If all meet the requirements, a nursing posture compliance mark is written. If there are any non-compliance items, a nursing posture deviation mark is written. The system checks whether the duration of each synchronization alignment period meets the duration requirement configured for the corresponding nursing action node. If the duration of the synchronization alignment period reaches the lower limit of the duration requirement configured for the corresponding nursing action node, a valid service time standard mark is written. The synchronization alignment periods with the valid service time standard mark are written to the corresponding nursing action node and summarized into a valid service time, forming a nursing action compliance verification record.

[0062] S4.4: For each nursing action node in the nursing action compliance verification record, set a status response window after the synchronization alignment period ends, according to the status achievement time limit in the nursing object's function status requirement, and match the nursing object status change in the status response window with the nursing object's function status requirement to form a nursing action status verification record. Specifically, the system reads the nursing object's status change category during the synchronization alignment period and extracts the changed status recorded in the nursing object's status change category; it reads the nursing object's target status maintenance category from the status response window after the synchronization alignment period ends, and verifies whether the changed status was formed and maintained within the status achievement time limit; if the changed status during the synchronization alignment period is consistent with the target status recorded in the nursing object's action status requirement, and the target status is maintained in the status response window, a nursing object action status establishment flag is written; if no target status change is identified, the target status is not formed within the status achievement time limit, or the target status is not maintained, a nursing object action status not established flag is written, forming a nursing action status verification record.

[0063] It should be noted that nursing action nodes that do not have synchronized time periods will not have a status response window set, and the unsynchronized alignment indicator will be retained. The time limit for achieving the desired state is derived from the state effectiveness requirements clearly recorded in the quality control specifications issued by the long-term care insurance authorities or agencies for each nursing action. It is also determined in conjunction with the usual time span from change to stability of the patient's state after the corresponding nursing action is performed. For example, the state requirement for assisting in turning over is that the patient changes from supine to lateral and remains in that position. The time limit for achieving this state can be set to one to three seconds after the action ends. The state requirement for assisting in getting up is that the patient's torso changes from supine to sitting and remains stable. The time limit for achieving this state can be set to two to four seconds after the action ends.

[0064] S4.5: Based on the nursing action compliance verification record and nursing action status verification record, identify nursing action nodes that deviate from the action sequence, have insufficient effective service time, or whose nursing object's role status is not established, and write corresponding verification identifiers. Record the location of nursing action nodes that are not synchronized and the observability of the point cloud for the corresponding time period, and generate a long-term care insurance service quality control judgment.

[0065] Specifically, based on the action sequence verification status of each nursing action node in the nursing action compliance verification record, an action sequence deviation verification flag is written when the action sequence verification status is "deviation"; the duration of each synchronization alignment period is read, and an insufficient effective service duration verification flag is written when the duration does not reach the minimum duration requirement configured for the corresponding nursing action node; the nursing action status verification record is read, and an incomplete nursing object status verification flag is written when the nursing object's role status is not established; nursing action nodes that have not formed a synchronization alignment period are extracted from the target alignment path, and the position of the nursing action node in the long-term care insurance service quality control task map is recorded. The corresponding time period is determined by the end time of the previous synchronization alignment period and the start time of the next synchronization alignment period. When it is before the first synchronization alignment period, the service start time recorded in the long-term care insurance service order is used as the starting point of the corresponding time period. When it is after the last synchronization alignment period, the service end time recorded in the long-term care insurance service order is used as the end point of the corresponding time period. The observability of the point cloud of the corresponding time period is read from the spatiotemporal recognition representation of nursing behavior. The verification flags, nursing action node positions, and point cloud observability are collected to generate a long-term care insurance service quality control judgment.

[0066] It should be noted that the upper limit of duration requirement is used to avoid obtaining a lower alignment cost during the dynamic time warping process for consecutive nursing events with abnormally long durations, and is not used as a criterion for determining insufficient effective service duration; when the quality control specification does not record the upper limit of duration requirement, the calculation branch exceeding the upper limit of duration requirement will not be executed for the duration deviation cost; when the lower limit of duration requirement is not set or is not greater than zero, or when the upper limit of duration requirement is lower than the lower limit of duration requirement, the corresponding duration requirement will be marked as a configuration abnormal and the duration verification of the corresponding nursing action node will be stopped.

[0067] S5. Based on the non-synchronized alignment of nursing action nodes, the observability of point cloud, and the changes in the status of the nursing subjects in the quality control judgment of long-term care insurance services, construct the assumptions that the nodes have been executed and the assumptions that the nodes have not been executed, reconstruct the synchronization alignment period and verify the consistency of the action status, distinguish between nursing action identification omissions, service action missingness and insufficient point cloud evidence, determine the source of abnormality and the level of nursing operation standardization, and form a long-term care insurance service quality control file.

[0068] S5.1: Extract the non-synchronized nursing action nodes, adjacent synchronized time periods, observability of point cloud, and changes in the status of the nursing object from the quality control judgment of long-term care insurance services. Construct the execution hypothesis and non-execution hypothesis of the node respectively. Fill the execution hypothesis into the corresponding node position and reconstruct the synchronized time period. Leave the non-execution hypothesis in the node vacancy. Verify the subsequent node connection relationship, effective service duration and function status requirements respectively to form dual hypothesis verification information. Specifically, based on the location of nursing action nodes, the system extracts non-synchronized nursing action nodes, adjacent synchronized periods, observability of point clouds, and changes in the state of the cared-for individual from the long-term care insurance service quality control assessment. For nodes that have been executed, the system retains the non-synchronized nursing action nodes and writes the nursing action category, duration requirement, effect status requirement, observability of point clouds, and changes in the state of the cared-for individual. For nodes that have not been executed, the system retains the location of the nursing action node, observability of point clouds, and changes in the state of the cared-for individual, while keeping nursing action nodes vacant. The system places the executed assumption at the corresponding nursing action node location, using the end time of the previous adjacent synchronized period as the starting point, and defines a synchronized period before the start time of the next adjacent synchronized period according to the corresponding duration requirement. The system connects the previous adjacent synchronized period, the defined synchronized period, and the next adjacent synchronized period according to the connection relationship between nursing action nodes. The system verifies the connection order of subsequent nursing action nodes under both the executed and non-executed assumptions, verifies the effective service duration based on the corresponding synchronized period, and checks the changes in the state of the cared-for individual against the corresponding effect status requirement, forming dual-assumption verification information.

[0069] S5.2: Based on the dual hypothesis verification information, if the observability of the point cloud does not reach the preset point cloud observability threshold, it is marked as insufficient point cloud evidence. If it reaches the preset point cloud observability threshold and the node has executed the hypothesis to restore the connection relationship and the consistency of the function state, it is marked as nursing action recognition omission. If the node has not executed the hypothesis to maintain subsequent alignment and the function state is not established, it is marked as service action missing, thus forming a nursing service abnormality classification record.

[0070] Specifically, based on the dual hypothesis verification information, the observability of the point cloud corresponding to the unsynchronized nursing action nodes is compared with the preset point cloud observability threshold. When the observability of the point cloud does not reach the preset point cloud observability threshold, the unsynchronized nursing action nodes are marked as having insufficient point cloud evidence. When the observability of the point cloud reaches the preset point cloud observability threshold, the connection relationship of subsequent nodes under the assumption that the node has executed, the effective service duration, and the changes in the state of the nursing object are verified. When the node has executed the assumption and restored the connection relationship, and the changes in the state of the nursing object meet the corresponding function state requirements, the unsynchronized nursing action nodes are marked as nursing action identification omissions. When the subsequent nursing action nodes under the assumption that the node has not executed remain synchronized and the changes in the state of the nursing object do not meet the corresponding function state requirements, the unsynchronized nursing action nodes are marked as service action missing, and an anomaly classification identifier is written according to the position of the nursing action node to form a nursing service anomaly classification record.

[0071] It should be noted that after the nursing service anomaly classification record is formed, a secondary anomaly verification is performed on the time period corresponding to the missing service action; when the nursing staff's trajectory is continuously within the perception area, and there is insufficient effective service time, deviation of nursing posture, or failure of the nursing object's action state, a suspected service perfunctory label is written; when the long-term care insurance service order records that the corresponding service item has been completed, and the observability of the point cloud reaches the point cloud observability threshold, but the corresponding nursing action, nursing role interaction process, and nursing object state change are not identified, a suspected false service label is written; the suspected service perfunctory label or suspected false service label is written as a secondary anomaly label into the nursing service anomaly classification record of the corresponding nursing action node, and associated with the corresponding verification time period, nursing staff trajectory, point cloud observability, and verification label; The observable threshold for point clouds is set based on the observable point cloud sequence for nursing roles (after manual review), the completeness of the point cloud for nursing staff, the completeness of the point cloud for nursing subjects, the continuity of the trajectories for nursing staff and nursing subjects, and the coverage of nursing roles. Under the condition that the millimeter-wave radar installation location, sensing area, and acquisition configuration remain consistent, the following criteria are selected: the number of nursing staff points and nursing subject points within the corresponding time period of the nursing action both reach the density threshold; the target point groups for nursing staff and nursing subjects in adjacent point cloud frames maintain trajectory connectivity; both nursing staff and nursing subjects have valid point cloud frames simultaneously; and the changes in nursing staff posture, nursing subject posture, and nursing role interaction are manually reviewed and aligned with the long-term care insurance service quality control. Continuous point cloud segments with consistent action content corresponding to nursing action nodes in the task graph are selected as sufficient point cloud evidence samples; continuous point cloud segments in which the number of nursing staff points or nursing object points within the corresponding time period of the nursing action do not reach the density threshold, the nursing staff trajectory or nursing object trajectory is interrupted, or the nursing staff and nursing object do not simultaneously have any of the valid point cloud frames are selected as insufficient point cloud evidence samples; the observability of point cloud for the two types of samples is read separately, and the boundary value that can distinguish the two types of samples is determined as the observability threshold of point cloud, and continuous point cloud segments corresponding to different nursing action categories, nursing object postures, and service site occlusion situations are used for verification. An exemplary point cloud observability threshold can be set to 0.6 to 0.8.

[0072] S5.3: Write the nursing service anomaly classification record into the long-term care insurance service quality control task map, take the node corresponding to the anomaly classification identifier as the candidate anomaly node, copy it to form a verification task map, replace the state change of the candidate anomaly node with the function state requirement along the node connection relationship, review the subsequent nodes, distinguish between independent anomaly sources and related anomalies, and form an anomaly source location record. Specifically, based on the location of nursing action nodes, the abnormal classification identifiers in the nursing service abnormality classification records are written into the corresponding nursing action nodes of the long-term care insurance service quality control task chart. Nursing action nodes with abnormal classification identifiers are identified as candidate abnormal nodes, and the long-term care insurance service quality control task chart with the abnormal classification identifiers written is copied to form a verification task chart. Candidate abnormal nodes are selected sequentially according to the node connection relationship, and the nursing object status change of the candidate abnormal node is replaced with the target status recorded in the corresponding action status requirement. Subsequent nursing action nodes are then reviewed along the node connection relationship. When the subsequent nursing action nodes maintain the connection relationship and the action status is established after replacement, the candidate abnormal node is identified as a unique... When an anomaly source is identified, the subsequent nursing action node that restores its function is determined as a related anomaly. If the function of the subsequent nursing action node is still not established after replacement, the corresponding subsequent nursing action node is retained as a candidate anomaly node and continues to be reviewed. When a candidate anomaly node is determined to be an independent anomaly source, the corresponding suspected service perfunctory label or suspected false service label is written into the anomaly source location record, and associated with the corresponding verification time period, nursing staff trajectory, point cloud observability, and verification label. When a candidate anomaly node is determined to be a related anomaly, the association between the secondary anomaly label and the independent anomaly source is retained, but the corresponding secondary anomaly label is not treated as an independent anomaly source, thus forming an anomaly source location record.

[0073] S5.4: Based on the independent abnormal sources and associated abnormal ranges in the abnormal source location record, combined with the long-term care insurance service quality control judgment, determine the nursing operation standard level according to the level judgment rules in the quality control specification, write the nursing operation standard level into the abnormal source location record and link it with the long-term care insurance service quality control judgment for filing, forming a long-term care insurance service quality control file.

[0074] Specifically, based on the anomaly source location record, independent anomalies and related anomalies are identified. The action sequence verification, effective service duration verification, and patient status verification are then performed according to the corresponding action node locations in the long-term care insurance service quality control assessment. From the quality control standards' level determination rules, the action sequence requirements, effective service duration requirements, patient status requirements, and permissible anomaly classifications for each nursing operation standard level are extracted. These are then checked level by level according to the level order recorded in the quality control standards. If all nursing action nodes within the independent anomaly source and related anomaly range meet the requirements corresponding to the current nursing operation standard level, the current nursing operation standard level is determined. If any nursing action nodes do not meet the requirements of the current nursing operation standard, the level is determined. When the requirements of the corresponding level are met, the process moves to the next nursing operation standard level for further verification. The determined nursing operation standard level is written into the anomaly source location record. Suspected perfunctory service indicators or suspected false service indicators corresponding to independent anomaly sources are identified as key review indicators. The key review indicators, corresponding nursing action nodes, verification time periods, observability of point clouds, nursing staff trajectory status, nursing action verification status, nursing role interaction verification status, changes in the status of nursing subjects, and service completion status are written into the anomaly source location record and archived in conjunction with the long-term care insurance service quality control judgment. For related anomaly nodes, only the corresponding independent anomaly source and related anomaly range are recorded, and they are not repeatedly written into the key review indicators, thus forming a long-term care insurance service quality control file.

[0075] It should be noted that the grading rules are the corresponding conditions for the nursing operation standard grading set by the quality control specifications for the verification status of nursing action nodes, including three levels: standard, basic standard, and non-standard. When all nursing action nodes have synchronous alignment marks, action sequence compliance marks, effective service time standard marks, and nursing patient status establishment marks, it is determined to be standard. When all nursing action nodes have synchronous alignment marks and action sequence compliance marks, but there are insufficient effective service time verification marks or nursing patient status establishment verification marks, it is determined to be basic standard. When there are non-synchronous alignment marks or action sequence deviation verification marks, it is determined to be non-standard. The corresponding conditions are checked in the order of non-standard, basic standard, and standard to determine the nursing operation standard grading.

[0076] This embodiment also provides a long-term care insurance service quality control system based on millimeter-wave point clouds, including: a service task module, used to collect long-term care insurance service orders and quality control specifications, write service locations and nursing roles into the service process, and convert quality control requirements extracted from the quality control specifications into nursing action nodes and connection relationships, while configuring duration requirements and function status requirements to construct a long-term care insurance service quality control task map; a point cloud perception module, used to activate millimeter-wave radar based on the service locations and nursing roles in the long-term care insurance service quality control task map, preprocess the continuously collected millimeter-wave point clouds and identify nursing staff and nursing subjects, and calculate the observability of the point clouds according to a preset observation window to form the observable point cloud time sequence of nursing roles; and a spatiotemporal recognition module, used to construct a PointNet++ and LSTM fusion model, inputting the observable point cloud time sequence of nursing roles into PointNet++ and LSTM. The LSTM fusion model extracts the spatial posture features of caregivers and patients, as well as the temporal features of nursing actions, and identifies changes in the patient's state to form a spatiotemporal representation of nursing behavior. The quality control verification module matches this spatiotemporal representation with the long-term care insurance service quality control task map, verifies the compliance and effective service duration of actions according to nursing action nodes and their connections, and verifies changes in the patient's state according to the requirements of the function state, generating a long-term care insurance service quality control judgment. The quality control archiving module extracts nursing action nodes, point cloud observability, and changes in the patient's state from the long-term care insurance service quality control judgment. Based on the point cloud observability and the requirements of the function state, it distinguishes between nursing action omissions, missing service actions, and insufficient point cloud evidence, forming a nursing service anomaly classification record. Combining node locations, it determines the nursing operation standard level and the source of the anomaly, forming a long-term care insurance service quality control archive.

[0077] In summary, this invention extracts the location and observability of nursing action nodes from the long-term care insurance service quality control judgment, and classifies non-synchronized nursing action nodes into service action missing and insufficient point cloud evidence, so that the spatiotemporal recognition representation of nursing behavior has a clear point cloud evidence basis; it confirms whether the nursing role point cloud and nursing role point cloud trajectory can support nursing action recognition based on the observability of the point cloud in the period to be verified, and can record the nursing operation standard level, the first abnormal node, the period to be verified and the abnormal classification label, so as to realize the node location, time period recording and evidence archiving of nursing abnormalities.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A quality control method for long-term care insurance services based on millimeter-wave point clouds, characterized in that, include: Collect long-term care insurance service orders and quality control standards, write service locations and nursing roles into the service process, and convert the quality control requirements extracted from the quality control standards into nursing action nodes and connection relationships. At the same time, configure duration requirements and function status requirements to construct a long-term care insurance service quality control task map. Based on the service locations and nursing roles in the long-term care insurance service quality control task map, millimeter-wave radar is activated to preprocess the continuously collected millimeter-wave point clouds and identify nursing staff and nursing subjects. At the same time, the observability of the point clouds is calculated according to the preset observation window to form the time sequence of observable point clouds for nursing roles. A PointNet++ and LSTM fusion model is constructed. The time sequence of observable point cloud of nursing roles is input into the PointNet++ and LSTM fusion model to extract the spatial pose features and time sequence features of nursing actions of nursing staff and nursing objects, and to identify the state changes of nursing objects, forming a spatiotemporal recognition representation of nursing behavior. The spatiotemporal identification and representation of nursing behavior is matched with the long-term care insurance service quality control task map. The compliance of the actions and the effective service duration are verified according to the nursing action nodes and connection relationships to form a nursing action compliance verification record. The changes in the status of the nursing object are verified according to the function status requirements to generate a long-term care insurance service quality control judgment. Based on the non-synchronized alignment of nursing action nodes, the observability of point clouds, and changes in the status of the nursing subjects in the quality control judgment of long-term care insurance services, we construct the assumptions that the nodes have been executed and the assumptions that the nodes have not been executed, reconstruct the synchronization alignment period and verify the consistency of the action status, distinguish between nursing action identification omissions, service action deficiencies, and insufficient point cloud evidence, determine the source of anomalies and the level of nursing operation standardization, and form a long-term care insurance service quality control file.

2. The long-term care insurance service quality control method based on millimeter-wave point clouds as described in claim 1, characterized in that, The steps for constructing the long-term care insurance service quality control task chart are as follows: Collect long-term care insurance service orders and quality control standards, analyze the service item execution relationship and arrange the service process, write the corresponding service location and nursing role, and form a long-term care insurance service process; According to the service items in the long-term care insurance service process, the corresponding quality control standards are retrieved, and the requirements for the execution of nursing actions, duration requirements, and the status requirements of the nursing subjects are extracted to form the service item quality control requirements; The nursing action execution requirements in the service project quality control requirements are converted into nursing action nodes one by one. The sequential constraints of the actions are extracted and the connection relationship between the nursing action nodes is established. The corresponding duration requirements and function status requirements are configured to the nursing action nodes to construct a long-term care insurance service quality control task map.

3. The long-term care insurance service quality control method based on millimeter-wave point clouds as described in claim 1, characterized in that, The steps for forming the observable point cloud sequence for the nursing role are as follows: According to the long-term care insurance service quality control task map, the service sites are matched with the corresponding millimeter-wave radars, the sensing areas are set up and a unified time and spatial coordinate reference is established. The corresponding millimeter-wave radars are started to continuously collect point cloud frames carrying the collection time and spatial attributes, and the original millimeter-wave point cloud sequence is formed by arranging them according to the collection time. Background reflection suppression, outlier removal and density filtering are performed on the original millimeter-wave point cloud sequence. Nursing roles are matched by target clustering and cross-frame tracking to identify nursing staff and nursing objects and form nursing role point cloud trajectories. The point cloud trajectory of the nursing role is divided into continuous point cloud segments according to the preset observation window. The observability of each continuous point cloud segment is calculated and written into the corresponding continuous point cloud segment to form the observable point cloud time sequence of the nursing role.

4. The long-term care insurance service quality control method based on millimeter-wave point clouds as described in claim 1, characterized in that, The steps for constructing the PointNet++ and LSTM fusion model are as follows: Collect the time sequence of observable point cloud of nursing roles that has been manually verified, and label the nursing action categories and nursing object status change categories according to the nursing action nodes and nursing object status requirements in the long-term care insurance service quality control task map to form a nursing behavior recognition training sample set; A local coordinate system is established using the centroid and body axis of the nursing object point cloud. The nursing staff point cloud and nursing object point cloud in the nursing behavior recognition training sample set are transformed into the local coordinate system. Spatial pose features are extracted by establishing the nursing staff PointNet++ branch and the nursing object PointNet++ branch respectively. The nursing role interaction features formed by the relative position and relative motion relationship between the nursing staff point cloud and the nursing object point cloud are fused with the observability of the corresponding point cloud to form a dual-role spatial training feature sequence. The dual-role space training feature sequence is input into the LSTM temporal feature extraction network. Nursing action recognition branch and nursing object state change recognition branch are set up. The model is trained according to the connection relationship between nursing action nodes, and a PointNet++ and LSTM fusion model is constructed.

5. The long-term care insurance service quality control method based on millimeter-wave point clouds as described in claim 4, characterized in that, The steps for forming a spatiotemporal recognition representation of nursing behavior are as follows: The observable point cloud of nursing roles is input into the PointNet++ and LSTM fusion model. Spatial pose features of nursing staff and nursing objects are extracted through the PointNet++ branches of nursing staff and nursing objects. The timing features of nursing actions are extracted through LSTM to form a joint recognition sequence of nursing actions and nursing object states. The joint recognition sequence of nursing actions and nursing object status is merged into continuous action intervals, the nursing object status changes corresponding to each nursing action are identified, and the corresponding point cloud observability is written into the form of spatiotemporal recognition representation of nursing behavior.

6. The long-term care insurance service quality control method based on millimeter-wave point clouds as described in claim 5, characterized in that, The steps for creating a compliance verification record for nursing actions are as follows: Organize the spatiotemporal recognition representation of nursing behaviors according to the collection time, merge adjacent spatiotemporal recognition representations of nursing behaviors with the same nursing action category and continuous time into continuous nursing events, and re-divide continuous nursing events when either the nursing action category changes or the time interval between adjacent collection times exceeds the window duration of the preset observation window, forming a continuous nursing event flow. The continuous nursing event flow is matched with the long-term care insurance service quality control task map. The nursing action category, start and end time and point cloud observability are read from the continuous nursing event flow and compared with the nursing action nodes, duration requirements and node connection relationship. The alignment cost is calculated and candidate alignment paths are generated by dynamic time warping. The candidate alignment path with the lowest cumulative alignment cost is determined as the target alignment path. The time interval of nursing events in the target alignment path that are consistent with the nursing action category and nursing action node and whose node transfer conforms to the node connection relationship is determined as the synchronous alignment period. The order of nursing actions is verified according to the node connection relationship, and the synchronous alignment period whose duration meets the corresponding time requirement is accumulated as the effective service time, forming a nursing action compliance verification record.

7. The long-term care insurance service quality control method based on millimeter-wave point clouds as described in claim 6, characterized in that, The steps for generating the quality control judgment for long-term care insurance services are as follows: For each nursing action node in the nursing action compliance verification record, a status response window is set after the synchronization alignment period ends, according to the status achievement time limit in the nursing object's function status requirement. The nursing object status change in the status response window is matched with the nursing object's function status requirement to form a nursing action status verification record. Based on the nursing action compliance verification record and nursing action status verification record, nursing action nodes with deviations in action sequence, insufficient effective service time, and invalid patient status are identified and corresponding verification identifiers are written. The location of nursing action nodes that are not synchronized and the observability of point cloud for the corresponding time period are recorded to generate a long-term care insurance service quality control judgment.

8. The long-term care insurance service quality control method based on millimeter-wave point clouds as described in claim 7, characterized in that, The steps for creating a nursing service exception classification record are as follows: From the quality control judgment of long-term care insurance services, we extract the nursing action nodes that are not synchronized, the adjacent synchronized time periods, the observability of point cloud and the changes in the status of the nursing object. We construct the node execution hypothesis and the node non-execution hypothesis respectively. We fill the node execution hypothesis into the corresponding node position and reconstruct the synchronized time period. We leave the node non-execution hypothesis with node vacancy. We verify the subsequent node connection relationship, effective service duration and function status requirements respectively to form dual hypothesis verification information. Based on the dual hypothesis verification information, when the observability of the point cloud does not reach the preset point cloud observability threshold, it is marked as insufficient point cloud evidence. When it reaches the preset point cloud observability threshold and the node has executed the hypothesis to restore the connection relationship and the consistency of the function state, it is marked as nursing action recognition omission. When the node has not executed the hypothesis to maintain subsequent alignment and the function state is not established, it is marked as service action missing, thus forming a nursing service abnormality classification record.

9. The long-term care insurance service quality control method based on millimeter-wave point clouds as described in claim 8, characterized in that, The steps for creating a long-term care insurance service quality control file are as follows: Write the nursing service anomaly classification record into the long-term care insurance service quality control task map, use the node corresponding to the anomaly classification identifier as the candidate anomaly node, copy it to form a verification task map, replace the state change of the candidate anomaly node with the function state requirement along the node connection relationship, review the subsequent nodes, distinguish between independent anomaly sources and related anomalies, and form an anomaly source location record. Based on the independent abnormal sources and associated abnormal ranges in the abnormal source location record, and in conjunction with the long-term care insurance service quality control judgment, the nursing operation standard level is determined according to the level judgment rules in the quality control specification. The nursing operation standard level is written into the abnormal source location record and archived in conjunction with the long-term care insurance service quality control judgment to form a long-term care insurance service quality control file.

10. A long-term care insurance service quality control system based on millimeter-wave point clouds, based on the long-term care insurance service quality control method based on millimeter-wave point clouds as described in any one of claims 1 to 9, characterized in that, It includes a service task module, a point cloud perception module, a spatiotemporal recognition module, a quality control verification module, and a quality control archiving module; The service task module is used to collect long-term care insurance service orders and quality control specifications, write service locations and nursing roles into the service process, and convert the quality control requirements extracted from the quality control specifications into nursing action nodes and connection relationships. At the same time, it configures duration requirements and function status requirements to construct a long-term care insurance service quality control task diagram. The point cloud perception module is used to activate millimeter-wave radar based on the service location and nursing role in the long-term care insurance service quality control task map, preprocess the continuously collected millimeter-wave point cloud and identify nursing staff and nursing objects, and calculate the observability of the point cloud according to the preset observation window to form the observable point cloud time sequence of nursing roles. The spatiotemporal recognition module is used to construct a PointNet++ and LSTM fusion model. The time sequence of the observable point cloud of nursing roles is input into the PointNet++ and LSTM fusion model to extract the spatial posture features and time sequence features of nursing actions of nursing staff and nursing objects, and to identify changes in the state of nursing objects, forming a spatiotemporal recognition representation of nursing behavior. The quality control verification module is used to match the spatiotemporal identification representation of nursing behavior with the long-term care insurance service quality control task map, verify the compliance of actions and effective service duration according to nursing action nodes and connection relationships, and verify the changes in the state of the nursing object according to the function state requirements, and generate a long-term care insurance service quality control judgment. The quality control archiving module is used to extract nursing action nodes, observability of point clouds, and changes in the status of nursing subjects from the quality control judgment of long-term care insurance services. Based on the observability of point clouds and the requirements of the function status, it distinguishes between nursing action omissions, service action deficiencies, and insufficient point cloud evidence, forming a nursing service anomaly classification record. Combined with the node position, it determines the nursing operation standard level and the source of the anomaly, forming a long-term care insurance service quality control archive.

Images