Integrated old-age care management system based on end-cloud collaboration

Abstract

The invention relates to the technical field of old-age service management, in particular to an integrated old-age service management system based on end-cloud collaboration, and aims to solve the problems that in the prior art, multi-source data of physical signs, medication and living habits cannot be integrated to predict the health change trend, abnormal fluctuation of four high cannot be perceived in advance, short-term risks cannot be pre-judged, and the service quality is poor. And a personalized intervention scheme cannot be formulated in combination with reinforcement learning. Multi-source data of physical signs, medication and living habits are integrated through the chronic disease management module, LSTM is used for predicting the health change trend, four-high abnormal fluctuation is perceived in advance, short-term risks are pre-judged, a personalized intervention scheme is formulated in combination with reinforcement learning, diet and exercise suggestions and medication reminding are covered, and the system has the advantages of being simple in structure and convenient to use. User compliance is improved by outputting custom prompts through natural languages, the system forms a complete closed loop, the system is suitable for long-term management of diabetes and hypertension diseases, the risk of acute events is reduced, and the self-management ability of patients is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pension service management, more particularly to an integrated all-in-one pension management system based on end-cloud collaboration. BACKGROUND

[0002] Traditional pension services have problems such as information islands, delayed responses, and resource mismatches. Health monitoring functions cannot be linked with emergency rescue systems. Home-based pension services are inefficient due to manual scheduling. Community pension institutions lack data-driven precision service capabilities. Although some existing systems use cloud platform architectures, terminal devices and cloud collaboration are insufficient, privacy protection mechanisms are not sound, and service resource scheduling lacks intelligent algorithm support.

[0003] Patent application CN119833136A discloses a smart pension service management system based on the Internet of Things, which includes a health detection module for using a sensor group and intelligent wearable devices to collect real-time physiological data of the elderly and add timestamp information to the collected data. A communication module is used to transmit and correct the collected data through a wireless local area network and 5G technology. An alarm module is used to monitor physiological data that exceeds the preset normal range. An information storage and health analysis module is provided. In the present application, data is transmitted synchronously through a wireless local area network and a 5G network. The data correction unit compares the data errors of the two networks and performs linear correction. This ensures the timeliness and stability of data transmission, meets different scene requirements, improves data accuracy, provides reliable basis for subsequent health analysis and decision-making, and effectively improves the precision and adaptability of the smart pension system in the data flow link. However, the above-mentioned reference patent optimizes data transmission, health analysis, and emergency response through end-cloud collaboration and intelligent algorithms, and comprehensively improves the data accuracy, service refinement, and emergency response capabilities of the smart pension system in multiple scenarios. However, it cannot integrate multi-source data such as vital signs, medication, and lifestyle habits to predict health trends, cannot detect abnormal fluctuations in four high-risk factors and predict short-term risks in advance, cannot develop personalized intervention plans based on reinforcement learning, cannot generate personalized training paths based on user age, physical ability, and chronic diseases, cannot accurately limit training intensity, cannot predict fatigue based on heart rate and movement trends, cannot adjust the intensity or prompt rest in time, and cannot improve rehabilitation efficiency and safety.

[0004] Therefore, we propose an integrated all-in-one pension management system based on end-cloud collaboration to address the above problems. SUMMARY

[0005] The application aims to provide an integrated end-cloud collaborative pension management system, which solves the problems of the prior art, such as inability to integrate multi-source data of signs, medication and living habits to predict health change trends, inability to detect abnormal fluctuations in four high values in advance and predict short-term risks, inability to develop individual intervention programs in combination with reinforcement learning, inability to generate individual training paths according to user age, physical ability and chronic diseases, inability to accurately limit training intensity, inability to predict fatigue according to heart rate and action trends, inability to adjust the degree of treatment or prompt rest in time, and reduction of rehabilitation efficiency and safety.

[0006] The application aims to provide an integrated end-cloud collaborative pension management system, which solves the problems of the prior art, such as inability to integrate multi-source data of signs, medication and living habits to predict health change trends, inability to detect abnormal fluctuations in four high values in advance and predict short-term risks, inability to develop individual intervention programs in combination with reinforcement learning, inability to generate individual training paths according to user age, physical ability and chronic diseases, inability to accurately limit training intensity, inability to predict fatigue according to heart rate and action trends, inability to adjust the degree of treatment or prompt rest in time, and reduction of rehabilitation efficiency and safety. The integrated end-cloud collaborative pension management system is applied to a pension management platform and comprises: A data acquisition and processing module is configured to acquire health monitoring data of an elderly user and perform preprocessing operations on the acquired health monitoring data. A chronic disease management module is configured to construct an individual health trajectory model based on the acquired health monitoring data, generate an intervention program in combination with dynamic risk prediction and reinforcement learning, and perform semantic mapping and natural language output based on a medical knowledge graph. A smart safety monitoring module is configured to acquire camera video streams, inertial sensor action data, door magnetic switch states and electrical appliance use timing data, detect falls, stillness and abnormal behaviors through an action recognition model and event fusion logic, and trigger multi-level early warning responses. A rehabilitation therapy module is configured to generate a rehabilitation training path according to user physical ability scores, chronic disease states and age, adjust training content through a graph optimization algorithm, evaluate actions in combination with posture recognition and IMU tracking, and predict fatigue states based on heart rate and action trends.

[0007] As a preferred embodiment of the application, the process of constructing an individual health trajectory model by the chronic disease management module based on the acquired health monitoring data comprises: The preprocessed health monitoring data is acquired, and the health monitoring data comprises physiological data, environmental data and daily behavior data, the physiological data comprises heart rate, blood pressure and blood oxygen saturation, the environmental data comprises temperature and air quality PM2.5, and the daily behavior data comprises step count and sleep quality. The standardized health data is modeled using a multivariate LSTM network, the model receives a historical health data window with a length of W days, each window contains D health indicators, and the LSTM unit outputs the hidden state at the current time . The LSTM model processes the input data at each time step, updates the internal state, and outputs the predicted values of each health indicator in the future T days after completing the processing of the entire window.

[0008] In a preferred embodiment of the present invention, the process by which the chronic disease management module combines dynamic risk prediction and reinforcement learning to generate intervention plans and performs semantic mapping and natural language output based on a medical knowledge graph includes: Based on the future T-day prediction results output by the LSTM model, an MLP model is constructed for risk prediction. The MLP model contains three fully connected layers. The input is the future T-day health indicator sequence output by the LSTM model, and the output is the daily health risk score for the future T days. , t∈{1,2,...,T}; The intervention strategy learning is carried out by using DoubleDQN combined with a priority experience replay mechanism. The state is composed of health indicators at the current and future T time steps. The action space consists of a set of discrete actions, each action representing a specific intervention behavior. The action numbers range from 0 to N-1, where N is the total number of actions. The reward value is composed of two factors: changes in health risks and the burden of interventions; Using the DoubleDQN strategy, the update formula is as follows: ; in It is 0.001. The value is 0.95, where s represents the state, a represents the action, and r represents the immediate reward. For the next state, For the target network; The target network parameters are synchronized from the main network every 100 training cycles. At each moment, the DQN network outputs the Q value of each action in the current state and selects the action with the largest Q value as the intervention suggestion. The reinforcement learning process is executed cyclically, and each execution includes the following steps: S1: Get the current user status; S2: Input the state into the DQN network, output the Q value corresponding to each action, and select the action corresponding to the largest Q value; S3: Apply the selected action and observe the next state, record the experience and update the network, then proceed to the next round of judgment; S4: Repeat S1-S3 to gradually generate intervention strategies that adapt to the current user state; The GPT-2 model is used to generate natural language suggestions. The input is the selected intervention action number and the user's basic information. The output is a complete natural language suggestion text.

[0009] In a preferred embodiment of the present invention, the process by which the intelligent safety monitoring module processes camera video streams, inertial sensor motion data, door magnetic switch status, and appliance usage timing data includes: The video data is acquired using a network high-definition camera, and the camera outputs a sequence of image frames as wherein each frame represents a fixed-resolution image matrix with dimensions , wherein H is the image height, W is the image width, and C is the number of color channels; An IMU module with a built-in three-axis accelerometer and a three-axis gyroscope is used; The three-axis acceleration vector is represented as ; The three-axis angular velocity vector is represented as ; The total acceleration value is obtained by taking the square root of the sum of the squares of the three directional accelerations; A wireless door magnetic sensor is used to detect the state of the door, and the state of the door is divided into two types: open or closed; The power P(t) of the electrical appliance is collected through the intelligent power metering socket, and the data of each device is composed of a timestamp and a power value , forming a time series ; The total power of the electrical appliance is the sum of the power of all connected devices. When the power changes from 0 to a certain set value, the power-on event is recorded, and when the power drops to 0, the power-off event is recorded.

[0010] As a preferred embodiment of the present application, the process of detecting falls, stillness and abnormal behaviors and triggering multi-level early warning responses by the intelligent safety monitoring module through the action recognition model and event fusion logic includes: A three-dimensional convolutional neural network is used to model the image sequence, and the model output is a behavior category, taking one of walking, stillness or falling as the value; The IMU data is analyzed using a sliding window, and the acceleration and angular velocity data are input into a preprocessing module to generate a standardized feature vector x. The feature vector x is input into a pre-trained SVM classifier to output a behavior category C, taking {still, walk, fall} as the value; A Bayesian network is used to fuse the three types of information of video, IMU and door magnetic, and the fusion result is a comprehensive judgment value, which is used to determine the final behavior category; A behavior baseline distribution μ(t) is constructed for each user, and the mean and standard deviation σ of the behavior indicators in the last k days are calculated based on a sliding window; If the absolute value of the difference between a behavior indicator and the mean exceeds three times the standard deviation, it is marked as an abnormal event; When the system detects that the user has not moved for a long time, and the acceleration change collected by the IMU sensor is lower than the set motion threshold, and the video analysis module does not identify valid activity behavior, and the current time is within the user's daily activity range, the system will trigger a first-level early warning response; When the IMU model determines a fall behavior, or the video model identifies a fall feature, and the confidence output by the fusion model does not reach the final determination standard, the system will trigger a secondary early warning response; When the fall probability output by the fusion model exceeds the system-set determination threshold, the system will trigger a tertiary early warning response.

[0011] As a preferred embodiment of the present application, the process of the rehabilitation physiotherapy module acquiring the user's physical fitness score, chronic disease status and age includes: Acquiring the user's physical fitness score, chronic disease status and age, and constructing an individualized training model through the user's physical fitness score, chronic disease status and age; According to the user's age value, a reference value for setting the upper limit of training intensity is calculated, which is obtained by subtracting the user's age from 220, and the result is taken as the maximum heart rate value of the user; The presence status of three chronic diseases is recorded: cardiovascular disease, diabetes, and osteoporosis. The status of each disease is identified in binary form: if a certain disease exists, it is represented by the number 1, and if a certain disease does not exist, it is represented by the number 0. Number 1 corresponds to cardiovascular disease, number 2 corresponds to diabetes, and number 3 corresponds to osteoporosis; The physical fitness score is composed of three indicators: cardiorespiratory endurance, muscle strength, and joint mobility. The final score is obtained by function fusion.

[0012] As a preferred embodiment of the present application, the process of the rehabilitation physiotherapy module generating a rehabilitation training path and adjusting the training content through a graph optimization algorithm includes: The training path is composed of multiple action nodes, each node representing a specific rehabilitation action. The system selects a set of actions from the pre-set action library to form an initial training path; Each dimension is weighted and summed according to a fixed weight to generate a target function value. The system sorts according to the target function value and selects the top n actions to form the initial training path; The training path is modeled as a weighted directed graph, with each node representing a training action and the edge representing the possibility of transitioning from one action to another; The system defines a cost function, considering the risk increment, intensity influence and recovery time between actions, and uses A* algorithm to search for the optimal path, so that all action combinations in the path meet the minimum cost condition.

[0013] As a preferred embodiment of the present application, the process of the rehabilitation physiotherapy module combining posture recognition and IMU tracking to evaluate actions and predicting fatigue status based on heart rate and action trend includes: The system combines video pose recognition with IMU data to evaluate the quality of user actions in real time. Pose recognition is achieved by extracting the coordinates of key points on the human body, defining standard actions as a set of key points. Given that the coordinates of each keypoint are (x, y, z), calculate the Euclidean distance between the current action and the standard action: ; in To represent the deviation of the i-th key point, To represent the coordinates of the key points of the current action, The coordinates of key points representing standard movements; Sum the deviations of all key points to obtain the overall deviation of the action: ; If the value exceeds the preset threshold, the action is deemed to have failed to meet the standard. By collecting heart rate and movement trend data in real time, the system predicts the user's fatigue state and adjusts the training intensity accordingly. It collects the current heart rate and resting heart rate, calculates the heart rate change rate, collects the action execution time, compares it with the recommended execution time, and calculates the action trend fatigue index. When the index exceeds the set threshold, the system reduces the intensity of subsequent actions or prompts the user to rest.

[0014] Compared with the prior art, the advantages of this invention are: In this invention, a chronic disease management module integrates multi-source data on vital signs, medication, and lifestyle habits. LSTM is used to predict health trends, detect abnormal fluctuations in the four highs (hypertension, hyperlipidemia, and hyperglycemia) in advance, and predict short-term risks. Reinforcement learning is combined to develop personalized intervention plans, covering dietary and exercise recommendations and medication reminders. Natural language output provides easy-to-understand prompts to improve user compliance. The system forms a complete closed loop, which is suitable for the long-term management of diabetes and hypertension, reduces the risk of acute events, and enhances patients' self-management ability. In this invention, a personalized training path is generated based on the user's age, physical fitness, and chronic diseases through a rehabilitation therapy module. A graph optimization algorithm is used to dynamically adjust the content and sequence of movements. The training intensity is limited by the maximum heart rate. An initial path is constructed by combining movement scores, and the execution sequence is optimized using the A* algorithm. The system integrates video and IMU data to evaluate movement quality, calculate deviation and stability, and judge the execution effect in real time. At the same time, fatigue is predicted based on heart rate and movement trends, and the treatment intensity is adjusted or rest is prompted in a timely manner to improve rehabilitation efficiency and safety. Attached Figure Description

[0015] Figure 1 This is a system block diagram of Embodiment 1 of the present invention; Figure 2 This is a system block diagram of Embodiment 2 of the present invention; Figure 3The step flowchart for the reinforcement learning process cycle in the application; Figure 4 The system block diagram of Example Three in the application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, and not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the application.

[0017] Example One: As shown in Figure 1 , Figure 3 and Figure 4 indicated, the integrated one-piece old-age management system based on end-cloud cooperation provided by the application is applied to an old-age management platform, and includes: a data acquisition and processing module, configured to acquire health monitoring data of an old user, and perform preprocessing operations on the acquired health monitoring data, the preprocessing operations including anomaly detection, filtering and denoising, time alignment, data down-sampling, and format standardization; By performing the preprocessing operations of anomaly detection, filtering and denoising, time alignment, data down-sampling, and format standardization on the acquired health monitoring data, the data quality is effectively improved, the sensor errors and environmental interference are eliminated, and the data accuracy is ensured. The time alignment mechanism is used to realize the time consistency of multi-source data, and support cross-device joint analysis. The data down-sampling reduces the amount of redundant information transmission and storage, and improves the system operation efficiency.

[0018] a chronic disease management module, configured to construct an individual health trajectory model based on the acquired health monitoring data, generate an intervention scheme by combining dynamic risk prediction and reinforcement learning, and perform semantic mapping and natural language output based on a medical knowledge graph; The process of constructing the individual health trajectory model by the chronic disease management module based on the acquired health monitoring data includes: acquiring the preprocessed health monitoring data, the health monitoring data including physiological data, environmental data, and daily behavior data, the physiological data including heart rate, blood pressure, and blood oxygen saturation, the environmental data including temperature and air quality PM2.5, and the daily behavior data including step count and sleep quality; using a multivariate LSTM network to model the standardized health data, the model receiving a historical health data window with a length of W days, each window containing D health indicators, and the LSTM unit outputting a hidden state at a current time , which is calculated as follows: ; wherein The output gate, The cell state, represents element-wise multiplication; The LSTM model processes the input data at each time step and updates the internal state. After completing the entire window processing, the model outputs the predicted values of the health indicators for the next T days. The prediction results serve as the input basis for the next stage of risk assessment; The process of generating intervention plans based on dynamic risk prediction and reinforcement learning and performing semantic mapping and natural language output based on the medical knowledge graph includes: Based on the prediction results of the next T days output by the LSTM model, an MLP model is constructed for risk prediction. The MLP model contains three fully connected layers, each using a ReLU activation function. The input is the sequence of health indicators for the next T days output by the LSTM model, with a dimension of TxD. The output is the daily health risk score for the next T days , t∈{1,2,...,T}, the risk score is directly calculated by the model, without subjective judgment or probability interpretation; DoubleDQN is used to learn intervention strategies combined with the priority experience replay mechanism. The state is composed of the current and future T time steps of health indicators. Each state is a fixed-length vector containing the predicted values of all health indicators within the specified time range; The action space is composed of a set of discrete actions, each representing a specific intervention behavior. The action number ranges from 0 to N−1, where N is the total number of actions. Each action is determined before training and is not allowed to be modified at runtime; The reward value is composed of two factors: the change in health risk and the burden brought by the intervention measures, with a weight setting of , ; DoubleDQN strategy is used, and the update formula is as follows: ; where is 0.001, is 0.95, s is the state, a is the action, r is the immediate reward, is the next state, is the target network; The experience pool capacity is set to 10000, and TDerror is used to calculate the priority for high-error experience; The target network parameters are synchronized from the main network once every 100 training cycles. At each time, the DQN network outputs the Q values of each action under the current state, and the action corresponding to the maximum Q value is selected as the intervention suggestion; The reinforcement learning process is executed in a loop, and each execution includes the following steps: S1: Obtain the current user state; S2: Input the state to the DQN network, output the Q value corresponding to each action, and select the action corresponding to the maximum Q value; S3: Apply the selected action and observe the next state, record the experience and update the network, and enter the next round of judgment; S4: Repeat S1-S3 to gradually generate an intervention strategy that adapts to the current user state; Use the GPT-2 model to generate natural language suggestions, input the selected intervention action number and the user's basic information, and output a complete natural language suggestion text; The basic structure is as follows: "According to the current health status, it is recommended to take the following intervention measures: XXX.", where "XXX" is automatically generated by the model according to the action number and context, and the text needs to be reviewed and confirmed before output; Support interactive dialogue mechanism, adjust the suggestion content according to the user feedback, the generation process is as follows: Receive the selected intervention action number; Get the user's age, gender and basic disease information; Build input prompt words; Call the GPT-2 model to generate suggestion text; Output the suggestion and wait for user feedback; If the user has questions or denies the suggestion, a new text will be generated; If the user has no objection, this round of suggestion generation is completed; Through the integration of multiple sources of data such as signs, medication and lifestyle habits by the chronic disease management module, the LSTM is used to predict the trend of health changes, to detect abnormal fluctuations in four high (high blood sugar, high blood pressure, high blood lipids and high uric acid) and to predict short-term risks in advance, to improve the accuracy of evaluation; DoubleDQN reinforcement learning is used to dynamically generate individualized intervention strategies, to enhance adaptability and practicality; Combined with the natural language model output structure, the interactive feedback mechanism is supported, and the user's understanding and acceptance are improved; The overall process is data-driven and closed-loop controllable, with good scalability, suitable for long-term chronic disease management.

[0019] Intelligent safety monitoring module, used for collecting camera video stream, inertial sensor action data, door magnetic switch state and electric appliance use timing data, detecting falls, stillness and abnormal behaviors through action recognition model and event fusion logic, triggering multi-level warning response; The process of the intelligent safety monitoring module processing camera video stream, inertial sensor action data, door magnetic switch state and electric appliance use timing data includes: Use a network HD camera to obtain video data, and the camera output image frame sequence is where each frame represents an image matrix with a fixed resolution, whose dimension is , where H is the image height, W is the image width, and C is the number of color channels. The video data is decoded by the local edge device and cached in the local encrypted storage. The transmission protocol uses RTSP, and the communication is based on the local area network. An IMU module with a built-in three-axis accelerometer and three-axis gyroscope is used, with a sampling frequency of 100 Hz. The three-axis acceleration vector is represented as . The three-axis angular velocity vector is represented as . The total acceleration value is obtained by taking the square root of the sum of the squares of the three directional accelerations. The sensor sends data to the central control module through the BLE protocol, and the receiving end uses a complementary filter to perform time synchronization and attitude estimation on the data. A wireless door magnetic sensor is used to detect the state of the door. The door magnetic sensor detects the state of the door once per second and uploads event records when the state changes. The state of the door is divided into two categories: open or closed. The system assigns a unique identifier to each door magnetic device to ensure that events can be traced. The power of electrical appliances P(t) is collected through intelligent power metering sockets. The data of each device is composed of a timestamp and a power value , forming a time series . The total power of electrical appliances is the sum of the power of all connected devices. When the power changes from 0 to a certain set value, a power-on event is recorded. When the power drops to 0, a power-off event is recorded. All devices are bound by MAC addresses and registered in the system. The intelligent safety monitoring module detects falls, stillness, and abnormal behaviors through action recognition models and event fusion logic and triggers a multi-level early warning response process, which includes: A three-dimensional convolutional neural network is used to model the image sequence. The model output is a behavior category, which can take one of three values: walking, stillness, or falling. The training process is completed using a large-scale video dataset, and the model parameters are determined through cross-validation. The IMU data is analyzed using a sliding window with a size of 2 seconds. The acceleration and angular velocity data are input into the preprocessing module to generate a standardized feature vector x. The feature vector x is input into a pre-trained SVM classifier, and the output is a behavior category C, which can take one of the following values: {still, walking, falling}. A Bayesian network is used to fuse the three types of information: video, IMU, and door magnetic. The fusion result is a comprehensive judgment value, which is used to determine the final behavior category. The joint determination probability calculation method is as follows: the fusion result is equal to the video determination result multiplied by the corresponding weight, plus the IMU determination result multiplied by the corresponding weight, plus the door magnet determination result multiplied by the corresponding weight, and the sum of each weight is 1; A behavior baseline distribution μ(t) is constructed for each user, and the mean and standard deviation σ of the behavior indicators in the last k days are calculated based on a sliding window; If the absolute value of the difference between a behavior indicator and the mean exceeds three times the standard deviation, it is marked as an abnormal event, and the event occurrence time is recorded; The system continuously updates the user behavior pattern and adjusts the model parameters according to the time type (weekday / weekend); When the system detects that the user has not moved for a long time, and the acceleration change collected by the IMU sensor is lower than the set motion threshold, and the video analysis module does not identify valid activity behavior, and the current time is within the user's daily activity range, the system will trigger a first-level warning response, and the response process is executed in the following order: first, start the local voice broadcast module and play the preset reminder sentence to prompt the user to confirm the status; then display the corresponding status prompt information on the bound user terminal interface; finally, the system automatically records this event and enters a confirmation waiting state, waiting for user feedback or subsequent state updates, all operations are completed locally and do not rely on remote communication or external devices; When the IMU model judges a fall behavior, or the video model identifies a fall feature, and the confidence level output by the fusion model does not meet the final determination standard, the system will trigger a second-level warning response, and the response process includes: first, activate the local sound and light alarm device to attract the user's attention; then start the audio collection module to monitor the environment sound for auxiliary judgment of the user's state; then update the event state on the user terminal interface and display the "suspected fall" prompt; finally, send event summary data to the emergency contact list, including timestamp, location identifier and confidence level, but not including specific action judgment results, all response actions are executed by the local computing module and do not involve external linkage or third-party platform calling; When the fall probability output by the fusion model exceeds the system's set determination threshold, the system will trigger a third-level warning response, and the response process is executed in the following order: first, trigger the local alarm signal to prompt the user that he has been determined to fall; then mark the event as "high confidence abnormality" and set it as a priority processing state; then push the complete event record to the local server or edge node, including the original data segment, model output result and timestamp; finally, display the event details and timeline on the user terminal interface for the user or caregiver to view, all operations are based on local network and event records are used for subsequent manual checking or system archiving to ensure traceability and integrity; The wisdom safety monitoring module fuses multi-source data (video, inertial sensor, etc.), uses 3D-CNN and SVM to identify behaviors, combines Bayesian network fusion judgment, constructs personalized behavior baseline and dynamically adjusts the model to automatically detect abnormalities, and early warning is divided into three levels from reminding to pushing complete records; data local processing is suitable for home-based care and health monitoring. The wisdom safety monitoring module fuses multi-source data (video, inertial sensor, etc.), uses 3D-CNN and SVM to identify behaviors, combines Bayesian network fusion judgment. Construct personalized behavior baseline and dynamically adjust the model to automatically detect abnormalities, early warning is divided into three levels from reminding to pushing complete records, data local processing is suitable for home-based care and health monitoring; The wisdom safety monitoring module further integrates two sub-modules, namely an on-site service sub-module and a daytime care sub-module; The functions of the on-site service sub-module are as follows: when the system triggers a three-level early warning or a continuous multiple two-level early warning and does not get user confirmation, the on-site service process is automatically activated, the user or family member can initiate a service request through the terminal, the system generates a service work order containing an event summary, including timestamp, location identification (room number or device ID), abnormal behavior type, risk level and original data segment summary, and pushes it to the handheld terminal of the nearest service personnel through the local area network; The system adopts a service personnel scheduling algorithm based on GPS positioning and task load, selects a response personnel, and supports a historical service matching mechanism, preferentially dispatches a nursing staff who has served the user to perform the task, after the service is completed, the nursing staff fills in a service report on the mobile terminal, including arrival time, departure time, user state evaluation and whether follow-up is needed, all records are uploaded to the local server for archiving, at the same time, when the on-site service is triggered, the system sends an event notification to the emergency contact, including the service personnel contact information and the expected arrival time, the scheduling algorithm in the above content is a technical means in the prior art, which will not be described in detail here.

[0020] The functions of the daytime care sub-module are as follows: the system counts the user's daily activity frequency, active time period, movement range and home appliance use habit through a sliding window, and analyzes the behavior rule in combination with time type and external factors, when the behavior trend changes are detected, the system generates care suggestions, including companion arrangement, social interaction, exercise guidance, rehabilitation training, diet and rest adjustment, mental health and sleep management; The system has a social interaction recommendation function, when it is found that the user is in a low active state for consecutive days, it automatically recommends to organize a small party, an interest group or a cultural and entertainment activity, and a periodic reminder can be set to guide the user to participate, the identified abnormal behavior is fed back to the daytime care personnel in real time, supporting them to carry out psychological counseling, life assistance or medical referral.

[0021] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that: like Figure 2 As shown, the rehabilitation and physiotherapy module is used to generate rehabilitation training paths based on the user's physical fitness score, chronic disease status and age. It adjusts the training content through graph optimization algorithms, combines posture recognition and IMU tracking to evaluate movements, and predicts fatigue status based on heart rate and movement trends. The process by which the rehabilitation and physiotherapy module obtains a user's physical fitness score, chronic disease status, and age includes: Obtain users' physical fitness scores, chronic disease status, and age, and build individualized training models based on these data. A reference value for setting the upper limit of training intensity is calculated based on the user's age. This reference value is obtained by subtracting the user's age from 220, and the result is used as the user's maximum heart rate value. This value is used to limit the heart rate to not exceed this set range during training. Record the presence status of three chronic diseases: cardiovascular disease, diabetes, and osteoporosis. The status of each disease is identified in binary form: if a disease exists, it is represented by the number 1, and if a disease does not exist, it is represented by the number 0. Number 1 corresponds to cardiovascular disease, number 2 corresponds to diabetes, and number 3 corresponds to osteoporosis. The risk level and intensity limit of the training movement are adjusted based on the value corresponding to each number. Cardiovascular disease present: There is a confirmed cardiovascular disease (such as coronary heart disease, hypertensive heart disease, valvular heart disease, etc.). Cardiovascular disease not present: There are no diagnostic records for any of the above-mentioned cardiovascular diseases; Diabetes is present if: fasting blood glucose ≥7.0 mmol / L or blood glucose level ≥11.1 mmol / L 2 hours after oral glucose tolerance test, or if there is a doctor's diagnosis record; Diabetes does not exist: The above criteria are not met, and there is no record of a diabetes diagnosis. Osteoporosis is present: T-score ≤ -2.5 as measured by dual-energy X-ray absorptiometry, or a fracture caused by osteoporosis has occurred in the past; Osteoporosis is not present: T-score greater than -2.5, and no history of osteoporosis-related fractures; The physical fitness score consists of three indicators: cardiorespiratory endurance, muscle strength, and joint mobility. The final score is obtained by fusing these indicators through a function that maps the results of standardized tests and outputs a value between 0 and 1 to measure the user's overall physical fitness level. Cardiorespiratory endurance Standardized testing: Using maximum oxygen uptake as a reference: Excellent: Greater than 40 mL / (kg·min); Good: 35-40 mL / (kg·min); Average: 30-35 mL / (kg·min); Poor: less than 30 mL / (kg·min); Muscle strength Test method: maximum number of repetitions with body weight-dependent movement, or single maximum lifting weight at a specific weight: Excellent: able to complete one repetition with a weight of more than 1.5 times body weight; Good: able to complete one repetition with a weight of 1.2-1.5 times body weight; Average: able to complete one repetition with a weight of 0.9-1.2 times body weight; Poor: less than 0.9 times body weight; Joint mobility The range of motion of the main joints (such as shoulders, elbows, hips, knees, and ankles) is measured using a protractor: Excellent: all key joint ranges of motion reach more than 80% of normal values; Good: 60%-80% of normal values; Average: 40%-60% of normal values; Poor: less than 40% of normal values; The process of generating a rehabilitation training path and adjusting the training content through graph optimization algorithm by the rehabilitation therapy module includes: The training path is composed of multiple action nodes, each node representing a specific rehabilitation action. The system selects a set of actions from the preset action library to form an initial training path. The system scores each action, with scoring dimensions including risk level, difficulty coefficient, and adaptability; Each dimension is weighted and summed according to fixed weights to generate a target function value. The weights are 0.4, 0.3, and 0.3, respectively, and do not change over time. The system sorts according to the target function value and selects the top n actions to form the initial training path; The training path is modeled as a weighted directed graph, with each node representing a training action and the edges representing the possibility of transitioning from one action to another; The system defines a cost function that considers the risk increment, intensity influence, and recovery time between actions. The weights of the cost function are consistent with those in the path generation stage, and the A* algorithm is used to search for the optimal path, ensuring that all action combinations in the path meet the minimum cost condition; The process of evaluating actions and predicting fatigue status based on heart rate and action trends by the rehabilitation therapy module in combination with posture recognition and IMU tracking includes: The system combines video gesture recognition and IMU data to evaluate the user's action execution quality in real time. Gesture recognition is achieved by extracting human key point coordinates, and the standard action is defined as a key point set Each key point coordinate is (x, y, z), and the Euclidean distance between the current action and the standard action is calculated: ; Wherein is the deviation of the i-th key point, is the key point coordinate of the current action, is the key point coordinate of the standard action; Sum all key point deviations to get the overall deviation of the action: ; If the value exceeds the preset threshold, it is determined that the action execution is not up to standard. The IMU device collects acceleration and angular velocity data to determine whether the action deviates from the expected trajectory and evaluate the stability and fluency of the action. By collecting heart rate and action trend data in real time, the user's fatigue state is predicted, and the training intensity is adjusted accordingly. The current heart rate and resting heart rate are collected, the heart rate change rate is calculated, the action execution time is compared with the recommended execution time, and the action trend fatigue index is calculated. When the index exceeds the set threshold, the system reduces the intensity of subsequent actions or prompts the user to rest. The rehabilitation therapy module generates a personalized training path according to the user's age, physical fitness score, and chronic disease status (cardiovascular disease, diabetes, osteoporosis), dynamically adjusts the action content with graph optimization algorithm to ensure safety and effectiveness; The system limits the training intensity by the maximum heart rate (220-age), constructs the initial path based on the weighted score (0.4:0.3:0.3) of action risk, difficulty and adaptability, and optimizes the action sequence using A* algorithm to realize scientific training process; Real-time evaluation of action quality through video gesture recognition and IMU sensor, calculation of key point deviation and action stability, if the deviation is out of standard, the action is determined to be not up to standard; At the same time, combined with the heart rate change rate and the action trend to predict the fatigue state, the intensity is adjusted in time or the user is prompted to rest, to prevent sports injuries, improve rehabilitation efficiency and user compliance, suitable for postoperative recovery, chronic disease management and home rehabilitation scenarios.

[0022] Embodiment three: The technical solution of the embodiment of the present application is different from the technical solutions of embodiment one and embodiment two: As shown in Figure 4 , on the basis of the above-mentioned integrated integrated pension management system based on end-cloud cooperation, the following three modules are further expanded: catering management module, supermarket management module and old university module, which comprehensively covers the daily life and spiritual and cultural needs of the elderly, and realizes the whole-chain intelligent support from health management to life service; The catering management module enables intelligent management of the entire dietary process for the elderly, including nutritional assessment, dietary planning, meal ordering and delivery, and meal feedback. The system establishes an individualized dietary model based on the user's gender, age, height, weight, physical signs, chronic disease status, and allergy history, combined with a medical nutrition database, and generates dietary recommendations that meet health requirements. The system automatically recommends meal plans based on seasonal changes, seasonal ingredients, user taste preferences, and the principle of nutritional balance, and provides alternative options. It supports connection to meal distribution centers or home kitchens to realize online ordering, timed delivery, and nutrition labeling. The catering management module analyzes dining conditions through image recognition technology, identifies food types and intake, calculates calories and nutrient content, and generates daily dietary reports. If abnormal intake is detected, the system triggers an alert mechanism and feeds back to the chronic disease management module for intervention. The catering management module supports the automatic generation of ingredient purchase lists based on inventory and recipes, and can be integrated with local fresh food platforms to enable one-click ordering; The implementation method is as follows: Calculate the total daily energy consumption based on basic information and activity intensity, and set a calorie limit. Set the proportions of carbohydrates, fats, and proteins according to health conditions and adjust them dynamically. Visual recognition technology is used to estimate the calories and nutrients of each meal, with a fixed margin of error.

[0023] The supermarket management module constructs an age-friendly smart retail system, covering product display, shopping recommendations, intelligent checkout and inventory management. The system integrates daily necessities consumption records, rehabilitation equipment usage frequency and lifestyle data to establish a personalized product recommendation model, recommending daily necessities, medicines, health products and assistive devices. The supermarket management module supports voice ordering, touch operation, and remote purchasing. It deploys RFID and visual recognition technologies to achieve unmanned shelf management and behavior trajectory analysis. It automatically reminds customers to replenish stock when inventory is below a set value to ensure stable supply. The system provides intelligent delivery route planning services, supports door-to-door delivery and logistics progress tracking, and uploads transaction data to the cloud for feedback to the security monitoring module for behavior judgment. The supermarket management module also supports member points, coupon distribution, health product recommendations, and integration with medical insurance payments; Implementation method: Recommended content is continuously optimized based on historical purchases, browsing behavior, and health status; Inventory is confirmed using both electronic tags and cameras to predict demand changes and prepare accordingly. Route planning integrates geographical location, order priority, and traffic conditions to optimize delivery efficiency.

[0024] The senior university module constructs a lifelong learning platform, improves the cognitive ability, social interaction and digital literacy of the elderly, systematically integrates online courses and offline activity resources, and covers health knowledge, interest cultivation, equipment operation, psychological adjustment and legal knowledge; The senior university module adopts an intelligent recommendation algorithm to generate a personalized course combination according to learning history, interest label, cognitive ability and free time, and dynamically adjusts the difficulty. Each course has a clear goal, chapter division and assessment mechanism, and an electronic certificate can be obtained after completing the learning; The system supports multi-person online classroom interaction, provides voice question and answer, bullet screen exchange, homework submission and teacher review functions, provides an offline activity registration channel, records participation information and archives it to the personal health record; A community exchange platform is set up to support interest group formation and work display. Learning rankings and activity notifications are regularly pushed, and the system has learning data analysis capabilities to calculate course completion rate, knowledge point mastery and interaction frequency, generate a learning effectiveness evaluation report, and assist in developing a personalized education intervention plan; Implementation: Course recommendation is based on interest, learning record, time arrangement and cognitive level sorting and screening; Learning achievements are evaluated by test banks and behavior data to generate comprehensive evaluation reports; Precedence dependency is set between courses, and users must complete the previous course to unlock the subsequent content; The catering management module, supermarket management module and senior university module are deeply integrated with the above-mentioned integrated end-cloud collaborative integrated pension management system to form a comprehensive pension service system covering "health monitoring-slow disease intervention-safety monitoring-rehabilitation training-diet nutrition-life service-spiritual culture". Each module supports edge-side real-time processing and cloud-side collaborative analysis to ensure response speed and data security, and is suitable for various application scenarios such as home-based care, community-based care and institutional care; The system realizes comprehensive care from physical health to mental health, from basic life to spiritual life through multi-modal data acquisition, AI modeling, natural language interaction and graph optimization algorithm core technologies, and builds a "technology + humanity" integrated smart pension management mode.

[0025] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes to the technical solutions and improvement concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. Integrated all-in-one pension management system based on end-cloud cooperation, applied to a pension management platform, characterized in that, Comprise: Data acquisition and processing module, for collecting health monitoring data of elderly users, and pre-processing the collected health monitoring data; Chronic disease management module, based on the collected health monitoring data, constructs an individual health trajectory model, combines dynamic risk prediction and reinforcement learning to generate an intervention scheme, and performs semantic mapping and natural language output based on a medical knowledge graph; Intelligent safety monitoring module, for collecting camera video stream, inertial sensor motion data, door magnetic switch state and electrical appliance use timing data, detecting falls, stillness and abnormal behaviors through motion recognition model and event fusion logic, and triggering multi-level warning response; Rehabilitation therapy module, for generating a rehabilitation training path according to user physical fitness score, chronic disease status and age, adjusting training content through graph optimization algorithm, evaluating motion based on posture recognition and IMU tracking, and predicting fatigue state based on heart rate and motion trend. 2.The integrated all-in-one old-age management system based on end-cloud cooperation of claim 1, wherein The process of constructing an individual health trajectory model based on the collected health monitoring data by the chronic disease management module comprises: Obtaining pre-processed health monitoring data, including physiological data, environmental data and daily behavior data, physiological data including heart rate, blood pressure and blood oxygen saturation, environmental data including temperature and air quality PM2.5, and daily behavior data including step count and sleep quality; The standardized health data is modeled using a multivariate LSTM network, which receives a window of W days of historical health data, each window containing D health indicators, and the LSTM unit outputs the hidden state at the current time ; The LSTM model processes the input data at each time step, updates the internal state, and after completing the entire window processing, the model outputs the predicted values of each health indicator in the next T days. 3.The integrated all-in-one old-age management system based on end-cloud cooperation of claim 2, wherein, The process of generating an intervention scheme by the chronic disease management module combining dynamic risk prediction and reinforcement learning, and performing semantic mapping and natural language output based on a medical knowledge graph comprises: Based on the future T-day prediction results output by the LSTM model, an MLP model is constructed for risk prediction, the MLP model includes three fully connected layers, the input is the future T-day health index sequence output by the LSTM model, and the output is the health risk score of each day in the future T days , t∈{1,2,...,T} Using DoubleDQN combined with priority experience replay mechanism to learn intervention strategies, the state is composed of current and future T time steps of health indicators; The action space is composed of a set of discrete actions, each action represents a specific intervention behavior, and the action number is from 0 to N-1, N is the total number of actions; The reward value is composed of two factors: health risk change and burden brought by intervention measures; Using DoubleDQN strategy, the update formula is as follows: ； wherein is 0.001, is 0.95, s is state, a is action, r is immediate reward, is next state, is target network; Target network parameters are synchronized from the main network once every 100 training periods, at each time, the DQN network outputs the Q value of each action under the current state, and the action corresponding to the maximum Q value is selected as the intervention suggestion; The reinforcement learning process is executed in a loop, each execution includes the following steps: S1: Obtain the current user state; S2: Input the state to the DQN network, output the Q value corresponding to each action, select the action corresponding to the maximum Q value; S3: Apply the selected action and observe the next state, record the experience and update the network, and enter the next round of judgment; S4: Repeat S1-S3 to gradually generate an intervention strategy adapted to the current user state; Using GPT-2 model to generate natural language suggestions, input is the selected intervention action number and user's basic information, output is a complete natural language suggestion text. 4.The integrated all-in-one old-age management system based on end-cloud cooperation of claim 1, wherein, The process of processing camera video stream, inertial sensor motion data, door magnetic switch state and electrical appliance use timing data by the intelligent safety monitoring module comprises: The video data is acquired using a network high-definition camera, and the camera outputs a sequence of image frames as where each frame represents a fixed-resolution image matrix with dimensions where H is the image height, W is the image width, and C is the number of color channels. Using IMU module with built-in three-axis accelerometer and three-axis gyroscope; The three-axis acceleration vector is represented as ; The three-axis angular velocity vector is represented as ; Total acceleration value is obtained by summing the square of acceleration in three directions and taking square root; Using wireless door magnetic sensor to detect the state of the door, the state of the door is divided into two kinds: open or close; The power P(t) of electrical appliances is collected through smart power metering sockets, and the data for each device is timestamped. and power value Composition, forming a time series ; The total power of the electrical appliance is the sum of the power of all connected devices. When the power changes from 0 to a certain set value, the power-on event is recorded. When the power drops to 0, the power-off event is recorded. 5.The integrated all-in-one old-age management system based on end-cloud cooperation according to claim 4, characterized in that, The process of detecting falls, stillness and abnormal behaviors and triggering multi-level early warning responses by the intelligent safety monitoring module through action recognition model and event fusion logic includes: Using three-dimensional convolutional neural network to model the image sequence, the model output is the behavior category, which takes one of walking, stillness or fall; Sliding window analysis is performed on the IMU data, and acceleration and angular velocity data are input into the preprocessing module to generate a standardized feature vector x. The feature vector x is input into the pre-trained SVM classifier to output the behavior category C, which takes the value of {still, walk, fall}; Using Bayesian network to fuse three types of information: video, IMU, door magnetic, the fusion result is a comprehensive judgment value, which is used to determine the final behavior category; For each user, construct a behavior baseline distribution μ(t), based on the sliding window statistics, the mean and standard deviation σ of the behavior index in the last k days are calculated; If the absolute value of the difference between a behavior index and the mean value exceeds three times the standard deviation, it is marked as an abnormal event; When the system detects that the user has not moved for a long time, and the acceleration change collected by the IMU sensor is lower than the set motion threshold, and the video analysis module does not identify valid activity behavior, and the current time is within the user's daily activity range, the system will trigger a first-level early warning response; When the IMU model judges that the behavior is falling, or the video model identifies the falling feature, and the confidence output by the fusion model does not reach the final judgment standard, the system will trigger a second-level early warning response; When the fall probability output by the fusion model exceeds the system's set judgment threshold, the system will trigger a third-level early warning response. 6.The end-cloud cooperation based integrated all-in-one old-age management system according to claim 1, characterized in that, The process of the rehabilitation physiotherapy module obtaining user physical fitness score, chronic disease status and age includes: Obtain user physical fitness score, chronic disease status and age, and construct individualized training model based on user physical fitness score, chronic disease status and age; According to the user's age value, a reference value for setting the upper limit of the training intensity is calculated. The reference value is obtained by subtracting the user's age from 220, and the result is the maximum heart rate value of the user; Record the existence state of three chronic diseases: cardiovascular disease, diabetes, osteoporosis. The state of each disease is identified in binary form: if a certain disease exists, use the number 1 to represent it, if a certain disease does not exist, use the number 0 to represent it. Number 1 corresponds to cardiovascular disease, number 2 corresponds to diabetes, and number 3 corresponds to osteoporosis; Physical fitness score is composed of three indicators: cardiorespiratory endurance, muscle strength and joint mobility. The final score is obtained by function fusion. 7.The integrated all-in-one old-age management system based on end-cloud cooperation of claim 6, wherein, The process of the rehabilitation physiotherapy module generating rehabilitation training path and adjusting training content through graph optimization algorithm includes: The training path is composed of multiple action nodes, each node representing a specific rehabilitation action, and the system selects a set of actions from a pre-set action library to form an initial training path; Each dimension is weighted and summed according to a fixed weight to generate a target function value, and the system ranks the target function values and selects the top n actions to form the initial training path; The training path is modeled as a weighted directed graph, with each node representing a training action and the edges representing the possibility of transitioning from one action to another; The system defines a cost function that considers the risk increment, intensity influence, and recovery time between actions, and uses the A* algorithm to search for the optimal path, ensuring that all action combinations in the path meet the minimum cost condition. 8.The integrated all-in-one old-age management system based on end-cloud cooperation of claim 7, wherein, The process of evaluating actions and predicting fatigue status based on heart rate and action trend prediction by the rehabilitation physiotherapy module combined with posture recognition and IMU tracking includes: The system combines video gesture recognition and IMU data to evaluate the user action execution quality in real time. The gesture recognition is achieved by extracting human key point coordinates, and the standard action is defined as a key point set Each key point coordinate is (x, y, z), and the Euclidean distance between the current action and the standard action is calculated: ； wherein is a deviation of the i-th key point, is a key point coordinate of the current motion, is a key point coordinate of the standard motion; Sum all key point deviations to obtain the overall deviation of the action: ； If the value exceeds the pre-set threshold, it is determined that the action execution is not up to standard; By collecting heart rate and action trend data in real time, the user's fatigue status is predicted, and the training intensity is adjusted accordingly. The current heart rate and resting heart rate are collected, the heart rate change rate is calculated, the action execution time is collected, and the recommended execution time is compared to calculate the action trend fatigue index. When the index exceeds the set threshold, the system reduces the intensity of subsequent actions or prompts the user to rest.

Images