An adaptive companionship system and method for elderly care service robots

By combining non-invasive sensing modules and dynamic baseline models, the problem of existing elderly care service robots relying on wearable devices has been solved, enabling low-cost, high-reliability monitoring of the elderly's condition and personalized early warning, thereby improving the intelligence and proactive prevention capabilities of elderly care services.

CN123103629APending Publication Date: 2026-11-17SUZHOU FUAI ELDERLY CARE SERVICE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611227717.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing elderly care service robots rely on wearable devices or high-precision sensors for data collection, which increases the cost of use and maintenance, reduces the acceptance and compliance of the elderly, and lacks personalized behavioral decision-making, resulting in a high false alarm rate and the inability to intervene in advance at the early stage of risk.

Method used

A non-invasive sensing module is used to acquire the elderly’s status data through wearable sensors, establish a dynamic baseline model, combine it with a prediction module to calculate the degree of deviation, generate a three-level early warning, and execute the corresponding care strategy through a decision module.

Benefits of technology

It has achieved low-cost, wearable-free elderly condition monitoring, reduced false alarm rate, improved the level of intelligence, enabled proactive prevention at the early stage of risk, and improved the elderly's acceptance and compliance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123103629A_ABST
    Figure CN123103629A_ABST
Patent Text Reader

Abstract

This application discloses an adaptive companionship system and method for elderly care service robots. A non-invasive sensing module acquires the state data of the elderly being accompanied through wearable sensors, including behavioral and physiological data. A baseline model generation module establishes a dynamic baseline model based on the elderly's historical state data during the initial monitoring phase. This model includes a mean and standard deviation that are smoothly updated over time. A prediction module compares the current state data with the dynamic baseline model, calculates the degree of deviation, and generates prediction results at three levels: normal, attentive, and warning. A decision-making module determines the target strategy from multiple companionship strategies based on the prediction results. An execution module executes the target strategy to drive the robot to perform corresponding actions. This application, through non-contact sensing and personalized dynamic baselines, reduces the false alarm rate, achieves accurate graded warnings and adaptive companionship for the elderly's state, and improves the intelligence and safety of elderly care services.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of elderly care and robotics, and in particular to an adaptive companionship system and method for elderly care service robots. Background Technology

[0002] In recent years, elderly care service robots have gradually become important tools to assist the lives of the elderly. Most existing elderly care service robots rely on preset programs or remote control commands to perform fixed tasks, such as delivering medication along a set route, playing music at set times, or issuing alarms after detecting a fall. These robots typically require the elderly to wear wearable devices such as wristbands or chest straps to collect physiological data, or rely on high-precision sensor arrays for environmental monitoring. However, wearable devices often cause discomfort for the elderly and are easily forgotten or lost, while high-precision sensor arrays are expensive and difficult to deploy on a large scale in nursing homes. Furthermore, the behavioral decisions of existing systems are mostly based on simple threshold judgments; once data exceeds a preset range, an alarm is immediately triggered, lacking the ability to adapt to individual differences and dynamic changes among the elderly.

[0003] Existing technologies have the following prominent problems: First, data collection relies on wearable devices or high-precision sensors, which not only increases the cost of use and maintenance difficulty, but also reduces the acceptance and compliance of the elderly. Second, the lack of personalized modeling of individual behavioral patterns of the elderly makes it impossible for the system to distinguish between normal physiological fluctuations and true abnormal states, which easily leads to a large number of false alarms. Third, the care strategies are singular and lagging, and can only respond after an event occurs, failing to intervene in advance at the nascent stage of risk, which seriously affects the level of intelligence and humanization of elderly care services.

[0004] Therefore, there is an urgent need for an adaptive companionship system for elderly care services that can collect elderly people's status data in a non-invasive, wear-free manner and make accurate predictions and graded early warnings based on a personalized dynamic baseline model. Summary of the Invention

[0005] To improve existing methods and systems, an adaptive companionship system and method for elderly care service robots are provided. The specific technical solution is as follows: In a first aspect, this application provides an adaptive companionship system for elderly care service robots, comprising: The non-invasive sensing module is configured to acquire the status data of the elderly being cared for through at least one wearable sensor. The status data includes behavioral data and physiological indicator data. The baseline model generation module is configured to establish a dynamic baseline model based on the historical status data of the elderly being cared for during the initial monitoring phase. The dynamic baseline model includes the mean and standard deviation that are updated smoothly over time. The prediction module, connected to the non-intrusive sensing module and the baseline model generation module, is configured to compare the current state data with the dynamic baseline model, calculate the degree of deviation, and generate prediction results based on the degree of deviation. The prediction results include three levels: normal, attention, and warning. The decision-making module, connected to the prediction module, is configured to determine the target strategy from multiple caregiving strategies based on the prediction results. The execution module, connected to the decision-making module, is configured to execute target strategies to drive the robot to perform corresponding actions.

[0006] In some embodiments, the non-invasive sensing module includes a ranging sensor and a camera; the ranging sensor is configured to emit ranging signals toward the chest or back of the elderly person being accompanied and to receive reflected signals, and the physiological index data is obtained by bandpass filtering and peak detection of the round-trip time series of the ranging signal to obtain respiratory rate; the camera is configured to acquire a full-body image of the elderly person being accompanied, and the behavioral data is obtained by detecting key points of the human skeleton and performing time-series analysis on the full-body image to obtain gait parameters and activity levels.

[0007] In some embodiments, the ranging sensor is an ultrasonic sensor or a laser ranging sensor. The ranging sensor is mounted on the front end of the robot and can adjust its pitch angle with the robot's head or robotic arm so that the ranging signal is always aligned with the chest area of ​​the elderly person being accompanied when the elderly person is in a sitting or lying position.

[0008] In some embodiments, the robot body is a movable chassis structure, and the non-invasive sensing module 1 also includes a follow control unit, which controls the robot to follow the elderly person being accompanied at a preset distance based on the image recognition results of the camera, and keeps the straight path between the distance measuring sensor and the chest of the elderly person being accompanied free of obstacles during the following process.

[0009] In some embodiments, the caregiving strategies stored in the decision-making module include: when the prediction result is normal, executing a routine caregiving strategy, which includes playing music or having a voice chat; when the prediction result is attention, executing a proactive greeting and close observation strategy, which includes the robot moving to a preset distance in front of the elderly person being cared for and asking if they need help; and when the prediction result is an early warning, executing an emergency notification to caregivers and a deceleration and following strategy, which includes the robot reducing its movement speed and turning on warning lights.

[0010] In some embodiments, the dynamic baseline model further includes an external factor correction submodule that shifts the mean and standard deviation based on the current time.

[0011] In some embodiments, when calculating the degree of deviation, the prediction module uses the cumulative deviation integral value within the sliding window as the basis for judgment. Only when the cumulative deviation integral value exceeds a preset threshold will a prediction result of the warning level be generated. The preset threshold is an adaptive threshold that is automatically adjusted based on the recent average deviation level of the elderly person being accompanied.

[0012] In some embodiments, the camera is also used to capture facial expression images of the elderly person being accompanied; the behavioral data also includes emotional states obtained through facial expression recognition, which are used to assist the prediction module in judging the level of attention.

[0013] In some embodiments, the baseline model generation module is also configured to periodically recalibrate the dynamic baseline model. The recalibration timing includes: every morning, after each major event, and automatically when no warning has occurred for a consecutive week; wherein major events include discharge from the hospital or medication adjustment.

[0014] Secondly, this application provides an adaptive companionship method for elderly care service robots, implemented in any of the above embodiments, comprising the following steps: S1: Acquire the status data of the elderly person being accompanied. The status data is acquired through at least one wearable sensor and includes behavioral data and physiological indicator data. S2: Based on the historical status data of the elderly being cared for during the initial monitoring phase, a dynamic baseline model is established. The dynamic baseline model includes the mean and standard deviation that are updated smoothly over time. S3: Compare the current status data with the dynamic baseline model, calculate the degree of deviation, and generate prediction results based on the degree of deviation. The prediction results include three levels: normal, attention, and warning. S4: Determine the target strategy from multiple caregiving strategies based on the prediction results; S5: Execute the target strategy to drive the robot to perform the corresponding action.

[0015] This application provides an adaptive companionship system for elderly care service robots, including a non-invasive sensing module, a baseline model generation module, a prediction module, a decision-making module, and an execution module. The non-invasive sensing module acquires behavioral and physiological data of the elderly through wearable sensors; the baseline model generation module establishes a dynamic baseline model containing the mean and standard deviation based on historical state data; the prediction module compares the current data with the baseline, calculates the degree of deviation, and generates three levels of prediction results: normal, attentive, and warning; the decision-making module selects a companionship strategy accordingly, and the execution module drives the robot to perform the corresponding actions. This application avoids the discomfort and loss problems of wearable devices through non-contact sensing, effectively distinguishes between normal fluctuations and abnormal states using personalized dynamic baselines, significantly reducing the false alarm rate. Simultaneously, the three-level warning system realizes a shift from passive response to proactive prevention, significantly improving the intelligence level of elderly care services and the elderly's sense of security. Attached Figure Description

[0016] Figure 1 This is a signaling diagram of the adaptive care system for the elderly care service robot proposed in this invention; Figure 2 This is a flowchart of the adaptive companionship method for elderly care service robots proposed in this invention.

[0017] In the diagram: 1. Non-invasive perception module; 2. Baseline model generation module; 3. Prediction module; 4. Decision module; 5. Execution module. Detailed Implementation

[0018] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0019] Firstly, this application provides an adaptive companionship system for elderly care service robots, such as... Figure 1 As shown, it includes: The non-invasive sensing module 1 is configured to acquire the status data of the elderly being accompanied through at least one wearable sensor. The status data includes behavioral data and physiological indicator data. Baseline model generation module 2 is configured to establish a dynamic baseline model based on the historical status data of the elderly being cared for during the initial monitoring phase. The dynamic baseline model includes the mean and standard deviation that are updated smoothly over time. Prediction module 3, connected to non-intrusive perception module 1 and baseline model generation module 2, is configured to compare the current state data with the dynamic baseline model, calculate the degree of deviation, and generate prediction results based on the degree of deviation. The prediction results include three levels: normal, attention, and warning. Decision module 4, connected to prediction module 3, is configured to determine the target strategy from multiple care strategies based on the prediction results; The execution module 5 is connected to the decision module 4 and is configured to execute the target strategy to drive the robot to perform corresponding actions.

[0020] This embodiment provides an adaptive companionship system for elderly care service robots. This system aims to monitor the elderly person's condition in real time through a non-invasive, wearable method, and to perform intelligent prediction and graded response based on a personalized dynamic baseline, thereby solving problems such as high false alarm rates, poor adaptability, and unsatisfactory user experience in existing technologies. The various components of this embodiment and their implementation methods are described in detail below with reference to the accompanying drawings.

[0021] The non-invasive sensing module 1 serves as the data entry point for the entire system. Its design aims to avoid the discomfort, forgetting to wear, or loss associated with traditional wearable devices for the elderly. This module is configured to acquire the elderly person's status data, including behavioral and physiological indicators, through at least one wearable sensor. In a preferred embodiment, the module employs a standard RGB camera and an ultrasonic ranging sensor. The camera captures full-body images of the elderly person and extracts behavioral data such as gait parameters and activity levels using a human skeletal keypoint detection algorithm. The ultrasonic sensor emits ultrasonic waves towards the elderly person's chest area and receives the reflected signals. By performing bandpass filtering and peak detection on the round-trip time series, respiratory rate is extracted as a physiological indicator. The advantage of this combination is that both the camera and the ultrasonic sensor are common robotic components, requiring no additional customized hardware. Furthermore, the ultrasonic ranging is unaffected by lighting conditions and can operate stably at night or in dim environments. In other feasible embodiments, a laser ranging sensor can be used instead of an ultrasonic sensor. Laser ranging offers higher accuracy but is slightly more expensive. Alternatively, a millimeter-wave radar module can be used, which can simultaneously detect breathing and heartbeat, but it consumes more power and is larger in size. A microphone array can also be used to obtain respiratory rate and activity level by analyzing breathing sounds and footsteps, but environmental noise may interfere with the signal. Regardless of the method, no device needs to be worn by the elderly, achieving truly non-invasive monitoring.

[0022] The baseline model generation module 2 aims to establish a personalized health status reference system for each elderly person. Due to significant differences in age, physical condition, and lifestyle among elderly individuals, a uniform threshold would inevitably lead to numerous false alarms or missed alarms. Therefore, this module is configured to build a dynamic baseline model based on the historical status data of the elderly person being cared for during the initial monitoring phase. Specifically, for the first three days after system deployment (the initial monitoring phase), the system only collects data without intervention, gathering behavioral and physiological data of the elderly person in a normal state. The mean and standard deviation of these data are then calculated to form an initial baseline. More importantly, this model is dynamic; its mean and standard deviation are updated smoothly over time, for example, using an exponentially weighted moving average, allowing the baseline to slowly adapt to the natural changes caused by aging or the recovery process. In this way, the system can distinguish whether "the elderly person is walking slower today than yesterday because they are tired" or "because their condition has worsened"—the former is a normal fluctuation, while the latter is abnormal. In other embodiments, the baseline model can also incorporate time segments, such as differentiating between morning, afternoon, and evening baselines to adapt to the elderly person's daily routine; or a weekday factor can be added, as activity patterns may differ between weekends and weekdays.

[0023] The prediction module 3 serves as the connecting hub between the non-invasive perception module 1 and the baseline model generation module 2, and is also the intelligent core of the entire system. This module is configured to compare the current state data with the dynamic baseline model, calculate the degree of deviation, and generate a prediction result based on the degree of deviation. The degree of deviation can be calculated using Z-scores (standardized score), which is the difference between the current value and the baseline mean divided by the standard deviation. However, to avoid false alarms caused by transient noise, the cumulative deviation integral value within a sliding window is preferred as the judgment criterion: an alert is triggered only when the deviation persists for a period of time and the cumulative amount exceeds a preset threshold. The prediction results are divided into three levels: normal, attention, and warning. Normal indicates that the current state is within the baseline range and no intervention is needed; attention indicates a moderate degree of deviation, which may pose a potential risk and requires closer observation; warning indicates a severe deviation, which is likely to be dangerous and requires immediate response. This hierarchical mechanism avoids a "one-size-fits-all" alarm approach, reducing unnecessary panic and the workload of caregivers. In other embodiments, machine learning classifiers (such as random forests or support vector machines) can also be used to classify multidimensional deviation features, but this increases computational complexity.

[0024] Decision module 4 determines the target strategy from multiple caregiving strategies based on the prediction results. This module is designed to translate prediction results into actual actions, achieving a closed loop from "problem identification" to "problem solving." Caregiving strategies are a set of behavioral rules pre-stored in the system. For example: when the prediction result is normal, a standard caregiving strategy is executed, including playing music, engaging in voice chat, or maintaining silent following; when the prediction result is attentive, a proactive greeting and close observation strategy is executed, with the robot moving one meter in front of the elderly person, asking "Do you need help?" via voice, and carefully observing the elderly person's expression and reaction; when the prediction result is a warning, an emergency notification to caregivers and a deceleration following strategy are executed, with the robot immediately sending an alarm message to the care station, reducing its own speed, and activating warning lights to prevent collisions with the elderly person. These strategies can be customized according to the specific management procedures of the nursing home. In other embodiments, strategies can also be generated online using reinforcement learning, but this requires a large amount of training data initially.

[0025] The execution module 5 is connected to the decision-making module 4 and is responsible for driving the robot to perform corresponding actions. For example, when music needs to be played, the execution module 5 controls the speaker to emit sound; when movement is required, the execution module 5 controls the motor to drive the chassis forward or turn; when an alarm needs to be sent, the execution module 5 pushes a message to the nursing station through the network interface. The presence of the execution module 5 makes the system no longer a purely information processing system, but a physical robot system capable of interacting with the environment.

[0026] In summary, this embodiment eliminates the constraints of wearable devices through the non-invasive sensing module 1, achieves personalized adaptation through the baseline model generation module 2, reduces the false alarm rate through the hierarchical mechanism of the prediction module 3, and forms a complete adaptive care loop through the decision-making module 4 and the execution module 5. The entire system improves the intelligence level of elderly care services in a low-cost and highly reliable manner, and is particularly suitable for centralized care scenarios such as nursing homes. It can promptly detect subtle changes in the elderly's physical condition without increasing the burden on caregivers, realizing a shift from passive response to proactive prevention.

[0027] In some embodiments, the non-invasive sensing module 1 includes a ranging sensor and a camera; the ranging sensor is configured to emit ranging signals toward the chest or back of the elderly person being accompanied and to receive reflected signals, and the physiological index data is obtained by bandpass filtering and peak detection of the round-trip time series of the ranging signal to obtain respiratory rate; the camera is configured to acquire a full-body image of the elderly person being accompanied, and the behavioral data is obtained by detecting key points of the human skeleton and performing time-series analysis on the full-body image to obtain gait parameters and activity level.

[0028] The non-invasive sensing module 1 uses a combination of a ranging sensor and a camera to collect elderly person's status data without wearable devices. An ultrasonic sensor is preferred as the ranging sensor, mounted on the front of the robot. Its pitch angle can be adjusted with the robot's head or robotic arm to ensure it is always aligned with the elderly person's chest area, whether they are sitting or lying down. The ultrasonic sensor emits a 40kHz pulse signal and receives the echo reflected from the chest surface. The distance is calculated by measuring the time difference between transmission and reception. Due to the periodic rise and fall of the chest caused by breathing, the distance value fluctuates regularly. The system performs bandpass filtering on the continuously acquired distance time series, filtering out slow drift below 0.1Hz and high-frequency noise above 2Hz, retaining the 0.1-0.5Hz breathing frequency band. Then, a peak detection algorithm is used to extract the start and end points of each breath, thereby calculating the respiratory rate. The advantages of this method are that ultrasound is unaffected by light, can work stably at night or in environments with curtains drawn, and is inexpensive. As another feasible implementation, laser rangefinders (such as single-point ToF lidar) can be used instead of ultrasound. These laser beams are narrower and more precise, capable of detecting even subtle chest cavity movements, but they are slightly more expensive and sensitive to the angle of the reflecting surface. A third implementation is to use a millimeter-wave radar module. This module can not only detect subtle chest cavity movements caused by respiration but also extract heart rate simultaneously through the micro-Doppler effect. However, it has higher power consumption and larger size, making it more suitable for high-end models.

[0029] The camera section uses a standard RGB camera mounted on the robot's head, covering the entire range of motion of the elderly person. The system extracts the coordinates of major joints such as the shoulders, hips, knees, and ankles from each frame of the image using human skeletal keypoint detection algorithms (such as OpenPose or MediaPipe), and then performs temporal analysis on the joint positions across multiple consecutive frames. Gait parameters, including cadence, stride length, and stride symmetry, are calculated by detecting the period and displacement of alternating foot contact; activity level is quantified by calculating the total distance traveled by the center point of the torso per unit time. This visual analysis method requires no markers worn by the elderly person, relies entirely on image processing technology, and the camera itself is a standard component for robot navigation and interaction, requiring no additional hardware investment. As an alternative, a depth camera (such as Intel RealSense) can be used instead of a standard RGB camera. Depth data can provide joint positions in three-dimensional space, reducing errors from two-dimensional projection, but the data processing volume is larger. Another option is to use an infrared thermal imaging camera, which can work in completely dark environments and also acquire body surface temperature information, but the resolution is usually lower and the cost is higher.

[0030] Through the aforementioned non-invasive sensing module 1, the system can continuously acquire behavioral and physiological data without disrupting the elderly person's normal life. Continuous monitoring of respiratory rate can detect respiratory abnormalities (such as breathing too fast or too slow) at an early stage, long-term tracking of gait parameters can identify early signs of Parkinson's disease or trends in muscle strength decline, and changes in activity level reflect the elderly person's overall vitality level. After these data are fed into the subsequent baseline model generation module 2 and prediction module 3, a personalized health profile for each elderly person can be established, thereby achieving accurate abnormality detection and graded early warning. Compared with traditional solutions that require wearing wristbands or chest straps, this wear-free method greatly improves the elderly person's acceptance and compliance, avoids problems such as lost devices and forgotten charging, and reduces the equipment procurement and maintenance costs for nursing homes.

[0031] In some embodiments, the ranging sensor is an ultrasonic sensor or a laser ranging sensor. The ranging sensor is mounted on the front end of the robot and can adjust its pitch angle with the robot's head or robotic arm so that the ranging signal is always aligned with the chest area of ​​the elderly person being accompanied when the elderly person is in a sitting or lying position.

[0032] The ranging sensor primarily uses an ultrasonic sensor, mounted on the robot's chin. The head is driven by an independent pitch servo, allowing for 45-degree vertical rotation. When the elderly person is seated, the robot's head remains horizontal or slightly tilted downwards, ensuring the ultrasonic beam is precisely focused on the center of the elderly person's chest. When the elderly person is lying down, the robot's head tilts upwards by approximately 30 degrees, keeping the beam perpendicular to the chest. To achieve automatic alignment, the robot uses a camera to identify key points on the human skeleton, calculates the position of the chest center, and then adjusts the head's pitch angle using a servo control algorithm, ensuring the ultrasonic sensor always points to that target point. This dynamic tracking mechanism ensures stable acquisition of respiratory signals regardless of the elderly person's posture.

[0033] As another feasible implementation, a laser rangefinder sensor can be used, fixed to the end gripper of a robotic arm. The robotic arm has multiple degrees of freedom, allowing it to adjust not only its pitch angle but also its length to accommodate different distances. When the elderly person is in bed, the robotic arm extends and bends, suspending the laser rangefinder sensor approximately 50 centimeters above the elderly person's chest, with the laser beam pointing vertically downwards. The advantage of this method is higher laser ranging accuracy, and the robotic arm's flexibility allows it to avoid obstacles such as bed rails. A third implementation involves fixing multiple ultrasonic sensors at different angles to a rotating bracket. A stepper motor drives the bracket to rotate, selecting the sensor closest to the elderly person's chest for operation. This method eliminates the need for a complex robotic arm but requires more sensors and calibration work.

[0034] This adjustable pitch angle design brings significant technological benefits. First, it solves the problem of signal loss caused by changes in the elderly's posture. Traditional fixed sensors lose their ranging signal when the elderly change from a sitting to a lying position, sometimes even projecting it towards the empty bed surface, causing interruptions in respiratory data. This solution ensures signal continuity through real-time tracking and adjustment, allowing respiratory monitoring to operate uninterrupted 24 / 7. Second, the automatic alignment mechanism reduces the need for manual intervention; caregivers do not need to adjust the robot's position each time, as the system automatically adapts, lowering the barrier to entry. Third, because the signal is always aligned with the chest area with the strongest reflection, the signal-to-noise ratio is higher, and the peak of the respiratory waveform is more pronounced, improving the accuracy of subsequent bandpass filtering and peak detection, thereby reducing the false alarm rate. Finally, this adjustment mechanism utilizes the robot's existing head or robotic arm degrees of freedom without adding extra hardware costs, achieving a balance between functionality and cost.

[0035] In some embodiments, the robot body is a movable chassis structure, and the non-invasive sensing module 1 also includes a follow control unit, which controls the robot to follow the elderly person being accompanied at a preset distance based on the image recognition results of the camera, and keeps the straight path between the distance measuring sensor and the chest of the elderly person being accompanied free of obstacles during the following process.

[0036] The robot body adopts a movable chassis structure with a dual-wheel differential drive, equipped with two active wheels and one omnidirectional driven wheel, enabling it to turn in place and move in any direction. The follow control unit runs on the robot's main controller, using images acquired by a camera to identify the elderly person's position and orientation in real time, calculating the distance and azimuth angle relative to the robot. In a preferred embodiment, the follow control unit employs a visual servo control algorithm, using the center of the elderly person's torso as the tracking target point, controlling the robot to follow the elderly person at a preset distance of 1.5 meters. During the following process, the control unit simultaneously uses the camera and ultrasonic obstacle avoidance sensors to map the environment in real time, identifying obstacles that may obstruct the straight path between the ranging sensor and the elderly person's chest, such as furniture, walls, or other people. Once an obstacle is detected, the control unit plans a detour path, bypassing the obstacle and returning to the preset distance and angle, ensuring that the ultrasonic or laser beam from the ranging sensor is always unobstructed and pointed towards the elderly person's chest.

[0037] As another feasible implementation, the chassis can adopt an omnidirectional wheel structure. Four Mecanum wheels enable the robot to move laterally, making it easier to maintain a straight path between the ranging sensor and the elderly person in narrow corridors without the need for sharp turns. In a third implementation, the following control unit no longer relies solely on vision but integrates vision and ultra-wideband (UWB) indoor positioning tags. A miniature UWB tag (which can be sewn onto clothing) is worn by the elderly person. The robot calculates the tag's position through multiple UWB base stations and combines the visual recognition results with Kalman filtering fusion to further improve following accuracy and anti-occlusion capability. However, to maintain the principle of wear-free operation, this solution is only used in special demand scenarios, and the preferred solution remains purely vision-based.

[0038] This mobile, follow-along structure offers several advantages. First, the robot can actively follow the elderly person's movements, freeing respiratory monitoring and gait analysis from fixed locations. The elderly can move freely within the room, go to the bathroom, or watch TV in the living room, while the robot maintains the optimal monitoring distance, enabling continuous monitoring around the clock and in all spaces. Second, the requirement to maintain an unobstructed straight path between the ranging sensor and the chest necessitates that the robot possess active obstacle avoidance and path replanning capabilities. This, in turn, enhances the robot's survivability in complex home environments, preventing it from losing its monitoring target due to obstructions from tables or chairs. Third, the preset distance (e.g., 1.5 meters) balances monitoring accuracy with the elderly person's psychological comfort. Too close, and the elderly person will feel monitored; too far, and the ranging signal weakens and the camera resolution decreases. This distance represents a compromise achieved through experimental optimization. Finally, because the robot is mobile, it can serve multiple elderly people simultaneously (through scheduled patrols or dispatch calls), improving equipment utilization and reducing the cost per bed in nursing homes. The entire system upgrades non-invasive perception from static fixed-point to dynamic close-fitting through visually guided autonomous following, truly realizing adaptive companionship where "wherever the robot is, the monitoring is there".

[0039] In some embodiments, the caregiving strategies stored in the decision module 4 include: when the prediction result is normal, executing a routine caregiving strategy, which includes playing music or having a voice chat; when the prediction result is attention, executing an active greeting and close observation strategy, which includes the robot moving to 1 meter in front of the elderly person being cared for and asking if they need help; and when the prediction result is a warning, executing an emergency notification to caregivers and a deceleration and following strategy, which includes the robot reducing its movement speed and turning on warning lights.

[0040] The caregiving strategies stored in decision module 4 respond differently based on the three levels output by prediction module 3. When the prediction result is normal, the regular caregiving strategy is executed. In a preferred embodiment, the robot plays a preset music library through its built-in speaker. The tracks can be automatically selected based on the elderly person's historical preferences, such as playing upbeat piano music in the morning and opera excerpts in the afternoon. Simultaneously, the robot activates its voice chat function, engaging in simple conversations with the elderly person through its natural language processing module, such as asking "How are you feeling today?" or recounting the day's news. Another feasible implementation is that the robot does not speak actively but remains silent, only responding when the elderly person actively calls out, thus minimizing disturbance to their quiet time. In a third implementation, the robot automatically switches to patrol mode, slowly inspecting the room, checking the condition of doors, windows, and floors, while continuously collecting data using a camera for subsequent analysis.

[0041] When the prediction result is "attention," a proactive greeting and close observation strategy is implemented. In a preferred embodiment, the robot uses a path planning algorithm to move to a distance of about 1 meter directly in front of the elderly person, faces them, and asks in a gentle tone through a voice synthesis module, "Do you need help? I notice you seem a little tired." Simultaneously, the robot zooms in with its camera, focusing on capturing the elderly person's facial expressions and eye condition, using micro-expression analysis to determine if there are signs of pain or anxiety. If the elderly person responds "I'm fine," the robot retreats to a normal distance and resumes its usual companionship strategy; if the elderly person does not respond or shows discomfort, the robot automatically increases the prediction level and sends a reminder to the caregiver. In another embodiment, the robot does not move its position but adjusts its head tilt angle to aim the camera at the elderly person's face, while playing soothing music or natural white noise to try to alleviate any possible tension. In a third embodiment, the robot proactively delivers a glass of warm water (if equipped with a robotic arm and cup holder), enhancing the interaction through physical care.

[0042] When the prediction result is a warning, an emergency notification to caregivers and a slowdown / following strategy are implemented. In a preferred embodiment, the robot immediately sends an alarm message to the control panel of the care station via Wi-Fi, including the elderly person's name, room number, warning level, and a summary of current status data. Simultaneously, it notifies the on-duty nurse's mobile phone via SMS or an app. At the same time, the robot limits the maximum speed of its drive motor to 50% of its normal value, reducing its movement speed to half of its walking speed to prevent secondary injuries to the elderly person from sudden acceleration or braking. The robot also activates the LED warning light located on its top, flashing yellow to alert those nearby. In another embodiment, the robot simultaneously activates a local buzzer to emit a soft but clear alert tone, guiding caregivers to quickly locate the elderly person. In a third embodiment, the robot automatically dials a preset emergency contact number, broadcasts a pre-recorded distress message via voice synthesis, and activates hands-free mode after the call is connected, facilitating direct communication between caregivers and the elderly person.

[0043] The beneficial effects of this tiered strategy are reflected in several aspects. The routine companionship strategy at the normal level maintains the elderly's normal daily rhythm, avoiding unnecessary intervention, while providing emotional comfort through music and conversation, helping to slow cognitive decline. The proactive greeting strategy at the attention level intervenes at the initial stage of risk, gathering more clues through close observation, avoiding the passive situation of responding only when the situation worsens. Moreover, the robot interacts by asking questions rather than giving commands, respecting the elderly's autonomy and reducing resistance. The emergency notification strategy at the warning level ensures the fastest possible professional rescue, while the deceleration following and warning lights minimize the risk of the robot causing accidents in emergency situations, while providing clear guidance for caregivers. The entire strategy system, from "no disturbance unless necessary" to "proactive care" and then to "emergency response," forms a progressive and humanized care chain that neither overly disturbs the elderly nor overlooks any potential dangers, significantly improving the refinement and intelligence of elderly care services.

[0044] In some embodiments, the dynamic baseline model further includes an external factor correction submodule that shifts the mean and standard deviation based on the current time, season, weather, and nursing home schedule.

[0045] The dynamic baseline model integrates an external factor correction submodule, specifically responsible for dynamically shifting and correcting the mean and standard deviation of the baseline. In a preferred embodiment, this submodule obtains the current system time from the robot operating system, retrieves local weather forecast data from a network API interface, and reads the daily schedule provided by the nursing home management system. The submodule maintains a multi-dimensional correction table, with each dimension corresponding to an external factor. For example, for the time dimension, a day is divided into six periods: early morning, morning, lunch break, afternoon, evening, and night, each assigned a different offset coefficient; for the season dimension, spring, summer, autumn, and winter each have independent offsets; for the weather dimension, sunny, cloudy, rainy, and snowy days correspond to different correction values; for the schedule dimension, meal times, activity times, rest times, and visiting times all have preset adjustment parameters. The submodule sums all factors corresponding to the current moment to obtain a comprehensive offset, then adds this offset to the baseline mean, and multiplies the standard deviation by a scaling factor. For example, in the early morning of winter, the basal metabolic rate of the elderly is slower, and their respiratory rate and activity level are usually lower. The submodule will lower the mean by a certain percentage and appropriately widen the standard deviation to avoid misjudging normal seasonal low activity as abnormal.

[0046] In another feasible implementation, the external factor correction submodule no longer relies on a preset fixed correction table, but instead adopts an online learning method. The submodule continuously collects historical data and automatically learns the influence weights of each external factor on the state data through multiple linear regression or gradient boosting tree models, and periodically updates the correction model. This approach is more adaptable, but requires accumulating a sufficient amount of training data. In a third implementation, the submodule uses a combination of a rule engine and fuzzy logic, taking time, season, weather, and schedule as fuzzy input variables. It outputs correction values ​​through an expert-defined fuzzy rule base (e.g., "If it is late at night and the weather is rainy, the mean will be adjusted more significantly"). This approach is highly interpretable, making it easier for nursing staff to understand and adjust the rules.

[0047] The benefits of the external factor correction submodule are significant. First, it eliminates systematic biases caused by objective factors such as time and season. For example, an elderly person's activity level is naturally higher in summer than in winter. Without correction, the system might frequently trigger "attention" or even "warning" alerts in winter, leading to numerous false alarms. After correction, the baseline fluctuates with natural patterns, and the elderly person's normal low activity level in winter will not be misjudged as abnormal. Second, weather factors have a significant impact on the elderly person's condition. On rainy days, the elderly person is prone to low mood, joint pain, and reduced activity. The correction submodule can identify this weather correlation, avoiding misinterpreting normal fluctuations caused by weather as health deterioration. Third, correction is particularly useful for adjusting the schedules of nursing homes. For example, during group mealtimes, the elderly person sits in the dining room, and their activity level drops sharply. Without correction, the system might mistakenly believe that the elderly person is inactive and in danger. After correction, the system knows that it is normal mealtime and will not trigger an alarm. Finally, this correction mechanism makes the dynamic baseline model more robust and accurate. The true abnormalities of individual elderly persons (such as sudden gait instability) can be highlighted from the corrected baseline, thereby improving the sensitivity and specificity of anomaly detection. The entire system thus possesses stronger environmental adaptability, maintaining stable monitoring performance under different seasons, weather conditions, and schedules, truly achieving reliable support 24 / 7, 365 days a year.

[0048] In some embodiments, when calculating the degree of deviation, the prediction module 3 uses the cumulative deviation integral value within the sliding window as the basis for judgment. Only when the cumulative deviation integral value exceeds a preset threshold will a prediction result of the warning level be generated. The preset threshold is an adaptive threshold that is automatically adjusted based on the recent average deviation level of the elderly person being accompanied.

[0049] The prediction module 3 uses the cumulative deviation integral value within a sliding window as the basis for judging the degree of deviation. In a preferred embodiment, the system maintains a circular buffer with a length of 30 seconds, collecting the difference between the current state data and the dynamic baseline mean once per second, and accumulating the absolute value of the difference into the cumulative sum of the buffer. When a new second of data enters, the oldest second of data is removed, and the cumulative sum is updated accordingly. Only when this cumulative sum exceeds a preset threshold does the prediction module 3 output the prediction result of the warning level. The preset threshold is not fixed, but is automatically adjusted according to the recent average deviation level of the elderly person being accompanied. Specifically, the system also maintains a 24-hour rolling average, recording the average deviation per minute over the past day, and then multiplies this average by a preset sensitivity coefficient to obtain the new threshold. For example, if the elderly person's recent condition is stable and the average deviation is small, the threshold will automatically decrease, so that even small, continuous deviations can be detected in time; if the elderly person has a large, continuous deviation due to a cold or other reasons, the threshold will automatically increase to avoid repeated alarms due to short-term fluctuations. Another feasible implementation is to use an exponentially weighted moving average instead of a fixed-length sliding window, giving more weight to recent data. This makes the cumulative deviation integral respond faster to sudden changes and less sensitive to slow drift. In a third implementation, the system uses a cumulative sum (CUSUM) algorithm, setting two cumulative values ​​to detect positive and negative deviations respectively. An alert is triggered when either cumulative value exceeds a dynamic threshold. This method is more sensitive to small but persistent shifts and is suitable for detecting chronic health deterioration.

[0050] This sliding window cumulative deviation integration combined with adaptive threshold mechanism brings significant benefits. First, sliding window integration effectively suppresses false alarms caused by transient noise. For example, an elderly person's occasional cough causes a sudden jump in respiratory data. If a single-point threshold is used, it is easy to trigger a false alarm. However, the integration method requires a sustained deviation over a period of time to reach the threshold. Transient events like coughing will not trigger an alarm, greatly reducing ineffective responses from caregivers. Second, adaptive threshold addresses the issue of varying sensitivities to deviation among different elderly individuals at different times. For agile elderly individuals, a slight decrease in activity may not be abnormal, but for a frail elderly person recently discharged from the hospital, the same degree of deviation may indicate danger. Adaptive threshold learns the elderly person's recent average deviation level and automatically adjusts the threshold to an appropriate range, avoiding the dilemma of fixed thresholds being either too sensitive or too insensitive. Third, this mechanism reduces the difficulty of system debugging and deployment. Caregivers do not need to manually set alarm thresholds for each elderly person. The system automatically learns and continuously optimizes after startup, truly achieving a smart, out-of-the-box experience. Finally, because the warning is only triggered when the cumulative deviation continues to exceed the limit, the elderly will not develop a sense of "crying wolf" due to frequent system alarms, and caregivers' trust in the warnings will increase, enabling them to respond quickly in real emergencies. The entire prediction module 3 therefore possesses extremely high robustness and practicality, serving as a key technological guarantee for the system's transition from laboratory prototype to actual deployment in nursing homes.

[0051] In some embodiments, the camera is also used to capture facial expression images of the elderly person being accompanied; the behavioral data also includes the emotional state obtained through facial expression recognition, which is used to assist the prediction module 3 in judging the level of attention.

[0052] While capturing full-body images, the camera also specifically captures facial expression images of the elderly person being accompanied. In a preferred embodiment, the system mounts a telephoto lens camera on the robot's head. When the robot is in a following state or actively greeting a person, it uses a face detection algorithm to lock onto the elderly person's facial region and captures facial images at a frequency of 10 frames per second. These images are fed into a pre-trained lightweight convolutional neural network trained on the FER2013 dataset, capable of outputting probability distributions for seven basic emotions: happiness, sadness, anger, fear, surprise, disgust, and neutrality. The system selects the emotion category with the highest probability as the current emotional state and records the confidence level of that emotion. The emotional state is incorporated into behavioral data and transmitted to prediction module 3 along with gait parameters, activity levels, etc. When processing attention levels, prediction module 3 additionally considers the emotional state: if physiological and behavioral data show a slight deviation (e.g., a slight decrease in walking speed), and the emotional state is sadness or fear, prediction module 3 raises the attention level to warning; if the emotional state is happiness or neutral, it maintains the attention level or lowers it to normal. This cross-modal fusion judgment significantly improves classification accuracy.

[0053] In another feasible implementation, instead of discrete emotion classification, the system uses a Facial Action Coding System (FACS) to extract the intensity of facial action units (AUs), such as the degree of mouth twitching and eyebrow furrowing. These continuous values ​​are then used as emotion features input into the auxiliary decision logic of prediction module 3. This approach is more sensitive to subtle facial expression changes, but requires slightly more computation. In a third implementation, the system combines speech emotion recognition with the acquisition of the elderly person's tone, speech rate, and volume via microphone. This data is then analyzed in conjunction with facial expressions to form a multimodal emotion judgment, further reducing the risk of misjudgment from a single modality.

[0054] This approach of incorporating facial expression emotional states has yielded several beneficial effects. First, emotional states provide an independent verification dimension for prediction module 3. When physiological and behavioral data present ambiguous signals (e.g., reduced activity in the elderly might be due to mood rather than physical discomfort), emotional states can help distinguish between psychological and physiological factors, avoiding misjudging low mood as a health crisis. Second, long-term tracking of emotional states can reveal trends in the elderly's mental health. For example, prolonged sadness or anger may indicate depressive tendencies, allowing the system to proactively alert caregivers and facilitate early intervention in mental health. Third, with emotional states assisting in attention level assessment, the system's responses to the elderly become more humane. For instance, when an elderly person exhibits fear, the robot prioritizes soothing strategies (such as playing calming music or offering gentle comfort) rather than mechanically executing standard greetings, enhancing user experience and trust. Finally, since facial expression recognition relies entirely on a camera without any additional sensors, it maintains the system's non-invasiveness and low cost, enabling seamless integration into existing hardware platforms.

[0055] In some embodiments, the baseline model generation module 2 is further configured to periodically recalibrate the dynamic baseline model. The recalibration timing includes: every morning, after each major event, and automatically when no warning has occurred for a consecutive week; wherein major events include discharge from the hospital or medication adjustment.

[0056] The baseline model generation module 2 is configured to periodically recalibrate the dynamic baseline model. Recalibration is performed automatically at midnight every day, after each major event, and when no warnings are detected for a consecutive week. Major events include discharge or medication adjustments. In a preferred embodiment, the system automatically triggers the recalibration process at 3:00 AM every day. At this time, the elderly are usually in deep sleep with minimal physical activity, and the collected respiratory rate and body movement data are most stable. The system extracts all valid data points from the past 24 hours, removes outliers caused by sensor obstruction or abnormal jumps, and then recalculates the mean and standard deviation. It then performs an exponentially weighted fusion with the old baseline, with the old baseline weighted at 0.3 and the newly calculated value weighted at 0.7, so that the baseline can both inherit historical trends and quickly respond to recent changes. When the system receives a discharge notice or medication change record from the nursing home management system, it immediately triggers an additional recalibration. At this time, the old baseline is directly reset to zero, and only the latest data collected after the event is used to rebuild the baseline, because discharge or medication change often means that the elderly's physical condition has fundamentally changed, and the old baseline has lost its reference value. If the system does not output any warning level results for seven consecutive days, it indicates that the current baseline may deviate from the actual condition of the elderly (too lenient or too strict). The system will automatically initiate a mild recalibration, incorporating the data from the past week into the baseline with higher weight, tightening the threshold range, and improving the sensitivity of subsequent anomaly detection.

[0057] Another feasible implementation uses a Bayesian online update method for recalibration. Instead of setting a fixed trigger time, the posterior distribution of the baseline is incrementally updated with each new data point collected. The system decides whether to accept the update based on the variance of the posterior distribution; if the variance is too large, the update is paused to prevent noise contamination. In a third implementation, the system maintains a 30-day sliding window. At midnight each day, the statistics of all data within the window are automatically calculated as the new baseline, and historical data outside the window are completely discarded. This method is simple to implement and can automatically adapt to aging trends up to one month long, but its response to sudden events is slow, requiring event-triggered calibration to compensate.

[0058] The benefits of this multi-time recalibration mechanism are significant. Daily recalibration at dawn ensures the baseline consistently reflects the elderly person's recent physical condition. For example, as age increases or rehabilitation progresses, respiratory rate and activity levels may change slowly; daily updates synchronize baseline drift, preventing systematic misjudgments caused by baseline stagnation. Immediate recalibration after major events prevents persistent false alarms caused by sudden changes in condition. For instance, if an elderly person's strength has not yet recovered after discharge and their activity level is significantly lower than before hospitalization, the system will continuously trigger alerts and even warnings if the baseline is not reset. After recalibration, the baseline matches the current reality, and false alarms disappear naturally. Automatic calibration after a week without warnings addresses the problem of an overly lenient baseline. Some elderly people may be in a low-activity state for a long time; as the system gradually adapts, it treats abnormalities as normal. Tightening the threshold restores sensitivity to subtle degradation. These three timing mechanisms complement each other, maintaining the timeliness and accuracy of the baseline model, enabling the system to operate stably for a long time without manual intervention, significantly reducing maintenance costs, and ensuring that the reliability of anomaly detection does not decay over time.

[0059] Secondly, this application provides an adaptive companionship method for elderly care service robots, implemented in any of the above embodiments, such as... Figure 2 As shown, it includes the following steps: S1: The status data of the elderly being accompanied is acquired through a non-invasive sensing module using at least one wearable sensor. The status data includes behavioral data and physiological indicator data. S2: Through the baseline model generation module, a dynamic baseline model is established based on the historical status data of the elderly being cared for during the initial monitoring stage. The dynamic baseline model includes the mean and standard deviation that are smoothly updated over time. S3: Through the prediction module, the current status data is compared with the dynamic baseline model, the degree of deviation is calculated, and the prediction results are generated based on the degree of deviation. The prediction results include three levels: normal, attention, and warning. S4: Through the decision-making module, determine the target strategy from multiple caregiving strategies based on the prediction results; S5: Through the execution module, the target strategy is executed to drive the robot to perform corresponding actions.

[0060] This embodiment provides an adaptive companionship method for elderly care service robots. This method is executed on the aforementioned system and achieves a complete closed loop from data collection to strategy execution through five steps. The following describes each step in detail with reference to the specific implementation method.

[0061] In step S1, a non-invasive sensing module uses at least one wearable sensor to acquire the elderly person's status data, including behavioral and physiological data. In a preferred embodiment, the sensing module uses a standard RGB camera and an ultrasonic ranging sensor. The camera captures a full-body image of the elderly person and extracts gait parameters and activity levels as behavioral data using a human skeletal key point detection algorithm. The ultrasonic sensor emits pulses toward the elderly person's chest and receives the reflected echoes, obtaining respiratory rate as physiological data by bandpass filtering and peak detection of the distance-time series. Throughout the process, the elderly person does not need to wear any devices, and the sensors are standard components of the robot. Another feasible implementation is to use a laser ranging sensor instead of ultrasonic, which offers higher accuracy but slightly increases the cost. A third implementation uses millimeter-wave radar, which can simultaneously detect respiration and heart rate, suitable for high-end models. The design intention of step S1 is to break free from the constraints of traditional wearable devices, lower the barrier to entry for the elderly, and simultaneously achieve zero-cost expansion using the robot's existing hardware.

[0062] In step S2, a dynamic baseline model is established using the baseline model generation module based on the historical status data of the elderly person being cared for during the initial monitoring phase. This model includes a mean and standard deviation that are smoothly updated over time. In a preferred embodiment, the system enters a silent learning mode for the first three days after deployment, collecting data without intervention, gathering behavioral and physiological data of the elderly person in a normal state, and calculating the initial mean and standard deviation. Subsequently, an exponentially weighted moving average is used to update the baseline daily, allowing it to slowly adapt to natural fluctuations caused by age, rehabilitation, or seasonal changes. Another implementation uses a sliding window statistics method with a window length of seven days, discarding the oldest day's data and adding the latest day's data each day for recalculation, achieving a smoother update. A third implementation introduces Bayesian online estimation, where the prior distribution is updated with the posterior distribution for each new data point, and the variance automatically adjusts the update rate. The purpose of establishing a personalized dynamic baseline is to create a unique health reference system for each elderly person, avoiding false alarms or false negatives caused by uniform thresholds, while preventing baseline aging through continuous updates.

[0063] In step S3, the prediction module compares the current state data with the dynamic baseline model, calculates the degree of deviation, and generates a prediction result based on the degree of deviation. The prediction result includes three levels: normal, attention, and warning. In a preferred embodiment, the system uses the cumulative deviation integral value within a sliding window as the judgment criterion. Specifically, a circular buffer with a length of 30 seconds is maintained, and the absolute value of the difference between the current value and the baseline mean is calculated and accumulated every second. A warning level is only output when the cumulative sum exceeds a preset threshold. The threshold itself is automatically adjusted based on the elderly person's recent average deviation level, for example, by multiplying the average deviation amount of the past 24 hours by a sensitivity coefficient as the new threshold. Another implementation uses a cumulative sum algorithm to monitor positive and negative deviations separately, which is more sensitive to small, persistent deviations. A third implementation uses a machine learning classifier (such as a random forest) to classify multidimensional deviation features, but requires more training data. The core value of the graded prediction mechanism lies in suppressing instantaneous noise false alarms, while adapting to the sensitivity differences of different elderly people through adaptive thresholds, so that caregivers can maintain trust in the warnings.

[0064] In step S4, the decision-making module determines the target strategy from multiple caregiving strategies based on the prediction results. In a preferred embodiment, the strategy library includes three levels of response: the normal level executes routine companionship strategies, such as playing music or engaging in voice chat; the attention level executes proactive greetings and close observation strategies, with the robot moving one meter in front of the elderly person to inquire if they need help and zooming in on their facial expressions; the warning level executes emergency notification to caregivers and a deceleration and following strategy, with the robot sending an alarm to the care station, reducing its movement speed, and activating warning lights. In another implementation, at the attention level, the robot can deliver a cup of warm water (if equipped with a robotic arm) to increase physical care. In a third implementation, at the warning level, the robot simultaneously dials a preset emergency number and activates hands-free calling. The design philosophy of the tiered strategy is "no disturbance when there is no need, and immediate response when there is a need," ensuring a rapid response in dangerous situations while minimizing interference.

[0065] In step S5, the execution module executes the target strategy to drive the robot to perform corresponding actions. In a preferred embodiment, the execution module controls the chassis movement via a motor driver, plays audio via a speaker, sends network messages via a Wi-Fi module, and illuminates warning lights via an LED controller. All actions are driven by a specific sequence of instructions issued by the decision module; the execution module is only responsible for underlying hardware scheduling. In another embodiment, the execution module also includes a robotic arm control interface for performing complex operations such as delivering items. The existence of the execution module transforms the decision result into a real impact in the physical world, completing a full closed loop from perception to action.

[0066] Overall, this method eliminates the burden of wearing protective gear through non-invasive sensing, solves the adaptability problem of uniform thresholds through personalized dynamic baselines, suppresses false alarms and improves warning reliability through hierarchical prediction, achieves humanized companionship through differentiated strategies, and completes closed-loop control through the execution module. These five interconnected steps form a complete intelligent chain from data collection to strategy execution, enabling elderly care service robots to provide 24 / 7, personalized, and highly reliable adaptive companionship without increasing the burden on the elderly or relying on expensive hardware. This method is particularly suitable for centralized care scenarios in nursing homes, significantly reducing the workload of caregivers while improving the elderly's sense of security and quality of life.

[0067] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0068] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive companionship system for elderly care service robots, characterized in that, include: The non-invasive sensing module (1) is configured to acquire the status data of the elderly being accompanied by at least one wearable sensor, the status data including behavioral data and physiological index data; The baseline model generation module (2) is configured to establish a dynamic baseline model based on the historical status data of the elderly being cared for during the initial monitoring stage. The dynamic baseline model includes the mean and standard deviation that are updated smoothly over time. The prediction module (3) is connected to the non-invasive sensing module (1) and the baseline model generation module (2), and is configured to compare the current state data with the dynamic baseline model, calculate the degree of deviation, and generate a prediction result based on the degree of deviation. The prediction result includes three levels: normal, attention and warning. The decision module (4) is connected to the prediction module (3) and is configured to determine the target strategy from multiple care strategies based on the prediction results; The execution module (5) is connected to the decision module (4) and is configured to execute the target strategy to drive the robot to perform corresponding actions.

2. The system according to claim 1, characterized in that, The non-invasive sensing module (1) includes a ranging sensor and a camera; the ranging sensor is configured to emit ranging signals toward the chest or back of the elderly person being accompanied and receive reflected signals, and the physiological index data is obtained by bandpass filtering and peak detection of the round-trip time series of the ranging signal to obtain respiratory frequency; the camera is configured to acquire a full-body image of the elderly person being accompanied, and the behavioral data is obtained by detecting key points of human skeleton and performing time-series analysis on the full-body image to obtain gait parameters and activity levels.

3. The system according to claim 2, characterized in that, The ranging sensor is an ultrasonic sensor or a laser ranging sensor. The ranging sensor is installed at the front end of the robot and can adjust the pitch angle with the robot's head or robotic arm so that the ranging signal is always aligned with the chest area of ​​the elderly person being accompanied when the elderly person is in a sitting or lying position.

4. The system according to claim 2, characterized in that, The robot body is a movable chassis structure. The non-invasive sensing module (1) also includes a follow control unit. This unit controls the robot to follow the elderly person being accompanied at a preset distance based on the image recognition results of the camera, and keeps the straight path between the distance measuring sensor and the chest of the elderly person being accompanied free of obstacles during the following process.

5. The system according to claim 1, characterized in that, The caregiving strategies stored in the decision module (4) include: when the prediction result is normal, a routine caregiving strategy is executed, which includes playing music or having a voice chat; when the prediction result is attention, an active greeting and close observation strategy is executed, which includes the robot moving to a preset distance in front of the elderly person being cared for and asking if they need help; when the prediction result is a warning, an emergency notification to caregivers and a deceleration and following strategy are executed, which includes the robot reducing its movement speed and turning on the warning lights.

6. The system according to claim 1, characterized in that, The dynamic baseline model also includes an external factor correction submodule, which adjusts the mean and standard deviation for offset based on the current time.

7. The system according to claim 1, characterized in that, When calculating the degree of deviation, the prediction module (3) uses the cumulative deviation integral value in the sliding window as the basis for judgment. Only when the cumulative deviation integral value exceeds the preset threshold will the prediction result of the warning level be generated. The preset threshold is an adaptive threshold that is automatically adjusted based on the recent average deviation level of the elderly person being cared for.

8. The system according to claim 2, characterized in that, The camera is also used to capture facial expression images of the elderly person being accompanied; the behavioral data also includes the emotional state obtained through facial expression recognition, which is used to assist the prediction module (3) in judging the level of attention.

9. The system according to claim 1, characterized in that, The baseline model generation module (2) is also configured to periodically recalibrate the dynamic baseline model. The recalibration timing includes: every morning, after each major event, and when no warning has occurred for a consecutive week. The major events include discharge from the hospital or medication adjustment.

10. An adaptive companionship method for an elderly care service robot, executed in the system described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Acquire the status data of the elderly person being accompanied. The status data is acquired through at least one wearable sensor and includes behavioral data and physiological indicator data. S2: Based on the historical status data of the elderly being cared for during the initial monitoring phase, a dynamic baseline model is established, which includes the mean and standard deviation that are updated smoothly over time. S3: Compare the current state data with the dynamic baseline model, calculate the degree of deviation, and generate a prediction result based on the degree of deviation; S4: Determine the target strategy from multiple caregiving strategies based on the prediction results; S5: Execute the target strategy to drive the robot to perform the corresponding action.