Internet-of-things management method for health care robot

Abstract

The invention discloses an internet-of-things management method of a health care robot, and relates to the technical field of health care robot management, an LSTM time sequence prediction algorithm is introduced on the basis of perception regulation and control to construct an LSTM model, and feature extraction is performed on time sequence data of temperature and distance sensors through an anomaly prediction module; a pre-trained LSTM (Long Short Term Memory) model is utilized to predict abnormal risk levels of future temperature and distance sensors, a heating module and a displacement module are controlled to operate based on the abnormal risk levels to implement differential linkage compensation, a pitch angle, a roll angle and an acceleration vector of the health care robot are calculated in real time, and safety early warning of corresponding levels is generated based on corresponding safety thresholds. And after the posture of the health care robot returns to normal, automatically entering a safety self-inspection mode, and recovering the initial regulation and control process after no abnormity exists. According to the management method, through deep fusion of multi-modal perception and an intelligent algorithm, robustness and safety risk controllability of the health care robot under the abnormal working condition in the health care process are achieved.

Description

Technical Field

[0001] This invention relates to the field of health and wellness robot management technology, and specifically to an IoT management method for health and wellness robots. Background Technology

[0002] Moxibustion is a traditional Chinese medicine external treatment method that uses the heat and medicinal effects generated by burning mugwort as a stimulus to act on specific acupoints or areas of the body to warm and tonify Yang energy, dispel cold and dampness, and dredge the meridians. Because it is relatively simple to perform and has a wide range of indications, it is especially suitable for people with a cold constitution, weakness, and chronic cold-dampness pain, and has become a common method for TCM health preservation and home healthcare.

[0003] In traditional moxibustion health care robots, the sensors are easily affected by factors such as dust blockage and component aging during long-term use, resulting in data drift or response delay. Existing solutions lack effective anomaly prediction and cross-modal compensation mechanisms, causing the accuracy of health care to decrease significantly over time, making it difficult to meet the comprehensive requirements of modern health care scenarios for equipment intelligence, reliability, and safety. Summary of the Invention

[0004] The purpose of this invention is to provide an IoT management method for health and wellness robots to address the shortcomings in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an IoT management method for a health and wellness robot, the management method comprising the following steps: S1: Data is collected and transmitted to the controller. The distance-temperature correlation algorithm model is used to determine the optimal temperature range and gradient change law corresponding to different distance intervals. The controller uses the PID algorithm to control the heating module and the displacement module in a coordinated manner. S2: Based on sensing and control, an LSTM time-series prediction algorithm is introduced to construct an LSTM model. The anomaly prediction module extracts features from the time-series data of temperature and distance sensors. The pre-trained LSTM model is used to predict the future anomaly risk level of temperature and distance sensors. Based on the anomaly risk level, the operation of the heating module and displacement module is controlled to implement differentiated linkage compensation. S3: Real-time calculation of the pitch angle, roll angle and acceleration vector of the health care robot, and generation of corresponding safety warnings based on the corresponding safety thresholds. After the health care robot's posture returns to normal, it automatically enters the safety self-check mode. After no abnormalities are found, the initial control process is restored.

[0006] Preferably, the differentiated linkage compensation between the heating module and the displacement module based on the abnormal risk level control includes the following steps: If a moderate anomaly is predicted in the distance sensor, the motor operating status data is retrieved, and the stepper motor step angle and transmission ratio are combined to deduce the actual displacement distance of the moxibustion head and the direction of deviation from the theoretical value. The distance-temperature correlation algorithm model is then used to deduce the temperature compensation amount corresponding to the deviation in reverse, and the heating power is adjusted in advance. If a moderate anomaly is predicted in the temperature sensor, the actual temperature range of the acupuncture point is calculated using a distance-temperature correlation algorithm model based on historical temperature-distance-ambient temperature three-dimensional correlation data, current motor displacement parameters, and ambient temperature compensation values. The heating power is then dynamically adjusted to the corresponding level.

[0007] Preferably, the anomaly prediction module extracts features from the time-series data of temperature and distance sensors, and uses a pre-trained LSTM model to predict the future anomaly risk level of the temperature and distance sensors, including the following steps: A continuous time-series data stream is acquired with a fixed sampling period to form a sliding time window. Each time, a sampling point is moved forward to generate training and inference samples. Multidimensional features are constructed for the original sequences within each time window, including distance data features, temperature data features, and cross-modal correlation features; Supervised learning was performed using a labeled historical anomaly dataset, with anomaly labels categorized into three levels: mild anomaly, moderate anomaly, and severe anomaly. The LSTM network structure uses two hidden layers. The input is a sequence of multidimensional feature vectors, and the output is the probability distribution of future abnormal risk levels.

[0008] Preferably, when the LSTM prediction or real-time rule-based determination indicates that the distance sensor has entered a moderate anomaly, a cross-modal compensation mechanism is activated to use motor operating data to infer the actual distance and complete temperature-linked correction. Read the stepper motor's operating status data, including forward and reverse direction indicators, pulse count, and cumulative running time. Based on the known stepper motor step angle and transmission ratio, calculate the theoretical displacement. Theoretical displacement = number of pulses × step angle / 360° × lead. Compare the theoretical displacement with the value reported by the distance sensor to obtain the direction and magnitude of the deviation. Substitute the deviation distance into the distance-temperature correlation algorithm model and calculate the amount of temperature change to be compensated according to the gradient change law learned by the model. By continuously comparing the actual displacement fed back by the motor encoder with the theoretical displacement calculated by the model, the difference is used as a new PID control error input to the displacement control module to dynamically adjust the position of the stepper motor.

[0009] Preferably, if the temperature sensor shows a moderate abnormality, the actual acupuncture point temperature is calculated using historical correlation data and cross-modal information, and dynamic power adjustment and subsequent calibration are implemented. We retrieve your three-dimensional correlation data of temperature, distance, and ambient temperature to form a training / reference library. Each sample record includes the measured temperature, actual distance, ambient temperature, and heating power level. Based on the current motor displacement parameters and ambient temperature compensation rules, the distance-temperature correlation algorithm model is invoked to reversely calculate the range of the actual temperature of the moxibustion point; Based on the calculated actual temperature range, the controller recalculates the target temperature and generates the corresponding PWM duty cycle to drive the heating module to output power, so that the temperature of the moxibustion point reaches the expected comfortable range.

[0010] Preferably, the distance data features include calculating the step change rate, mean drift magnitude, and fluctuation variance; the temperature data features include calculating the variance fluctuation, short-term gradient change, and the degree of persistent deviation from the ambient temperature difference; and the cross-modal correlation features include introducing the time delay correlation between distance changes and temperature changes.

[0011] Preferably, generating a corresponding level of security alert based on the corresponding security threshold includes the following steps: In the first-level warning, if the pitch angle θ > the first pitch angle threshold or the roll angle φ > the roll angle threshold, the user is prompted to adjust the posture of the health care robot. In the secondary intervention, if the pitch angle θ > the second pitch angle threshold or the linear acceleration component a_z in the Z-axis direction > the first component, attitude compensation is initiated. In a Level 3 emergency, if the pitch angle θ > the third pitch angle threshold or the linear acceleration component a_z in the Z-axis direction > the second component, it is determined to be an unexpected overturning / large-scale abnormal movement, and the three-break, one-lock, and one-alarm operation is executed.

[0012] Preferably, the three-disconnection-one-lock-one-alarm operation includes the following steps: The power supply circuit to the heating element is disconnected by a relay, a brake signal is sent to the stepper motor driver, the servo control signal is disabled, the current position is maintained, and an emergency alarm message is sent to the cloud health management platform via the MQTT protocol. The emergency alarm message includes the health care robot ID, abnormal timestamp, posture data snapshot, and temperature / distance curve, and is simultaneously pushed to the user APP and medical staff terminal and activates local audible and visual alarms.

[0013] Preferably, once the health care robot's posture returns to normal, it automatically enters a safety self-check mode, including the following steps: Verify the zero point of the temperature sensor and the validity of the distance sensor; The motor and heating module were gradually unlocked, and the displacement accuracy and temperature stability were verified through trial operation. Once no abnormalities are detected, the initial control procedure will be resumed.

[0014] Preferably, the controller uses a PID algorithm to control the heating module and the displacement module in a coordinated manner, including: The heating module uses a PWM signal to adjust the on / off frequency of the solid-state relay, thereby dynamically changing the power of the heating element. For the displacement module, the stepper motor / servo motor is driven to adjust the position of the moxibustion head so that the distance and temperature are within the comfortable coupling range defined by the distance-temperature correlation algorithm model.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention achieves forward-looking prediction of sensor failures by extracting features from time-series data of temperature and distance sensors (such as distance step change rate and temperature variance fluctuation) and predicting the future anomaly risk level (mild jump, moderate drift / delay, severe failure) using a pre-trained model. This counteracts the interference of sensor anomalies on the accuracy of health and wellness care, significantly improves the equipment's anti-disturbance capability and reliability under long-term use or complex environments (such as sensor performance degradation and environmental interference), and avoids control inaccuracies caused by single-point sensor failures.

[0016] This invention relies on a distance-temperature correlation algorithm model trained based on massive human thermal comfort experimental data to clarify the optimal temperature range and gradient change law for different distance intervals, so that the temperature output is deeply matched with the physiological needs of the human body. Then, through the linkage control of the heating module (PWM-regulated solid-state relay) and the displacement module (stepper motor / servo motor position adjustment) by the PID algorithm, a seamless adaptation of temperature to distance changes is achieved. For example, when the user's body moves slightly and the distance changes, the temperature can be dynamically adjusted accordingly, effectively maintaining the stability of the sensation and the accuracy of health care during the moxibustion process, and solving the problem of decreased comfort caused by distance-temperature coupling mismatch in traditional moxibustion equipment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the management method of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example: This example provides an IoT management method for a senior care robot. Please refer to [link / reference]. Figure 1 As shown, the management method includes the following steps: S1: Deploys three core sensors: a high-precision temperature sensor (to collect the temperature of the moxibustion point and the environment), an infrared ranging sensor (to simultaneously monitor the distance between the moxibustion head and the human skin), and a six-axis gyroscope (to sense the posture of the health and wellness robot). All data is filtered and noise-reduced before being synchronously transmitted to the ESP32-C5 controller. Temperature and distance data serve as the core basis for regulation and control, and are integrated into a distance-temperature correlation algorithm model. This model is trained based on massive amounts of human thermal comfort experimental data, clearly defining the optimal temperature range (e.g., 38-52℃) and gradient change law corresponding to different distance intervals (e.g., 20-50mm). (e.g., for every 5mm increase in distance, the temperature needs to be increased by 1.5-2℃ to maintain a stable sensation.) The controller implements linkage control for two types of actuators based on a PID algorithm: for the heating module, the switching frequency of the solid-state relay is adjusted through a PWM signal to dynamically change the power of the heating element; for the displacement module, the stepper motor / servo motor is driven to adjust the position of the moxibustion head, ensuring that the distance and temperature are always within the comfortable coupling range defined by the distance-temperature correlation algorithm model. For example, when the distance between the user and the robot increases from 30mm to 40mm due to slight body movement, the temperature is simultaneously increased from 45℃ to 48℃; conversely, when the distance decreases to 25mm, the temperature is reduced to 42℃, achieving seamless temperature adjustment as the distance changes. Simultaneously, a six-axis gyroscope monitors the robot's posture in real time. If a tilt angle >15° or a non-axial acceleration >0.5g (indicating a risk of tipping over) is detected, a warning threshold is immediately triggered, paving the way for subsequent safety responses.

[0021] S2: Building upon sensing and control, an LSTM time-series prediction algorithm is introduced to construct a robust LSTM model encompassing anomaly prediction, cross-modal compensation, and error correction, addressing the interference of potential sensor failures on the accuracy of healthcare services. The anomaly prediction module extracts features from the time-series data of temperature and distance sensors (e.g., step rate of change in distance data, variance volatility of temperature data), and uses a pre-trained LSTM model to predict the anomaly risk level within the next 3-5 seconds (mild: occasional data jumps; moderate: continuous drift / response delay; severe: complete failure). Differentiated, interconnected compensation is implemented for different sensor anomalies. For moderate anomalies in the distance sensor (e.g., data drift of ±3mm due to dust obstruction of the infrared sensor): retrieve motor operating status data (forward and reverse direction, number of pulses, cumulative running time), combine it with the stepper motor step angle (e.g., 1.8° / step) and transmission ratio (e.g., 10:1), and reverse-calculate the actual displacement distance of the moxibustion head and the direction of deviation from the theoretical value (e.g., theoretical displacement of 10mm, actual displacement of only 8mm); call the distance-temperature correlation algorithm model to reverse-calculate the temperature compensation amount corresponding to this deviation (e.g., if the actual distance is 2mm larger than the theoretical distance, the temperature needs to be increased by 1℃), and adjust the heating power in advance; further offset the sensor error through displacement closed-loop PID correction to ensure that the distance-temperature linkage is not affected by sensor anomalies.

[0022] For moderate temperature sensor anomalies (e.g., NTC thermistor readings are 2°C lower due to aging): Based on the three-dimensional correlation data of temperature-distance-ambient temperature over the past 7 days (sampled hourly to build a database of 100,000+ samples), current motor displacement parameters (converted to actual distance), and ambient temperature compensation values ​​(e.g., when the room temperature is 25°C, the moxibustion point temperature needs to be corrected by an additional 1°C), the actual temperature range of the moxibustion point is estimated through the distance-temperature correlation algorithm model (e.g., if the sensor displays 43°C, the actual temperature should be 45°C), and the heating power is dynamically adjusted to the corresponding level. At the same time, the temperature deviation trend during abnormal periods (e.g., 1.5°C lower per hour) is recorded, a calibration command is generated and temporarily stored, and the calibration program is automatically triggered when the sensor is maintained to avoid the accumulation of anomalies affecting long-term accuracy.

[0023] S3: Relying on the deployed six-axis gyroscope, the system calculates the pitch angle (θ), roll angle (φ), and acceleration vector (a_x, a_y, a_z) of the health and wellness robot in real time, and sets three levels of safety thresholds. Level 1 warning (θ > 10° or φ > 10°, prompting the user to adjust the robot's posture), Level 2 intervention (θ > 15° or a_z > 0.3g, initiating posture compensation), Level 3 emergency (θ > 20° or a_z > 0.5g, determined as accidental tipping / significant abnormal movement). Once the Level 3 emergency threshold is triggered, the ESP32-C5 controller executes a three-break, one-lock, one-alarm operation: The system cuts off the heating power supply (by disconnecting the power supply circuit to the heating element via a relay to ensure no continuous heat source); locks the motor (by sending a brake signal to the stepper motor driver to prevent the moxibustion head from shifting due to inertia); locks the displacement module (by disabling the servo control signal to maintain the current position and avoid secondary collisions); sends emergency alarm information (including the health and wellness robot ID, abnormal timestamp, posture data snapshot, and temperature / distance curve) to the cloud-based health management platform via the MQTT protocol, and simultaneously pushes it to the user's APP and the medical staff's terminal; and activates the local audible and visual alarm (high-frequency buzzer sound + flashing red LED light) to alert on-site personnel to take action.

[0024] Once the health care robot's posture returns to normal (θ < 5° and a_z < 0.1g for 10 seconds), it will automatically enter the safety self-check mode. First, verify the zero point of the temperature sensor (short-circuit the temperature measuring pin to detect the reference voltage) and the validity of the distance sensor (emit infrared signal to detect the reflection intensity). Then, gradually unlock the motor and heating module, and verify the displacement accuracy and temperature stability through low-speed test run (displacement speed reduced to 10% of the rated value). After confirming that there are no abnormalities, restore the initial control process.

[0025] In this application, the health and wellness robot system adopts a three-layer collaborative architecture of cloud-edge-terminal. Its core lies in the deep integration of automated execution hardware, cloud-based intelligent scheduling and data analysis platform and standardized business front-end to form a complete solution.

[0026] The terminal execution layer (automated moxibustion robot) consists of an intelligent moxibustion and health care robot equipped with an ESP32-C5 controller and integrating multiple sensors and actuators. Its role is to replace manual techniques, ensuring the precise, automated, and safe execution of the moxibustion process. In the edge communication and control layer, the health care robot maintains a real-time and stable connection with the cloud via the MQTT protocol for command and data flow. The LVGL graphical interface (developed based on ESP-IDF) running on the health care robot provides basic status monitoring and emergency intervention interfaces, ensuring basic controllability in complex network environments. The cloud platform layer (server and data hub) uses Go language to build a high-concurrency, highly reliable backend service cluster. Its role is to act as the system brain, unifying resource scheduling, process monitoring, data analysis, and business logic processing. The user interaction layer uses the UniappX framework to develop customer and merchant mini-programs, providing a standardized, low-barrier-to-entry user interface, serving as a unified entry point connecting users, merchants, and system services.

[0027] In this embodiment, the management scheme is described in detail as follows: Phase S1 focuses on building a high-precision multimodal perception system for moxibustion therapy scenarios and establishing a data-driven closed-loop control architecture to achieve dynamic adaptation of distance and temperature and pre-sensing of safety status.

[0028] In one embodiment disclosed in this application, the physical deployment and electrical integration of three types of key sensors are completed: The high-precision temperature sensor uses a composite solution of contact thermocouple and infrared non-contact sensor, which are respectively attached to the moxibustion head treatment end face (contact type, used to directly collect the temperature of the moxibustion point tissue) and the outer ring array of the device (non-contact type, used to synchronously monitor the ambient background temperature). The sampling frequency is set to 10Hz to ensure dynamic response accuracy. The infrared ranging sensor adopts a miniature module based on the triangulation principle and is integrated at the center of the front end of the moxibustion head. The transmitting and receiving optical axes are perpendicular to the treatment end face, and the effective ranging range is 10-100mm (covering the interaction distance of the natural undulation of human skin). The sampling frequency is synchronously 10Hz. The six-axis gyroscope uses a low-noise MEMS inertial measurement unit (IMU) embedded in the geometric center of the main control cabin of the device. It communicates with the ESP32-C5 controller through the I2C bus and outputs raw data of three-axis angular velocity (±2000dps range) and three-axis linear acceleration (±16g range) at a frequency of 100Hz. The analog / digital signals from the three types of sensors are processed by independent signal conditioning circuits (including amplification and bias correction) and then uniformly input into the ADC interface (temperature, distance) and SPI / I2C interface (gyroscope) of the ESP32-C5 controller to form a standardized multi-source data acquisition channel.

[0029] To ensure the reliability of subsequent adjustments, all sensor data must undergo two stages of preprocessing: The first stage is hardware-level filtering. Temperature and distance signals are first passed through an RC low-pass filter (cutoff frequency 50Hz) to suppress high-frequency electromagnetic interference, while gyroscope data undergoes preliminary detrending and baseline drift correction through the built-in digital motion processor (DMP). The second stage is software-level filtering and noise reduction. A sliding window mean filter (window length N=5, corresponding to a 500ms time window) is used to perform an arithmetic average of temperature and distance data from 5 consecutive sampling points to eliminate instantaneous impulse noise. The gyroscope data is then filtered using a Kalman filter (the state variables are estimated values ​​of angular velocity and acceleration, the observed values ​​are the original sampled values, and the process noise covariance matrix is ​​initialized according to the nominal noise density specified in the sensor manual) to improve the stability of attitude calculation. The preprocessed data is timestamped (accuracy ±1ms) and synchronously transmitted to the controller's memory buffer through the ESP32-C5's dual-core asynchronous processing architecture (Core0 is responsible for data acquisition and preprocessing, and Core1 is responsible for algorithm calculation and control output), forming a multi-dimensional state snapshot stream of temperature, distance, and attitude.

[0030] In one embodiment disclosed in this application, temperature and distance data serve as the core control basis and need to be integrated with a distance-temperature correlation algorithm model trained based on massive human thermal comfort experimental data. The model construction process is as follows: A controlled experiment was conducted by recruiting no fewer than 200 healthy subjects (covering different ages, genders, and BMI indices). Under controlled conditions (temperature 25±1℃, humidity 50±5%RH), an orthogonal test combination of 20-50mm intervals (5mm step size) and 35-55℃ temperature intervals (1℃ step size) was set. Subjects' thermal comfort scores (using a 7-point scale, 1 = discomfort from cold, 7 = discomfort from heat) and skin surface temperatures (monitored by an infrared thermal imager) were recorded. Subsequently, three-dimensional correlation features between distance intervals, temperature intervals, and thermal comfort scores were extracted through feature engineering. After removing abnormal score samples (such as samples where subjects actively reported interference from non-temperature factors), a random forest regression algorithm was used to train the model. A thermal comfort score ≥5 was used as the comfort label. The model learned the optimal temperature range and gradient change patterns corresponding to different distance intervals. The specific training process of the distance-temperature correlation algorithm model is existing technology and will not be elaborated upon in this application.

[0031] The model's processing logic can be described as follows: Receive the real-time distance value D (unit: mm) output by the current infrared ranging sensor, and divide D into intervals (e.g., 20≤D<25→interval A, 25≤D<30→interval B, ..., 45≤D≤50→interval F); based on the interval to which D belongs, retrieve the target temperature range [T_min, T_max] corresponding to that interval from the model knowledge base (e.g., interval B (25-30mm) corresponds to [40℃, 46℃]); calculate the offset ΔD=D-D0 of the current D relative to the reference distance D0 (taking the lower limit of the interval, e.g., D0=25mm in interval B), and apply the gradient coefficients learned by the model (e.g., every 5mm offset corresponds to...). The target temperature increment ΔT is calculated as follows: ΔT = (ΔD / 5) × K (K ∈ [1.5, 2], the specific value can be dynamically adjusted according to user preferences or real-time thermal comfort score feedback, the default value is the median 1.75). Combined with the base temperature T_base of the current interval (take the midpoint of the interval, e.g., T_base = (40+46) / 2 = 43℃ for interval B), the final target temperature T_target = T_base + ΔT is generated (in the example, ΔD = 5mm (D = 30mm), then ΔT = 1.75℃, T_target = 44.75℃, rounded to 45℃). Through this logic, the model realizes the mapping from the original distance data to the target temperature, ensuring that the temperature output at different distances conforms to the physiological characteristics of human thermal comfort.

[0032] In one embodiment disclosed in this application, the controller implements PID linkage control on the heating module and the displacement module based on the target temperature T_target and the real-time acquired actual temperature T_actual (after filtering), as well as the target distance D_target (initial setting or manual fine-tuning by the user, defaulting to 30mm) and the actual distance D_actual. 1) Heating module control (PWM-regulated solid-state relay) Temperature error E_temp = T_target - T_actual; Incremental PID algorithm is called (to avoid integral saturation) to calculate control increment Δu_temp = Kp × (E_temp - E_temp_prev) + Ki × E_temp + Kd × (E_temp - 2 × E_temp_prev + E_temp_prev2), where Kp (proportional coefficient), Ki (integral coefficient), and Kd (derivative coefficient) are tuned using the Ziegler-Nichols method (example values: Kp = 2.5, Ki = 0.1, Kd = 1.0), and E_temp_prev and E_temp_prev2 are the temperature errors of the previous two measurements, respectively; The control increment Δu_temp is converted into PWM duty cycle adjustment (range 0-100%), which is output to the solid-state relay drive circuit through the LEDPWM controller of ESP32-C5 to control the on / off frequency of the heating element (the higher the duty cycle, the longer the heating element is powered on, and the greater the power), until the temperature error E_temp approaches 0 (control accuracy ±0.5℃).

[0033] 2) Displacement module control (stepper motor / servo motor position adjustment) Distance error E_dist = D_target - D_actual (Note: When D_actual deviates from D_target, the position of the moxibustion head needs to be adjusted to make D_actual approach D_target, and T_target is adjusted in conjunction with it); Similarly, the incremental PID algorithm is used to calculate the control increment Δu_dist = Kp' × (E_dist - E_dist_prev) + Ki' × E_dist + Kd' × (E_dist - 2 × E_dist_prev + E_dist_prev2) (Example parameters: Kp' = 1.2, Ki' = 0.05, Kd' = 0.8); Δu_dist is converted into stepper motor step command (or servo motor angle command), and the moxibustion head is controlled by the driver to move along the axis (perpendicular to the skin direction) until E_dist approaches 0 (control accuracy ±1mm).

[0034] Example of linkage: When the user's slight body movement causes D_actual to increase from 30mm (interval B) to 40mm (interval D), the model first calculates ΔD=40-30=10mm (spanning two 5mm steps), the gradient increment ΔT=10 / 5×1.75=3.5℃, and the original T_target=45℃ is adjusted to 48.5℃ (rounded to 49℃); at the same time, the displacement module detects E_dist=30-40=-10mm (the actual distance is greater than the target), the PID controller drives the stepper motor to retract the moxibustion head, so that D_actual falls back to around 30mm, ensuring that the distance and temperature are always within the comfortable coupling range defined by the model (such as the distance of 25-30mm corresponding to [40℃, 46℃] in interval B).

[0035] S2: Aiming to improve the system's anti-disturbance capability and stability in health and wellness precision under complex operating environments, this phase introduces a Long Short-Term Memory (LSTM) time-series prediction algorithm to establish a robust enhancement model with anomaly prediction, cross-modal compensation, and error correction capabilities. This addresses potential performance degradation or transient failures of temperature and distance sensors, preventing temperature-distance coupling mismatch caused by single-point sensor failure, thereby ensuring the continuity and safety of the moxibustion therapy process. This phase can be divided into four main modules: Temporal feature extraction and anomaly risk level prediction, cross-modal compensation and closed-loop correction for moderate anomalies in distance sensors, multi-source estimation and dynamic calibration for moderate anomalies in temperature sensors, and anomaly trend recording and subsequent self-maintenance triggering.

[0036] In one embodiment disclosed in this application, to predict the risk of sensor anomalies in the next few seconds, it is necessary to first abstract features from the historical time-series data of temperature and distance sensors, and then perform sequence learning and classification prediction using a pre-trained LSTM model. The processing logic is as follows: A continuous time-series data stream is acquired with a fixed sampling period (temperature and distance are both 10Hz), forming a sliding time window (the window length is set to 50 sampling points, corresponding to a duration of 5 seconds). Each time, one sampling point is slid forward to generate training and inference samples.

[0037] Multidimensional feature construction is performed on the original sequence within each time window, mainly including: Distance data characteristics: Calculate step change rate (the proportion of absolute difference between adjacent sampling points exceeding a preset threshold), mean drift magnitude (the difference between the mean in the window and the mean in the previous window), and fluctuation variance (the variance value of the data in the window).

[0038] Temperature data characteristics: Calculate variance fluctuation (reflecting temperature stability), short-term gradient change (temperature difference at the beginning and end of the window divided by the time span), and the degree of persistent deviation from the ambient temperature difference (the number of times the temperature difference has the same sign in multiple consecutive windows).

[0039] Cross-modal correlation features: Introduce the time delay correlation between distance changes and temperature changes (e.g., the number of lag frames in the temperature response as the distance increases).

[0040] Supervised learning was performed using a labeled historical anomaly dataset, with anomaly labels categorized into three levels: Mild: Occasional data fluctuations (briefly exceeding reasonable limits but quickly recovering); Moderate: Continuous drift or response delay (multiple consecutive windows exhibit unidirectional offset or response lag); Severe: Complete failure (data remains constant or unresponsive for an extended period).

[0041] The LSTM network structure uses two hidden layers (the number of neurons in each layer is optimized according to the data scale, for example 64 and 32). The input is the above multidimensional feature vector sequence, and the output is the probability distribution of the abnormal risk level in the next 3-5 seconds.

[0042] During the inference phase, the current window features are calculated in real time and input into the model to obtain the risk level prediction results for the next few seconds, and the results are provided as flags (risk_level∈{mild, moderate, severe}) for subsequent compensation strategies.

[0043] To predict the risk of sensor anomalies within the next few seconds, distance and temperature time-series data are acquired using a fixed sampling frequency of 10Hz. In one analysis, a sliding window with a length of 50 sampling points (5 seconds) is selected. For example, the distance data within the window is [30, 30.2, 30.5, 31, 35, 36, ...] (unit: mm). The step rate of change is first calculated: The percentage of adjacent points whose absolute difference exceeds a preset threshold (1 mm) is calculated. For example, if 3 pairs of 50 points have a difference exceeding 1 mm, then the step change rate = 3 / 49 ≈ 0.061. The mean drift is the difference between the mean of the current window and the mean of the previous window. Assuming the mean of the current window is 32.6 mm and the mean of the previous window is 30.0 mm, then the mean drift = 32.6 - 30.0 = 2.6 mm. The variance is the variance of the data within the window, calculated as follows: σ 2 ≈4.84 (standard deviation ≈2.2mm). The temperature window data is [45, 45.1, 45.3, 46, 47, 50, ...] (unit: °C), and the variance fluctuation is the temperature variance within the window, denoted as σ. 2_T≈2.25 (standard deviation≈1.5℃); the short-term gradient change is the temperature difference between the beginning and end of the window divided by the time span. The initial temperature is 45℃, the final temperature is 50℃, and the time span is 5 seconds. Therefore, the gradient change = (50-45) / 5 = 1.0℃ / s; the degree of continuous deviation from the ambient temperature refers to the number of times the temperature difference sign is consistent in consecutive windows. If the difference between the current window and the previous window relative to the ambient temperature of 25℃ is positive and this is the case for 5 consecutive windows, then this feature = 5.

[0044] Cross-modal correlation features, such as the number of lag frames in the temperature response after the distance increases: if the distance is detected to increase suddenly at second t, and the temperature only rises at second t+0.6, with a sampling period of 0.1 seconds, then the number of lag frames = 0.6 / 0.1 = 6 frames. These multidimensional features are combined into vectors and input into a pre-trained two-layer LSTM (64 and 32 neurons) for supervised learning. The labels are anomaly levels. For example, the model infers that the probability distribution of anomalies in the next 3-5 seconds is {mild: 0.1, moderate: 0.7, severe: 0.2}. The category with the highest probability is taken as moderate risk. Therefore, the risk_level flag is set to moderate during the inference stage. Subsequently, the distance sensor drift compensation strategy is triggered: the number of motor pulses and the step angle of 1.8° / step and the transmission ratio of 10:1 are retrieved to infer the actual displacement. If the theoretical displacement is 10mm but the sensor reports 12mm, the deviation is +2mm. According to the distance-temperature gradient law (each 5mm deviation corresponds to an adjustment of 1.5-2℃), it is calculated that the temperature needs to be adjusted by about 0.6-0.8℃, and the heating power is adjusted in advance to achieve the linkage between prediction and compensation.

[0045] In one embodiment disclosed in this application, when the LSTM prediction or real-time rule determines that the distance sensor has entered a moderate anomaly (such as infrared ranging data drift of ±3mm due to dust obstruction), a cross-modal compensation mechanism is activated to use motor operation data to back-calculate the actual distance and complete temperature-linked correction. Read the stepper motor's operating status data, including forward / reverse direction indicators, pulse count, and cumulative running time. Based on the known stepper motor step angle (e.g., 1.8° / step) and transmission ratio (e.g., 10:1), calculate the theoretical displacement: Theoretical displacement = Pulse count × Step angle / 360° × Lead (converted to linear displacement using the transmission ratio). Compare the theoretical displacement with the value reported by the distance sensor to determine the direction and magnitude of the deviation (Example: Theoretical displacement 10mm, sensor reported 12mm, deviation +2mm, indicating the sensor reading is too high).

[0046] Substitute the deviation distance into the distance-temperature correlation algorithm model integrated in S1, and calculate the amount of temperature change to be compensated according to the gradient change law learned by the model (e.g., 1.5–2℃ adjustment for every 5mm deviation). Example: If the actual distance is 2mm smaller than the sensor indication value, the original target temperature needs to be increased by about 0.6–0.8℃. The controller adjusts the heating power in advance to avoid insufficient temperature at the moxibustion point due to sensor drift.

[0047] By continuously comparing the actual displacement fed back by the motor encoder with the theoretical displacement calculated by the model, the difference is used as a new PID control error input to the displacement control module to dynamically adjust the position of the stepper motor so that the actual distance approaches the target distance.

[0048] When LSTM prediction or real-time rules determine that the distance sensor has entered a moderate anomaly, such as infrared ranging data drift of ±3mm due to dust obstruction, the system activates the cross-modal compensation mechanism, first reading the stepper motor's operating status data: Assuming the forward and reverse directions are considered forward, the pulse count is 200 steps, and the cumulative running time is used for verification but not directly involved in this calculation, given a step angle of 1.8° / step, a transmission ratio of 10:1, and a lead (linear displacement per revolution of the lead screw) of 10mm, the theoretical displacement calculation formula is: Theoretical Displacement = Pulse Count × Step Angle ÷ 360° × Lead × Transmission Ratio Correction. The transmission ratio correction is that the lead screw moves 10mm per revolution of the motor, meaning the linear displacement per step of the motor is (1.8° ÷ 360°) × 10mm = 0.05mm. Multiplying this by the transmission ratio of 10:1 gives the actual displacement per step = 0.05mm × 10 = 0.5mm. Therefore, the theoretical displacement = 200 steps × 0.5mm / step = 100mm. Assuming the distance sensor reports a value of 102mm, the deviation = sensor value – theoretical displacement = 102mm – 100mm = +2mm, indicating that the sensor reading is too high, meaning the actual distance is 2mm less than the sensor reading.

[0049] Substituting this deviation into the distance-temperature correlation algorithm model of S1, the model gradient law is that every 5mm deviation corresponds to a temperature adjustment of 1.5–2℃. The temperature change ΔT is calculated proportionally as (deviation ÷ 5) × adjustment coefficient. Taking the median value of 1.75℃ / 5mm, ΔT = (2 ÷ 5) × 1.75 = 0.7℃, meaning an additional adjustment of approximately 0.7℃ to the original target temperature. Based on this, the controller increases the heating power in advance to avoid insufficient temperature at the acupuncture point. Subsequently, the actual displacement fed back by the motor encoder (assuming feedback is 99.5mm) is continuously compared with the theoretical displacement of 100mm calculated by the model, resulting in a new error = 99.5mm – 100mm = –0.5mm. This error is used as the PID control input, adjusting the stepper motor pulse output to bring the displacement closer to the target. After several iterations, the actual distance can be stabilized within the target range, thus maintaining accurate distance-temperature coupling even when the sensor is drifting.

[0050] In one embodiment disclosed in this application, for moderate anomalies in temperature sensors (such as NTC thermistor aging causing a systematic 2°C drop in readings), historical correlation data and cross-modal information are used to estimate the true acupuncture point temperature, and dynamic power adjustment and post-calibration are implemented. We retrieved hourly temperature-distance-ambient temperature correlation data from the past 7 days to construct a training / reference library of approximately 100,000+ samples. Each sample record includes the measured temperature, actual distance, ambient temperature, and heating power level.

[0051] Based on the current motor displacement parameters (the actual distance is calculated through transmission) and the ambient temperature compensation rules (e.g., an additional 1℃ correction is needed when the room temperature is 25℃), the distance-temperature correlation algorithm model is called to inversely calculate the possible range of the actual temperature of the moxibustion point. Example: The sensor displays 43℃, but the model, combining the actual distance and room temperature, calculates that the actual temperature should be around 45℃, then it is determined that there is a system deviation of about 2℃.

[0052] Based on the calculated actual temperature range, the controller recalculates the target temperature and generates the corresponding PWM duty cycle, driving the heating module to output appropriate power so that the temperature of the moxibustion point reaches the expected comfortable range.

[0053] The system continuously records temperature deviation trends during abnormal periods (e.g., a rate of change of 1.5°C lower per hour), generates calibration instructions, and stores them in a temporary queue of non-volatile memory. When the sensor enters maintenance mode (e.g., when the user restarts the device or the system detects that maintenance conditions are met), the calibration procedure is automatically executed. By using a standard heat source or a known temperature environment as a reference, the calibration parameters of the sensor are updated to eliminate the impact of accumulated errors on long-term accuracy.

[0054] For moderate anomalies in temperature sensors (such as NTC thermistor aging causing a systematic 2°C underestimation of the reading), the system first retrieves hourly temperature-distance-ambient temperature three-dimensional correlation data from the past 7 days, forming a training / reference library of approximately 100,000+ samples. Each sample includes measured temperature, actual distance, ambient temperature, heating power level, and user thermal comfort rating. Assuming the current motor displacement is converted to an actual distance of 30mm and the ambient temperature is 25°C, according to the ambient temperature compensation rule, an additional 1°C correction is needed when the room temperature is 25°C. The distance-temperature correlation algorithm model integrated in S1 is called. This model corresponds to a target temperature of 44°C (base value) + 1°C (ambient compensation) = 45°C within a 30mm distance range. However, the temperature sensor displays 43°C in real time, indicating a system deviation of approximately 2°C. The controller recalculates the target temperature (taking the median of 45°C) based on the calculated actual temperature range [44.5°C, 45.5°C] and maps it to the PWM duty cycle. Assuming a linear relationship between temperature and duty cycle, with each 1°C corresponding to a 5% duty cycle change, and a reference temperature of 40°C corresponding to a 50% duty cycle, then the target duty cycle = 50% + (45°C – 40°C) × 5% = 75%. This generates a corresponding PWM signal to drive the heating module, bringing the acupuncture point temperature closer to the true comfort range. Simultaneously, the system continuously records the temperature deviation trend during abnormal periods. For example, if the sensor readings for five consecutive hours are 43°C, 41.5°C, 40°C, 38.5°C, and 37°C, and the deviation decreases by 1.5°C relative to the true value per hour, then the deviation change rate = 1.5°C / h. A calibration command is generated and stored in a non-volatile memory temporary queue. When the sensor enters maintenance mode (e.g., the user restarts the device or the system detects a deviation exceeding the threshold for 24 consecutive hours), the calibration procedure is automatically executed. The sensor output was measured to be 48℃ under a standard heat source (known temperature 50℃). The calibration correction value was calculated as: Actual temperature – Sensor reading = 50℃ – 48℃ = 2℃. The sensor calibration formula was then updated to: Calibration temperature = Original sensor reading + 2℃. This eliminated the impact of accumulated errors on long-term accuracy and ensured the accuracy of subsequent temperature measurements.

[0055] The LSTM prediction module runs continuously in the background, without affecting the real-time performance of the main control loop; the compensation strategy is activated only when a moderate or severe anomaly is detected. The compensation process is imperceptible to the user; heating and displacement control still use the S1 PID algorithm interface, only the input target value has undergone cross-modal correction. If the LSTM prediction indicates a severe anomaly (complete sensor failure), the system will, in addition to triggering the S1 safety warning (such as gyroscope attitude anomaly), add an emergency stop to heating or force entry into standby mode to prevent the risk of loss of control.

[0056] S3: Focusing on attitude safety monitoring and emergency protection during equipment operation, it relies on the deployed six-axis gyroscope (MEMSIMU) to capture the attitude angle and acceleration vector of the equipment in three-dimensional space in real time, constructs a three-level progressive safety threshold system, and formulates a complete closed-loop response logic from early warning prompts, attitude compensation, emergency protection to safety recovery.

[0057] In one embodiment disclosed in this application, a six-axis gyroscope outputs three-axis angular velocities (ω_x, ω_y, ω_z) and three-axis linear accelerations (a_x_raw, a_y_raw, a_z_raw) at a frequency of 100Hz. The system obtains usable attitude and acceleration information through the following processing logic: The Mahony complementary filtering algorithm is used to fuse angular velocity and acceleration data, filtering out angular velocity integral drift and accelerometer dynamic interference. The logic of the filtering process is as follows: Integrating the angular velocity yields a short-time attitude estimate; using the gravity vector direction from the accelerometer under static or quasi-static conditions, the pitch angle θ and roll angle φ are calculated; these two are then fused according to weights to output a stable attitude angle sequence (update period synchronized with IMU sampling). The output results are real-time pitch angle θ (rotation angle about the horizontal axis) and roll angle φ (rotation angle about the vertical axis), both expressed in degrees, with positive and negative signs indicating the tilt direction.

[0058] The raw acceleration data is processed to remove the gravity component: the components of gravity along each axis are calculated using the current attitude angle and subtracted from the raw data to obtain the pure motion acceleration vector (a_x, a_y, a_z). The resultant acceleration magnitude is calculated as |a| = sqrt(a_x). 2 +a_y 2 +a_z 2 (This is used to assess whether there is a sudden impact or fall.)

[0059] The six-axis gyroscope outputs three-axis angular velocities (ω_x, ω_y, ω_z) and three-axis linear accelerations (a_x_raw, a_y_raw, a_z_raw) at a frequency of 100Hz. The system uses the Mahony complementary filter algorithm for processing: assuming that within a certain sampling period, the angular velocities ω_x = 5° / s, ω_y = –3° / s, and ω_z = 0.5° / s, and the integration time interval Δt = 0.01s (corresponding to 100Hz), the short-time attitude change is obtained by integrating the angular velocities. Converting Δθ≈ω_x×Δt=5×0.01=0.05° and Δφ≈ω_y×Δt=–3×0.01=–0.03°, and summing these values ​​to the attitude from the previous moment, we obtain short-time attitude estimates θ_est=θ_prev+0.05° and φ_est=φ_prev–0.03°. Simultaneously, under static conditions, the accelerometer measures a_x_raw=0.1g, a_y_raw=–0.05g, and a_z_raw=0.98g (g≈9.8m / s²). 2 The gravity components along each axis are calculated using attitude angles: the gravity vector in the body coordinate system is [g·sinθ, –g·sinφ, g·cosθ·cosφ]. Assuming θ=2° and φ=–1°, the gravity components are approximately [9.8×sin2°, –9.8×sin(–1°), 9.8×cos2°×cos(–1°)]≈[0.342m / s²]. 2 0.171m / s 2 9.796m / s 2 Subtracting this from the original acceleration gives the pure acceleration a_x = 0.1g × 9.8 – 0.342 ≈ 0.98 – 0.342 = 0.638 m / s². 2, a_y=–0.05g×9.8–0.171≈–0.49–0.171=–0.661m / s 2 , a_z=0.98g×9.8–9.796≈9.604–9.796=–0.192m / s 2 ; Then, the attitude estimates from short-time attitude estimation and those from acceleration estimation are fused together with weights (e.g., giving higher weights to acceleration estimation results to suppress integral drift), outputting stable attitude angles θ≈1.8° and φ≈–0.9°. The resultant acceleration amplitude |a|=√(a_x) is then calculated. 2 +a_y 2 +a_z 2 ) = √(0.638 2 +(–0.661) 2 +(–0.192) 2 )≈√(0.407+0.437+0.037)=√0.881≈0.939m / s 2 (Approximately 0.096g), this small value indicates no significant impact; if an impact causes a_z_raw to suddenly increase to 2g, after removing gravity, a_z≈2g×9.8–9.796≈19.6–9.796=9.804m / s 2 The resultant acceleration |a|≈√(0.1) 2 +0.05 2 +9.804 2 )≈9.805m / s 2 (Approximately 1.0g), combined with an attitude angle θ>20°, a Level III emergency safety response can be triggered.

[0060] In one embodiment disclosed in this application, a three-level progressive threshold is set based on the attitude angle and acceleration amplitude, and the processing logic is as follows: Level 1 Warning (Informative): When θ>10° or φ>10°, the device posture is determined to deviate from the ideal vertical position. The user is prompted to adjust the device posture via a pop-up window in the user's APP and the device status light (a solid yellow light). This level does not interfere with the control process, but only provides information prompts.

[0061] Secondary intervention (proactive compensation): When θ > 15° or a_z > 0.3g (g ≈ 9.8 m / s²), 2 When the system detects a significant posture deviation or an initial impact, it temporarily increases the PID gain of the displacement module to enhance the moxibustion head's ability to maintain its position; it limits the maximum duty cycle of the heating power to a safe limit (e.g., 70%) to reduce the risk of localized overheating caused by posture changes; and it continuously prompts the user interface to stabilize the device.

[0062] Level 3 Emergency (Protection Execution): When θ > 20° or a_z > 0.5g, the system determines it as an unexpected overturning or a large abnormal movement, and immediately triggers the following three-break, one-lock, and one-alarm operation: The system controls the relay to disconnect the power supply circuit to the heating element, ensuring no continuous heat source output; sends a brake signal to the stepper motor driver to immediately lock the motor shaft, preventing the moxibustion head from continuing to move due to inertia; disables the servo motor control signal to maintain the current mechanical position and avoid the risk of secondary collisions or burns; sends a structured emergency message to the cloud-based health management platform via the MQTT protocol, including the device's unique ID, anomaly trigger timestamp, current posture data snapshot (θ, φ, a_x, a_y, a_z), and temperature and distance change curves for the last 10 minutes (taken from S1 / S2 cache); this message is simultaneously pushed to the bound user's APP and authorized medical personnel's terminals; and activates a high-frequency buzzer (e.g., 4kHz) and a flashing red LED (1s cycle) to alert on-site personnel to take immediate action.

[0063] In one embodiment disclosed in this application, when the device attitude returns to a safe range (θ < 5° and a_z < 0.1g, and remains so for more than 10 seconds), the system automatically enters a safety self-test mode, which verifies the functions of key subsystems in a strict sequence: Short-circuit the temperature measurement pin to a known reference voltage (such as ground or internal reference), read the ADC value and compare it with the factory zero-point calibration value; if the deviation exceeds the allowable tolerance (such as ±0.5℃ corresponding to voltage deviation), mark the sensor as needing recalibration. Emitter an infrared ranging signal and detect the reflected intensity return value; if the return value is lower than the sensitivity threshold or fluctuates too much in multiple measurements, determine that the sensor is contaminated or has a hardware malfunction, temporarily disable the distance closed-loop control and replace it with motor displacement back-propagation.

[0064] First, unlock the motor brake, then perform a small-range reciprocating displacement test at 10% of the rated speed to verify the displacement accuracy (error from the target position within ±1mm). Next, unlock the heating module and operate in low-power mode, monitoring whether the temperature change is stable and within the expected range (fluctuation not exceeding ±0.5℃ / min). If any abnormality is detected in any step, stop the recovery process and remain in safe mode, awaiting manual inspection. After all self-test items pass, clear the emergency status flag, restore the normal temperature-distance linkage control and LSTM anomaly prediction functions established by S1 / S2, and the system enters normal operation.

[0065] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0066] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for IoT management of a health and wellness robot, characterized in that: The management method includes the following steps: S1: Data is collected and transmitted to the controller. The distance-temperature correlation algorithm model is used to determine the optimal temperature range and gradient change law corresponding to different distance intervals. The controller uses the PID algorithm to control the heating module and the displacement module in a coordinated manner. S2: Based on sensing and control, an LSTM time-series prediction algorithm is introduced to construct an LSTM model. The anomaly prediction module extracts features from the time-series data of temperature and distance sensors. The pre-trained LSTM model is used to predict the future anomaly risk level of temperature and distance sensors. Based on the anomaly risk level, the operation of the heating module and displacement module is controlled to implement differentiated linkage compensation. S3: Real-time calculation of the pitch angle, roll angle and acceleration vector of the health care robot, and generation of corresponding safety warnings based on the corresponding safety thresholds. After the health care robot's posture returns to normal, it automatically enters the safety self-check mode. After no abnormalities are found, the initial control process is restored.

2. The IoT management method for a health and wellness robot according to claim 1, characterized in that: Based on the abnormal risk level control, the heating module and displacement module are operated in a differentiated linkage compensation manner, including the following steps: If a moderate anomaly is predicted in the distance sensor, the motor operating status data is retrieved, and the stepper motor step angle and transmission ratio are combined to deduce the actual displacement distance of the moxibustion head and the direction of deviation from the theoretical value. The distance-temperature correlation algorithm model is then used to deduce the temperature compensation amount corresponding to the deviation in reverse, and the heating power is adjusted in advance. If a moderate anomaly is predicted in the temperature sensor, the actual temperature range of the acupuncture point is calculated using a distance-temperature correlation algorithm model based on historical temperature-distance-ambient temperature three-dimensional correlation data, current motor displacement parameters, and ambient temperature compensation values. The heating power is then dynamically adjusted to the corresponding level.

3. The IoT management method for a health and wellness robot according to claim 2, characterized in that: The anomaly prediction module extracts features from the time-series data of temperature and distance sensors, and uses a pre-trained LSTM model to predict the future anomaly risk level of the temperature and distance sensors. This includes the following steps: A continuous time-series data stream is acquired with a fixed sampling period to form a sliding time window. Each time, a sampling point is moved forward to generate training and inference samples. Multidimensional features are constructed for the original sequences within each time window, including distance data features, temperature data features, and cross-modal correlation features; Supervised learning was performed using a labeled historical anomaly dataset, with anomaly labels categorized into three levels: mild anomaly, moderate anomaly, and severe anomaly. The LSTM network structure uses two hidden layers. The input is a sequence of multidimensional feature vectors, and the output is the probability distribution of future abnormal risk levels.

4. The IoT management method for a health and wellness robot according to claim 2, characterized in that: When the LSTM prediction or real-time rule determines that the distance sensor has entered a moderate anomaly, the cross-modal compensation mechanism is activated, using motor operation data to infer the actual distance and complete temperature-linked correction. Read the stepper motor's operating status data, including forward and reverse direction indicators, pulse count, and cumulative running time. Based on the known stepper motor step angle and transmission ratio, calculate the theoretical displacement. Theoretical displacement = number of pulses × step angle / 360° × lead. Compare the theoretical displacement with the value reported by the distance sensor to obtain the direction and magnitude of the deviation. Substitute the deviation distance into the distance-temperature correlation algorithm model and calculate the amount of temperature change to be compensated according to the gradient change law learned by the model. By continuously comparing the actual displacement fed back by the motor encoder with the theoretical displacement calculated by the model, the difference is used as a new PID control error input to the displacement control module to dynamically adjust the position of the stepper motor.

5. The IoT management method for a health and wellness robot according to claim 2, characterized in that: If the temperature sensor shows moderate abnormality, the actual acupuncture point temperature is calculated using historical correlation data and cross-modal information, and dynamic power adjustment and subsequent calibration are implemented. We retrieve your three-dimensional correlation data of temperature, distance, and ambient temperature to form a training / reference library. Each sample record includes the measured temperature, actual distance, ambient temperature, and heating power level. Based on the current motor displacement parameters and ambient temperature compensation rules, the distance-temperature correlation algorithm model is invoked to reversely calculate the range of the actual temperature of the moxibustion point; Based on the calculated actual temperature range, the controller recalculates the target temperature and generates the corresponding PWM duty cycle to drive the heating module to output power, so that the temperature of the moxibustion point reaches the expected comfortable range.

6. The IoT management method for a health and wellness robot according to claim 3, characterized in that: Distance data features include calculating step change rate, mean drift magnitude, and fluctuation variance; temperature data features include calculating variance volatility, short-term gradient change, and the degree of persistent deviation from the ambient temperature difference; cross-modal correlation features include introducing the time delay correlation between distance changes and temperature changes.

7. The IoT management method for a health and wellness robot according to claim 1, characterized in that: Based on the corresponding security threshold, a security alert of the appropriate level is generated, including the following steps: In the first-level warning, if the pitch angle θ > the first pitch angle threshold or the roll angle φ > the roll angle threshold, the user is prompted to adjust the posture of the health care robot. In the secondary intervention, if the pitch angle θ > the second pitch angle threshold or the linear acceleration component a_z in the Z-axis direction > the first component, attitude compensation is initiated. In a Level 3 emergency, if the pitch angle θ > the third pitch angle threshold or the linear acceleration component a_z in the Z-axis direction > the second component, it is determined to be an unexpected overturning / large-scale abnormal movement, and the three-break, one-lock, and one-alarm operation is executed.

8. The IoT management method for a health and wellness robot according to claim 7, characterized in that: Performing the "three disconnections, one lock, one alarm" operation includes the following steps: The power supply circuit to the heating element is disconnected by a relay, a brake signal is sent to the stepper motor driver, the servo control signal is disabled, the current position is maintained, and an emergency alarm message is sent to the cloud health management platform via the MQTT protocol. The emergency alarm message includes the health care robot ID, abnormal timestamp, posture data snapshot, and temperature / distance curve, and is simultaneously pushed to the user APP and medical staff terminal and activates local audible and visual alarms.

9. The IoT management method for a health and wellness robot according to claim 8, characterized in that: Once the health and wellness robot returns to a normal posture, it will automatically enter a safety self-check mode, including the following steps: Verify the zero point of the temperature sensor and the validity of the distance sensor; The motor and heating module were gradually unlocked, and the displacement accuracy and temperature stability were verified through trial operation. Once no abnormalities are detected, the initial control procedure will be resumed.

10. The IoT management method for a health and wellness robot according to claim 1, characterized in that: The controller uses a PID algorithm to control the heating module and the displacement module in a coordinated manner, including: The heating module uses a PWM signal to adjust the on / off frequency of the solid-state relay, thereby dynamically changing the power of the heating element. For the displacement module, the stepper motor / servo motor is driven to adjust the position of the moxibustion head so that the distance and temperature are within the comfortable coupling range defined by the distance-temperature correlation algorithm model.

Images