Joint rehabilitation aid control method and system based on multi-parameter intelligent perception

Abstract

The invention discloses a joint rehabilitation assisting device control method and system based on multi-parameter intelligent sensing. A plurality of biomechanical parameters of the joint part are collected in real time through the multi-parameter sensing module; carrying out preprocessing and feature extraction on the collected various biomechanical parameter data to obtain joint state feature information; based on the joint state feature information, a preset recognition rule or a recognition model is used for recognizing the current joint state; selecting a target working mode corresponding to the recognized joint state from a plurality of preset working modes according to the recognized joint state; the controller module is used for driving a plurality of mutually independent adjusting areas of the dynamic adjusting execution module and exerting an effect on the joint according to a specific pressure distribution or supporting strategy; and continuously receiving feedback data of the multi-parameter sensing module, and dynamically correcting the pressure distribution or support strategy based on the feedback data. And through multi-parameter sensing, joint state recognition and partition dynamic adjustment, self-adaptive supporting and nursing of the joints are achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and in particular to a control method and system for joint rehabilitation assistive devices based on multi-parameter intelligent sensing. Background Technology

[0002] Existing joint rehabilitation assistive devices mostly use elastic fabrics, fixed support strips, or mechanical limiting structures, lacking real-time monitoring and dynamic adjustment, and have limited personalization capabilities; their intervention for joint rehabilitation is imprecise and untimely, lacking control methods that can dynamically adjust according to the joint's condition in real time. Specifically, the following problems exist: First, there is a conflict between static support and dynamic physiology. The physiological state of a joint is dynamic, even fluctuating significantly within a single day. However, existing joint rehabilitation assistive devices provide fixed support, pressure distribution, and rigidity, failing to dynamically adjust according to joint swelling, activity level, or diurnal variations. Long-term use can easily cause local discomfort and even restricted blood flow; resulting in insufficient pressure during periods of swelling and inappropriate restraint during periods of rest. Prolonged use not only leads to poor comfort but may also exacerbate the risk of local ischemia, nerve compression, or muscle atrophy. Second, there is a lack of objective quantification and feedback mechanisms. Rehabilitation assessment and management of joint diseases heavily rely on the subjective experience of clinicians and patient self-reports. Existing joint rehabilitation assistive devices lack data collection capabilities, failing to provide objective, continuous quantitative data on joint status, and cannot proactively warn of abnormal situations such as acute inflammatory attacks. This results in a lack of data support for rehabilitation management and delayed intervention measures. Third, they lack functionality and personalization. Existing joint rehabilitation assistive devices have fixed functions and control methods from manufacturing to disposal; they cannot be self-adjusted according to the user's specific condition, anatomical differences, or life scenarios; and they cannot meet the requirements of individualized precision medicine. Summary of the Invention

[0003] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a joint rehabilitation assistive device control method and system based on multi-parameter intelligent sensing. Through multi-parameter sensing, joint state recognition, and zoned dynamic adjustment, it achieves adaptive support and care for the joint, improving wearing comfort and usage safety.

[0004] On one hand, embodiments of the present invention provide a joint rehabilitation assistive device control method based on multi-parameter intelligent sensing, including: The multi-parameter sensing module collects various biomechanical parameters of the joint in real time, including joint circumference change parameters, pressure distribution parameters of the wearing area, and local skin temperature parameters of the joint. The collected biomechanical parameter data are preprocessed and feature extracted to obtain joint state feature information; Based on the joint state feature information, the current joint state is identified using preset recognition rules or recognition models. The joint state includes inflammatory and swollen state, stable resting state, and active state. Based on the identified joint status, select the corresponding target working mode from a variety of preset working modes; Based on the target working mode, control commands are generated, and the controller module drives multiple independent adjustment zones of the dynamic adjustment execution module to apply force to the joint with a specific pressure distribution or support strategy. The system continuously receives feedback data from the multi-parameter sensing module and dynamically adjusts the pressure distribution or support strategy based on the feedback data to form a closed-loop control of sensing-judgment-execution.

[0005] According to some embodiments of the present invention, the preprocessing and feature extraction of the collected biomechanical parameter data includes: The validity of the various biomechanical parameter data is verified, abnormal data is removed, and the time series data is smoothed and filtered. The various biomechanical parameter data of different dimensions were normalized. Extract the magnitude or rate of change of the joint circumference change parameter, the uniformity characteristics of the pressure distribution parameter of the wearing area, and the absolute value or trend of the local skin temperature parameter of the joint.

[0006] According to some embodiments of the present invention, the step of identifying the current state of the joint based on the joint state feature information using preset recognition rules or recognition models includes: When the change in joint circumference exceeds the first threshold, the local skin temperature of the joint is higher than the preset reference value, or the pressure distribution in the wearing area is abnormally concentrated, the joint is determined to be in an inflammatory and swollen state. When the joint circumference change tends to be stable, the pressure distribution in the wearing area is uniform, and the change in local skin temperature of the joint is less than the second threshold, the joint is determined to be in a stable resting state. When the joint circumference changes periodically and the pressure distribution in the wearing area changes dynamically over time, the joint is determined to be in a state of motion.

[0007] According to some embodiments of the present invention, selecting a target working mode from a preset plurality of working modes based on the identified joint state includes: Predict the type of action a user is about to perform based on historical kinematic data; Based on the identified joint state and the predicted action type, a target working mode is selected.

[0008] According to some embodiments of the present invention, the controller module drives multiple independent adjustment zones of the dynamic adjustment execution module to apply force to the joint with a specific pressure distribution or support strategy, including: When an inflammatory and swollen condition is identified, the pressure distribution strategy is to reduce localized concentrated pressure and apply uniform coverage. When the condition is identified as a stable resting state, apply gentle, continuous circumferential pressure. When identified as a state of motion activity, it provides enhanced support in key stress areas.

[0009] According to some embodiments of the present invention, the dynamic correction of pressure distribution or support strategy based on the feedback data includes: A user-personalized baseline model is pre-built, which is constructed by recording the user's baseline physiological parameters and subjective feedback data under standard actions; Based on the user's personalized baseline model and the feedback data, the control parameters of each working mode are optimized through machine learning algorithms to achieve personalized adaptive adjustment.

[0010] According to some embodiments of the present invention, after the step of dynamically correcting the pressure distribution or support strategy based on the feedback data to form a closed-loop control of perception-judgment-execution, the method further includes: The desensitized joint status data and adjustment logs were encrypted and uploaded to the cloud platform. Receive customized working mode parameters from the cloud platform.

[0011] In another aspect, embodiments of the present invention provide a joint rehabilitation assistive device for implementing the aforementioned control method for joint rehabilitation assistive devices based on multi-parameter intelligent sensing. This includes: A flexible wearable body for fixing to a user's joint area; A multi-parameter sensing module is disposed inside the flexible wearable body and includes a flexible pressure sensor array, a joint circumference change sensing sensor and a skin temperature sensor, used to collect joint state-related parameters. A dynamic adjustment execution module is disposed within the flexible wearable body. The dynamic adjustment execution module includes a multi-zone pneumatic pressurization unit and a shape memory alloy driving unit, which are used to apply zoned adjustable pressure or support to the joints. The controller module is electrically connected to the multi-parameter sensing module and the dynamic adjustment execution module, respectively. It is used to process the collected joint state-related parameters, control the working state of the dynamic adjustment execution module, identify joint state, and switch working modes.

[0012] In another aspect, embodiments of the present invention provide a joint rehabilitation assistive device control system based on multi-parameter intelligent sensing. The system includes: The aforementioned joint rehabilitation assistive devices; The mobile terminal has an application installed to display joint status, set preferences, and receive prompts; The cloud platform is used to store and analyze user data, provide remote monitoring and intervention interfaces, train better state recognition models based on user data, and distribute model update files to joint rehabilitation assistive devices.

[0013] The joint rehabilitation assistive device control method and system based on multi-parameter intelligent sensing according to embodiments of the present invention have at least the following beneficial effects: Compared to existing technologies, this invention represents a paradigm shift from passive tools to proactive intelligent rehabilitation partners, resulting in significant multi-dimensional benefits. First, in terms of application mode, it breaks the limitations of traditional rehabilitation equipment confined to training periods. Through a flexible wearable design and multi-parameter sensing modules, it achieves 24 / 7 continuous monitoring and integrated intervention of the user's joint status, seamlessly integrating precise rehabilitation into daily life and sleep, greatly improving the continuity and convenience of rehabilitation. Second, it achieves a significant breakthrough in intervention precision. Through a zoned dynamic adjustment execution module, it can implement precise targeted physical output to the joint surface, achieving a leap from overall force application to local programmable control, qualitatively improving both intervention precision and safety. Third, in terms of intelligent control, it not only possesses high-response real-time closed-loop control to ensure safety, but also introduces a motion prediction-based proactive protection mechanism and millisecond-level synchronization capability with physiological activities. Simultaneously, through a long-term data-driven adaptive optimization algorithm, the joint rehabilitation assistive device can continuously learn and personalize itself to adapt to the user's unique responses, achieving continuous evolution that becomes increasingly user-friendly with use. Fourth, we will build a multi-party collaborative intelligent rehabilitation ecosystem. Through a cloud platform, we will realize the construction of user digital twins, the remote and precise prescription delivery by physicians, and the evolution of group models based on federated learning, achieving a complete intelligent closed loop of "perception-recognition-decision-regional execution-dual closed-loop optimization-cloud collaboration". We will focus on wearable, continuous, and adaptive physical therapy.

[0014] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a joint rehabilitation assistive device based on multi-parameter intelligent sensing according to an embodiment of the present invention; Figure 3 This is a functional block diagram of a joint rehabilitation assistive device control system based on multi-parameter intelligent sensing according to an embodiment of the present invention; Figure 4 This is a detailed flowchart of the joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to an embodiment of the present invention. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0018] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," "exceeding," etc. are understood to exclude the stated number, and "above," "below," "within," etc. are understood to include the stated number. If "first," "second," etc. are used in the description, they are only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance or implicitly indicating the number of indicated technical features or the order of the indicated technical features.

[0019] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] This embodiment provides a control method for a joint rehabilitation assistive device based on multi-parameter intelligent sensing. The joint rehabilitation assistive device includes a multi-parameter sensing module 120, a dynamic adjustment execution module 130, and a controller module 140. Please refer to [link to relevant documentation]. Figure 1 The joint rehabilitation assistive device control method based on multi-parameter intelligent sensing mainly includes steps S101~S106: S101. The multi-parameter sensing module 120 collects various biomechanical parameters of the joint in real time, including joint circumference change parameters, pressure distribution parameters of the wearing area, and local skin temperature parameters of the joint.

[0021] S102. Preprocess and extract features from the collected biomechanical parameter data to obtain joint state feature information.

[0022] S103. Based on joint state feature information, identify the current joint state using preset recognition rules or recognition models. The joint state includes inflammatory and swollen state, stable resting state, and movement activity state.

[0023] S104. Based on the identified joint state, select the corresponding target working mode from a variety of preset working modes.

[0024] S105. Based on the target working mode, control commands are generated, and the controller module 140 drives multiple independent adjustment zones of the dynamic adjustment execution module 130 to apply action to the joint with a specific pressure distribution or support strategy.

[0025] S106. Continuously receive feedback data from the multi-parameter sensing module 120, and dynamically correct the pressure distribution or support strategy based on the feedback data to form a closed-loop control of sensing-judgment-execution.

[0026] In step S102 above, the collected biomechanical parameter data are preprocessed and feature extracted, including: The validity of various biomechanical parameter data was verified, outlier data was removed, and the time series data was smoothed and filtered. Normalize the data of various biomechanical parameters with different dimensions; Extract the magnitude or rate of change of joint circumference parameters, the uniformity characteristics of pressure distribution parameters in the wearing area, and the absolute value or trend of local skin temperature parameters in the joint.

[0027] In step S103 above, the current state of the joint is identified based on joint state feature information using preset recognition rules or recognition models, including: When the change in joint circumference exceeds the first threshold, the local skin temperature of the joint is higher than the preset reference value, or the pressure distribution in the wearing area is abnormally concentrated, the joint is determined to be in an inflammatory and swollen state. When the joint circumference change tends to be stable, the pressure distribution in the wearing area is uniform, and the change in local skin temperature of the joint is less than the second threshold, the joint is determined to be in a stable resting state. When the joint circumference changes periodically and the pressure distribution in the wearing area changes dynamically over time, the joint is determined to be in a state of motion.

[0028] Based on historical kinematic data, the system predicts the type of action the user will perform and selects a working mode based on the identified joint status and the predicted action type.

[0029] In some embodiments of the present invention, the controller module 140 drives multiple independent adjustment zones of the dynamic adjustment execution module 130 to apply action on the joint with a specific pressure distribution or support strategy, including: When an inflammatory and swollen condition is identified, the pressure distribution strategy is to reduce localized concentrated pressure and apply uniform coverage. When the condition is identified as a stable resting state, apply gentle, continuous circumferential pressure. When identified as a state of motion activity, it provides enhanced support in key stress areas.

[0030] In some embodiments of the present invention, dynamic correction of pressure distribution or support strategy based on feedback data includes: A personalized baseline model for users is pre-built. The personalized baseline model is constructed by recording the user's baseline physiological parameters and subjective feedback data under standard actions. Based on user-specific baseline models and feedback data, machine learning algorithms are used to optimize the control parameters of each working mode, achieving personalized adaptive adjustment.

[0031] In some embodiments of the present invention, after the step of dynamically correcting the pressure distribution or support strategy based on feedback data to form a closed-loop control of perception-judgment-execution, the method further includes: The desensitized joint status data and adjustment logs were encrypted and uploaded to the cloud platform 300. Receive customized working mode parameters issued by the cloud platform 300.

[0032] In some embodiments of the present invention, closed-loop control includes short-term closed-loop feedback and long-term adaptive optimization. Short-term closed-loop feedback includes real-time data acquisition, construction of multiple parallel feedback control loops, and effect evaluation and strategy fine-tuning. The long-term adaptive optimization includes data accumulation and personalized baseline establishment, optimization based on reinforcement learning, optimization algorithm and model update, and cloud-based collaborative learning evolution.

[0033] Please see Figure 2 This embodiment also provides a joint rehabilitation assistive device 100 for implementing the above-mentioned joint rehabilitation assistive device control method based on multi-parameter intelligent sensing, including: A flexible wearable body 110 is used to fix it to the user's joint area; The multi-parameter sensing module 120 is disposed inside the flexible wearable body and includes a flexible pressure sensor array 121, a joint circumference change sensing sensor 122, and a skin temperature sensor 123, which are used to collect joint state-related parameters.

[0034] The dynamic adjustment execution module 130 is disposed in the flexible wearable body. The dynamic adjustment execution module 130 includes a multi-zone pneumatic pressurization unit 131 and a shape memory alloy drive unit 132, and has multiple independent adjustment zones for applying zoned adjustable pressure or support to the joint. The controller module 140 is electrically connected to the multi-parameter sensing module 120 and the dynamic adjustment execution module 130, respectively. It is used to process the collected joint state-related parameters and control the working state of the dynamic adjustment execution module 130, identify joint state recognition and switch working modes.

[0035] It should be noted that the multi-parameter sensing module 120 may also include an inertial measurement unit, a microenvironment humidity sensor, and a bioimpedance sensor. The microenvironment humidity sensor and the bioimpedance sensor are used to monitor the humidity of the microenvironment between the protective gear and the skin. Combined with temperature parameters, they can indirectly assess the level of inflammatory activity or wearing comfort, providing an additional dimension for adjustment strategies. The bioimpedance sensor measures the impedance characteristics of local tissues through a weak current. Edema causes changes in tissue impedance, allowing for a more direct and quantitative assessment of the degree of joint swelling, reflecting the accumulation of tissue fluid better than a single circumference measurement. The dynamic adjustment execution module 130 may also include an electroactive polymer drive unit and a magnetorheological fluid support strip, enabling zonal and multidimensional physical outputs (such as pressure, stiffness, and deformation) on the joint surface. For example, during gait, only the weight-bearing area is pressurized and supported, while other areas remain relaxed, providing the ability for zonal targeted intervention; millisecond-level real-time intervention synchronized with the gait cycle (such as applying pressure at the moment of heel strike) enables precise control triggered by events. The controller module 140 is equipped with a microprocessor.

[0036] Please see Figure 3 This embodiment also provides a joint rehabilitation assistive device control system based on multi-parameter intelligent sensing, including: The aforementioned joint rehabilitation assistive devices 100; Mobile terminal 200, with an APP application installed for displaying joint status, setting preferences and receiving prompts; The cloud platform 300 is used to store and analyze user data, provide remote monitoring and intervention interfaces, train a better state recognition model based on user data, and distribute model update files to joint rehabilitation assistive devices.

[0037] Please see Figure 4 The joint rehabilitation assistive device control method and system based on multi-parameter intelligent sensing provided in this invention includes, but is not limited to, the following steps: S201. Collect and preprocess various biomechanical parameter data of the joint. The microprocessor in controller module 140 polls and acquires raw data from all sensors at a fixed frequency (e.g., 10-100Hz). The acquired raw data is preprocessed to ensure quality.

[0038] (1) Validity verification: Determine whether the data is within a physiologically reasonable range, such as skin temperature 20-42°C, and remove obviously abnormal outliers.

[0039] (2) Filtering and noise reduction: Digital filters are used to smooth the signal and suppress high-frequency noise introduced by motion artifacts, electromagnetic interference, etc.

[0040] (3) Normalization: Data from different sensors with different dimensions (such as ohms, kilopascals, degrees Celsius, degrees / second) are mapped to a unified numerical range (such as 0-1) through an algorithm to eliminate the influence of dimensions and prepare for subsequent fusion analysis.

[0041] S202, Data Feature Extraction and Intelligent Joint Status Recognition (1) Feature Extraction: Representative feature values ​​are extracted from the preprocessed time series data. For example, short-term (e.g., 5-minute) slopes of change are extracted from joint circumference data; a rapid rise indicates acute swelling. Entropy or standard deviation is calculated from the pressure distribution matrix to assess the uniformity of pressure distribution. The moving average and difference from the individual baseline are calculated from skin temperature data. Gait cycles are identified from inertial measurement unit data through peak detection, and gait frequency and joint range of motion are calculated.

[0042] (2) State Recognition: Based on the extracted feature vectors, a multi-source information fusion algorithm is used for state classification. A rule-based expert system or a lightweight machine learning model (such as a classifier trained in the cloud and deployed locally) is employed. The identified joint states include: Inflammatory swelling phase: The criteria for determination are: circumferential change rate > threshold A, skin temperature > baseline + threshold B, and / or bioimpedance < threshold C. This is a high-priority care status.

[0043] Stable resting period: The criteria for this period are that the inertial measurement unit shows no significant movement and the rate of change of all physiological parameters is below the threshold. This is the typical state at night or when sitting still.

[0044] Motion activity phase: The determination condition is that the inertial measurement unit detects regular periodic motion signals. This state can be further subdivided into sub-modes such as low-speed walking, fast walking, and going up and down stairs.

[0045] S203. Match the working mode and determine the control strategy. Based on the identified joint status, the optimal operating mode is selected from a pre-defined mode library. Operating modes are a predefined set of control strategies for each region of the execution module.

[0046] (1) Resting care mode (corresponding to stable resting period): The control goal is to promote circulation and comfort. Strategy: Apply a low (e.g., 10-15 mmHg) and uniform static pressure to all airbags to simulate a gentle massage to promote lymphatic drainage.

[0047] (2) Dynamic Support Mode (corresponding to the active phase of movement): The control objective is primarily to provide dynamic protection. Strategy: During the gait support phase, momentarily increase the airbag pressure in the corresponding areas of the medial and lateral ligaments (e.g., increase to 20-30 mmHg); during the swing phase, fully release the pressure to reduce movement resistance. For going up and down stairs, the support at the quadriceps attachment point can be adjusted accordingly.

[0048] (3) Swelling Relief Mode (corresponding to the inflammatory swelling stage): The control goal is to reduce swelling and avoid pressure. Strategy: Automatically reduce the local pressure in the area with the most obvious swelling, while applying slightly higher pressure to the proximal area of ​​the limb to form a pressure gradient from the distal to the proximal end, which more effectively promotes tissue fluid return.

[0049] S204, Partition Execution and Real-time Output The microprocessor translates the pattern strategy into specific actuator control commands, driving the various partitions of the dynamic adjustment execution module 130 to work together and output precise mechanical action.

[0050] S205, Closed-Loop Feedback and Adaptive Optimization This is an ongoing process that involves continuously monitoring the feedback signal from the multi-parameter sensing module 120 while outputting control signals.

[0051] (1) Short-term closed loop: Determine whether the current output has achieved the expected effect. For example, in dynamic support mode, if the pressure sensor feedback shows that the support is still insufficient, the system will fine-tune and increase the target pressure value in that area.

[0052] (2) Long-term adaptation: Long-term recording of user data, such as joint state sequences, execution parameters, and subjective comfort / pain scores provided by users through the APP. Through machine learning algorithms (such as reinforcement learning), control parameters in each mode are gradually optimized to build a unique personalized model for the user, achieving adaptive optimization that becomes more and more user-friendly over time.

[0053] S206, Remote Interaction and Cloud Platform Collaboration Data is encrypted and uploaded to the cloud platform 300. Doctors can view patients' recovery trends, receive abnormal alarms, and remotely customize or fine-tune the working mode parameters of protective gear for specific patients, thus realizing remote medical intervention.

[0054] It should be noted that through feature extraction and intelligent joint state recognition in step S202 above, information is extracted from the raw data, and the current physiological and functional state of the joint is ultimately understood. The massive, heterogeneous low-level sensor data collected by the multi-parameter sensing module 120 is transformed into high-level state labels with clear clinical or physiological significance. This process can be divided into two closely linked stages: feature extraction and state recognition.

[0055] Phase 1: Multi-source data preprocessing and feature extraction Before state recognition, robust features capable of effectively distinguishing different joint states must first be extracted from the raw, noisy sensor data stream. Different feature extraction strategies need to be employed for different types of parameters.

[0056] 1. Data Preprocessing Basics All sensor data undergoes rigorous preprocessing before feature extraction to provide a "clean" data foundation for subsequent analysis. This includes: Data validity verification and outlier removal: Set a reasonable physiological range (such as joint skin temperature, which is usually 28°C-40°C), and automatically identify and remove obvious abnormal data points caused by sensor momentary failure or strong electromagnetic interference.

[0057] Signal filtering and smoothing: Digital signal processing techniques, such as moving average filtering or low-pass filtering (cutoff frequency is usually set to 1-5Hz), are used to suppress high-frequency noise (such as small fluctuations caused by muscle tremors and walking vibrations) and preserve the true physiological signal trend.

[0058] Data alignment and synchronization: Since there may be slight differences in the sampling rate and response time of different sensors, the system adopts a timestamp alignment algorithm to ensure that the multi-parameter data collected at the same time can be correctly matched, laying the foundation for multi-source fusion.

[0059] 2. Feature extraction methods for different parameters (1) Feature extraction of joint circumference / deformation data Absolute value and relative change: Calculate the absolute change of the current circumference value relative to the user's personalized baseline (such as the resting circumference upon waking in the morning). This is a direct indicator of the degree of swelling.

[0060] Rate of change (first derivative): Calculates the slope of change in circumference over a specific time window (e.g., the past 30 minutes). During the acute inflammatory phase, the rate of change accelerates significantly (rapid swelling); during the stable phase, the rate approaches zero; during the recovery phase, it may be negative (swelling subsides). This is a key characteristic distinguishing between the acute and chronic phases.

[0061] Fluctuation amplitude and stability: Calculate the variance or standard deviation of the circumference data under resting conditions. A small variance indicates that the joint is in a stable state; a large variance may mean that the measurement is unstable or that there is joint micro-movement.

[0062] (2) Feature extraction of pressure distribution data Mean pressure and peak pressure: Calculate the average and maximum pressure values ​​for the entire pressure sensor array or a specific region of interest (such as both sides of the patella). Excessively high peak pressure is a risk indicator for discomfort and pressure sores.

[0063] Pressure distribution uniformity index: Entropy calculation: Treat the pressure distribution map as a probability distribution and calculate its information entropy. The higher the entropy value, the more uniform the pressure distribution; the lower the entropy value, the more concentrated the pressure is at a few points, indicating an uneven distribution.

[0064] Central pressure shift: Calculates the coordinates of the center of gravity in relation to the pressure distribution. During knee joint movement, the center of gravity trajectory reflects gait symmetry. Abnormal varus / valgus deviation can cause the center of gravity to consistently shift to one side.

[0065] (3) Feature extraction of skin temperature data Absolute temperature value: Directly read the absolute value of the skin temperature sensor 123.

[0066] Temperature difference ΔT from a personalized baseline: Calculates the difference between the current temperature and the user's baseline temperature in a healthy resting state. Temperature difference ΔT is a more sensitive indicator of inflammation than absolute temperature because it eliminates the influence of individual differences in basal metabolic rate.

[0067] Temperature change trend: Perform linear fitting on temperature data over a period of time to observe whether it is in an upward, downward or stable trend.

[0068] (4) Feature extraction of inertial measurement unit data: Joint angles: By fusing accelerometer and gyroscope data, complementary filtering or Kalman filtering algorithms are used to estimate the flexion / extension angles of the joint in real time and calculate the range of motion of the joint.

[0069] Movement frequency and rhythm: Step frequency (steps / minute) or movement rhythm is calculated by detecting periodic peaks in acceleration data sequences. The stability and symmetry of step frequency are important indicators for assessing motor function.

[0070] Motion recognition features: Temporal features (such as mean, variance, and root mean square) and frequency domain features (main frequency and frequency component energy obtained through Fast Fourier Transform (FFT)) are extracted from inertial measurement unit (IMU) data. These feature vectors can be used in machine learning models to distinguish different activity modes such as walking, running, and climbing stairs, which is the core of motion state recognition.

[0071] (5) Feature extraction of bioimpedance data Impedance amplitude: Reflects the total fluid content of a tissue. A decreased impedance value usually indicates tissue edema and increased fluid.

[0072] Impedance phase angle: provides information about the integrity of tissue and cell structure, and can change under certain pathological conditions.

[0073] Phase 2: Intelligent Joint State Recognition Based on Multi-Feature Fusion After extracting a series of feature values ​​from various types of data, these features are combined into a multi-dimensional feature vector and input into the state recognition engine. This embodiment of the invention employs a hybrid strategy combining a rule engine and a lightweight machine learning model, balancing reliability and adaptability.

[0074] 1. Rule-based state recognition For states with clear characteristics and logical consistency, pre-defined expert rules are used for rapid judgment, resulting in low computational resource consumption and fast response time. It offers strong interpretability and is used to define specific scenarios. Example rules are as follows: Rule 1: Identify the inflammatory and swollen phase IF ((circumference change rate > threshold R1) AND (skin temperature difference ΔT > threshold T1)) OR (bioimpedance value < threshold Z1) THEN Status = Inflammatory and swollen phase, Confidence level = High.

[0075] Logical explanation: This rule captures the typical characteristics of acute inflammation—rapid swelling accompanied by an increase in local skin temperature. Bioimpedance, as an auxiliary verification, improves the accuracy of the judgment.

[0076] Rule 2: Identify stable resting periods IF (Inertial Measurement Unit Motion Energy < Threshold E1) AND (Absolute Value of Circumference Change Rate < Threshold R2) AND (Pressure Distribution Entropy > Threshold_Entropy1) THEN Status = Stable resting period, Confidence level = High.

[0077] Logical explanation: No significant movement, stable circumference, and uniform pressure distribution are typical characteristics of a resting state.

[0078] Rule 3: Identify the "Active Period" IF (Inertial Measurement Unit detects significant periodic signal) AND (Joint Range of Motion ROM > Threshold ROM1) THEN State = Active phase, Confidence level = High.

[0079] Logical explanation: The core criterion is the existence of regular joint movements that reach a certain range.

[0080] 2. State recognition based on machine learning models For complex, ambiguous, or emerging states that are difficult to cover by rules, the system employs a lightweight classifier pre-trained in the cloud or on large datasets. Commonly used models include support vector machines, decision tree ensembles (such as random forests), or small neural networks. They are highly adaptable and suitable for complex scenarios.

[0081] Model input: This is the feature vector composed of all the feature values ​​extracted in the previous steps, such as: [ΔC, dC / dt, average pressure, pressure entropy, ΔT, motion frequency, ROM, impedance value, ...].

[0082] Model output: In addition to state labels (such as "inflammatory swelling phase"), it usually includes a probability or confidence level (e.g., 85% probability of inflammatory swelling phase, 15% probability of stable resting phase). This probabilistic output makes decision-making more flexible.

[0083] Model Training and Updates: The initial model is trained using a large amount of labeled data, derived from clinical observations, user feedback, or expert judgment. The system supports online learning or periodic model updates, enabling it to adapt to new patterns or optimize recognition performance.

[0084] 3. Multi-source information fusion and final decision-making The outputs of the rules engine and the machine learning model are fused at the decision level, and the arbitration module makes the final state determination.

[0085] High confidence level: When the rule engine gives a high confidence level judgment and it is consistent with the output of the machine learning model, the result is adopted directly.

[0086] In cases of conflict or low confidence: When the rules conflict with the model results, or when both have low confidence levels, the arbitration module may adopt a conservative strategy, such as prioritizing the inflammatory and swollen phase that requires attention; or it may initiate a more complex discrimination process. Alternatively, it can send a simple question to the user via the app, such as "Are you currently experiencing joint pain or discomfort?", using user feedback as a crucial basis for the final judgment. This demonstrates the intelligence of human-machine collaboration.

[0087] The state recognition process receives feature information from S2, and the output joint state directly drives the pattern matching decision in S4. Through the above-mentioned refined feature engineering and advanced fusion recognition algorithm, the system ensures that it can accurately understand the needs of the joints, just like an experienced rehabilitation therapist.

[0088] It should be noted that while the joint state intelligent recognition in step S202 above, based on recognition rules, is intuitive, its performance heavily relies on prior expert knowledge and struggles to handle complex, ambiguous, or unknown states. In contrast, methods based on machine learning models can learn complex patterns and relationships from the data itself, exhibiting stronger generalization and adaptability. The following details the entire process of building, training, deploying, and optimizing this machine learning model.

[0089] 1. Selection of Machine Learning Models The relationship between joint status and multi-parameter features is often highly nonlinear. For example, slight swelling accompanied by specific temperature change patterns and pressure distributions may indicate the onset of inflammation. Such subtle patterns are difficult to exhaustively describe with simple judgment rules, but machine learning models can automatically learn and handle complex nonlinear relationships. The model can automatically learn the importance of different features for state classification (i.e., feature weights) without manual setting, automatically weighting features. For example, it was found that the spectral characteristics of joint angular velocity are more important than average pressure in distinguishing between going up and down stairs and walking on flat ground. Through fine-tuning with individual user data, the model can adapt to the unique physiological characteristics and behavioral patterns of different users, achieving truly personalized recognition and adapting to individual differences. Given the computational power, memory, and power consumption limitations of the Joint Rehabilitation Assistive Device 100, a lightweight and highly efficient machine learning model was selected, specifically including: (1) Support Vector Machine: It is particularly suitable for classification problems with small sample size and high-dimensional data. It finds the optimal classification hyperplane in the feature space through the radial basis function (RBF) kernel, has strong generalization ability, and is suitable as a model for initial deployment.

[0090] (2) Random Forest Ensemble Learning Algorithm: This algorithm classifies data by constructing multiple decision trees and combining their voting results. It can effectively prevent overfitting, is insensitive to missing features, and can provide a ranking of feature importance, offering good interpretability.

[0091] (3) Lightweight Neural Networks: If computational resources permit, simple neural networks such as multilayer perceptrons (MLP) or one-dimensional convolutional neural networks (1D-CNN) can be used. 1D-CNN is good at automatically extracting local features from time-series data such as inertial measurement units.

[0092] In practical applications, the random forest ensemble learning algorithm is used as the main model because it achieves a good balance between accuracy, speed and interpretability.

[0093] 2. Model building process (1) Feature engineering and dataset construction All features extracted in the previous stage (joint circumference change rate, pressure entropy, temperature trend, motion frequency, etc.) are combined into a feature vector. Each feature vector corresponds to a snapshot of the joint state at a specific time point or within a time window.

[0094] Data labeling: Obtaining a large amount of labeled data is crucial for supervised learning. Label sources include user feedback, clinical records, and heuristic rules. User feedback is provided through an app prompting users to manually label their current state, such as "I feel swollen now" or "I am walking." Clinical records are used in clinical trials, where rehabilitation physicians add labels to data segments based on clinical assessments (such as the degree of swelling and inflammatory markers). Heuristic rules initially use a high-performance rule engine to automatically label a portion of high-quality data as a supplement to the training set.

[0095] Dataset partitioning: The labeled dataset is divided into a training set (for model learning), a validation set (for tuning hyperparameters and selecting the best model), and a test set (for final evaluation of model performance).

[0096] (2) Model training Model training was completed on the cloud platform 300.

[0097] Feature standardization: Standardize the features of the training set (such as Z-score standardization) to bring all features to the same order of magnitude, thereby accelerating model convergence and improving performance.

[0098] Model learning involves inputting standardized feature vectors and their corresponding labels into a selected algorithm (such as a random forest). The algorithm iteratively optimizes and automatically adjusts its internal parameters, aiming to minimize the difference between the predicted results and the true labels. In the case of a random forest, this involves growing multiple decision trees that can make accurate predictions.

[0099] Hyperparameter tuning: Adjusting the model's hyperparameters (such as the number of trees in a random forest, tree depth, etc.) based on validation set performance to find the optimal configuration. This process is automated using techniques such as grid search or random search.

[0100] (3) Model evaluation and compression Performance evaluation: Evaluate the performance of the final model on the test set using metrics such as accuracy, precision, and F1 score.

[0101] Model Compression and Conversion: The trained model is compressed and optimized for operation on the resource-constrained joint rehabilitation assistive device 100. Quantization: Model parameters are converted from 32-bit floating-point numbers to 8-bit integers, significantly reducing model size and memory usage with minimal impact on accuracy. Pruning: Neurons or weights that contribute little to the model output are removed, simplifying the model structure. Finally, the model is converted to a format suitable for embedded inference, such as TensorFlow Lite.

[0102] 3. Model Deployment and Real-time Inference on Embedded Devices The compressed model file is downloaded to the memory of the controller module 140 of the joint rehabilitation assistive device 100. The real-time inference process includes: Feature input: After preprocessing and feature extraction, the data of each new time window forms a feature vector for the current moment.

[0103] Forward propagation: The microprocessor calls the built-in lightweight inference engine, such as TensorFlow Lite Micro, to input the feature vectors into the loaded machine learning model.

[0104] State prediction: The model performs high-speed calculations (forward propagation) and outputs a probability distribution vector. For example: [Inflammatory swelling phase: 0.75, Stable resting phase: 0.10, Active phase: 0.15], which means that the model is 75% confident that it is currently in the inflammatory swelling phase.

[0105] Decision output: The state with the highest probability is taken as the final identification result, such as the inflammatory and swollen stage. At the same time, this probability value (e.g., 0.75) can be used as the confidence level for subsequent decision-making modules.

[0106] 4. Online learning and model evolution By introducing online learning or federated learning mechanisms, the system can be made capable of continuous evolution.

[0107] Online learning: When users correct the joint state recognition results through the app, this correction, along with the original feature vector, forms a valuable new labeled sample. These samples are periodically encrypted and uploaded to the cloud platform 300. After aggregating anonymous feedback from all users, the cloud platform 300 incrementally updates the model and then pushes the new model version to all users.

[0108] Federated learning: a privacy-preserving technology. Model training and updates are performed locally on the joint rehabilitation assistive device 100, with only the updated model parameters uploaded to the cloud platform 300 for aggregation. This protects user privacy while enabling the collective evolution of the model.

[0109] 5. Collaboration with the rules engine Build a hybrid system where a rules engine and a machine learning model work together, rather than a complete replacement.

[0110] The rule engine handles well-defined scenarios: For scenarios with clear logic and high certainty, such as when the inertial measurement unit detects complete stillness, the rule engine makes a quick determination with low power consumption and fast response.

[0111] Machine learning models handle ambiguous scenarios: When features are atypical, rules conflict, or new situations arise, machine learning models are used for comprehensive judgment, leveraging their powerful pattern recognition capabilities.

[0112] Confidence Arbitration: When the rule and model outputs are inconsistent, compare their confidence levels. Adopt the one with the higher confidence level. If the confidence levels are similar, a conservative strategy or a manual feedback mechanism can be used.

[0113] The joint status recognition based on machine learning models represents a complete technological system, encompassing data preparation, model selection, training optimization, embedded deployment, and continuous evolution. This design transforms joint rehabilitation assistive devices from machines executing fixed programs into intelligent rehabilitation partners capable of continuous learning, adaptation, and increasing precision.

[0114] It should be noted that the working mode matching and decision-making in step S203 above involves a complex, multi-level decision-making process. Its goal is to select the optimal intervention strategy for the current specific situation from a pre-set mode library and generate executable instructions.

[0115] 1. Building the pattern library Each working mode targets one or more specific joint states, clearly defining the control objectives, parameters, and timing for each independent region within the dynamic adjustment execution module 130. The mode library is constructed based on a digital crystallization of biomechanics, rehabilitation medicine, and clinical experience; it is a strategic digital encapsulation. The core working modes are detailed below: (1) Mode A: Resting care mode Target state: Stable resting period (such as nighttime sleep, prolonged sitting).

[0116] Control objectives: Promote venous blood and lymphatic return, relieve mild edema, maximize comfort, and save energy.

[0117] Specific strategies: Pressure strategy: Apply a low and constant pressure (e.g., 10-15 mmHg) to all pressurized areas. This pressure is below the mean capillary pressure and is designed to provide a gentle, circumferential embrace, promoting tissue fluid return without obstructing arterial blood supply.

[0118] Timing strategy: Apply pressure continuously, or use a slow cycle strategy, such as applying pressure for 10 minutes and resting for 2 minutes, to simulate a gentle massage effect.

[0119] Triggering conditions: Step S202 is identified as a stable resting state, and the inertial measurement unit shows no motion for more than a certain time threshold.

[0120] (2) Mode B: Dynamic support mode Target state: the period of physical activity and its sub-states (such as walking, going up and down stairs).

[0121] Control objective: To provide timely, appropriate, and adequate dynamic stability protection during joint movement, reduce abnormal forces, and prevent injury.

[0122] Specific strategies: Event Triggering and Phase Control: The core of this mode is synchronization with the movement rhythm. The system triggers actions at key moments based on the gait phase calculated in real time by the inertial measurement unit (e.g., heel strike, full foot flattening, toe lift). For example, at the beginning of the gait cycle's support phase (when the heel just touches the ground and weight-bearing begins), the airbag pressure in the corresponding area of ​​the medial and lateral collateral ligaments is rapidly increased within milliseconds (e.g., 25-30 mmHg) to compensate for insufficient muscle stability; during the swing phase (leg lift-off and swing), the pressure is rapidly reduced to a minimum (e.g., 5 mmHg) to reduce motion resistance.

[0123] Sub-mode differentiation: Different strategies are adopted for different activities. When going up and down stairs, the focus is on providing support at the quadriceps attachment point, and the pressure curve is fine-tuned according to the mechanical differences between going up stairs (mainly eccentric contraction) and going down stairs (mainly concentric contraction).

[0124] (3) Mode C: Swelling Relief Mode Target state: Inflammatory and swollen phase.

[0125] Control objectives: Avoid compressing the area with the most severe swelling, while promoting centripetal return of tissue fluid through gradient pressure to eliminate swelling.

[0126] Specific strategies: Pressure gradient strategy: The dynamic adjustment execution module 130 generates a pressure gradient that increases sequentially from the distal to the proximal end of the limb (e.g., 15 mmHg at the ankle -> 20 mmHg at the calf -> 18 mmHg below the knee). This physical gradient powers the return of lymph and interstitial fluid, simulating professional lymphatic drainage techniques.

[0127] Intelligent avoidance: Based on the pressure distribution cloud map, it automatically identifies the hot spots with the most obvious swelling and actively reduces pressure in the area to avoid secondary damage.

[0128] Waveform pressurization: Wave-like sequential pressurization can be used to further promote liquid flow.

[0129] (4) Abnormal prompt mode Target state: When abnormal parameters are identified, but it is insufficient or uncertain whether they should be classified into the above modes (e.g., a sudden increase in temperature but no obvious swelling).

[0130] Control objectives: Handle with caution, prioritize alarms, and rely on user or remote doctor judgment.

[0131] Specific strategies: Control action: May switch to a conservative pressure setting, such as uniform low pressure mode C.

[0132] User interaction: Immediately alert the user via vibration and display a warning message on the app, such as "An abnormal increase in joint skin temperature has been detected. Do you feel uncomfortable? Please confirm." At the same time, the abnormal data is marked, encrypted, and uploaded to the cloud platform 300 for doctors to review.

[0133] 2. Matching and Decision Logic The intelligent mapping from state to action involves matching decisions that are not simply a one-to-one lookup table, but a comprehensive judgment process that includes priority, confidence, and context.

[0134] Basic mapping: Maintain a state-mode mapping table. For example: stable resting state -> resting care mode; active state -> dynamic support mode; inflamed and swollen state -> swelling reduction mode.

[0135] Handling Complex States and Priorities: Joints may be in complex states simultaneously (e.g., starting to walk with slight swelling). In this case, the decision engine follows preset priority rules. Typically, the inflammatory and swollen state has the highest priority because it relates to acute symptom management; followed by the active state; and finally the stable resting state. Therefore, the system may prioritize the swelling reduction mode, but will incorporate the phase-aware characteristics of the dynamic support mode, forming a hybrid strategy.

[0136] Confidence-based decision making: The machine learning model outputs a confidence score, and the decision engine adopts different strategies based on this score. High confidence level (e.g., >85%): Directly adopt the mapping pattern.

[0137] Medium confidence level (e.g., 60% - 85%): This can be implemented on a trial basis, that is, the mode is applied but feedback signals are monitored more frequently and prepared for quick adjustments.

[0138] Low confidence (e.g., <60%): May trigger an anomaly alert mode, or revert to a rule-based conservative judgment and actively seek user feedback.

[0139] 3. Personalized strategy library and adaptive optimization To truly achieve personalization, the mode library is not fixed; users can fine-tune the intensity of modes, such as low, medium, and high pressure levels, through the app. The system learns and remembers user preferences, recording the improvement in joint parameters (such as swelling and temperature) after degrading a certain mode strategy under specific conditions. For example, data analysis revealed that for user A, using slow-circuit compression at night was more effective in reducing morning swelling than constant compression; by optimizing the parameters of this user's resting care mode, a unique optimal strategy was gradually formed. This reflects the system's evolution from executing instructions to optimizing strategies, achieving therapeutic-oriented adaptive optimization.

[0140] The system matches the corresponding working mode based on the recognition results, demonstrating intelligent decision-making logic. It receives joint status recognition results and drives the dynamic adjustment execution module 130. The entire decision-making process fully considers individual differences and real-time context, ensuring precise and comfortable intervention. Medical knowledge, biomechanical principles, and personalized user needs are encoded into a sophisticated, dynamic, and evolvable decision-making system. This ensures that the intervention behavior of joint rehabilitation assistive devices is not only automated but also intelligent, adaptable, and safe.

[0141] It should be noted that the task of step S204 above is to translate the high-level instructions of intelligent decision-making into a series of precise and coordinated low-level control commands, drive the various independent areas of the dynamic adjustment execution module 130, and complete the final physical output.

[0142] 1. Generation and translation of control commands Instruction parsing: The microprocessor receives instructions from step S203, which include the target operating mode and its parameters.

[0143] Command generation: The microprocessor, based on its built-in driver library, parses the instruction into a specific sequence of hardware operation instructions. It maps logical region numbers to specific hardware driver channels and converts the target pressure value into control variables for the actuator.

[0144] Timing synchronization: If the action needs to be synchronized with the motion event, the microprocessor will continuously read the inertial measurement unit data and issue the execution command at a predefined trigger point (such as when the acceleration feature of the heel touching the ground is detected).

[0145] 2. Coordinated control strategy for multiple execution units Partition control allows for the independent or coordinated operation of multiple execution units simultaneously.

[0146] (1) Precise control of the pneumatic pressurization unit Taking a three-zone pneumatic pressurization unit as an example, the control logic is as follows: Independent pressure closed-loop control: Each air chamber is an independent pressure closed-loop control system.

[0147] Drive components: Each air chamber is controlled by a miniature air pump (responsible for inflation and pressurization) and a normally closed electromagnetic exhaust valve (responsible for exhaust and depressurization).

[0148] Feedback element: Each chamber integrates a miniature pressure sensor for real-time monitoring of the current pressure value.

[0149] Control Cycle: When the microprocessor receives a command to pressurize a certain area, it starts the corresponding air pump and closes the exhaust valve. The pressure sensor provides real-time feedback of the current pressure value to the microprocessor. The microprocessor compares the current pressure value with the target pressure. When the current pressure value is greater than or equal to the target pressure, the air pump is immediately shut off. To prevent overshoot, a PID control algorithm can be used to reduce the air pump power in advance when approaching the target pressure, or pulse inflation can be used. When pressure needs to be maintained, the microprocessor monitors the natural pressure decay and performs minor pressure replenishment. When pressure needs to be reduced, the microprocessor opens the exhaust valve and can control the valve's opening and closing time as needed, or use PWM control to achieve different depressurization rates such as fast, medium, and slow.

[0150] (2) Other control characteristics Shape memory alloys: The control quantity is typically current or voltage. By controlling the energizing time and current magnitude, the amount of shrinkage and the force generated by the shape memory alloy can be precisely controlled. Special attention needs to be paid to duty cycle control to prevent overheating, and integrated temperature sensors may be necessary for protection.

[0151] Electroactive polymers require high-voltage, low-current driving circuits. By controlling the applied electric field strength, their deformation can be controlled, enabling high-precision micro-motion.

[0152] Magnetorheological fluid support strip: By changing the current in the electromagnet, the magnetic field strength is altered, thereby continuously and rapidly adjusting the viscosity of the magnetorheological fluid and the macroscopic rigidity of the support strip. The response time can reach the millisecond level.

[0153] 3. Real-time output and system response Hard real-time requirement: The latency of the entire process, from gait phase recognition to the actuator generating target support force, must be controlled within tens of milliseconds; otherwise, the support will lag and lose its protective significance. This requires the microprocessor to have high interrupt response priority and efficient code. Real-time performance is a core requirement, especially in dynamic support mode.

[0154] Coordinated output: The system not only controls individual regions but also needs to coordinate the actions of multiple regions. For example, in gait, regions 1 and 2 need to be rapidly pressurized simultaneously, while other regions remain relaxed. This requires the microprocessor to have the ability to control multiple I / O ports or peripherals in parallel.

[0155] The dynamic adjustment execution module 130 is driven by control commands to perform corresponding actions. The execution effect is applied to the joint through pressure / support output and is immediately monitored by the multi-parameter sensing module 120, forming a closed-loop feedback. Intelligent decision-making is transformed into precise physical action on the joint through precise hardware control.

[0156] 4. Security monitoring and fault tolerance During execution, the system continuously monitors for safety to ensure that the output remains within a safe range.

[0157] Pressure safety threshold: Set an absolute pressure limit (e.g., 40 mmHg) to prevent excessive pressure from causing circulatory problems. Once the pressure sensor reading in any area exceeds this threshold, the microprocessor will immediately force all air pumps to shut down and open the exhaust valve, regardless of any instructions.

[0158] Actuator status monitoring: Monitors the air pump motor current, shape memory alloy drive unit 132 temperature, etc., to determine whether faults such as stall or short circuit have occurred.

[0159] Fail-safe mode: Once a serious fault is detected (such as sensor failure or actuator malfunction), the system immediately switches to a safe mode, for example, resetting all actuators to minimum pressure or rigidity and issuing a critical alert to the user via the app. This is a crucial "fail-safe" design.

[0160] It should be noted that the S205 steps described above can be divided into short-term closed-loop feedback and long-term adaptive optimization, ensuring both immediate safety and long-term effectiveness. Constantly monitoring the execution effect and continuously optimizing the next action is a parallel, continuous background process. Each intervention is not the end point, but rather the starting point for the next, even better intervention.

[0161] 1. Short-term closed-loop feedback Short-term closed-loop focuses on the immediate effect of the current control cycle. Its main goal is to ensure that the currently executed actions can accurately and safely achieve the set control objectives, and to make real-time corrections and ensure safety.

[0162] (1) Real-time collection of feedback data While the control commands are executed in step S204, the multi-parameter sensing module 120 is also continuously working, collecting the latest joint status data. This data constitutes the source of feedback information: Feedback from the actuator itself: For example, the pressure sensor readings in each chamber of the pneumatic unit directly reflect whether the output pressure has reached the target value.

[0163] Joint physiological response feedback: Joint circumference change sensing sensor 122, skin temperature sensor 123, and inertial measurement unit continuously monitor changes in the joint's physiological state. For example, after applying gradient pressure in the swelling relief mode, does the joint circumference show a decreasing trend? Does the skin temperature begin to drop? (2) Feedback control loop Multiple parallel feedback control loops are constructed, the most important of which is the pressure closed-loop control, and its working principle is as follows: Set a target: In step S203, determine the target pressure value for a certain area (e.g., 28 mmHg).

[0164] Execution output: Drive the air pump and valve to pressurize to the target pressure via step S204.

[0165] Measurement feedback: The pressure sensor in this area measures the actual pressure value in real time.

[0166] Calculation error: The microcontroller's microprocessor continuously calculates the error between the target pressure and the actual pressure.

[0167] Corrected output: The microprocessor uses a PID control algorithm to dynamically adjust the output based on the error.

[0168] Proportional (P) term: Provides a fast response that is proportional to the magnitude of the error.

[0169] Integral (I) term: Eliminates steady-state errors (such as pressure decay caused by slow leakage).

[0170] Differential (D) term: suppresses overshoot and makes the pressure smoothly approach the target value.

[0171] Through a rapid, low-level closed loop, the system ensures that the output mechanical stimulus is accurate and stable, unaffected by factors such as airbag elasticity or slight air leakage.

[0172] (3) Effect evaluation and strategy fine-tuning On a slightly longer timescale (e.g., 5 minutes), assess the intervention effect of the current work pattern on joint physiological state.

[0173] Evaluation metrics: The system monitors the trends in key physiological parameters. For example, in "swelling relief mode," the system will monitor whether the rate of change in circumference changes from positive to negative, and whether the uniformity of pressure distribution improves. Strategy fine-tuning: If the effect is not as expected, such as slow swelling reduction, and it is within the safe parameter range of the current mode, a small-scale strategy optimization will be automatically performed. For example, the target value of gradient pressure may be appropriately increased by 5%, or the frequency of pressurization cycles may be increased. This fine-tuning is a self-optimization of the system based on short-term feedback.

[0174] Closed-loop feedback is the cornerstone of precise control in a system. The feedback signal originates from the change in joint state after the action is performed and is captured by the multi-parameter sensing module 120, thus forming a complete loop.

[0175] 2. Long-term adaptive optimization The goal is to make the devices increasingly suitable for the wearer over time, thus achieving personalized rehabilitation and the learning and evolution of personalized models.

[0176] (1) Data accumulation and personalized baseline establishment The system will continuously and permanently encrypt and store the following data: Contextual data: joint state recognition results, user activity scenarios.

[0177] Intervention data: The working mode used and its specific parameters (such as pressure value, time series).

[0178] Results data: Changes in key joint physiological parameters (such as the degree of swelling reduction and skin temperature changes) over a period of time after intervention, as well as subjective feedback submitted by users through the APP application (such as comfort scores and pain levels).

[0179] (2) Optimization based on reinforcement learning Essentially, it involves learning the optimal policy within a Markov decision process. The entire interaction process can be modeled as follows: Status: The joint status identified in step S202 (e.g., "moderate inflammation and swelling").

[0180] Action: The mode and parameters selected in step S203 (e.g., "swelling relief mode, intensity level 2").

[0181] Reward: A quantified value calculated based on the intervention's effectiveness. An example reward function is as follows: r = w1 * (swelling reduction) + w2 * (comfort score) - w3 * (energy expenditure); Here, r is the reward score, and w1, w2, and w3 are weights, indicating whether more emphasis is placed on the swelling reduction effect, user comfort, or energy saving.

[0182] (3) Optimization of algorithm and model update The accumulated data sequences are analyzed periodically. A policy function π is learned such that, in a specific state s, the action 'a' that yields the highest long-term cumulative reward is selected. Lightweight reinforcement learning algorithms, such as Q-learning or its deep neural network variant DQN, are employed. A Q-value table or neural network is progressively updated, recording the expected long-term value of taking action 'a' in state s. Through learning, it is discovered that for the current user, such as in a state of mild swelling at night, using the resting care mode at intensity level 1 yields a higher overall reward than intensity level 2. This is because the user is more sensitive to the pressure at intensity level 2 and scores a lower comfort level. Therefore, the decision-making strategy is automatically optimized to prioritize intensity level 1 in similar future states.

[0183] (4) Cloud-based collaborative collective evolution To protect user privacy and leverage collective intelligence, a federated learning framework is employed. Each user's joint rehabilitation assistive device 100 performs personalized training locally using its own data, and the incremental updates of the local model are encrypted and uploaded to the cloud platform 300. The cloud platform 300 aggregates model updates from thousands of joint rehabilitation assistive devices 100, generating a more powerful and universal global model. The improved global model is then distributed to each joint rehabilitation assistive device 100 and merged with the local model. In this way, all users can benefit from the collective data without sharing any sensitive personal health data, achieving collective evolution.

[0184] Short-term closed-loop feedback ensures the accuracy and safety of each intervention, while long-term adaptive optimization enables personalized and continuous improvement of the intervention strategy. This transforms the Joint Rehabilitation Assistive Device 100 from a tool for executing fixed procedures into an intelligent rehabilitation partner capable of deeply interacting with and growing alongside the user's physiological state. This dual closed-loop design constitutes a disruptive advantage over all existing joint braces.

[0185] It should be noted that in step S206 above, by connecting the joint rehabilitation assistive device 100, the mobile terminal 200, and the cloud platform 300 into a collaborative network, the limitations of the joint rehabilitation assistive device 100 itself in terms of computing power, storage, and knowledge are overcome; this represents a leap from individual intelligence to collective intelligence and remote precision medicine. The cloud platform 300 possesses powerful data storage and computing capabilities, enabling it to perform tasks that the joint rehabilitation assistive device 100 cannot accomplish. For example: (1) Personal health digital twin: Create a continuous, dynamically updated electronic health record for each authorized user. It includes long-term trend data on their joint status, such as "average daily swelling index", "correlation between nocturnal resting heart rate variability and pain onset", and "joint stability score under different activity modes".

[0186] (2) Group data analysis: Mining and analyzing the desensitized group data is the key to achieving collective evolution.

[0187] (3) Comparative analysis of therapeutic effects: For different users with similar conditions (such as those who have undergone knee surgery), the dynamic support mode pressure parameters are most effective in improving gait symmetry.

[0188] (4) Early warning model: Discover common precursor features in the group through machine learning. For example, if a specific pattern of slight increase in skin temperature at night combined with small fluctuations in circumference is found, it will often develop into acute inflammation after 24-48 hours; then this model will be updated and distributed to all relevant users of the joint rehabilitation assistive device 100 to achieve predictive intervention.

[0189] The following section uses knee braces as an example to explain in detail how to perform real-time, dynamic, and precise physical intervention on joint conditions in daily life, using a typical application scenario.

[0190] Scenario 1: Nighttime sleep (00:00 - 06:00) Sensing: The inertial measurement unit shows that the attitude is stable and there is no movement; the circumference, pressure, and temperature parameters fluctuate slightly and steadily within the normal range.

[0191] Recognition: The system determines that the state is stable and at rest.

[0192] Decision and Execution: Activate the resting care mode. Control the pneumatic unit to maintain all airbags at a constant low pressure level (e.g., 12 mmHg). This pressure is below the capillary closure pressure, providing a gentle wrapping sensation while effectively promoting the centripetal return of venous blood and lymph from the lower limbs, relieving mild edema that may occur at night, and allowing the user to wear it almost unnoticed.

[0193] Feedback: The system detected stable pressure and no abnormal parameters, so this mode will be maintained.

[0194] Scene 2: Getting out of bed and walking in the morning (06:30) Sensing: The inertial measurement unit detects the body changing from a lying to a sitting position, and then detects regular gait signals. Pressure sensors show the impact force transmitted to the knee joint when the heel strikes the ground.

[0195] Recognition: The system immediately recognizes the state as a movement activity and further determines it as a low-speed walking sub-mode.

[0196] Decision and Execution: Instantaneous switch to dynamic support mode. Based on real-time gait phase (calculated in real-time by inertial measurement unit data), the microprocessor rapidly inflates the air bladders corresponding to the medial and lateral collateral ligaments to 25 mmHg approximately 50 milliseconds before heel strikes the ground (early support phase) to provide additional stability during weight-bearing. As the foot leaves the ground (swing phase), the air bladder is quickly deflated to 5 mmHg to reduce drag. This process is fast and precise, providing the user with consistent support rather than abrupt mode switching.

[0197] Feedback: The system fine-tunes the inflation volume based on pressure feedback to ensure support.

[0198] Scene 3: Acute joint swelling in the afternoon (2:00 PM) Sensing: The circumference sensor detected an increase of 8mm in leg circumference 10cm above the knee within 1 hour; the temperature sensor showed a 1.5°C increase in local skin temperature; and the user reported mild pain through the APP application.

[0199] Identification: The features conform to the rules for determining the state of inflammation and swelling.

[0200] Decision and Execution: Triggering the swelling reduction mode. First, the system automatically reduces the airbag pressure in the area with the most pronounced swelling (determined by the pressure distribution map) to avoid compression. Second, it initiates a "gradient compression program": controlling the airbag to apply pressure in a wave-like sequence from the ankle to the groin (e.g., ankle 15mmHg -> calf 18mmHg -> below the knee 20mmHg -> above the knee 18mmHg), simulating professional lymphatic drainage techniques to promote tissue fluid return. Simultaneously, a notification is sent to the user via the app: "Joint swelling detected; you have been switched to swelling reduction mode. Please rest." Feedback and Remote Interaction: Data such as the time, duration, and parameter changes of this acute event were encrypted and uploaded to the cloud platform 300. The rehabilitation physician, seeing the system alert on their computer and reviewing the data, sent a command to the user's knee brace via the cloud platform 300: "Execute the preset intensive treatment for 45 minutes tonight from 8:00 PM to 10:00 PM." Upon receiving the command, the knee brace automatically executed the intensive treatment plan at the specified time.

[0201] The embodiments of the present invention have the following beneficial effects: 1. A paradigm shift from passive tools to active partners: The equipment is no longer a cold object, but an intelligent rehabilitation partner that can sense, think, and act.

[0202] 2. Full-cycle, personalized rehabilitation management: covering the entire rehabilitation cycle from acute anti-inflammatory and swelling reduction, to stable functional maintenance, and then to dynamic protection during the active phase, and can be continuously optimized based on personal data.

[0203] 3. Data-driven precision medicine: It provides doctors and patients with unprecedented continuous and objective joint health data, enabling rehabilitation decisions to shift from experience-driven to data-driven.

[0204] 4. Improve treatment adherence and quality of life: By enhancing comfort and intelligence, users are more willing to wear the device for extended periods, thus ensuring the continuity of rehabilitation effects and ultimately improving the user's quality of life.

[0205] By deeply integrating multi-parameter status recognition, dynamic partitioning execution, millisecond-level real-time closed-loop management, and long-term data-driven optimization, an intelligent control system is formed that can truly understand joint status and make precise responses. This achieves a leap from monitoring and alarming to proactive intervention, providing all-weather, fully automated, and personalized rehabilitation management capabilities. Through systematic innovation in intervention methods, intelligent control, and application models, it offers users superior rehabilitation outcomes, enhanced safety, and a better overall experience; it also drives the evolution of rehabilitation medical services towards digital, personalized, and remote management models.

[0206] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A control method for a joint rehabilitation assistive device based on multi-parameter intelligent sensing, wherein the joint rehabilitation assistive device includes a multi-parameter sensing module, a dynamic adjustment execution module, and a controller module, characterized in that, include: The multi-parameter sensing module collects various biomechanical parameters of the joint in real time, including joint circumference change parameters, pressure distribution parameters of the wearing area, and local skin temperature parameters of the joint. The collected biomechanical parameter data are preprocessed and feature extracted to obtain joint state feature information; Based on the joint state feature information, the current joint state is identified using preset recognition rules or recognition models. The joint state includes inflammatory and swollen state, stable resting state, and active state. Based on the identified joint status, select the corresponding target working mode from a variety of preset working modes; Based on the target working mode, control commands are generated, and the controller module drives multiple independent adjustment zones of the dynamic adjustment execution module to apply force to the joint with a specific pressure distribution or support strategy. The system continuously receives feedback data from the multi-parameter sensing module and dynamically adjusts the pressure distribution or support strategy based on the feedback data to form a closed-loop control of sensing-judgment-execution.

2. The joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to claim 1, characterized in that, The preprocessing and feature extraction of the collected biomechanical parameter data includes: The validity of the various biomechanical parameter data is verified, abnormal data is removed, and the time series data is smoothed and filtered. The various biomechanical parameter data of different dimensions were normalized. Extract the magnitude or rate of change of the joint circumference change parameter, the uniformity characteristics of the pressure distribution parameter of the wearing area, and the absolute value or trend of the local skin temperature parameter of the joint.

3. The joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to claim 2, characterized in that, The step of identifying the current state of the joint based on the joint state feature information using preset recognition rules or recognition models includes: When the change in joint circumference exceeds the first threshold, the local skin temperature of the joint is higher than the preset reference value, or the pressure distribution in the wearing area is abnormally concentrated, the joint is determined to be in an inflammatory and swollen state. When the joint circumference change tends to be stable, the pressure distribution in the wearing area is uniform, and the change in local skin temperature of the joint is less than the second threshold, the joint is determined to be in a stable resting state. When the joint circumference changes periodically and the pressure distribution in the wearing area changes dynamically over time, the joint is determined to be in a state of motion.

4. The joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to claim 1, characterized in that, The step of selecting a target working mode from a set of preset working modes based on the identified joint state includes: Predict the type of action a user is about to perform based on historical kinematic data; Based on the identified joint state and the predicted action type, a target working mode is selected.

5. The joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to claim 4, characterized in that, The controller module drives multiple independent adjustment zones of the dynamic adjustment execution module to apply force to the joint with a specific pressure distribution or support strategy, including: When an inflammatory and swollen condition is identified, the pressure distribution strategy is to reduce localized concentrated pressure and apply uniform coverage. When the condition is identified as a stable resting state, apply gentle, continuous circumferential pressure. When identified as a state of motion activity, it provides enhanced support in key stress areas.

6. The joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to claim 1, characterized in that, The dynamic adjustment of pressure distribution or support strategy based on the feedback data includes: A user-personalized baseline model is pre-built, which is constructed by recording the user's baseline physiological parameters and subjective feedback data under standard actions; Based on the user's personalized baseline model and the feedback data, the control parameters of each working mode are optimized through machine learning algorithms to achieve personalized adaptive adjustment.

7. The joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to claim 1, characterized in that, After the step of dynamically correcting the pressure distribution or support strategy based on the feedback data to form a closed-loop control of perception-judgment-execution, the method further includes: The desensitized joint status data and adjustment logs were encrypted and uploaded to the cloud platform. Receive customized working mode parameters from the cloud platform.

8. The joint rehabilitation assistive device control method based on multi-parameter intelligent sensing according to claim 7, characterized in that, The closed-loop control includes short-term closed-loop feedback and long-term adaptive optimization. The short-term closed-loop feedback includes real-time data acquisition, construction of multiple parallel feedback control loops, effect evaluation and strategy fine-tuning. The long-term adaptive optimization includes data accumulation and personalized baseline establishment, optimization based on reinforcement learning, optimization algorithm and model update, and cloud-based collaborative learning evolution.

9. A joint rehabilitation assistive device, characterized in that, The method for controlling joint rehabilitation assistive devices based on multi-parameter intelligent sensing as described in any one of claims 1 to 8 includes: A flexible wearable body for fixing to a user's joint area; A multi-parameter sensing module is disposed inside the flexible wearable body and includes a flexible pressure sensor array, a joint circumference change sensing sensor and a skin temperature sensor, used to collect joint state-related parameters. A dynamic adjustment execution module is disposed within the flexible wearable body. The dynamic adjustment execution module includes a multi-zone pneumatic pressurization unit and a shape memory alloy driving unit, which are used to apply zoned adjustable pressure or support to the joints. The controller module is electrically connected to the multi-parameter sensing module and the dynamic adjustment execution module, respectively. It is used to process the collected joint state-related parameters, control the working state of the dynamic adjustment execution module, identify joint state, and switch working modes.

10. A joint rehabilitation assistive device control system based on multi-parameter intelligent sensing, characterized in that, include: The joint rehabilitation assistive device according to claim 9; The mobile terminal has an application installed to display joint status, set preferences, and receive prompts; The cloud platform is used to store and analyze user data, provide remote monitoring and intervention interfaces, train better state recognition models based on user data, and distribute model update files to joint rehabilitation assistive devices.

Images