Rehabilitation training action evaluation method based on artificial intelligence and internet of things

Abstract

This invention provides a rehabilitation training movement assessment method based on artificial intelligence and the Internet of Things, belonging to the field of movement assessment technology. The method includes: synchronously collecting multimodal raw data streams of users performing rehabilitation training movements through an IoT sensor unit cluster; performing spatiotemporal alignment, filtering and denoising, and feature decomposition on the data to obtain standardized multidimensional movement feature time-series data; extracting quantitative assessment parameters in real time based on a rehabilitation medicine knowledge model; inputting the parameters into a pre-trained movement quality assessment and compensation pattern recognition model to generate intelligent assessment results containing specific compensation pattern labels and root cause inferences; fusing the intelligent results with historical data to generate a visualized assessment report; and dynamically adjusting the parameters of the personalized rehabilitation prescription for the next cycle based on trend analysis. This invention achieves automated, quantitative, and real-time assessment of the intrinsic quality, hidden compensation, and kinetic chain synergy of training movements without on-site therapist supervision.

Description

Technical Field

[0001] This invention relates to the field of motion assessment technology, and in particular to a method for assessing rehabilitation training motions based on artificial intelligence and the Internet of Things. Background Technology

[0002] In the field of rehabilitation medicine, movement quality assessment is a core component in guiding patients to conduct effective and safe training. Taking common post-shoulder joint rehabilitation as an example, patients need to perform long-term exercises at home, such as lateral shoulder raises, to restore function. However, current technology struggles to provide accurate and objective assessments in unsupervised home settings. Current mainstream assessment methods heavily rely on the on-site observation and experience of rehabilitation therapists. Therapists assess training effectiveness and identify compensatory movements (such as shoulder shrugging and trunk tilting to complete arm raises) by visually observing the patient's range of motion, speed, and changes in body posture. This method has significant limitations. First, the assessment results are highly subjective, and different therapists may have different judgments. Second, the human eye cannot accurately quantify the detailed parameters of movement, such as instantaneous changes in joint angular velocity, the activation sequence of specific muscles, and millimeter-level shifts in the center of gravity. More importantly, some hidden, subtle compensatory patterns or minute abnormalities in neuromuscular control are easily missed by the naked eye alone. This can lead to patients training in incorrect movement patterns, not only delaying the rehabilitation process but also potentially causing secondary injuries or developing incorrect movement patterns. In recent years, some wearable devices based on single-type sensors have emerged. However, these devices are limited in scope, measuring only isolated parameters such as angle or muscle activation intensity, and cannot establish the correlation between kinematics, dynamics, and neural control during movement execution. For example, the device may detect insufficient arm raising angle but cannot automatically distinguish whether this is due to limited joint mobility or a compensatory strategy of shrugging due to pain or insufficient muscle strength. Furthermore, it cannot identify the underlying causes of this compensation (such as trapezius overactivation or core instability). This "data-only, strategy-less" assessment is static and isolated, failing to meet the needs of precision rehabilitation for multi-dimensional, process-oriented, and causal analysis of movement execution quality. Therefore, there is an urgent need for a rehabilitation training movement assessment method that can deeply integrate multi-source sensor data, utilize artificial intelligence technology to simulate the thinking of rehabilitation experts, and achieve automated, quantitative, real-time assessment in a home environment, revealing the underlying strategies and root causes of movement execution, to address the shortcomings of existing technologies. Summary of the Invention

[0003] The purpose of this invention is to provide a rehabilitation training movement assessment method based on artificial intelligence and the Internet of Things, in order to solve a series of problems in existing technologies, such as: the high subjectivity and difficulty in quantification caused by the heavy reliance on human experience in assessment; the single assessment dimension of traditional equipment, which can only record surface mechanical counts and cannot assess the intrinsic quality of movements or identify hidden compensatory patterns; the presentation of instantaneous static results in assessments, which contradicts the long-term dynamic and temporal requirements of rehabilitation training; and the isolated assessment perspective, which cannot analyze from the overall kinetic chain level. The specific technical solution of this invention is as follows: This invention provides a method for assessing rehabilitation training movements based on artificial intelligence and the Internet of Things, comprising the following steps: S10, through a cluster of IoT sensor units configured in key parts of the user's body and the training environment, synchronously collects multimodal raw data streams during the execution of the target rehabilitation training movements; S20 performs spatiotemporal alignment, filtering and denoising, and feature decomposition on the multimodal raw data stream to obtain standardized multi-dimensional action feature time-series data; S30, based on a rehabilitation medicine knowledge model, extracts quantitative evaluation parameters in real time from multi-dimensional motion feature time-series data; S40, input the quantitative evaluation parameters into the pre-trained action quality evaluation and compensation pattern recognition model to generate intelligent evaluation results containing specific compensation pattern labels and root cause inferences; S50 integrates intelligent evaluation results with historical training data to generate a visual evaluation report that includes multi-dimensional scores, qualitative descriptions, risk warnings, and improvement points. Based on trend analysis in the current and historical evaluation reports, S60 dynamically adjusts the personalized rehabilitation prescription parameters for the next training cycle through a preset rule engine or recommendation algorithm.

[0004] Furthermore, in step S10, the IoT sensing unit cluster includes an inertial measurement unit worn on the user's body, an electromyography sensor, a pressure sensor, and a depth camera deployed in the environment. Step S20 includes filtering and attitude calculation of inertial measurement unit data to obtain joint angle time-series data; processing electromyography sensor signals to obtain muscle activation intensity time-series signals; calculating the movement trajectory of the pressure center from pressure sensor data; and extracting the three-dimensional spatial coordinates of key human body points from depth camera data.

[0005] Furthermore, the quantitative evaluation parameters in step S30 include the target joint range of motion achievement rate, compensatory motion amplitude, muscle activation timing disorder, motion trajectory smoothness, and kinetic chain compensation index. In step S40, the motion quality assessment and compensation pattern recognition model is a gradient boosting decision tree model, and the input is a feature vector composed of quantitative assessment parameters. The model output includes an overall action quality score and the classification probability of specific compensation patterns; The model integrates a rule-based root cause inference module, which is used to generate root cause inferences based on the logical relationships between quantitative evaluation parameters; In step S50, the visual evaluation report includes: The quantitative evaluation parameters are compared with historical and target values ​​in the form of a five-dimensional radar chart. List the identified compensation patterns and their underlying causes inferences in words; Risk warnings are triggered based on the overall action quality score and the probability of compensation patterns.

[0006] Further, step S60 includes: Retrieve historical user evaluation data sequences to analyze trends in overall action quality scores and the probability of compensation patterns; If the progress is determined to be stable, the target angle or resistance load for the next training cycle will be automatically increased. If a plateau or a specific compensatory pattern persists, targeted corrective exercises will be automatically inserted into the next training cycle.

[0007] Furthermore, step S10 also includes: enabling a dedicated shoulder joint trainer integrating an encoder and an inertial sensing unit, as well as a visual recognition channel based on a mobile terminal camera, to form an auxiliary sensing network in parallel with the IoT sensing unit cluster. A dedicated shoulder joint trainer is used to collect data on the number of rotations, angular velocity, and posture of its handle; The visual recognition channel is used to extract the coordinates of two-dimensional skeletal key points of the user's joints in real time; S20 also includes: establishing a multi-source data spatiotemporal alignment and feature enhancement module, used to fuse and process data from IoT sensor unit clusters, dedicated shoulder joint trainers and visual recognition channels; The stability coefficient of the ball-swinging motion was calculated from the angular velocity data of the dedicated shoulder joint trainer. Perform cross-modal data verification by comparing the joint angles estimated by visual recognition with the joint angles calculated by the inertial measurement unit, and assigning information weights based on the difference.

[0008] Furthermore, the S30 also includes the following new quantitative evaluation parameters extracted based on data from a dedicated shoulder joint trainer and a visual recognition channel: Load adaptability quantification metrics are used to assess the degree of match between the current training load and the user's capabilities; The direction-switching neural control index is used to quantify the user's control efficiency when switching action directions.

[0009] Furthermore, S40 also includes: the motion quality assessment and compensation pattern recognition model adds fusion analysis of load adaptability quantitative indicators and direction switching neural control index; The model adds a forward inference branch that performs rapid pre-analysis based on trainer data, and inputs the inference results as meta-information into the model's backbone network. The model's output layer now includes a device interaction credibility score, used to quantify the quality of the external data sources relied upon in this evaluation.

[0010] Furthermore, the S50 also includes an integrated equipment interaction and adaptability analysis section in the visualization evaluation report. This section includes dynamic curves showing the changes in load adaptability quantitative indicators and ball-swing motion stability coefficient over training time, as well as specific text comments on the ability to switch directions.

[0011] Furthermore, the S60 also includes expanded prescription dynamic adjustment rules based on parameters obtained from a dedicated shoulder joint trainer and a visual recognition channel: If the user's continuous load adaptability quantification index shows that the load is too light and the action quality score is high, the system recommends upgrading the training load. If the user's direction-switching neural control index remains low, the system will automatically insert an adaptive neural control training sub-prescription in the next training cycle.

[0012] Furthermore, S40 also includes: when the model determines that there are multiple compensation modes and their probabilities are close, a multi-stage joint inference strategy is used for arbitration, including calculating the distance between the input feature vector and the pre-stored typical compensation mode prototype feature vector, selecting the two candidate prototypes with the smallest distance; retrieving secondary features that are strongly correlated with the candidate prototypes and comparing them; and determining the candidate prototype with higher matching degree as the final recognition result. And / or, when the smoothness of the movement trajectory increases abnormally over multiple consecutive movement cycles, fatigue state correlation analysis is performed. By analyzing the relationship between the slope of the smoothness deterioration and the slope of the decrease in the median frequency of the electromyographic signal, the cause of movement instability can be distinguished as either neural control disorder or muscle fatigue.

[0013] The beneficial effects of this invention are as follows: By deeply integrating IoT sensing and artificial intelligence analysis, this invention constructs a complete technical closed loop from synchronous acquisition of multi-source data and deep feature extraction to intelligent pattern recognition and decision-making intervention. Without the need for on-site supervision by therapists, it achieves automated, quantitative, and real-time assessment of the intrinsic quality, hidden compensation, temporal changes, and kinetic chain synergy of rehabilitation training movements. It can not only output intelligent assessment results containing specific compensation pattern labels and root cause inferences, advancing the assessment from simple right-or-wrong judgments to the pathophysiological interpretation of movement execution strategies, but also dynamically adjust subsequent rehabilitation prescriptions based on historical trends. This fundamentally solves the pain points of traditional methods in assessment, such as subjectivity, single dimension, and delayed feedback, providing a reliable technical foundation for personalized and precise rehabilitation. Attached Figure Description

[0014] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the method described in Embodiment 1 of the present invention. Detailed Implementation

[0016] 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, and 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.

[0017] The concepts involved in this application will first be described with reference to the accompanying drawings. It should be noted that the following descriptions of various concepts are only for the purpose of making the content of this application easier to understand and do not constitute a limitation on the scope of protection of this application; furthermore, the embodiments and features in the embodiments of this application can be combined with each other unless otherwise specified. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] Example 1

[0019] like Figure 1 As shown, this invention provides a rehabilitation training movement assessment method based on artificial intelligence and the Internet of Things, including the following steps: S10, through a cluster of IoT sensor units configured in key parts of the user's body and the training environment, synchronously collects multimodal raw data streams during the execution of the target rehabilitation training movements; S20, perform spatiotemporal alignment, filtering and denoising, and feature decomposition on the multimodal raw data stream to obtain standardized multi-dimensional action feature time-series data; S30, based on the rehabilitation medicine knowledge model, extract at least five interrelated quantitative evaluation parameters in real time from the multi-dimensional motion feature time-series data; S40, the quantitative evaluation parameters are input into the pre-trained action quality evaluation and compensation pattern recognition model to generate an intelligent evaluation result containing specific compensation pattern labels and root cause inferences. S50, integrates the intelligent evaluation results with historical training data to generate a visual evaluation report that includes multi-dimensional scores, qualitative descriptions, risk warnings and improvement points; Based on trend analysis in the current and historical evaluation reports, S60 dynamically adjusts the personalized rehabilitation prescription parameters for the next training cycle through a preset rule engine or recommendation algorithm.

[0020] The principle and effect of the above technical solution are as follows: This invention provides a rehabilitation training movement assessment method that deeply integrates Internet of Things perception and artificial intelligence analysis. By constructing a complete technical closed loop from multi-source data synchronous acquisition to deep feature extraction and then to intelligent pattern recognition and decision intervention, it realizes automated, quantitative and real-time assessment of the intrinsic quality, hidden compensation, temporal changes and kinetic chain synergy of training movements without the need for on-site supervision by therapists. It fundamentally solves the problems of subjective, single-dimensional and static isolation in existing technology assessments.

[0021] In one embodiment, S10 includes: S101, configure an IoT sensing unit cluster, the cluster including: a nine-axis inertial measurement unit worn on the distal end of the upper arm on the user's training side, with a sampling frequency of 100Hz; a surface EMG electromyography sensor attached to the upper trapezius and middle deltoid on the training side, with a sampling frequency of 1000Hz; a pressure-sensing insole placed on the sole of the user's foot, with a sampling frequency of 50Hz; and a depth camera deployed 3 meters directly in front of the user, with a sampling frequency of 30fps.

[0022] S102 uses a hardware timestamp-based synchronization protocol to transmit the raw data streams collected by the nine-axis inertial measurement unit, surface EMG electromyography sensor, pressure-sensing insole, and depth camera to the edge computing gateway in real time via Bluetooth 5.2 and Wi-Fi 6 links for time synchronization and alignment, forming a multimodal raw data packet under a unified time reference.

[0023] The principle and effect of the above technical solution are as follows: Through a carefully configured sensor cluster, comprehensive and synchronous capture of motion kinematics, muscle activation dynamics, overall balance and center of gravity changes, and visual spatial trajectory data is achieved. The combination of high frequency and multimodal approaches provides a complete data foundation for subsequently revealing hidden neuromuscular control abnormalities.

[0024] In one embodiment, S20 includes: S201 uses a fourth-order Butterworth low-pass filter to filter the raw acceleration and angular velocity data of the nine-axis inertial measurement unit, with a cutoff frequency of 20Hz to eliminate high-frequency noise. Then, a gradient descent fusion algorithm is used to calculate the attitude, and the Euler angles of the upper arm in the coronal, sagittal and horizontal planes of the human body are calculated in real time.

[0025] S202 performs full-wave rectification and bandpass filtering on the raw signal from the surface EMG electromyography sensor, and calculates its linear envelope to obtain the activation intensity time-series signal of each target muscle.

[0026] S203 calculates the movement trajectory and speed of the pressure center within the foot plane per second based on the pressure-sensing insole data.

[0027] S204 uses a lightweight human skeleton key point detection model to extract the three-dimensional spatial coordinates of key anatomical landmarks such as the acromion, upper edge of the sternum, and greater trochanter in real time from the depth camera data stream.

[0028] The principle and effect of the above technical solution are as follows: by performing targeted preprocessing and feature calculation on raw data from different sources, the original voltage and pixel signals are transformed into standardized feature data with clear biomechanical significance, such as joint angles, muscle activation levels, and center of gravity trajectories, providing direct and high-quality data input for subsequent quantitative analysis.

[0029] In one embodiment, S30 includes: S301, Calculate the target joint range of motion achievement rate: When the user performs a shoulder lateral raise, extract the peak angle θ of a single movement cycle from the time-series data of the upper arm's tilt angle in the coronal plane. peak Define the activity level achievement rate R. ROM =(θ peak / θ prescription ) * 100%, where θ prescription The target angle set for the rehabilitation prescription, such as 90 degrees.

[0030] S302, Calculate the compensatory motion amplitude: From the time-series data of the vertical displacement of the acromion point extracted from the depth camera, calculate the maximum vertical displacement D of the acromion point relative to the initial resting position within a single lateral raise movement cycle.shoulder Define the shrug amplitude as D. shoulder If D shoulder If the deviation is greater than 10mm, significant shoulder shrugging compensation is considered to exist. Simultaneously, the lateral shift L of the body's center of gravity to the opposite side during the movement is calculated from the trajectory of the center of pressure. shift .

[0031] S303, Calculate the temporal disorder of muscle activation: From the electromyographic envelopes of the middle deltoid and upper trapezius, identify the starting time t when the activation intensity of each muscle exceeds three standard deviations of its resting value. deltoid and t trapezius Calculate the muscle activation timing difference ΔT = t trapezius - t deltoid In the standard mode, the middle deltoid should be preferentially activated, meaning ΔT should be positive. If ΔT < -50ms, it is considered premature activation of the trapezius muscle.

[0032] S304, Calculating the smoothness of the movement trajectory: From the time-series data of the upper arm's angular velocity in the coronal plane, calculate the normalized jerk (the integral of the modulus of jerk) of the angular velocity within a single movement cycle, i.e., the smoothness index S. The higher the S value, the less smooth the movement, and the more likely it is to have jitter or stuttering. The calculation formula is S = (1 / 2) *∫(jerk) 2 ) dt, where jerk is the derivative of angular acceleration.

[0033] S305, Calculate the kinetic chain compensation index: This index C chain For comprehensive parameters, C chain = 0.5(D shoulder / 15mm)+ 0.3(L shift / 20mm) + 0.2*(absolute value of negative ΔT / 100ms), is used to quantify the degree of diffusion of compensatory motion along the kinetic chain. Where D shoulder L represents the acromion elevation amplitude. shift ΔT represents the lateral offset of the center of gravity, and ΔT represents the time difference of muscle activation.

[0034] The principle and effect of the above technical solution are as follows: Five quantitative parameters are defined to characterize the quality of movement from different dimensions. These parameters not only focus on the result of the movement (such as range of motion), but also reveal the process and strategy of the movement (such as compensatory amplitude, muscle timing, and smoothness). They also link local joints with the overall posture, thereby constructing a comprehensive, multi-layered and computable evaluation index system.

[0035] In one embodiment, S40 includes: S401, a lightweight gradient boosting decision tree model is constructed as the action quality assessment and compensation pattern recognition model. This model uses the five quantitative assessment parameters (R) calculated in step S30. ROM D shoulder , ΔT, S, C chain ) as the input feature vector.

[0036] S402, the input feature vector is input into the model, and the model outputs a five-dimensional vector, which are: overall action quality score (0-100 points), and four specific compensation mode classification probabilities: P1 (scapularis-dominant shrug compensation), P2 (trunk tilt compensation), P3 (neuromuscular control instability), and P4 (action initiation mode error).

[0037] S403, the model integrates a rule-based root cause inference module. When P1 > 0.7 and ΔT < -50ms, a root cause inference is added to the evaluation results: overactivity of the upper trapezius / delayed activation of the middle deltoid, suggesting increased scapular stability training and strengthening of middle deltoid isolation; when P2 > 0.6 and C chain When the score is >0.5, the additional root cause inference is: insufficient core stability or improper use of hip joint strategies. It is recommended to assess and strengthen the ipsilateral hip abductor muscles.

[0038] The principle and effect of the above technical solution are as follows: By integrating machine learning models with medical rule knowledge, not only can a comprehensive score of action quality be achieved, but more importantly, it can automatically identify specific and clinically common abnormal compensation patterns and provide possible functional root cause inferences based on the correlation between parameters, thus deepening the assessment from what happened to why it happened, providing a direct basis for precise intervention.

[0039] In one embodiment, S50 includes: S501 generates a five-dimensional radar chart, with the five dimensions corresponding to the five quantitative evaluation parameters in S30, and compares the current value with the historical average and the prescription target value in the same chart.

[0040] S502, in the assessment report, lists in textual description the main compensatory patterns identified (e.g., significant shoulder shrugging compensation was detected, with a probability of 85%) and their root cause inferences (e.g., possibly stemming from overactivation of the upper trapezius muscle).

[0041] S503 triggers different levels of risk warnings based on the overall action quality score and the probability of compensation patterns: if the score is below 60 and P1 or P2 is greater than 0.8, a high-risk warning (red) is triggered, prompting immediate correction; if the score is between 60 and 80, a medium-risk warning (yellow) is triggered, prompting attention; if the score is above 80, positive feedback (green) is given, prompting maintenance.

[0042] In one embodiment, S60 includes: S601 retrieves recent training evaluation reports from the cloud database and extracts historical sequences of overall movement quality scores, probabilities of major compensation patterns, and kinetic chain compensation indices.

[0043] S602, Analyze historical trends: If the overall action quality score shows a continuous upward trend and the probability of the main compensation mode is below 0.3 for three consecutive training sessions, then it is determined that the user has made stable progress in this action.

[0044] S603, based on trend determination results, automatically adjusts the prescription through a rule engine: if it is determined to be stable progress, the training target angle θ for that action in the next cycle is adjusted. prescription Increase the temperature by 5 degrees or increase the resistance load by 10%. If a plateau or a specific compensatory pattern persists, a targeted corrective exercise (such as isometric contraction of the scapula to target overactivation of the upper trapezius) will be automatically inserted into the training plan of the next cycle, and the intensity of the main training movement will be reduced.

[0045] In one embodiment, in S40, if the action quality assessment and the compensation pattern recognition model determine that there are multiple compensation patterns with similar probabilities, a multi-stage joint reasoning strategy is used for arbitration. The specific steps are as follows: SA41: Extract the input feature vector and calculate its Euclidean distance to the prototype feature vectors of several typical compensatory patterns pre-stored in the model.

[0046] SA42: Select the two prototype patterns with the smallest Euclidean distance, and retrieve secondary features that are strongly correlated with these two prototype patterns. The secondary features include: the slope of the rising edge of the electromyographic envelope and the instantaneous velocity of the pressure center at the peak of the action.

[0047] SA43: Recalculate the value of the current data on the secondary feature and compare it with the typical secondary feature values ​​of the two candidate prototypes.

[0048] SA44: The candidate prototype with higher matching degree of secondary features is determined as the final compensation pattern recognition result.

[0049] The proposed multi-stage joint reasoning strategy effectively addresses potential misjudgments by the model in complex, hybrid compensation scenarios. By introducing deeper and more discriminative biomechanical features for secondary arbitration, it significantly improves the accuracy and robustness of pattern recognition.

[0050] In S40, if the motion trajectory smoothness S in the quantitative evaluation parameter increases abnormally over multiple consecutive motion cycles, then fatigue state correlation analysis is performed. The specific steps of the fatigue state correlation analysis are as follows: SB41: Record the smoothness S value for each action cycle in the current training group and its order within the group.

[0051] SB42: Plot the action on the x-axis and the smoothness S on the y-axis, fit a linear regression line, and calculate its slope K. s .

[0052] SB43: Simultaneously, calculate the slope K of the median frequency decrease of the electromyographic signals of the upper trapezius muscle acquired by the electromyography sensor within the same time period. mf .

[0053] SB43: If K is detected s > 0.15 (per-cycle smoothness deterioration rate) and K mf If the rate is less than -0.08 (median frequency decline rate per cycle), it is determined that the current decline in movement quality is mainly caused by muscle fatigue. The evaluation report will include a note: "Movement instability is related to muscle fatigue. It is recommended to appropriately extend the rest time between sets to distinguish it from instability caused by neural control disorder."

[0054] Specifically, the median frequency refers to the frequency point in the electromyographic signal power spectrum that divides the total power into two equal parts, and its decrease with muscle fatigue is a recognized physiological phenomenon.

[0055] In its specific implementation, this invention constructs a complete evaluation system from microscopic muscle electrophysiological signals to macroscopic spatial motion trajectories, from instantaneous movement quality to long-term trends, through the aforementioned tightly coupled steps. In an unsupervised environment, it achieves automated identification of deeper issues such as whether shoulder shrugging compensation is caused by insufficient scapular control or core instability, and whether movement tremors stem from poor neural control or muscle fatigue. This advances rehabilitation training evaluation from simple right-or-wrong judgments to the pathophysiological interpretation of movement execution strategies, laying a technical foundation for personalized and precise rehabilitation.

[0056] Example 2

[0057] The present invention also provides a rehabilitation training movement assessment method based on artificial intelligence and the Internet of Things. Step S10 in Embodiment 1 further includes deploying a dedicated shoulder joint trainer integrated with a high-precision encoder and an inertial sensing unit, and simultaneously enabling a visual recognition channel based on a mobile terminal camera to form a low-cost auxiliary perception and verification network that runs parallel to the high-precision wearable sensing network in Embodiment 1.

[0058] Specifically, the dedicated shoulder joint trainer integrates a 1024 PPR photoelectric encoder and a six-axis inertial measurement unit (IMU) inside its cylindrical, non-slip handle. The IMU's sampling frequency is configured at 50 Hz to collect the absolute number of rotations, real-time angular velocity, and rotation direction of the trainer's handle. The handle is connected to colored balls via a 0.8-meter steel cable. The balls come in four sizes: ball #1 (0.5 kg, red), ball #2 (0.75 kg, blue), ball #3 (1.0 kg, green), and ball #4 (1.25 kg, black). The trainer connects to an application on the user's mobile terminal via Bluetooth 5.2 to upload real-time data on rotation count, direction, real-time angular velocity, and handle posture.

[0059] Meanwhile, the visual recognition channel is implemented by calling the front-facing camera of the user's mobile terminal, which captures RGB video streams at a rate of 30 frames per second; in the application, a human pose estimation model based on the MobileNetV2 lightweight neural network is integrated to extract the pixel coordinates of two-dimensional skeletal key points of the user's shoulder, elbow, wrist and hip joints in real time.

[0060] S20 also includes a multi-source data spatiotemporal alignment and feature enhancement module for fusing and processing data from the dedicated shoulder joint trainer, the visual recognition channel, and the high-precision wearable sensor network.

[0061] The module first synchronizes and aligns the Bluetooth data stream of the trainer, the visual key point data stream, and the multimodal data stream at the millisecond level using network time protocol and hardware timestamp.

[0062] The module then calculates a key derived parameter from the real-time angular velocity data uploaded by the trainer: the ball-swing motion stability coefficient Cs, calculated as Cs = 1 - (std(ω)). window ) / mean(ω window )), where ω window The data sequence represents the angular velocity collected within a sliding time window, with a window width of 3.0 seconds and a sliding step size of 0.1 seconds. `std` represents the standard deviation function, and `mean` represents the average value function. A value closer to 1 indicates smaller angular velocity fluctuations and more stable force application.

[0063] The module also performs cross-modal data verification: it verifies the angle θ between the user's upper arm and torso in the coronal plane, as estimated by the visual recognition model. v The same included angle θ calculated by the inertial measurement unit worn on the upper arm in Embodiment 1. i Compare; if the absolute difference |θ v - θ i If the temperature exceeds 8 degrees for 5 consecutive seconds, a data reliability verification flag is triggered, and this value is assigned to θ in subsequent fusion calculations. v With θ i Assign confidence weights of 0.3 and 0.7 respectively.

[0064] Step S30 in Embodiment 1 further includes extracting the following newly added quantitative evaluation parameters based on the physical and visual data provided by the dedicated shoulder joint trainer: S306, Calculate the load fit quantification index L: This index is used to assess the degree of matching between the current training load and the user's muscle strength and control ability. First, extract the peak angular velocity ω within a single complete movement cycle from the trainer data. peak The target angular velocity for the current ball-swing weight and movement type is set as ω in the rehabilitation prescription. target The formula for calculating the load adaptability quantification index L is: L = (mean(ω)) / (ω) peak ) / ω target ) * Cs. Where, mean(ω peak L is the average of the peak angular velocities of the last 5 movement cycles, and Cs is the stability coefficient of the ball-swing motion calculated during the same period. A value between 0.9 and 1.1 indicates good load matching; L > 1.2 indicates that the load is too light and fails to form an effective training stimulus; L < 0.8 indicates that the load may be too heavy or there is serious compensation that prevents the target speed from being reached.

[0065] S307, Calculate the direction-switching neural control index N: This index quantifies the neuromuscular control efficiency of the user when switching directions during forward and reverse rotation training. At the instant t0 when the trainer detects a change in rotation direction, extract the angular velocity data within the time window [t0-0.3 seconds, t0+0.3 seconds]. Calculate the smoothness of the angular velocity increase after the direction change: N = 1 - (|ω(t0+0.1s) - ω(t0+0.05s)| / max(|ω window |) ). Where ω(t) represents the angular velocity at time t, max(ω window N represents the maximum absolute value of the angular velocity within this time window. The closer the N value is to 1, the smoother the acceleration process after the direction switch and the more precise the control; an N value below 0.7 indicates hesitation or explosive compensation in the switching control.

[0066] Step S40 further includes: the motion quality assessment and compensation pattern recognition model, adding a fusion analysis of the load adaptability quantification index L and the direction switching neural control index N.

[0067] The model adds a forward inference branch, which first performs a rapid pre-analysis based on the trainer data (L, Cs, direction switching points). For example, if it is identified that during a 30-second abduction training session, the Cs value linearly decreases from an initial 0.92 to 0.75, while the L value decreases from 1.0 to 0.7, this branch infers with high confidence that "progressive muscle fatigue leads to a decrease in control stability," and inputs this inference as meta-information into the model's backbone network, guiding the model to give higher weight to changes in muscle fiber recruitment patterns in electromyographic signals.

[0068] The model's output layer now includes a device interaction credibility score D, calculated using the following formula: D = (T connection / T total ) * (N validpose / N totalpose ), where T connection To maintain a stable Bluetooth connection duration during training, T total N represents the total training duration. validpose N represents the number of valid pose frames output by the visual recognition module. totalpose The total number of frames. The D value ranges from 0 to 1 and is used to quantify the quality of the external data source on which this evaluation is based. When D < 0.8, the evaluation conclusion will include a note that "data integrity is limited, and it is recommended to review the data under the condition that the sensors are fully worn."

[0069] Step S50 also includes: In the generated evaluation report, an integrated device interaction and adaptability analysis section is included. This section contains dynamic curves, with the horizontal axis representing the training time axis and the vertical axis simultaneously displaying the change trajectory of the load adaptability quantitative index L and the ball-swing motion stability coefficient Cs. Different colored backgrounds are used to mark intervals such as "good adaptability", "light load", and "fatigue accumulation". This section also outputs specific comments on the ability to switch directions in structured text, such as: In this training, the average value of the direction switching neural control index N is 0.85, which is excellent, and the switching process is decisive and smooth; or the average value of the direction switching neural control index N is 0.65, which shows obvious lag, and it is recommended to conduct low-speed, unloaded direction perception exercises.

[0070] Step S60 further includes: expanding the prescription dynamic adjustment rules based on new parameters obtained from the dedicated shoulder joint trainer and the visual recognition channel.

[0071] Extended Rule 1: If a user's data from three consecutive training days shows that the daily average value of their load adaptability quantitative index L is consistently greater than 1.15, and the overall movement quality score for the corresponding training objective is higher than 88 points, then the prescription adjustment engine will push a suggestion to the user and rehabilitation therapist through the application: the current load (such as blue ball No. 2) has been adapted, and it is recommended to try a higher level load (such as green ball No. 3) in the next training session to maintain the effectiveness of the training stimulus.

[0072] Extended Rule 2: If, during the execution of a training prescription that includes direction switching, the system detects that the user's direction switching neural control index N is consistently below 0.7, the system will automatically generate and insert a 5-minute adaptive neural control training sub-prescription in the next training cycle. This sub-prescription requires the user to consciously practice direction switching at an extremely slow speed (target angular velocity set to 30% of the normal value) using 50% of the current load, aiming to enhance motion perception and control precision.

[0073] Building upon the comprehensive and high-precision assessment system constructed in Example 1, this invention introduces tightly coupled dedicated device sensing nodes and lightweight visual verification channels. It deeply mines the biomechanical and neural control information contained in the physical signals of the dedicated training device itself (such as angular velocity fluctuations and transient responses to direction switching), defining new assessment dimensions with clear rehabilitation semantics, such as the ball-swinging motion stability coefficient Cs and the direction-switching neural control index N, transforming simple devices into intelligent assessment terminals. It constructs a flexible assessment architecture with master-slave verification and smooth degradation between a high-precision multimodal network and a low-cost dual-channel network. Through data verification and confidence fusion mechanisms, it ensures high reliability of assessment results even in resource-constrained scenarios, greatly expanding the system's applicability and user compliance. It achieves a two-way intelligent closed loop between assessment results and training prescriptions, not only evaluating movements but also directly driving personalized adjustments to training parameters (such as load weight and training mode) based on device interaction data (such as the load adaptability index L). This constitutes a complete solution covering both professional and home scenarios, and connecting the entire chain from assessment to intervention.

[0074] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A rehabilitation training action evaluation method based on artificial intelligence and Internet of Things, characterized in that, Includes the following steps: S10, through a cluster of IoT sensor units configured in key parts of the user's body and the training environment, synchronously collects multimodal raw data streams during the execution of the target rehabilitation training movements; S20 performs spatiotemporal alignment, filtering and denoising, and feature decomposition on the multimodal raw data stream to obtain standardized multi-dimensional action feature time-series data; S30, based on a rehabilitation medicine knowledge model, extracts quantitative evaluation parameters in real time from multi-dimensional motion feature time-series data; S40, input the quantitative evaluation parameters into the pre-trained action quality evaluation and compensation pattern recognition model to generate intelligent evaluation results containing specific compensation pattern labels and root cause inferences; S50 integrates intelligent evaluation results with historical training data to generate a visual evaluation report that includes multi-dimensional scores, qualitative descriptions, risk warnings, and improvement points. Based on trend analysis in the current and historical evaluation reports, S60 dynamically adjusts the personalized rehabilitation prescription parameters for the next training cycle through a preset rule engine or recommendation algorithm.

2. The method of claim 1, wherein, In step S10, the IoT sensing unit cluster includes an inertial measurement unit worn on the user's body, an electromyography sensor, a pressure sensor, and a depth camera deployed in the environment. Step S20 includes filtering and attitude calculation of the inertial measurement unit data to obtain joint angle time-series data; processing the electromyography sensor signal to obtain muscle activation intensity time-series signal; and calculating the movement trajectory of the pressure center from the pressure sensor data. Extract the three-dimensional spatial coordinates of key human body points from depth camera data.

3. The method of claim 1, wherein, The quantitative evaluation parameters in step S30 include the target joint range of motion achievement rate, compensatory motion amplitude, muscle activation timing disorder, movement trajectory smoothness, and kinetic chain compensation index. In step S40, the motion quality assessment and compensation pattern recognition model is a gradient boosting decision tree model, and the input is a feature vector composed of quantitative assessment parameters. The model output includes an overall action quality score and the classification probability of specific compensation patterns; The model integrates a rule-based root cause inference module, which is used to generate root cause inferences based on the logical relationships between quantitative evaluation parameters; In step S50, the visual evaluation report includes: The quantitative evaluation parameters are compared with historical and target values ​​in the form of a five-dimensional radar chart. List the identified compensation patterns and their underlying causes inferences in words; Risk warnings are triggered based on the overall action quality score and the probability of compensation patterns.

4. The method of claim 1, wherein, Step S60 includes: Retrieve historical user evaluation data sequences to analyze trends in overall action quality scores and the probability of compensation patterns; If the progress is determined to be stable, the target angle or resistance load for the next training cycle will be automatically increased. If a plateau or a specific compensatory pattern persists, targeted corrective exercises will be automatically inserted into the next training cycle.

5. The method of claim 1, wherein, Step S10 also includes: enabling a dedicated shoulder joint trainer integrating an encoder and an inertial sensing unit, as well as a visual recognition channel based on a mobile terminal camera, to form an auxiliary sensing network that runs parallel to the IoT sensing unit cluster. A dedicated shoulder joint trainer is used to collect data on the number of rotations, angular velocity, and posture of its handle; The visual recognition channel is used to extract the coordinates of two-dimensional skeletal key points of the user's joints in real time; S20 also includes: establishing a multi-source data spatiotemporal alignment and feature enhancement module, used to fuse and process data from IoT sensor unit clusters, dedicated shoulder joint trainers and visual recognition channels; The stability coefficient of the ball-swinging motion was calculated from the angular velocity data of the dedicated shoulder joint trainer. Perform cross-modal data verification by comparing the joint angles estimated by visual recognition with the joint angles calculated by the inertial measurement unit, and assigning information weights based on the difference.

6. The method of claim 1, wherein, The S30 also includes data from a dedicated shoulder joint trainer and visual recognition channel, from which the following new quantitative evaluation parameters are extracted: Load adaptability quantification metrics are used to assess the degree of match between the current training load and the user's capabilities; The direction-switching neural control index is used to quantify the user's control efficiency when switching action directions.

7. The method of claim 1, wherein, S40 also includes: the motion quality assessment and compensation pattern recognition model adds fusion analysis of load adaptability quantitative indicators and direction switching neural control index; The model adds a forward inference branch that performs rapid pre-analysis based on trainer data, and inputs the inference results as meta-information into the model's backbone network. The model's output layer now includes a device interaction credibility score, used to quantify the quality of the external data sources relied upon in this evaluation.

8. The method of claim 1, wherein, S50 also includes an integrated equipment interaction and adaptation analysis section in the visualization assessment report. This section contains dynamic curves showing the changes in load adaptability quantitative indicators and ball-swing motion stability coefficient over training time, as well as specific text comments on the ability to switch directions.

9. The method of claim 1, wherein, S60 also includes expanded prescription dynamic adjustment rules based on parameters obtained from a dedicated shoulder joint trainer and a visual recognition channel: If the user's continuous load adaptability quantification index shows that the load is too light and the action quality score is high, the system recommends upgrading the training load. If the user's direction-switching neural control index remains low, the system will automatically insert an adaptive neural control training sub-prescription in the next training cycle.

10. The method of claim 1 or 3, wherein S40 also includes: when the model determines that there are multiple compensation modes and their probabilities are close, a multi-stage joint inference strategy is used for arbitration, including calculating the distance between the input feature vector and the pre-stored typical compensation mode prototype feature vector, selecting the two candidate prototypes with the smallest distance; retrieving secondary features that are strongly correlated with the candidate prototypes and comparing them; and determining the candidate prototype with the higher comparison consistency as the final recognition result. And / or, when the smoothness of the movement trajectory increases abnormally over multiple consecutive movement cycles, fatigue state correlation analysis is performed. By analyzing the relationship between the slope of the smoothness deterioration and the slope of the decrease in the median frequency of the electromyographic signal, the cause of movement instability can be distinguished as either neural control disorder or muscle fatigue.

Images