Data analysis system for multi-joint exercise rehabilitation monitoring

Abstract

The present invention relates to the field of data analysis technology, and discloses a data analysis system for multi-joint motion rehabilitation monitoring. The system collects three-dimensional motion data, electromyographic activity signals and spatial position data of multiple joints in real time through a multimodal sensor array. The system includes data acquisition, preprocessing, spatiotemporal analysis and result output modules. The preprocessing module removes noise, normalizes and extracts features from the original data; the spatiotemporal analysis module constructs a monitoring model based on a graph spatiotemporal network, and uses an intelligent chip to analyze the spatiotemporal characteristics of multimodal data through dynamic joint relationship modeling. The result output module generates a visual report including motion coordination assessment, injury risk prediction and rehabilitation progress judgment, and feeds back to the rehabilitation equipment in real time. In addition, the system also has distributed storage and real-time feedback functions, which improves the accuracy and efficiency of rehabilitation monitoring and solves the problems of single data processing and inaccurate spatiotemporal feature capture in the existing technology.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a data analysis system for multi-joint motion rehabilitation monitoring. Background Art

[0002] In the field of rehabilitation medicine, multi-joint motion rehabilitation monitoring of patients is key to evaluating treatment outcomes and developing personalized rehabilitation plans. However, traditional monitoring methods often rely on manual observation and recording, which is not only inefficient but also easily affected by subjective factors. With the continuous development of sensor technology and data analysis technology, the use of multimodal sensor arrays to collect multi-joint motion data in real time and analyze it in combination with advanced algorithms has become an important means to improve the accuracy and efficiency of rehabilitation monitoring. However, most existing data analysis systems can only process single-modal data, lack the ability to comprehensively analyze multimodal data, and have difficulty accurately capturing the spatiotemporal characteristics of joint motion, thus limiting their application in multi-joint motion rehabilitation monitoring. Summary of the Invention

[0003] In view of the above-mentioned technical deficiencies, the purpose of the present invention is to provide a data analysis system for multi-joint motion rehabilitation monitoring to solve the problems of single data processing and inaccurate capture of spatiotemporal features in the prior art.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] In a first aspect, the present invention provides a data analysis system for multi-joint motion rehabilitation monitoring, the system comprising:

[0006] The data acquisition module collects three-dimensional motion data, myoelectric activity signals and spatial position data of multiple joints in real time through a multimodal sensor array;

[0007] The data preprocessing module is used to remove noise, normalize and extract multimodal features from the data collected by the data acquisition module;

[0008] The spatiotemporal analysis module builds a multi-joint motion rehabilitation monitoring model based on a graph spatiotemporal network. It uses an intelligent chip to perform spatiotemporal feature analysis on the multimodal data preprocessed by the data preprocessing module through dynamic joint relationship modeling.

[0009] The result output module generates a visual report containing movement coordination assessment, injury risk prediction and rehabilitation progress judgment based on the analysis results of the spatiotemporal analysis module, and feeds it back to the rehabilitation equipment terminal in real time.

[0010] Preferably, in a possible implementation manner of the first aspect, the multimodal sensor array includes:

[0011] Inertial measurement unit, used to obtain the three-dimensional acceleration, angular velocity and angular acceleration of the joint;

[0012] A surface electromyography sensor for collecting electromyographic activity signals at a sampling rate of not less than 1000 Hz;

[0013] Optical motion capture markers record the spatial position and motion trajectory of joints with sub-millimeter accuracy.

[0014] Preferably, in a possible implementation of the first aspect, the data preprocessing module includes:

[0015] The data cleaning unit uses median filtering and wavelet threshold denoising algorithm to eliminate high-frequency noise;

[0016] Normalization unit, which maps data to the [0,1] interval through minimum-maximum normalization;

[0017] The feature extraction unit combines principal component analysis with time-frequency domain analysis methods to extract the joint motion amplitude, frequency spectrum entropy and root mean square value of the electromyographic activity signal.

[0018] Preferably, in a possible implementation of the first aspect, the multi-joint movement rehabilitation monitoring model is constructed using a graph structure, and the graph structure construction process includes:

[0019] Model N joints of the human body as a set of graph nodes , node features are multimodal data ,in represents a real matrix, d is the feature dimension, and T is the time step;

[0020] Define the adjacency matrix , edge weight By physical connection strength Relevance to exercise Jointly determine:

[0021]

[0022] Where N is the number of joints, is an adjustable parameter, The correlation between joint motion trajectory and myoelectric activity signal is calculated through mutual information. and are the i-th and j-th nodes, and For the i and j Node features.

[0023] Preferably, in a possible implementation of the first aspect, the multi-joint motion rehabilitation monitoring model includes:

[0024] Input layer, receiving preprocessed multimodal time series data;

[0025] Multimodal data fusion layer, which uses a multi-head self-attention mechanism to fuse inertial data, electromyographic activity signals, and optical data;

[0026] The spatiotemporal graph convolutional network layer consists of K stacked ST-GCN modules, each of which performs spatial graph convolution. and time-expanded convolution, the expansion factor increases with the network depth Increasing, where is the activation function, is an adjacency matrix with self-connection, is the degree matrix, For the Layer spatial graph convolution, For the Layer trainable weight matrix;

[0027] Adaptive joint relationship modeling layer, which generates the potential joint association matrix through a differentiable graph learning algorithm;

[0028] The multi-task output layer consists of two branches, which output the injury risk probability and rehabilitation progress score respectively.

[0029] Preferably, in a possible implementation of the first aspect, the structure of the spatiotemporal graph convolutional network layer includes:

[0030] Dynamic spatial convolution module with learnable weight matrix , combined with the dynamic adjacency matrix Update node characteristics:

[0031]

[0032] Where ReLU represents the activation function, BN is the batch normalization layer, is the current node feature, is the updated feature, is the input feature dimension, is the output feature dimension;

[0033] Multi-scale temporal convolution module uses causal convolution to extract joint motion timing features ;

[0034] Residual connection structure, the output of each ST-GCN module is , where Dropout is the random loss ratio, It is the input of the ST-GCN module.

[0035] Preferably, in a possible implementation of the first aspect, extracting features from the preprocessed multimodal data includes:

[0036] Align inertial data, electromyographic activity signals, and optical data by time window;

[0037] Generate joint features through multimodal data fusion layer;

[0038] Extracting spatiotemporal features through the spatiotemporal graph convolutional network layer , the spatiotemporal features are current motion features.

[0039] Preferably, in a possible implementation manner of the first aspect, the content output to the result output module includes:

[0040] Joint movement synchronization index ,in and are the angles of the i-th joint and the j-th joint at time t respectively;

[0041] Injury risk probability , for The transpose of represents the weight vector of the damage risk prediction branch;

[0042] Rehabilitation Progress Score ,in is the dynamic time warping distance, is the maximum allowed difference threshold, Standard recovery characteristics.

[0043] Preferably, in a possible implementation manner of the first aspect, the result output module includes:

[0044] 3D motion trajectory simulation unit, reconstructing the 3D model of joint motion based on optical capture data ,in is the joint angle parameter;

[0045] The damage risk prediction unit outputs the damage risk probability according to the result of the damage risk branch;

[0046] The rehabilitation progress evaluation unit compares the DWT distance between the current motion characteristics and the standard rehabilitation characteristics and outputs a rehabilitation progress score.

[0047] Preferably, in a possible implementation manner of the first aspect, the system further includes:

[0048] Distributed storage unit, using Hadoop architecture to store raw data, pre-processed data and analysis results;

[0049] Real-time feedback unit, when joint movement deviation Greater than the preset deviation When the resistance parameter of the rehabilitation equipment is adjusted by the PID controller, the resistance parameter adjustment formula is:

[0050]

[0051] in is the angle error, is the current joint motion angle, To set a recovery trajectory, is the proportional gain coefficient, is the integral gain coefficient, represents the time integral of the historical deviation, represents the rate of change of deviation, is the differential gain coefficient.

[0052] The beneficial effect of this invention lies in achieving full-dimensional dynamic analysis and precise feedback for multi-joint motion rehabilitation monitoring through the fusion of heterogeneous data from a multimodal sensor array and deep feature mining using a graph spatiotemporal network. The system combines inertial sensing, electromyographic signals, and optical capture technologies to establish a cross-modal data acquisition and collaborative analysis mechanism, overcoming the limitations of traditional single-modality monitoring and comprehensively capturing the biomechanical characteristics of joint motion and muscle co-activation patterns.

[0053] A spatiotemporal network model based on a dynamic graph structure adaptively constructs functional coupling relationships between joints through dual weight optimization of physical connections and motion associations. This significantly improves the ability to analyze multi-joint collaborative features in complex motion patterns, effectively addressing the inadequacy of traditional linear models in modeling nonlinear spatiotemporal features. A spatiotemporal graph convolutional network, combined with multi-scale causal convolution and residual learning, deeply extracts cross-joint spatial correlation features of motion trajectories while ensuring temporal causality, enabling early identification and dynamic tracking of abnormal movement patterns during rehabilitation exercises. The application of intelligent chips and the coordinated design of a distributed storage architecture with a real-time feedback mechanism ensure efficient processing and low-latency transmission of massive multi-source data, providing immediate intervention and dynamic adjustment capabilities for rehabilitation training. The system's output visual assessment report integrates multi-dimensional indicators such as movement coordination, injury risk, and rehabilitation progress, and incorporates three-dimensional motion trajectory reconstruction technology to intuitively present the biomechanical state of joint motion, assisting clinical decision-making. Through adaptive graph learning and multi-task joint optimization, the system significantly enhances its dynamic adaptability to individualized rehabilitation progress, providing a reliable basis for the development and optimization of precise rehabilitation plans.

[0054] The present invention achieves breakthroughs in data fusion depth, spatiotemporal feature modeling accuracy, and real-time feedback, effectively solving core problems in existing technologies such as single data processing, inaccurate spatiotemporal feature capture, and delayed feedback, and promoting rehabilitation monitoring from static assessment to dynamic intelligent regulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 A structural diagram of a data analysis system for multi-joint motion rehabilitation monitoring is provided for this application.

[0057] Explanation of the accompanying figures: 1-data acquisition module, 2-data preprocessing module, 3-spatiotemporal analysis module, 4-result output module. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] Example 1: Figure 1 As shown, the present invention provides a data analysis system for multi-joint motion rehabilitation monitoring, the system comprising:

[0060] The data acquisition module 1 collects three-dimensional motion data, myoelectric activity signals and spatial position data during multi-joint motion in real time through a multi-modal sensor array.

[0061] Specifically, the multimodal sensor array adopts a heterogeneous sensor fusion architecture, including inertial measurement units, surface electromyography sensors, and optical motion capture markers.

[0062] The inertial measurement unit is used to obtain the three-dimensional acceleration, angular velocity and angular acceleration of the joint. In this embodiment, a 9-axis MEMS inertial sensor of model Xsens MTw Awinda is used, which integrates a three-axis accelerometer (range ), three-axis gyroscope (range ) and a triaxial magnetometer. These sensors are installed on the distal femur (anatomical landmark: 10 cm above the lateral epicondyle of the femur) and proximal tibia (lateral to the tibial tuberosity) in the lower limbs, and on the distal humerus (5 cm proximal to the olecranon process) and radial styloid process in the upper limbs. A sampling rate of 200 Hz is used to capture gait frequencies ≤ 15 Hz, meeting the Nyquist theorem. Data is transmitted in real time in quaternion format, and sensor data is fused using Kalman filtering.

[0063] Surface electromyography sensors are used to collect muscle electrical activity signals. Muscle electrical activity signals were collected at a high sampling rate. In this example, the Delsys Trigno system was used for EMG data acquisition, with an electrode spacing of 20 mm and a common-mode rejection ratio of ≥100 dB. Attachments were made to the rectus femoris muscle of the lower limb (1 / 3 of the line connecting the anterior superior iliac spine to the superior edge of the patella), the long head of the biceps femoris (midpoint of the line connecting the ischial tuberosity and the fibular head), and the erector spinae muscle of the trunk (2 cm lateral to the L3 spinous process). Data were collected at a sampling rate of 2000 Hz and a bandwidth of 20-450 Hz, and RMS values were calculated in real time.

[0064] Optical motion capture markers record the spatial position and motion trajectory of joints with submillimeter accuracy. In this example, an eight-camera infrared optical system, model Qualisys Oqus 7+, was used, with markers having a diameter of 14 mm. Markers were located on the anterior superior iliac spine, greater trochanter, lateral epicondyle, tibial tuberosity, and lateral malleolus of the lower limbs, and on the acromion, lateral epicondyle of the humerus, radial styloid process, and base of the third metacarpal bone of the upper limbs. Data acquisition was performed with a static positioning error of ≤0.2 mm, a dynamic tracking error of ≤0.5 mm, and a sampling rate of 120 Hz.

[0065] This embodiment also utilizes a multimodal data synchronization mechanism, including hardware synchronization, software alignment, and coordinate system 1. Hardware synchronization uses the PXIe-6674T timing module to generate a 10MHz reference clock, which triggers all sensors synchronously via the BNC interface. Software alignment utilizes the IEEE 1588 precision time protocol for data streams, ensuring time deviations of less than 5ms across multiple sources. Coordinate system 1 establishes a global coordinate system (with the origin at the projection of the first metatarsal head and the Z axis pointing vertically upward). A calibration plate is used to achieve a rigid transformation between the IMU's local coordinate system and the optical system's global coordinate system.

[0066] For the collected kinematic data, inertial data compensation is performed and the gravity component is eliminated using the quaternion differential method. The formula is: ,in is the original output three-axis acceleration, is the rotation matrix, which represents the rotation transformation of the IMU's current posture relative to the global coordinate system. is the gravitational acceleration vector, To eliminate the influence of gravity, three-axis acceleration was measured. Optical data was filtered using a fourth-order Butterworth low-pass filter. Motion artifacts were removed from the electromyographic activity signal using adaptive filtering based on IMU acceleration data, and muscle activation was detected using a dual-threshold method. Joint angles were calculated using Euler angles based on the human skeletal model recommended by the International Society for Sport (ISB). Foot pressure switches were used to mark the contact phase and swing phase to divide the gait cycle.

[0067] The data preprocessing module 2 is used to remove noise, normalize and extract multimodal features from the original data.

[0068] Specifically, the data preprocessing module 2 includes a data cleaning unit, a normalization unit and a feature extraction unit.

[0069] The data cleaning unit first performs median filtering on the three-dimensional acceleration and angular velocity data collected by the inertial measurement unit, using a sliding median filter with a window length of 5 sampling points (corresponding to a 25ms time window) to eliminate the spike noise caused by the sensor's transient interference. For the electromyographic activity signal collected by the surface electromyographic sensor, a three-layer decomposition and soft threshold denoising algorithm based on the Daubechies 4 wavelet basis is used to set the threshold (in is the estimated noise standard deviation, and N is the signal length), effectively suppressing high-frequency noise while preserving muscle activation characteristics. The optical motion capture data was processed with a fourth-order Butterworth low-pass filter with a cutoff frequency set to 6 Hz to eliminate high-frequency jitter noise while preserving key motion features during the gait cycle.

[0070] The normalization unit uses a minimum-maximum scaling method to independently map each channel of multimodal sensor data to the [0, 1] range. For the triaxial acceleration data from the inertial measurement unit, the extreme value reference is updated every 10 minutes to dynamically adapt to sensor drift. The normalization range of the electromyographic activity signal is truncated based on the baseline noise level at rest (±3 times the standard deviation) to avoid outlier interference. The normalization formula is: ,in To prevent division by zero errors, the normalized data is stored in IEEE 754 floating-point format to maintain accuracy.

[0071] The feature extraction unit combines principal component analysis (PCA) with a joint time-frequency domain analysis method. First, multimodal data (including IMU acceleration, sEMG RMS values, and optical joint angles) are temporally aligned and spatially registered. The data stream is segmented using a sliding window (window length 200ms, 50% overlap). Time-domain features are extracted from joint motion amplitude (RMS value of the acceleration signal) and integrated electromyographic activity (iEMG) values. Frequency-domain features are calculated using power spectral entropy and the signal power spectral density estimated using the Welch method, with a segment length of 256 points and Hanning window weighting. Principal component analysis, using a cumulative variance contribution criterion of ≥95%, reduces the high-dimensional feature vector to 12 dimensions to eliminate data redundancy. Feature fusion generates a joint feature matrix containing joint kinematics, dynamics, and muscle activation patterns.

[0072] Multimodal data synchronization is hardware-triggered using the 10MHz reference clock of the PXIe-6674T timing module. Software-level data stream alignment utilizes the IEEE 1588 Precision Time Protocol, keeping time deviation within 5ms. During coordinate normalization, the global coordinate system of the optical capture system is used as the reference (origin: projection of the first metatarsal head, Z-axis perpendicular to the ground). Singular Value Decomposition (SVD) is used to calculate the rigid transformation matrix of the IMU's local coordinate system, ensuring spatial consistency of motion parameters.

[0073] In this implementation, the data cleaning unit's median filtering and wavelet denoising algorithms are implemented in embedded C++ code, using the Intel MKL library to accelerate matrix operations. The normalization and feature extraction modules run in Python, utilizing the Scipy and Scikit-learn libraries for real-time processing. The processed feature data is encapsulated in JSON format and uploaded to a cloud-based analysis platform via a wireless transmission module.

[0074] The spatiotemporal analysis module 3 builds a multi-joint motion rehabilitation monitoring model based on the graph spatiotemporal network, and uses an intelligent chip to perform spatiotemporal feature analysis on the pre-processed multimodal data through dynamic joint relationship modeling.

[0075] Specifically, this embodiment uses the NVIDIA Jetson AGX Orin model chip to achieve efficient spatiotemporal modeling by dynamically allocating GPU (processing joint topology spatial convolution) and CPU (updating dynamic adjacency matrix and sensor scheduling) tasks.

[0076] The multi-joint exercise rehabilitation monitoring model is constructed using a graph structure. The graph structure construction process includes: modeling N joints of the human body as a set of graph nodes , node features are multimodal data , where d is the feature dimension and T is the time step. Define the adjacency matrix , edge weight By physical connection strength Relevance to exercise Jointly determine:

[0077]

[0078] in is an adjustable parameter, The correlation between joint motion trajectory and myoelectric activity signal is calculated through mutual information.

[0079] The multi-joint motion rehabilitation monitoring model includes an input layer, a multimodal data fusion layer, a spatiotemporal graph convolutional network layer, an adaptive joint relationship modeling layer, and a multi-task output layer.

[0080] The input layer receives preprocessed multimodal time series data , where N is the number of joints, d is the feature dimension, and T is the time window length. The pre-processed multimodal data input process includes: , myoelectric activity signals and optical data Align by time window; generate joint features through cross-modal fusion layer ; Extract spatiotemporal features through spatiotemporal graph convolutional network .

[0081] The multimodal data fusion layer uses a multi-head self-attention mechanism to fuse inertial data Q, electromyographic activity signal K, and optical data V:

[0082]

[0083] in, is the dimension scaling factor of the key vector.

[0084] In this embodiment, the multimodal data fusion layer is implemented using the MultiHeadCrossAttention class. Its core function is to fuse heterogeneous data from the inertial measurement unit (IMU), surface electromyography (sEMG), and optical capture systems. By defining a linear transformation layer (nn.Linear) for the query (Q), key (K), and value (V) vectors, the data from each modality is mapped to a unified feature space. The scaled dot product attention calculation is performed using PyTorch's built-in nn.MultiheadAttention component, where the query vector is generated from the inertial data, and the key and value vectors correspond to the electromyographic activity signal and optical data, respectively. The multi-head mechanism captures features from multiple semantic spaces by running four independent attention heads in parallel (n_heads=4), ultimately maintaining the 64-dimensional fused feature vector as the output.

[0085] The spatiotemporal graph convolutional network layer consists of K stacked ST-GCN modules, each of which performs spatial graph convolution. and time-expanded convolution, the expansion factor increases with the network depth Increasing, where is the activation function, is an adjacency matrix with self-connection, is the degree matrix, For the Layer spatial graph convolution, For the Layer trainable weight matrix.

[0086] The structure of the spatiotemporal graph convolutional network layer includes a dynamic spatial convolution module, a multi-scale temporal convolution module, and a residual connection structure.

[0087] Dynamic spatial convolution module with learnable weight matrix , combined with the dynamic adjacency matrix Update node characteristics: , where BN is the batch normalization layer.

[0088] Multi-scale temporal convolution module, using dilation factors Causal convolution to extract joint motion timing features: , k is the temporal convolution layer number, and d is the dilation factor of causal convolution.

[0089] Residual connection structure, the output of each ST-GCN module is , where BN is batch normalization, Dropout is the random loss ratio, It is the input of the ST-GCN module.

[0090] In this embodiment, the spatiotemporal graph convolution is implemented by the STGCNBlock class, which includes dual feature processing in the spatial and temporal domains. The spatial graph convolution uses the GCNConv graph convolution layer to propagate node features based on predefined physical connection edge indices (such as the lower limb hip-knee-ankle chain connection). Its adjacency matrix generates anatomically constrained binary connection relationships through the _get_physical_edges method. The temporal convolution part uses nn.Conv1d to construct causal dilated convolution, and the dilation factor increases exponentially with the number of network layers (the kth layer is ), padding (padding=dilation) is used to ensure temporal causality. Spatial convolution is performed frame by frame by looping over time steps (for t in range(T)), followed by dilated convolution along the temporal dimension. Finally, the output features are processed using batch normalization (nn.BatchNorm1d) and ReLU activation. The residual connection structure (x_temporal + x) preserves the original input information to mitigate the vanishing gradient problem.

[0091] Adaptive joint relationship modeling layer, generating potential joint association matrix through differentiable graph learning algorithm In this embodiment, the AdaptiveGraphLearner class implements the learning function of dynamic joint association weights. This module receives the node features output by the spatiotemporal graph convolution and calculates the potential association strength between joints through a fully connected network (nn.Sequential). Specifically, for each pair of joint nodes The concatenated feature vectors are fed into a two-layer MLP (32-dimensional hidden layer) and normalized to [0, 1] edge weights using a Sigmoid function. A double loop is used to traverse all joint pairs to generate an N×N weight matrix, which is then linearly blended with the predefined physical connection matrix in a 7:3 ratio to form the final adjacency matrix.

[0092] The multi-task output layer consists of two branches, which output the damage risk probability respectively. and rehabilitation progress score. Injury risk probability , is the weight vector of the injury risk prediction branch; rehabilitation progress score ,in is the dynamic time warping distance, is the maximum allowed difference threshold, Standard recovery characteristics.

[0093] The result output module 4 generates a visual report including movement coordination assessment, injury risk prediction and rehabilitation progress judgment, and feeds it back to the rehabilitation equipment terminal in real time.

[0094] Specifically, the content input by the result output module 4 includes the joint movement synchronization index ,in and are the angles of the i-th joint and the j-th joint at time t, respectively, where T is the length of the time window; the injury risk probability , is the weight vector of the injury risk prediction branch; rehabilitation progress score ,in is the dynamic time warping distance, is the maximum allowed difference threshold, Standard recovery characteristics.

[0095] The result output module 4 includes a three-dimensional motion trajectory simulation unit, an injury risk prediction unit, and a rehabilitation progress assessment unit.

[0096] The 3D motion trajectory simulation unit reconstructs the 3D model of human motion using the inverse kinematics algorithm based on the coordinate data of the marker points of the optical capture system and the joint angle parameters of the inertial measurement unit. A personalized skeletal model is constructed through the OpenSim biomechanics platform, and the spatial trajectories of key marker points such as the lateral epicondyle of the femur and the radial styloid process are input into the model. The weighted least squares method is used to solve the joint rotation angle and generate a 3D motion simulation animation. The simulation results are rendered using the Unity3D engine, which supports perspective rotation, motion speed adjustment (0.1-2 times the speed) and key frame marking functions. Abnormal motion segments (in this embodiment, synchronization scoring) are also included. ) is automatically highlighted in red. Using the touch interface, you can select a specific joint (such as the affected knee) to view a 6DOF motion parameter curve, including trends over time for flexion / extension, valgus / valgus, and rotation angles.

[0097] The damage risk prediction unit integrates the risk probability output by the spatiotemporal analysis module 3 Combined with real-time biomechanical parameters, a dynamic risk heat map is generated. When the system automatically triggers the three-level warning mechanism: Level 1 warning Display a yellow warning icon on the visual interface; Level 2 warning Send speed reduction instructions to rehabilitation equipment via CAN bus; three-level warning The training equipment is immediately paused and an audible and visual alarm sounds. Risk prediction results are correlated with biomechanical parameters (such as peak knee contact force and erector spinae activation delay). A Gaussian mixture model is used to cluster high-risk movement patterns, and risk areas (such as the anterior cruciate ligament stress concentration zone) are annotated in the 3D simulation model.

[0098] The rehabilitation progress assessment unit uses the dynamic time warping (DTW) algorithm to calculate the patient's motion characteristics Standard recovery curve Similarity, output standardized score The standard curve library contains 200 groups of healthy subjects' data, which are stored by age (±5 years), BMI (±3), and gender. The most appropriate reference group is automatically matched when the score is calculated. The evaluation results are visualized in the form of radar charts, and comparative analysis is conducted from six dimensions such as movement coordination, muscle balance, and joint mobility. When three consecutive evaluations are performed, the results are compared. When the improvement is less than 5%, the system automatically generates training program adjustment suggestions, including increasing the difficulty of virtual reality tasks or adjusting exoskeleton assistance parameters.

[0099] The visual report generation engine uses a template design to automatically compile the analysis results into a structured medical report. The report body includes:

[0100] Movement coordination assessment: Display the synchronization matrix of each joint pair Abnormal values (e.g., hip-knee synchrony < 0.5) were annotated with clinical explanations (e.g., "suggesting abnormal flexion-extension coordination pattern").

[0101] Injury risk distribution: Combine heat maps and probability curves to mark high-risk periods and their corresponding motion phases (e.g., heel strike peak at 63% of the gait cycle);

[0102] Rehabilitation Progress Tracking: Display Time-varying curves are overlaid with clinical scale results such as the Fugl-Meyer score for dual verification. Report output supports the PDF / A medical archive format and the HL7 FHIR standard interface, and is stored on the medical cloud platform using 256-bit AES encryption to ensure data privacy compliance with GDPR regulations.

[0103] Example 2: The present invention provides a data analysis system for multi-joint exercise rehabilitation monitoring, the system also includes a distributed storage unit and a real-time feedback unit

[0104] Specifically, the distributed storage unit utilizes a hybrid architecture based on the Hadoop ecosystem to achieve efficient storage and parallel processing of multimodal data. Raw sensor data (including IMU raw acceleration, sEMG time-domain signals, and optical marker coordinates) is stored in the HDFS distributed file system in Apache Avro binary format. The data block size is set to 128MB to optimize MapReduce processing performance, and the replication factor is configured to 3 to ensure fault tolerance. Preprocessed feature data and spatiotemporal analysis results are stored in an HBase columnar database, with a row key of "patient ID_timestamp_joint number" to support millisecond-level time range queries. To meet the needs of real-time clinical analysis, the Alluxio in-memory acceleration layer caches frequently accessed data for the past 24 hours, with a LRU strategy managing cache replacement. For data security, AES-256 encryption is used for data at rest, combined with the Kerberos protocol for fine-grained access control, complying with HIPAA and GDPR healthcare data privacy regulations. The data lifecycle management module automatically migrates raw data older than 180 days to Glacier low-cost storage, while maintaining permanent access to feature data and assessment results.

[0105] The real-time feedback unit realizes the adaptive resistance adjustment of the rehabilitation equipment through the fuzzy PID controller. Its core algorithm is ,in is the angle error, is the current joint motion angle, To set a recovery trajectory, is the proportional gain coefficient, which determines the response strength of the system to the current deviation. is the integral gain coefficient, which determines the system's ability to eliminate steady-state errors. represents the time integral of the historical deviation, Indicates the deviation change rate, reflecting the joint movement trend, is the differential gain coefficient, which determines the system's ability to suppress motion mutations.

[0106] In this embodiment, the gain coefficient is dynamically adjusted according to the patient's motor ability: when the rehabilitation progress score When the conservative control strategy is adopted ( , , );when When , switch to the aggressive strategy ( , , The angle deviation data is smoothed by Kalman filtering, and the sampling interval is 5ms to ensure real-time control.

[0107] The resistance adjustment of the rehabilitation equipment is realized through the EtherCAT industrial bus, and the control instruction format complies with the CiA 402 motion control specification. and lasts for 200ms), the system implements three-level braking strategy:

[0108] Primary adjustment: Send the 0x6041 status word command to linearly increase the device resistance torque to Calculated value;

[0109] Intermediate intervention: If the deviation is not corrected within 500ms, the control mode switch 0x6040 is triggered to enable position-force hybrid control;

[0110] Emergency brake: When When the device is in the state of being stopped, the device joints are immediately locked through the 0x6042 fast stop instruction.

[0111] The system integration and communication architecture uses the OPC UA protocol to implement data interaction between modules and defines the following information model:

[0112] Device node: contains real-time parameters such as joint angle, resistance torque, control mode, etc.

[0113] Patient node: stores historical assessment reports, personalized control parameters and safety thresholds;

[0114] Alarm node: records the timestamp, triggering cause and device response log of the braking event.

[0115] Real-time data streams are distributed through Apache Kafka message queues, and the topic partitioning strategy is allocated by patient ID hash to ensure that sensor data, control instructions and feedback signals of the same patient maintain sequential consistency.

[0116] The safety monitoring mechanism incorporates a dual-redundancy design: the primary controller uses an NI cRIO-9045 real-time processor running VxWorks, while the backup controller is based on a Raspberry Pi 4. Heartbeat packets ensure failover (with a failover latency of <50ms). All control commands are digitally signed (using the ECDSA algorithm) to prevent unauthorized parameter modifications. A device self-test module performs full system diagnostics every 30 minutes, checking sensor communication latency, storage unit CRC checksums, and actuator impedance tests.

[0117] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A data analysis system for multi-joint exercise rehabilitation monitoring, characterized in that: The system comprises: The data acquisition module collects three-dimensional motion data, myoelectric activity signals and spatial position data of multiple joints in real time through a multimodal sensor array; The data preprocessing module is used to remove noise, normalize and extract multimodal features from the data collected by the data acquisition module; The spatiotemporal analysis module builds a multi-joint motion rehabilitation monitoring model based on a graph spatiotemporal network. It uses an intelligent chip to perform spatiotemporal feature analysis on the multimodal data preprocessed by the data preprocessing module through dynamic joint relationship modeling. The multi-joint movement rehabilitation monitoring model is constructed using a graph structure, and the graph structure construction process includes: Model N joints of the human body as a set of graph nodes , node features are multimodal data ,in represents a real matrix, d is the feature dimension, and T is the time step; Define the adjacency matrix , edge weight By physical connection strength Relevance to exercise Jointly determine: Where N is the number of joints, is an adjustable parameter, The correlation between joint motion trajectory and myoelectric activity signal is calculated through mutual information. and are the i-th and j-th nodes, and For the i and j Node features; The multi-joint motion rehabilitation monitoring model includes: Input layer, receiving preprocessed multimodal time series data; Multimodal data fusion layer, which uses a multi-head self-attention mechanism to fuse inertial data, electromyographic activity signals, and optical data; The spatiotemporal graph convolutional network layer consists of K stacked ST-GCN modules, each of which performs spatial graph convolution. and time-expanded convolution, the expansion factor increases with the network depth Incremental, where is the activation function, is an adjacency matrix with self-connection, is the degree matrix, For the Layer spatial graph convolution, For the Layer trainable weight matrix; Adaptive joint relationship modeling layer, which generates the potential joint association matrix through a differentiable graph learning algorithm; The multi-task output layer consists of two branches, which output the injury risk probability and rehabilitation progress score respectively; The structure of the spatiotemporal graph convolutional network layer includes: Dynamic spatial convolution module with learnable weight matrix , combined with the dynamic adjacency matrix Update node characteristics: Where ReLU represents the activation function, BN is the batch normalization layer, is the current node feature, is the updated feature, is the input feature dimension, is the output feature dimension; Multi-scale temporal convolution module uses causal convolution to extract joint motion timing features ; Residual connection structure, the output of each ST-GCN module is , where Dropout is the random loss ratio, It is the input of the ST-GCN module; The result output module generates a visual report containing movement coordination assessment, injury risk prediction and rehabilitation progress judgment based on the analysis results of the spatiotemporal analysis module, and feeds it back to the rehabilitation equipment terminal in real time.

2. The data analysis system according to claim 1, wherein: The multimodal sensor array comprises: Inertial measurement unit, used to obtain the three-dimensional acceleration, angular velocity and angular acceleration of the joint; A surface electromyography sensor for collecting electromyographic activity signals at a sampling rate of not less than 1000 Hz; Optical motion capture markers record the spatial position and motion trajectory of joints with sub-millimeter accuracy.

3. The data analysis system according to claim 1, wherein: The data preprocessing module includes: The data cleaning unit uses median filtering and wavelet threshold denoising algorithm to eliminate high-frequency noise; Normalization unit, which maps data to the [0,1] interval through minimum-maximum normalization; The feature extraction unit combines principal component analysis with time-frequency domain analysis methods to extract the joint motion amplitude, frequency spectrum entropy and root mean square value of the electromyographic activity signal.

4. The data analysis system according to claim 1, wherein: Feature extraction of multimodal data after preprocessing includes: Align inertial data, electromyographic activity signals, and optical data by time window; Generate joint features through multimodal data fusion layer; Extracting spatiotemporal features through the spatiotemporal graph convolutional network layer , the spatiotemporal features are current motion features.

5. The data analysis system according to claim 4, wherein: The output to the result output module includes: Joint movement synchronization index ,in and are the angles of the i-th joint and the j-th joint at time t respectively; Injury risk probability , for The transpose of represents the weight vector of the damage risk prediction branch; Rehabilitation Progress Score ,in is the dynamic time warping distance, is the maximum allowed difference threshold, Standard recovery characteristics.

6. The data analysis system according to claim 5, wherein: The result output module includes: 3D motion trajectory simulation unit, reconstructing the 3D model of joint motion based on optical capture data ,in is the joint angle parameter; The damage risk prediction unit outputs the damage risk probability according to the result of the damage risk branch; The rehabilitation progress evaluation unit compares the DWT distance between the current motion characteristics and the standard rehabilitation characteristics and outputs a rehabilitation progress score.

7. The data analysis system according to claim 1, wherein: The system further comprises: Distributed storage unit, using Hadoop architecture to store raw data, pre-processed data and analysis results; Real-time feedback unit, when joint movement deviation Greater than the preset deviation When the resistance parameter of the rehabilitation equipment is adjusted by the PID controller, the resistance parameter adjustment formula is: in is the angle error, is the current joint motion angle, To set a recovery trajectory, is the proportional gain coefficient, is the integral gain coefficient, represents the time integral of the historical deviation, represents the rate of change of deviation, is the differential gain coefficient.

Images