Arm motion monitoring and health recognition method and system based on wearable sensor

Abstract

The invention discloses an arm movement monitoring and health recognition method and system based on a wearable sensor. The method comprises the steps that original multi-axis signals in the arm movement process of a user are collected; carrying out analysis, calibration and low-pass filtering on the original multi-axis signal, and calculating a three-dimensional attitude angle and a motion track of an arm in real time by adopting an improved Kalman filtering algorithm; segmenting the continuous motion data into a standardized motion sequence by adopting a hybrid motion segmentation algorithm; performing time-frequency domain analysis on the standardized action sequence to obtain a time domain feature and a frequency domain feature; extracting attitude track features; the time domain features, the frequency domain features and the attitude track features are combined into a high-dimensional feature matrix, and low-dimensional feature vectors are generated after dimension reduction; and inputting the low-dimensional feature vector into a pre-trained health assessment model, and outputting an arm motion function score, abnormal type early warning and rehabilitation guidance. The system can realize remote guidance of doctors and household rehabilitation monitoring of patients, and has the advantages of high precision, low cost and high real-time performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of health monitoring, and particularly relates to a method and system for arm movement monitoring and health identification based on wearable sensors. Background Art

[0002] As the core part of the human body for performing complex operations, the integrity of the arm's motor function directly affects the quality of daily life and professional ability. In recent years, problems such as sports injuries (such as tennis elbow, rotator cuff tear), postoperative rehabilitation needs (such as joint stiffness after fracture), and optimization of sports performance (such as standardization of athletes' movements) have become increasingly prominent. Traditional health monitoring methods usually rely on regular examinations in hospitals and clinics, often lacking sufficient real-time performance and convenience, and being costly, making it difficult to meet the needs of different groups for continuous and personalized health management.

[0003] In recent years, with the rapid development of sensing technology, microelectronics technology, wireless communication technology, and data processing technology, it has become possible to use wearable devices to monitor the movement state of the arm in real time. This method can provide movement data of the arm in a natural state, providing a new means for the health assessment and rehabilitation treatment of the arm.

[0004] However, although there are already some wearable devices for health monitoring on the market at present, most devices are limited by the insufficient accuracy of single sensors and the lack of multi-sensor fusion algorithms, resulting in large errors in attitude angle calculation; and data processing mostly relies on local terminals, making it difficult to achieve real-time cloud analysis and remote medical collaboration. Summary of the Invention

[0005] Object of the Invention: The technical problem to be solved by the present invention is to provide a method and system for arm movement monitoring and health identification based on wearable sensors in view of the deficiencies of the prior art.

[0006] To solve the above technical problem, in the first aspect, a method for arm movement monitoring and health identification based on wearable sensors is disclosed, including:

[0007] Step 1, multi-axis signal acquisition: acquiring the original multi-axis signals during the movement of the user's arm, where the multi-axis signals include three-axis acceleration, three-axis angular velocity, and three-axis magnetic field signals;

[0008] Step 2, signal preprocessing and sensor fusion: parsing, calibrating, and low-pass filtering the original multi-axis signals, and for the low-pass filtered multi-axis signals, using an improved Kalman filtering algorithm with adaptive dynamic noise covariance to fuse the data of the accelerometer, gyroscope, and magnetometer, and calculating the three-dimensional attitude angles (pitch angle, roll angle, yaw angle) and movement trajectory of the arm in real time;

[0009] Step 3, Action Cycle Segmentation: Based on the variation law of the attitude angle, the hybrid action segmentation algorithm is used to segment the time series data of the three-dimensional attitude angle and the motion trajectory of the arm into standardized action sequences, and each sequence corresponds to a single complete action;

[0010] Step 4, Multi-dimensional Feature Extraction: Perform time-domain analysis (such as maximum value, mean value, variance, etc.) and frequency-domain analysis (such as FFT energy spectrum, main frequency component, etc.) on the standardized action sequence to obtain time-domain features and frequency-domain features; Extract attitude trajectory features from the three-dimensional attitude angle and the motion trajectory, and the attitude trajectory features include range of motion (ROM), severity of the motion state (Jerk value), trajectory symmetry, activity angle change rate, multi-joint motion phase difference, asymmetric fluctuation of the acceleration signal, low-frequency vibration in the stationary state, etc.

[0011] Step 5, Feature Fusion and Dimensionality Reduction: Combine the time-domain features, frequency-domain features and attitude trajectory features into a high-dimensional feature matrix, and generate a low-dimensional feature vector through principal component analysis (PCA, Principal Component Analysis);

[0012] Step 6, Health Identification: Input the low-dimensional feature vector into the pre-trained health assessment model, and output the arm movement function score, abnormal type warning and rehabilitation guidance.

[0013] Furthermore, the improved Kalman filtering algorithm with adaptive dynamic noise covariance in Step 2 includes:

[0014] When the arm moves violently, dynamically adjust the process noise covariance Q k = Q0·(1 + α·Jerk);

[0015] Among them, Q0 represents the static noise covariance, which is obtained by calibrating the wearable sensor; α is the dynamic adjustment coefficient, and the value range is [0,1], which is calibrated through experiments; Jerk represents the derivative of the acceleration, which characterizes the "urgency" of the motion. The more violently the arm moves, the larger the Jerk value, and the more significant the sensor noise;

[0016] When there is magnetic interference, dynamically adjust the magnetometer observation noise covariance R mag = R mag.base ·(1 + β·σ M );

[0017] Among them, R mag.base represents the magnetometer base noise covariance, β represents the magnetic interference sensitivity coefficient, which is calibrated through experiments, and σ M represents the standard deviation of the magnetic field intensity; when σ M is greater than the set threshold, it is determined that there is magnetic interference.

[0018] The core problem of traditional Kalman filtering in arm movement monitoring stems from its fixed noise covariance assumption. Kalman filtering describes the system dynamics and sensor error characteristics through the process noise covariance matrix Q and the observation noise covariance matrix R. However, arm movement is highly dynamic:

[0019] During strenuous exercise: the noise of the accelerometer and gyroscope increases significantly (such as muscle tremors and sensor jitter);

[0020] When there is a sudden change in movement (such as a rapid arm swing): the fixed Q cannot adapt to the instantaneous noise change, resulting in the accumulation of prediction errors and an increase in the deviation of the pitch angle and roll angle;

[0021] Magnetic interference environment: ferromagnetic substances in the environment (such as metal instruments) will interfere with the magnetometer data. If the observation weight is not dynamically adjusted, a yaw angle error will be introduced.

[0022] In step 2, the Jerk value is first introduced into the noise covariance adjustment of Kalman filtering, which is more in line with the characteristics of human movement; the intensity of the arm movement state is identified through the Jerk value, and the noise dynamics are adjusted in real time according to the Jerk value, adjusting Q k , increasing the weight of the process noise during strenuous exercise to respond faster to sensor measurement values, avoiding the accumulation of prediction errors, and thus significantly reducing the attitude angle error during strenuous exercise. The magnetic interference is detected in real time through the magnetic field standard deviation, avoiding relying on fixed thresholds or offline calibration; when magnetic interference appears in the environment, the corresponding gain component K of the magnetometer mag decreases, and its weight is reduced; the estimation of the attitude angle will rely more on the fusion result of the accelerometer and gyroscope rather than the interfered magnetometer data, thereby reducing its impact on the attitude angle estimation. The above algorithm has low complexity and is suitable for embedded deployment.

[0023] Furthermore, step 3 includes:

[0024] Step 3-1, rough segmentation: dynamically set the sliding window length, detect the peak value of the pitch angle change rate within the sliding window, and preliminarily divide the action interval; the sliding window length L = L base +η·var(θ pitch ), where L base represents the basic window length, covering regular actions; var(θ pitch ) represents the variance of the pitch angle within the current window, reflecting the intensity of the action amplitude change; η represents the adjustment coefficient calibrated through training data, used to amplify and reduce the influence of the pitch angle variance;

[0025] Step 3-2, Fine Segmentation: Only match the preset action template within the initially divided action interval to obtain candidate boundary points; use the DBSCAN algorithm to eliminate isolated noise points among the matched candidate boundary points based on density clustering, and the final boundary is determined by the center point of the maximum density cluster to enhance robustness.

[0026] Fixed-threshold segmentation is sensitive to the motion speed and is prone to misjudging action boundaries. For example, the signal amplitudes generated by fast waving and slow waving are quite different, and the fixed threshold may lead to boundary misjudgments (such as missed detections or over-segmentation). Although the traditional DTW (Dynamic Time Warping) can effectively match the similarity of time series, its computational complexity is high and it is difficult to meet the requirements of real-time applications. In Step 3-1, using the local extreme points of the pitch angle change rate as action interval markers can greatly reduce the subsequent calculation range and avoid processing the entire sequence; if the action amplitude changes violently (such as fast waving), the sliding window automatically expands to capture the complete action; otherwise, the window shrinks to reduce redundant calculations, and the dynamic window adapts to speed changes to reduce missed detections. In Step 3-2, only match the preset template (such as the standard arm waving action) within the roughly segmented interval, which reduces the computational complexity and improves the boundary accuracy.

[0027] Furthermore, in Step 6, the pre-trained health assessment model is a machine learning model, and the training method of the machine learning model includes:

[0028] Step 6.1, the user performs different arm movements, including grasping, lifting, waving, and rotating actions. Execute Steps 1 - 5 to obtain an arm movement dataset, divide the arm movement dataset into a training set and a test set through five-fold cross-validation, and optimize the hyperparameters of the machine learning model using grid search;

[0029] Step 6.2, on the premise of the patient's informed consent, the doctor clinically evaluates the arm motor function according to the Fugl-Meyer scoring standard and divides the grades;

[0030] Step 6.3, compare the performances of the support vector machine (SVM, Support Vector Machine), random forest (RandomForest), and convolutional neural network (CNN, Convolutional Neural Network), and select the model with a five-fold cross-validation AUC mean ≥ 0.85 and a standard deviation ≤ 0.05 for deployment.

[0031] Furthermore, Step 6 includes:

[0032] Input the low-dimensional feature vector into the model selected in Step 6.3, and output the arm motor function score, abnormal type warning, and rehabilitation guidance;

[0033] Real-time monitor the arm movement data, with the delay from data acquisition to the output of abnormal type warning ≤ 200 ms, and realize the recognition of abnormal arm movement types or motor functions;

[0034] Verify the reliability of the model through the cross-validation of doctor evaluation and machine learning results, calculate the consistency coefficient based on the independent test set. If the consistency coefficient Kappa ≥ 0.80, it is determined that the doctor's diagnosis result is consistent with the model prediction.

[0035] Further, when the model selected in step 6.3 is a support vector machine, its kernel function is a radial basis function kernel (RBF, Radial Basis Function kernel), and the regularization parameter C ∈ [0.1, 10];

[0036] When the model selected in step 6.3 is a random forest, the number of decision trees ≥ 100, and the maximum depth ≤ 15;

[0037] When the model selected in step 6.3 is a convolutional neural network, the network structure includes ≤ 3 convolutional layers and is deployed on an edge computing device.

[0038] In a second aspect, a wearable sensor-based arm movement monitoring and health recognition system is disclosed, which is characterized by including a signal acquisition module, a data parsing module, a data processing module, a feature extraction and fusion module, a health assessment module, and a model training module.

[0039] The signal acquisition module is used to acquire the original multi-axis signals during the arm movement of the user, and the original multi-axis signals include three-axis acceleration, three-axis angular velocity, and three-axis magnetic field signals;

[0040] The data parsing module is used to parse, calibrate, and low-pass filter the original multi-axis signals;

[0041] The data processing module is used to, for the low-pass filtered multi-axis signals, adopt an improved Kalman filtering algorithm to fuse the data of the accelerometer, gyroscope, and magnetometer, and calculate the three-dimensional attitude angle and movement trajectory of the arm in real time; based on the change law of the attitude angle, adopt a hybrid action segmentation algorithm to segment the time series data of the three-dimensional attitude angle and movement trajectory of the arm into standardized action sequences, and each sequence corresponds to a single complete action;

[0042] The feature extraction and fusion module is used to perform time-domain analysis and frequency-domain analysis on the standardized action sequences to obtain time-domain features and frequency-domain features; extract attitude trajectory features from the three-dimensional attitude angle and movement trajectory; merge the time-domain features, frequency-domain features, and attitude trajectory features into a high-dimensional feature matrix, and generate a low-dimensional feature vector through principal component analysis;

[0043] The health assessment module is used to input the low-dimensional feature vector into a pre-trained health assessment model and output the arm movement function score, abnormal type warning, and rehabilitation guidance;

[0044] The model training module is used to train the health assessment model according to the training data set to generate a pre-trained health assessment model; the health assessment model is a machine learning model.

[0045] In a third aspect, an arm movement monitoring and health identification device based on a wearable sensor is disclosed, which is characterized by including a main control chip, a power supply module, a sensor module, a Bluetooth module, a wearable arm guard unit, a cloud server, and a mobile terminal. The main control chip is connected to the sensor module and the Bluetooth module and encapsulated into the wearable arm guard unit;

[0046] The main control chip is used to control the sensor module to collect signals, perform signal preprocessing (calibration, filtering), execute Kalman filter attitude solution, and control the Bluetooth module to achieve real-time data transmission;

[0047] The power supply module is used to supply power to the main control chip, the sensor module, and the Bluetooth module;

[0048] The sensor module is used to collect the original multi-axis signals during the arm movement. The original multi-axis signals include three-axis acceleration, three-axis angular velocity, and three-axis magnetic field signals;

[0049] The Bluetooth module is used to wirelessly transmit the data processed by the main control chip to the mobile terminal,

[0050] The mobile terminal is used to control the start and stop of the sensor module, data encoding and decoding, and data interaction with the cloud server;

[0051] The cloud server is used to pre-deploy a pre-trained health assessment model, parse and decode the data of the request sent by the mobile terminal, process the data using the pre-trained health assessment model, generate the arm movement function score, abnormal type warning, and rehabilitation guidance, and encode and return them to the mobile terminal; the pre-trained health assessment model is a machine learning model.

[0052] Further, the sensor module is a nine-axis inertial sensor, including a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer.

[0053] Further, the wireless communication in the Bluetooth module uses the low-power Bluetooth protocol (BLE, Bluetooth Low Energy), and the wearable arm guard unit is a medical arm guard or a sports arm guard.

[0054] Compared with the existing technology, the beneficial effects of the present invention are:

[0055] (1) Precision improvement: Multi-sensor fusion. By combining a nine-axis sensor with an improved Kalman filter algorithm with adaptive dynamic noise covariance, the problem of insufficient accuracy of a single sensor is solved, and the three-dimensional attitude angle error < 2.5°.

[0056] (2) Real-time performance improvement: Data is transmitted to the cloud server for processing to achieve remote model training and health assessment, with a delay ≤ 200 ms.

[0057] (3) Cost reduction: The device adopts a modular design, supports home use by patients, and reduces medical costs.

[0058] This application can overcome the limitations of the prior art and achieve more accurate, comfortable, and real-time health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The following further specifically describes the present invention in conjunction with the drawings and specific embodiments, and the above and / or other advantages of the present invention will become clearer.

[0060] Figure 1 It is a flowchart of a method for arm movement monitoring and health recognition based on a wearable sensor provided by an embodiment of the present invention;

[0061] Figure 2 It is a flowchart of a model training method in a method for arm movement monitoring and health recognition based on a wearable sensor provided by an embodiment of the present invention;

[0062] Figure 3 It is a detailed flowchart of a method for arm movement monitoring and health recognition based on a wearable sensor provided by an embodiment of the present invention;

[0063] Figure 4 It is an architecture diagram of a system for arm movement monitoring and health recognition based on a wearable sensor provided by an embodiment of the present invention;

[0064] Figure 5 It is a schematic diagram of a device for arm movement monitoring and health recognition based on a wearable sensor provided by an embodiment of the present invention; BRIEF DESCRIPTION OF THE DRAWINGS:

[0066] 1. Main control chip; 2. Power supply module; 3. Sensor module; 4. Bluetooth module; 5. Wearable arm unit; 6. Cloud server; 7. Mobile terminal. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] The following further describes the specific embodiments of the present invention in conjunction with the drawings and specific examples:

[0068] The present application provides an arm motion monitoring and health identification method and system based on wearable sensors, which can be applied to application scenarios such as sports injury rehabilitation, postoperative functional assessment, and athlete movement optimization.

[0069] Embodiment 1:

[0070] like Figure 1 and Figure 3 As shown, the first embodiment of the present application discloses an arm motion monitoring and health identification method based on a wearable sensor, comprising:

[0071] Step 1, multi-axis signal acquisition:

[0072] The nine-axis inertial sensor (including a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer) collects multi-axis motion signals of the user in real time during grasping, lifting, swinging, and rotating arm movements, including three-axis acceleration, three-axis angular velocity, and three-axis magnetic field signals;

[0073] Step 2: Signal preprocessing, sensor fusion and attitude calculation:

[0074] The original multi-axis signal is analyzed, calibrated and low-pass filtered with a cutoff frequency of 20Hz to eliminate high-frequency noise such as electromyographic signals. For the multi-axis signal after low-pass filtering, the improved Kalman filter algorithm based on adaptive dynamic noise covariance is used to fuse the multi-axis motion signal (fusing accelerometer, gyroscope and magnetometer data), calculate the three-dimensional attitude angle and motion trajectory of the arm in real time, eliminate the noise interference of a single sensor, and improve the accuracy of attitude angle calculation;

[0075] The improved Kalman filter algorithm of the adaptive dynamic noise covariance includes:

[0076] When the arm moves violently, the process noise covariance Q is dynamically adjusted k =Q0·(1+α·Jerk); where Q0 represents the static noise covariance, which is obtained through wearable sensor calibration. In the specific implementation process, data can be collected in a stationary state to calculate Q0; α is the dynamic adjustment coefficient, with a value range of [0,1], which is calibrated through experiments. In the specific implementation process, users can perform arm swings at different speeds, record the relationship between the Jerk value and the attitude angle error, and fit the optimal α through the least squares method; Jerk represents the derivative of acceleration, which characterizes the "rush" of movement. The more intense the arm movement, the larger the Jerk value, and the more significant the sensor noise. The calculation formula is as follows:

[0077]

[0078] Among them, a x , a y , a zis the acceleration value of three axes, and the derivative is calculated by the difference method.

[0079] Kalman gain K k The calculation formula is:

[0080]

[0081] Among them, P pred is the prior error covariance matrix, P pred =FPF T +Q k , F is the state transfer matrix, P is the posterior error covariance matrix of the previous moment, H is the observation matrix, and R is the observation noise covariance matrix. In the case of intense motion, Q can be increased by adjusting α. k , so that P pred Increase, under the premise that the numerator increase is greater than the denominator increase, the Kalman gain K k As it increases, the algorithm relies more on the observed values ​​(sensor measurements) and reduces the trust in the predicted values, thereby suppressing the accumulation of prediction errors during intense movement and improving the accuracy of attitude angle estimation.

[0082] Dynamically adjust the magnetometer observation noise covariance R when magnetic interference exists mag =R mag.base ·(1+β·σ M );

[0083] Among them, R mag.base represents the basic noise covariance of the magnetometer, β represents the magnetic interference sensitivity coefficient, which can be taken as 2.0; σ M represents the standard deviation of magnetic field intensity; when σ M When it is greater than the set threshold, it is determined that magnetic interference exists. M The calculation formula is as follows:

[0084]

[0085] Among them, M i is the three-axis magnetic field intensity vector of the i-th sampling point (i is the sampling point index in the sliding window, i = 1, 2, ..., N), μ M is the mean value of the magnetic field intensity vector of all sampling points in the sliding window, N is the sliding window length of magnetic interference detection, and N=50 can be taken, that is, covering 0.5 seconds of data, and the sampling frequency is 100 Hz.

[0086] Magnetic interference time R mag Increasing will reduce the weight of the magnetometer in the Kalman gain and reduce its influence on the attitude angle estimation. k The calculation formula of the observation noise covariance matrix R is as follows:

[0087]

[0088] When magnetic interference appears in the environment, σ M increases, thus causing R mag to increase; when R mag increases, the acceleration observation noise covariance R acc and the angular velocity observation noise covariance R gyro remain unchanged. The system considers that the observation data of the magnetometer has greater noise, and the corresponding gain component K mag of the magnetometer decreases, and its weight is reduced; the estimation of the attitude angle will rely more on the fusion result of the accelerometer and the gyroscope, rather than the disturbed magnetometer data.

[0089] When the wearable sensor uses the STMicroelectronics LSM9DS1 nine-axis inertial sensor with a sampling frequency of 100 Hz, multi-axis signals of two actions, namely, quickly swinging the arm horizontally (simulating strenuous exercise) and slowly drawing a circle (simulating regular exercise), are collected. The three-dimensional attitude angles calculated by comparing the traditional Kalman filter and the improved Kalman filter algorithm with adaptive dynamic noise covariance in this embodiment are shown in Table 1 for the error.

[0090] Table 1 Three-dimensional attitude angle error

[0091]

[0092] As can be seen from Table 1: 1. Dynamic Q adjustment: significantly reduces the pitch angle and roll angle errors during strenuous exercise; 2. Magnetic interference compensation: reduces the heading angle error by 40%, especially when close to metal objects; 3. Real-time performance: the delay only increases by 2 ms, fully meeting the real-time monitoring requirements (≤200 ms).

[0093] Step 3, action cycle segmentation:

[0094] Based on the attitude angle change rule (such as the periodic fluctuation of the pitch angle), the time series data of the three-dimensional attitude angle and the motion trajectory of the arm are segmented into standardized action sequences using the hybrid action segmentation algorithm. Each sequence corresponds to a single complete action (such as a wave or a grasp), including:

[0095] Step 3-1, rough segmentation: Dynamically set the sliding window length, and detect the peak value of the pitch angle change rate within the sliding window to initially divide the action interval (such as "the start of arm waving" and "the end of arm waving"); the sliding window length L = L base +η·var(θ pitch ), where L base represents the basic window length, covering regular actions; var(θ pitch) represents the pitch angle variance within the current window, reflecting the intensity of the change in the action amplitude; η represents the adjustment coefficient calibrated through training data, which is used to amplify and reduce the influence of the pitch angle variance;

[0096] Step 3-2, fine segmentation:

[0097] Lightweight DTW: Only match the preset action template within the initially divided action interval to obtain candidate boundary points;

[0098] DBSCAN optimization of the boundary: Use the DBSCAN algorithm to cluster based on density to eliminate isolated noise points among the candidate boundary points matched by DTW. The final boundary is determined by the center point of the maximum density cluster, improving the robustness.

[0099] Table 2 shows the effects of obtaining the standardized action sequence by using the fixed threshold segmentation and the rough segmentation + fine segmentation method of this embodiment. It can be seen from Table 2 that the rough segmentation + fine segmentation method of this embodiment makes: 1. The segmentation accuracy increases: DBSCAN effectively filters out noise, and DTW template matching improves the boundary accuracy. 2. The recall rate increases: The dynamic window adapts to speed changes and reduces missed detections. 3. The real-time performance is enhanced: Lightweight DTW + dynamic window optimization improves the calculation efficiency.

[0100] Table 2 Segmentation Effects

[0101]

[0102] Step 4, multi-dimensional feature extraction:

[0103] Perform time-domain analysis on the standardized action sequence to extract multi-dimensional features including the maximum value, minimum value, mean value, variance, kurtosis, skewness, coefficient of variation, and speed indicators (such as average angular velocity, peak velocity); perform frequency-domain analysis on the standardized action sequence to obtain the FFT energy spectrum, main frequency component, and the proportion of specific frequency energy.

[0104] Extract pose trajectory features from the three-dimensional pose angle and motion trajectory; the pose trajectory features include the range of motion (ROM) of joints, the severity of the motion state (Jerk value), trajectory symmetry, activity angle change rate, multi-joint motion phase difference, asymmetric fluctuation of the acceleration signal, low-frequency vibration in the stationary state, etc.

[0105] Range of motion (ROM) of joints: Calculated by the difference between the maximum value and the minimum value of the pose angle;

[0106] Severity of the motion state (Jerk value): Calculated by the root mean square value of the derivative of acceleration;

[0107] Trajectory symmetry: Quantified by the trajectory mirror symmetry parameter (such as the cosine similarity of the left and right trajectories);

[0108] Activity angle change rate: Calculated by the first derivative of the attitude angle time series;

[0109] Multi-joint movement phase difference: Analyzed by the phase correlation of the attitude angles of multiple sensors (such as double arm guards);

[0110] Asymmetric fluctuation of acceleration signal: Analyzed jointly by the skewness and kurtosis of the acceleration time-domain signal;

[0111] Low-frequency vibration in the stationary state: The energy ratio of the 0.5 - 3 Hz frequency band is extracted through frequency-domain analysis.

[0112] Step 5, Feature fusion and dimensionality reduction:

[0113] Merge the time-domain features, frequency-domain features, and attitude trajectory features (kinematic parameters) into a high-dimensional feature matrix, and perform dimensionality reduction on the high-dimensional features through principal component analysis (PCA) to eliminate redundant information and construct a low-dimensional feature vector; In the specific implementation process, the principal components with a cumulative variance contribution rate ≥ 95% can be retained through principal component analysis to generate an 8 - 12-dimensional low-dimensional feature vector.

[0114] Step 6, Health identification:

[0115] Input the low-dimensional feature vector into a pre-trained health assessment model, and output the arm movement function score, abnormal type warning, and rehabilitation guidance. In the specific implementation process, the arm movement function score can be from 0 to 100 points, and the abnormal types include one or more of tremor, insufficient range of motion, bradykinesia, movement discontinuity, movement trajectory deviation, joint stiffness, abnormal muscle tone, movement coordination disorder, postural tremor, and inertial overshoot. The abnormal types can be determined in the following ways:

[0116] Tremor: Detect abnormal vibrations in specific frequency bands through the frequency-domain energy spectrum (FFT energy spectrum).

[0117] Insufficient range of motion: Quantitatively judged by the joint range of motion (ROM) parameter.

[0118] Bradykinesia: The time-domain speed index is lower than the threshold.

[0119] Movement discontinuity: The Jerk value reflects abnormal movement smoothness.

[0120] Movement trajectory deviation: Detect the deviation of the trajectory symmetry parameter.

[0121] Joint stiffness: Analyze the activity angle change rate.

[0122] Abnormal muscle tone: Asymmetric fluctuation of the acceleration signal.

[0123] Movement coordination disorder: Analyze the multi-joint movement phase difference.

[0124] Postural tremor: Detection of low-frequency vibration in the static state.

[0125] Inertial overshoot: Monitoring of the peak angular velocity exceeding the limit.

[0126] The pre-trained health assessment model in Step 6 is a machine learning model, such as Figure 2 As shown, the training method of the machine learning model includes:

[0127] Step 6.1, the user performs different arm movements, including grasping, lifting, waving, and rotating actions. Execute Steps 1 - 5 to obtain the arm movement dataset, divide the training set and the test set from the arm movement dataset through five-fold cross-validation, and optimize the hyperparameters of the machine learning model using grid search;

[0128] Step 6.2, with the patient's informed consent, the doctor clinically evaluates the arm movement function according to the Fugl-Meyer scoring standard and divides it into grades;

[0129] Step 6.3, compare the performance of support vector machine, random forest, and convolutional neural network, and select the model with the mean AUC of five-fold cross-validation ≥ 0.85 and the standard deviation ≤ 0.05 for deployment.

[0130] When the model selected in Step 6.3 is a support vector machine, its kernel function is a radial basis kernel function, and the regularization parameter C ∈ [0.1, 10];

[0131] When the model selected in Step 6.3 is a random forest, the number of decision trees ≥ 100, and the maximum depth ≤ 15;

[0132] When the model selected in Step 6.3 is a convolutional neural network, the network structure contains ≤ 3 convolutional layers and is deployed on an edge computing device, and the edge computing device can use NVIDIA Jetson Nano, Canaan Technology K210, etc.

[0133] Step 6 includes:

[0134] Input the low-dimensional feature vector into the model selected in Step 6.3, and output the arm movement function score, abnormal type warning, and rehabilitation guidance;

[0135] Real-time monitor the arm movement data, and the delay from data acquisition to the output of the abnormal type warning ≤ 200ms, to realize the recognition of abnormal arm movement types or movement functions;

[0136] Verify the reliability of the model through the cross-validation of the doctor's evaluation and the machine learning results, calculate the consistency coefficient based on the independent test set, and if the consistency coefficient Kappa ≥ 0.80, determine that the doctor's diagnosis result is consistent with the model prediction.

[0137] Example 2:

[0138] As Figure 4 shown, the second embodiment of the present application discloses an arm movement monitoring and health recognition system based on a wearable sensor. This system can be applied to the arm movement monitoring and health recognition method based on a wearable sensor described in Example 1, and includes a signal acquisition module, a data parsing module, a data processing module, a feature extraction and fusion module, a health assessment module, and a model training module:

[0139] The signal acquisition module is used to collect multi-axis motion signals during the movement of the user's arm. The multi-axis signals include three-axis acceleration, three-axis angular velocity, and three-axis magnetic field signals;

[0140] The data parsing module is used to parse, calibrate, and perform low-pass filtering on the original multi-axis signals;

[0141] The data processing module is used to use an improved Kalman filtering algorithm for the multi-axis signals after low-pass filtering, fuse the data of the accelerometer, gyroscope, and magnetometer, and calculate the three-dimensional attitude angle and motion trajectory of the arm in real time; based on the change law of the attitude angle, use a hybrid action segmentation algorithm to segment the time series data of the three-dimensional attitude angle and motion trajectory of the arm into standardized action sequences, and each sequence corresponds to a single complete action;

[0142] The feature extraction and fusion module is used to perform time-domain analysis and frequency-domain analysis on the standardized action sequences to obtain time-domain features and frequency-domain features; extract attitude trajectory features from the three-dimensional attitude angle and motion trajectory; merge the time-domain features, frequency-domain features, and attitude trajectory features into a high-dimensional feature matrix, and generate a low-dimensional feature vector through principal component analysis;

[0143] The health assessment module is used to input the low-dimensional feature vector into a pre-trained health assessment model, and output an arm movement function score, an abnormal type warning, and rehabilitation guidance;

[0144] The model training module is used to train the health assessment model according to the training data set to generate a pre-trained health assessment model; the health assessment model is a machine learning model.

[0145] Example 3:

[0146] As Figure 5As shown, the third embodiment of the present application discloses an arm movement monitoring and health identification device based on a wearable sensor. This device can be applied to the method for arm movement monitoring and health identification based on a wearable sensor described in the first embodiment, and includes a main control chip 1, a power supply module 2, a sensor module 3, a Bluetooth module 4, a wearable arm guard unit 5, a cloud server 6, and a mobile terminal 7. The main control chip 1 is connected to the sensor module 3 and the Bluetooth module 4 and encapsulated into the wearable arm guard unit 5.

[0147] The main control chip 1 is used to control the sensor module to perform signal acquisition, signal preprocessing (calibration, filtering), execute Kalman filter attitude solution, and control the Bluetooth module to achieve real-time data transmission. In the specific implementation process, the main control chip 1 can use the low-power Bluetooth SoC chip nRF52840 of Nordic Semiconductor, the ultra-low-power MCU STM32L4R9 of STMicroelectronics, or the C2640R2F of Texas Instruments (supporting BLE 5.1). Nordic Semiconductor nRF52840 is preferred, and the others are alternatives.

[0148] The power supply module 2 is used to supply power to the main control chip 1, the sensor module 3, and the Bluetooth module 4. In the specific implementation process, the power supply module can use the CoinPower CP1254 series of VARTA or the thin battery SDI Li-ion 702030 of Samsung. The VARTA CoinPower CP1254 series is preferred.

[0149] The sensor module 3 is used to collect the original multi-axis signals during the arm movement. The multi-axis signals include three-axis acceleration, three-axis angular velocity, and three-axis magnetic field signals. In the specific implementation process, the sensor module can use the LSM9DS1 of STMicroelectronics or the MPU-9250 of InvenSense. STMicroelectronics LSM9DS1 is preferred.

[0150] The Bluetooth module 4 is used to send the data processed by the main control chip 1 to the mobile terminal 7 through wireless communication. In the specific implementation process, the Bluetooth module can use the BGM220 of Silicon Labs.

[0151] The mobile terminal 7 is used to control the start and stop of the sensor module 3, perform data encoding and decoding, and interact with the cloud server 6; the data encoding and decoding includes the mobile terminal 7 encoding the data sent by the Bluetooth module 4 and sending it to the cloud server 6, the cloud server 6 decoding it after receiving, processing it and then encoding and sending it to the mobile terminal 7, and the mobile terminal 7 decoding it and presenting it to the user. In the specific implementation process, the control operation can be performed through the health monitoring APP installed on the mobile terminal 7, and the interface of the health monitoring APP is simple and meets the operation requirements of ordinary users and medical staff.

[0152] The cloud server 6 is used to deploy a pre-trained health assessment model in advance, parse and decode the data of the request sent by the mobile terminal 7, process the data using the pre-trained health assessment model, generate an arm movement function score, an abnormal type warning and rehabilitation guidance, and encode and return them to the mobile terminal 7; the pre-trained health assessment model is a machine learning model.

[0153] The sensor module 3 is a nine-axis inertial sensor, including a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer, which collects multi-axis motion signals of the user's arm during movement; the wireless communication in the Bluetooth module 4 uses the low-power Bluetooth protocol BLE; the wearable arm unit 5 is a medical armband or a sports armband, ensuring convenient wearing, and the wearable arm unit conforms to the ergonomic design to avoid discomfort.

[0154] In the specific implementation, the present application provides a computer storage medium and a corresponding data processing unit. Among them, the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, it can run the invention content of a method for monitoring arm movement and health recognition based on wearable sensors provided by the present invention and some or all of the steps in each embodiment. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0155] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of a computer program and its corresponding general hardware platform. Based on such an understanding, the technical solutions in the embodiments of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a computer program, that is, a software product. The computer program software product can be stored in the storage medium, including several instructions for causing a device (which can be a personal computer, a server, a single-chip microcomputer, a MUU, or a network device, etc.) including a data processing unit to execute the methods described in each embodiment or some parts of the embodiments of the present invention.

[0156] The present invention provides a method and system for arm movement monitoring and health identification based on wearable sensors. There are many methods and ways to specifically implement this technical solution. The above description is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. Each component not clearly defined in this embodiment can be implemented by using the prior art.

Claims

1. A wearable sensor-based arm motion monitoring and health recognition method, characterized in that: The following steps are involved: Step 1, multi-axis signal acquisition: collecting original multi-axis signals during the user's arm movement, wherein the multi-axis signals include three-axis acceleration, three-axis angular velocity and three-axis magnetic field signals; Step 2, signal preprocessing and sensor fusion: the original multi-axis signal is analyzed, calibrated and low-pass filtered, and the improved Kalman filter algorithm with adaptive dynamic noise covariance is used for the multi-axis signal after low-pass filtering, and the accelerometer, gyroscope and magnetometer data are integrated to calculate the three-dimensional posture angle and motion trajectory of the arm in real time; Step 3, action cycle segmentation: Based on the change law of posture angle, a hybrid action segmentation algorithm is used to segment the time series data of the three-dimensional posture angle and motion trajectory of the arm into standardized action sequences, each sequence corresponding to a single complete action; Step 4, multi-dimensional feature extraction: perform time domain analysis and frequency domain analysis on the standardized action sequence to obtain time domain features and frequency domain features; extract posture trajectory features from the three-dimensional posture angle and motion trajectory; Step 5, feature fusion and dimensionality reduction: merge the time domain features, frequency domain features and posture trajectory features into a high-dimensional feature matrix, and generate a low-dimensional feature vector through principal component analysis; Step 6, health identification: input the low-dimensional feature vector into the pre-trained health assessment model to output the arm motor function score, abnormal type warning and rehabilitation guidance.

2. According to claim 1, a wearable sensor-based arm movement monitoring and health identification method is characterized in that: The improved Kalman filter algorithm of adaptive dynamic noise covariance described in step 2 includes: When the arm moves violently, the process noise covariance Q is dynamically adjusted k =Q0·(1+α·Jerk); Where Q0 represents the static noise covariance, which is obtained through wearable sensor calibration; α is the dynamic adjustment coefficient, with a value range of [0,1], which is calibrated through experiments; Jerk represents the derivative of acceleration; Dynamically adjust the magnetometer observation noise covariance R when magnetic interference exists mag =R mag.base ·(1+β·σ M ); Among them, R mag.base represents the basic noise covariance of the magnetometer, β represents the magnetic interference sensitivity coefficient, which is calibrated by experiments, and σ M represents the standard deviation of magnetic field intensity; when σ M When it is greater than the set threshold, it is determined that magnetic interference exists.

3. The arm movement monitoring and health identification method based on wearable sensors according to claim 2 is characterized in that: Step 3 includes: Step 3-1, rough segmentation: dynamically set the sliding window length, detect the peak value of the pitch angle change rate in the sliding window, and preliminarily divide the action interval; the sliding window length L = L base +η·var(θ pitch ), where L base Indicates the basic window length, covering conventional actions; var(θ pitch ) represents the variance of the pitch angle in the current window, reflecting the intensity of the change in the action amplitude; η represents the adjustment coefficient calibrated by the training data, which is used to amplify or reduce the influence of the pitch angle variance; Step 3-2, fine segmentation: match the preset action template only within the initially divided action interval to obtain candidate boundary points; use the DBSCAN algorithm based on density clustering to eliminate isolated noise points in the matched candidate boundary points, and the final boundary is determined by the center point of the maximum density cluster.

4. The arm movement monitoring and health identification method based on wearable sensors according to claim 3 is characterized in that: The pre-trained health assessment model in step 6 is a machine learning model, and the training method of the machine learning model includes: Step 6.1, the user performs different arm movements, including grasping, lifting, swinging and rotating movements, and executes steps 1 to 5 to obtain an arm movement dataset, divides the arm movement dataset into a training set and a test set by five-fold cross validation, and uses grid search to optimize the hyperparameters of the machine learning model; Step 6.2: With the patient's informed consent, the doctor conducts a clinical assessment of the arm motor function and grading according to the Fugl-Meyer scoring standard; In step 6.3, the performance of support vector machine, random forest and convolutional neural network is compared, and the model with the mean AUC of 5-fold cross validation ≥ 0.85 and standard deviation ≤ 0.05 is selected for deployment.

5. The arm movement monitoring and health identification method based on wearable sensors according to claim 4 is characterized in that: Step 6 includes: Input the low-dimensional feature vector into the model selected in step 6.3, and output the arm motor function score, abnormality type warning and rehabilitation guidance; Real-time monitoring of arm movement data, with a delay of ≤200ms from data collection to abnormal type warning output, to identify abnormal arm movement types or motor functions; The reliability of the model was cross-validated by doctor evaluation and machine learning results, and the consistency coefficient was calculated based on an independent test set. If the consistency coefficient Kappa ≥ 0.80, the doctor's diagnosis result was judged to be consistent with the model prediction.

6. The arm movement monitoring and health identification method based on wearable sensors according to claim 5 is characterized in that: When the model selected in step 6.3 is a support vector machine, its kernel function is a radial basis kernel function, and the regularization parameter C∈[0.1,10]; When the model selected in step 6.3 is random forest, the number of decision trees is ≥ 100 and the maximum depth is ≤ 15; When the model selected in step 6.3 is a convolutional neural network, the network structure contains ≤3 convolutional layers and is deployed on an edge computing device.

7. An arm movement monitoring and health recognition system based on wearable sensors, characterized in that: It includes signal acquisition module, data analysis module, data processing module, feature extraction and fusion module, health assessment module and model training module. The signal acquisition module is used to collect original multi-axis signals during the user's arm movement, wherein the original multi-axis signals include three-axis acceleration, three-axis angular velocity and three-axis magnetic field signals; The data analysis module is used to analyze, calibrate and low-pass filter the original multi-axis signal; The data processing module is used to use an improved Kalman filter algorithm to fuse the accelerometer, gyroscope and magnetometer data on the multi-axis signal after low-pass filtering to calculate the three-dimensional posture angle and motion trajectory of the arm in real time; based on the change law of the posture angle, a hybrid action segmentation algorithm is used to segment the time series data of the three-dimensional posture angle and motion trajectory of the arm into standardized action sequences, each sequence corresponding to a single complete action; The feature extraction and fusion module is used to perform time domain analysis and frequency domain analysis on the standardized action sequence to obtain time domain features and frequency domain features; Extract posture trajectory features from 3D posture angles and motion trajectories; combine time domain features, frequency domain features, and posture trajectory features into a high-dimensional feature matrix, and generate a low-dimensional feature vector through principal component analysis; The health assessment module is used to input the low-dimensional feature vector into the pre-trained health assessment model and output the arm motor function score, abnormal type warning and rehabilitation guidance; The model training module is used to train the health assessment model according to the training data set to generate a pre-trained health assessment model; the health assessment model is a machine learning model.

8. An arm movement monitoring and health identification device based on wearable sensors, characterized in that: It includes a main control chip, a power supply module, a sensor module, a Bluetooth module, a wearable arm guard unit, a cloud server and a mobile terminal. The main control chip is connected with the sensor module and the Bluetooth module and packaged into the wearable arm guard unit; The main control chip is used to control the sensor module to perform signal acquisition, signal preprocessing, Kalman filter attitude solution, and control the Bluetooth module to achieve real-time data transmission; The power supply module is used to supply power to the main control chip, the sensor module and the Bluetooth module; The sensor module is used to collect original multi-axis signals during arm movement, wherein the original multi-axis signals include three-axis acceleration, three-axis angular velocity and three-axis magnetic field signals; The Bluetooth module is used to send the data processed by the main control chip to the mobile terminal through wireless communication; The mobile terminal is used to control the start and stop of the sensor module, data encoding and decoding, and data interaction with the cloud server; The cloud server is used to deploy a pre-trained health assessment model in advance, parse and decode data from requests sent by mobile terminals, process the data using the pre-trained health assessment model, generate arm motor function scores, abnormality type warnings and rehabilitation guidance, and encode them and return them to the mobile terminal; the pre-trained health assessment model is a machine learning model.

9. The arm movement monitoring and health identification device based on wearable sensors according to claim 8, characterized in that: The sensor module is a nine-axis inertial sensor, including a three-axis accelerometer, a three-axis gyroscope and a three-axis magnetometer.

10. The arm movement monitoring and health identification device based on wearable sensors according to claim 8, characterized in that: The wireless communication in the Bluetooth module adopts the low-power Bluetooth protocol.

Images