Intelligent motion assessment and damage early warning method and system based on multi-source biological signal fusion

Abstract

The invention provides an intelligent motion evaluation and damage early warning method and system based on multi-source biological signal fusion, and is applied to the technical field of medical data processing. The method comprises the following steps: acquiring a multi-modal biological signal acquired based on a multi-source sensing node; carrying out self-adaptive preprocessing on the multi-modal biological signals, eliminating motion artifacts by adopting an improved spectral subtraction formula, realizing multi-rate synchronization through a Lagrange kernel, triggering node vibration by virtue of signal quality evaluation to prompt electrode reattachment, and generating preprocessed signal features; and constructing cross-modal fusion features by using multi-scale cavity convolution and the like, and outputting results such as a motion mode and the like. A target depth model is trained in combination with signal optimization and data enhancement and the like, a motion pattern classification result, joint load estimation data and a damage risk scoring result are processed based on the target depth model, damage early warning information is generated, and accurate and dynamic motion evaluation and damage early warning are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data processing, in particular to an intelligent motion evaluation and injury warning method and system based on multi-source biosignal fusion. BACKGROUND

[0002] The current motion monitoring technology faces three technical bottlenecks: first, a single modality sensing system (such as using only an accelerometer) cannot fully reflect the complex interaction between neuromuscular control and biomechanical load during motion. For example, a traditional inertial measurement unit cannot distinguish between acceleration signals generated by active muscle contraction and passive joint motion. Second, multi-source signal fusion methods mostly use simple feature-level concatenation, ignoring the dynamic coupling relationship between different physiological signals. Experimental data shows that this processing method can introduce 15-20% cross-interference error due to the phase difference between electromyographic signals and inertial data. Third, the biomechanical modeling of existing systems is too simplified, usually using linear regression methods to estimate joint torque, with a prediction error of up to 30-45% in dynamic motion.

[0003] At the signal processing level, the existing technology has two key defects: motion artifact removal algorithms mostly use time-domain adaptive filtering, but cannot effectively separate the respiratory interference (0.1-0.3Hz) from the true R-wave component (1-2Hz) in ECG signals. In addition, multi-source data synchronization mainly relies on hardware timestamps, and when the sampling rate difference exceeds 5 times (such as 500Hz EMG and 30Hz PPG), the interpolation error can cause 0.8-1.2ms timing misalignment.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide an intelligent motion evaluation and injury warning method and system based on multi-source biosignal fusion, which at least partially overcomes the problems existing in the prior art, and realizes dynamic injury warning through multi-source sensing and deep model fusion. Multi-modal signals are acquired, features are generated after preprocessing, cross-modal fusion features are constructed, relevant information is output, signals are optimized and models are trained, comprehensive warnings are generated, and the problems of single modality limitation and signal interference are solved.

[0006] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the application.

[0007] According to an aspect of the present application, a method for intelligent motion assessment and injury warning based on multi-source biosignal fusion is provided, comprising: acquiring multi-modal biosignals collected based on multi-source sensing nodes, wherein the multi-source sensing nodes include trunk nodes, limb nodes and peripheral nodes, and the multi-modal biosignals include inertial dynamics, electromyographic activity and hemodynamic signals; performing adaptive preprocessing on the multi-modal biosignals, using an improved spectral subtraction formula to eliminate motion artifacts, implementing multi-rate synchronization through a Lagrange kernel, and triggering node vibration prompt re-pasting electrodes with the help of signal quality assessment, to generate preprocessed signal features; based on the preprocessed signal features, using multi-scale hollow convolution, gated collaborative attention and bidirectional gated recurrent unit to construct a feature fusion layer, to generate cross-modal fusion features; processing the cross-modal fusion features to generate motion pattern classification results, joint load estimation data and injury risk score results; processing the original sensor data and the signals in the preprocessing process, using three-step static-dynamic-static calibration and temperature drift compensation to calibrate the sensor, and through IVA-Wiener-wavelet soft threshold-Kalman-NMF three-level suppression, to generate signal features after multi-level suppression of motion artifacts; based on the original data and the preprocessed signal features, cross-modal fusion features, decision reasoning intermediate features and signals after multi-level suppression of motion artifacts, using data augmentation strategy, dynamic weighted loss function and Lookahead-AdamW optimizer, processing a pre-set deep model to generate a target deep model; based on the target deep model, processing the motion pattern classification results, joint load estimation data and injury risk score results to generate injury warning information.

[0008] In another aspect of the present application, an intelligent motion evaluation and injury warning device based on multi-source biosignal fusion is characterized in that it comprises: an acquisition module for acquiring multi-modal biosignals collected based on multi-source sensing nodes; a processing module for adaptively pre-processing the multi-modal biosignals, eliminating motion artifacts by using an improved spectral subtraction formula, realizing multi-rate synchronization through a Lagrange kernel, and triggering node vibration prompt re-pasting of electrodes by means of signal quality evaluation, to generate signal features after pre-processing; based on the signal features after pre-processing, a feature fusion layer is constructed by using multi-scale hollow convolution, gated collaborative attention, and bidirectional gated recurrent unit to generate cross-modal fusion features; the cross-modal fusion features are processed to generate motion pattern classification results, joint load estimation data, and injury risk score results; the original acquisition data of the sensor and the signals in the pre-processing process are processed, the sensor is calibrated by using three-step static-dynamic-static calibration and temperature drift compensation, and the signal features after multi-level suppression of motion artifacts are generated through IVA-Wiener-wavelet soft threshold-Kalman-NMF three-level suppression; based on the original data recorded by multi-center heterogeneous devices and the signal features after pre-processing, the cross-modal fusion features, the intermediate features in decision reasoning, and the signals after multi-level suppression of motion artifacts, a data enhancement strategy, a dynamic weighted loss function, and a Lookahead-AdamW optimizer are used to process a preset deep model to generate a target deep model; the motion pattern classification results, the joint load estimation data, and the injury risk score results are processed based on the target deep model to generate injury warning information.

[0009] According to still another aspect of the present application, an electronic device is characterized by comprising: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to implement the above-mentioned intelligent motion evaluation and injury warning method based on multi-source biosignal fusion.

[0010] The intelligent motion evaluation and injury warning method and system based on multi-source biosignal fusion provided by the present application realize dynamic injury warning through multi-source sensing and deep model fusion. First, multi-modal signals such as inertial dynamics, electromyography, and hemodynamics of the torso, limbs, and peripheral nodes are acquired, and features are generated after pre-processing such as artifact removal by an improved spectral subtraction formula and synchronization by a Lagrange kernel. Then, cross-modal fusion features are constructed by using multi-scale hollow convolution and gated collaborative attention. Through classification, load estimation, and risk scoring models, motion patterns and other information are output, and the signals are optimized by sensor calibration and multi-level suppression. Finally, a target deep model is trained by combining data enhancement, and comprehensive warning information is generated, solving the limitations of single modalities and signal interference, and realizing precise and dynamic evaluation and warning.

[0011] It should be understood that the general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 A flow chart of a method for intelligent motion assessment and injury warning based on multi-source biosignal fusion is shown according to an embodiment of the present application.

[0013] Figure 2 A structural schematic diagram of a device for intelligent motion assessment and injury warning based on multi-source biosignal fusion is shown according to an embodiment of the present application. DETAILED DESCRIPTION

[0014] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to explain and illustrate the present application, and are not used to limit the present application.

[0015] The method for intelligent motion assessment and injury warning based on multi-source biosignal fusion according to the exemplary embodiments of the present application is described below in conjunction with Figure 1 It should be noted that the following application scenarios are only shown for the purpose of facilitating the understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect. On the contrary, the embodiments of the present application are applicable to any applicable scenarios.

[0016] In one embodiment, the present application further provides a method and system for intelligent motion assessment and injury warning based on multi-source biosignal fusion. Figure 1 A flow chart of a method for intelligent motion assessment and injury warning based on multi-source biosignal fusion according to an embodiment of the present application is shown.

[0017] S101, acquiring multi-modal biosignals collected by multi-source sensing nodes.

[0018] In one embodiment, the multi-source sensing nodes are divided into three categories according to body parts and are deployed at different positions to comprehensively collect signals: trunk nodes, mainly deployed in the core areas of the trunk such as the chest and waist, for monitoring the overall motion state of the trunk and the activity of the core muscle group. A sensing node is pasted at the median position of the waist, which can collect inertial dynamics signals of the trunk such as pitch and rotation, reflecting the amplitude and speed of actions such as bending over and turning around. Limb nodes are deployed on the limbs (such as the thigh, lower leg, upper arm, and forearm), focusing on capturing the motion trajectory and muscle activity of the limbs. A node is installed at the quadriceps muscle on the front side of the thigh, which can collect electromyographic signals when the knee joint is flexed and extended, and judge the muscle contraction strength. Peripheral nodes are located at the ends of the limbs such as hands and feet, for monitoring peripheral circulation and fine motion related signals. A sensing node is worn at the wrist, which can collect hemodynamic signals (such as pulse wave) and record inertial dynamics data of wrist rotation.

[0019] Multi-modal bio-signals cover three types of key physiological and motion information, which are collected by corresponding sensing nodes: inertial dynamics signals, used to describe the motion state of body parts (such as acceleration, angular velocity, displacement, etc.), collected by the inertial measurement unit (IMU) of each node. The lower leg acceleration signal is collected by the limb node during running, and motion parameters such as step frequency and stride length can be analyzed; the angular velocity signal is collected by the trunk node when bending, and the speed and stability of the action can be judged.

[0020] Electromyographic signals reflect the electrical activity of muscles during muscle contraction, which are collected by electrodes attached to the surface of the muscle. When performing dumbbell bicep curls, the electromyographic signal of the upper arm biceps is collected by the limb node, and the muscle strength can be judged by the signal intensity change; the electromyographic activity of the abdominal muscles is collected by the trunk node to evaluate the participation of the core muscle group.

[0021] Hemodynamic signals are used to monitor blood flow-related parameters (such as pulse, blood oxygen saturation, etc.), mainly collected by optical sensors on peripheral nodes. During long-term exercise (such as a marathon), the pulse wave signal of the finger is collected by the peripheral node, and the heart rate change and blood circulation state can be analyzed; the fatigue level during exercise can be evaluated by the blood flow signal at the wrist node.

[0022] The signals collected by different nodes complement each other to form a complete motion evaluation data chain: inertial dynamics signals from the trunk node + electromyographic signals from the limb node: can jointly analyze the coordination of the action (such as the matching degree of trunk stability and leg muscle force during deep squat). Hemodynamic signals from peripheral nodes + inertial dynamics signals from limb nodes: can correlate exercise intensity and physiological load (such as when running fast, the lower leg inertial signal reflects the exercise intensity, and the finger blood flow signal reflects the cardiopulmonary load). Through the distributed deployment of multi-source nodes and the fusion of multi-modal signals, physiological and mechanical characteristics during exercise can be comprehensively captured, providing basic data for subsequent motion evaluation and injury warning.

[0023] S102, adaptive preprocessing of multi-modal bio-signals, using an improved spectral subtraction formula to eliminate motion artifacts, implementing multi-rate synchronization through Lagrange kernel, and triggering node vibration prompt electrode reattachment with the help of signal quality assessment, generating preprocessed signal features.

[0024] In one embodiment, the multi-modal bio-signals are adaptively preprocessed, and an improved spectral subtraction formula S proc (ξ)=(S sig (ξ) 2 -γS bkg (ξ) 1 / 2 e iψ(ξ) Eliminate motion artifacts and restore ECG P-QRS-T waveforms at 15 km / h.proc (ξ) is the output signal after preprocessing, that is, the effective biological signal (such as electrocardiogram signal, electromyogram signal, etc.) after removing motion artifacts. ξ is the independent variable, representing the frequency or time variable of the signal (according to the signal processing scene, it can represent the parameters in the frequency domain or time domain).

[0025] S sig (ξ) is the original signal component containing effective information, that is, the real biological signal (such as the signal corresponding to the P-QRS-T waveform in the electrocardiogram signal) that needs to be preserved. S bkg (ξ) is the background noise or motion artifact signal, that is, the interference component (such as the artifact generated by body shaking during motion) that needs to be removed. γ is an adjustment factor used to balance the weight of the signal and the artifact, and by dynamically setting the control strength of the artifact removal, it avoids excessive removal that leads to loss of effective signal. e iψ(ξ) is a phase-preserving term, where ψ(ξ) is the phase information of the original signal. This parameter is used to preserve the phase characteristics of the signal while removing artifacts, ensuring the waveform integrity of the processed signal (such as the phase characteristics of the electrocardiogram signal, which are crucial for heart rate analysis)

[0026] By improving the spectral subtraction formula, the original biological signal is processed to remove the interference signal (motion artifact) generated by body shaking, muscle tremor, etc. during motion, and to preserve the true physiological and motion signal. When the user runs at a speed of 15 km / h, the electrocardiogram (ECG) signal collected by the trunk node is easily affected by body shaking, resulting in baseline drift or waveform distortion. After applying the formula, the artifact and the true electrocardiogram signal can be effectively separated, and the P wave, QRS complex, and T wave (P-QRS-T waveform) can be clearly recovered, accurately reflecting the cardiac electrical activity.

[0027] By dynamically adjusting the order d = [log2ρ] + 1 of the variable-order pseudo-Laguerre kernel, multi-rate synchronization is achieved, and the 500Hz EMG and 30Hz PPG mixed sampling are aligned to sub-millisecond level. By dynamically adjusting the order of the variable-order pseudo-Laguerre kernel, the time sequence misalignment problem caused by the difference in sampling rate of different sensors is solved, and the time alignment of multi-modal signals is achieved. The electromyogram (EMG, sampling rate 500Hz) collected by the limb node and the photoplethysmogram (PPG, sampling rate 30Hz) collected by the peripheral node have a large difference in sampling rate, and the original data has a time deviation. After adjusting the order by this method, the time alignment accuracy of the two signals can be improved to sub-millisecond level, ensuring the accurate time correspondence between electromyographic activity and blood flow changes.

[0028] The signal quality evaluation triggers the node vibration to prompt the electrode reattachment, and generates the preprocessed signal features. The signal quality (such as signal strength, signal-to-noise ratio, etc.) is monitored in real time, and when the signal quality is lower than the threshold (such as the electrode loosening leading to the muscle electromyogram blur), the sensing node vibration is triggered to prompt the user to reattach the electrode, ensuring the continuity and reliability of signal acquisition. When the user performs dumbbell training, the electromyogram electrode of the upper arm muscle is loosened due to large amplitude of motion, resulting in a sharp drop in the signal-to-noise ratio of the collected electromyogram. The system detects the abnormality through signal quality evaluation, controls the limb node vibration, reminds the user to reattach the electrode, and avoids invalid data acquisition.

[0029] The above three steps together constitute an adaptive preprocessing process of multi-modal biological signals, which eliminates artifacts, synchronizes data, and ensures signal quality, and finally generates preprocessed signal features, providing a high-quality data basis for subsequent cross-modal feature fusion, motion evaluation, and injury warning.

[0030] S103, based on the preprocessed signal features, a feature fusion layer is constructed using multi-scale hollow convolution, gated collaborative attention, and bidirectional gated recurrent unit to generate cross-modal fusion features.

[0031] In one embodiment, based on the preprocessed signal features, multi-scale hollow convolution is used to capture the step frequency micro-variation within a 1.25s window to realize real-time fatigue inflection point prompting, and generate multi-scale spatiotemporal features. By using multi-scale hollow convolution to capture the subtle changes in step frequency within a 1.25s time window, the fatigue inflection point in motion is identified in real time, and multi-scale spatiotemporal features containing information of different time scales and spatial dimensions are generated. During long-distance running, the system continuously monitors the step frequency signal collected by the lower limb sensing node. When the step frequency within a 1.25s window decreases from a stable 3 steps / s to 2.8 steps / s, it can be judged that the user has entered the fatigue inflection point, and a real-time reminder is issued to adjust the exercise intensity.

[0032] The gated collaborative attention is used to judge the synchronization degree of the electromyogram peak value and the joint angular acceleration by the formula ω ab = σ(P T tanh(Qz MG + Qz MU )) to generate cross-modal correlation weight features. ω ab is the cross-modal correlation weight, which is used to measure the synchronization degree between the electromyogram and the joint angular acceleration signal, and the value range is [0, 1], and the closer the value is to 1, the higher the synchronization. σ is an activation function (such as Sigmoid function), which is used to map the output result to the interval [0, 1] to standardize the weight value. P is a learnable weight matrix, which is used to perform linear transformation on the input features to extract key correlation information. P Tis the transpose of weight matrix P. tanh is the hyperbolic tangent activation function, which is used to perform a nonlinear transformation on the input features to enhance the model's ability to capture complex correlations. Q is a learnable transformation matrix that maps features from different modalities to the same dimensional space for fusion computation. MG is the feature vector of pre-processed electromyography (EMG) signals, containing information such as muscle contraction strength and peak timing. MU is the feature vector of pre-processed joint angular acceleration signals, reflecting the rate of change of joint movement speed.

[0033] The synchronization degree between the peak value of the electromyography signal and the joint angular acceleration is calculated by the formula, generating a weight feature that reflects the correlation strength of different modalities of signals, highlighting key correlation information. In the squat movement, the peak value of the electromyography signal (from the thigh muscle) should be synchronized with the change in the knee joint angular acceleration. If the calculated ω ab is low, it indicates that the synchronization between the two is poor (such as the timing of muscle exertion being out of sync with joint movement), which the system will consider as one of the features of non-standard movement.

[0034] The bidirectional gated recurrent unit is used to track a 30-minute movement sequence and identify abnormal movement patterns by the formula g t = σ(A g [s t-1 ; y t ] + c g ), generating time-dependent features. g t is the gated output at the current time (t), used to control the fusion ratio of the previous time information and the current input information, affecting the update of the time-dependent features. σ is the activation function (such as the Sigmoid function), which maps the output to the [0, 1] interval, realizing the gating mechanism (0 represents complete blocking, and 1 represents complete passing). A g is the weight matrix of the gating unit, used to perform linear transformation on the input time-dependent features. s t-1 is the hidden state at the previous time (t-1), containing time-dependent information of the historical movement sequence (such as the movement pattern in the previous 30 minutes).

[0035] y t is the input feature at the current time (t), coming from pre-processed multi-modal biosignals (such as inertial dynamics, electromyography signals, etc.). [s t-1 ; y t ] is the joint feature vector formed by concatenating the hidden state at the previous time and the current input feature. c gThe bias term of the gating unit is used to adjust the baseline of the linear transformation and enhance the flexibility of the model. The above parameters work together to respectively realize the quantification of the cross-modal signal correlation strength (gating collaborative attention) and the tracking of the long-time motion pattern (bidirectional gating recurrent unit), providing key support for the generation of cross-modal fusion features.

[0036] The formula is used to continuously track the motion sequence for 30 minutes, identify the abnormal motion pattern, and generate the time sequence feature reflecting the time sequence dependency. In basketball training, the system tracks the shooting motion sequence within 30 minutes. When the shooting motion amplitude (from inertial dynamics signal) at a certain time suddenly deviates from the stable range of the previous 10 minutes, and the abnormality lasts for 3 consecutive time steps, the technology can identify this abnormal motion pattern and mark it as a potential injury risk feature.

[0037] The feature fusion layer is constructed based on the multi-scale spatio-temporal feature, cross-modal correlation weight feature and time sequence dependency feature to generate the cross-modal fusion feature. The above multi-scale spatio-temporal feature, cross-modal correlation weight feature and time sequence dependency feature are fused to form a comprehensive cross-modal fusion feature containing spatio-temporal information, modal correlation and time sequence change, providing comprehensive input for subsequent motion pattern classification, joint load estimation, etc. In marathon sports, the fused features contain not only the spatio-temporal information of step frequency change, the correlation weight of electromyography and joint motion, but also the action time sequence trend within 30 minutes, which can fully support the evaluation of running posture standardization, knee joint load and fatigue injury risk. Through multi-dimensional feature extraction and fusion, the physiological and mechanical signals in the sports process are deeply analyzed, providing a key feature basis for intelligent sports evaluation and injury warning.

[0038] S104, processing the cross-modal fusion feature to generate a motion pattern classification result, joint load estimation data and injury risk score result.

[0039] In one implementation, the cross-modal fusion feature is processed, and a motion pattern classification model is used to generate a motion pattern classification result based on the formula The motion action type is identified to generate a motion pattern classification result, providing demonstration and feedback for action switching. Pr(φ=j|v) is the probability of the action type being j under the input feature v, which is used to determine which mode the current action belongs to (e.g., running for j=1 and deep squat for j=2). φ is a feature mapping function that converts the cross-modal fusion feature v into a high-dimensional feature space. v is the cross-modal fusion feature, which contains the spatio-temporal, correlation and time sequence information of multi-source biological signals.

[0040] e is a natural constant (about 2.71828), which is used as the base of the exponential function to map the linear transformation result to the probability space, so that the output value meets the non-negativity and normalization requirements of the probability distribution. For m j The dot product of the weight vector of action type j and φ(v) (the result after high-dimensional feature mapping) highlights the features that are discriminative for action type j through linear combination. For example, when recognizing "running (j = 1)", the weights of features such as step frequency and arm swing amplitude are emphasized. j The bias term for action type j, used to adjust the baseline of different action categories in probability calculation. If the "deep squat (j = 2)" has a larger d j When the features are similar, the model will be more inclined to judge as a deep squat action.

[0041] Based on the preset formula, the cross-modal fusion features are analyzed to identify the type of the current motion (such as running, deep squat, shooting, etc.), providing a reference for action specification evaluation and switching. When the user is training the upper limbs, the system identifies the current action as "dumbbell bicep curl" through the model, and compares it with the standard action pattern. If it finds that the elbow is excessively bent (deviating from the standard angle by more than 15°), it generates an action correction prompt (such as "keep the elbow slightly bent and avoid excessive bending") to help the user adjust the action.

[0042] The cross-modal fusion features are processed, and the joint load estimation model is based on the formula The actual force on the joint is calculated, and joint load estimation data is generated to monitor joint load to prevent injury. act The actual force on the joint (such as the load on the knee joint or ankle joint) reflects the true burden on the joint during exercise. F0 is the reference value based on the physiological structure of the human body. sig norm The standardized signal features (such as muscle electrical signal intensity and inertial dynamics parameters) are used to quantify exercise intensity. χ, λ opt The model parameters are decay coefficient and optimal rate parameters, respectively, used to fit the relationship between load and exercise intensity. v is the exercise speed feature (such as action speed and step frequency), which affects the dynamic change of joint load.

[0043] Based on the preset formula, the actual force on the joint (such as the knee joint, ankle joint, or elbow joint) during exercise is calculated, and the load change is monitored in real time to prevent injury due to excessive load. When a soccer player performs a sudden stop and turn action, the system estimates the knee joint load through the model. If the load exceeds 5 times the body weight (the safety threshold is 4 times the body weight), an overload warning (such as "knee joint load is too high, please reduce the frequency of sudden stop") is generated immediately to avoid meniscus injury.

[0044] The cross-modal fusion features are processed, and the injury risk scoring model is based on the formula Dynamic monitoring of pressure at specific sites, generating injury risk score results, risk threshold timeout intelligent intervention prompt. k The injury risk score at the kth moment, the higher the value, the greater the risk of injury. Q is the size of the time window (such as the previous 5 moments), used for cumulative analysis of recent pressure changes. η is the decay coefficient (0<η<1), which makes the recent data (k-q is smaller) have a greater impact on the score. q The site pressure value at the qth moment (such as joint pressure, muscle tension). ζ q The safety threshold at the qth moment, based on human tolerance, exceeding the threshold contributes to the risk score.

[0045] Based on the preset formula, the pressure of a specific site (such as the waist, shoulder) is analyzed in time sequence, and the risk score is dynamically generated. When the score exceeds the preset threshold, an intervention suggestion is automatically pushed. Office staff who perform the action of bending over to move objects for a long time, the system continuously monitors the waist pressure through this model. If the pressure value exceeds the safety threshold for 3 times in 5 consecutive measurements, the risk score reaches 85 points (threshold is 70 points), and an intervention prompt is pushed (such as "pause moving, do waist stretching; suggest using auxiliary tools to reduce waist stress").

[0046] Based on the motion pattern classification results, joint load estimation data, and injury risk score results, the cross-modal fusion features are processed to generate core decision data for motion assessment and injury warning. The above motion pattern classification results, joint load estimation data, and injury risk score results are integrated to form core decision data that comprehensively reflects the motion state and injury risk, providing a basis for subsequent injury warning. The core decision data for marathon runners may include: "motion pattern: long run (standard 80 points); knee joint load: 3.2 times body weight (safe); injury risk score: 65 points (waist, close to threshold)", based on which the system makes a comprehensive judgment and pushes the suggestion "maintain current running posture, pay attention to waist relaxation".

[0047] Through hierarchical analysis of cross-modal fusion features, precise recognition of motion actions, quantitative monitoring of joint loads, and dynamic assessment of injury risks are achieved, providing direct and reliable decision support for intelligent motion assessment and injury warning.

[0048] S105, process the sensor raw data and the signal in the pre-processing process, use three-step static-dynamic-static calibration and temperature drift compensation to calibrate the sensor, and through IVA-Wiener-wavelet soft threshold-Kalman-NMF three-level suppression, generate signal features after motion artifact multi-level suppression.

[0049] In one embodiment, the sensor raw data and the signal in the preprocessing process are processed, a three-step static-dynamic-static calibration and temperature drift compensation method is used to calibrate the sensor, the residual error of calibration is less than 62 ns, and the calibrated sensor data is generated. Focus on optimizing the raw data collected by the sensor and the preprocessed signal, and finally generate high-quality signal features by califying and multi-stage filtering to suppress motion artifacts. Specifically as follows: through the "static-dynamic-static" three-step calibration method combined with temperature drift compensation, the measurement deviation of the sensor caused by temperature change and hardware error is eliminated, the data accuracy is ensured, and the residual error of calibration is controlled within 62 ns. When calibrating the inertial sensor deployed on the calf, first let the sensor stand for 10 minutes at room temperature (25℃) (static stage), record the initial zero point offset; then let the user run for 3 minutes (dynamic stage), collect error data in the motion state; finally, stand for 5 minutes again (static stage), calculate the temperature drift coefficient (for example, if the temperature increases by 1℃, the error increases by 0.2 ns) combined with the data of the previous two stages, correct through the compensation algorithm, and finally reduce the residual error of calibration to 58 ns to generate calibrated sensor data.

[0050] The calibrated sensor data is processed, and the initial artifact separation is realized through IVA-Wiener filtering to generate a first-stage suppression signal. The independent components in the mixed signal are separated by IVA (Independent Vector Analysis), and the signal is denoised by Wiener filtering to preliminarily strip the motion artifacts (such as high-frequency interference caused by muscle tremor). The calibrated electromyographic signal is mixed with artifacts generated by leg vibration during running, the effective components and vibration noise of electromyography are separated by IVA algorithm, and then the noise energy is suppressed by Wiener filtering to retain the peak value and baseline characteristics of the electromyographic signal, thereby generating a first-stage suppression signal (the artifact energy is reduced by about 40%).

[0051] The first-stage suppression signal is processed, and the high-frequency noise interference is further eliminated through wavelet soft threshold filtering to generate a second-stage suppression signal. The signal is decomposed into different frequency components through wavelet transform, and soft threshold processing is applied to high-frequency noise components (noise less than the threshold is set to zero), thereby further eliminating the high-frequency interference (such as electronic noise introduced by the sensor circuit) remaining after the first-stage suppression. The 50Hz power frequency interference (high-frequency noise) still exists in the heart rate signal after the first-stage suppression, the high-frequency components are processed through wavelet soft threshold filtering, the interference peaks are filtered out, the heart rate waveform (low-frequency signal) is smoother, and a second-stage suppression signal is generated.

[0052] The second-stage suppression signal is processed, and the signal component optimization and artifact deep stripping are realized through Kalman-NMF algorithm to generate a third-stage suppression signal. The Kalman filter (predicting and correcting the signal trend) and NMF (Non-negative Matrix Factorization, separating the basis vectors of the signal and artifacts) are combined to realize the signal component optimization and deep stripping of artifacts, and the true physiological signal characteristics are retained.

[0053] There are still some respiratory-related artifacts (low-frequency interference) in the joint acceleration signal after secondary suppression. The Kalman filter predicts the motion trend of the acceleration, and then uses NMF to decompose the independent basis vector of the artifact and remove it, finally generating a tertiary suppression signal (artifact residual <5%).

[0054] Based on the calibrated sensor data and the first, second, and third suppression signals, the multi-level motion artifact suppression processing is completed, and the signal features after multi-level motion artifact suppression are generated. The calibrated original data and the tertiary suppression signal are integrated, and the signal features reflecting the true physiological and motion state are generated through feature extraction algorithms (such as time domain statistics and frequency domain energy analysis).

[0055] Based on the calibrated data and the tertiary suppressed electromyography and acceleration signals, features such as "muscle contraction peak time" and "joint motion acceleration average" are extracted. These features can be directly used for subsequent joint load estimation and injury risk assessment (such as determining whether the muscle effort is abnormal by the electromyography peak intensity). The conversion from raw sensor data to high-quality signal features is realized, effectively solving the motion artifact interference problem, and providing a reliable data foundation for subsequent cross-modal fusion and model training.

[0056] S106, based on the multi-center heterogeneous recording of the original data and the preprocessed signal features, cross-modal fusion features, decision reasoning intermediate features, and motion artifact multi-level suppressed signals, using data enhancement strategy, dynamic weighted loss function, and Lookahead-AdamW optimizer, the preset deep model is processed to generate the target deep model.

[0057] In one embodiment, based on the multi-center heterogeneous recording of the original data and the preprocessed signal features, cross-modal fusion features, decision reasoning intermediate features, and motion artifact multi-level suppressed signals, the training data is expanded using data enhancement strategy, including time stretching, bandwidth replacement, and Cross-SensorCutMix, to generate an enhanced training data set. Through data enhancement, dynamic loss adjustment, and parameter optimization, the preset deep model is trained, and finally the target deep model capable of accurately assessing the motion state and warning injury is generated. For multi-center heterogeneous recording of original data and various features (preprocessed signals, cross-modal fusion features, etc.), time stretching, bandwidth replacement, Cross-SensorCutMix, and other strategies are used to expand the training data, improving the generalization ability of the model (adapt to signal differences of different scenes and devices).

[0058] Time Stretch, stretch the 30-minute EMG signal sequence of marathon runners to 35 minutes (keep the action trend unchanged), or compress it to 25 minutes, simulate data of different exercise rhythms, and enhance the adaptability of the model to changes in the time dimension; Bandwidth Replacement, randomly adjust the bandwidth of the acceleration signal collected by the inertial sensor in the 10-50Hz frequency band (such as replacing the step frequency signal bandwidth from 20-30Hz to 15-25Hz when running), simulate the frequency band differences of different sensors;

[0059] Cross-Sensor CutMix, mix the upper limb EMG signal collected by sensor A with the lower limb inertial signal collected by sensor B (such as keeping 70% of the upper limb signal and 30% of the lower limb signal), simulate the signal coupling scenario when multiple sensors are cooperatively collected. Through the above strategies, the original 1000 training data can be expanded to 3000, forming an enhanced training data set.

[0060] A dynamic weighted loss function is used to dynamically adjust the model training loss, generating adaptive loss weights. According to the error of different tasks (such as exercise mode classification, joint load estimation) in the model training process, the loss weight is dynamically adjusted, so that the model pays more attention to the task with higher error, and the overall precision is improved.

[0061] In the early stage of model training, the error of exercise mode classification (15%) is much higher than that of joint load estimation (5%), and the dynamic weighting function will assign a higher weight (such as 0.7) to the classification task and a lower weight (such as 0.3) to the load estimation; As the training progresses, when the classification error drops to 8% and the load estimation error rises to 10%, the weight is automatically adjusted to 0.4 for classification and 0.6 for load estimation, ensuring that the model optimizes the high-error task.

[0062] The Lookahead-AdamW optimizer is used to optimize the model parameters, generating optimized model parameters. Combined with AdamW (Adam algorithm with weight decay, to prevent overfitting) and Lookahead (lookahead mechanism, to improve convergence stability), the model parameters are iteratively optimized to make the parameters closer to the global optimal solution.

[0063] When training the joint load estimation sub-model, AdamW is responsible for gradient updating of the current parameters (such as weight matrix, bias term), and Lookahead predicts the "candidate parameters" of the next step based on the trend of the current parameters every 5 iteration steps, and selects the better solution to update the model. For example, when the updated parameters of AdamW cause the knee joint load prediction error to rise, Lookahead will backtrack to more stable candidate parameters to avoid the model falling into a local optimum.

[0064] The target deep model is generated by training the preset deep model based on the enhanced training data set, adaptive loss weight and optimized model parameters. The enhanced training data set is input into the preset deep model (such as a deep neural network), the loss value is calculated in combination with the adaptive loss weight, the parameters are iteratively updated by the Lookahead-AdamW optimizer, and the target deep model is generated until the evaluation index (such as accuracy and error rate) of the model on the validation set meets the standard.

[0065] The preset deep model is a neural network including a feature extraction layer, a fusion layer and an output layer, inputs the enhanced “electromyography + inertia + blood flow” fusion features, calculates classification loss (movement mode) and regression loss (joint load) through a dynamic loss function, and then optimizes the parameters by using Lookahead-AdamW. After 100 rounds of training, the classification accuracy of the model for the movement mode is increased from 75% to 92%, and the estimation error of the joint load is reduced from 12% to 4%, reaching the preset standard to generate the target deep model. The trained target deep model can effectively cope with the heterogeneity of multi-source signals, noise interference and individual differences, and provide high-precision model support for subsequent movement evaluation and injury warning.

[0066] In S107, the movement mode classification result, the joint load estimation data and the injury risk score result are processed based on the target deep model to generate injury warning information.

[0067] In an implementation, the movement mode classification result is processed based on the target deep model to identify the action standardization and generate an action correction prompt to provide action demonstration and feedback for the user. The processing of the movement mode classification result based on the target deep model is mainly to realize the identification, quantification and correction guidance of action deviation through the precise comparison between multi-source biological signal analysis and standard action template. The specific process and examples are as follows: The analysis of the target deep model depends on the movement mode classification result in the early stage, which is derived from the cooperative collection of multi-source sensing nodes: the inertial dynamics signals (acceleration and angular velocity) collected by the trunk nodes (such as the waist and back) are used to capture the overall body posture (such as the trunk inclination angle during the plank support and the spine curvature degree during the bending); the electromyography activity signals collected by the limb nodes (such as the arms and legs) are used to judge the muscle activation timing and intensity (such as the activation state of the triceps brachii and core muscle group during the plank support); the peripheral nodes assist in providing physiological load reference during the movement (such as the heart rate change reflecting whether the movement intensity affects the action stability). For example, in the “plank support” movement, the system first identifies the current action as “plank support” by the movement mode classification model, and calls the standard template of the action (including the trunk angle range, core muscle group activation intensity threshold and other parameters).

[0068] The target deep model compares the motion pattern classification result with the standard template through the following dimensions to identify the action deviation, spatial posture deviation: based on the inertial signal of the torso node, the difference between the spatial coordinates of the key parts of the body (such as shoulders, hips, ankles) and the standard coordinates is calculated, and the angle and displacement deviation (such as the angle between the trunk and the ground exceeding the standard range of 5°-0° during the plank support) is quantified; muscle force deviation: analyze the electromyographic signal of the limb node to determine whether the muscle activation sequence and intensity meet the standard (such as the core muscle group activation lagging behind the arm muscle during the plank support, causing the trunk to collapse); timing consistency deviation: through the timing features processed by the bidirectional gated recurrent unit, identify the action rhythm abnormality (such as the time ratio of squatting to standing during the squat deviating from the standard ratio of 1:1). Taking "plank support" as an example, the model calculates that the user's hip position is 10 cm lower than the shoulder (the standard is the same height of hip and shoulder), and combines the electromyographic signal to find that the rectus abdominis activation intensity is insufficient (only 60% of the standard value), and comprehensively determines the action deviation of "trunk collapse and core muscle insufficient".

[0069] For the identified deviation, the target deep model generates multi-level correction prompts and provides multi-channel feedback to quantitatively guide: clearly point out the specific value of the deviation and the adjustment target (such as "the trunk needs to be raised by 5 cm, making the shoulder-hip-ankle a straight line"); force prompt: combined with electromyographic signal analysis, guide the activation of key muscles (such as "tighten the transverse abdominis muscle and feel the inward contraction of the abdomen"); multi-modal feedback: synchronously push the standard action video (show the correct posture), real-time posture comparison chart (mark the deviation part), and vibration prompt (such as the waist sensor vibration reminding to adjust the trunk angle). For example, for the deviation of plank support, the system generates a comprehensive prompt: text: "trunk tilt 10°, please raise the hip to the same height as the shoulder, and continuously tighten the core"; video: play a standard plank support demonstration by a professional trainer, highlighting the trunk and core position; real-time monitoring: during the adjustment process, the trunk node signal is updated in real time, and when the angle returns to the standard range, the "action is correct, continue to maintain" is displayed. Through the above process, the target deep model realizes the closed loop from signal analysis to action correction, helps users optimize actions in motion, and reduces the risk of injury caused by non-standard actions.

[0070] The joint load estimation data is processed based on the target depth model, the change of joint stress is monitored in real time, and a load overrun warning is generated to prevent joint strain events. The generation of joint load estimation data is based on the fusion analysis of multiple source signals, which provides the input basis for the target depth model. Inertial dynamics signals: acceleration and angular velocity data collected by limb nodes (such as sensors near the knee joint), which are used to calculate the motion amplitude and speed of the joint during the action (such as the flexion and extension angle and rotation rate of the knee joint when stopping suddenly and changing direction); electromyographic activity signals: electromyographic signals of the quadriceps femoris on the front side of the thigh and the hamstrings on the back side, which reflect muscle contraction strength and indirectly derive the power source of joint stress (such as the positive correlation between muscle exertion peak and joint load);

[0071] Combined with human anatomy parameters (such as joint size and muscle attachment points), the motion signals are converted into actual joint load (such as converting acceleration signals into impact force borne by the knee joint by parameters such as lower limb length and body weight). Taking the "sudden stop and change direction" action of a basketball player as an example, the system calculates that the knee joint load at the moment of landing = body weight x impact force coefficient, where the impact force coefficient is determined by the vertical acceleration peak (3.2g) in the inertial signal and the muscle exertion coefficient (1.4), and finally the load is 4.5 times the body weight.

[0072] The safety threshold in the target depth model is not a fixed value, but is dynamically adjusted based on multiple factors. Action type: different movements have different load tolerance of joints (such as slow running, the knee joint safety threshold is 2.5 times the body weight, while the threshold of explosive actions such as sudden stop and change direction is increased to 3.5 times, because the muscle synergistic protection is stronger); individual differences: combined with user age, exercise experience, and previous injury history adjustment (such as young athletes' knee joint threshold can be slightly higher than that of middle-aged and elderly users, and the threshold of those with ligament injury history is reduced by 10%); real-time state: dynamically corrected according to the length of the movement (such as after 1 hour of continuous movement, muscle fatigue leads to decreased protection ability, and the threshold is automatically reduced by 5%-10%). For example, for the sudden stop and change direction action of a basketball player, the system default safety threshold is 3.5 times the body weight, and if the player has a recent record of knee strain, the threshold will be adjusted to 3.2 times, further reducing the risk of injury.

[0073] When the joint load exceeds the safety threshold, the target depth model generates an alert through a multi-level response mechanism. Instant reminder: Attract attention through vibration (3 short vibrations) of the wearable device (such as a knee sensor), and display a text alert (such as "knee load 4.5 times body weight, exceeding threshold 3.5 times") on the companion APP or smartwatch simultaneously; Cause analysis: Explain the cause of the load exceeding the limit (such as "knee joint inner buckle angle reaches 15° when changing direction, causing concentrated load"); Alternative solution: Recommend low-load actions (such as "use sliding movement instead, reduce knee joint twisting"), with action diagrams attached; Follow-up suggestion: Suggest short-term intervention measures (such as "suspend high-intensity training, apply cold compress to the knee after 5 minutes").

[0074] In an example, after receiving the alert, the basketball player can adjust the movement pattern according to the suggestion to avoid repeated exposure of the knee joint to excessive load, reducing the risk of cruciate ligament and meniscus injury. Through the above process, the target depth model realizes the "monitoring-evaluation-warning-intervention" closed loop of joint load, and converts biomechanical signals into executable protection strategies, providing precise injury prevention support for high-intensity sports.

[0075] Based on the target depth model, the injury risk score results are processed, combined with the risk threshold to generate injury warning prompts, and when the risk exceeds the limit, the intelligent intervention suggestion of ice compress is pushed. The core of the target depth model processing the injury risk score results is to realize the dynamic determination of risk level and the generation of targeted intervention suggestions through multi-dimensional signal time series analysis. The specific process and examples are as follows:

[0076] The injury risk score results are based on long-term tracking and fusion analysis of multi-source biological signals, providing a basis for the model to make judgments. The inertial dynamics signal of the trunk node records the bending angle, rotation amplitude and action frequency of the waist (such as the flexion angle of the lumbar spine when a worker bends over, the number of times the worker bends over per hour), reflecting the mechanical load accumulated by the waist. The electromyographic activity signal monitors the electrical activity intensity and fatigue of core muscles such as the psoas major muscle and erector spinae muscle (such as a decrease in electromyographic signal amplitude and a decrease in frequency after sustained effort), and assesses the protection ability of muscles to the spine. The hemodynamic signal is the pulse wave signal collected through the end node (such as the wrist), which analyzes the blood circulation state of the waist muscles (such as a decrease in blood flow velocity indicating local ischemia and accumulation of metabolic products).

[0077] Taking a worker as an example, the system calculates that the worker bends over 60 times in 1 hour (far exceeding the safety threshold of 40 times / hour), the psoas major muscle electromyographic signal fatigue index is 0.7 (threshold 0.5), and the lumbar flexion angle exceeds 30° multiple times (safe range <20°), and the system generates a lumbar risk score of 82 points.

[0078] The preset threshold is not a fixed value, but is dynamically adjusted in combination with multiple factors to ensure the accuracy of risk judgment. Part characteristics: the tolerance of different body parts is significantly different (such as the waist threshold set to 70 points due to the need for load bearing; the ankle joint has a large range of motion, and the threshold is set to 80 points); occupation / sports type: adjustment for high-risk scenarios (such as the waist threshold of a porter being 10% lower than that of the general population due to long-term repetitive bending movements); individual baseline: calibration combined with user historical data (such as a worker who previously had a waist score of 65 points experiencing pain, with his personal threshold lowered to 65 points). For example, the waist risk threshold for the general population is 70 points, while for porters who engage in long-term bending work, the system maintains the threshold at 70 points, but is more sensitive to trends in scores close to the threshold (such as an increase from 60 points to 82 points within 1 hour), triggering early warnings.

[0079] When the risk score exceeds the threshold, the target deep model generates multi-level intervention suggestions and pushes them through multiple channels, immediate stop prompts: explicitly terminate high-risk actions (such as "stop carrying objects immediately to avoid cumulative load on the waist"); physical intervention measures: recommend methods to alleviate acute injuries (such as "use an ice pack to cold compress the waist for 15 minutes, with a 2-hour interval between each time to reduce local inflammation"); rehabilitation action guidance: provide targeted stretching / relaxation programs (such as "waist around-the-ring stretching: slowly rotate clockwise and counterclockwise for 5 turns each, with an appropriate amplitude to avoid causing pain"); follow-up monitoring suggestions: prompt follow-up considerations (such as "avoid bending and bearing weight for 24 hours, and seek medical attention if there is a sharp pain").

[0080] In the example, after receiving the warning, the porter can reduce the likelihood of waist tissue damage through immediate intervention, while promoting recovery through subsequent rehabilitation actions, forming a "risk identification-intervention-rehabilitation" closed-loop protection. The target deep model converts abstract risk scores into actionable protection strategies, enabling early prevention of chronic injuries, especially for professional groups or sports enthusiasts who engage in long-term repetitive specific movements.

[0081] The action correction prompt, the load overrun warning and the injury early warning prompt are integrated and processed to generate comprehensive injury early warning information, and dynamic prediction and timely intervention of sports injury are realized. The action correction prompt, the load overrun warning and the injury early warning prompt are integrated to form comprehensive information covering action specification, load control and immediate intervention, and full-process injury prevention is realized. The comprehensive early warning information of the marathon runner includes the following: action correction: "when the right foot lands, the ankle is inwardly buckled, it is suggested to adjust the posture so that the toe points forward"; load warning: "the load on the left knee joint is greater than 2.8 times the body weight for 5 minutes in a row, it is suggested to reduce the pace"; intervention suggestion: "the current calf muscle risk score is 68 points (close to the threshold of 70 points), it is suggested to massage and relax after reaching the supply station". The system pushes the integrated information in real time through the APP to help the runner dynamically adjust the exercise state. The target deep model converts the dispersed evaluation results into executable early warning information, realizes the closed loop from "passive monitoring" to "active intervention", and effectively reduces the risk of sports injury.

[0082] The present application realizes dynamic injury early warning of sports through multi-source sensing and deep model fusion. Multi-modal signals such as inertial dynamics, electromyography, hemodynamics collected by the trunk, limbs and peripheral nodes are obtained; the features are generated through improved spectral reduction formula deartifacting, Lagrange kernel synchronization and other pretreatments; multi-scale hollow convolution, gate collaborative attention and other methods are used to construct cross-modal fusion features; sports mode, joint load and injury risk are output through classification model, load estimation model and risk score model; the signals are optimized through sensor calibration and multi-level suppression; the target deep model is trained through data enhancement, dynamic loss function and optimizer, and finally comprehensive early warning information such as action correction, load warning and injury intervention is generated. Through multi-modal signal fusion and deep model optimization, the system solves the problems of single-modal limitation and signal interference, realizes the precision and dynamics of sports evaluation and injury early warning.

[0083] In an embodiment, as shown in Figure 2 The present application also provides an intelligent sports evaluation and injury early warning device based on multi-source biosignal fusion, which comprises:

[0084] The acquisition module 201 is used for acquiring multi-modal biosignals collected based on multi-source sensing nodes;

[0085] The processing module 202 is configured to perform adaptive preprocessing on the multi-modal biological signals, remove motion artifacts by using an improved spectral subtraction formula, realize multi-rate synchronization by using a Lagrange kernel, and trigger a node vibration prompt to reattach electrodes by using signal quality evaluation, to generate preprocessed signal features; based on the preprocessed signal features, use multi-scale hollow convolution, gated collaborative attention, and bidirectional gated recurrent units to construct a feature fusion layer, to generate cross-modal fusion features; process the cross-modal fusion features to generate motion pattern classification results, joint load estimation data, and injury risk score results; process the original sensor acquisition data and signals in the preprocessing process, calibrate the sensor by using three-step static-dynamic-static calibration and temperature drift compensation, and generate motion artifact multi-level suppressed signal features by using IVA-Wiener-wavelet soft threshold-Kalman-NMF three-level suppression; based on the original data recorded by the multi-center heterogeneous recording device and the preprocessed signal features, cross-modal fusion features, decision reasoning intermediate features, and motion artifact multi-level suppressed signals, use data augmentation strategies, dynamic weighted loss functions, and Lookahead-AdamW optimizers to process a pre-set deep model, to generate a target deep model; and based on the target deep model, process the motion pattern classification results, joint load estimation data, and injury risk score results, to generate injury warning information.

[0086] The computer-readable storage medium provided by the above embodiments of the present application has the same beneficial effects as the methods used, run, or implemented by the application programs stored therein, based on the same inventive concept as the intelligent motion evaluation and injury warning method based on multi-source biological signal fusion provided by the embodiments of the present application.

[0087] Each of the embodiments in the present application is described in a related manner, and the same or similar parts of each of the embodiments can be referred to. Each of the embodiments focuses on the differences from other embodiments. In particular, the intelligent motion evaluation and injury warning method based on multi-source biological signal fusion, the electronic device, the electronic equipment, and the readable storage medium are basically similar to the above-described embodiments of the intelligent motion evaluation and injury warning method based on multi-source biological signal fusion, so the description is relatively simple, and the relevant parts can be referred to the above-described embodiments of the intelligent motion evaluation and injury warning method based on multi-source biological signal fusion.

Claims

1. A method for intelligent motion assessment and injury early warning based on multi-source biological signal fusion, characterized in that, include: Acquire multimodal biological signals based on multi-source sensing nodes, including trunk nodes, limb nodes and peripheral nodes, and multimodal biological signals including inertial dynamics, electromyographic activity and hemodynamic signals; Adaptive preprocessing of multimodal biological signals is performed, using an improved spectral subtraction formula to eliminate motion artifacts, multi-rate synchronization through the Lagrange kernel, and signal quality assessment to trigger node vibration to prompt re-attach electrodes, thereby generating preprocessed signal features. Based on the preprocessed signal features, a feature fusion layer is constructed using multi-scale dilated convolution, gated cooperative attention, and bidirectional gated recurrent units to generate cross-modal fusion features. The cross-modal fusion features are processed to generate motion pattern classification results, joint load estimation data, and injury risk score results. The original data acquired by the sensor and the signals during the preprocessing process are processed. The sensor is calibrated by three-step static-dynamic-static calibration and temperature drift compensation. The signal characteristics after multi-level suppression of motion artifacts are generated through IVA-Wiener-wavelet soft threshold-Kalman-NMF three-level suppression. Based on the raw data recorded by multiple centers heterogeneously and the preprocessed signal features, cross-modal fusion features, intermediate features for decision reasoning and signals after multi-level suppression of motion artifacts, a data augmentation strategy, a dynamic weighted loss function and a Lookahead-AdamW optimizer are used to process the preset depth model and generate the target depth model. Based on the target depth model, the motion pattern classification results, joint load estimation data, and injury risk score results are processed to generate injury early warning information.

2. The method as described in claim 1, characterized in that, Adaptive preprocessing of multimodal biological signals is performed, employing an improved spectral subtraction formula to eliminate motion artifacts, using the Lagrange kernel to achieve multi-rate synchronization, and leveraging signal quality assessment to trigger node vibration to prompt electrode reattachment, generating preprocessed signal features, including: Adaptive preprocessing of multimodal biological signals is performed using an improved spectral subtraction formula S. proc (ξ)=(S sig (ξ) 2 -γS bkg (ξ) 1 / 2 e iψ(ξ) Motion artifacts were eliminated, and the P-QRS-T waveform of the electrocardiogram was restored under the condition of 15km / h. Multi-rate synchronization is achieved by dynamically adjusting the order d = [log2ρ] + 1 using a variable pseudo-Lagrange kernel, enabling sub-millisecond alignment of 500Hz EMG and 30Hz PPG mixing sampling. By using signal quality assessment to trigger node vibration prompts for electrode reattachment, preprocessed signal characteristics are generated.

3. The method as described in claim 2, characterized in that, Based on the preprocessed signal features, a feature fusion layer is constructed using multi-scale dilated convolution, gated cooperative attention, and bidirectional gated recurrent units to generate cross-modal fused features, including: Based on the preprocessed signal features, multi-scale dilated convolution is used to capture the step frequency micro-changes within a 1.25s window to achieve real-time prompting of fatigue inflection point and generate multi-scale spatiotemporal features. Using gated collaborative attention via formula ω ab =σ(P T tanh(Qz MG +Qz MU Determine the synchronization degree between peak electromyography (EMG) values ​​and joint angular acceleration, and generate cross-modal correlation weight features; Using a bidirectional gated loop unit via formula g t =σ(A g [s t-1 ;y t ]+c g Continuously track a 30-minute motion sequence and identify abnormal action patterns to generate time-dependent features; A feature fusion layer is constructed based on multi-scale spatiotemporal features, cross-modal association weight features, and time-series dependent features to generate cross-modal fused features.

4. The method as described in claim 3, characterized in that, The cross-modal fusion features are processed to generate motion pattern classification results, joint load estimation data, and injury risk scores, including: Cross-modal fusion features are processed and classified using a motion pattern classification model based on a formula. It identifies the types of movement actions, generates movement pattern classification results, and provides demonstrations and feedback for action switching; The cross-modal fusion features are processed, and the joint load estimation model is based on the formula. Calculate the actual stress on the joint and generate joint load estimation data to monitor joint load and prevent sprain events; The cross-modal fusion features are processed and applied to the damage risk scoring model based on the formula. Dynamically monitor pressure at specific locations, generate damage risk scores, and intelligently push intervention prompts when the risk threshold is exceeded; Based on motion pattern classification results, joint load estimation data, and injury risk score results, the cross-modal fusion features are processed to generate core decision data for motion assessment and injury early warning.

5. The method as described in claim 1, characterized in that, The sensor's raw data acquisition and preprocessing signals are processed. A three-step static-dynamic-static calibration and temperature drift compensation method is used to calibrate the sensor. A three-level suppression process—IVA-Wiener-wavelet soft threshold-Kalman-NMF—is employed to generate signal characteristics after multi-level suppression of motion artifacts, including: The sensor's raw data and signals during the preprocessing process are processed. The sensor is calibrated using a three-step static-dynamic-static calibration and temperature drift compensation method to ensure that the calibration residual is <62ns, and calibrated sensor data is generated. The calibrated sensor data is processed, and initial artifact separation is achieved through IVA-Wiener filtering to generate a first-level suppression signal; The primary suppression signal is processed, and high-frequency noise interference is further eliminated by wavelet soft threshold filtering to generate the secondary suppression signal; The secondary suppression signal is processed, and the Kalman-NMF algorithm is used to optimize signal components and remove artifacts at a depth of 3, generating a tertiary suppression signal. Based on the calibrated sensor data and the first, second, and third level suppression signals, multi-level motion artifact suppression processing is completed, generating the signal characteristics after multi-level motion artifact suppression.

6. The method as described in claim 5, characterized in that, Based on raw data recorded from multiple centers heterogeneously, preprocessed signal features, cross-modal fusion features, intermediate features for decision inference, and signals after multi-level suppression of motion artifacts, a data augmentation strategy, a dynamic weighted loss function, and the Lookahead-AdamW optimizer are used to process a pre-defined depth model to generate a target depth model, including: Based on the raw data recorded by multiple centers heterogeneously and the preprocessed signal features, cross-modal fusion features, intermediate features for decision reasoning and signals after multi-level suppression of motion artifacts, data augmentation strategies are used to expand the training data, including time scaling, bandwidth permutation and Cross-SensorCutMix, to generate an augmented training dataset. A dynamic weighted loss function is used to dynamically adjust the model training loss and generate adaptive loss weights. The Lookahead-AdamW optimizer was used to optimize the model parameters, generating optimized model parameters; Based on the enhanced training dataset, adaptive loss weights, and optimized model parameters, a preset deep model is trained to generate a target deep model.

7. The method as described in claim 6, characterized in that, Based on the target depth model, the motion pattern classification results, joint load estimation data, and injury risk score results are processed to generate injury early warning information, including: The motion pattern classification results are processed based on the target depth model to identify the norms of the movements and generate movement correction prompts, providing users with movement demonstrations and feedback. The target depth model is used to process the joint load estimation data, monitor the changes in joint force in real time and generate overload warnings to prevent joint strain events. The damage risk score results are processed based on the target depth model, and damage warning prompts are generated by combining the risk threshold judgment. When the risk exceeds the limit, intelligent intervention suggestions for ice application are pushed. The system integrates and processes movement correction prompts, overload warnings, and injury early warnings to generate comprehensive injury early warning information, enabling dynamic prediction and timely intervention of sports injuries.

8. A smart motion assessment and injury early warning device based on multi-source biosignal fusion, characterized in that, The device includes: The acquisition module is used to acquire multimodal biological signals collected by multi-source sensing nodes; The processing module adaptively preprocesses multimodal biological signals, employing an improved spectral subtraction formula to eliminate motion artifacts, using Lagrange kernels for multi-rate synchronization, and triggering node vibrations based on signal quality assessment to prompt electrode reattachment, generating preprocessed signal features. Based on these preprocessed features, a feature fusion layer is constructed using multi-scale dilated convolution, gated collaborative attention, and bidirectional gated recurrent units to generate cross-modal fusion features. These cross-modal fusion features are then processed to generate motion pattern classification results, joint load estimation data, and injury risk scores. Finally, the module processes the raw sensor data and signals acquired during preprocessing, employing a three-step static-dynamic-static calibration process. The sensor is calibrated with temperature drift compensation. A three-level suppression process (IVA-Wiener-wavelet soft threshold-Kalman-NMF) is used to generate signal features after multi-level suppression of motion artifacts. Based on the raw data recorded by multiple centers heterogeneously, the preprocessed signal features, cross-modal fusion features, intermediate features for decision inference, and the signal after multi-level suppression of motion artifacts, a data augmentation strategy, a dynamic weighted loss function, and a Lookahead-AdamW optimizer are employed to process the preset depth model and generate a target depth model. Based on the target depth model, the motion pattern classification results, joint load estimation data, and injury risk score results are processed to generate injury warning information.

9. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the intelligent motion assessment and injury warning method based on multi-source biosignal fusion as described in any one of claims 1 to 7 by executing the executable instructions.

Images