Muscle activation quality joint evaluation method and system based on rhythm cycle normalization

Abstract

This application relates to a method and system for joint assessment of muscle activation quality based on rhythmic cycle normalization, belonging to the field of multidimensional motion quality assessment. Addressing the problem that multi-source signals cannot be directly compared due to differences in physical duration across different cycles in periodic motion, this application uses the rhythmic cycle as a unified phase benchmark, mapping surface electromyography (EMG) signals and inertial sensing signals of different physical durations to a unified normalized phase domain, constructing a standardized phase framework for signal comparability across cycles and speeds. Under this unified phase framework, quality assessment is performed on dimensions such as muscle activation timing, peak phase position, left-right symmetry, periodic stability, and fatigue evolution, outputting a joint assessment result including multi-dimensional quantitative scores and grading results, achieving a refined and interpretable assessment of motion quality.

Description

Technical Field

[0001] This application relates to the field of multidimensional motion quality assessment, specifically to a method and system for joint assessment of muscle activation quality based on rhythmic cycle normalization. Background Technology

[0002] With increased public awareness of health and the development of wearable technology, various sports monitoring products have been widely used. However, most existing products can only output macroscopic sports parameters such as cadence, pace, power, or simple repetitions. While these parameters have some training reference value, they essentially only focus on the "quantity" of exercise—how fast or how many repetitions—and cannot deeply assess the "quality" of exercise. For example, they fail to address core issues that are closer to the body's functional state, such as whether the force application pattern of each movement is correct, whether the left and right limbs are coordinated and symmetrical, whether the movement process is stable, and whether physiological fatigue is accumulating. This results in users and coaches lacking effective tools to assess exercise quality, optimize movement patterns, and prevent injuries.

[0003] To achieve deeper analysis of motion quality, some advanced research approaches attempt to combine surface electromyography (sEMG) signals with inertial measurement unit (IMU) signals. However, existing techniques have significant drawbacks: First, there is a lack of a unified time reference for organizing the analysis of these signals. In periodic movements such as running and cycling, without using a single step or pedal cycle as a unified normalization framework, it is difficult to directly compare the activation timing and amplitude of the same muscle under different rhythms, leading to a lack of stability and reliable physiological interpretation of the analysis results. Second, rhythm detection and quality analysis are functionally separated. Inertial sensors are typically used only to measure rhythm, while the characteristic analysis of EMG signals (such as calculating integrated EMG values) exists as an isolated indicator. There is no unified evaluation system driven by rhythmic cycles and with intrinsic data correlation between the two. This separation makes it difficult to transform complex sensor signals into real-time, easily understandable, multi-dimensional evaluation results for ordinary users or rehabilitation scenarios, greatly limiting their application value.

[0004] Therefore, the industry hopes to develop an evaluation scheme that can use movement rhythm as the core time reference, perform periodic normalization and joint modeling of electromyography and inertial signals, and output multi-dimensional movement quality scores. Summary of the Invention

[0005] This application provides a method and system for joint evaluation of muscle activation quality based on rhythmic cycle normalization, in order to at least solve the above-mentioned technical problems existing in the prior art.

[0006] According to a first aspect of this application, a method for joint assessment of muscle activation quality based on rhythmic cycle normalization is provided, comprising: Step 1, based on the start and end times of each cycle in a continuous motion rhythmic cycle sequence, performing cycle truncation and time coordinate normalization mapping on the surface electromyography (EMG) signals and inertial sensing signals collected within the corresponding time period to obtain normalized cycle data; Step 2, based on the normalized cycle data, performing multidimensional feature extraction on the surface EMG signals of the target muscle group within each normalized cycle to obtain EMG feature vectors, wherein the EMG feature vectors include intensity features, temporal features, periodic stability features, left-right symmetry features, and fatigue-related features; Step 3 Third, based on normalized periodic data, kinematic co-feature extraction is performed on the inertial sensing signals corresponding to each period to obtain inertial consistency feature vectors; fourth, cross-dimensional fusion calculation is performed on the electromyographic feature vectors and inertial consistency feature vectors to obtain multidimensional scoring results, which include quantitative scores for rhythm quality, activation intensity, movement stability, left-right symmetry, and fatigue risk dimensions; fifth, based on preset level threshold classification rules and continuous window time series analysis methods, level mapping and trend evolution identification are performed on the multidimensional scoring results to generate a joint evaluation result containing level classification results, risk level results, and trend indication information.

[0007] According to a second aspect of this application, a joint assessment system for muscle activation quality based on rhythmic period normalization is provided, comprising: a signal normalization module, used to perform periodic interception and time coordinate normalization mapping on surface electromyography (EMG) signals and inertial sensing signals collected within corresponding time periods based on the start and end times of each period in a continuous motion rhythmic periodic sequence to obtain normalized periodic data; an EMG feature extraction module, used to perform multi-dimensional feature extraction on the surface EMG signals of the target muscle group within each normalized period based on the normalized periodic data to obtain EMG feature vectors, the EMG feature vectors including intensity features, temporal features, periodic stability features, left-right symmetry features, and fatigue-related features; and an inertial consistency processing module. The system consists of three modules: a kinematic feature extraction module for extracting kinematic co-features from inertial sensing signals corresponding to each cycle based on normalized periodic data to obtain inertial consistency feature vectors; a multidimensional scoring module for cross-dimensional fusion of electromyographic feature vectors and inertial consistency feature vectors to obtain multidimensional scoring results, including quantitative scores for rhythm quality, activation intensity, movement stability, left-right symmetry, and fatigue risk; and an evaluation result generation module for mapping and identifying trends in the multidimensional scoring results based on preset level threshold classification rules and continuous window time series analysis methods, generating a joint evaluation result that includes level classification results, risk level results, and trend indication information.

[0008] Compared with existing technologies, this application constructs a joint assessment method and system for muscle activation quality based on rhythmic cycle normalization. Addressing the fundamental obstacle in existing technologies where the physical duration differences of different movement cycles prevent direct comparison of multi-source signals, this application uses the rhythmic cycle as a unified phase benchmark, mapping cycle data with varying durations due to factors such as speed variations and fatigue accumulation to standardized normalized phase coordinates, thus enabling comparability of signals across cycles and speeds. Within this unified phase framework, this application further systematically assesses the quality of muscle activation timing, peak phase position, left-right symmetry, cycle stability, and fatigue evolution trends. Combined with the collaborative analysis of inertial signals in the phase domain, it outputs a joint assessment result including multi-dimensional quantitative scores, grade levels, and trend indications. This application achieves refined and interpretable assessment of movement quality and is applicable to various periodic movement scenarios such as running, cycling, and rehabilitation training. Attached Figure Description

[0009] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, wherein: in the drawings, the same or corresponding reference numerals denote the same or corresponding parts.

[0010] Figure 1 This is a flowchart of a joint assessment method for muscle activation quality based on rhythmic cycle normalization according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of the joint assessment method for muscle activation quality based on rhythmic cycle normalization according to an embodiment of this application; Figure 3 This is a flowchart of step three in the joint assessment method for muscle activation quality based on rhythmic cycle normalization according to the embodiments of this application; Figure 4 This is a flowchart of step five in the joint assessment method for muscle activation quality based on rhythmic cycle normalization according to the embodiments of this application; Figure 5 This is a block diagram of a joint assessment system for muscle activation quality based on rhythmic cycle normalization, according to an embodiment of this application. Detailed Implementation

[0011] To further illustrate the technical means and effects adopted by this application in order to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of this application is provided in conjunction with the accompanying drawings and preferred embodiments.

[0012] To address the shortcomings in the aforementioned technical field, this application provides a joint assessment method for muscle activation quality based on rhythmic cycle normalization, such as... Figure 1 and Figure 2 As shown, Figure 1 This is a flowchart of a joint assessment method for muscle activation quality based on rhythmic cycle normalization according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow of the joint evaluation method for muscle activation quality based on rhythmic cycle normalization according to an embodiment of this application.

[0013] Specifically, in step one, based on the start and end times of each cycle in the continuous motion rhythmic cycle sequence, the surface electromyography (EMG) signals and inertial sensing signals collected within the corresponding time period are periodically truncated and normalized to obtain normalized cycle data. It should be understood that during periodic sports activities, such as running or cycling, although the movement patterns are repetitive, the physical duration of each independent movement cycle, such as each step or each pedaling motion, varies subtly and dynamically. This inconsistency in cycle duration caused by factors such as speed, fatigue, and terrain leads to a loss of comparability between EMG and kinematic signals in different cycles on the absolute time axis. Direct feature extraction and analysis would introduce significant errors, making it impossible to stably assess muscle activation patterns and movement quality. Therefore, step one in this technical solution constructs a unified, physical duration-independent relative time reference, mapping all movement cycles to standardized phase coordinates from 0% to 100%. This fundamentally eliminates the interference caused by differences in cycle duration, providing a stable and reliable data foundation for subsequent analysis of muscle function consistency, symmetry, and fatigue across cycles, speeds, and even individuals. It is a core prerequisite for achieving accurate and interpretable quantitative assessment of exercise quality.

[0014] In one specific implementation, step one includes: based on the start and end timestamps of each cycle in the continuous motion rhythm cycle sequence, performing temporal alignment and cycle segmentation on the surface electromyography signals and inertial sensing signals within the corresponding time period to obtain an original set of cycle segments; based on a preset number of normalized fixed sampling points and the physical duration of each original cycle segment, performing proportional phase coordinate mapping on each segment in the original set of cycle segments to obtain a cycle segment with mapped coordinates; and generating normalized cycle data based on the cycle segments with mapped coordinates.

[0015] The process is as follows: The initial input data for processing consists of two parts: a continuous motion rhythm cycle sequence, and surface electromyography (EMG) signals and inertial sensing signals acquired synchronously. The continuous motion rhythm cycle sequence is a set of structured time-stamped data that records the precise start and end times of each repetitive motion cycle during continuous movement. This sequence can be obtained through real-time analysis of inertial sensing signals, or it can be provided by an independent external rhythm detection module, an external sensing system, or a manual annotation system. This application does not limit the specific method of obtaining the rhythm cycle sequence. For example, in a running scenario, data from an inertial measurement unit (including a triaxial accelerometer and a triaxial gyroscope) installed on the foot is used to identify the precise time point of each heel strike using peak detection or zero-point crossover algorithms. The time points between two consecutive heel strikes constitute one gait cycle. The specific data structure of this sequence can be represented as a timestamp array, where each element contains the start and end timestamps of a period, for example, [{start: 10.35s, end: 11.10s}, {start: 11.10s, end: 11.83s},...]. Surface electromyography (EMG) signals and inertial sensing signals are high-frequency sampled continuous time-series data streams. For example, EMG signals record the electrical activity amplitude of target muscle groups such as the quadriceps and gastrocnemius muscles at a sampling rate of 1000 Hz, while inertial sensing signals record the acceleration and angular velocity of limb segments at a sampling rate of 200 Hz. These two signals are strictly time-synchronized during acquisition via hardware or software.

[0016] In the first processing stage, namely sequence alignment and periodic segmentation, the program iterates through the received continuous motion rhythm periodic sequence. For the k-th period in the sequence, its start timestamp is extracted. With end timestamp Using these two timestamps as boundaries, precise segmentation is performed within the synchronous, continuous surface electromyography (SEMG) and inertial sensing signal data streams. For example, if the start timestamp of the k-th gait cycle is 10.35 seconds and the end timestamp is 11.10 seconds, then the physical duration of this cycle is 0.75 seconds. The processing program will precisely extract all SEMG signal sampling points within this 0.75-second time window, i.e., 0.75s × 1000Hz = 750 points, and inertial sensing signal sampling points, i.e., 0.75s × 200Hz = 150 points. The extracted signal data from each channel within this cycle is then compared with the physical duration of the cycle. Structured encapsulation is performed to form a set of original periodic segments.

[0017] In the second processing stage, namely the proportional phase coordinate mapping, a fixed number of normalized sampling points N needs to be pre-defined to construct a relative percentage progress axis proportionally distributed from 0% to 100%. The selection of this value needs to balance signal resolution and computational efficiency. In one embodiment, N=201, which is equivalent to resolving the progress of each motion cycle into 201 equidistant phase nodes, with a resolution of 0.5%. This setting means that regardless of whether the actual physical duration of a cycle is 0.75 seconds or 0.73 seconds, after normalization, it will be uniformly represented by 201 data points. Subsequently, for each segment in the original set of cycle segments, based on its own physical duration... Given a preset value of N, a new set of target mapping timestamps on the original physical timeline is calculated. Specifically, for the i-th normalized sampling point in the k-th period, where i ranges from 0 to N-1, its corresponding target mapping timestamp on the original physical timeline is calculated. It can be calculated using the following linear mapping formula: In this formula, This represents the target mapping timestamp corresponding to the ith normalized sampling point in the kth period; It is the starting physical timestamp of the k-th cycle; It is the index of the normalized sampling points, counting from 0; It is the preset normalized fixed total number of sampling points, i.e., 201; This is the original physical duration of the k-th cycle. Essentially, this formula stretches or compresses the discrete indices from 0 to N-1 proportionally to a length of... Within the physical time interval. Taking the aforementioned period of 0.75 seconds as an example, its 101st normalized sampling point, i=100, represents the target mapped timestamp corresponding to 50% of the period phase, which is 10.35s + 100 / 200 × 0.75s = 10.725s. Through this calculation, the precise projection of the 201 normalized phase points within this period onto the original physical time has been found. Binding the resulting array of target mapped timestamps with the corresponding original period segment data yields the period segment with mapped coordinates.

[0018] In the third processing stage, namely, interpolation and resampling to generate normalized periodic data, the program will use the periodic segments with mapped coordinates generated in the previous step to complete the final signal reconstruction. In one specific implementation, generating normalized periodic data based on the periodic segments with mapped coordinates includes: interpolating and resampling the signal amplitude at the corresponding positions of each target mapped timestamp in the periodic segments to obtain normalized periodic data. This is because the calculated target mapped timestamps... Typically, the time points of the target mapping will not perfectly coincide with the discrete sampling time points of the original signal; therefore, an interpolation algorithm is needed to estimate the signal amplitude at that time point. This embodiment uses linear interpolation. Specifically, for a given target mapping timestamp... First, find two real sampling points that are immediately adjacent in time in the original signal sequence, according to... The amplitude is calculated proportionally based on the relative position within this interval. For example, for surface electromyography (EMG) signals, the sampling interval is 0.001 seconds. If a target mapping timestamp is 10.7256 seconds, and the two nearest real sampling points before and after it are at 10.725 seconds (amplitude 0.15 mV) and 10.726 seconds (amplitude 0.18 mV), the interpolation result is 0.15 + (0.18 - 0.15) × ((10.7256 - 10.725) / (10.726 - 10.725)) = 0.168 mV. Repeating this interpolation process for all 201 target mapping timestamps in one cycle generates a new time series of fixed length 201 points for each channel of the EMG and inertial sensing signals. This new sequence is detached from absolute physical time, and the index of its data points (0 to 200) directly corresponds to the relative phase of the motion cycle (0% to 100%), resulting in normalized periodic data.

[0019] In particular, a key step in the joint assessment of muscle activation quality based on the normalization of movement rhythm cycles is to convert each movement cycle segment, which has a different physical duration, into a normalized cycle data with a uniform fixed length, in order to facilitate feature comparison and model analysis across cycles. The above implementation typically relies on standard linear interpolation algorithms; however, this method contains a fundamental assumption that the signal changes linearly between any two adjacent original sampling points. When processing typical bioelectrical signals such as surface electromyography (SEMG), this assumption reveals its fundamental flaws. SEMG signals, especially in periodic movements, exhibit strong non-stationarity and impulsive characteristics: the signal amplitude changes drastically within a very short time at the beginning and end of muscle exertion; at the peak of muscle force output, the signal forms a sharp convex arc. These key kinematic event regions all possess extremely high nonlinear characteristics (i.e., high curvature). Linear interpolation, unable to perceive the local curvature of the signal, forcibly straightens these important nonlinear segments, leading to systematic signal morphology distortion. This manifests as flattening of activation peaks (resulting in underestimated intensity) and passivation of transient gradients in the start / end phases (leading to inaccurate timing localization). This distortion is not random noise but a deterministic bias directly related to the inherent nonlinearity of the signal. It propagates and amplifies downstream feature extraction processes, ultimately compromising the reliability and accuracy of the assessment results. Therefore, to address the structural defects in morphological fidelity in the core region of electromyography activation caused by linear approximation, a curvature-compensated adaptive interpolation process is introduced, capable of dynamically sensing and compensating for signal nonlinearity. This curvature-compensated adaptive interpolation estimates the local curvature field of periodic segments with mapped coordinates and adaptively compensates for standard linear interpolation based on the estimation results. Finally, it generates high-quality normalized periodic data through fidelity verification.

[0020] In another specific implementation, normalized periodic data is generated based on periodic segments with mapped coordinates, including: Local curvature field estimation is performed on periodic segments with mapped coordinates to obtain a local curvature field sequence. It should be understood that the accuracy of interpolation is closely related to the degree of local nonlinearity of the signal, and the second derivative of the signal (i.e., curvature) is the most direct physical quantity to measure its local bending or nonlinearity. By pre-quantifying and estimating the local curvature of the signal around each target mapping point before interpolation, a precise and adaptive correction basis can be provided for subsequent interpolation compensation, generating a curvature field sequence that corresponds one-to-one with the target mapping timestamp, much like terrain. Figure 1Generally, this accurately depicts the nonlinear topography of the original signal at each phase point, making the interpolation process no longer a blind first-order approximation, but an intelligent reconstruction that can perceive the signal morphology. Specifically, the input to this stage is a periodic segment with mapped coordinates, which contains the original high-frequency sampled signal sequence and a set of target mapped timestamps that need to be interpolated on the original time axis. Where k is the period index and i is the normalized sampling point index. For the i-th target mapping timestamp in the k-th period... First, locate the preceding real sampling point in the original sampling sequence that is immediately adjacent to it in time. and the next real sampling point Subsequently, using and its adjacent sampling points and The second derivative of the signal at that location, i.e., the local curvature, is approximately estimated using the three-point central difference method. The calculation formula is as follows: In this formula, It is the estimated local curvature value; , and These are the signal amplitudes recorded at these three consecutive real sampling points; It is the average physical time interval between sampling points within the neighborhood of these three points, i.e. For example, when analyzing a running gait cycle, the target is mapped to a timestamp. =10.7514 seconds falls within a region of peak electromyography (EMG). The original sampling data for the vicinity are as follows: =10.750 seconds, =0.40 millivolts; =10.751 seconds, =0.55 millivolts; =10.752 seconds, =0.45 millivolts. The sampling interval is 0.001 seconds, therefore =0.001 seconds. Substituting into the formula, we get: Unit: millivolts per second 2 This large negative value precisely quantifies the sharp downward curvature of the signal at that point. Repeating this calculation for all target-mapped timestamps yields the local curvature field sequence for that cycle. In the signal morphology of the actual motion cycle, this curvature field sequence exhibits distinct regional differentiation characteristics: in the region of the muscle's primary activation peak, The large negative value reflects the severity of the signal's convex curve bending; in the ramp region of the onset transition, κ... It exhibits high-frequency fluctuations with alternating positive and negative values, reflecting the rapid reversal of the signal curvature direction; while in the resting, flat region, The value approaches zero, indicating that the signal has almost no nonlinear bending component at this point.

[0021] Curvature-compensated adaptive interpolation resampling is performed on the local curvature field sequence and the periodic segments with mapped coordinates to obtain a curvature-compensated interpolation signal sequence. This step aims to utilize the local curvature information generated in the previous step to dynamically compensate the standard linear interpolation result, correcting its approximation error in high curvature regions. The core idea is to upgrade the interpolation model from first-order linear to second-order (quadratic or parabolic), allowing the interpolation curve to better track the curvature of the original signal. Thus, in the electromyography peak region, the interpolation result is actively raised to counteract the peak truncation effect; in abrupt transition regions such as onset / offset, the interpolation curve can better fit its step shape, improving timing accuracy; and in flat resting regions, the compensation term automatically decays, avoiding unnecessary corrections. Finally, a curvature-compensated interpolation signal sequence that is morphologically closer to the original signal is obtained. Specifically, the local curvature field sequence and the periodic segments with mapped coordinates generated in the previous step are received. For the i-th target mapped timestamp in the k-th period... The amplitude of the interpolated signal after compensation It is calculated using a formula that includes linear fundamental components and a second-order curvature compensation factor: The first part on the right side of the equation is the standard linear interpolation result. The second part is the second-order curvature compensation factor, which adjusts the local curvature. With a quadratic term Multiplication. The quadratic term is in and The value is zero at the given point and reaches an extreme value between the two, its sign is zero. The combination of these factors determines the direction and magnitude of the compensation. The above formula superimposes a second-order radian correction term driven by local curvature onto the original linear interpolation's first-order chord approximation. In the compensation factor... The product of these terms is always negative within the subinterval (because...). Its absolute value reaches its maximum at the midpoint of the sub-interval. When the target point falls into the convex peak region of electromyographic activation... The result is a large negative value. Multiplying two negative values ​​causes the compensation term to automatically generate a positive correction, which precisely compensates for the peak amplitude loss caused by the truncation of the linear interpolation due to the substitution of arcs for chords. When the target point is in a resting flat region, As the value approaches zero, the entire compensation term automatically decays to a negligible level. At this point, the formula degenerates into a standard linear interpolation result, thus avoiding over-correction of low-curvature regions. Continuing with the previous example, the target timestamp... =10.7514 seconds, =10.751 seconds, =10.752 seconds. Linear interpolation fundamental component = millivolts. Second-order curvature compensation factor = millivolts. Finally, the interpolation result after curvature compensation. =0.51 + 0.03 = 0.54 millivolts. This result is closer to the 0.51 millivolts obtained through linear interpolation. The peak value at that location is 0.55 mV, which effectively corrects the underestimation of the peak value.

[0022] Point-by-point fidelity verification and normalization encapsulation are performed on the curvature-compensated interpolation signal sequence and the local curvature field sequence to obtain normalized periodic data. Although second-order compensation can significantly improve the interpolation accuracy in most regions, residual errors may still exist in a few regions where the signal has more complex third-order or higher-order nonlinear characteristics (such as transient jumps at the superposition of motion artifacts). This step sets up a quality control checkpoint to perform a final verification of the fidelity of each interpolation point, identifying and handling potential extreme outliers. Its effect is to provide a quantitative quality assurance for the entire normalized output, controlling the residual morphological distortion within a preset tolerance range, and ensuring that the data delivered to the downstream feature extraction stage has extremely high reliability and consistency. Specifically, the curvature-compensated interpolation signal sequence and the local curvature field sequence generated in the previous step are received. For each interpolation point, its maximum residual approximation error is estimated based on interpolation theory. This error is related to the third derivative of the signal, and its theoretical upper bound can be estimated by the following formula: Among them, the third derivative of the signal By analyzing the curvature field sequence The approximation is obtained by performing another difference. The program will then set an upper bound on the residual error calculated for each interpolation point. It is compared with a preset fidelity tolerance threshold. This threshold is set according to the signal accuracy requirements of the application scenario; for example, it can be set to 0.5% of the signal dynamic range. In most areas, after second-order compensation... The error will be far below the threshold, thus passing the verification. If the error exceeds the limit at a certain point, it indicates that there is a complex shape that the second-order model cannot capture. This can trigger an alarm flag or enable a higher-order (such as cubic spline) interpolation strategy to locally reconstruct the point. All verified (or repaired) interpolation points are uniformly packaged and encapsulated in a one-dimensional array of equal length with a fixed number of normalized sampling points (N=201). The final output is normalized periodic data that achieves a systematic improvement in morphological fidelity.

[0023] Specifically, in step two, based on normalized periodic data, multidimensional features are extracted from the surface electromyography (EMG) signals of the target muscle group within each normalized period to obtain EMG feature vectors. These EMG feature vectors include intensity features, temporal features, periodic stability features, left-right symmetry features, and fatigue-related features. Correspondingly, compared to traditional monitoring schemes that only provide macroscopic kinematic indicators such as cadence and speed, a deeper assessment of muscle activation status can better reveal the intrinsic quality of movement. The previous step established a unified time benchmark for cross-period comparisons through normalization, but this only solves the data organization problem. The normalized signal itself is still an abstract waveform and cannot directly answer key functional questions such as whether force exertion is sufficient, whether the movement is stable, whether left-right symmetry is maintained, and whether fatigue has accumulated. In order to transform these abstract and standardized time-series data into quantitative indicators that can intuitively reflect the physiological functional state and have clear physical meaning, step two in this technical solution decodes the core physiological information that represents activation intensity, temporal sequence, periodic stability, limb symmetry and fatigue evolution trend from the normalized electromyographic signal, thereby constructing a rich and structured electromyographic feature vector.

[0024] In one specific implementation, step two includes: based on the surface electromyography (EMG) signal amplitude sequence in the normalized periodic data, performing root mean square energy calculation and activation sequence identification on the activation signals of the target muscle group in each normalized periodic segment to obtain a single-cycle basic feature set; based on the feature value sequence of multiple consecutive periodic segments in the single-cycle basic feature set, performing cross-cycle statistics on the periodic variation coefficient, left-right intensity asymmetry index, and fatigue evolution slope to obtain a cross-cycle high-level feature set; and performing feature splicing and dimensional standardization processing on the single-cycle basic feature set and the cross-cycle high-level feature set to obtain an EMG feature vector.

[0025] The process is as follows: In the first processing stage, namely generating the single-cycle basic feature set, the processing program iterates through each cycle in the normalized periodic data. Taking the k-th cycle as an example, it first isolates the surface electromyographic signal sequence of the target muscle group, such as the left quadriceps femoris, within that cycle. This is a one-dimensional array of length N, which is 201 in this example, denoted as... Subsequently, two types of core fundamental features are calculated based on this sequence. The first type is intensity features, used to quantify the overall activation level of the muscle within that cycle. A core intensity metric is the root mean square (RMS) energy value, which reflects the average power of the signal. Its calculation formula is: In this formula, Let be the root mean square value of the k-th period. To normalize the number of sampling points to 201, This represents the electromyographic signal amplitude at the j-th normalized sampling point within the k-th cycle. The calculation process squares, averages, and then takes the square root of the signal amplitude for the entire cycle to obtain a scalar that stably represents the total force exertion during that cycle. For example, in a running cycle, the calculated amplitude of the left quadriceps femoris... The value is 0.23 mV. The second type is temporal features, used to describe the activation time points and duration of muscles within a cycle. To obtain this feature, the original normalized electromyography (EMG) signal first needs to be preprocessed to obtain its envelope, using a method of full-wave rectification followed by low-pass filtering. After obtaining the envelope, an adaptive activation threshold is set to distinguish between active and inactive segments. This threshold can be set as a fixed percentage of the peak value of the EMG signal envelope within the cycle, such as 20%, to accommodate signal variations under different intensities. The muscle activation start time, Onset, is defined as the normalized time point when the signal envelope first crosses the threshold upwards, while the activation end time, Offset, is defined as the time point when the signal envelope last crosses the threshold downwards. For example, if the peak value of the quadriceps EMG envelope in a certain cycle is 0.45 mV, then the activation threshold is 0.09 mV. The program detects that the envelope first exceeds 0.09 mV at 12.5% ​​of the cycle phase corresponding to the 25th sampling point, and finally falls back below this threshold at 77.5% of the cycle phase corresponding to the 155th sampling point. Therefore, the activation start time of this cycle is 12.5%, and the end time is 77.5%. The intensity features (such as RMS values) and temporal features (such as start and end times) calculated for this cycle are structurally combined to form a single-cycle basic feature set, for example, {RMS:0.23, Onset:12.5, Offset:77.5}. In specific application scenarios, these temporal features have the potential for further drill-down analysis. For example, in running scenarios, the system can further analyze the pre-activation and delay of muscles before ground contact, during the support phase, and before and after the swing phase; in cycling scenarios, the system can further analyze the proportion of the main activation peak falling within the power exertion phase and the pullback / lift phase, thereby determining whether the pedaling power window has shifted.

[0026] In the second processing stage, namely the generation of cross-cycle high-level feature sets, the program will no longer be limited to a single cycle, but will focus on the dynamic relationships and evolutionary trends between consecutive cycles. The input to this stage is a series of single-cycle basic feature sets generated in the previous stage. The program calculates macro-features through a sliding historical data window. The first type of macro-feature is the cycle-period stability feature, used to quantify the consistency of movement patterns. A key indicator is the cycle-period coefficient of variation (CV), which is obtained by calculating the ratio of the standard deviation of a basic feature, such as the RMS value, to its mean value within the most recent M cycles (e.g., M=10). A lower CV value means that the muscle activation intensity is very stable within consecutive cycles, indicating good movement repeatability. The second type of macro-feature is the left-right symmetry feature, used to assess the balance of both limbs, which is crucial for injury risk warning and rehabilitation process monitoring. Taking the left-right intensity asymmetry index as an example, its calculation requires simultaneously acquiring the RMS values ​​of the corresponding muscles on the left and right sides (e.g., the left and right quadriceps) under the same movement cycle k, denoted as... and The calculation formula is as follows: In this formula, This is the intensity symmetry index for the k-th cycle. The formula calculates the absolute difference in RMS values ​​between the left and right sides and normalizes it to the larger of the two, presenting the result as a percentage. An index close to 0% indicates highly symmetrical force exertion on both sides, while a higher value suggests possible lateral compensation. For example, if the RMS of the left quadriceps is 0.23 mV and the right quadriceps is 0.28 mV in a gait cycle, the symmetry index is |0.23-0.28| / max(0.23,0.28)×100%=17.9%. The third type of macroscopic feature is fatigue-related features, used to quantify the evolution of muscle state during prolonged exercise. The fatigue evolution slope is one of the core indicators. This indicator is obtained by performing linear regression analysis on a physiological indicator sequence over a longer sliding window, such as the most recent L=50 cycles. A typical fatigue indicator is the median frequency (MDF), which can be calculated by performing a Fourier transform on the electromyography signal for each normalized cycle. Muscle fatigue is usually accompanied by a decrease in MDF. The processing program performs least-squares linear fitting on the MDF value sequence of the most recent 50 cycles to obtain a trend line. The slope of this trend line is the fatigue evolution slope. A significant negative slope value, such as -0.15 Hz / cycle, quantitatively reveals that the muscle is in the process of fatigue accumulation. By aggregating the calculated inter-cycle coefficient of variation, left-right intensity asymmetry index, and fatigue evolution slope, a high-level feature set across cycles is generated.

[0027] In the third processing stage, namely the concatenation and standardization of the feature set, the program integrates the single-cycle basic feature set generated in the first two stages with the cross-cycle high-level feature set. Following a pre-defined dimensional order, such as intensity features, temporal features, periodic stability features, left-right symmetry features, and fatigue-related features, these numerical indicators are vertically concatenated to form a high-dimensional original feature vector. Since these features have different units and numerical ranges (e.g., millivolts, percentages, Hertz / cycle), direct use can lead to imbalances in certain dimensional weights during model training. Therefore, dimensional standardization is necessary. This embodiment uses the Z-score standardization method. For example, if the currently calculated RMS value is 0.23 millivolts, while the pre-stored population mean for this indicator is 0.20 millivolts and the standard deviation is 0.05 millivolts, then its Z-score standardized value is (0.23-0.20) / 0.05 = +0.6. After performing this standardization operation on each feature in the concatenated high-dimensional vector, the final electromyographic feature vector with all dimensions at a uniform scale is obtained.

[0028] Specifically, in step three, based on normalized periodic data, kinematic coordination features are extracted from the inertial sensing signals corresponding to each period to obtain an inertial consistency feature vector. It is understandable that analyzing only surface electromyography (EMG) signals, while revealing the muscle's activation intensity, timing, and fatigue state, is essentially limited to interpreting neuromuscular commands and cannot directly assess whether these commands achieve optimal coordination with the limb's actual movement. Efficient movement requires not only muscles to activate at the correct time and intensity but also that this activation be precisely coupled with the limb's kinematic processes, such as speed, angle, and timing. Without a quantitative assessment of this neuromechanical coupling, it is impossible to distinguish between effective force application and ineffective or mistimed force application. Therefore, step three in this technical solution establishes a feature dimension for evaluating the coordination consistency between muscle activation and limb movement through synchronous analysis of inertial sensing signals. This fundamentally solves the core problem of whether muscles exert force at the correct time for the correct movement, thereby expanding the assessment of movement quality from a single physiological level to a functional coordination level.

[0029] Figure 3 The flowchart below shows step three of the joint assessment method for muscle activation quality based on rhythmic period normalization according to the embodiments of this application. Figure 3As shown, in one specific implementation, step three includes: Step three-one, based on normalized periodic data, performing local extremum optimization on the surface electromyography signal envelope and inertial sensing signal within each normalized period to obtain a set of characteristic extremum points within the period; Step three-two, based on the electromyography activation peak index and inertial phase marker index in the set of characteristic extremum points within the period, calculating the percentage phase deviation of the relative position difference between the two on the normalized time axis to obtain a single-period deviation feature set; Step three-three, based on the phase deviation value sequence of multiple consecutive single-period deviation feature sets within the sliding window, performing standard deviation statistics on the deviation distribution within the window and concatenating it with the transient deviation value to obtain an inertial consistency feature vector.

[0030] The process is as follows: The input to the processing is the normalized periodic data generated in step one. This data is a one-dimensional array of fixed length N (201 in this example) points in the k-th period, corresponding to the surface electromyography (EMG) signal of the target muscle group and the inertial sensing signal of the limb segment related to the function of that muscle group. For example, in running applications, if the EMG signal comes from the rectus femoris muscle on the front of the thigh, the inertial sensing signal analyzed in conjunction with it can be the angular velocity signal of the movement along the sagittal plane recorded by the inertial measurement unit at the tibia on the same side of the lower leg. In the first processing step, namely the generation of the characteristic extreme point set within the period, the program processes the data of the k-th normalized period in parallel. On the one hand, for the EMG signal sequence of this period, its smooth signal envelope is first extracted by full-wave rectification and low-pass filtering algorithms. Then, optimization is performed in the entire envelope sequence of length N to locate its peak, that is, the time point with the greatest muscle activation intensity. The index number of this peak point on the normalized time axis is recorded as the EMG activation peak point index. This index represents the precise phase in which the muscle exerts maximum force during that cycle. For example, in a running gait cycle, if the peak force exertion of the rectus femoris occurs at the 145th sampling point, then... = 145, which corresponds to a periodic phase of 145 / (201-1)×100%=72.5%. On the other hand, local extrema optimization is performed simultaneously on the inertial sensing signals within this period, such as the angular velocity signal sequence of the lower leg. In a typical gait cycle, the angular velocity reaches a peak when the lower leg swings forward, corresponding to a critical kinematic node where the lower limb prepares for the next landing. The program finds the maximum value in this angular velocity sequence and records its index on the normalized time axis as an index for the inertial phase marker. For example, if the maximum forward swing angular velocity of the lower leg occurs at the 150th sampling point within the same gait cycle, then... =150, which corresponds to 150 / (201-1)×100%=75% of the periodic phase. By structurally binding these two key point indices located in the same normalized time domain, i.e. {EMG activation peak point index: 145, inertial phase marker point index: 150}, the characteristic extreme point set of the k-th period is formed.

[0031] In the second processing stage, namely the generation of the single-cycle deviation feature set, the core task of the program is to quantify the temporal deviation of the two key event points mentioned above. The input to this stage is the set of characteristic extreme points within the cycle generated in the previous stage. For the k-th cycle, the difference between the electromyographic activation peak point index and the inertial phase marker point index is calculated to obtain the phase index deviation. In the example above, this value is 145-150=-5. While this original index difference reflects the temporal relationship, its absolute value is affected by the number of normalization points N. To obtain a more universal and interpretable metric, it needs to be converted to a relative percentage phase deviation. The calculation formula is as follows: In this formula, the denominator This represents the total index span from 0% to 100% phase. Substituting the example values, we get... =(-5) / 200×100%=-2.5%. This negative value indicates that the peak of electromyographic activation is 2.5% of a cycle ahead of the kinematic landmark, which may be an efficient pre-activation mode; conversely, a positive value indicates that muscle exertion is delayed. Encapsulating this calculated scalar value, which quantifies the neuromechanical synergy of the current cycle, as {phase deviation: -2.5%}, yields the single-cycle deviation feature set.

[0032] In the third processing stage, namely the generation of the inertial consistency feature vector, the analysis scale is expanded from a single cycle to multiple consecutive cycles, aiming to evaluate the stability of the cooperative mode. This stage continuously receives and caches the single-cycle deviation feature set output from the previous stage, and maintains a sliding observation window of fixed length W, for example, set to the most recent 20 cycles. The window stores a sequence of phase deviation values ​​for the most recent W cycles, for example {-2.5%, -2.8%, -2.3%, ..., -2.6%}. The standard deviation of this sequence is then calculated. This value quantifies the volatility or stability of recent coordination patterns. A low standard deviation, such as 0.45%, indicates highly consistent and stable muscle-limb coordination patterns, while a higher value may suggest fatigue, motor compensation, or a decline in control. This cross-cycle coordination stability index constitutes the second dimension of the feature vector. Finally, the transient phase deviation value of the current k-th cycle is... For example, -2.5, compared to the just calculated cross-cycle collaborative stability index. For example, a one-dimensional tensor concatenation is performed using 0.45. This generates a two-dimensional inertial consistency feature vector [-2.5, 0.45]. This vector not only describes the immediate coordination between the force exerted and the action in the current cycle, but also contains information on the stability of this coordination pattern over a recent period.

[0033] Specifically, step four involves cross-dimensional fusion of the electromyographic (EMG) feature vectors and the inertial consistency feature vectors to obtain a multidimensional scoring result. This multidimensional scoring result includes quantitative scores for rhythm quality, activation intensity, movement stability, left-right symmetry, and fatigue risk. It should be understood that the preceding steps, through normalization and hierarchical analysis of the original sensor signals, have successfully extracted two sets of feature vectors rich in content and with clear physical meaning: one set is the EMG feature vector, which details the activation state of the muscles themselves, and the other set is the inertial consistency feature vector, which precisely quantifies the synergistic relationship between muscle activation and limb movement. However, these feature vectors are essentially high-dimensional, standardized sets of values. For non-professional users and even sports coaches, they are too complex and abstract to directly form an intuitive and comprehensive judgment of movement quality. To transform these low-level features, scattered across different dimensions and representing different physiological and mechanical phenomena, into a set of application-oriented and easily understood top-level evaluation indicators, step four in this technical solution uses intelligent weighting and mapping to integrate and refine the complex feature inputs into quantitative scores corresponding to core dimensions such as rhythm quality, activation intensity, and stability.

[0034] In one specific implementation, step four includes: vertically concatenating the electromyography feature vector and the inertial consistency feature vector along the column dimension and smoothing outliers to obtain a joint input feature matrix; performing linear mapping on the joint input feature matrix based on the fusion weight matrix and the bias vector to obtain a multidimensional quantized score set; and performing value range normalization and compression on the original scores of each dimension in the multidimensional quantized score set to obtain a multidimensional scoring result.

[0035] The process is as follows: In the first processing stage, namely the construction of the joint input feature matrix, the primary task is to ensure strict spatiotemporal alignment of the two input vectors, i.e., verifying whether they originate from the same motion cycle k. After confirming consistency, the two one-dimensional vectors are vertically concatenated along the column dimension, also known as block matrix concatenation. This operation combines the two independent feature sets, the electromyography feature vector and the inertial consistency feature vector, into a single feature vector of a higher dimension containing global information. The form is as follows: In this formula, This represents the joint input feature matrix constructed at the k-th evaluation node; it is a... A column vector of dimension. For this node 3D electromyographic feature vector, Then it is the corresponding 3D inertial consistency feature vector. After concatenation, to enhance the robustness of the model, further... Outlier smoothing is performed on the elements in the model. For example, median filtering or limiting the rate of change is used to identify and correct extreme feature values ​​that exceed reasonable physiological ranges caused by instantaneous sensor noise, ensuring the stability and reliability of the input model's features. After this processing step, the output is a clean, complete, and information-fused joint input feature matrix.

[0036] In the second processing step, namely the solution of the multidimensional quantized score set, the joint input feature matrix generated in the previous step is... As input, a pre-defined feature fusion calculation model is loaded. In one implementation, this model is a pre-trained multivariate linear weighted network. The core of this network consists of a fusion weight matrix. With a bias vector Composition. Its goal is to... An S-dimensional input feature space is linearly mapped to an S-dimensional output scoring space, where S is the preset number of final evaluation dimensions, such as 5, corresponding to rhythm quality, activation intensity, movement stability, left-right symmetry, and fatigue risk. The solution process follows the matrix multiplication and addition operations as follows: Analysis of this formula reveals that, It is output. Original score vector. Weight matrix. The dimension is Each element within it They all represent the j-th input feature, from The contribution weight to the i-th final scoring dimension, such as action stability. Bias vector. The dimension is This is used to provide the base offsets for each rating dimension. Weight matrix With bias vector The values ​​are obtained through supervised machine learning training. The training process requires a large-scale labeled dataset containing raw sensor data from athletes of varying skill levels under different conditions, along with gold-standard scores from biomechanics experts or experienced coaches for these movement segments. During training, the sensor data is processed into a joint input feature matrix, which is then used as the model input. The model is continuously adjusted using optimization algorithms such as gradient descent, with the expert scores as the target. and The numerical value minimizes the error, such as mean squared error, between the model's predicted raw score and the expert's score. Through this process, the model learns how to combine and calculate an assessment score with expert cognitive level based on the complex underlying features. The output of this stage is a multi-dimensional quantified score set containing the unconstrained raw scores of each assessment dimension, such as {rhythm quality: 15.2, activation intensity: -8.3,...}.

[0037] In the third processing stage, namely the generation of multidimensional scoring results, the task is to convert the raw scores from the previous stage—which have an unbounded numerical range and are not intuitive—into standardized, easily understood percentage scores. This stage involves processing the raw scores for each dimension in the multidimensional quantified score set. ,in This represents independent range normalization and compression for a specific evaluation dimension. In this embodiment, a parameterized logic function (Sigmoid function) is used to implement this nonlinear mapping. The specific calculation formula is as follows: In this formula, It is a dimension The final output ranges from 0 to The ratings between them. This is the upper limit of the rating scale, set to 100. It is a dimension The input is the original score. and It is for this specific rating dimension Two key parameters: The gain or slope parameter controls the steepness of the curve near the center point, thus determining the sensitivity of the score. The center point parameter represents the baseline point of the original score that can be mapped to 50 points. These two parameters are empirically determined based on the statistical distribution of the original scores in the massive dataset during training. For example, for the action stability dimension, by analyzing the distribution of the original scores of all training samples, the mean or median can be set as the center point, and the gain can be adjusted so that most good samples score between 70 and 90 points, while poor samples score below 40 points. For example, if a sample's original action stability score is 15.2, and the normalization parameter for this dimension is set to... =100, =0.2, =10.0, then its final score is After performing this normalization operation on all S dimensions of the raw scores, these continuous scores, standardized to the 0-100 range, are structured and encapsulated to generate multidimensional scoring results, such as {rhythm quality: 85.1, activation intensity: 62.4, movement stability: 73.9, left-right symmetry: 92.0, fatigue risk: 33.5}. It should be noted that the above scoring calculation can be used not only for offline deep analysis on the server side but also perfectly adapted to real-time monitoring scenarios at the edge. In real-time applications, the system can use a sliding update mechanism to continuously refresh the scoring results by single step, every few steps, every few cadences, or according to a preset fixed time window. The output multidimensional scoring results can be directly pushed and displayed on smart terminal devices, such as smartphones, sports watches, cycling computers, rehabilitation monitoring platforms, or specific medical assessment terminals. Furthermore, these quantitative scores can also be cascaded to drive features such as sound prompts, haptic feedback, automatic adjustment of the running belt speed on smart treadmills, or directly connected to remote rehabilitation systems to generate adaptive intervention suggestions.

[0038] Specifically, in step five, based on preset level threshold classification rules and continuous window time series analysis methods, the multidimensional scoring results are mapped to levels and their trend evolution is identified, generating a joint evaluation result that includes level classification results, risk level results, and trend indication information. In other words, although the multidimensional scoring results output in the previous steps precisely quantify the specific levels of athletic performance in multiple dimensions such as activation intensity and symmetry in the form of continuous scores from 0 to 100, they are essentially still a dynamically changing numerical stream. For end-users in non-professional fields, an isolated number, such as a motion stability score of 73.9, is not intuitively good or bad, and it is difficult to quickly understand its underlying kinematic meaning. More importantly, a snapshot-style score at a single point in time cannot reveal the evolutionary pattern of this indicator during the exercise process. For example, the potential risks and intervention measures required for a symmetry that is stable at 70 points are completely different from those for a symmetry that rapidly drops from 90 points to 70 points. Therefore, step five in this technical solution introduces hierarchical mapping and time series analysis to give these continuous quantitative scores an intuitive qualitative interpretation with clear business meaning, transforming abstract score values ​​into easily understandable levels and risk labels, and actively identifying their dynamic trends, ultimately generating a joint evaluation result that is complete, highly readable, and has decision-guiding value.

[0039] Figure 4 This is a flowchart of step five in the joint assessment method for muscle activation quality based on rhythmic cycle normalization according to an embodiment of this application. Figure 4As shown, in one specific implementation, step five includes: step five-one, based on a preset stepped judgment threshold, performing segmented condition judgment and discrete label mapping on the continuous scores of each dimension in the multidimensional scoring results to obtain a discrete level and risk state set; step five-two, based on the continuous score time series within the sliding historical window, performing least squares linear regression fitting and gradient trigger judgment on the score evolution of each dimension in the multidimensional scoring results to obtain trend indication information; step five-three, performing structured merging and serialized encapsulation of the discrete level and risk state set and trend indication information to obtain a joint evaluation result.

[0040] The process is as follows: In the first processing stage, namely the generation of discrete levels and risk state sets, the program will apply different preset threshold rules to determine the condition of each dimension's score in the multidimensional scoring results, based on its attribute (gain or risk). These tiered judgment thresholds are determined based on statistical analysis of exercise data from a large-scale reference population, combined with expert experience. For example, by analyzing the distribution of movement stability scores of thousands of healthy runners, percentiles such as 25%, 50%, and 75% can be used as boundaries to divide the levels into medium, good, and excellent. Specifically, for the performance gain dimension, such as a movement stability score of 73.9, the preset tiered judgment thresholds might be: {Excellent: >90, Good: 70-90, Medium: 50-70, Poor: <50}. The program compares 73.9 with this rule, determines that it falls within the range of 70 to 90, and therefore maps it to the discrete performance level label of good. For negative monitoring dimensions, such as a fatigue risk score of 33.5, the preset safety boundary thresholds might be: {High Risk: >70, Medium Risk: 40-70, Low Risk: <40}. The program compares 33.5 with this rule, determines that it is less than 40, and therefore maps it to the discrete risk warning label "Low Risk". This mapping process is repeated for all scoring dimensions, and the resulting levels and risk labels are bound as key-value pairs, ultimately generating a discrete set of levels and risk states for that evaluation node, for example, {Movement Stability Level: "Good", Left-Right Symmetry Level: "Excellent", Fatigue Risk Level: "Low Risk"}.

[0041] In the second processing stage, namely the generation of trend information, the program focuses on the time-series dynamics of the scoring. This stage continuously receives and caches the multi-dimensional scoring results output from step four, and maintains a length of [missing information] in memory. For example, a sliding historical data window with 60 evaluation nodes can be set. Once the window is full, a data path is formed for each rating dimension within the window. A time series of continuous scores was then used. Subsequently, a univariate linear regression was performed on this time series using the least squares method to calculate the gradient of the recent evolution of the score in that dimension, i.e., the slope of the regression line. The evolution slope of each evaluation dimension within the time window The calculation formula is: In this formula, Let d be the slope of the evolution of dimension d; It is the window length; It is the time index of the i-th data point within the window, ranging from 1 to... ; Is dimension d in The score for each moment; It is the average value of the time index; It is the average score of dimension d within the window. This formula precisely quantifies the average rate and direction of change of the score over time. The slope is calculated. The program then compares this with a preset trend trigger threshold. This threshold is also based on statistical analysis of a large amount of data to determine what degree of slope change has actual physiological significance, rather than random fluctuation. For example, for the fatigue risk dimension, the preset positive surge threshold might be +0.5 points / cycle, meaning that when the rate of increase in fatigue score exceeds this value, fatigue should be considered to be accumulating significantly. If the fatigue risk slope calculated in the current window is +0.68, then this value exceeds the threshold of +0.5, and the program will activate a preset semantic description template to generate a text string such as "fatigue risk is rising rapidly." For the left-right symmetry dimension, if its negative decay threshold is -0.25 points / cycle, and the calculated slope is -0.31, then a prompt indicating that symmetry is showing a downward trend will be generated. The output of this step is these structured trend prompts generated by gradient trigger judgment, representing the dynamic evolution law. The system has a rich built-in trend-based diagnostic vocabulary. In addition to the examples mentioned above, when the model identifies a specific combination of mutations in underlying features, it can also generate textual prompts with high medical and kinesiological intuition, including but not limited to rhythmic stability but decreased local activation quality, enhanced right-sided limb compensation, and fluctuations in core control.

[0042] In the third processing stage, namely the merging and encapsulation of the joint evaluation results, the program will synchronously receive two sets of data from the first two stages, belonging to the same evaluation node: one set is a discrete set of levels and risk states describing the current static state, and the other set is a trend indication information describing recent dynamic changes. The program will spatially merge and serialize these two data structures to construct a unified payload format that is easy for external systems to call and parse, such as a JSON object. An example of a final generated joint evaluation result is as follows: {"Evaluation Node":k,"Level Grading Result":{"Movement Stability":"Good","Left-Right Symmetry":"Excellent"},"Risk Level Result":{"Fatigue Risk":"Low Risk"},"Trend Indication Information":["Symmetry shows a downward trend","Fatigue Risk is rapidly increasing"]}. This structured data contains both the qualitative rating of each indicator at the current moment and indicates its recent evolution direction. The information is complete and highly readable, providing end users, coaches, or rehabilitation therapists with a comprehensive, intuitive, and clearly guiding exercise quality diagnostic report. This provides an objective and quantitative basis for real-time technical adjustments, training load management, and injury risk warnings.

[0043] In summary, the joint evaluation method for muscle activation quality based on rhythmic cycle normalization, as described in this application, is elucidated. It uses the rhythmic cycle as a unified phase benchmark to construct a standardized framework that makes motion cycle data of different physical durations comparable within the normalized phase domain. Under this unified phase framework, a systematic quality assessment is performed on muscle activation timing, peak phase position, left-right symmetry, periodic stability, and fatigue evolution. Combined with the collaborative analysis of inertial signals in the phase domain, a joint evaluation result including multi-dimensional quantitative scores, grade levels, and trend indicators is output. This application has good scenario adaptability and can be applied to various periodic exercise scenarios, such as analyzing muscle activation timing and left-right symmetry during the support and swing phases of running, evaluating the force window and pedaling symmetry within the cadence cycle during cycling, and monitoring activation differences and compensatory trends between the affected and healthy sides during rehabilitation training.

[0044] Figure 5 The following is a block diagram of a joint assessment system for muscle activation quality based on rhythmic period normalization according to an embodiment of this application, as shown below. Figure 5As shown, the muscle activation quality joint assessment system 100 based on rhythmic cycle normalization includes: a signal normalization module 110, used to perform periodic interception and time coordinate normalization mapping on the surface electromyography (EMG) signals and inertial sensing signals collected in the corresponding time period based on the start and end times of each cycle in the continuous motion rhythmic cycle sequence to obtain normalized cycle data; an EMG feature extraction module 120, used to perform multi-dimensional feature extraction on the surface EMG signals of the target muscle group in each normalized cycle based on the normalized cycle data to obtain EMG feature vectors, the EMG feature vectors including intensity features, temporal features, periodic stability features, left-right symmetry features, and fatigue-related features; and an inertial consistency processing module 130. The system is used to extract kinematic co-features from inertial sensing signals corresponding to each cycle based on normalized periodic data to obtain inertial consistency feature vectors; the multidimensional scoring module 140 is used to perform cross-dimensional fusion calculation of electromyographic feature vectors and inertial consistency feature vectors to obtain multidimensional scoring results, which include quantitative scores for rhythm quality, activation intensity, movement stability, left-right symmetry, and fatigue risk dimensions; the evaluation result generation module 150 is used to perform level mapping and trend evolution identification on the multidimensional scoring results based on preset level threshold classification rules and continuous window time series analysis methods, and generate a joint evaluation result containing level classification results, risk level results, and trend prompt information.

[0045] Here, those skilled in the art will understand that the specific operations of each step in the above-described joint assessment system for muscle activation quality based on rhythmic cycle normalization have been referenced above. Figures 1 to 4 The description of the joint assessment method for muscle activation quality based on rhythmic cycle normalization is detailed here, and therefore, its repeated description will be omitted.

[0046] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any way. Although this application has been disclosed above with reference to preferred embodiments, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the scope of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A joint assessment method for muscle activation quality based on rhythmic cycle normalization, characterized in that, include: Step 1: Based on the start and end times of each cycle in the continuous motion rhythm cycle sequence, perform cycle extraction and time coordinate normalization mapping on the surface electromyography signals and inertial sensing signals collected in the corresponding time period to obtain normalized cycle data. Step 2: Based on normalized periodic data, multidimensional feature extraction is performed on the surface electromyography (EMG) signals of the target muscle group in each normalized period to obtain EMG feature vectors. The EMG feature vectors include intensity features, temporal features, periodic stability features, left-right symmetry features, and fatigue-related features. Step 3: Based on the normalized periodic data, perform kinematic cooperative feature extraction on the inertial sensing signals corresponding to each period to obtain the inertial consistency feature vector. Step 4: Perform cross-dimensional fusion calculation on the electromyographic feature vector and the inertial consistency feature vector to obtain multidimensional scoring results. The multidimensional scoring results include quantitative scores for rhythm quality, activation intensity, movement stability, left-right symmetry, and fatigue risk dimensions. Step 5: Based on the preset level threshold classification rules and continuous window time series analysis method, perform level mapping and trend evolution identification on the multidimensional scoring results to generate a joint evaluation result that includes level classification results, risk level results and trend prompt information.

2. The method for joint assessment of muscle activation quality based on rhythmic cycle normalization according to claim 1, characterized in that, Step one includes: Based on the start and end timestamps of each cycle in the continuous motion rhythm cycle sequence, the surface electromyography signals and inertial sensing signals in the corresponding time period are time-series aligned and cycle segments are cut to obtain the original cycle segment set. Based on the preset normalized fixed number of sampling points and the physical duration of each original periodic segment, each segment in the original periodic segment set is subjected to proportional phase coordinate mapping to obtain a periodic segment with mapped coordinates. Normalized periodic data is generated based on periodic segments with mapped coordinates.

3. The method for joint assessment of muscle activation quality based on rhythmic cycle normalization according to claim 1, characterized in that, Step two includes: Based on the surface electromyography signal amplitude sequence in the normalized periodic data, the root mean square energy of the activation signal of the target muscle group in each normalized periodic segment is calculated and the activation time sequence is identified to obtain the single-period basic feature set. Based on the feature value sequence of multiple consecutive period segments in the single-period basic feature set, cross-period statistics are performed on the period variation coefficient, left and right side intensity asymmetry index and fatigue evolution slope to obtain a cross-period high-level feature set. The single-cycle basic feature set and the cross-cycle high-level feature set are spliced ​​and standardized to obtain the electromyographic feature vector.

4. The method for joint assessment of muscle activation quality based on rhythmic cycle normalization according to claim 1, characterized in that, Step three includes: Based on normalized periodic data, local extremum optimization is performed on the surface electromyography signal envelope and inertial sensing signal within each normalized period to obtain the set of characteristic extremum points within the period. Based on the electromyographic activation peak index and the inertial phase marker index in the feature extreme point set within the cycle, the percentage phase deviation is calculated on the relative position difference between the two on the normalized time axis to obtain the single-cycle deviation feature set. Based on the phase deviation value sequence of multiple consecutive single-cycle deviation feature sets within a sliding window, the standard deviation of the deviation distribution within the window is statistically analyzed and concatenated with the transient deviation value to obtain the inertial consistency feature vector.

5. The method for joint assessment of muscle activation quality based on rhythmic cycle normalization according to claim 1, characterized in that, Step four includes: The electromyography feature vector and the inertial consistency feature vector are vertically concatenated along the column dimension and outlier smoothing is performed to obtain the joint input feature matrix; Based on the fusion weight matrix and bias vector, a linear mapping solution is performed on the joint input feature matrix to obtain a multidimensional quantized score set; The original scores of each dimension in the multidimensional quantified score set are normalized and compressed to obtain the multidimensional scoring results.

6. The method for joint assessment of muscle activation quality based on rhythmic cycle normalization according to claim 1, characterized in that, Step five includes: Based on the preset step-by-step judgment threshold, the continuous scores of each dimension in the multidimensional scoring results are segmented and conditionally judged and discretely labeled to obtain discrete levels and risk status sets. Based on the continuous score time series within the sliding history window, least squares linear regression fitting and gradient triggering determination are performed on the score evolution of each dimension in the multidimensional scoring results to obtain trend indication information. The discrete levels, risk status sets, and trend indication information are structured, merged, and serialized to obtain joint evaluation results.

7. The method for joint assessment of muscle activation quality based on rhythmic cycle normalization according to claim 2, characterized in that, Based on periodic segments with mapped coordinates, normalized periodic data is generated, including: interpolating and resampling the signal amplitude at the corresponding position of each target mapped timestamp in the periodic segment with mapped coordinates to obtain normalized periodic data.

8. The method for joint assessment of muscle activation quality based on rhythmic period normalization according to claim 2, characterized in that, Based on periodic segments with mapped coordinates, normalized periodic data is generated, including: Local curvature field estimation is performed on periodic segments with mapped coordinates to obtain a local curvature field sequence; Curvature-compensated adaptive interpolation resampling is performed on the local curvature field sequence and the periodic segment with mapped coordinates to obtain the curvature-compensated interpolation signal sequence; Point-by-point fidelity verification and normalization encapsulation are performed on the curvature compensation interpolation signal sequence and the local curvature field sequence to obtain normalized periodic data.

9. A joint assessment system for muscle activation quality based on rhythmic cycle normalization, characterized in that, include: The signal normalization module is used to perform periodic extraction and time coordinate normalization mapping on the surface electromyography signals and inertial sensing signals collected in the corresponding time period based on the start and end times of each period in the continuous motion rhythm periodic sequence to obtain normalized periodic data. The electromyography feature extraction module is used to extract multidimensional features from the surface electromyography signals of the target muscle group in each normalized cycle based on normalized cycle data to obtain electromyography feature vectors. The electromyography feature vectors include intensity features, temporal features, cycle stability features, left-right symmetry features, and fatigue-related features. The inertial consistency processing module is used to extract kinematic cooperative features from the inertial sensing signals corresponding to each period based on normalized periodic data to obtain inertial consistency feature vectors. The multidimensional scoring module is used to perform cross-dimensional fusion calculation of electromyographic feature vectors and inertial consistency feature vectors to obtain multidimensional scoring results. The multidimensional scoring results include quantitative scores for rhythm quality, activation intensity, movement stability, left-right symmetry, and fatigue risk dimensions. The assessment result generation module is used to perform level mapping and trend evolution identification on multidimensional scoring results based on preset level threshold classification rules and continuous window time series analysis methods, and generate a joint assessment result that includes level classification results, risk level results and trend prompt information.

Images