A muscle state analysis method based on surface electromyography time-frequency image recognition

Abstract

The application discloses a muscle state analysis method based on surface electromyogram time-frequency image recognition. The method collects surface electromyogram signals of target muscles, and synchronously collects electrode contact impedance, wearing detection state and time stamp; before formal analysis, resting electromyogram data, baseline contraction electromyogram data and current environmental noise data of the current subject are acquired to form individual and environmental adaptive threshold; the original electromyogram sampling data is calibrated in zero point, gain and baseline voltage to obtain input end electromyogram value data; the input end electromyogram value data is divided into windows and surface electromyogram time-frequency images are generated; the signal reliability value is calculated in combination with contact impedance, power frequency interference, baseline drift, saturated sampling, time-frequency texture distortion and sentinel channel environmental reference signal. When the recognition result abnormally jumps or the reliability is insufficient, reverse verification is triggered to improve the accuracy, comparability and traceability of muscle state analysis under different subjects and different environments.

Description

Technical Field

[0001] This invention relates to the field of muscle state analysis technology, specifically a muscle state analysis method based on surface electromyography time-frequency image recognition. Background Technology

[0002] Existing surface electromyography (EMG) acquisition and analysis schemes typically employ a processing flow of electrode acquisition, differential amplification, filtering, analog-to-digital conversion, feature extraction, and state recognition. With the development of image recognition technology, some schemes convert one-dimensional EMG time-series signals into time-frequency images, and then extract time-frequency texture features using image recognition models to determine the type of movement, muscle activation intensity, or fatigue level. Compared to simple time-domain or frequency-domain features, time-frequency images can simultaneously express the changes in EMG signals in both time and frequency, thus possessing better state representation capabilities.

[0003] However, existing methods for muscle state analysis based on surface electromyography (SEMG) time-frequency image recognition still have the following shortcomings. Current methods typically assume the input signal is a valid EEMG signal and only perform fixed filtering or simple impedance judgment at the front end. In actual acquisition, electrode loosening, wire swaying, skin sweating, motion artifacts, and environmental power frequency interference can all alter the amplitude, baseline, and spectral structure of the acquired signal, causing the time-frequency image to produce energy enhancement, low-frequency drift, or texture abrupt changes similar to real muscle activity. If the subsequent recognition model still outputs a definite state result, it is easy to misjudge acquisition abnormalities as increased muscle activation, changes in movement, or increased fatigue, and there are significant individual differences among different subjects. Skin impedance, subcutaneous fat thickness, muscle cross-sectional area, hair condition, sweating status, muscle fiber direction, and force application habits all affect the amplitude and spectral distribution of surface EEMG signals. Under the same muscle physiological state, the EEMG amplitude acquired by different subjects may show significant differences. If a uniform fixed threshold is used for muscle activation or fatigue judgment, it is difficult to take into account different subjects, and different acquisition environments have a significant impact on surface EEMG signals. Laboratory power supply conditions, grounding status, surrounding electromagnetic equipment, ambient temperature and humidity, conductor layout, and the subject's range of motion all affect power frequency interference, harmonic interference, low-frequency drift, and random noise levels. Existing methods often suppress interference by fixing filter parameters, but fixed filtering cannot distinguish whether energy changes appearing in the current window originate from genuine muscle activity or from sudden increases in ambient noise or motion artifacts. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a muscle state analysis method based on surface electromyography time-frequency image recognition, which solves the problems in the prior art where fixed thresholds are difficult to adapt to different subjects, fixed filters are difficult to adapt to different environments, signal quality lacks quantitative evaluation, time-frequency image recognition results lack credibility constraints, and fatigue state determination is prone to abnormal jumps.

[0005] This invention proposes a muscle state analysis method based on surface electromyography time-frequency image recognition, comprising the following steps: S1. Obtain the individual baseline data of the current subject and the environmental interference data of the current collection environment; S2. Acquire surface electromyographic signals of the target muscle and simultaneously obtain electrode contact impedance, wearing detection status and sampling timestamp to form raw electromyographic sampling data; S3. Based on the device calibration parameters, convert the raw electromyography sampling data into input electromyography value data. S4. Perform windowing and time-frequency transformation on the input electromyography data to generate a surface electromyography time-frequency image; S5. Determine the individual and environmental adaptive threshold sets based on the individual baseline data and environmental interference data, and calculate the signal confidence value of each data window by combining the electrode contact impedance, wearing detection status, baseline drift, saturation sampling ratio and time-frequency texture distortion. S6. Input the surface electromyography time-frequency image, the standardized features obtained from the electromyography data at the input end, and the signal confidence value into the muscle state recognition model, and output the muscle activation intensity, fatigue level, or movement state. S7. When the output results meet the abnormal conditions, perform reverse verification on the corresponding data window and store the analysis data, recognition results and reverse verification flags.

[0006] A method for muscle state analysis based on surface electromyography time-frequency image recognition, characterized by comprising the following steps: S1. Obtain the individual baseline data of the current subject and the environmental interference data of the current collection environment; S2. Acquire surface electromyographic signals of the target muscle and simultaneously obtain electrode contact impedance, wearing detection status and sampling timestamp to form raw electromyographic sampling data; S3. Based on the device calibration parameters, convert the raw electromyography sampling data into input electromyography value data. S4. Perform windowing and time-frequency transformation on the input electromyography data to generate a surface electromyography time-frequency image; S5. Determine the individual and environmental adaptive threshold sets based on the individual baseline data and environmental interference data, and calculate the signal confidence value of each data window by combining the electrode contact impedance, wearing detection status, baseline drift, saturation sampling ratio and time-frequency texture distortion. S6. Input the surface electromyography time-frequency image, the standardized features obtained from the electromyography data at the input end, and the signal confidence value into the muscle state recognition model, and output the muscle activation intensity, fatigue level, or movement state. S7. When the output results meet the abnormal conditions, perform reverse verification on the corresponding data window and store the analysis data, recognition results and reverse verification flags.

[0007] Specifically, the individual baseline data mentioned in S1 includes the current subject's resting electromyography data, baseline contractile electromyography data, and skin and electrode contact impedance data; The environmental interference data includes at least one of the following: environmental noise data under electrode short-circuit condition, environmental noise data under resting wearing condition, or environmental reference signal collected by the sentry channel.

[0008] Specifically, the device calibration parameters include zero-point offset, actual amplification factor, analog-to-digital conversion reference voltage, and reference voltage drift; the input electromyography data is obtained in the following manner: Zero-point data is collected when the detection electrode and the reference electrode are short-circuited, and the zero-point offset is determined based on the sampled value of the stable segment. Input a calibration signal with a known amplitude and a known frequency to the acquisition input terminal, and determine the actual amplification factor based on the acquired output amplitude and the known amplitude; For each raw sampled value, it is first converted into the output voltage based on the analog-to-digital conversion reference voltage and the number of quantization bits, then the voltage value corresponding to the zero offset is subtracted, and then divided by the actual amplification factor to obtain the input electromyography voltage value. The input electromyography voltage value is saved as input electromyography data in the microvolt range.

[0009] Specifically, the surface electromyography time-frequency image is a multi-layer time-frequency image, which includes at least an electromyography energy layer, a quality mask layer, and an environmental noise reference layer; The electromyographic energy layer is generated by performing short-time spectral transformation or wavelet transform on the input electromyographic data, and is used to represent the time-frequency energy distribution of the target muscle. The quality mask layer is generated based on electrode contact impedance, wear detection status, signal confidence value, saturation sampling ratio, and baseline drift, and is used to mark low confidence time locations and abnormal acquisition time locations. The environmental noise reference layer is generated based on environmental interference data or environmental reference signals collected by the sentry channel, and is used to represent the time-frequency distribution of noise unrelated to the target muscle in the current collection environment.

[0010] Specifically, the individual and environment adaptive threshold set includes at least the contact impedance warning threshold, power frequency interference deduction threshold, baseline drift deduction threshold, muscle activation threshold, fatigue spectrum migration threshold, and signal credibility effective threshold. The contact impedance warning threshold is determined based on the median of the stable segment contact impedance under the current resting state of the subject. The power frequency interference reduction threshold is determined based on the power frequency interference amplitude under electrode short-circuit state or resting wearing state; The muscle activation threshold is determined based on the current subject's resting root mean square value and baseline contraction intensity; The fatigue spectrum migration threshold is determined based on the center of the reference spectrum under the current subject's reference contraction state. The effective threshold for signal reliability is determined based on the signal reliability distribution of stable data windows in resting electromyography data and baseline contractile electromyography data.

[0011] Specifically, the sentry channel has a switchable acquisition mode; In static electromyography experiments, the sentinel channel is connected to an impedance network that matches the input impedance of the target electromyography channel, and is used to acquire reference signals of electromagnetic interference from the acquisition circuit and the current environment. In free-movement electromyography experiments, the sentinel channel is placed in an area with low muscle activity to collect motion artifact reference signals caused by wire swaying, overall human movement, and electrode slippage relative to the skin. When the target electromyography channel and the sentinel channel exhibit synchronous power frequency enhancement, synchronous low-frequency oscillation, or synchronous spike interference within the same data window, this data window is marked as an environmental interference risk window or a motion artifact risk window.

[0012] Specifically, the muscle state recognition model reads the multi-layer time-frequency image and extracts the effective frequency band energy distribution of electromyography in the electromyography energy layer, the low confidence region location in the quality mask layer, and the noise texture location in the environmental noise reference layer, respectively. When the electromyography energy layer shows energy enhancement and the quality mask layer does not mark the anomaly at the corresponding time position, and the environmental noise reference layer does not show synchronous enhancement at the corresponding time position, the energy enhancement is taken as a candidate feature of real muscle activity. When the electromyography energy layer and the environmental noise reference layer show synchronous enhancement at the corresponding time position, or when the quality mask layer is marked abnormally at the corresponding time position, reduce the weight of this data window in muscle state recognition or output a state pending review.

[0013] Specifically, the standardized features include at least one of standardized muscle activation values, spectral migration values, and fatigue trend values; The standardized muscle activation value is determined based on the root mean square value of the current data window, the current subject's resting root mean square value, and the current subject's baseline contraction intensity. The spectral shift value is determined based on the spectral center position of the current data window and the baseline spectral center under the current subject baseline contraction state; The fatigue trend value is determined based on the changes in the root mean square value of the continuous data window, the changes in the integral electromyography value, the downward shift of the spectrum center, and the duration of sustained contraction.

[0014] Specifically, the abnormal conditions include a jump in muscle state results within a preset time, a signal confidence value lower than the effective threshold for signal confidence, or standardized features exceeding the physiologically reasonable range corresponding to the current subject. The reverse verification includes reading the original electromyography sampling data, electrode contact impedance, wearing detection status, input electromyography value data, surface electromyography time-frequency image, environmental interference data, and individual and environmental adaptive threshold sets of the corresponding data window, and re-determining whether the data window has abnormal contact impedance, abnormal power frequency interference, abnormal baseline drift, abnormal saturation sampling, or abnormal time-frequency texture. If any of the above abnormalities exist, the data window is marked as an abnormal acquisition window. If no such abnormality exists, the standardized feature calculation and muscle state recognition are re-executed.

[0015] Specifically, this also includes individual parameter binding and storage, and model retraining steps; The individual baseline data of the current subject, the environmental interference data of the current acquisition environment, the equipment calibration parameters, the individual and environmental adaptive threshold sets, the surface electromyography time-frequency images, the signal confidence value, the muscle state results, and the reverse verification mark are bound and stored in the same experimental data package; When the experimental data package is used for subsequent model retraining, data windows with signal confidence values ​​lower than the effective threshold of signal confidence are removed, and data windows marked as abnormal acquisition in reverse verification are used as negative samples or low-weight samples. The muscle state recognition model is updated using the retained data window and its corresponding individual baseline data and environmental interference data.

[0016] The beneficial effects of this invention are as follows: (1) This invention establishes individual baseline data of the current subject and environmental interference data of the current collection environment before formal analysis, and generates an individual and environment adaptive threshold set accordingly, so that muscle state recognition no longer depends on a uniform fixed threshold. This can reduce the impact of different subjects’ skin impedance, muscle strength, sweating state, as well as power frequency interference, baseline drift and motion artifacts in different experimental environments on the recognition results.

[0017] (2) The present invention calculates the signal confidence value for each data window and triggers reverse verification when the muscle state result changes abnormally, the confidence is insufficient or the standardized features exceed the reasonable range. It can re-evaluate, correct or mark abnormal windows, thereby improving the accuracy, stability, comparability and traceability of muscle activation intensity, fatigue degree and movement state analysis. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Figure 1 This is a schematic diagram of the process of the present invention.

[0020] Figure 2 This is a timing diagram for the data processing of the present invention. Detailed Implementation

[0021] To make the objectives, technical means, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0023] Example 1: As Figures 1-2 As shown in this embodiment, the muscle state analysis method based on surface electromyography time-frequency image recognition is used to analyze the activation intensity, fatigue level, or movement state of target muscles. Target muscles include forearm flexors, forearm extensors, biceps brachii, triceps brachii, quadriceps femoris, gastrocnemius, or other muscles whose electromyographic signals can be collected through surface electrodes.

[0024] In this embodiment, the positive and negative terminals of the detection electrode are arranged along the direction of the target muscle fibers, and the reference electrode is placed in the bony region or a region with weak muscle activity near the target muscle. The positive and negative terminals of the detection electrode are used to form a differential input, and the reference electrode is used to provide a stable reference potential. The acquisition circuit amplifies, filters, and performs analog-to-digital conversion on the differential input signal, and simultaneously acquires the electrode contact impedance, wearing detection status, and sampling timestamp to form raw electromyographic sampling data.

[0025] Furthermore, before formal analysis, individual baseline data of the current subject and environmental interference data of the current acquisition environment are obtained. Individual baseline data is used to characterize the current subject's resting electromyography level, baseline contraction level, and skin and electrode contact status; environmental interference data is used to characterize the levels of power frequency interference, harmonic interference, baseline drift, low-frequency disturbances, and random noise in the current acquisition environment. Based on the above data, the system establishes an individual and environmental adaptive threshold set, enabling the subsequent analysis process to adapt to the current subject and the current environment.

[0026] Therefore, the method of this application can avoid directly using a uniform fixed threshold to judge the muscle state of all subjects. For subjects with high skin impedance, weak muscle strength, or high resting noise, the system uses the subject's own resting baseline and baseline contraction data as a judgment reference; for environments with strong power frequency interference or obvious motion artifacts, the system uses the current environmental interference data as a quality evaluation reference, thereby improving the stability of the recognition results.

[0027] In this embodiment, the raw electromyography (EMG) sampling data is not directly input into the muscle state recognition model. Instead, it is first converted into input EMG values ​​based on device calibration parameters. These parameters include zero-point offset, actual amplification factor, analog-to-digital conversion reference voltage, and reference voltage drift. Through this conversion, the system transforms the digital sampling values ​​output by the analog-to-digital converter into microvolt-level input EMG voltage values, reducing the impact of gain differences and reference voltage differences between different acquisition devices on subsequent analysis.

[0028] Furthermore, the system performs windowing processing on the input electromyography (EMG) data. Each data window corresponds to a specific time period, and the system performs time-frequency transformation on each data window to generate a surface EMG time-frequency image. This surface EMG time-frequency image is used to represent the time-frequency energy changes of the target muscle within the corresponding time period. Subsequently, the system combines electrode contact impedance, wearing detection status, baseline drift, saturation sampling ratio, time-frequency texture distortion degree, and environmental interference data to calculate the signal confidence value for each data window.

[0029] In this embodiment, the muscle state recognition model receives surface electromyography (EMG) time-frequency images, standardized features obtained from the input EMG data, and signal confidence values. The standardized features include at least one of standardized muscle activation values, spectral shift values, and fatigue trend values. The model outputs muscle activation intensity, fatigue level, or movement state. When the output meets an abnormal condition, the system triggers reverse verification processing and stores the analyzed data, recognition results, and reverse verification flags.

[0030] Example 2: In this example, the individual baseline data includes the current subject's resting electromyography (EMG) data, baseline contractile EMG data, and skin and electrode contact impedance data. The individual baseline data is used to establish an EMG analysis reference for the current subject, avoiding direct impact of EMG amplitude differences between different subjects on state determination.

[0031] In this embodiment, after electrode placement, the system requires the subject to maintain a relaxed target muscle state and collect resting electromyography (EMG) data for at least 30 seconds. After removing transitional data from the initial acquisition phase, the system divides the remaining resting EMG data into multiple data windows and calculates the root mean square (RMS), integrated EMG value, baseline mean, and baseline drift for each window. The system sorts the RMS values ​​of each window from smallest to largest and takes the value in the middle of the sort as the current subject's resting RMS value; it also sorts the integrated EMG values ​​of each window from smallest to largest and takes the value in the middle of the sort as the current subject's resting integrated EMG value; and it uses the stable range of the baseline mean for each window as the resting baseline drift range.

[0032] Furthermore, the subject performs a baseline contraction movement, and the system collects at least 10 seconds of baseline contraction electromyographic (EMG) data. The baseline contraction movement is pre-defined in the experimental protocol and can be an isometric contraction, fist clenching, leg raising, or other standard movements corresponding to the target muscle. The system performs windowing processing on the baseline contraction EMG data, calculating the root mean square value, integrated EMG value, peak frequency, and spectral center position for each window. The system determines the baseline contraction intensity, baseline spectral center, and baseline peak frequency from the stable windows.

[0033] The baseline contraction intensity is determined by the median of the root mean square values ​​within the stabilization window. The baseline spectral center is determined by the center position of each frequency component within the stabilization window after energy weighting. The baseline peak frequency is determined by the frequency position with the highest energy within the stabilization window. These parameters are used in subsequent calculations of standardized muscle activation values, spectral migration values, and fatigue trend values.

[0034] In some embodiments, the system simultaneously acquires skin and electrode contact impedance in both the resting and baseline contraction states. The system sorts the stable contact impedance data from smallest to largest and takes the impedance value at the middle position as the contact impedance baseline value for the current subject. This contact impedance baseline value is used to subsequently determine the contact impedance warning threshold and participates in signal reliability calculation.

[0035] Therefore, this embodiment establishes an analytical reference by using the subject's own resting data and baseline contraction data, so that muscle state determination does not rely on a uniform fixed amplitude threshold. When different subjects perform the same action but have different original electromyographic amplitudes, the system can still perform standardized comparisons based on their respective baseline contraction intensities.

[0036] Example 3: In this example, environmental interference data is used to characterize interference components in the current acquisition environment that are unrelated to the target muscle. Environmental interference data includes at least one of the following: environmental noise data under electrode short-circuit conditions, environmental noise data under resting wearing conditions, or environmental reference signals acquired through the sentinel channel.

[0037] In this embodiment, environmental noise data is collected for at least 10 seconds with the detection electrodes short-circuited. Since there is no real electromyographic potential difference at the input when the detection electrodes are short-circuited, the collected data mainly reflects the noise of the acquisition circuit, power supply ripple, power frequency electromagnetic interference, and environmental harmonic interference. The system performs spectrum analysis on the environmental noise data to determine the amplitude of the 50 Hz power frequency interference, the amplitude of the 100 Hz harmonic interference, and the level of random noise.

[0038] Furthermore, at least 30 seconds of resting ambient noise data is collected while the electrodes are worn and the target muscles are relaxed. This data includes not only circuit noise and ambient electromagnetic noise, but also low-frequency disturbances and baseline drift caused by skin and electrode contact conditions. Based on this data, the system determines the baseline drift range, low-frequency disturbance range, and time-frequency texture background under the combined effects of the current environment and current skin contact conditions.

[0039] In another specific implementation, the system employs a sentinel channel. This sentinel channel is not used to acquire effective contraction information of the target muscle, but rather to acquire environmental reference signals. The system compares the environmental reference signals acquired by the sentinel channel with the signals from the target electromyography (EMG) channel within the same time window to identify synchronously occurring power frequency interference, low-frequency oscillations, or spike interference.

[0040] As can be seen, in this embodiment, the system can establish an interference benchmark corresponding to the current environment by using electrode short-circuit data, resting wear data, and sentry channel data. During the formal data acquisition process, the system does not simply determine whether a certain noise component exceeds a fixed threshold, but rather determines whether it is abnormally enhanced relative to the current environmental benchmark, thereby reducing false alarms and false negatives caused by different experimental environments.

[0041] Example 4: In this example, the input electromyography (EMG) values ​​are calculated from the original EMG sampling data using device calibration parameters. These parameters include zero-point offset, actual amplification factor, analog-to-digital conversion reference voltage, and reference voltage drift.

[0042] In this embodiment, the zero-point offset is determined using zero-point data with the detection electrode and reference electrode short-circuited. The system collects at least 10 seconds of zero-point data, discards transitional data from the initial stage, and averages the sampled values ​​during the stable phase to obtain the zero-point average value. The system then converts this zero-point average value into a zero-point offset voltage based on the analog-to-digital conversion reference voltage and the quantization bit depth. This zero-point offset voltage is used to eliminate fixed errors caused by amplifier circuit input offset and reference voltage offset.

[0043] Furthermore, the actual amplification factor is determined by inputting a calibration signal with a known amplitude and frequency. After the calibration signal is input to the acquisition front end, the system acquires the amplified output waveform and calculates the output amplitude during the stable phase of the waveform. The actual amplification factor is obtained by dividing the output amplitude by the known input amplitude. Subsequent acquired data are all converted using this actual amplification factor, rather than directly using the nominal amplification factor.

[0044] For each original sampled value, the system first multiplies it by the analog-to-digital conversion reference voltage, then divides it by the maximum value of the quantization range to obtain the output voltage. The system then subtracts the zero-point offset voltage and the reference voltage drift compensation from the output voltage to obtain the corrected output electromyography (EMG) voltage. Subsequently, the system divides the corrected output EMG voltage by the actual amplification factor to obtain the input EMG voltage value. This input EMG voltage value is stored in the microvolt range as input EMG data.

[0045] As shown above, through the calibration and conversion described, the system can unify the original sampled values ​​obtained under different devices, different amplification factors, and different analog-to-digital conversion reference voltages to the scale of the input electromyography value. This processing improves the consistency of values ​​in subsequent time-frequency image generation, standardized feature calculation, and model recognition.

[0046] Example 5: In this example, the surface electromyography (EMG) time-frequency image is a multi-layer time-frequency image. This multi-layer time-frequency image includes at least an EMG energy layer, a quality mask layer, and an environmental noise reference layer.

[0047] In this embodiment, the electromyography (EMG) energy layer is generated by performing short-time spectral transform or wavelet transform on the input EMG data. The system first divides the input EMG data into windows according to a preset window length, with each data window corresponding to a time interval. After removing the DC component from each data window, a window function is applied for weighting, followed by a short-time spectral transform or wavelet transform to obtain the correspondence between time position, frequency position, and energy intensity. The system maps the energy intensity to pixel intensity, forming an EMG energy layer where the horizontal axis represents time, the vertical axis represents frequency, and the pixel intensity represents EMG energy.

[0048] Furthermore, the quality mask layer is generated based on electrode contact impedance, wear detection status, signal confidence value, saturation sampling ratio, and baseline drift. If a data window corresponding to a certain time position has invalid contact impedance, invalid wear detection status, excessive saturation sampling ratio, or excessive baseline drift, then that time position is marked as an abnormal acquisition time position in the quality mask layer. If the signal confidence value of that window is lower than the valid signal confidence threshold, but does not meet the direct invalidation condition, then that time position is marked as a low-confidence time position.

[0049] Additionally, the environmental noise reference layer is generated based on environmental interference data or environmental reference signals collected through the sentinel channel. For data collected through the sentinel channel, the system uses the same windowing and time-frequency transformation methods as the target EMG channel to generate a noise time-frequency image. For environmental interference data obtained in the electrode short-circuit state and the resting wearing state, the system generates a baseline noise time-frequency distribution for the current environment. This layer is used to represent the noise time-frequency distribution unrelated to the target muscle.

[0050] In this embodiment, when the muscle state recognition model reads multi-layer time-frequency images, it extracts the effective frequency band energy distribution of electromyography (EMG) in the EMG energy layer, the location of low-confidence regions in the quality mask layer, and the location of noise textures in the environmental noise reference layer. When the EMG energy layer shows energy enhancement, and the quality mask layer does not mark anomalies at the corresponding time position, and the environmental noise reference layer does not show synchronous enhancement at the corresponding time position, the system uses this energy enhancement as a candidate feature of real muscle activity. If the EMG energy layer and the environmental noise reference layer enhance synchronously, or the quality mask layer marks anomalies at the corresponding time position, the system reduces the weight of this data window in muscle state recognition, or outputs a pending review status.

[0051] Example 6: In this example, the individual and environment adaptive threshold set includes at least the contact impedance warning threshold, power frequency interference deduction threshold, baseline drift deduction threshold, muscle activation threshold, fatigue spectrum migration threshold, and signal credibility effective threshold.

[0052] The contact impedance warning threshold is determined based on the median of the stable contact impedance in the current resting state of the subject. The system sorts the stable contact impedance values ​​from smallest to largest, takes the median as the contact impedance benchmark, and uses the value obtained by increasing this benchmark by 30% as the contact impedance warning threshold. If the increased value exceeds 50 kΩ, the contact impedance warning threshold is set to 50 kΩ. If the contact impedance is less than 5 kΩ or greater than 50 kΩ during the actual data acquisition process, the corresponding data window is directly identified as a contact anomaly window.

[0053] Furthermore, the power frequency interference deduction threshold is determined based on the power frequency interference amplitude under electrode short-circuit conditions or resting wearing conditions. The system uses the 50 Hz interference amplitude in the stable environmental noise window as the environmental power frequency reference. When the 50 Hz interference amplitude in the formal acquisition window exceeds twice the environmental power frequency reference, or exceeds the preset absolute upper limit, the signal reliability value of that window is deducted.

[0054] The muscle activation threshold is determined based on the current subject's root mean square (RMS) resting value and baseline contraction intensity. The system compares the difference between the current window's RMS value and the current subject's RMS resting value with the difference between the baseline contraction intensity and the RMS resting value. When the former reaches 20% or more of the latter, the window is considered to have met the muscle activation condition. This threshold can vary with different subjects' muscle response levels and resting noise levels.

[0055] In one alternative implementation, the fatigue spectral shift threshold is determined based on the baseline spectral center under the current subject's baseline contraction state. During formal data acquisition, if the spectral center of multiple consecutive valid data windows continuously decreases relative to the baseline spectral center, and the root mean square value or integrated electromyographic value continuously increases, the system determines that the fatigue trend is increasing. If only a single window shows a spectral decrease, but it is not accompanied by a continuous increase in amplitude, the fatigue level is not directly increased.

[0056] In addition, the effective threshold for signal reliability is determined based on the signal reliability distribution of stable data windows in resting electromyography (EMG) data and baseline contractile EMG data. The system filters data windows with effective impedance, no saturation sampling, and no abnormal enhancement of power frequency interference. The signal reliability values ​​of these windows are calculated, and the lowest stable value in the reliability distribution of stable windows is taken as the effective threshold for signal reliability. When this threshold is lower than 60 points, 60 points is taken as the lowest effective threshold.

[0057] Example 7: In this example, the signal confidence value is generated on a percentage basis. Each data window corresponds to a signal confidence value, which represents the confidence level of the result after the window enters the muscle state recognition model.

[0058] In this embodiment, when there are sampling points within the data window with a contact impedance less than 5 kΩ or greater than 50 kΩ, the system sets the signal reliability value of that data window to 0. A contact impedance less than 5 kΩ indicates that there may be a conductive dielectric bridging, electrode short circuit, or abnormally moist skin surface between the detection electrodes; a contact impedance greater than 50 kΩ indicates poor electrode contact, loose electrodes, or insufficient skin pretreatment. All of these conditions will result in the acquired signal not reliably representing the target muscle activity.

[0059] Furthermore, when the contact impedance is between 5 kΩ and 50 kΩ, the system sets the initial signal reliability score to 100 points. The system then sequentially checks the residual power frequency interference, residual harmonic interference, baseline drift, saturation sampling ratio, time-frequency texture distortion, and sentry channel synchronization interference, deducting points according to the degree of abnormality. The deducted score becomes the signal reliability value for that window.

[0060] Among them, the residual power frequency interference is determined by analyzing the spectral energy near 50 Hz and its harmonics; the baseline drift is determined by comparing the changes in the baseline average value of adjacent windows; the saturation sampling ratio is determined by statistically analyzing the proportion of the number of sampling points that reach the upper or lower limit of analog-to-digital conversion to the total number of sampling points in the window; and the degree of time-frequency texture distortion is determined by comparing the texture difference between the current window time-frequency image and the resting reference time-frequency image or the reference contracted time-frequency image.

[0061] In addition, the synchronization interference of the sentinel channel is determined by comparing the time-frequency changes of the target electromyography channel and the sentinel channel within the same data window. If both channels synchronously exhibit power frequency enhancement, low-frequency oscillation, or spike interference, the system considers that the window to have environmental interference or motion artifact risk and reduces the signal confidence value.

[0062] Example 8: In this example, the sentinel channel has switchable acquisition modes, including a static electromyography (EMG) experiment mode and a free-motion EMG experiment mode. The function of the sentinel channel is to provide environmental reference signals that are independent of the target muscle contraction.

[0063] In static electromyography (EMG) experiments, the sentinel channel is connected to an impedance network that matches the input impedance of the target EMG channel. At this time, the sentinel channel does not receive actual muscle contraction signals; its acquisition results primarily reflect electromagnetic interference, power supply ripple, and random noise in the acquisition circuit and the current environment. If the target EMG channel and the sentinel channel simultaneously exhibit power frequency enhancement within the same data window, the system marks this window as an environmental interference risk window.

[0064] Furthermore, in free-motion electromyography (EMG) experiments, the sentinel channel is positioned in an area with low muscle activity. This area is affected synchronously with human movement by wire swaying, overall movement, and skin slippage, but does not correspond to the main force exertion area of ​​the target muscle. If the target EMG channel and the sentinel channel exhibit low-frequency oscillations or spike interference synchronously, the system marks this window as a motion artifact risk window.

[0065] Therefore, this embodiment provides a synchronous reference through the sentinel channel, enabling the system to distinguish between real muscle activity and abnormal signals from non-muscle sources. When the target electromyography channel shows energy enhancement in the effective frequency band of electromyography, but the sentinel channel does not show synchronous enhancement, this change is retained as a candidate feature of real muscle activity; when both enhance synchronously, this change is identified as environmental interference or motion artifact risk.

[0066] Example 9: In this example, the standardized features include at least one of standardized muscle activation value, spectral migration value, and fatigue trend value.

[0067] The standardized muscle activation value is determined based on the root mean square (RMS) value of the current data window, the current subject's resting RMS value, and the current subject's baseline contraction intensity. The system first calculates the difference between the current window RMS value and the resting RMS value, then calculates the difference between the baseline contraction intensity and the resting RMS value. Finally, the ratio of the former to the latter is calculated to obtain the standardized muscle activation value. This value represents the proportion of the current muscle activity intensity relative to the current subject's baseline contraction level.

[0068] Furthermore, the spectral shift value is determined based on the spectral center position of the current data window and the baseline spectral center under the current subject's baseline contraction state. The system divides the difference between the baseline spectral center and the current window's spectral center by the baseline spectral center to obtain the spectral shift value. When the spectral shift value continues to increase while the standardized muscle activation value remains at a high level, the system considers the target muscle to have a tendency to increase fatigue.

[0069] The fatigue trend value is determined based on changes in the root mean square (RMS) value, integrated electromyography (EMG) value, downward shift of the spectral center, and duration of sustained contraction within a continuous data window. If the RMS value or integrated EMG value continuously increases across multiple consecutive valid data windows, while the spectral center continuously shifts downward relative to the baseline spectral center, and the target muscle is in a state of sustained contraction, the fatigue trend value increases. If the target muscle is in a relaxed state and the RMS value returns to near the resting baseline, the fatigue trend value decreases.

[0070] In this embodiment, the muscle state recognition model takes surface electromyography (EMG) time-frequency images, standardized features, and signal confidence values ​​as input. Based on the effective frequency band energy distribution of EMG in the time-frequency image, standardized muscle activation values, spectral shift values, fatigue trend values, and signal confidence values, the model outputs muscle activation intensity, fatigue level, or movement state. If the signal confidence value is lower than the effective signal confidence threshold, the model does not output a definitive muscle state result but instead outputs a state pending verification.

[0071] Example 10: In this example, abnormal conditions include a jump in the level of muscle state results within a preset time, a signal confidence value lower than the effective threshold of signal confidence, or a standardized feature exceeding the physiologically reasonable range corresponding to the current subject.

[0072] Among these, a jump in muscle state results within a preset time period refers to a fatigue level that jumps by more than two levels within an adjacent second, or a movement state that repeatedly switches between multiple consecutive data windows. Standardized features exceeding the physiologically reasonable range corresponding to the current subject refers to standardized muscle activation values ​​exceeding twice the reference level corresponding to the current subject's baseline contraction intensity, and the corresponding time-frequency image not showing a sustained increase in the effective electromyographic frequency band; or spectral shift values ​​showing a large change in a single window, but without a continuous trend between preceding and following windows.

[0073] Furthermore, when abnormal conditions are met, the system triggers reverse verification processing. This reverse verification process reads the original electromyography (EMG) sampling data, electrode contact impedance, wearing detection status, input EMG values, surface EMG time-frequency images, environmental interference data, and individual / environment adaptive threshold sets from the corresponding data window. The system then re-evaluates whether the window exhibits abnormal contact impedance, abnormal power frequency interference, abnormal baseline drift, abnormal saturation sampling, or abnormal time-frequency texture.

[0074] If any of the above anomalies are found after reassessment, the system marks the data window as an abnormal acquisition window and marks the corresponding muscle state result as unsuitable for direct statistical analysis. If no anomalies are found, the system re-executes standardized feature calculation and muscle state recognition, and uses the re-recognition result as the final muscle state result for that window.

[0075] Therefore, this embodiment enables backend anomaly detection to trigger frontend data quality review. This method differs from a one-way pipeline-style identification process and can reduce misjudgments caused by loose electrodes, swaying wires, sudden increases in environmental noise, or short-term motion artifacts.

[0076] Example 11: In this example, the system binds and stores the current subject's individual baseline data, the environmental interference data of the current acquisition environment, the device calibration parameters, the individual and environmental adaptive threshold sets, the surface electromyography time-frequency image, the signal confidence value, the muscle state results, and the reverse verification mark into the same experimental data package.

[0077] Individual baseline data includes resting root mean square value, resting integrated electromyography value, resting baseline drift range, baseline contraction intensity, baseline spectral center, baseline peak frequency, and contact impedance baseline value. Environmental interference data includes power frequency interference amplitude, harmonic interference amplitude, random noise level, low-frequency disturbance range, and time-frequency texture background. Equipment calibration parameters include zero-point offset, actual amplification factor, analog-to-digital conversion reference voltage, and reference voltage drift.

[0078] Furthermore, when the experimental data package is used for subsequent model retraining, the system first removes data windows whose signal confidence values ​​are below the effective threshold for signal confidence. For data windows marked as abnormal acquisition during reverse verification, the system uses them as negative samples or low-weight samples to train the model to identify time-frequency texture features corresponding to electrode loosening, wire swaying, skin sweating, motion artifacts, and abnormal environmental noise. For data windows with high signal confidence, passing reverse verification, and stable muscle state results, the system uses them as high-weight samples to update the muscle state recognition model.

[0079] In addition, the experimental data package is also used for subsequent experimental verification. When a muscle state result is questioned, the system can trace back the original sampling data, input electromyography data, time-frequency image, quality mask layer, environmental noise reference layer, signal confidence value, and reverse verification flag corresponding to that result. Thus, researchers can determine whether the result originates from real muscle activity or from abnormal acquisition or environmental interference.

[0080] Table 1. Comparison of data between the present invention and traditional methods

[0081] As shown in Table 1, the comprehensive muscle state recognition accuracy rate represents the proportion of correct judgments made by the system regarding muscle activation intensity, fatigue level, and movement state. Traditional methods, relying primarily on fixed thresholds and single-layer time-frequency images, are prone to misinterpreting noise enhancement, motion artifacts, or abnormal electrode contact as changes in muscle state, resulting in an accuracy rate of 83.1%. The method of this invention introduces the current subject's individual baseline data and current environmental interference data before recognition, and combines signal confidence values ​​and multi-layer time-frequency images during the recognition process, improving the recognition accuracy to 93.2%. Traditional methods lack environmental reference and reverse verification mechanisms, resulting in an average false positive rate of 20.0%. The method of this invention identifies non-muscle source interference through quality mask layers, environmental noise reference layers, and sentinel channels, and deweights or re-verifies low-confidence data windows, reducing the false positive rate to 6.4%. Low-quality windows include data windows with abnormal contact impedance, excessive baseline drift, excessively high power frequency residue, excessively high saturation sampling ratio, and significant time-frequency texture distortion. Traditional methods typically rely solely on impedance thresholds or filtering results to determine data validity, resulting in a detection rate of 62.5%. This invention's method comprehensively considers electrode contact impedance, wearing detection status, environmental interference data, baseline drift, saturation sampling ratio, time-frequency texture distortion degree, and sentinel channel synchronization interference to calculate signal reliability values, increasing the low-quality window detection rate to 92.1%. A smaller value indicates better cross-subject consistency. Traditional methods use a uniform fixed threshold, which is difficult to adapt to differences in skin impedance, muscle strength, subcutaneous fat thickness, and exertion habits among different subjects, resulting in a coefficient of variation of 24.6%. This invention's method establishes an individualized baseline based on the current subject's resting electromyography (EMG) data and baseline contractile EMG data, and calculates standardized muscle activation values ​​and spectral migration values, reducing the cross-subject result coefficient of variation to 10.8%. Traditional methods typically classify individual time-frequency image windows independently, making them susceptible to transient noise, posture changes, or motion artifacts, leading to abnormal jumps in fatigue levels within a short period, with a test result of 14 times / hour. This invention combines changes in the root mean square value of a continuous data window, changes in integrated electromyography (EMG) values, the downward shift of the spectral center, and the duration of sustained contraction to determine fatigue trends. It also triggers reverse verification upon abnormal jumps, reducing the number of abnormal jumps to 3 times per hour. Traditional methods typically only save the original waveform or the final identification result, making it difficult to determine whether the anomaly originates from electrode contact, environmental interference, data saturation, model misidentification, or actual muscle state changes; therefore, the traceability rate is only 35.0%. This invention binds and stores the original EMG sampling data, input EMG value data, surface EMG time-frequency images, signal confidence values, individual-environment adaptive threshold sets, muscle state results, and reverse verification markers, increasing the traceability rate of abnormal data to 96.4%.

[0082] In summary, the above comparative results show that the method of the present invention can improve the accuracy of surface electromyography time-frequency image recognition, reduce the misjudgment rate in complex interference scenarios, improve the ability to detect low-quality data, and enhance the continuity of fatigue state analysis and the traceability of abnormal data under different subjects, different environments and conditions with acquisition disturbances.

[0083] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for muscle state analysis based on surface electromyography time-frequency image recognition, characterized in that, Includes the following steps: S1. Obtain the individual baseline data of the current subject and the environmental interference data of the current collection environment; S2. Acquire surface electromyographic signals of the target muscle and simultaneously obtain electrode contact impedance, wearing detection status and sampling timestamp to form raw electromyographic sampling data; S3. Based on the device calibration parameters, convert the raw electromyography sampling data into input electromyography value data. S4. Perform windowing and time-frequency transformation on the input electromyography data to generate a surface electromyography time-frequency image; S5. Determine the individual and environmental adaptive threshold sets based on the individual baseline data and environmental interference data, and calculate the signal confidence value of each data window by combining the electrode contact impedance, wearing detection status, baseline drift, saturation sampling ratio and time-frequency texture distortion. S6. Input the surface electromyography time-frequency image, the standardized features obtained from the electromyography data at the input end, and the signal confidence value into the muscle state recognition model, and output the muscle activation intensity, fatigue level, or movement state. S7. When the output results meet the abnormal conditions, perform reverse verification on the corresponding data window and store the analysis data, recognition results and reverse verification flags.

2. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 1, characterized in that, The individual baseline data mentioned in S1 includes the current subject's resting electromyography data, baseline systolic electromyography data, and skin and electrode contact impedance data; The environmental interference data includes at least one of the following: environmental noise data under electrode short-circuit condition, environmental noise data under resting wearing condition, or environmental reference signal collected by the sentry channel.

3. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 2, characterized in that, The device calibration parameters include zero-point offset, actual amplification factor, analog-to-digital conversion reference voltage, and reference voltage drift; the input electromyography data is obtained as follows: Zero-point data is collected when the detection electrode and the reference electrode are short-circuited, and the zero-point offset is determined based on the sampled value of the stable segment. Input a calibration signal with a known amplitude and a known frequency to the acquisition input terminal, and determine the actual amplification factor based on the acquired output amplitude and the known amplitude; For each raw sampled value, it is first converted into the output voltage based on the analog-to-digital conversion reference voltage and the number of quantization bits, then the voltage value corresponding to the zero offset is subtracted, and then divided by the actual amplification factor to obtain the input electromyography voltage value. The input electromyography voltage value is saved as input electromyography value data in the microvolt range.

4. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 1, characterized in that, The surface electromyography time-frequency image is a multi-layer time-frequency image, which includes at least an electromyography energy layer, a mass mask layer, and an environmental noise reference layer. The electromyographic energy layer is generated by performing short-time spectral transformation or wavelet transform on the input electromyographic data, and is used to represent the time-frequency energy distribution of the target muscle. The quality mask layer is generated based on electrode contact impedance, wear detection status, signal confidence value, saturation sampling ratio, and baseline drift, and is used to mark low confidence time locations and abnormal acquisition time locations. The environmental noise reference layer is generated based on environmental interference data or environmental reference signals collected by the sentry channel, and is used to represent the time-frequency distribution of noise unrelated to the target muscle in the current collection environment.

5. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 1, characterized in that, The individual and environment adaptive threshold set includes at least the contact impedance warning threshold, power frequency interference deduction threshold, baseline drift deduction threshold, muscle activation threshold, fatigue spectrum migration threshold, and signal credibility effective threshold. The contact impedance warning threshold is determined based on the median of the stable segment contact impedance under the current resting state of the subject. The power frequency interference reduction threshold is determined based on the power frequency interference amplitude under electrode short-circuit state or resting wearing state; The muscle activation threshold is determined based on the current subject's resting root mean square value and baseline contraction intensity; The fatigue spectrum migration threshold is determined based on the center of the reference spectrum under the current subject's reference contraction state. The effective threshold for signal reliability is determined based on the signal reliability distribution of stable data windows in resting electromyography data and baseline contractile electromyography data.

6. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 1, characterized in that, The sentry channel has a switchable acquisition mode; In static electromyography experiments, the sentinel channel is connected to an impedance network that matches the input impedance of the target electromyography channel, and is used to acquire reference signals of electromagnetic interference from the acquisition circuit and the current environment. In free-movement electromyography experiments, the sentinel channel is placed in an area with low muscle activity to collect motion artifact reference signals caused by wire swaying, overall human movement, and electrode slippage relative to the skin. When the target electromyography channel and the sentinel channel exhibit synchronous power frequency enhancement, synchronous low-frequency oscillation, or synchronous spike interference within the same data window, this data window is marked as an environmental interference risk window or a motion artifact risk window.

7. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 6, characterized in that, The muscle state recognition model reads the multi-layer time-frequency image and extracts the effective frequency band energy distribution of electromyography in the electromyography energy layer, the low confidence region location in the quality mask layer, and the noise texture location in the environmental noise reference layer, respectively. When the electromyography energy layer shows energy enhancement and the quality mask layer does not mark the anomaly at the corresponding time position, and the environmental noise reference layer does not show synchronous enhancement at the corresponding time position, the energy enhancement is taken as a candidate feature of real muscle activity. When the electromyography energy layer and the environmental noise reference layer show synchronous enhancement at the corresponding time position, or when the quality mask layer is marked abnormally at the corresponding time position, reduce the weight of this data window in muscle state recognition or output a state pending review.

8. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 1, characterized in that, The standardized features include at least one of standardized muscle activation values, spectral migration values, and fatigue trend values; The standardized muscle activation value is determined based on the root mean square value of the current data window, the current subject's resting root mean square value, and the current subject's baseline contraction intensity. The spectral shift value is determined based on the spectral center position of the current data window and the baseline spectral center under the current subject baseline contraction state; The fatigue trend value is determined based on the changes in the root mean square value of the continuous data window, the changes in the integral electromyography value, the downward shift of the spectrum center, and the duration of sustained contraction.

9. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 1, characterized in that, The abnormal conditions include a jump in the level of muscle state results within a preset time, a signal confidence value lower than the effective threshold of signal confidence, or a standardized feature exceeding the physiologically reasonable range corresponding to the current subject. The reverse verification includes reading the original electromyography (EMG) sampling data, electrode contact impedance, wearing detection status, input EMG value data, surface EMG time-frequency image, environmental interference data, and individual and environmental adaptive threshold sets of the corresponding data window, and re-determining whether the data window has abnormal contact impedance, abnormal power frequency interference, abnormal baseline drift, abnormal saturation sampling, or abnormal time-frequency texture. If any of the above abnormalities exist, the data window is marked as an abnormal acquisition window. If no such abnormality exists, the standardized feature calculation and muscle state recognition are re-executed.

10. The muscle state analysis method based on surface electromyography time-frequency image recognition according to claim 1, characterized in that, It also includes individual parameter binding and storage, and model retraining steps; The individual baseline data of the current subject, the environmental interference data of the current acquisition environment, the equipment calibration parameters, the individual and environmental adaptive threshold sets, the surface electromyography time-frequency images, the signal confidence value, the muscle state results, and the reverse verification mark are bound and stored in the same experimental data package; When the experimental data package is used for subsequent model retraining, data windows with signal confidence values ​​lower than the effective threshold of signal confidence are removed, and data windows marked as abnormal acquisition in reverse verification are used as negative samples or low-weight samples. The muscle state recognition model is updated using the retained data window and its corresponding individual baseline data and environmental interference data.

Images