Surface electromyography signal processing device and processing method for same

Abstract

Provided is a surface electromyography signal processing method for extracting an electromyography component from a surface electromyography signal measured on the skin of a subject over a prescribed time, and removing a noise component. The method comprises: a decomposition step for dividing the measured surface electromyography signal into a plurality of epochs at prescribed unit times, and decomposing each of the epochs into a plurality of spectral components and weight components indicating weights of the respective epochs in each spectral component; an electromyography component identification step for comparing a reference spectral distribution determined by modeling the electromyography with the spectral components, and identifying, as an electromyography component, a spectral component similar to one reference spectral component included in the reference spectral distribution; a filter generation step for generating a filter for reducing a contribution rate of the noise component to each epoch by using, as the noise component, a spectral component other than the electromyography component among the spectral components decomposed in the decomposition step; a noise component removal step for removing the noise component of each epoch by using the generated filter; and a reconstruction step for recombining each epoch after processing in the noise component removal step, and reconstructing an electromyography signal from which the noise component is removed.

Description

Surface electromyography signal processing device and processing method thereof

[0001] The present invention relates to a method and apparatus for processing surface electromyography (sEMG) signals, and more particularly to a method and apparatus for extracting electromyography components and reducing noise components.

[0002] Techniques are known for effectively separating electromyographic signals and external force (impact) information from signals detected by a single sensor (for example, Patent Document 1).

[0003] Japanese Patent Publication No. 2018-134189

[0004] Surface electromyography (EMG) signals are used to measure the electrical activity of muscles and reflect the state of muscle contraction. However, these signals are affected by many external noise sources, so denoising the signals is crucial for accurate analysis. Conventional techniques use filtering and signal decomposition to remove noise, but these techniques have not been sufficient to properly separate EMG components from noise components.

[0005] The present invention has been made in view of the above-mentioned points. Its objective is to provide a surface electromyography signal processing device and a processing method that can appropriately separate a surface electromyography signal into electromyography components and noise components.

[0006] The surface electromyography signal processing method and processing apparatus according to the present invention are for extracting electromyography components from surface electromyography signals measured over a predetermined time period on the skin of a subject and removing noise components, comprising: a decomposition step (spectral component-weight component decomposition step) or decomposition unit that divides the measured surface electromyography signal into a plurality of epochs at predetermined unit time intervals, and decomposes each of the epochs into a plurality of spectral components and a weight component that indicates the respective weight of each of the epochs in each of the spectral components; an electromyography component identification step or identification unit that compares the spectral components with a reference spectral distribution predetermined by modeling electromyography in advance, and identifies spectral components similar to one reference spectral component included in the reference spectral distribution as electromyography components; and a filter generation step (noise component reduction filter generation step) or filter generation unit that generates a filter that reduces the contribution rate of the noise component to each of the epochs, treating spectral components other than the electromyography components among the decomposed spectral components as noise components. The system includes: a noise component removal step (noise component removal signal processing step) or signal processing unit that removes the noise component from each of the epochs using the generated filter; and a reconstruction step (clean surface electromyography signal reconstruction step) or reconstruction unit that recombines each of the epochs after the noise component removal process and reconstructs the surface electromyography signal from which the noise component has been removed.

[0007] Surface electromyography signals can be appropriately separated into electromyographic components and noise components.

[0008] This is a functional block diagram showing the functions of the surface electromyography signal processing device 10. This is a block diagram illustrating the concept of the process of decomposing a signal into spectral components and weighted components. This is a diagram showing an example of a signal from the surface electromyography signal to its decomposition into spectral components and weighted components. This is a list showing a matrix (Equation 1) representing the spectral components H, a matrix (Equation 2) representing the weighted components W, an equation (Equation 3) showing the relationship between spectral components H and weighted components W and the epoch, and an equation (Equation 4) showing the matrix representing the epoch. This is a graph showing 10 reference spectral components M1 to M10 as an example. This is a diagram comparing spectral components (H1, H2, ...) and reference spectral components M (M1, M2, ...). This is a flowchart showing the process of determining electromyography components and generating filters by non-negative matrix factorization. This is a graph showing the process of generating filters according to frequency. This is a block diagram showing the process of generating noise-removed surface electromyography signal data by noise removal filter (noise component reduction filter), inverse FFT, and reconstruction. This is a diagram showing an example of highlighting.

[0009] <<<<<Outline of this Embodiment>>>> <<<First Feature>>> According to the first feature, a surface electromyography signal processing method for extracting electromyography components from a surface electromyography signal measured on the skin of a subject over a predetermined period of time and removing noise components, comprising: a decomposition step of dividing the measured surface electromyography signal into a plurality of epochs at predetermined unit time intervals, and decomposing each of the epochs into a plurality of spectral components and a weight component indicating the respective weight of each of the epochs in each of the spectral components; an electromyography component identification step of comparing the spectral components with a reference spectral distribution predetermined by modeling electromyography in advance, and identifying a spectral component similar to one reference spectral component included in the reference spectral distribution as an electromyography component; a filter generation step of generating a filter that reduces the contribution rate of the noise component to each of the epochs, treating the spectral components other than the electromyography component among the spectral components decomposed in the decomposition step as noise components; and a noise component removal step of removing the noise component to each of the epochs using the generated filter, A surface electromyography signal processing method is provided, which includes a reconstruction step of recombining each of the epochs after the processing of the noise component removal step to reconstruct the surface electromyography signal from which the noise component has been removed.

[0010] Decomposition based on non-negative matrix factorization allows for the separation of electromyographic (EMG) components from noise components, enabling more accurate extraction of surface EMG than conventional filtering methods. Furthermore, comparison with a reference spectral distribution allows for the selection of spectral components similar to the modeled EMG, reducing misinterpretations.

[0011] Because it generates noise reduction filters on an epoch-by-epoch basis, it can adapt to time-varying noise. Furthermore, by applying filters that consider the noise contribution rate for each frequency, it can appropriately remove different types of noise in each frequency band. Compared to general fixed filters (bandpass, lowpass, etc.), it can dynamically reduce noise components, resulting in cleaner surface electromyography signals.

[0012] Since decomposition is performed epoch by epoch, short-term changes in muscle activity can also be analyzed in detail.

[0013] Since automatic identification and removal of noise components can be achieved, manual signal processing is reduced, and the method can be applied to automatic analysis and large-scale data analysis. Since the filter is optimized for each epoch, it can adapt to various situations without fine-tuning of settings.

[0014] By reconstructing the signal after noise removal, a clean surface myoelectric potential signal in the time domain can be obtained.

[0015] <<<<Second Feature>>>> The second feature is that, in the first feature, the myoelectric potential component identifying step identifies, as the myoelectric potential component, a spectral component whose similarity between the reference spectral component and the spectral component exceeds a similarity threshold.

[0016] By comparing with the reference spectrum, myoelectric potential components can be clearly distinguished from noise. Furthermore, incorrect identification can be reduced.

[0017] <<<<Third Feature>>>> The third feature is that, in the first feature, the decomposition into the spectral components and the weight components decomposed in the decomposition step is performed using a non-negative matrix factorization algorithm.

[0018] The non-negative matrix factorization algorithm expresses a surface myoelectric potential signal as a linear combination of only non-negative components, and therefore can appropriately reflect the physical characteristics of the surface myoelectric potential signal.

[0019] <<<<Fourth Feature>>>> The fourth feature is that, in the second feature, the similarity threshold is determined according to characteristics of the measured surface myoelectric potential signal in order to optimize the identification accuracy of the myoelectric potential component.

[0020] Since the threshold is determined according to the characteristics of the surface myoelectric potential signal, it is less susceptible to the influence of the environment and signal quality. In a high-noise environment, the threshold can be set high to ensure the purity of the surface myoelectric potential signal. Since the threshold can be adjusted according to the intensity of the myoelectric potential, the invention can be applied to various measurement conditions.

[0021] <<<Fifth Feature>>> The fifth feature is characterized in that, based on the fourth feature, the similarity threshold is determined in accordance with the myoelectric potential for defining the reference spectral distribution, the myoelectric potential model for defining the reference spectral distribution, or the measurement conditions of the surface myoelectric potential signal, so as to optimize the identification accuracy of the myoelectric potential component.

[0022] Here, "the myoelectric potential for defining the reference spectral distribution" refers to differences in the muscles themselves, such as differences between humans and animals, or differences in target regions even for the same organism. Furthermore, "the myoelectric potential model for defining the reference spectral distribution" refers to differences between models (such as differences in theoretical formulas for modeling). "Characteristics of the surface myoelectric potential signal" refers to differences in measurement devices, measurement conditions, hardware, and the like.

[0023] <<<Sixth Feature>>> The sixth feature is characterized in that, based on the first feature, the reference spectral distribution is generated in advance by reproducing modeled myoelectric potential through simulation.

[0024] The use of the reference spectrum optimized by simulation can improve the extraction accuracy of surface myoelectric potential signals.

[0025] <<<Seventh Feature>>> The seventh feature is that, in the second feature, the electromyographic component identification step includes: a selection step of selecting a selection number of spectral components for similarity determination from the plurality of spectral components for each of the plurality of epochs; a similarity calculation step of calculating the similarity between each of the spectral components decomposed according to the selection number and a reference spectral component included in the reference spectral distribution; a spectral component existence determination step of determining whether or not there is a spectral component whose similarity of each of the decomposed spectral components exceeds the similarity threshold; an iterative process of repeatedly executing the selection step, the similarity calculation step, and the spectral component existence determination step while changing the selection number; and an electromyographic component number determination step of determining the spectral components that exceed the similarity threshold as the electromyographic components and ending the iterative process, according to the number of spectral components that exceed the similarity threshold obtained by executing the iterative process and the similarity of the spectral components that exceed the similarity threshold.

[0026] By repeatedly performing identification while varying the number of selections, the optimal decomposition result can be obtained depending on the situation. It does not depend on a fixed number of spectral components and can be optimized according to the characteristics of the data.

[0027] By repeatedly performing similarity assessments, misclassification of spectral components containing noise can be minimized. Instead of making an immediate judgment based on a single threshold, iterative processing to find the optimal number of spectral components allows for the extraction of highly accurate surface electromyographic signals.

[0028] Furthermore, it prevents excessive breakdown of spectral components and avoids the over-decomposition of muscle fibers.

[0029] <<<Feature 8>>> The eighth feature is that, in the first feature, each of the epochs is generated by dividing the measured surface electromyographic signal so that some time overlaps.

[0030] This allows for the capture of continuous signal changes without loss of time information between epochs. Because signal information near the end of an epoch can also be considered, rapid changes (such as the start and end of muscle contraction) can be appropriately captured.

[0031] <<<Ninth Feature>>> The ninth feature is that, in the first feature, before processing the decomposition step, a frequency data generation step is further provided, in which a fast Fourier transform is performed on the measured surface electromyographic signal to generate frequency domain data.

[0032] While electromyographic components and noise components are difficult to distinguish in the time domain, they exhibit characteristic spectral distributions in the frequency domain, allowing for more accurate identification.

[0033] <<<Tenth Feature>>> The tenth feature is that, in the first feature, the reconstruction step converts the surface electromyographic signal for frequency in each epoch into a surface electromyographic signal for time in each epoch using an inverse fast Fourier transform.

[0034] By converting the surface electromyography (EMG) signal, filtered in the frequency domain, into a time domain, a clean surface EMG signal can be obtained. This allows for the generation of an appropriate time signal for each epoch, minimizing gaps between epochs.

[0035] <<<Feature No. 11>>> The eleventh feature is a step in the tenth feature of displaying a reconstructed surface electromyographic signal, further comprising a display step of displaying the period during which the electromyographic component exceeds a predetermined display threshold in a manner different from other periods.

[0036] Visually highlighting periods of high electromyography activity allows for intuitive analysis.

[0037] <<<Feature No. 12>>> The twelfth feature is that, in the eleventh feature, the display step displays the period in which the proportion to which the electromyographic component contributes to the reconstructed surface electromyographic signal exceeds the display threshold in a manner different from other periods.

[0038] This allows for clear identification of the period in which electromyography (EMG) is dominant within the surface EMG signal, making analysis easier.

[0039] <<<Feature No. 13>>> According to Feature No. 13, a surface electromyography signal processing device is provided for extracting electromyography components from a surface electromyography signal measured over a predetermined period of time and removing noise components, comprising: a decomposition unit that divides the measured surface electromyography signal into a plurality of epochs at predetermined unit time intervals, and decomposes each of the epochs into a plurality of spectral components and a weighting component that indicates the weight of each of the epochs in each of the spectral components; an identification unit that compares the spectral components with a reference spectral distribution predetermined by modeling electromyography in advance, and identifies a spectral component similar to one reference spectral component included in the reference spectral distribution as an electromyography component; a filter generation unit that generates a filter that reduces the contribution rate of the noise component to each of the epochs, treating the spectral components other than the electromyography component among the spectral components decomposed by the decomposition unit as noise components; a signal processing unit that removes the noise component using the generated filter; and a reconstruction unit that recombines each of the epochs after processing by the signal processing unit to reconstruct a clean surface electromyography signal.

[0040] <<<<<Details of this embodiment>>>>> The embodiment will be described below with reference to the drawings.

[0041] <<<<Surface electromyography signal processing device 10>>>> <Hardware configuration of surface electromyography signal processing device 10> The surface electromyography signal processing device 10 is composed of a processor, RAM, ROM, HDD, etc. (Figure 1). The surface electromyography signal processing device 10 may also have I / O (input / output interface), communication I / F (communication interface device), etc.

[0042] <Processor> A processor (such as a CPU) is a central processing unit that performs calculations for signal processing and other tasks.

[0043] <RAM (Random Access Memory)> RAM is a memory device used as a temporary storage area for data. RAM temporarily stores values ​​such as variables that are used when a program is executed.

[0044] <ROM (Read-Only Memory)> ROM is a memory device used to store program code and configuration information. Specifically, ROM stores various programs and constants for executing various processes (for example, the process shown in Figure 7).

[0045] <HDD (Hard Disk Drive)> An HDD is a storage device for long-term data storage. It stores programs executed in RAM and various other types of data.

[0046] <<<Functional Configuration of Surface Electromyography Signal Processing Device 10>>> The surface electromyography signal processing device 10 executes a surface electromyography signal processing method.

[0047] The surface electromyography signal processing device 10 includes a spectral component-weighting component decomposition unit 100, an electromyography component identification unit 200, a noise component reduction filter generation unit 300, a noise component removal signal processing unit 400, and a clean surface electromyography signal reconstruction unit 500.

[0048] <<Spectral component-weight component decomposition unit 100>> The spectral component-weight component decomposition unit 100 performs a process (decomposition step) that decomposes the surface electromyographic signal (Figure 3(a)) into spectral components and weight components.

[0049] The spectral component-weight component decomposition unit 100 divides the measured surface electromyographic signal (Figure 3(a)) into multiple epochs (e.g., epoch 1 to epoch E) at regular intervals w (Figures 2 and 3(b)). The epochs have an overlap length o. The overlap length o is usually set to w / 2. The overlap length o affects the trade-off between time accuracy and computation time. A small w improves time accuracy but increases the computational load.

[0050] The spectral component-weight component decomposition unit 100 uses a Fast Fourier Transform (FFT) to convert each epoch into a spectrum (frequency domain) (for example, spectrum 1 to spectrum E) (Figure 3(c)). Next, the spectral component-weight component decomposition unit 100 decomposes the spectrum into multiple spectral components H and multiple weight components W by non-negative matrix factorization (NMF) processing (Figure 3(d)). By decomposing the spectrum into multiple spectral components H and multiple weight components W, the frequency characteristics of the surface electromyographic signal and the characteristics of its weights can be extracted.

[0051] The multiple spectral components H are H1, H2, ... and are generally represented by the matrix shown in (Equation 1) in Figure 4. The multiple weight components W are generally represented by the matrix shown in (Equation 2) in Figure 4.

[0052] The spectral component H represents the frequency component of each epoch (time segment). The spectral component H includes the spectral characteristics of both the surface electromyographic signal and the noise component. The weighting component W indicates the weight of how much each epoch contributes to each spectral component.

[0053] By non-negative matrix factorization, the epoch is expressed as a linear combination of spectral components H and weight components W (see (Equations 3) and (4) in Figure 4). For example, the e-th epoch e is given by: Epoch e = H1 × We １ +H2×We ２ +...+HX×We Ｘ This is expressed as follows: X is the maximum number of components in the linear combination, and when determining similarity, it becomes the number of selections used for similarity determination.

[0054] The non-negative matrix factorization process is performed multiple times, as shown in step S713. Because the spectral component H and weight component W are non-negative after the non-negative matrix factorization process, the physical characteristics of the electromyographic signal can be appropriately reflected.

[0055] <<Electromyographic Component Identification Unit 200>> The electromyographic component identification unit 200 performs a process (electromyographic component identification step) to identify electromyographic components from the spectral component H and weight component W decomposed by the spectral component-weight component decomposition unit 100.

[0056] A myoelectric potential component identification unit 200 includes a spectrum component selection unit 210, a similarity calculation unit 220, a spectrum component presence determination unit 230, and a myoelectric potential component number determination unit 240.

[0057] FIG. 7 is a flowchart showing myoelectric potential component identification processing executed by the myoelectric potential component identification unit 200.

[0058] <Spectrum Component Selection Unit 210> The spectrum component selection unit 210 determines a selection number X (step S711 in FIG. 7), executes non-negative matrix factorization processing according to the selection number X, and generates X spectral components H for each epoch from spectrum 1 to spectrum E (step S713 in FIG. 7).

[0059] For example, if the selection number X for similarity determination is 2 (X=2), the spectrum component selection unit 210 generates two spectral components H1 and H2. In the case of five epochs from epoch 1 to epoch 5, each of the five epochs from epoch 1 to epoch 5 includes the two spectral components H1 and H2, and a weight component W １１ , W １２ , W ２１ , W ２２ , W ３１ , W ３２ , W ４１ , W ４２ , W ５１ , W5 ２ by linear combination, it is determined that: Epoch 1 = H1 × W １１ + H2 × W １２ Epoch 2 = H1 × W ２１ + H2 × W ２２ Epoch 3 = H1 × W ３１ + H2 × W ３２ Epoch 4 = H1 × W ４１ + H2 × W ４２ Epoch 5 = H1 × W ５１ + H2 × W ５２

[0060] Furthermore, the spectral component selection unit 210 generates three spectral components H1, H2, and H3 if the number of selections X for similarity determination is 3 (X = 3). In the case of five epochs 1 to 5, each of the five epochs 1 to 5 consists of three spectral components H1, H2, and H3 and a weight component W. １１ , W １２ , W ２１ , W ２２ , W ３１ , W ３２ , W ４１ , W ４２ , W ５１ , W ５２ Through a linear combination, Epoch 1 = H1 × W １１ +H2×W １２ +H3×W １３ Epoch 2 = H1 × W ２１ +H2×W ２２ +H3×W ２３ Epoch 3 = H1 × W ３１ +H2×W ３２ +H3×W ３３ Epoch 4 = H1 × W ４１ +H2×W ４２ +H3×W ４３ Epoch 5 = H1 × W ５１ +H2×W ５２ +H3×W ５３ This is determined.

[0061] The spectral component selection unit 210 sets the initial value of the number of selections X for similarity determination to 2, and increases the number of selections X for similarity determination by 1 each time until a predetermined condition is met (step S719 in Figure 7). Since the spectral component H and weight component W are initialized with random values, they will have different values ​​each time the non-negative matrix factorization process is performed.

[0062] <Reference Spectral Distribution> Figure 5 is a graph showing an example of a reference spectral distribution. Surface electromyography (EMG) signals are potentials based on muscle activity and are distributed within a specific frequency range (usually 10 Hz to 500 Hz). Therefore, the spectral distribution of surface EMG signals has certain characteristics and depends on the potential propagation velocity of muscle fibers. When a muscle contracts, the potential propagates through the muscle fibers. This propagation velocity varies from person to person and from muscle to muscle.

[0063] Based on this conduction velocity, the modeled electromyographic potential can be reproduced through simulation to generate a spectrum of the surface electromyographic signal. This spectral distribution of the modeled surface electromyographic signal is designated as the reference spectral distribution (see Figure 5). The reference spectral distribution shown in Figure 5 consists of 10 reference spectral components M1 to M10. By comparing spectral component H with the reference spectral component M, it is possible to determine whether spectral component H is an electromyographic component or a noise component. The reference spectral distribution is generated in advance and stored in an HDD or RAM.

[0064] <Similarity Calculation Unit 220> The similarity calculation unit 220 calculates the similarity between each of the generated spectral components H and the reference spectral component M included in the reference spectral distribution (step S715 in Figure 7). The similarity calculation unit 220 compares the generated spectral components (H1, H2, ...) with the reference spectral component M (M1, M2, ...) (Figure 6). The correlation coefficient is used for this comparison. The correlation coefficient indicates the similarity between two spectral components and the reference spectral component, and its value ranges from -1 (negative correlation) to 1 (perfect correlation).

[0065] If the correlation coefficient is close to 1, spectral component H is determined to be a surface electromyographic signal. On the other hand, if the correlation coefficient is 0 or negative, spectral component H is determined to be a noise component. The correlation coefficient of each spectral component (H1, H2, ...) with a reference spectral component M (M1, M2, ...) is calculated, and the spectral component H showing the highest correlation is selected as the surface electromyographic signal. Based on the correlation coefficient, spectral component H is distinguished as either an electromyographic component or a noise component.

[0066] <Spectral Component Existence Determination Unit 230> The spectral component existence determination unit 230 determines whether there are spectral components whose similarity to each of the selected spectral components H and each of the reference spectral components M exceeds the similarity threshold (step S717 in Figure 7). If there are no spectral components whose similarity exceeds the similarity threshold, the spectral component existence determination unit 230 increases the number of selections X (step S719 in Figure 7), returns to step S713, and generates spectral components for the number of selections X by non-negative matrix factorization processing. This allows the process to be repeated.

[0067] <Electromyographic Component Number Determination Unit 240> The electromyographic component number determination unit 240 determines the spectral components that exceed the similarity threshold as electromyographic components, based on the number of spectral components that exceed the similarity threshold and the similarity of the spectral components that exceed the similarity threshold.

[0068] Specifically, the electromyographic component number determination unit 240 performs the following process. The electromyographic component number determination unit 240 determines whether there is only one spectral component whose similarity exceeds the similarity threshold (Step S721 in Figure 7). When the electromyographic component number determination unit 240 determines that there is only one spectral component whose similarity exceeds the similarity threshold, it increases the number of selections X (Step S719 in Figure 7) and returns to Step S713.

[0069] The electromyographic component number determination unit 240 determines that there is more than one spectral component whose similarity exceeds the similarity threshold, and then determines the optimal number of spectral components X as electromyographic components (step S721 in Figure 7).

[0070] <<Specific Example>> The following is a specific example of the processing performed by the spectral component selection unit 210, the similarity calculation unit 220, the spectral component presence determination unit 230, and the electromyography component number determination unit 240.

[0071] Assume the reference spectral distribution consists of four reference spectral components M (M1, M2, M3, M4). Also, assume the threshold for similarity determination is 0.5. If the number of selections X for similarity determination is 2 (X=2), then two spectral components H1 and H2 are generated.

[0072] For the first spectral component H1, the correlation coefficient between H1 and M1 is 0.4, the correlation coefficient between H1 and M2 is 0.6, the correlation coefficient between H1 and M3 is 0.4, and the correlation coefficient between H1 and M4 is 0.2. For the second spectral component H2, the correlation coefficient between H2 and M1 is 0.1, the correlation coefficient between H2 and M2 is 0.2, the correlation coefficient between H2 and M3 is 0.1, and the correlation coefficient between H2 and M4 is 0.05.

[0073] In this case, the highest correlation coefficient is 0.6, which is the correlation coefficient between H1 and M2, and is greater than the threshold for similarity determination. Also, the correlation coefficients for H2 are all smaller than the threshold for similarity determination. For this reason, H1 is determined to be the electromyographic component, and H2 is determined to be the noise component.

[0074] <If only one electromyographic component is found> If only one electromyographic component is found, increase the number of selections X for similarity determination by 1, perform the non-negative matrix factorization process again, and generate (update) the spectral component H of selections X from spectrum 1 to spectrum E for each epoch.

[0075] Specifically, the number of selections X for similarity determination is set to 3 (X=3), generating three spectral components H1, H2, and H3. The correlation coefficients between these three spectral components H1, H2, and H3 and the reference spectral components M1, M2, M3, and M4 are then calculated. Based on the correlation coefficients, spectral component H is distinguished from electromyographic components or noise components.

[0076] <When two or more electromyographic components are found> If two or more electromyographic components are found, the optimal number of selections X is determined. Increasing the number of selections X will result in the discovery of two or more electromyographic components. If the number of selections X becomes too large, the electromyographic components will be further and excessively broken down into smaller components. In other words, they will be broken down into different types of muscle fibers, so it is necessary to determine the optimal number of selections X. Normally, surface electromyographic signals reflect the mixed activity of slow-twitch muscle fibers (Type 1) (low frequency band 10 Hz to 100 Hz) and fast-twitch muscle fibers (Type 2) (high frequency band 100 Hz to 500 Hz), but if the number of selections X becomes large, they will be broken down into different types of muscle fibers.

[0077] Therefore, it is necessary to determine the optimal number of selections X. We monitor the change in the correlation coefficient of the electromyographic components when the number of selections X is increased. For example, if the correlation coefficient for spectral component H3 is 0.9 when the number of selections X = 3, it is highly likely that the electromyographic components have been separated to the optimal level. Therefore, we determine that the number of selections X = 3.

[0078] <<<Noise component reduction filter generation unit 300>>> The electromyography component identification unit 200 can remove noise component spectral components H from the spectral components H generated by the non-negative matrix factorization process and extract spectral components H consisting only of electromyography components.

[0079] From the spectral components H generated by the non-negative matrix factorization process, a filter can be generated that reduces the contribution rate of the noise component from the noise component H (Figure 9).

[0080] <Types of Noise> Surface electromyographic signals contain different types of noise, each affecting different frequency ranges. For example, power supply noise has a main frequency range of 50 / 60 Hz and is periodic noise originating from power lines. White Gaussian noise has a main frequency range of the entire frequency band and has constant energy across the entire band. Mechanical vibration noise has a main frequency range of 10-100 Hz and is caused by environmental vibrations. High-frequency noise has a main frequency range of 250 Hz and above and is caused by interference from electronic equipment. Therefore, adaptive filtering for each frequency is necessary, rather than a single fixed filter.

[0081] <Noise Contribution Rate per Frequency> The noise contribution rate (f) at a specific frequency f in each epoch is calculated. The noise contribution rate (f) indicates how much energy at each frequency f is contributed by the noise component.

[0082] <Example for 35Hz> For example, for Epoch e at 35Hz, let's assume that the spectral components H1, H2, H3, H4, and H5 each have the following power values. Also, let's assume that spectral component H3 is determined to be the electromyographic component, and the remaining spectral components H1, H2, H4, and H5 are determined to be noise, with the following power values: Power of H1 (noise component) = 45 Power of H2 (noise component) = 25 Power of H3 (electromyographic component) = 103 Power of H4 (noise component) = 12 Power of H5 (noise component) = 1 (Figure 8).

[0083] Therefore, the sum of the powers of the noise components (H1, H2, H4, H5) is 45 + 25 + 12 + 1 = 83, and the sum of the powers of the spectral components H1, H2, H3, H4, H5 is 183.

[0084] In this case, the noise contribution rate at 35 Hz for epoch e is: Noise contribution rate (35 Hz) = 80 / 183 = 0.4371. This indicates that 43.71% of the total power at 35 Hz for epoch e originates from noise, and the remaining 56.29% is from surface electromyography signals.

[0085] <Calculation of filter coefficients for each frequency> The filter value (f) for each frequency f is calculated as follows: Filter value (f) = 1 - Noise contribution rate (f).

[0086] <Examples> When the frequency is 35 Hz, if the noise contribution rate (35 Hz) = 0.4371, the filter value (f) = 1 - 0.4371 = 0.5629 When the frequency is 50 Hz, if the noise contribution rate (50 Hz) = 0.9, the filter value (f) = 1 - 0.9 = 0.1 When the frequency is 100 Hz, if the noise contribution rate (100 Hz) = 0.2, the filter value (f) = 1 - 0.2 = 0.8 When the frequency is 250 Hz, if the noise contribution rate (250 Hz) = 0.05, the filter value (f) = 1 - 0.05 = 0.95.

[0087] Applying this filter reduces the following: 43.71% reduction at 35 Hz, 90% reduction at 50 Hz (power supply noise), 20% reduction at 100 Hz (high surface electromyography signal component), and only 5% reduction at 250 Hz (almost clean surface electromyography signal).

[0088] In this way, the filter value (f) can be determined for all frequencies, and the filter can be generated.

[0089] <<<Noise component removal signal processing unit 400>>> The noise component removal signal processing unit 400 removes the noise component from each of the epochs using the generated filter.

[0090] The noise component removal signal processing unit 400 multiplies the spectrum e of epoch e by the filter value (f) to determine the spectrum e' after noise removal (Figure 9).

[0091] <<<Clean Surface Electromyography Signal Reconstruction Unit 500>>> The clean surface electromyography signal reconstruction unit 500 recombines each epoch after the processing of the noise component removal step and reconstructs the surface electromyography signal from which the noise component has been removed.

[0092] First, the clean surface electromyography signal reconstruction unit 500 performs an inverse Fourier transform on the denoised spectrum e' to reconstruct the time-domain epoch e (Figure 9).

[0093] Furthermore, the clean surface electromyography signal reconstruction unit 500 reconstructs the cleaned epoch e' in the same order as the original epoch e. ​​In doing so, it synthesizes each epoch while considering the overlap length o to generate a final surface electromyography signal from which noise components have been removed (Figure 9).

[0094] <<<Highlight Display>>> This process displays the reconstructed surface electromyography (EMG) signal, showing the period during which the EMG component exceeds a predetermined display threshold in a different manner from other periods. Additionally, the period during which the proportion of the reconstructed surface EMG signal contributed by the EMG component exceeds a display threshold is displayed in a different manner from other periods.

[0095] Since surface electromyography (EMG) signals contain noise, highlighting areas where the EMG signal contributes significantly to the noise makes it possible to visualize areas where the EMG signal is dominant.

[0096] Highlighting allows for the identification of significant time intervals (epochs) in surface electromyography (EMG). By highlighting epochs in which the EMG signal exceeds a display threshold, patterns of muscle activity and important events can be easily visualized. For example, as shown in Figure 10, epochs in which the display threshold has been exceeded can be easily visualized by distinguishing between areas where the EMG signal is displayed (highlighted) and areas where the EMG signal is not displayed.

[0097] The contribution rate of the surface electromyography (EMG) signal in epoch e, after removing noise components, is determined. The contribution rate of the surface EMG signal represents the proportion of the EMG component to the total power in epoch e after removing noise components.

[0098] <Display Threshold> The display threshold determines the percentage of electromyography (EMG) components that must be exceeded to be considered an important epoch. For example, setting the display threshold to 0.5 will highlight EMG components with a contribution rate of 50% or more. Setting the display threshold to 0.7 will highlight EMG components with a contribution rate of 70% or more, indicating a stricter condition for highlighting.

[0099] Multiple different display thresholds may be set, and the display method may differ for each display threshold. For example, each display threshold may be displayed in a different color. The difference in the contribution rate of the surface electromyographic component to the noise should be displayed in a way that is easily visible.

[0100] <<<<<Scope of Embodiments>>>>> As described above, this embodiment has been presented. However, the descriptions and drawings that constitute part of this disclosure should not be understood as limiting. Various embodiments not described herein are included.

[0101] This technology can appropriately separate surface electromyography (EMG) signals into EMG components and noise components, and can also be applied to reconstruct surface EMG signals from which the noise component has been removed. Cross-reference of related applications

[0102] This application claims priority over Japanese Patent Application No. 2025-052186, filed with the Japan Patent Office on 26 March 2025, all of which disclosures are incorporated herein by reference in their entirety.

Claims

1. A surface electromyography signal processing method for extracting electromyography components from a surface potential signal measured on a subject's skin over a predetermined period of time and removing noise components, comprising: a decomposition step of dividing the measured surface potential signal into a plurality of epochs at predetermined unit time intervals, and decomposing each of the epochs into a plurality of spectral components and a weight component indicating the respective weight of each of the spectral components for each of the epochs; an electromyography component identification step of comparing the spectral components with a reference spectral distribution predetermined by modeling electromyography in advance, and identifying a spectral component similar to one reference spectral component included in the reference spectral distribution as an electromyography component; a filter generation step of generating a filter that reduces the contribution rate of the noise component to each of the epochs, treating the spectral components other than the electromyography component among the spectral components decomposed in the decomposition step as noise components; a noise component removal step of removing the noise component from each of the epochs using the generated filter; and a reconstruction step of recombining each of the epochs after processing in the noise component removal step to reconstruct the surface electromyography signal from which the noise component has been removed. A method for processing surface electromyographic signals, including the method described above.

2. The surface electromyographic signal processing method according to claim 1, wherein the electromyographic component identification step is to identify the spectral component whose similarity to the reference spectral component of the reference spectral distribution exceeds a similarity threshold as the electromyographic component.

3. The surface electromyographic signal processing method according to claim 1, wherein the decomposition into spectral components and weight components performed by the decomposition step is carried out using a non-negative matrix factorization algorithm.

4. The surface electromyographic signal processing method according to claim 2, wherein the similarity threshold is determined according to the characteristics of the measured surface electromyographic signal in order to optimize the identification accuracy of the electromyographic component.

5. The surface electromyographic signal processing method according to claim 4, wherein the similarity threshold is determined according to the electromyographic potential for determining the reference spectral distribution, the model of the electromyographic potential for determining the reference spectral distribution, or the measurement conditions of the surface electromyographic signal, in order to optimize the identification accuracy of the electromyographic components.

6. The surface electromyographic signal processing method according to claim 1, wherein the reference spectral distribution is pre-generated by reproducing the modeled electromyographic potential through simulation.

7. The surface electromyographic signal processing method according to claim 2, wherein the electromyographic component identification step includes: a selection step of selecting a selection number of spectral components for similarity determination from the plurality of spectral components for each of the plurality of epochs; a similarity calculation step of calculating the similarity between each of the spectral components decomposed according to the selection number and a reference spectral component included in the reference spectral distribution; a spectral component existence determination step of determining whether or not there are spectral components whose similarity among each of the decomposed spectral components exceeds the similarity threshold; an iterative process of repeatedly executing the selection step, the similarity calculation step, and the spectral component existence determination step while changing the selection number; and an electromyographic component number determination step of determining the spectral components that exceed the similarity threshold as the electromyographic components and terminating the iterative process, according to the number of spectral components that exceed the similarity threshold obtained by executing the iterative process and the similarity of the spectral components that exceed the similarity threshold.

8. The surface electromyography signal processing method according to claim 1, wherein each of the epochs is generated by dividing the measured surface electromyography signal so that some time overlaps.

9. The surface electromyography signal processing method according to claim 1, further comprising a frequency data generation step of performing a fast Fourier transform on the measured surface electromyography signal before processing the decomposition step to generate frequency domain data.

10. The surface electromyography signal processing method according to claim 1, wherein the reconstruction step converts the surface electromyography signal for frequency in each epoch into a surface electromyography signal for time in each epoch in the table using an inverse fast Fourier transform.

11. A surface electromyographic signal processing method according to claim 10, further comprising a display step of displaying a reconstructed surface electromyographic signal, wherein the period during which the electromyographic component exceeds a predetermined display threshold is displayed in a manner different from other periods.

12. The surface electromyographic signal processing method according to claim 11, wherein the display step displays the period in which the proportion of the reconstructed surface electromyographic signal to which the electromyographic component contributes exceeds the display threshold in a manner different from other periods.

13. A surface electromyography signal processing device for extracting electromyography components from a surface electromyography signal measured on the skin of a subject over a predetermined period of time and removing noise components, comprising: a decomposition unit that divides the measured surface electromyography signal into a plurality of epochs at predetermined unit time intervals, and decomposes each of the epochs into a plurality of spectral components and a weighting component that indicates the respective weight of each of the epochs in each of the spectral components; an identification unit that compares the spectral components with a reference spectral distribution predetermined by modeling electromyography in advance, and identifies a spectral component similar to one reference spectral component included in the reference spectral distribution as an electromyography component; a filter generation unit that generates a filter that reduces the contribution rate of the noise component to each of the epochs, treating the spectral components other than the electromyography component among the spectral components decomposed by the decomposition unit as noise components; a signal processing unit that removes the noise component to each of the epochs using the generated filter; and a reconstruction unit that recombines each of the epochs after processing by the signal processing unit and reconstructs the surface electromyography signal from which the noise component has been removed. A surface electromyography signal processing device comprising:

14. The surface electromyographic signal processing device according to claim 13, wherein the identification unit identifies the spectral component whose similarity to the reference spectral component of the reference spectral distribution exceeds a similarity threshold as the electromyographic component.

15. The surface electromyographic signal processing device according to claim 13, wherein the decomposition into spectral components and weight components performed by the decomposition unit is carried out using a non-negative matrix factorization algorithm.

Images