A training method, application method and related system for a blind separation model of electromyographic signals

Abstract

This invention discloses a training method, application method, and related system for a blind electromyography (EMG) signal separation model, relating to the field of signal processing. The training method includes: acquiring sample data; whitening and expanding the sample data; decomposing the whitened and expanded data using the gCKC algorithm to obtain IPTs; standardizing the sample data and IPTs; inputting the standardized data into the EMG signal blind separation model to output a motor unit action potential sequence; determining the loss value based on the motor unit action potential sequence, IPTs, and a loss function; and optimizing the parameters of the EMG signal blind separation model based on the loss value to obtain a trained EMG signal blind separation model. This invention proposes a sequence-to-sequence deep learning method that does not require data preprocessing and can achieve high accuracy. Furthermore, temporal convolutional networks offer fast training speeds, fewer parameters, and convolutional models are better suited for EMG signal problems.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a training method, application method and related system for a blind separation model of electromyographic signals. Background Technology

[0002] Electromyography (EMG) signals are physiological electrical signals that contain information about motor intention. The motor information carried by these signals can be extracted through EMG signal decomposition. Blind separation of surface EMG signals is a signal processing technique aimed at isolating the firing time sequence of motor units (MUs) from surface EMG signals. Blind EMG signal separation can be seen as the reverse process of EMG formation. Its main purpose is to restore the EMG signal to its basic components—the action potential sequence of motor units. This allows for the identification and classification of individual motor units within the EMG signal, thereby obtaining the number, type, firing information, and waveform information of the motor units involved in the muscle activity. These MU firing time pulse sequences can be used in neurophysiological research, the diagnosis of central nervous system dysfunction, and the establishment of accurate and intuitive human-computer interfaces (HMIs).

[0003] Electromyography (EMG) signal decomposition allows researchers to obtain recruitment, firing, and waveform information of individual motor units, thus fundamentally understanding the activity state of central motor neurons. By studying the firing relationships between various motor units, the collaborative relationships between muscles can be directly observed. When muscle diseases occur, studying changes in the firing patterns of motor units can aid in traditional experience-based clinical diagnosis. In rehabilitation applications, EMG signal decomposition results help therapists clearly observe the recovery of patients' motor function, providing a reliable reference for rehabilitation treatment. In the field of motor control, the results of EMG signal decomposition can even be directly used as the most basic muscle movement commands to achieve control functions. Furthermore, EMG signal decomposition can analyze the behavioral relationships between various motor units during muscle contraction, which will more effectively study the working characteristics within and between muscles. When muscle contraction force is relatively large, calculating the overall information of the decomposed signal to analyze the degree of muscle contraction during actual movement and further assess the normality of neuromuscular system function has significant practical implications.

[0004] In recent years, the rapid development of computer technology has propelled the application of electromyography (EMG) signal decomposition in engineering fields such as gait analysis, prosthetic control, and human-computer interaction. An increasing number of applications require the rapid, even online, decomposition of EMG signals. Consequently, more and more decomposition methods need to be developed to meet these requirements. Research on EMG signal decomposition algorithms can significantly advance clinical rehabilitation medicine and healthcare electronics, and is of great research importance.

[0005] Currently, decomposition methods for high-density surface electromyography (HD-sEMG) mainly rely on blind source separation (BSS) techniques. Among these techniques, the Kernel Compensation (CKC) algorithm proposed by Holobar and Zazula, along with its gradient optimization counterpart gCKC, is the most commonly used method.

[0006] CKC is a blind source separation method based on the non-Gaussian and non-linear properties of electromyography (EMG) signals. The CKC algorithm first models the EMG signal acquisition system with multiple acquisition channels as a linear time-invariant convolutional MIMO system. In this model, the EMG signal is represented as a convolution of the firing time series of motor units and action potentials. Then, the firing time series relationship of motor units is inferred using a linear minimum mean square error model. Finally, an activity index is introduced to solve for the cross-correlation vector and calculate the firing time series. This method can extract the recruitment and firing information of motor units directly from the EMG signal itself without involving complex convolution matrix calculations. Validated using artificially synthesized and real-world EMG signals, the CKC algorithm can decompose more than thirty motor units when the signal-to-noise ratio is high. The CKC algorithm achieves blind source separation by maximizing the core entropy of the output signal. This method makes no assumptions about the distribution of the input signal, thus exhibiting excellent performance when processing mixed non-Gaussian and non-linear source signals.

[0007] Gradient Convolution Kernel Compensation (gCKC) is a generalization of the CKC algorithm. This method improves upon CKC by optimizing the calculation of the cross-correlation vector in the signal model. It introduces a first-order differentiable gradient function related to the firing time series to iteratively update the cross-correlation vector, and uses the convergence value of this cross-correlation vector to calculate the firing time series of motor units. The initialization of the cross-correlation vector follows the same rules as CKC, while the update rule uses a gradient function to update and calculate the convergence value. The effectiveness of this method in decomposing surface electromyography (EMG) signals depends on the selection of the gradient function, which must satisfy certain convergence and stability requirements. Similarly, gCKC was validated using both simulated and real-world EMG signals. Experiments showed that compared to CKC, gCKC can decompose more motor units even with low signal-to-noise ratios in EMG signals, while maintaining almost unchanged decomposition accuracy, indicating higher robustness. The gCKC algorithm can handle not only non-Gaussian and nonlinear mixed source signals, but also source signals with various distributions, thus improving the applicability and flexibility of the CKC algorithm.

[0008] These methods all model the surface sEMG signal acquisition system as a multi-channel linear time-invariant convolutional multiple-input multiple-output system, representing the surface sEMG signal as a convolution of the motor unit discharge time series and action potentials. Subsequently, a linear minimum mean square error model is used to estimate the neural innervation pulse sequence (IPTs) of the motor unit activity represented by the sparse time series, with the same sampling rate as the HD-sEMG signal.

[0009] All the aforementioned studies on surface electromyography (SEMG) signal decomposition methods are primarily applied to offline analysis because these methods are slow and cannot meet the requirements for real-time decomposition. Furthermore, these methods rely on complex mathematical models and iterative solution processes. Before decomposition, the HD-sEMG signal needs to be expanded to convert the convolutional mixture into a linear transient mixture, and spatial decorrelation is performed through whitening transform to improve the convergence properties of the optimization program. These preprocessing steps introduce computational complexity, latency, and sensitivity to noise.

[0010] With the development of deep learning, attempts have gradually begun to apply it to the decomposition of surface electromyography (sEMG) signals. Yue Wen et al. treated the sEMG decomposition problem as a binary classification task and trained a deep convolutional neural network (CNN) using the recognition results of the CKC algorithm. Specifically, they used the MUAP (Muscle Activity Scale) and firing time series of the muscle signal as input, classified them using the CNN, and output the probability of whether each motor unit fired. Then, they reconstructed the muscle signal based on these probabilities and obtained a more accurate decomposition result through iterative optimization. This method can perform fast decomposition with a small window, achieving low-latency processing.

[0011] However, classification models using CNNs to decompose data in windows cannot accurately identify dense muscle motor unit (MU) firings. Furthermore, training becomes impossible when MU firings are very dense, with firings occurring within every window. This is because the method identifies data in windows; when MU firing density is high, multiple firings may occur within a single window, leading to inaccurate identification. Simultaneously, sequence-to-sequence models use GRU networks with a serial structure, resulting in long training times, high memory consumption during training, and a large number of model parameters.

[0012] Therefore, how to design a blind electromyography signal separation model that does not require data preprocessing, has a fast training speed, fewer parameters, and can achieve high accuracy has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0013] The purpose of this invention is to provide a training method, application method and related system for a blind separation model of electromyographic signals. This invention is based on a sequence-to-sequence training framework of temporal convolutional networks, which can efficiently and accurately decompose high-density surface electromyographic signals into motor unit action potential sequences.

[0014] To achieve the above objectives, the present invention provides the following solution:

[0015] In a first aspect, the present invention provides a training method for a blind electromyography signal separation model, comprising:

[0016] Acquire sample data; the sample data is high-density surface electromyography (EMG) signals.

[0017] The sample data is whitened and expanded to obtain the whitened and expanded data.

[0018] The whitened and expanded data were decomposed using the gCKC algorithm to obtain the neural innervation pulse sequence.

[0019] The sample data and the neural innervation pulse sequence are standardized to obtain standardized data.

[0020] The standardized data is input into the electromyographic signal blind separation model, which outputs a motor unit action potential sequence. The motor unit action potential sequence is determined by the noise power ratio of gCKC running at a given signal-to-noise ratio. The electromyographic signal blind separation model is a model based on a temporal convolutional network.

[0021] The loss value is determined based on the motor unit action potential sequence, the nerve innervation pulse sequence, and the loss function.

[0022] The parameters of the blind electromyography signal separation model are optimized based on the loss value to obtain a trained blind electromyography signal separation model.

[0023] Optionally, before inputting the standardized data into the blind electromyography signal separation model, the training method further includes:

[0024] The sample data is segmented into window samples using a sliding window method.

[0025] Optionally, the sample data and the neural innervation pulse sequence are standardized to obtain standardized data, specifically including:

[0026] The sample data and the neural innervation pulse sequence were standardized using the mean and standard deviation of the sample data based on the z-score method.

[0027] Optionally, the loss function is the mean squared error loss function.

[0028] Secondly, the present invention provides an application method for a blind separation model of electromyographic signals, including:

[0029] Acquire real experimental high-density surface electromyography signals.

[0030] The experimentally acquired high-density surface electromyography (EMG) signals are input into the EMG signal blind separation model to obtain the predicted motor unit action potential sequence; the EMG signal blind separation model is a model trained according to the training method of the EMG signal blind separation model described above.

[0031] Thirdly, the present invention provides a training system for a blind separation model of electromyographic signals, comprising:

[0032] The sample acquisition unit is used to acquire sample data; the sample data is high-density surface electromyography signals.

[0033] The whitening and expansion unit is used to whiten and expand the sample data to obtain whitened and expanded data.

[0034] The decomposition unit is used to decompose the whitened and expanded data using the gCKC algorithm to obtain the neural innervation pulse sequence.

[0035] The standardization unit is used to standardize the sample data and the neural innervation pulse sequence to obtain standardized data.

[0036] The input unit is used to input the standardized data into the electromyographic signal blind separation model and output the motor unit action potential sequence; the motor unit action potential sequence is determined by the noise power ratio of gCKC running at a given signal-to-noise ratio; the electromyographic signal blind separation model is a model based on a temporal convolutional network.

[0037] The loss value determination unit is used to determine the loss value based on the motor unit action potential sequence, the nerve innervation pulse sequence, and the loss function.

[0038] The training unit is used to optimize the parameters of the blind electromyography signal separation model based on the loss value, so as to obtain a trained blind electromyography signal separation model.

[0039] Optionally, the training system further includes:

[0040] Segmentation unit, used to segment the sample data into window samples using a sliding window method.

[0041] Optionally, the standardization unit includes:

[0042] The standardization module is used to standardize the sample data and the neural innervation pulse sequence based on the z-score method using the mean and standard deviation of the sample data.

[0043] Optionally, the loss function is the mean squared error loss function.

[0044] Fourthly, the present invention provides an application system for a blind separation model of electromyographic signals, comprising:

[0045] The signal acquisition unit is used to acquire real experimental high-density surface electromyography signals.

[0046] The prediction unit is used to input the real-collected experimental high-density surface electromyography (EMG) signals into the EMG signal blind separation model to obtain the predicted motor unit action potential sequence; the EMG signal blind separation model is a model trained according to the training method of the EMG signal blind separation model described above.

[0047] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0048] This invention provides a training method, application method, and related system for a blind separation model of electromyographic signals. It proposes a sequence-to-sequence deep learning method that eliminates the need for data preprocessing and achieves high accuracy. Furthermore, temporal convolutional networks offer fast training speeds, require fewer parameters, and are better suited for electromyographic signal problems. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart illustrating a training method for a blind electromyography signal separation model provided in Embodiment 1 of the present invention;

[0051] Figure 2 This is a schematic diagram of the training process of a blind electromyography signal separation model provided in Embodiment 1 of the present invention;

[0052] Figure 3 This is a schematic diagram of the sample data acquisition process provided in Embodiment 1 of the present invention;

[0053] Figure 4 This is a flowchart of an application method for a blind electromyography signal separation model provided in Embodiment 2 of the present invention;

[0054] Figure 5 This is a schematic diagram of the structure of a training system for a blind electromyography signal separation model provided in Embodiment 3 of the present invention;

[0055] Figure 6 This is a schematic diagram of the structure of an application system for a blind separation model of electromyographic signals provided in Embodiment 4 of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0057] The purpose of this invention is to provide a blind separation model training method, application method, and related system for electromyography (EMG) signals, used to efficiently and accurately decompose HD-sEMG signals into IPTs. The IPTs of each unit (MU) obtained after decomposition by the CKC algorithm are considered as continuous time series. A sequence-sequence training architecture from surface EMG signals to IPT series is constructed. A supervised deep learning algorithm learns how to obtain the IPT time series of each MU from unwhitened HD-sEMG signals, thus decomposing unseen HD-sEMG signals into sampling point precision release time series. A novel network for solving time series problems, namely the Temporal Convolutional Network (TCN), is used for sequence-sequence learning. The TCN network adopts a parallel structure and parameter sharing mechanism, outperforming recurrent neural networks (RNNs) in terms of training time and model parameters. It has been widely used in various sequence modeling tasks such as speech recognition, natural language processing, time series prediction, and video classification, achieving satisfactory results.

[0058] This invention uses a TCN network to decompose HD-sEMG signals into IPT (Intra-Pulse Transmission) framework. The TCN network is trained using the time series output by the gCKC algorithm, and a threshold-based method is used to convert the IPT sequence into a firing time series. This decomposition method has been validated on simulated data and real experimental high-density surface electromyography (HD-sEMG) signals.

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] Example 1:

[0061] like Figure 1As shown, this embodiment provides a training method for a blind electromyography signal separation model, including:

[0062] S1: Acquire sample data; the sample data is high-density surface electromyography (EMG) signals.

[0063] S2: Whiten and expand the sample data to obtain whitened and expanded data.

[0064] S3: The whitened and expanded data are decomposed using the gCKC algorithm to obtain the neural innervation pulse sequence.

[0065] S4: Standardize the sample data and the neural innervation pulse sequence to obtain standardized data.

[0066] S5: Input the standardized data into the electromyographic signal blind separation model and output the motor unit action potential sequence; the motor unit action potential sequence is determined by the noise power ratio of gCKC running at a given signal-to-noise ratio; the electromyographic signal blind separation model is a model based on a temporal convolutional network.

[0067] S6: Determine the loss value based on the motor unit action potential sequence, the nerve innervation pulse sequence, and the loss function.

[0068] S7: Optimize the parameters of the blind electromyography signal separation model based on the loss value to obtain a trained blind electromyography signal separation model.

[0069] The specific training process is as follows: Figure 2 As shown: First, electromyography (EMG) signal data is collected, and IPT sequences are generated using the unsupervised gCKC algorithm. These sequences are then paired with unwhitened HD-sEMG signals for supervised training. The HD-sEMG signals are z-score normalized. Next, the mean squared error between the output sequence and the IPT sequence is calculated and backpropagated through a TCN network.

[0070] As an optional implementation method provided in this embodiment, the sample data collection process is as follows: Figure 3 As shown, (a) illustrates the placement of the high-density electromyography (HD-EMG) electrodes. A two-dimensional viscous grid consisting of 64 electrodes was placed above the medial head (GM) and lateral head (GL) of the gastrocnemius muscle. A two-dimensional viscous grid consisting of 32 electrodes was placed on the medial portion of the soleus iliac muscle (SOL), below the gastrocnemius tendon junction. (b) shows an isometric plantar flexion task performed on an ergonomic measuring instrument (knee fully extended, ankle flexion angle 10 degrees). (c) shows the HD-EMG signals recorded using the 64-electrode grid.

[0071] The data collection process was as follows: Participants lay prone on a custom-designed force gauge consisting of a bench (TRE-50k, Dacell) equipped with foot pedals and torque sensors. The participants' knees were fully extended, and their ankles were securely fixed at a 10° plantar flexion angle (0° being the foot perpendicular to the lower leg). After a series of submaximal contraction warm-up exercises, participants performed three maximal isometric voluntary contractions (MVC), each lasting 3–5 seconds, with a 120-second rest between each. In this embodiment, the maximum torque value during each contraction was obtained from a 250ms moving average window and used as a reference value for maximum torque. Participants then performed three contractions at 10%, 30%, and 50% MVC intensities, in a randomized order. These contractions included a 5-second rise phase, a 15-second (50% MVC) or 20-second (10% and 30% MVC) plateau phase, and a 5-second fall phase. In summary, this embodiment analyzed signals lasting 75 seconds (50% MVC) and signals lasting 90 seconds (10% and 30% MVC). To reduce the impact of muscle fatigue on the HD-EMG signal, the contraction process was divided into two parts, lasting either 60 seconds (10% MVC) or 120 seconds (30% and 50% MVC). The torque signal was digitized using a 2048Hz acquisition system and low-pass filtered (Butterworth 3rd order, cutoff frequency 20Hz). The same acquisition system was also used for HD-EMG signal acquisition (EMG-Quattrocento; 400-channel EMG amplifier, OT Bioetronica, Italy).

[0072] HD-EMG signal decomposition of the experimentally recorded signals was performed using the CKC algorithm implemented in the DEMUSE tool software developed in MATLAB. Prior to decomposition, the EMG signals underwent bandpass filtering (Butterworth 2nd order, 20-750Hz). After automatic MU identification, all MU spikes were visually inspected and manually edited to avoid false positives (FP, labeled artifacts) or false negatives (FN, unlabeled discharges). In this embodiment, only MUs with an impulse noise ratio (PNR) > 30dB were retained for further analysis.

[0073] To generate 10 seconds of simulated data for each signal-to-noise ratio (SNR), this embodiment divides each simulated dataset into an 8-second training set, a 1-second validation set, and a 10-second test set. The gCKC algorithm is used to decompose the training data into IPTs, where the cutoff pulse signal-to-noise ratio (PNR) is set to 30 dB as the threshold for source acceptance.

[0074] In step S3, the innervation pulse sequence (IPT) of the j-th motor unit can be directly estimated from the extended electromyographic signal:

[0075]

[0076] Where X(n) = [x1(n), ..., x M (n)] T For measurement signals with M channels, x j (n) represents the signal of the j-th channel at the n-th sampling point. It is an extended form of X(n), s j (n) is the IPT sequence of the j-th MU. It is an IPT sequence of N MUs, s(n) = [s1(n), ..., s2(n)]. N The extended form of [(n)]. This represents the cross-correlation vector between the extended firing time pulse sequence and the surface electromyography signal, also known as the separation vector. Let E(·) represent the autocorrelation matrix of the extended surface electromyography signal itself, and E(·) represent the expected value. Due to the cross-correlation vector... It is unknown; the gCKC algorithm uses measurement vectors. To approximate it, and continuously iterate and update it.

[0077]

[0078]

[0079] Where E represents the learning rate and k represents the number of iterations. express The gradient of the gradient. After obtaining the estimated IPT through iteration, it can be converted into a timestamp through a thresholding process.

[0080] As an optional implementation method provided in this embodiment, step S4 specifically includes:

[0081] The sample data and the neural innervation pulse sequences are standardized using the z-score method based on the mean and standard deviation of the sample data. This process ensures that the data have similar scale and distribution, facilitating better model learning of data features. The sample data includes training, validation, and test sets. For each data point x and IPT data, the following formula is used for transformation:

[0082] z = (x - mean) / σ.

[0083] Where z is the standardized value, mean is the mean of the training set, and σ is the standard deviation of the training set.

[0084] Before inputting the standardized data into the blind electromyography signal separation model, the training method further includes: segmenting the sample data into window samples using a sliding window method. Each window has a width of 2000 samples and a stride of 20. This generates a series of window samples for model training. At each update, 32 samples are randomly selected from the generated window samples as a mini-batch for training. This batch training method improves training efficiency and convergence speed. By standardizing HD-sEMG and IPT data, and using the sliding window and mini-batch training methods, data can be better prepared and a more accurate model can be trained.

[0085] This embodiment uses a temporal convolutional network for training. The mean squared error between the output sequence and the IPT sequence is calculated, and the error signal is backpropagated through a TCN network. That is, the loss function in this embodiment is the mean squared error loss function. The network output undergoes further processing by the temporal convolutional layers to generate the predicted sequence.

[0086] The TCN network structure consists of two convolutional layers, each with 32 filters (nb_filters) and a kernel size of 3. Training uses the Adam optimization algorithm with a learning rate of 0.001, Betas parameters of 0.9 and 0.999, and epsilon parameter of 10e-8. To prevent overfitting, this embodiment incorporates a 20% dropout probability on all TCN units, improving the network's robustness to electrode faults and eliminating over-reliance on specific channels. Error signals are backpropagated using a mean squared error loss function.

[0087] Before training, this embodiment uses a sliding window method to segment the HD-sEMG signal, with a window width of 2000 samples and a stride of 20. Each update uses a minibatch of 32 samples. To prevent overfitting, we use an early stopping strategy. Specifically, training stops when the model loss function on the validation set no longer decreases for 20 consecutive epochs.

[0088] For each dataset, the TCN network is trained on a training set of unwhitened and unfiltered HD-sEMG signals. The number of accepted MU trains is determined by the noise power ratio (PNR) of running gCKC at a given signal-to-noise ratio (SNR). The TCN network is trained using the IPT output generated by gCKC. For class predictions on the test set, this embodiment uses an iterative method to select threshold parameters to convert them to timestamps, maximizing the RoA (Rate of Agreement), similar to the output of gCKC. This thresholding technique is also used for predictions on experimental data.

[0089] In summary, this invention proposes a sequence-to-sequence deep learning method that does not require data preprocessing and can achieve high accuracy. Furthermore, temporal convolutional networks offer fast training speeds, require fewer parameters, and are better suited for electromyographic signal problems.

[0090] Example 2:

[0091] like Figure 4 As shown, this embodiment provides an application method for a blind separation model of electromyographic signals, including:

[0092] A1: Obtain the actual high-density surface electromyography signals collected in the experiment.

[0093] A2: Input the real-collected experimental high-density surface electromyography (EMG) signals into the EMG signal blind separation model to obtain the predicted motor unit action potential sequence; the EMG signal blind separation model is a model trained according to the EMG signal blind separation model training method described in Example 1.

[0094] It should be noted that after acquiring the actual experimental high-density surface electromyography (EMG) signals, the actual acquired experimental high-density surface EMG signals need to be whitened and expanded to obtain whitened and expanded data. The gCKC algorithm is used to decompose the whitened and expanded data to obtain the nerve innervation pulse sequence. Then, the actual acquired experimental high-density surface EMG signals and nerve innervation pulse sequence are standardized to obtain standardized data. The standardized data is then input into the EMG signal blind separation model to obtain the predicted motor unit action potential sequence.

[0095] Example 3:

[0096] like Figure 5 As shown, this embodiment provides a training system for a blind separation model of electromyographic signals, including:

[0097] The sample acquisition unit M1 is used to acquire sample data; the sample data is high-density surface electromyography signals.

[0098] The whitening and expansion unit M2 is used to whiten and expand the sample data to obtain whitened and expanded data.

[0099] Decomposition unit M3 is used to decompose the whitened and expanded data using the gCKC algorithm to obtain the neural innervation pulse sequence.

[0100] The standardization unit M4 is used to standardize the sample data and the neural innervation pulse sequence to obtain standardized data.

[0101] Input unit M5 is used to input the standardized data into the electromyographic signal blind separation model and output the motor unit action potential sequence; the motor unit action potential sequence is determined by the noise power ratio of gCKC running at a given signal-to-noise ratio; the electromyographic signal blind separation model is a model based on a temporal convolutional network.

[0102] The loss value determination unit M6 is used to determine the loss value based on the motor unit action potential sequence, the nerve innervation pulse sequence, and the loss function.

[0103] Training unit M7 is used to optimize the parameters of the blind electromyography signal separation model based on the loss value, so as to obtain a trained blind electromyography signal separation model.

[0104] As an optional implementation provided in this embodiment, the training system further includes:

[0105] Segmentation unit, used to segment the sample data into window samples using a sliding window method.

[0106] As an optional implementation provided in this embodiment, the standardization unit includes:

[0107] The standardization module is used to standardize the sample data and the neural innervation pulse sequence based on the z-score method using the mean and standard deviation of the sample data.

[0108] As an optional implementation method provided in this embodiment, the loss function is the mean squared error loss function.

[0109] Example 4:

[0110] like Figure 6 As shown, this embodiment provides an application system for a blind separation model of electromyographic signals, including:

[0111] The signal acquisition unit N1 is used to acquire the experimental high-density surface electromyography signals that were actually collected.

[0112] The prediction unit N2 is used to input the real-collected experimental high-density surface electromyography (EMG) signals into the EMG signal blind separation model to obtain the predicted motor unit action potential sequence; the EMG signal blind separation model is a model trained according to the EMG signal blind separation model training method described in Example 1.

[0113] It should be noted that after acquiring the actual experimental high-density surface electromyography (EMG) signals, the actual acquired experimental high-density surface EMG signals need to be whitened and expanded to obtain whitened and expanded data. The gCKC algorithm is used to decompose the whitened and expanded data to obtain the nerve innervation pulse sequence. Then, the actual acquired experimental high-density surface EMG signals and nerve innervation pulse sequence are standardized to obtain standardized data. The standardized data is then input into the EMG signal blind separation model to obtain the predicted motor unit action potential sequence.

[0114] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0115] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A training method for a blind electromyography signal separation model, characterized in that, include: Acquire sample data; the sample data is high-density surface electromyography (EMG) signals. The sample data is whitened and expanded to obtain whitened and expanded data; The whitened and expanded data were decomposed using the gCKC algorithm to obtain the neural innervation pulse sequence; The sample data and the nerve innervation pulse sequence are standardized to obtain standardized data. The standardized data is input into the electromyography signal blind separation model, which outputs a motor unit action potential sequence; the motor unit action potential sequence is determined by the noise power ratio of gCKC running at a given signal-to-noise ratio; the electromyography signal blind separation model is a model based on a temporal convolutional network; The loss value is determined based on the motor unit action potential sequence, the nerve innervation pulse sequence, and the loss function; The parameters of the blind electromyography signal separation model are optimized based on the loss value to obtain a trained blind electromyography signal separation model.

2. The training method for a blind electromyography signal separation model according to claim 1, characterized in that, Before inputting the standardized data into the blind electromyography signal separation model, the training method further includes: The sample data is segmented into window samples using a sliding window method.

3. The training method for a blind electromyography signal separation model according to claim 1, characterized in that, The sample data and the nerve innervation pulse sequence are standardized to obtain standardized data, specifically including: The sample data and the neural innervation pulse sequence were standardized using the mean and standard deviation of the sample data based on the z-score method.

4. The training method for a blind electromyography signal separation model according to claim 1, characterized in that, The loss function is the mean squared error loss function.

5. An application method for a blind separation model of electromyographic signals, characterized in that, include: Acquire real-world high-density surface electromyography (EMG) signals from experiments; The experimentally acquired high-density surface electromyography (EMG) signals are input into the EMG signal blind separation model to obtain the predicted motor unit action potential sequence; the EMG signal blind separation model is a model trained by the training method of the EMG signal blind separation model according to any one of claims 1-4.

6. A training system for a blind separation model of electromyographic signals, characterized in that, include: The sample acquisition unit is used to acquire sample data; The sample data consisted of high-density surface electromyography (EMG) signals. The whitening and expansion unit is used to whiten and expand the sample data to obtain whitened and expanded data. The decomposition unit is used to decompose the whitened and expanded data using the gCKC algorithm to obtain the neural innervation pulse sequence; A standardization unit is used to standardize the sample data and the neural innervation pulse sequence to obtain standardized data; The input unit is used to input the standardized data into the electromyography signal blind separation model and output the motor unit action potential sequence; the motor unit action potential sequence is determined by the noise power ratio of gCKC running at a given signal-to-noise ratio; the electromyography signal blind separation model is a model based on a temporal convolutional network. The loss value determination unit is used to determine the loss value based on the motor unit action potential sequence, the nerve innervation pulse sequence, and the loss function; The training unit is used to optimize the parameters of the blind electromyography signal separation model based on the loss value, so as to obtain a trained blind electromyography signal separation model.

7. The training system for a blind electromyography signal separation model according to claim 6, characterized in that, The training system also includes: Segmentation unit, used to segment the sample data into window samples using a sliding window method.

8. The training system for a blind electromyography signal separation model according to claim 6, characterized in that, The standardization unit includes: The standardization module is used to standardize the sample data and the neural innervation pulse sequence based on the z-score method using the mean and standard deviation of the sample data.

9. The training system for a blind electromyography signal separation model according to claim 6, characterized in that, The loss function is the mean squared error loss function.

10. An application system for a blind separation model of electromyographic signals, characterized in that, include: The signal acquisition unit is used to acquire real experimental high-density surface electromyography signals; The prediction unit is used to input the real-collected experimental high-density surface electromyography (EMG) signals into the EMG signal blind separation model to obtain the predicted motor unit action potential sequence; the EMG signal blind separation model is a model trained by the training method of the EMG signal blind separation model according to any one of claims 1-4.

Images