Wearable myoelectric prosthesis control method, terminal, storage medium and system

Abstract

The invention provides a wearable myoelectric prosthesis control method, a terminal, a storage medium and a system, and the method comprises the steps: carrying out the wavelet packet decomposition of a surface electromyogram signal for the surface electromyogram signal of each preset muscle position of a target body, and obtaining a plurality of sub-bands and the decomposition coefficients thereof; variational mode decomposition is carried out on the surface electromyogram signals to obtain a plurality of intrinsic mode functions; extracting a first feature corresponding to each sub-band based on the decomposition coefficient; extracting a second feature of each intrinsic mode function; extracting time domain features of the surface electromyogram signals; inputting the first feature, the second feature and the time domain feature into a TabPFN model to obtain an action classification result of the target body; and controlling the myoelectric prosthesis based on the action classification result. According to the method, the wavelet packet decomposition method and the variation mode decomposition method are fused for feature extraction, and the TabPFN model is combined, so that a signal complex mode can be learned more accurately when the high-dimensional nonlinear electromyographic signals are processed, and the action recognition accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for human prostheses, and in particular to a control method, terminal, storage medium, and system for a wearable myoelectric prosthesis. Background Technology

[0002] In today's society, physical disabilities impose a heavy burden on countless individuals and families, and have become a key factor hindering the comprehensive and harmonious development of society. The loss of the hand, in particular, has a devastating impact on the daily lives of patients. Traditional purely mechanical prostheses cannot accurately reproduce finger gestures and other movements, and impose additional burdens on users. However, the muscles at the ends of a disabled person's severed limb, which were originally used to control the limb, have not completely lost their function; they can still generate electromyographic (EMG) signals. These EMG signals are proactive, meaning they are generated before movement occurs. Therefore, if we can identify these EMG signals and thus determine the intention behind the movement, we can achieve highly real-time prosthetic control.

[0003] Existing prosthetic control solutions typically use traditional signal processing algorithms, such as Fast Fourier Transform and Time Domain Decomposition. These methods have limitations in feature extraction of electromyographic signals, which are non-stationary random signals, resulting in poor signal feature extraction quality and difficulty in meeting the requirements of fine motor skills. Summary of the Invention

[0004] This invention provides a control method, terminal, storage medium, and system for wearable myoelectric prostheses to address the problem of poor fine motor control in existing prostheses.

[0005] In a first aspect, embodiments of the present invention provide a control method for a wearable myoelectric prosthesis, comprising:

[0006] Acquire surface electromyographic signals at multiple preset muscle locations on the target body;

[0007] For the surface electromyography (EMG) signal at each preset muscle location, wavelet packet decomposition is performed on the surface EMG signal to obtain multiple sub-bands and the decomposition coefficients of each sub-band; and variational mode decomposition is performed on the surface EMG signal to obtain multiple intrinsic mode functions.

[0008] Based on the decomposition coefficients of multiple sub-bands of the surface electromyography signal, the features corresponding to each sub-band are extracted and used as the first feature; the features of each intrinsic mode function of the surface electromyography signal are extracted and used as the second feature; the time domain features of the surface electromyography signal are extracted.

[0009] The first feature, second feature, and temporal feature corresponding to the surface electromyography signals of all preset muscle locations are input into the TabPFN model to obtain the action classification results of the target body.

[0010] Based on the action classification results, the wearable myoelectric prosthesis is controlled to perform corresponding actions.

[0011] In a second aspect, embodiments of the present invention provide a terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the control method for a wearable myoelectric prosthesis as described in any possible implementation of the first aspect above.

[0012] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the control method for a wearable myoelectric prosthesis as described in the first aspect or any possible implementation thereof.

[0013] Fourthly, embodiments of the present invention provide a wearable myoelectric prosthesis system, comprising: a wearable myoelectric prosthesis, a signal acquisition device, and a terminal as described in the second aspect above;

[0014] The signal acquisition device includes an FPGA development board, a differential surface electromyography (SEMG) sensor, and an analog-to-digital converter (ADC). The differential SEMG sensor is used to acquire surface electromyography (SEMG) analog signals from multiple preset muscle locations on the target body and send the SEMG analog signals to the ADC. The ADC is used to convert the SEMG analog signals to digital signals to obtain surface electromyography (SEMG) signals. The FPGA development board is used to filter the SEMG signals and send the filtered SEMG signals to the terminal via a serial port. The terminal communicates with the wearable electromyography prosthesis via a serial port.

[0015] This invention provides a control method, terminal, storage medium, and system for a wearable myoelectric prosthesis. The method first acquires surface electromyography (SEM) signals at multiple preset muscle locations on a target body. Then, for each preset muscle location's SEM signal, wavelet packet decomposition is performed to obtain multiple sub-bands and their decomposition coefficients. Variational mode decomposition is then performed on the SEM signal to obtain multiple intrinsic mode functions (EMFs). Based on the decomposition coefficients of the multiple sub-bands, features corresponding to each sub-band are extracted and used as first features. Features of each EMF of the SEM signal are extracted and used as second features. Temporal features of the SEM signal are extracted. Finally, the first, second, and temporal features corresponding to the SEM signals at all preset muscle locations are input into a TabPFN model to obtain the target body's action classification result. Based on the action classification result, the wearable myoelectric prosthesis is controlled to perform corresponding actions. This embodiment integrates wavelet packet decomposition and variational mode decomposition methods, overcoming the limitations of traditional signal processing algorithms. Wavelet packet decomposition performs multi-band fine division of the original electromyographic signal, achieving good localization in the time and frequency domains and comprehensively capturing signal details. Variational mode decomposition, with its high noise resistance and frequency resolution, efficiently separates signal components. The combination of the two provides rich information for feature extraction. Combined with the TabPFN model, the above feature extraction methods can more accurately learn complex signal patterns when processing high-dimensional nonlinear electromyographic signals, improving the accuracy of action recognition. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art 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.

[0017] Figure 1 This is a schematic diagram of the structure of the wearable myoelectric prosthesis system provided in an embodiment of the present invention;

[0018] Figure 2 This is a flowchart illustrating the implementation of the control method for a wearable myoelectric prosthesis provided in this embodiment of the invention.

[0019] Figure 3 This is a schematic diagram of the graphical user interface of the wearable myoelectric prosthesis system provided in an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of the control device for a wearable myoelectric prosthesis provided in an embodiment of the present invention;

[0021] Figure 5 This is a schematic diagram of the terminal provided in an embodiment of the present invention. Detailed Implementation

[0022] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments will be described below in conjunction with the accompanying drawings.

[0024] Figure 1 This is a schematic diagram of the structure of a wearable myoelectric prosthetic system provided in an embodiment of the present invention. Figure 1 As shown, the wearable myoelectric prosthesis system includes: a wearable myoelectric prosthesis 30, a signal acquisition device 10, and a terminal 5;

[0025] The signal acquisition device 10 includes an FPGA development board 11, a differential surface electromyography sensor 13, and an analog-to-digital converter 12;

[0026] The differential surface electromyography sensor is used to collect surface electromyography analog signals at multiple preset muscle locations of the target body, and send the surface electromyography analog signals to the analog-to-digital converter 12;

[0027] The analog-to-digital converter 12 is used to convert the surface electromyography analog signal into a digital signal to obtain a surface electromyography signal.

[0028] The FPGA development board 11 is used to filter the surface electromyography (EMG) signal and send the filtered EMG signal to the terminal 5 via a serial port; the terminal 5 communicates with the wearable EMG prosthesis 30 via a serial port.

[0029] In this embodiment, the signal acquisition device 10 includes a differential surface electromyography (sEMG) sensor 13, an analog-to-digital converter (ADC) 12, and an FPGA development board 11. The differential sEMG sensor 13 acquires non-invasive surface electromyography (sEMG) signals from the supinator, brachioradialis, pronator teres, and flexor carpi ulnaris muscles of the forearm. Using the differential sEMG sensor 13, the signal can be amplified up to 13,000 times. The FPGA development board 11 has parallel processing capabilities and can use two parallel dual-channel high-speed ADCs 12 to meet sampling rate requirements of over 1000Hz.

[0030] Specifically, the FPGA development board 11, with its abundant I / O resources and powerful parallel processing capabilities, provides hardware support for high-speed real-time signal acquisition; the analog-to-digital converter 12 (ADC) uses two dual-channel 10-bit parallel interface high-speed analog-to-digital converters, whose high resolution ensures effective capture of weak sEMG signal changes; four electronic differential sEMG sensors are selected to acquire surface electromyographic signals of the supinator, brachioradialis, pronator teres, and flexor carpi ulnaris muscles of the forearm.

[0031] The signal acquisition device 10 consists of external physical circuitry and internal logic circuitry. The external physical circuitry's analog-to-digital converter 12 is connected to the differential surface electromyography (sEMG) sensor 13, converting the sEMG analog signal into a digital signal. The signal acquisition device 10 also includes flash memory, buttons, LEDs, and a serial communication module. The flash memory stores a firmware program, the buttons allow external interaction, the LEDs display the operating status, and the serial communication module converts USB and TTL levels. The digitized sEMG signal is received by the ADC signal receiving module of the FPGA development board 11. The FPGA development board 11 includes a Butterworth bandpass filter, a single notch filter, and a serial communication module. The sEMG signal is first filtered by the Butterworth bandpass filter to remove invalid frequency components, then by the single notch filter to suppress power frequency interference, and finally sent to the terminal 5 by the serial communication module.

[0032] This embodiment implements a Butterworth digital bandpass filter and notch filter on an FPGA development board 11 using the hardware description language VHDL. It also includes a serial communication module configured with a baud rate of 500,000, LED control, a button module, and an analog-to-digital converter 12. Top-level solid modeling is performed using block diagrams in Quartus Prime 24.1std, timing constraints are applied to the system, and synthesis and routing are performed before finalization to the onboard Flash memory. In practical use, the signal acquisition device 10 can first output four channels of calibration data. After the system completes calibration, pressing the onboard button will officially start the signal acquisition device 10.

[0033] Terminal 5 is used to implement the control method of the wearable myoelectric prosthesis 30 provided in this embodiment. Terminal 5 is connected to the signal acquisition device 10 and the wearable myoelectric prosthesis 30 through serial communication, and is used to control the wearable myoelectric prosthesis 30 to perform corresponding fine movements.

[0034] The wearable myoelectric prosthesis 30 can be a mechanical forearm, which consists of an open-source robotic hand uHand 33, a serial bus servo motor 32, and a control board 31. The serial bus servo motor 32, as the driving component, is directly connected to the mechanical structure of the open-source robotic hand. The rotation angle of the serial bus servo motor 32 controls the joint movement of the uHand, such as finger flexion and extension, and wrist rotation. The serial bus servo motor 32 is also connected to the control board 31 of the mechanical forearm. The control board 31 receives the status feedback of the serial bus servo motor 32 through a circuit interface and sends driving commands. The control board 31 is connected to a terminal 5 through a serial port to realize command transmission with the terminal 5.

[0035] See Figure 2 The diagram illustrates the implementation flowchart of the control method for the wearable myoelectric prosthesis provided in this embodiment of the invention, which is described in detail below:

[0036] S101: Acquire surface electromyographic signals at multiple preset muscle locations on the target body.

[0037] In this embodiment, terminal 5 receives surface electromyography signals collected by signal acquisition device 10 to represent seven actions, including clenching fist, extending, releasing wrist, retracting wrist, turning wrist to the left, turning wrist to the right, and raising thumb, and stores them in memory channel by channel, with each channel containing M points as a sample.

[0038] In one possible implementation, the specific implementation process of S101 includes:

[0039] The signal acquisition device 10 acquires surface electromyography (EMG) signals from multiple preset muscle locations of the target body. The signal acquisition device 10 includes an FPGA development board 11, a differential surface EMG sensor 13, and an analog-to-digital converter 12. The differential surface EMG sensor 13 is used to acquire simulated surface EMG signals from multiple preset muscle locations of the target body and send the simulated surface EMG signals to the analog-to-digital converter 12. The analog-to-digital converter 12 is used to convert the simulated surface EMG signals into analog signals to obtain surface EMG signals. The FPGA development board 11 is used to filter the surface EMG signals.

[0040] S102: For the surface electromyography (EMG) signal at each preset muscle location, perform wavelet packet decomposition on the surface EMG signal to obtain multiple sub-bands and the decomposition coefficients of each sub-band; and perform variational mode decomposition on the surface EMG signal to obtain multiple intrinsic mode functions.

[0041] In this embodiment, traditional signal feature engineering mostly extracts time-domain features, such as zero-crossing rate and peak value, and frequency-domain features, such as average frequency and absolute frequency of a specific frequency band. Surface electromyography (EMG) signals are non-stationary random signals, and their statistical characteristics change over time; therefore, simple time-domain or frequency-domain analysis methods cannot be used to extract features. According to the Wald decomposition theorem, generally, generalized stationary signals can be decomposed into a superposition of deterministic and random signals. Extending this further, any signal can be considered a superposition of any sub-signals. Therefore, this embodiment decomposes the EMG signal into multiple locally stationary or time-frequency focused sub-signals according to certain rules, and then performs time-frequency feature extraction on these sub-signals, which enhances the accuracy, representativeness, and effectiveness of signal feature extraction.

[0042] This embodiment uses wavelet packet decomposition and variational mode decomposition to decompose sEMG into several sub-signals, and then extracts time-domain statistical features from the sub-signals. By combining the advantages of the two methods, the adaptability and noise resistance of feature engineering are enhanced, which helps to suppress mode aliasing.

[0043] Specifically, common signal decomposition methods include Empirical Mode Decomposition (EMD), Local Mean Decomposition (LMD), and Ensemble Empirical Mode Decomposition (EEMD). EMD is an adaptive decomposition method based on the local extremum characteristics of a signal. It decomposes the original signal into Integrated Mode Factors (IMFs) through a sieving process. However, this method suffers from mode aliasing, meaning that the same IMF may contain signal components of different scales, leading to distorted decomposition results. LMD, on the other hand, decomposes the signal into a series of physically meaningful product functions (PFs) by calculating the local mean and envelope signal, exhibiting better stability when processing complex signals compared to EMD. EEMD, as an improved algorithm of EMD, effectively suppresses mode aliasing by adding white noise multiple times and performing ensemble averaging, improving the accuracy and reliability of the decomposition. However, it still suffers from slow computation speed.

[0044] This embodiment combines wavelet packet decomposition and variational mode decomposition. First, wavelet packet decomposition is used to perform preliminary multi-band segmentation of the sEMG signal, providing a structured signal framework for subsequent processing. Simultaneously, variational mode decomposition is used to further refine the sEMG, fully leveraging its ability to suppress mode aliasing. After signal decomposition, statistical features such as mean, variance, and kurtosis are extracted from the obtained sub-signals. These features reflect the amplitude distribution and variation characteristics of the signal from different perspectives. By complementing the advantages of the two decomposition methods—utilizing the multi-band analysis capability of wavelet packet decomposition and combining it with the anti-mode aliasing characteristics of variational mode decomposition—not only is the adaptive processing capability of feature engineering for complex biomedical signals enhanced, but the effectiveness and reliability of features are also improved, laying a solid foundation for subsequent applications such as sEMG-based action pattern recognition.

[0045] In one possible implementation, the specific implementation process of S102 includes:

[0046] A third-order Dobessi wavelet was selected as the wavelet basis function. A six-layer full-tree decomposition mode was adopted. By scaling and translating the wavelet basis function, the surface electromyography signal was decomposed into multiple sub-bands, and the decomposition coefficients corresponding to each sub-band were obtained. The sub-bands include high-frequency detail sub-bands and low-frequency approximation sub-bands. The decomposition coefficients include high-frequency detail coefficients and low-frequency approximation coefficients.

[0047] In this embodiment, wavelet packet decomposition (WPD), based on wavelet transform (WT) theory, can accurately decompose signals across all frequency components. While WT analyzes different scales and temporal locations of signals by scaling and translating a wavelet basis function, WPD decomposes not only low-frequency components but also high-frequency components simultaneously, producing a fine and uniform time-frequency plane division covering the entire frequency range. This provides higher resolution high-frequency information than WT, which is highly advantageous for capturing rapidly changing transient components in surface electromyography (EMG) signals.

[0048] Specifically, wavelet packet decomposition uses the third-order Dobessi wavelet as the wavelet basis to perform a 6-layer full tree decomposition on the surface electromyography (EMG) signal of each channel, generating low-frequency approximation coefficients and high-frequency detail coefficients. The low-frequency approximation coefficients represent the low-frequency part of the signal, preserving the overall trend and main features of the signal and reflecting the macroscopic structure of the signal; the high-frequency detail coefficients represent the high-frequency part of the signal, preserving the detailed features of the signal and reflecting the microscopic features of the signal. Wavelet packet decomposition is used to extract all subbands of all layers of the surface EMG signal.

[0049] Variational Mode Decomposition (VMD) decomposes a signal into a series of Intrinsic Mode Functions (IMFs) with finite bandwidths. Each IMF oscillates around a center frequency and has a definite physical meaning. VMD solves a variational optimization problem to ensure that the sum of all IMFs equals the original signal, and that the estimated bandwidth of each IMF is minimized, meaning that energy is concentrated as much as possible across frequencies. The center frequency and bandwidth of VMD are automatically determined by the algorithm, requiring no preset basis functions, and are insensitive to noise, allowing for adaptive feature extraction of surface electromyography (EMG) signals.

[0050] The preset number of modes for variational modal decomposition is 5, the fidelity constraint is 0, a large modal bandwidth is allowed, there is no noise suppression, and the center frequency is uniformly initialized. The convergence accuracy can be set to 1x10. -7 Finally, the surface electromyography signal of each channel was decomposed into 5 intrinsic mode functions.

[0051] After signal decomposition, the surface electromyography signal is decomposed into multiple nearly stationary sub-signals / sub-bands, greatly reducing the interference of the time-varying characteristics of the signal on feature extraction.

[0052] S103: Based on the decomposition coefficients of multiple sub-bands of the surface electromyography signal, extract the features corresponding to each sub-band and use them as the first feature; extract the features of each intrinsic mode function of the surface electromyography signal and use them as the second feature; extract the time domain features of the surface electromyography signal.

[0053] In this embodiment, based on wavelet packet decomposition and variational mode decomposition, multi-dimensional feature extraction is performed on each sub-signal / sub-band from the perspective of time domain and statistical characteristics.

[0054] In one possible implementation, the first feature includes median, skewness, maximum autocorrelation, and inter-subband ratio; the specific implementation process of S103 includes:

[0055] For each subband corresponding to the surface electromyography signal, the median, kurtosis, skewness, and maximum autocorrelation value of the decomposition coefficients of the subband are calculated; the maximum autocorrelation value is the maximum value of the autocorrelation function corresponding to the subband.

[0056] Based on formula Calculate the inter-subband ratio of the high-frequency detail subbands corresponding to the electromyographic signal on this surface;

[0057] Based on formula Calculate the inter-subband ratio of the low-frequency approximate subband corresponding to the electromyography signal on this surface;

[0058] Among them, K WPD_1K represents the inter-subband ratio of this high-frequency detail subband. WPD_2 a represents the inter-subband ratio of the low-frequency approximate subband. n d represents the mean of the high-frequency detail subband of the nth layer; n This represents the mean of the low-frequency approximate subband of the nth layer; n represents the number of layers in the wavelet decomposition.

[0059] The time-domain features include root mean square value, root mean square amplitude, skewness, kurtosis, and impulse characteristics.

[0060] Specifically, the median of each sub-band is the median value of the corresponding sub-band amplitude, reflecting the center of the signal energy distribution. Skewness is used to quantify the asymmetry of the signal distribution; a skewness greater than zero indicates right skewness, and a skewness less than zero indicates left skewness. Kurtosis is used to quantify the steepness of the distribution. Kurtosis greater than a preset kurtosis value indicates a sharp peak and heavy tail, such as in impact signals; kurtosis less than a preset kurtosis value indicates a gentle distribution, and this preset kurtosis value is between 2 and 5, preferably 3. The maximum autocorrelation value can be used to detect the energy concentration of the sub-band, and also corresponds to the signal periodicity and time delay characteristics. The inter-sub-band ratio feature is used to compare the energy between sub-bands and establish cross-scale correlations.

[0061] The root mean square (RMS) value of surface electromyography (SEMG) signals is the square root of the mean of the squares of the signal amplitudes representing the SEMG signals. It is used to characterize the energy level of the SEMG signals and is more sensitive to larger amplitude values. The root square amplitude value is the square root of the mean of the absolute values ​​of the signal amplitudes representing the SEMG signals. It is used to enhance the sensitivity to small amplitude signals, avoid the over-amplification of large values ​​by the squaring operation, and focus more on reflecting the average amplitude level of the signal.

[0062] In one possible implementation, the second feature includes median, skewness, maximum autocorrelation, and inter-subband ratio; the specific implementation process of S103 includes:

[0063] For each intrinsic mode function corresponding to the surface electromyography signal, the median, kurtosis, skewness, and maximum autocorrelation value of the intrinsic mode function are calculated; the maximum autocorrelation value is the maximum value of the intrinsic mode function.

[0064] Based on formula Calculate the inter-signal ratio of the intrinsic mode function corresponding to the surface electromyography signal;

[0065] Among them, K VMD c represents the ratio between sub-signals of the intrinsic mode function. m This represents the mean of the m-th eigenmode function in the decomposition; m represents the preset number of decomposition modes.

[0066] Specifically, the median of the intrinsic mode functions (IMFs) obtained from variational mode decomposition is extracted to reflect the center of signal energy distribution. Skewness is used to quantify the asymmetry of the signal distribution of the sub-signals corresponding to the IMFs; a skewness greater than zero indicates right skewness, and a skewness less than zero indicates left skewness. Kurtosis is used to quantify the steepness of the distribution of the sub-signals corresponding to the IMFs. Kurtosis greater than a preset kurtosis value indicates a sharp peak and heavy tail, such as in impulse signals; kurtosis less than a preset kurtosis value indicates a gentle distribution, and this preset kurtosis value is between 2 and 5, preferably 3. The maximum autocorrelation value can be used to detect the degree of energy concentration among the sub-signals corresponding to the IMFs, and also corresponds to the signal periodicity and time delay characteristics. The ratio characteristics among the sub-signals corresponding to the IMFs are used to compare the energy between the sub-signals and establish cross-scale correlations.

[0067] As can be seen from the above embodiments, by calculating skewness and kurtosis, the asymmetry and steepness of the signal distribution can be quantified; by extracting the maximum value of the autocorrelation function, the periodicity and correlation within the signal can be evaluated; ratios can be used to reflect differences in signal components; root mean square (RMS) and root square (RMS) amplitudes can be calculated to measure signal strength; kurtosis can capture signal impulse characteristics; and peak value and median can be extracted to describe signal extrema and central tendency. Finally, the above statistical features are integrated and output, and WPD+VMD+time domain multi-domain fusion is used to achieve information complementarity and avoid feature omission; VMD adaptive frequency band decomposition is adopted to adapt to the non-stationary characteristics of sEMG signals and reduce mode mixing; features are manually designed and statistical quantities and time domain features are automatically extracted, thereby enhancing stability, covering more comprehensive signal characteristics, and providing a data foundation for subsequent pattern recognition.

[0068] S104: Input the first feature, second feature and temporal feature corresponding to the surface electromyography signals of all preset muscle locations into the TabPFN model to obtain the action classification result of the target body.

[0069] In one possible implementation, the specific implementation process of S104 includes:

[0070] The first feature, second feature and temporal feature corresponding to the surface electromyography signals of all preset muscle locations are spliced ​​together to obtain multi-domain fusion features;

[0071] The multi-domain fusion features are input into the TabPFN model to obtain the action classification result of the target body;

[0072] The input layer of the TabPFN model employs a hybrid numerical feature encoder that integrates a multilayer perceptron and a bucket embedding mechanism.

[0073] Specifically, the multi-domain fusion feature is formed by fusing WPD, VMD, and time-domain features to create a multi-dimensional feature vector of N tabular data, rather than time-domain data. To improve computational efficiency, this embodiment can also perform dimensionality reduction and downsampling through one-dimensional convolution after multi-domain fusion feature extraction, thereby improving computational efficiency and reducing the risk of overfitting. The one-dimensional convolution includes 4 input channels, 1 output channel, a kernel size of 3, and a stride that can be set to 1.

[0074] In this embodiment, due to the high real-time requirements of this system and the high computational complexity of VMD iterative solution, the convergence accuracy is set to 1x10⁻¹⁰. -7 To balance speed and accuracy, this embodiment employs the TabPFN model, which performs exceptionally well on small sample tabular data and maintains excellent performance even with a large number of features and a small sample size.

[0075] TabPFN (Tabular Predictive Framework Network) is a deep learning model specifically designed for tabular data. Its core is based on the self-attention mechanism of the Transformer architecture, significantly improving generalization ability in small-sample scenarios by incorporating pre-trained prior knowledge. Traditional deep learning models often suffer performance degradation when processing tabular data due to features being sparse and high-dimensional. TabPFN, however, models dynamic feature interactions, automatically learning complex relationships between features and supporting mixed categorical and numerical data. Furthermore, its pre-training-fine-tuning paradigm allows the model to converge quickly on limited labeled data, making it particularly suitable for scenarios where large amounts of data are difficult to obtain.

[0076] During the instantiation phase of TabPFN, to improve computational efficiency, the pre-trained model configuration was adjusted to enable GPU acceleration to balance performance and resource consumption, while ignoring pre-training limitations to ensure the model is fully adapted to the specific dataset. In the fitting phase, the preprocessed feature data was divided into training and test sets in a 4:1 ratio, and the training set was used to optimize the model. Thanks to TabPFN's efficient algorithm design and parallel computing capabilities, the model can complete the fitting process within 1 second. After training, the test set was used to comprehensively evaluate the model's performance, outputting quantitative metrics such as accuracy, recall, precision, F1 score, Coen-Kappa score, and Matthews correlation coefficient, and generating a confusion matrix to visually display the classification results. Finally, the trained model was saved in joblib format for easy subsequent use. In the inference phase, the saved model was quickly loaded using the joblib library, enabling real-time prediction of new multi-domain fusion features, achieving efficient pattern recognition tasks.

[0077] After training, the TabPFN model can adapt to the electromyographic signal characteristics of different individuals, achieving personalized movement pattern recognition. The TabPFN model can automatically learn complex patterns and features in electromyographic signals and has excellent processing capabilities for high-dimensional, nonlinear electromyographic signal data. Compared to traditional machine learning algorithms, such as Support Vector Machines (SVM) and Artificial Neural Networks (ANN), TabPFN has significant advantages in processing speed and accuracy, enabling more accurate recognition of user movements.

[0078] As shown in Table 1, the performance of different models is compared. As can be seen from Table 1, TabPFN still performs well with a small sample size. It can quickly fit the data by collecting only a small number of samples from individual users, laying the foundation for system customization and personalization.

[0079] Table 1

[0080]

[0081]

[0082] In one embodiment, the TabPFN model provided in this embodiment introduces a hybrid numerical feature encoder into the input layer of the original TabPFN model. The hybrid numerical feature encoder achieves semantic representation of numerical features by integrating a multilayer perceptron (Management and Leadership Pathway, MLP) and a bucket embedding mechanism, which can enhance the model's feature representation capability, optimize computational efficiency, and improve generalization ability.

[0083] Specifically, the hybrid numerical feature encoder comprises two processing paths. The multilayer perceptron includes a linear layer (1D input, 16D output), a ReLU activation function, and another linear layer (16D input, embedding dimension output). It maps single-dimensional numerical features to a high-dimensional semantic space, capturing the continuous distribution features of numerical values. The bucketing embedding path maps numerical features to equally spaced buckets, using an embedding layer (nn.Embedding) to learn the discrete semantic representation of the buckets. Finally, the two outputs are summed to form an embedding vector that fuses continuous and discrete features, which is then flattened and concatenated to form a unified input representation.

[0084] S105: Control the wearable myoelectric prosthesis to perform corresponding actions based on the action classification results.

[0085] In one possible implementation, the specific implementation process of S105 includes:

[0086] The corresponding action execution instruction is determined based on the action classification result;

[0087] The wearable myoelectric prosthesis is controlled to perform corresponding actions based on the action execution command.

[0088] Specifically, terminal 5 generates corresponding action execution instructions based on the action classification results (such as clenching a fist, extending a hand, raising a thumb, etc.) output by the TabPFN model, and establishes a connection with the control board 31 of the robotic forearm via serial communication. The serial communication program is written based on Python's PySerial library. Terminal 5 encodes the action execution instructions into a string with a specific format and sends it to the control board 31 of the robotic forearm via serial communication. After receiving the action execution instructions, control board 31 parses the target rotation angle of the corresponding serial bus servo motor 32, drives the serial bus servo motor 32 to rotate to the specified angle, and thus drives the open-source robotic arm uHand to complete the corresponding action. Control board 31 can provide feedback to terminal 5 via the serial port on the operating status of the serial bus servo motor 32, such as whether it is in position or malfunctioning, to ensure the accuracy of action execution.

[0089] In this embodiment, the open-source robotic arm uHand provides rich interfaces and flexible structural design, which can be seamlessly integrated with the control board 31 and the serial bus servo motor 32, supporting a variety of complex hand movements, such as grasping, pinching, pronation and supination, etc., to meet the needs of different scenarios; the baud rate and other parameters of the serial communication are consistent with the overall system communication configuration, and are adapted to the serial communication parameters of the FPGA acquisition device and the terminal 5, ensuring the real-time performance and stability of command transmission.

[0090] In one possible implementation, the method provided in this embodiment further includes:

[0091] A graphical user interface is set up, and different system states are displayed modularly in the graphical user interface. The system states include the connection status between various devices, the waveforms of surface electromyography signals at various preset muscle locations, and the action classification results. The horizontal axis of the waveform is time, and the vertical axis is the signal amplitude.

[0092] Specifically, this embodiment uses a graphical user interface (GUI) designed based on the PySide6 framework to achieve multi-system adaptation and functional integration. The GUI intuitively displays device connections, signal status, and action classification results, providing a personalized training entry point to facilitate users' real-time monitoring and operation of the system. The mechanical forearm combines the open-source uHand robotic arm with a serial bus servo motor 32, achieving stable and precise control through serial communication. Combined with rich interfaces and a flexible structural design, it can meet the needs of complex hand movements in various scenarios, realizing a complete closed loop from signal acquisition, processing, and recognition to action execution and result feedback, providing users with an integrated and convenient operating experience.

[0093] Figure 3 A schematic diagram of the graphical user interface provided in this embodiment is shown, such as... Figure 3As shown, the graphical user interface modularly displays the signal waveform of surface electromyography (EMG) signals, the status of the signal acquisition device 10, the status of the control device, and the label display, enabling users to intuitively understand and operate the entire system.

[0094] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0095] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0096] Figure 4 A schematic diagram of the control device for a wearable myoelectric prosthesis provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below:

[0097] like Figure 4 As shown, the control device 100 for the wearable myoelectric prosthesis includes:

[0098] Signal acquisition module 110 is used to acquire surface electromyographic signals at multiple preset muscle locations of the target body;

[0099] The signal decomposition module 120 is used to perform wavelet packet decomposition on the surface electromyography signal at each preset muscle location to obtain multiple sub-bands and the decomposition coefficients of each sub-band; and to perform variational mode decomposition on the surface electromyography signal to obtain multiple intrinsic mode functions.

[0100] The feature extraction module 130 is used to extract the features corresponding to each sub-band based on the decomposition coefficients of multiple sub-bands of the surface electromyography signal and use them as the first feature; extract the features of each intrinsic mode function of the surface electromyography signal and use them as the second feature; and extract the time domain features of the surface electromyography signal.

[0101] The action recognition module 140 is used to input the first feature, second feature and temporal feature corresponding to the surface electromyography signals of all preset muscle locations into the TabPFN model to obtain the action classification result of the target body;

[0102] The prosthesis control module 150 is used to control the wearable myoelectric prosthesis to perform corresponding actions based on the action classification results.

[0103] In one possible implementation, the signal decomposition module 120 includes:

[0104] A third-order Dobessi wavelet was selected as the wavelet basis function. A six-layer full-tree decomposition mode was adopted. By scaling and translating the wavelet basis function, the surface electromyography signal was decomposed into multiple sub-bands, and the decomposition coefficients corresponding to each sub-band were obtained. The sub-bands include high-frequency detail sub-bands and low-frequency approximation sub-bands. The decomposition coefficients include high-frequency detail coefficients and low-frequency approximation coefficients.

[0105] In one possible implementation, the first feature includes median, skewness, maximum autocorrelation, and inter-subband ratio; the feature extraction module 130 includes:

[0106] For each subband corresponding to the surface electromyography signal, the median, kurtosis, skewness, and maximum autocorrelation value of the decomposition coefficients of the subband are calculated; the maximum autocorrelation value is the maximum value of the autocorrelation function corresponding to the subband.

[0107] Based on formula Calculate the inter-subband ratio of the high-frequency detail subbands corresponding to the electromyographic signal on this surface;

[0108] Based on formula Calculate the inter-subband ratio of the low-frequency approximate subband corresponding to the electromyography signal on this surface;

[0109] Among them, K WPD_1 K represents the inter-subband ratio of this high-frequency detail subband. WPD_2 a represents the inter-subband ratio of the low-frequency approximate subband. n d represents the mean of the high-frequency detail subband of the nth layer; n This represents the mean of the low-frequency approximate subband of the nth layer; n represents the number of layers in the wavelet decomposition.

[0110] The time-domain features include root mean square value, root mean square amplitude, skewness, kurtosis, and impulse characteristics.

[0111] In one possible implementation, the signal acquisition module 110 includes:

[0112] The signal acquisition device 10 acquires surface electromyography (EMG) signals from multiple preset muscle locations of the target body. The signal acquisition device 10 includes an FPGA development board 11, a differential surface EMG sensor 13, and an analog-to-digital converter 12. The differential surface EMG sensor 13 is used to acquire simulated surface EMG signals from multiple preset muscle locations of the target body and send the simulated surface EMG signals to the analog-to-digital converter 12. The analog-to-digital converter 12 is used to convert the simulated surface EMG signals into analog signals to obtain surface EMG signals. The FPGA development board 11 is used to filter the surface EMG signals.

[0113] In one possible implementation, the action recognition module 140 includes:

[0114] The first feature, second feature and temporal feature corresponding to the surface electromyography signals of all preset muscle locations are spliced ​​together to obtain multi-domain fusion features;

[0115] The multi-domain fusion features are input into the TabPFN model to obtain the action classification result of the target body;

[0116] The input layer of the TabPFN model employs a hybrid numerical feature encoder that integrates a multilayer perceptron and a bucket embedding mechanism.

[0117] In one possible implementation, the prosthesis control module 150 includes:

[0118] The corresponding action execution instruction is determined based on the action classification result;

[0119] The wearable myoelectric prosthesis is controlled to perform corresponding actions based on the action execution command.

[0120] In one possible implementation, the control device 100 for the wearable myoelectric prosthesis further includes:

[0121] A graphical display module is used to set up a graphical user interface and modularly display different system states on the graphical user interface; the system states include the connection status between various devices, the waveforms of surface electromyography signals at various preset muscle locations, and the action classification results; the horizontal axis of the waveform is time, and the vertical axis is the signal amplitude.

[0122] Figure 5 This is a schematic diagram of a terminal provided in an embodiment of the present invention. Figure 5 As shown, the terminal 5 in this embodiment includes: a processor 50, a memory 51, and a computer program 52 stored in the memory 51 and executable on the processor 50. When the processor 50 executes the computer program 52, it implements the steps in the control method embodiments of the various wearable myoelectric prostheses described above, for example... Figure 2 Steps S101 to S105 are shown. Alternatively, when the processor 50 executes the computer program 52, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 4 The functions of modules 110 to 150 are shown.

[0123] For example, the computer program 52 may be divided into one or more modules / units, which are stored in the memory 51 and executed by the processor 50 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program 52 in the terminal 5.

[0124] The terminal 5 can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The terminal 5 may include, but is not limited to, a processor 50 and a memory 51. Those skilled in the art will understand that... Figure 5 This is merely an example of terminal 5 and does not constitute a limitation on terminal 5. It may include more or fewer components than shown, or combine certain components, or different components. For example, the terminal may also include input / output devices, network access devices, buses, etc.

[0125] The processor 50 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0126] The memory 51 can be an internal storage unit of the terminal 5, such as a hard disk or memory of the terminal 5. The memory 51 can also be an external storage device of the terminal 5, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the terminal 5. Furthermore, the memory 51 can include both internal storage units and external storage devices of the terminal 5. The memory 51 is used to store the computer program and other programs and data required by the terminal. The memory 51 can also be used to temporarily store data that has been output or will be output.

[0127] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0128] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0129] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0130] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0131] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0132] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0133] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the control methods for the wearable myoelectric prostheses described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0134] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A control method for a wearable myoelectric prosthesis, characterized in that, include: Acquire surface electromyographic signals at multiple preset muscle locations on the target body; For the surface electromyography (EMG) signal at each preset muscle location, wavelet packet decomposition is performed on the surface EMG signal to obtain multiple sub-bands and the decomposition coefficients of each sub-band; The surface electromyography signal was then subjected to variational mode decomposition to obtain multiple intrinsic mode functions; Based on the decomposition coefficients of multiple sub-bands of the surface electromyography signal, the features corresponding to each sub-band are extracted and used as the first feature; Features of each intrinsic mode function of the surface electromyography signal are extracted and used as the second feature; temporal features of the surface electromyography signal are extracted. The first feature, second feature, and temporal feature corresponding to the surface electromyography signals of all preset muscle locations are input into the TabPFN model to obtain the action classification results of the target body. Based on the action classification results, the wearable myoelectric prosthesis is controlled to perform corresponding actions.

2. The control method for a wearable myoelectric prosthesis according to claim 1, characterized in that, The surface electromyography signal is subjected to wavelet packet decomposition to obtain multiple sub-bands and decomposition coefficients for each sub-band, including: A third-order Dobessi wavelet was selected as the wavelet basis function. A six-layer full-tree decomposition mode was adopted. By scaling and translating the wavelet basis function, the surface electromyography signal was decomposed into multiple sub-bands, and the decomposition coefficients corresponding to each sub-band were obtained. The sub-bands include high-frequency detail sub-bands and low-frequency approximation sub-bands. The decomposition coefficients include high-frequency detail coefficients and low-frequency approximation coefficients.

3. The control method for a wearable myoelectric prosthesis according to claim 1, characterized in that, The first feature includes median, skewness, maximum autocorrelation, and inter-subband ratio; The decomposition coefficients of multiple sub-bands based on the surface electromyography signal are used to extract features corresponding to each sub-band, including: For each subband corresponding to the surface electromyography signal, the median, kurtosis, skewness, and maximum autocorrelation value of the decomposition coefficients of the subband are calculated; the maximum autocorrelation value is the maximum value of the autocorrelation function corresponding to the subband. Based on formula Calculate the inter-subband ratio of the high-frequency detail subbands corresponding to the electromyographic signal on this surface; Based on formula Calculate the inter-subband ratio of the low-frequency approximate subband corresponding to the electromyography signal on this surface; Among them, K WPD_1 K represents the inter-subband ratio of this high-frequency detail subband. WPD_2 a represents the inter-subband ratio of the low-frequency approximate subband. n d represents the mean of the high-frequency detail subband of the nth layer; n This represents the mean of the low-frequency approximate subband of the nth layer; n represents the number of layers in the wavelet decomposition. The time-domain features include root mean square value, root mean square amplitude, skewness, kurtosis, and impulse characteristics.

4. The control method for a wearable myoelectric prosthesis according to claim 1, characterized in that, The acquisition of surface electromyographic signals at multiple preset muscle locations of the target body includes: The method involves acquiring surface electromyography (EMG) signals from multiple preset muscle locations on the target body using a signal acquisition device. The signal acquisition device includes an FPGA development board, a differential EMG sensor, and an analog-to-digital converter (ADC). The differential EMG sensor acquires simulated EMG signals from multiple preset muscle locations on the target body and sends these simulated signals to the ADC. The ADC performs analog-to-digital conversion on the simulated EMG signals to obtain actual EMG signals. The FPGA development board performs filtering on the EMG signals.

5. The control method for a wearable myoelectric prosthesis according to claim 1, characterized in that, The first feature, second feature, and temporal feature corresponding to the surface electromyography signals of all preset muscle locations are input into the TabPFN model to obtain the action classification result of the target body, including: The first feature, second feature and temporal feature corresponding to the surface electromyography signals of all preset muscle locations are spliced ​​together to obtain multi-domain fusion features; The multi-domain fusion features are input into the TabPFN model to obtain the action classification result of the target body; The input layer of the TabPFN model employs a hybrid numerical feature encoder that integrates a multilayer perceptron and a bucket embedding mechanism.

6. The control method for a wearable myoelectric prosthesis according to claim 1, characterized in that, The control of the wearable myoelectric prosthesis to perform corresponding actions based on the action classification results includes: The corresponding action execution instruction is determined based on the action classification result; The wearable myoelectric prosthesis is controlled to perform corresponding actions based on the action execution command.

7. The control method for a wearable myoelectric prosthesis according to claim 1, characterized in that, The method further includes: A graphical user interface is set up, and different system states are displayed modularly in the graphical user interface. The system states include the connection status between various devices, the waveforms of surface electromyography signals at various preset muscle locations, and the action classification results. The horizontal axis of the waveform is time, and the vertical axis is the signal amplitude.

8. A terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the control method for the wearable myoelectric prosthesis as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the control method for the wearable myoelectric prosthesis as described in any one of claims 1 to 7.

10. A wearable myoelectric prosthetic system, characterized in that, include: Wearable myoelectric prosthesis, signal acquisition device, and terminal as described in claim 8; The signal acquisition device includes an FPGA development board, a differential surface electromyography sensor, and an analog-to-digital converter; The differential surface electromyography sensor is used to collect surface electromyography analog signals at multiple preset muscle locations of the target body, and send the surface electromyography analog signals to the analog-to-digital converter; The analog-to-digital converter is used to convert the surface electromyography analog signal into a digital signal to obtain the surface electromyography signal. The FPGA development board is used to filter the surface electromyography signal; The filtered surface electromyography signal is then sent to the terminal via a serial port. The terminal communicates with the wearable myoelectric prosthesis via a serial port.

Images