Intelligent artificial limb control algorithm based on EMG electromyographic signals

Abstract

The invention belongs to the field of intelligent prostheses, and particularly relates to an intelligent prosthetic control algorithm based on EMG electromyographic signals, which comprises the following steps: S1, electromyographic signal acquisition: acquiring the electromyographic signals of the forearm of a user through a multi-channel surface electromyographic sensor, the electromyographic signals being sEMG signals; s2, signal preprocessing; according to the time-frequency feature enhancement and mixed learning architecture, the system can realize more accurate and stable small sample gesture recognition, the processing capability and classification performance of the non-stationary electromyographic signals are remarkably improved, and in addition, due to the fact that deep features and discriminative manual features are fused and a lightweight network and a multi-modal decision mechanism are combined, the recognition efficiency of the non-stationary electromyographic signals is improved. A user can realize reliable and smooth gesture interaction in a low-delay and low-data-dependence embedded environment, and does not need to depend on a large amount of labeled data or complex computing resources, so that the whole method has higher adaptability and practicability.

Description

Technical Field

[0001] This invention relates to the field of intelligent prosthetics technology, and in particular to an intelligent prosthetic control algorithm based on EMG electromyography signals. Background Technology

[0002] Intelligent prosthetic control technology based on surface electromyography (sEMG) is an important research direction in the fields of rehabilitation engineering and neuromechanical interfaces. It provides a highly intuitive and natural solution for the reconstruction of motor function for upper limb amputees. The core value of this technology lies in the non-invasive acquisition of electrophysiological signals generated by residual muscle groups during contraction using electrodes attached to the skin surface of the residual limb. These signals are then decoded in real time using signal processing technology and machine learning algorithms to accurately identify the user's movement intentions. Based on this, the system can drive the bionic prosthesis to perform corresponding movements, which not only significantly improves the user's control experience of the prosthesis and daily living self-care ability, but also further promotes the development of intelligent rehabilitation equipment towards a more intelligent and humanized direction.

[0003] The current mainstream technical solutions can be mainly divided into the following two categories: one is based on traditional machine learning methods, such as support vector machines (SVM) and linear discriminant analysis (LDA). These methods usually rely on manually extracted time-domain features, such as mean absolute value (MAV), wavelength (WL), and zero-crossing rate (ZC), and realize motion intention recognition by classifying these features into patterns. This type of solution has the advantages of relatively simple model structure, low computational burden, and easy deployment to embedded systems; however, it also has some obvious limitations, including limited ability to express manual features, sensitivity to individual user differences, poor cross-user generalization performance, and difficulty in effectively handling complex continuous motion intentions. The second type is end-to-end solutions based on deep learning, typically represented by convolutional neural networks (CNN), long short-term memory networks (LSTM), and their hybrid models. The significant advantage of this type of method is that it can automatically learn deep, highly abstract feature representations from raw or preprocessed sEMG signals, avoiding tedious and potentially subjectively biased manual feature engineering, and often achieving higher recognition accuracy under ideal conditions.

[0004] However, this powerful capability comes with serious challenges: First, deep models typically require large-scale, high-quality training data, meaning that each user often needs to provide hundreds of repetitive data collections for each gesture, a tedious and time-consuming process that places a heavy burden on users and restricts the rapid, personalized deployment of the model. More importantly, the complex computational graph structure and huge number of parameters of the model cause its inference speed to face bottlenecks on resource-constrained embedded prosthetic controllers, with inference latency reaching 100 milliseconds or even higher. This significant latency far exceeds the threshold required for real-time, natural interaction (usually <100ms), severely affecting the smoothness of control and user experience, making it difficult to meet the actual needs of daily real-time control of prostheses. Therefore, we propose an intelligent prosthetic control algorithm based on EMG electromyography signals. Summary of the Invention

[0005] The intelligent prosthetic control algorithm based on EMG electromyography signals proposed in this invention solves the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The intelligent prosthetic control algorithm based on EMG electromyography signals includes the following steps: S1: Electromyography (EMG) signal acquisition: Electromyography (EMG) signals of the user's forearm are acquired through a multi-channel surface electromyography (SEMG) sensor. S2: Signal preprocessing: The raw sEMG signal obtained from S1 is highly susceptible to various noise contaminations. Therefore, it is necessary to preprocess the raw sEMG signal using a bandpass filter of 0.3Hz-500Hz and notch filtering algorithms of 50Hz and 100Hz. This can preserve the effective sEMG signal while reducing the impact of 50Hz power frequency interference. S3: Time-frequency map calculation: Combining deep learning methods, time-frequency analysis is performed on the preprocessed sEMG signal to segment each channel signal into a 100ms window length and a 20ms step length, and the one-dimensional signal is converted into a two-dimensional time-frequency map through continuous wavelet transform, thereby constructing a time-frequency representation of the multi-channel electromyography signal and obtaining a multi-channel time-frequency map, which provides input for subsequent deep learning feature extraction; S4: Manual feature extraction: Extract six types of time-domain feature parameters within the same signal window. The six types of time-domain feature parameters include root mean square (RMS), wavelength (WL), zero-crossing rate (ZC), slope sign change (SSC), mean absolute value (MAV), and variance (VAR). Then, evaluate the discriminative ability of each feature based on the training set data, and select the top n features with the largest variance to form the optimal feature subset, so as to improve the generalization ability of the model and reduce the computational burden. S5: Dual-branch model construction and training: The model is trained based on deep learning and handcrafted features respectively, as detailed below: S501: For the deep learning branch, the multi-channel time-frequency graph is input into the lightweight network ShuffleNet V2, and the model is trained using the Adam optimizer and cross-entropy loss function. The model data is divided into training and test sets in a 7:3 ratio. The model training is iterated for 40 rounds, and the model parameters with the highest accuracy on the test set are selected as the final output of this branch. S502: For the handcrafted feature branch, the selected handcrafted features are input into the support SVM vector machine, and the optimal splitting hyperplane in the high-dimensional feature space is learned by maximizing the classification margin. S6: Predict fusion smoothing, which includes the following steps: S601: An adaptive weighted fusion algorithm based on entropy is used to fuse the prediction results of the deep learning branch and the handmade feature branch; S602: Building upon S601, a time-weighted window smoothing mechanism is introduced. The processing interval is 140ms of data from the current moment. Multiple data windows are extracted with a window length of 100ms and a step size of 10ms, and predictions are performed on each window to obtain a series of probability distribution results. These results are then subjected to exponentially decaying weighted smoothing filters, with different weights assigned based on the timestamp of each window. Predictions closer to the current moment have higher weights, effectively smoothing out instantaneous jitter while maintaining system response sensitivity. S603: Based on S602, a dual-frame threshold state machine is also introduced to judge the output probability, so as to filter out instantaneous jitter signals and low-confidence fuzzy classification results. Only when a continuous high-confidence result appears is it considered a reliable output. S7: Action command transmission: Only when the system continuously outputs high-confidence classification results exceeding a set threshold and maintains stability for a certain period of time, will the final control command be sent to the prosthetic terminal via the Bluetooth module, thereby achieving reliable and stable gesture output at the hardware level.

[0007] Preferably, in S1, the electrode positions of the multi-channel surface electromyography (SEMG) sensor should preferentially cover the key muscle groups most relevant to the fine motor movements of the fingers and wrist, to ensure stable capture of electromyographic activity patterns sufficient to distinguish different hand gesture intentions. The key muscle groups most relevant to the fine motor movements of the fingers and wrist are mainly the lateral wrist flexor, flexor digitorum superficialis, extensor digitorum, and flexor pollicis longus.

[0008] Preferably, in S2, noise pollution refers to one or more of power frequency interference, motion artifacts, electrode-skin contact noise, and other bioelectrical signal crosstalk.

[0009] Preferably, in step S3, the continuous wavelet transform preferably uses Morlet wavelet as the basis function to decompose the sEMG signal into multi-scale components and generate wavelet coefficient spectra to effectively capture transient features in the electromyography signal. The generated time-frequency image is uniformly normalized into a 224×224 pixel grayscale image, and the time-frequency images of multiple channels are superimposed.

[0010] Preferably, in S4, the root square RMS represents the average power or amplitude of the signal within the window time, distinguishing between muscle contraction intensity and detection activation state; Wavelength WL represents the sum of the absolute differences between adjacent signal samples within the window, which helps to distinguish between smooth gestures and rapid shaking gestures; Zero crossing rate (ZC) represents the number of times a signal crosses the zero level within a window. High-frequency signals cross the zero line more frequently than low-frequency signals, so it can help distinguish between fast-switching gestures and slow-maintaining gestures. The slope sign change SSC represents the number of times the signal slope direction changes, which helps to distinguish muscle activities with similar amplitudes but different oscillation patterns; The Mean Absolute Value (MAV) represents the average of the absolute values ​​of the signal, and it complements the Root Mean Square (RMS) to reflect the overall amplitude level of the signal. Variance (VAR) measures how much a signal value deviates from its mean. It provides information about the signal's "stability" and helps distinguish between dynamic and static gestures.

[0011] Preferably, in step S601, the adaptive weighted fusion algorithm dynamically allocates weights based on the information entropy of the output probability of each branch, assigning higher weights to branches with lower entropy values, thereby enhancing the reliability of the overall decision.

[0012] Preferably, the information entropy of the output probability of each branch is calculated using the following formula: , where p is the predicted probability corresponding to each gesture category; The formula for the fusion weights is: The formula for the fusion probability is: .

[0013] Compared with existing technologies, the time-frequency feature enhancement and hybrid learning architecture of this invention enables the system to achieve more accurate and stable small-sample gesture recognition, significantly improving the processing capability and classification performance of non-stationary electromyographic signals. In addition, by integrating deep features and discriminative handcraft features, and combining lightweight networks and multimodal decision mechanisms, users can achieve reliable and smooth gesture interaction in low-latency, low-data-dependency embedded environments without relying on a large amount of labeled data or complex computing resources, making the overall method more adaptable and practical. Attached Figure Description

[0014] Figure 1This is a flowchart of the intelligent prosthetic control algorithm based on EMG electromyography signals proposed in this invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] Reference Figure 1 The intelligent prosthetic control algorithm based on EMG electromyography signals includes the following steps: S1: Electromyography (sEMG) signal acquisition: Electromyography (sEMG) signals of the user's forearm are acquired through a multi-channel surface electromyography (SEMG) sensor. The electrode positions of the multi-channel SEMG sensor should preferentially cover the key muscle groups most relevant to fine motor skills of the fingers and wrist to ensure stable capture of electromyography activity patterns that can distinguish different hand gestures. The key muscle groups most relevant to fine motor skills of the fingers and wrist are mainly the lateral wrist flexor, flexor digitorum superficialis, extensor digitorum, and flexor pollicis longus. S2: Signal preprocessing: The raw sEMG signal obtained from S1 is highly susceptible to various noise contaminations, including power frequency interference, motion artifacts, electrode-skin contact noise, and crosstalk from other bioelectrical signals. Therefore, the raw sEMG signal needs to be preprocessed using bandpass filtering (0.3Hz-500Hz) and notch filtering (50Hz and 100Hz) algorithms to retain the effective sEMG signal while reducing the impact of 50Hz power frequency interference. S3: Time-Frequency Map Calculation: Combining deep learning methods, time-frequency analysis is performed on the preprocessed sEMG signal. Each channel signal is segmented into a 100ms window with a 20ms step size. The one-dimensional signal is converted into a two-dimensional time-frequency map through continuous wavelet transform, thereby constructing a time-frequency representation of the multi-channel electromyography signal. The continuous wavelet transform preferably uses Morlet wavelet as the basis function to decompose the sEMG signal into multi-scale components and generate wavelet coefficient spectra to effectively capture transient features in the electromyography signal. The generated time-frequency map is uniformly normalized into a 224×224 pixel grayscale image. The time-frequency maps of multiple channels are superimposed to obtain a multi-channel time-frequency map, which provides input for subsequent deep learning feature extraction. S4: Manual Feature Extraction: Six types of temporal feature parameters are extracted within the same signal window. These six types of temporal feature parameters include Root Mean Square (RMS), Wavelength (WL), Zero Crossing Rate (ZC), Slope Sign Change (SSC), Mean Absolute Value (MAV), and Variance (VAR). RMS represents the average power or amplitude of the signal within the window time, distinguishing muscle contraction intensity and detection activation state. Wavelength (WL) represents the sum of the absolute differences between adjacent signal samples within the window, helping to distinguish between smooth gestures and rapid jittering gestures. The Zero Crossing Rate (ZC) represents the number of times the signal crosses the zero level within the window; high-frequency signals cross the zero line more frequently than low-frequency signals, therefore it can... The system helps distinguish between rapidly switching gestures and slowly sustained gestures; the slope sign change (SSC) represents the number of times the signal slope direction changes, which helps distinguish muscle activities with similar amplitudes but different oscillation patterns; the mean absolute value (MAV) represents the average value of the absolute value of the signal, which complements the root mean square (RMS) and reflects the overall amplitude level of the signal; the variance (VAR) represents the degree to which the signal value deviates from its mean, which provides information on the "stability" of the signal and helps distinguish between dynamic and static gestures. Then, based on the training set data, the discriminative ability of each feature is evaluated, and the top n features with the largest variance are selected to form the optimal feature subset to improve the model's generalization ability and reduce the computational burden. S5: Dual-branch model construction and training: The model is trained based on deep learning and handcrafted features respectively, as detailed below: S501: For the deep learning branch, the multi-channel time-frequency graph is input into the lightweight network ShuffleNet V2, and the model is trained using the Adam optimizer and cross-entropy loss function. The model data is divided into training and test sets in a 7:3 ratio. The model training is iterated for 40 rounds, and the model parameters with the highest accuracy on the test set are selected as the final output of this branch. S502: For the handcrafted feature branch, the selected handcrafted features are input into the support SVM vector machine, and the optimal splitting hyperplane in the high-dimensional feature space is learned by maximizing the classification margin. S6: Predict fusion smoothing, which includes the following steps: S601: An adaptive weighted fusion algorithm based on entropy is used to fuse the prediction results of the deep learning branch and the handcrafted feature branch. The adaptive weighted fusion algorithm dynamically allocates weights according to the information entropy of the output probability of each branch, assigning higher weights to branches with lower entropy values, thereby enhancing the reliability of the overall decision. The information entropy of the output probability of each branch is calculated using the following formula: Where p is the predicted probability corresponding to each gesture category; the formula for the fusion weight is: The formula for the fusion probability is: ; S602: Building upon S601, a time-weighted window smoothing mechanism is introduced. The processing interval is 140ms of data from the current moment. Multiple data windows are extracted with a window length of 100ms and a step size of 10ms, and predictions are performed on each window to obtain a series of probability distribution results. These results are then subjected to exponentially decaying weighted smoothing filters, with different weights assigned based on the timestamp of each window. Predictions closer to the current moment have higher weights, effectively smoothing out instantaneous jitter while maintaining system response sensitivity. S603: Based on S602, a dual-frame threshold state machine is also introduced to judge the output probability, so as to filter out instantaneous jitter signals and low-confidence fuzzy classification results. Only when a continuous high-confidence result appears is it considered a reliable output. S7: Action command transmission: Only when the system continuously outputs high-confidence classification results exceeding the set threshold and maintains stability for a certain period of time, the final control command is sent to the prosthetic terminal via the Bluetooth module, thereby achieving reliable and stable gesture output at the hardware level. In this embodiment, the time-frequency feature enhancement and hybrid learning architecture enables the system to achieve more accurate and stable small-sample gesture recognition, significantly improving the processing capability and classification performance of non-stationary electromyographic signals. Furthermore, by integrating deep features and discriminative handcrafted features, and combining lightweight networks and multimodal decision-making mechanisms, users can achieve reliable and smooth gesture interaction in a low-latency, low-data-dependency embedded environment without relying on a large amount of labeled data or complex computing resources, making the overall method more adaptable and practical.

[0017] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent prosthetic limb control algorithm based on EMG electromyography signals, characterized in that, Includes the following steps: S1: Electromyography (EMG) signal acquisition: Electromyography (EMG) signals of the user's forearm are acquired through a multi-channel surface electromyography (SEMG) sensor. S2: Signal preprocessing: The raw sEMG signal obtained from S1 is highly susceptible to various noise contaminations. Therefore, it is necessary to preprocess the raw sEMG signal using a bandpass filter of 0.3Hz-500Hz and notch filtering algorithms of 50Hz and 100Hz. This can preserve the effective sEMG signal while reducing the impact of 50Hz power frequency interference. S3: Time-frequency map calculation: Combining deep learning methods, time-frequency analysis is performed on the preprocessed sEMG signal to segment each channel signal into a 100ms window length and a 20ms step length, and the one-dimensional signal is converted into a two-dimensional time-frequency map through continuous wavelet transform, thereby constructing a time-frequency representation of the multi-channel electromyography signal and obtaining a multi-channel time-frequency map, which provides input for subsequent deep learning feature extraction; S4: Manual feature extraction: Extract six types of time-domain feature parameters within the same signal window. The six types of time-domain feature parameters include root mean square (RMS), wavelength (WL), zero-crossing rate (ZC), slope sign change (SSC), mean absolute value (MAV), and variance (VAR). Then, evaluate the discriminative ability of each feature based on the training set data, and select the top n features with the largest variance to form the optimal feature subset, so as to improve the generalization ability of the model and reduce the computational burden. S5: Dual-branch model construction and training: The model is trained based on deep learning and handcrafted features respectively, as detailed below: S501: For the deep learning branch, the multi-channel time-frequency graph is input into the lightweight network ShuffleNet V2, and the model is trained using the Adam optimizer and cross-entropy loss function. The model data is divided into training and test sets in a 7:3 ratio. The model training is iterated for 40 rounds, and the model parameters with the highest accuracy on the test set are selected as the final output of this branch. S502: For the handcrafted feature branch, the selected handcrafted features are input into the support SVM vector machine, and the optimal splitting hyperplane in the high-dimensional feature space is learned by maximizing the classification margin. S6: Predict fusion smoothing, which includes the following steps: S601: An adaptive weighted fusion algorithm based on entropy is used to fuse the prediction results of the deep learning branch and the handmade feature branch; S602: Building upon S601, a time-weighted window smoothing mechanism is introduced. The processing interval is 140 ms of data from the current moment. Multiple data windows are extracted with a window length of 100 ms and a step size of 10 ms, and predictions are performed on each window to obtain a series of probability distribution results. These results are then subjected to exponentially decaying weighted smoothing filters, assigning different weights based on the timestamp of each window. Predictions closer to the current moment have higher weights, effectively smoothing out instantaneous jitter while maintaining system response sensitivity. S603: Based on S602, a dual-frame threshold state machine is also introduced to judge the output probability, so as to filter out instantaneous jitter signals and low-confidence fuzzy classification results. Only when a continuous high-confidence result appears is it considered a reliable output. S7: Action command transmission: Only when the system continuously outputs high-confidence classification results exceeding a set threshold and maintains stability for a certain period of time, will the final control command be sent to the prosthetic terminal via the Bluetooth module, thereby achieving reliable and stable gesture output at the hardware level.

2. The intelligent prosthetic limb control algorithm based on EMG electromyography signals according to claim 1, characterized in that, In S1, the electrode positions of the multi-channel surface electromyography sensor should preferentially cover the key muscle groups most relevant to the fine movements of the fingers and wrists, to ensure that it can stably capture electromyographic activity patterns that are sufficient to distinguish different hand gesture intentions. The key muscle groups most relevant to the fine movements of the fingers and wrists are mainly the lateral wrist flexor, flexor digitorum superficialis, extensor digitorum, and flexor pollicis longus.

3. The intelligent prosthetic limb control algorithm based on EMG electromyography signals according to claim 1, characterized in that, In S2, noise pollution refers to one or more of the following: power frequency interference, motion artifacts, electrode-skin contact noise, and crosstalk of other bioelectrical signals.

4. The intelligent prosthetic limb control algorithm based on EMG electromyography signals according to claim 1, characterized in that, In S3, the continuous wavelet transform preferably uses the Morlet wavelet as the basis function to decompose the sEMG signal into multi-scale components and generate a wavelet coefficient spectrum to effectively capture the transient features in the electromyography signal. The generated time-frequency image is uniformly normalized into a 224×224 pixel grayscale image, and the time-frequency images of multiple channels are superimposed.

5. The intelligent prosthetic control algorithm based on EMG electromyography signals according to claim 1, characterized in that, In S4, the root square RMS represents the average power or amplitude of the signal within the window time, distinguishing between muscle contraction intensity and detection activation state. Wavelength WL represents the sum of the absolute differences between adjacent signal samples within the window, which helps to distinguish between smooth gestures and rapid shaking gestures; Zero crossing rate (ZC) represents the number of times a signal crosses the zero level within a window. High-frequency signals cross the zero line more frequently than low-frequency signals, so it can help distinguish between fast-switching gestures and slow-maintaining gestures. The slope sign change SSC represents the number of times the signal slope direction changes, which helps to distinguish muscle activities with similar amplitudes but different oscillation patterns; The Mean Absolute Value (MAV) represents the average of the absolute values ​​of the signal, and it complements the Root Mean Square (RMS) to reflect the overall amplitude level of the signal. Variance (VAR) measures how much a signal value deviates from its mean. It provides information about the "stability" of the signal and helps distinguish between dynamic and static gestures.

6. The intelligent prosthetic limb control algorithm based on EMG electromyography signals according to claim 1, characterized in that, In step S601, the adaptive weighted fusion algorithm dynamically allocates weights based on the information entropy of the output probability of each branch, assigning higher weights to branches with lower entropy values, thereby enhancing the reliability of the overall decision.

7. The intelligent prosthetic limb control algorithm based on EMG electromyography signals according to claim 6, characterized in that, The information entropy of the output probability of each branch is calculated using the following formula: , where p is the predicted probability corresponding to each gesture category; The formula for the fusion weights is: ; The formula for the fusion probability is: .

Images