A robust electromyography pattern recognition method against electrode shift and signal loss

Abstract

The application discloses a myoelectric pattern recognition method robust to electrode shift and signal loss. The method first simulates the features of the electrode shift visual angle through the steps of feature map interpolation, translation and down-sampling, and obtains the features of the mask visual angle by randomly setting zero to individual channel data, then inputs the features of the electrode shift visual angle and the mask visual angle into two branches respectively by constructing a twin self-encoder network, and trains a robust myoelectric feature reconstruction model by using multiple reconstruction error backpropagation. The application can realize high-precision myoelectric gesture classification under the condition that the double influencing factors of electrode shift and signal loss are superimposed, and provides a new solution for the problems and challenges faced by the myoelectric pattern recognition technology in practical application.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing and human-computer interaction, and in particular to a robust electromyography pattern recognition method that overcomes the simultaneous problems of electrode offset and signal loss in electromyography wristbands. Background Technology

[0002] Electromyography (EMG) pattern recognition is a technology that decodes intentional movements from surface electromyography (SEMG) signals and converts them into specific control commands. This technology involves the detection, processing, and application of weak electrophysiological signals, and can establish an important neural-computer interface with broad application prospects in prosthetic control and human-computer interaction. Wearable EMG bracelets for detecting hand gestures have become a novel and ideal human-computer interaction method, but this also presents greater challenges to the robustness of EMG pattern recognition. Based on the sensing principle and characteristics of surface electromyography sensors, electrode misalignment and signal loss are the main interference factors reducing the robustness of EMG pattern recognition in daily use of EMG bracelets.

[0003] Electrode misalignment primarily stems from the rotational shift of the contact position between the electrodes and the skin during repeated wear of EMG bracelets, leading to variations in the recorded data. Existing research on overcoming electrode misalignment interference can be broadly categorized into three types: data augmentation-based, transfer learning-based, and data modeling-based. Data augmentation-based methods collect data from multiple wearing positions and train the classifier using a richer dataset. This approach effectively improves the robustness of gesture recognition, but the resulting data collection burden is unacceptable to users. Thanks to advancements in deep learning technology, transfer learning-based methods utilize the limited data provided by users when re-wearing the EMG bracelet to quickly calibrate the original model and achieve good gesture classification accuracy. Furthermore, data modeling based on the spatial distribution of the electrodes can effectively simulate data under electrode misalignment conditions, leading to two solutions: data augmentation of the training set and data calibration of the test set. Data modeling-based methods can significantly improve the resistance to electrode misalignment interference in EMG gesture recognition systems without requiring any additional data or only a small amount of calibration data.

[0004] Signal loss primarily stems from friction or poor contact between the electrodes and the skin, leading to abnormal signals recorded in some channels. Effective methods for recovering missing data are needed to address this issue. Data recovery methods for abnormal channels can be categorized into "smoothing"-based and "learning"-based methods. "Smoothing"-based methods include using the average of neighboring channel data as the data for the abnormal channel, applying median filtering to the signal, and using different interpolation methods to recover the missing signal. These methods are traditional solutions and are often performed during signal preprocessing. "Learning"-based methods employ an encoder-decoder model architecture, inputting abnormal data into a neural network with the aim of obtaining an output closer to normal data. These methods demonstrate better data recovery performance than traditional methods and exhibit higher robustness in electromyography (EMG) pattern recognition tasks.

[0005] While the aforementioned solutions to the problems of electrode misalignment and signal loss have made encouraging progress, current EMG pattern recognition technology still falls short of the needs of consumer-grade intelligent human-computer interaction applications. The key issue is that existing methods primarily address only one type of interference. In reality, electrode misalignment and data loss often coexist in everyday applications, and the feasibility of existing solutions is significantly reduced when multiple interference factors are combined. To advance the practical application of EMG pattern recognition technology, there is an urgent need to develop new solutions that can simultaneously address the challenges of multiple interference factors. Summary of the Invention

[0006] To address the shortcomings of current electromyography (EMG) pattern recognition technology, this invention proposes an EMG pattern recognition method robust to electrode offset and signal loss, aiming to maintain good gesture recognition accuracy even when interference from electrode offset and signal loss exists simultaneously. This provides a new solution to the problems and challenges faced by EMG pattern recognition technology in practical applications.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a robust electromyography pattern recognition method for electrode offset and signal loss, characterized by comprising the following steps: Step 1: Construct the original electromyography feature dataset : use Channel electromyography bracelet data acquisition 10 subjects The surface electromyography (EMG) signals of various hand gestures were analyzed, and sliding window and active segment detection were performed on the EMG signals to obtain... The dataset of electromyography (EMG) data samples was divided into a training set and a test set. This represents the number of samples in the training set of electromyography (EMG) data. The number of electromyography data samples in the test set represents the number of samples in the test set. ; Extracting from each electromyography data sample The original electromyography feature dataset is obtained by analyzing time-domain features. ,in, The training set representing the original electromyographic features, and ,in, Representing the One original electromyography feature training sample, Representing the One original electromyography feature training sample The true label, Represents the original electromyographic characteristic test set, and ,in, Representing the One original electromyography characteristic test sample, Representing the Original electromyography characteristic test samples The real label, and ; Step 2: Construct a high-resolution electromyography feature training set from the electrode offset perspective. : The original electromyographic feature training set Each training sample in the dataset undergoes third-order polynomial interpolation along the channel dimension to obtain a high-resolution electromyography feature training set. ,in, Representing the High-resolution electromyography feature training samples, and Then, the high-resolution electromyography feature training set was used. Each training sample in the lateral translation Each pixel is filled with the original pixel that was shifted out, thus obtaining a high-resolution electromyographic feature training set from the electrode offset perspective. ; Step 3: Set the number of samples in each batch to... The original electromyographic feature training set High-resolution electromyographic feature training set with electrode offset perspective The data are divided into several batches; the raw electromyographic feature data in the current batch are denoted as... ; Step 4: Downsample all high-resolution electromyographic feature training samples from all electrode offset perspectives in the current batch to obtain electromyographic feature training data from all electrode offset perspectives in the current batch. ,in, Represents the first in the current batch Electromyographic feature training samples from each electrode offset perspective, and , represent The real label, and ; Step 5: Electromyographic feature data of electrode offset angle in the current batch. All training samples are randomly zeroed out in Q channels to obtain the electromyography (EMG) feature data from the masked perspective in the current batch. ,in, The first mask viewpoint in the current batch One electromyographic feature sample, and , represent The real label, and ; Step 6, , as well as Each sample in the dataset is transformed into a one-dimensional format and denoted as an electromyographic feature vector at the electrode offset perspective. Electromyographic feature vectors from a masked perspective and the original electromyographic feature vector ; Step 7: Construct a twin autoencoder network, containing two branches, each used for... and The process is performed to obtain the electromyographic feature reconstruction vectors from the electrode offset perspective in the current batch. Electromyography feature reconstruction vector from the mask perspective ; Step 8: Train the Siamese autoencoder network by inputting different batches of data into it, following the process in steps 4-7, and calculate the total loss function for each batch. The network parameters are updated by backpropagation until the total loss converges, thus obtaining a robust electromyography feature data reconstruction model. Step 9: Use a robust electromyography (EMG) feature data reconstruction model on the original EMG feature training set. All samples are reconstructed to obtain a new electromyography feature data training set, and the SVM classifier is trained using the new electromyography feature data training set to obtain a trained SVM classifier. Step 10: From the original electromyography characteristic test set Test samples are randomly selected and input into a robust electromyography feature data reconstruction model for reconstruction. The results are then input into the trained SVM classifier to obtain the predicted labels of the test samples.

[0008] The electromyography pattern recognition method robust to electrode offset and signal loss described in this invention is characterized in that the first branch of the twin autoencoder network is composed of an input layer, an encoding layer, a hidden representation layer, a decoding layer and an output layer in sequence, and the structure of the second branch is completely consistent with the structure of the first branch and shares the weights.

[0009] The total loss function in step 8 Create it using the following steps: Construct the current batch of data according to equations (1), (2), and (3) respectively. and Reconstruction error between , and Reconstruction error between , and Reconstruction error ; (1) (2) (3) In equations (1)-(3), , , Represent , , The Electromyographic feature reconstruction vector of a sample, electromyographic feature reconstruction vector from a masked perspective, and original electromyographic feature vector; Construct the total loss function for the current batch according to equation (4). ; (4).

[0010] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the robust electromyography pattern recognition method, and the processor is configured to execute the program stored in the memory.

[0011] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to perform the steps of the robust electromyography pattern recognition method.

[0012] Compared with the prior art, the beneficial effects of the present invention are reflected in: 1. The method of the present invention utilizes electromyographic feature maps of electrode offset viewpoint and occlusion viewpoint, combined with the designed twin autoencoder network, to enable the model to learn the ability to reconstruct normal data from data disturbed by different factors, thereby maintaining higher classification accuracy in gesture recognition tasks.

[0013] 2. In this invention, steps 4 and 5 utilize the electrode distribution pattern of the electromyography wristband to effectively simulate the data after electrode offset through interpolation, random translation, and downsampling. Based on this, masking operations are randomly performed on the data of several channels according to the actual situation, taking into account various situations of signal loss, making the trained electromyography feature reconstruction model more generalizable, thereby significantly improving the robustness of the electromyography gesture interaction system.

[0014] 3. The twin autoencoder network proposed in this invention has low input feature dimensionality and a simple hidden layer structure, resulting in a small overall model parameter count, low computational burden, and easy deployment on low-power terminals. Furthermore, the autoencoder's training is independent of the gesture classification process and can be viewed as a feature learner at the front end of the classifier. This design allows for more flexible integration with other methods, making it highly practical in electromyography pattern recognition and human interaction.

[0015] In summary, this invention combines robustness, practicality, and scalability, and is of great significance for promoting the practical application of electromyography pattern recognition technology. Attached Figure Description

[0016] Figure 1 This is an overall framework diagram of the robust electromyography pattern recognition method of the present invention; Figure 2 This is a schematic diagram of the six gestures involved in the present invention; Figure 3 This is a schematic diagram of the electromyography feature map simulation process from the electrode offset perspective of the present invention; Figure 4 This is a schematic diagram of electromyographic feature simulations for six different occlusion perspectives involved in the present invention. Detailed Implementation

[0017] In this embodiment, a robust electromyography (EMG) pattern recognition method for electrode offset and signal loss involves acquiring surface EMG signals from hand gestures, training a twin autoencoder, and reconstructing abnormal EMG feature data simultaneously affected by electrode offset and signal loss into original EMG data, thereby achieving robust gesture recognition. The overall framework of the proposed method is as follows: Figure 1 As shown, specifically, it includes the following steps: Step 1: Construct the original electromyography feature dataset : use Channel electromyography bracelet data acquisition 10 subjects The surface electromyography (EMG) signals of various hand gestures were analyzed, and a sliding window with a window length of 200 ms and a step size of 80 ms was applied. The amplitude thresholding method was then used to detect the active segment. The dataset of electromyography (EMG) data samples was divided into a training set and a test set. This represents the number of samples in the training set of electromyography (EMG) data. The number of electromyography data samples in the test set represents the number of samples in the test set. In this embodiment of the invention, the signal sampling rate is 650 Hz. =8, =10, =6, hand gestures include: clenching a fist, extending the palm, extending the wrist, flexing the wrist, radial deviation, and ulnar deviation. Illustrations of the six hand gestures are shown below. Figure 2 As shown, the training and test datasets for surface electromyography (EMG) signals were collected in six batches. First, at the initial wearing position of the EMG bracelet, eight measurements were taken for each type of gesture, each held for five seconds with moderate comfort. This data was used as the training set. Then, the EMG bracelet was removed and put back on, and the bracelet was placed back in the initial position, rotated one centimeter to the left, and two centimeters to the right. The offset distance was measured manually with a soft measuring tape, with an error range within two millimeters. In these five batches, three measurements were taken for each type of gesture, each held for five seconds with moderate comfort. This data was used as the test set.

[0018] Extracting from each electromyography data sample In this embodiment of the invention, a time-domain feature is described. =4, the four time-domain features are: Mean Absolute Value (MAV), Waveform Length (WL), Zero Crossings (ZC), and Slope Sign Changes (SSC), and the calculation formulas are as follows: ①Mean Absolute Value (MAV): (1) In equation (1), This is the first activity segment The One sample, This is the number of sample points for each activity segment. It refers to the number of active segments.

[0019] ② Waveform Length (WL) provides information about the waveform complexity within each segment. It is the cumulative length of the waveform over the segment, defined as: (2) In equation (2), The resulting value is the total length of each active segment waveform, which can be represented as a measure of waveform amplitude, frequency, and duration within a single parameter.

[0020] ③ Zero Crossings (ZC): A simple frequency measurement can be obtained by calculating the number of times a waveform crosses zero. A threshold must be included in the zero-crossings calculation. To reduce zero-crossings caused by noise. Given two consecutive samples. and If equation (3) is satisfied, the zero-crossing count is incremented by one.

[0021] and (3) ④ Slope Sign Changes (SSC): Slope sign changes are the number of times the slope sign of a statistical signal changes per unit time. Threshold This is used to reduce noise interference with the slope sign change number. It is achieved by recording three consecutive values ​​of the surface electromyography signal. , and If it satisfies equation (4), then the number of changes in the slope sign is increased by one.

[0022]

[0023] and (4) After the above steps, the original electromyography feature dataset is obtained. ,in, The training set representing the original electromyographic features, and ,in, Representing the One original electromyography feature training sample, Representing the One original electromyography feature training sample The true label, Represents the original electromyographic characteristic test set, and ,in, Representing the One original electromyography characteristic test sample, Representing the Original electromyography characteristic test samples The real label, and .

[0024] Step 2: Construct a high-resolution electromyography feature training set from the electrode offset perspective. : Training set of primitive electromyographic features Each training sample in the dataset undergoes third-order polynomial interpolation along the channel dimension to obtain a high-resolution electromyography feature training set. ,in, Representing the High-resolution electromyography feature training samples, and Then, the high-resolution electromyography feature training set was used. Each training sample in the lateral translation Each pixel is filled with the original pixel that was shifted out, thus obtaining a high-resolution electromyographic feature training set from the electrode offset perspective. In this embodiment, =360, .

[0025] Step 3: Set the number of samples in each batch to... ,in, =64; Training set of original electromyographic features High-resolution electromyographic feature training set with electrode offset perspective The data are divided into several batches; the raw electromyographic feature data in the current batch are denoted as... .

[0026] Step 4: Downsample all high-resolution electromyographic feature training samples from all electrode offset perspectives in the current batch to obtain electromyographic feature training data from all electrode offset perspectives in the current batch. ,in, Represents the first in the current batch Electromyographic feature training samples from each electrode offset perspective, and , represent The real label, and The overall diagram of steps 2 and 4 is as follows: Figure 3 As shown, these steps can effectively simulate the characteristic data after electrode offset.

[0027] Step 5: Electromyographic feature data of electrode offset angle in the current batch. In each training sample, the data of Q channels are randomly set to zero. ; Obtain electromyographic feature data from the masked perspective in the current batch. ,in, The first mask viewpoint in the current batch One electromyographic feature sample, and , represent The real label, and For random The data in each channel is set to zero, including six cases: one channel, two non-adjacent channels, two adjacent channels, three non-adjacent channels, two adjacent channels and one additional non-adjacent channel, three adjacent channels, etc. Figure 4 As shown.

[0028] Step 6, , as well as Each sample in the dataset is transformed into a one-dimensional format and denoted as an electromyographic feature vector at the electrode offset perspective. Electromyographic feature vectors from a masked perspective and the original electromyographic feature vector This allows the input to the Siamese autoencoder network for training.

[0029] Step 7: Construct a Siamese autoencoder network containing two branches. The first branch consists of an input layer, an encoding layer, a hidden representation layer, a decoding layer, and an output layer. The structure of the second branch is completely identical to that of the first branch, and they share weights. The number of neurons in the encoding layer, hidden representation layer, and decoding layer of the Siamese autoencoder network are 512, 128, and 512, respectively.

[0030] Step 8, and The inputs are fed into two branches of the twin autoencoder network, respectively, to obtain the electromyographic feature reconstruction vectors from the electrode offset perspective in the current batch. Electromyography feature reconstruction vector from the mask perspective ; Step 9: Construct the current batch of data according to equations (1), (2), and (3) respectively. and Reconstruction error between , and Reconstruction error between , and Reconstruction error ; (1) (2) (3) In equations (1)-(3), , , Represent , , The The electromyographic feature reconstruction vector of each electromyographic feature sample, the electromyographic feature reconstruction vector from the masked perspective, and the original electromyographic feature vector.

[0031] Step 10: Construct the total loss function for the current batch according to equation (4). To measure the model's ability to reconstruct electromyographic signals affected by electrode offset and signal loss; (4).

[0032] Step 11: Set the learning rate parameter to 0.001 and the weight decay parameter to 0.00001. Using the Nadam optimizer, train the Siamese autoencoder network by inputting different batches of data according to the process of steps 4-10. Calculate the total loss function in different batches and backpropagate to update the network parameters until the total loss converges, thereby obtaining a robust electromyography feature data reconstruction model.

[0033] Step 12: Use a robust electromyography (EMG) feature data reconstruction model on the original EMG feature training set. All samples are reconstructed to obtain a new electromyography (EMG) feature data training set, and an SVM classifier is trained using the new EMG feature data training set to obtain a trained SVM classifier.

[0034] Step 13: Considering different testing scenarios, from the original electromyography feature test set Samples are randomly selected from the data and it is randomly decided whether to perform masking processing as in step 5. Regardless of whether the currently selected sample has been masked, it continues to be input into the robust electromyography feature data reconstruction model for reconstruction. The results are then input into the trained SVM classifier to obtain the predicted label for each test sample.

[0035] Therefore, in step 13, in addition to the test scenario with both interference factors, three other test scenarios were considered: ideal conditions, electrode offset interference only, and signal loss interference only. Specifically, the test data used in the ideal conditions test scenario is the test data collected at the initial position without performing the random occlusion processing in step 5; the test data used in the electrode offset interference only test scenario is the test data collected at all offset positions without performing the random occlusion processing in step 5; the test data used in the signal loss interference only test scenario is the test data collected at the initial position and masked according to step 5; and the test data used in the dual interference factors test scenario is the test data collected at all offset positions and with the random occlusion processing in step 5 performed.

[0036] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0037] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0038] Furthermore, to verify the advanced robustness of this invention, the gesture recognition accuracy of the proposed method was compared with two existing classical methods, and the gesture recognition accuracy was statistically analyzed in four test scenarios. Specifically, the interpolation method recovers the missing signal through linear interpolation between neighboring channels, and the interpolation & data augmentation method, based on the interpolation method, expands the training dataset by 5 times by simulating the data after electrode offset according to steps 2 and 4 of this invention. The final experimental results are shown in Table 1. In all four test scenarios, the gesture recognition accuracy achieved by the proposed method is higher than that of the two comparative methods.

[0039] Table 1. Average gesture recognition accuracy of the three methods in four test scenarios.

[0040] In summary, this invention makes a pioneering exploration to simultaneously solve the problems of electrode offset and signal loss in electromyography (EMG) pattern recognition, and can achieve high-precision EMG gesture recognition even under the superposition of dual interference factors, effectively improving the robustness of EMG pattern recognition systems and significantly outperforming existing methods. The method proposed in this invention is easy to implement and highly portable, providing a new solution to the robustness challenges faced by EMG pattern recognition technology in practical applications.

Claims

1. A robust electromyographic pattern recognition method for electrode offset and signal loss, characterized in that, Includes the following steps: Step 1: Construct the original electromyography feature dataset : use Channel electromyography bracelet data acquisition 10 subjects The surface electromyography (EMG) signals of various hand gestures were analyzed, and sliding window and active segment detection were performed on the EMG signals to obtain... The dataset of electromyography (EMG) data samples was divided into a training set and a test set. This represents the number of samples in the training set of electromyography (EMG) data. The number of electromyography data samples in the test set represents the number of test set samples. ; Extracting from each electromyography data sample The original electromyography feature dataset is obtained by analyzing time-domain features. ,in, The training set representing the original electromyographic features, and ,in, Representing the One original electromyography feature training sample, Representing the Original electromyography feature training samples The true label, Represents the original electromyographic characteristic test set, and ,in, Representing the One original electromyography characteristic test sample Representing the Original electromyography characteristic test samples The real label, and ; Step 2: Construct a high-resolution electromyography feature training set from the electrode offset perspective. : The original electromyographic feature training set Each training sample in the dataset undergoes third-order polynomial interpolation along the channel dimension to obtain a high-resolution electromyography feature training set. ,in, Representing the High-resolution electromyography feature training samples, and Then, the high-resolution electromyography feature training set was used. Each training sample in the lateral translation Each pixel is filled with the original pixel that was shifted out, thus obtaining a high-resolution electromyographic feature training set from the electrode offset perspective. ; Step 3: Set the number of samples in each batch to... The original electromyographic feature training set High-resolution electromyographic feature training set with electrode offset perspective The data are divided into several batches; the raw electromyographic feature data in the current batch are denoted as... ; Step 4: Downsample all high-resolution electromyographic feature training samples from all electrode offset perspectives in the current batch to obtain electromyographic feature training data from the electrode offset perspectives in the current batch. ,in, Represents the first in the current batch Electromyographic feature training samples from each electrode offset perspective, and , represent The real label, and ; Step 5: Electromyographic feature data of electrode offset angle in the current batch. All training samples are randomly zeroed out in Q channels to obtain the electromyography (EMG) feature data from the masked perspective in the current batch. ,in, The first mask viewpoint in the current batch One electromyographic feature sample, and , represent The real label, and ; Step 6, , as well as Each sample in the dataset is transformed into a one-dimensional format and denoted as an electromyographic feature vector at the electrode offset perspective. Electromyographic feature vectors from a masked perspective and the original electromyographic feature vector ; Step 7: Construct a twin autoencoder network, containing two branches, each used for... and The process is performed to obtain the electromyographic feature reconstruction vectors from the electrode offset perspective in the current batch. Electromyography feature reconstruction vector from the mask perspective ; Step 8: Train the Siamese autoencoder network by inputting different batches of data into it, following the process in steps 4-7, and calculate the total loss function for each batch. The network parameters are updated by backpropagation until the total loss converges, thus obtaining a robust electromyography feature data reconstruction model. Step 9: Use a robust electromyography (EMG) feature data reconstruction model on the original EMG feature training set. All samples are reconstructed to obtain a new electromyography feature data training set, and the SVM classifier is trained using the new electromyography feature data training set to obtain a trained SVM classifier. Step 10: From the original electromyography characteristic test set Test samples are randomly selected and input into a robust electromyography feature data reconstruction model for reconstruction. The results are then input into the trained SVM classifier to obtain the predicted labels of the test samples.

2. The method for robust electromyographic pattern recognition in the face of electrode offset and signal loss according to claim 1, characterized in that, The first branch of the twin autoencoder network consists of an input layer, an encoding layer, a hidden representation layer, a decoding layer, and an output layer. The structure of the second branch is completely identical to that of the first branch, and they share the same weights.

3. The electromyographic pattern recognition method robust to electrode offset and signal loss according to claim 1, characterized in that, The total loss function in step 8 Create it using the following steps: Construct the current batch of data according to equations (1), (2), and (3) respectively. and Reconstruction error between , and Reconstruction error between , and Reconstruction error ; (1) (2) (3) In equations (1)-(3), , , Represent , , The Electromyographic feature reconstruction vector of a sample, electromyographic feature reconstruction vector from a masked perspective, and original electromyographic feature vector; Construct the total loss function for the current batch according to equation (4). ; (4)。 4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the electromyography pattern recognition method according to any one of claims 1-3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the electromyography pattern recognition method according to any one of claims 1-3.

Images