Deep Learning for Electromyographic Lower-Limb Motion Signal Classification Using Residual Learning

Electromyographic (EMG) signals have gained popularity for controlling prostheses and exoskeletons, particularly in the field of upper limbs for stroke patients. However, there is a lack of research in the lower limb area, and standardized open-source datasets of lower limb EMG signals, especially recording data of Asian race features, are scarce. Additionally, deep learning algorithms are rarely used for human motion intention recognition based on EMG, especially in the lower limb area. In response to these gaps, we present an open-source benchmark dataset of lower limb EMG with Asian race characteristics and large data volume, the JJ dataset, which includes approximately 13,350 clean EMG segments of 10 gait phases from 15 people. This is the first dataset of its kind to include the nine main muscles of human gait when walking. We used the processed time-domain signal as input and adjusted ResNet-18 as the classification tool. Our research explores and compares multiple key issues in this area, including the comparison of sliding time window method and other preprocessing methods, comparison of time-domain and frequency-domain signal processing effects, cross-subject motion recognition accuracy, and the possibility of using thigh and calf muscles in amputees. Our experiments demonstrate that the adjusted ResNet can achieve significant classification accuracy, with an average accuracy rate of 95.34% for human gait phases. Our research provides a valuable resource for future studies in this area and demonstrates the potential for ResNet as a robust and effective method for lower limb human motion intention pattern recognition.


I. INTRODUCTION
D ISABILITY can result from various diseases, including stroke which has become the leading cause of death in Hong Kong.Survivors of stroke often suffer from varying degrees of disability that impact their independence.Traffic accidents can also cause disability, with the inability to walk as usual having a significant impact on an individual's confidence and independence, while increasing the burden of social care.Robotics and artificial intelligence offer a promising solution to increase the autonomy of people living with disabilities, enabling users to seamlessly interact with robots to complete daily tasks with increased independence [1], [2], [3], [4].
A common approach to transmitting information about the intention of human movement is through electromyography (EMG) signals.EMG measures the electrical activity of muscles in response to nerve stimulation and has been widely used in the identification of user intention to potentially control assistive devices such as smart wheelchairs, exoskeletons, and prosthetic devices over the past few decades [5].The current intensity curve over time, measured by attaching electrodes to the skin surface, is called surface electromyography(sEMG), and is commonly used in medicine to examine nerve and muscle excitation and conduction functions to determine the functional state of peripheral nerves and neurons as well as the muscle itself.This method is a relatively simple way to measure muscle activity.The research and development of sEMG is an active discipline, especially in the upper limb area [6].Researchers have made many good results, such as in human gesture recognition and arm posture estimation, however, as for the lower limb area, there is still a long way to go [7], [8], [9].There are two main challenges.One of the challenges in deep learning-based research is to obtain a reasonably large dataset.To produce a prosthetic knee joint product, it needs to accommodate as many data as possible.Therefore, we need to construct a large dataset which contains the data from various motions of people.There are seldom well-structured and standardized open source datasets in this area.Furthermore, many prosthetic products are produced according to physical condition of the western people.Currently, almost all of the prosthetic knee products are located in Europe.We aim to develop an AI-based prosthetic knee joint for Asian users especially Hong Kong users.Therefore, to construct a large dataset which suits for users in Hong Kong feature is of key importance [10], [11].On the other hand, ResNet is very popular in deep learning.Deeper neural networks are more difficult to train.Microsoft Research Asia(MSRA) present a residual learning framework to ease the training of networks that are substantially deeper than those used previously.They explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions.They provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth [12]

II. RELATED WORK
In the field of sEMG pattern recognition, researchers have primarily focused on upper limb EMG signals, specifically hand gesture recognition.Using the NinaPro database, which consists of hand movement data from 40 intact and 11 amputee subjects, researchers developed a CNN-based system that achieved a 10.18% increase in classification accuracy averaged over five testing sessions compared to the unrecalibrated classifier [13].These results demonstrate the effectiveness of deep learning methods over traditional machine learning methods in the sEMG domain, as has been proposed in other studies [14], [15], [16].
In the lower limb area, EMG signals from elderly individuals have shown that the voltage amplitudes of EMG signals vary widely with age.Thus, researchers have focused on the postural effectiveness of muscular activities in the stability challenges of knee joint movements [17].Additionally, researchers have proposed a locomotion simulation method for lower limb EMG-prosthesis based on OpenGL [18].They used a non-linear extension of the Kalman filter to predict knee and ankle joint kinematics from lower limb muscle activation patterns during overground locomotion.The results demonstrated a high mean r-value for each of the six subjects, ranging from 0.38 to 0.92, with corresponding SNR values between 2.8 dB and 10.8 dB [19].
There are many open source benchmark datasets of sEMG and EEG signals.EEGdenoiseNet is a benchmark EEG dataset that is suited for training and testing DL-based denoising models, as well as for performance comparisons across models.EEGdenoiseNet contains 4514 clean EEG segments, 3400 ocular artifact segments and 5598 muscular artifact segments, allowing users to synthesize contaminated EEG segments with the ground-truth clean EEG [20].Ninapro dataset, containing kinematic and sEMG data from the upper limbs of 27 intact subjects while performing 52 finger, hand and wrist movements of interest.Eventually it makes a huge benchmark dataset [21].
In recent years, There are many deep learning method.scientists use VGGNet and others to explore the BCI data [22], [23], [24], [25].After MSRA introduced Resnet, soon ResNet was also used in the sEMG field [26], [27], [28].However, We regret to find that all these works are done based on the upper limbs, like hand gesture [29], and elbow [30].ResNet shows its strong power to automatically extract the valid feature of EMG instead of a manual way.These wonderful examples of ResNet's works in the upper limb area provide valuable experience for us to carry out work in the lower limb area.

III. SEMG DATASET
One of the major contributions of this article is to provide a new, publicly available, sEMG-based lower limb motion recognition dataset, referred to as the JJ Dataset, which covers 13350 clean EMG segments of 10 gait phases from 15 people.The data acquisition protocol was approved by the Human Research Ethics Committee (HREC) of HKU (HREC Reference Number: EA210538 from 10-01-2022 to 09-01-2025) and informed consent was obtained from all participants.sEMG signals were collected using OpenBCI (up to 16channels, 1000 Hz sampling rate), a high cost-effectiveness board.All data is collected in an indoor experimental environment to ensure its quality.The acquired EMG is amplified and converted into a digital signal and transferred to a PC for subsequent data processing and experiments.Then transfer the data to matlab via LSL(lab streaming layer) [31].The signal acquisition amplifier adopts a mature analog integrated front-end solution with high commonmode rejection ratio and high input impedance, which can acquire weak microvoltage EMG signals and has good EMG acquisition performance.Also compact in size, it is a portable high performance EMG signal amplifier for motion scenarios.Thus, this dataset will have higher competitiveness in practical application scenarios.
To the best of our knowledge, JJ dataset is the largest dataset published utilizing the commercially available OpenBCI, and it strongly fills the gap in this very cutting-edge field and gives researchers a new data platform.It is our sincere hope that JJ dataset could become a very useful tool in the sEMG-based lower limb motion recognition community.
Fifteen healthy Asian volunteers participated in this study.For JJ dataset, we used Ag/Agcl gel-based electrode.This is a non-invasive experiment in which the gel electrode provides good contact between the subject's skin and the electrode, and thus has a significantly higher signal-to-noise ratio than the dry electrode.On the other hand, the gel electrode requires skin scraping and cleaning before and after the experiment, which makes the experiment more tedious than dry electrodes.In fact, the accuracy and robustness of gel electrodes in surface EMG signal acquisition against motion artifacts are better than that of dry electrodes [32].Additionally, while the recommended sampling rate is 700Hz [13].Excellent hardware conditions within the frequency measurement range of the device provide a guarantee for the quality of the JJ dataset.
In this study, we selected nine major muscles on the left leg of the human body for sEMG signal collection, and all locations were calibrated according to clinical medical standards, which makes the JJ dataset more realistic and reliable than other similar datasets.
In this study, We take into account the experience of European SENIAM project to select the muscle [33].The SENIAM project (Surface ElectroMyoGraphy for the Non-Invasive Assessment of Muscles) is a European concerted action in the Biomedical Health and Research Program (BIOMED II) of the European Union.We finally selected nine major muscles on the left leg for the experiment as shown in Fig. 2, three muscles on the calf and six muscles on the thigh.After cleaning the skin by alcohol to remove impurities like oils, sEMG electrodes are located on their approximate location.There are Tibial anterior, 2 Gastrocnemius on the calf, and Rectus femoris, Vastus lateralis, Semitendinosus adductor longus, biceps femoris longus and vastus medialis on the thigh.At the same time, we placed the reference electrode at the knee of the left leg, as close to the bone as possible.Fig. 3 displays classes of gait phase.
To recognize motion intentions, different gait events in a gait circle should be recorded.We define 10 gait events with a sitting event as resting state.These nine events form a gait circle.They are Stand, Sit, Sit leg extension, Stand with leg back, Stand with leg up, Front kick, Back kick, Side leg lift, Squat and Tiptoe stand.Each motion is recorded 89 segments and each time record 2000 data points.This device has a sampling rate at 1000Hz.So, each data segment includes 2 seconds of signal.Usually, the amount of experimental data of sEMG is a bit small, which is because the lower limb sEMG experiments require higher efforts and will be more strenuous than the upper extremity experiments, But our JJ dataset is large enough.It is known that deep learning methods have high requirements for sample data volume, so in this experiment we spent longer time and conducted more than ten times the usual number of repetitions to finally complete the recording of the original JJ dataset.

IV. METHOD A. Data Pre-Processing
In terms of data pre-processing, we employed principal component analysis (PCA) to reduce the dimensionality of the time domain data.PCA is a technique that maps n-dimensional features to k-dimensions, where the new features are orthogonal and also called principal components.By reconstructing k-dimensional features based on the original n-dimensional features, PCA effectively reduces the dimensionality of the data.Specifically, the first new axis is chosen in the direction of the largest variance in the original data, and the subsequent new axes are chosen in the planes orthogonal to the preceding axes that maximize the variance.Through this process, we obtained n such axes, but found that most of the variance was contained in the first k axes, while the remaining axes contained almost zero variance.Therefore, we only retained the first k axes that contained most of the variance, effectively reducing the dimensionality of the data.This is equivalent to keeping only the dimensional features that contain most of the variance and ignoring the dimensional features that contain almost zero variance.
Moreover, we utilized classical methods for data preprocessing, such as Fast Fourier Transform (FFT), which converts the data from the time domain to the frequency domain.FFT is a fast method for calculating the sequence discrete Fourier transform (DFT), which transforms a finite sequence into the frequency domain of the original input sequence.This enables easy analysis or feature extraction in the frequency domain.Specifically, given a finite sequence x (n) [0,M-1] with M points, FFT is able to efficiently calculate the DFT of the sequence.the DFT is In addition, the JJ dataset is labeled with the static marking method to facilitate data processing, and the data are divided into approximately a four to one ratio.Via PCA, we transferred the size of the data to 10 × 10 ×16 (or 9) pictures.These pre-processing methods effectively improve the quality and usability of the data for subsequent analysis and modeling.
The sliding time window method is a commonly used technique for processing sEMG data, and it is better suited for practical application scenarios.Previous research on similar sEMG upper-limb data processing typically uses a time window of 100 to 250ms.However, using a window that is too large can result in the loss of details, as longer time windows may not capture rapid changes or short-term trends in the data.This can lead to data smoothing, blurring the actual variation in the data.Conversely, using a window that is too small may increase noise and instability, as shorter time windows may not  capture trends and patterns in the data and may be subject to random fluctuations.Therefore, it is important to balance these factors and determine the optimal time window size based on the data being studied and the objectives.
In this experiment, after testing different window sizes, a time window of 200ms was selected for data processing, as it achieved good performance.Then we computes timedomain features for each window of data and normalizes the resulting feature vectors.The optimal window size may vary depending on the specific data being analyzed and the objectives of the study.The normalized feature vectors are stored in a four-dimensional matrix, with dimensions corresponding to the number of samples, the number of windows, the number of features, and the number of sEMG channels.The resulting feature vectors are normalized using the normalization function and concatenated into a threedimensional matrix, which represents the feature vectors for a single window of data.Then we store it in the appropriate location within the final matrix, creating a final dataset that can be used for machine learning applications.The dataset was divided into a train set and a test set with a 4:1 ratio using the Scikit-learn library.These considerations will be further discussed in the discussion section.
Overall, using an appropriate time window size is crucial for accurately capturing trends and patterns in sEMG data while minimizing noise and instability.The sliding time window method is a practical and effective technique for processing sEMG data in various application scenarios.

B. Classic sEMG Classification
In previous machine learning studies, the most commonly used tradtional classifiers for EMG signals were Non-linear Logistic Regression (NLR), Multi-Layer Perceptron (MLP), and Support Vector Machine(SVM).NLR is a linear and binary supervised classification algorithm that calculates the class membership and possibility.MLP is a type of Artificial Neural Network (ANN) that is also a supervised algorithm.The network architecture consists of an input layer, one or more hidden layers, and an output layer with one neuron for each class to classify.SVM is a linear and binary supervised classification algorithm that considers only dichotomous distinction between two classes [34].
However, the majority of research results demonstrate that SVM achieves the highest values for both classification performance and number of classification parameters, followed by MLP, and then by NLR [35], [36].This suggests that SVM may be the most effective classifier for EMG signal analysis, as it is able to accurately distinguish between different classes while minimizing the number of parameters required for classification.Nonetheless, the selection of a classifier should depend on the specific research objectives, as well as the characteristics of the dataset being analyzed.Additionally, other classification algorithms may also be useful for EMG signal analysis, and further research is needed to explore their potential.
Feature extraction is a crucial step in SVM for transforming raw data into feature vectors that can be used by classifiers to improve their accuracy [37].In EMG signal processing, the appropriate selection of features is critical for achieving   4. Sliding time window method in this experiment.The whole time domain sEMG data was divided into multiple samples, each consisting of 9 channels and a size of 2000ms.In this study, the sliding time window approach was applied to each sample using a window size of 200ms and a step of 50ms.Consequently, every sample was segmented into 37 parts, each with a size of 200ms and 9 channels.From each part, five time domain features were extracted, and the resulting data constituted the final dataset for analysis.accurate classification results.This selection should be based on specific application scenarios and signal characteristics, with an appropriate number of features selected to avoid overfitting or decreased accuracy.In this study, five time-domain features (RMS, MAV, VAR, ZC and WAMP) and two frequency domain features (Peak Frequency and PSD) were selected.To ensure stable and accurate division of the dataset while running the model, a random seed was set during the division process.
To optimize the accuracy of the SVM model, the two parameters, γ and the penalty parameter C, are significant for the model's performance.γ is a special parameter in the RBF kernel function that determines the distribution of the data after mapping it to the new feature space.The number of support vectors is affected by the value of γ , with fewer support vectors for larger values of γ and more support vectors for smaller values of γ .The number of support vectors affects the speed of training and prediction.The penalty parameter C determines the tolerance to errors, with higher values of C being less tolerant of errors and more prone to overfitting, while smaller values of C may result in underfitting.In addition, the number of iterations needs to be considered due to the large amount of data, as higher iteration Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.In this study, we chose 10 subjects as a dataset and iterated through different values of C, γ , and kernel functions using the parameter grid search method to obtain the optimal parameters for training the model.The optimal parameters were found to be C=10, γ =0.1, and kernel='rbf'.
After comparing the parameters and kernel functions of SVM for our 10 subjects, we achieved an average accuracy of 72.53%.This demonstrates the effectiveness of our SVM model in accurately classifying EMG signals, and the importance of selecting appropriate parameters and kernel functions to optimize the model's performance.

C. Deep Learning Method
1) Adjusted Resnet: Deep learning is a machine learning technique that has been widely used in various domains, such as computer vision, speech recognition, and medical image analysis.The most commonly used deep learning method is based on artificial neural networks (ANN), especially convolutional neural networks (CNN).One of the advantages of CNN is that it requires less pre-processing compared to other recognition algorithms.
In this study, we implemented ResNet-18, which is a deep learning model similar to CNN but has a different structure.ResNet was first proposed by Kaiming He in 2015 [12], and it has been used in various domains, such as computer vision and speech recognition.
One of the main challenges with deep neural networks is the problem of accuracy degradation which occurs when the accuracy of a model becomes saturated and then starts to decrease rapidly without overfitting.This problem is common in deep networks because the gradient becomes smaller and smaller during backpropagation of the loss function, hindering the process of parameter update.
ResNet addresses this problem by introducing a residual structure that connects the layer and two layers before it.This allows the parameters to transfer to the front network successfully.The residual structure consists of a shortcut connection that bypasses one or more layers and connects the input directly to the output of the residual block.This allows the network to learn the residual mapping instead of the full mapping, which makes it easier to optimize the network and reduces the problem of accuracy degradation.
We used the Adam optimizer and Cross-entropy loss function for training our adjusted Resnet model.The accuracy of each training was recorded, and we took the average accuracy of 20 trainings as the final accuracy.To make the model more adaptable to our JJ dataset, we made several improvements, as shown in Fig. 6.Firstly, we reduced the number of parameters in the model by using smaller kernels and a deeper network.This can prevent overfitting and improve the generalization ability of the model.Secondly, we made the model capable of handling smaller input image sizes, which can speed up training and processing of smaller images.Lastly, we believe that by using a deeper network, the model can extract more features and achieve higher accuracy on specific datasets.
In our experiments, the ResNet-18 model with these improvements achieved an accuracy of 95.34% on our JJ dataset.This demonstrates the effectiveness of our approach in adapting the ResNet-18 model to EMG signal classification, and highlights the importance of making appropriate modifications to deep learning models to fit specific datasets and tasks.
2) VGG-16: VGGnet is a convolutional neural network algorithm that was first proposed in the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) in 2014 by the University of Oxford [38].It has shown excellent performance in object detection and classification tasks.VGG 16 is the most common variation of VGG, which contains five stages of convolution.Each convolutional stage consists of two or three convolutional layers and one pooling layer.The number of convolutional kernels increases stage by stage, but the size of each convolutional kernel is always 3 × 3.At the end of VGG-16, there are two fully-connected layers to output the result.
In this study, we used a modified version of VGG 16 for EMG signal classification.Since VGG-16 contains many convolutional and pooling layers, we increased the size of a single channel of input data by reshaping it from 10 × 10×16 to 40 × 40×1.We also reduced the number of convolutional kernels and removed the last three pooling layers of VGG-16.
Experiments were carried out with VGG-16 using our dataset as input.We used Adam as the optimizer and Crossentropy as the loss function.We performed 20 trainings with 100 epochs each.The accuracy of each training was recorded, and took the average accuracy of the 20 trainings as the final accuracy.The training result is 80.08% which is lower than that achieved with ResNet in our study.

V. DISCUSSION
Unless otherwise specified, the overall recognition rates mentioned in this article are cross-subject ones.Cross-subject research itself is a very important research topic, as it has obvious importance in promoting the generalization of research in this area of study.In the vast majority of EEG data, cross-subject research is a huge challenge, while sEMG data is relatively more concise and clean, which also poses some difficulties in similar upper-limb signal processing work.One of the primary difficulties is the differences in signal amplitude and morphology among different subjects, which can affect the accuracy of classification and recognition.Another challenge is the nonlinearity and non-stationarity of EMG signals, with amplitude and frequency changing over time.This variability may impact the model's generalization ability.Usually, similar work involves identifying each individual separately, and finally integrating all recognition rates at the level of statistical methods to obtain the final overall recognition rate.
In this study, we directly amalgamated the dataset for processing and achieved favorable results.This demonstrates that analogous data can be directly merged and utilized in future research.Additionally, we conducted a cross-subject recognition rate verification using the Leave-one-out (LOO) method to assess the performance of ResNet.Subjects 1 to 14 were used as the training set, while subject 15 was used as the test set.The final recognition rate obtained was 68.01%, which is of considerable significance.
In this confusion matrix, the class labels are denoted as C 1 , C 2 , . . .C 10 .Since there are 10 classes of gesture.The vertical axis of the confusion matrix is utilized to signify the true class of the input sample, while the horizontal axis delineates the predicted class generated by our model.The count of each prediction scenario are documented with respect to the corresponding column i and row j, and can represented as follows, where C i, j represents the number of predictions that true class is C i and predicted class is C j , y t stands for the true class for each sample and y p stands for the predicted class.N is the total number of input sample.The prediction outcomes of the cross-subject task were presented in a confusion matrix, which illustrated the distribution of predicted and actual labels.As depicted in Figure 7, sitting leg extension and sit were frequently mispredicted as front kick and sit.This may be attributed to the similarity of force muscles between sitting and standing gestures and common features among different subjects.Resolving the confusion between the sitting and standing states of subjects is a promising area for future research.Further investigation is needed to enhance crosssubject accuracy and improve the robustness of the proposed model.
To verify the accuracy of ResNet-18, we also made a comparison with VGG-16, CNN, ANN and SVM by using sliding time window method.Table I show that in all cases ResNet 18 is better than models.With the sliding time window method,there will be a certain decrease in recognition rate, but it is not very big.ResNet-18 can still reach 91.65% recognition rate, significantly surpassing the effectiveness of traditional machine learning, proved its excellent performance in this more practical situation.At the same time, this also indicates that there is still room for improvement in this method, and in practical applications, such an effect will definitely be showcased.
The main disadvantage of this low-cost method is reflected in the performance of frequency domain data.The sampling coverage of high-frequency information is not sufficient, Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.resulting in the loss of some information, which is clearly reflected in the recognition effect of frequency data, only about 70%.Therefore, the time-domain performance of this dataset is significantly better than the frequency-domain performance, and in the future, it will be more inclined to optimize methods suitable for time-domain data.
In addition, we simulated a hypothetical scenario that only considers the impact of thigh data on recognition rate.Assuming that the disabled person is missing their calf, so they can only show the thigh muscles raw data, we test the recognition rate difference between original recognition and using only the 6-channel JJ data of the thigh muscles mentioned above.In general, the recognition rate using only thigh data is still relatively stable, but the recognition effect of the LOO method cross-subject is poor, and there are significant fluctuations depending on the individual, with recognition rates ranging from 30 to 70 percent.This result demonstrates the feasibility of using only thigh muscle data, providing samples for disabled individuals, and also indicating that cross-subject testing remains a significant challenge To train this model, hyperparameters should be adjust times to times.Or the network will easily get overfitted.The main hyperparameters that should be adjusted are learning rate, decay, epoch, batch size.First parameter that should be adjusted is learning rate.By reducing it slowly, the loss function will converge.The exist of decay determines the loss function will not vibrate too much when training.The batch size determines how much data is trained at the same time.Larger batch size may reduce the generalization but smaller batch size will cause slow convergence speed.Epoch can be adjusted base on the record of training process.There is even a reduce overfitting way by stopping training in advance.
This study uses 3D input.Compared to 2D input, Kernel detects the relationship between channels, which may not have too much significance, so we use 3D input to get better result, but the accuracy of Resnet is not far beyond the Other deep learning methods.This EMG pattern maybe mot suit the Resnet enough.Open-sourced datasets like our JJ dataset are valuable resource which allow direct comparison of different algorithms.So, it is very important to continue to enrich the JJ dataset.Limitations of this study include but are not limited to post hoc, offline data analysis; therefore, we have not yet demonstrated the ability to recognize and classify gait phases in real time.
In future work, we will extend the JJ dataset to include gait phases of lower limb amputees to improve human motion recognition in this population.However, this endeavor encounters several obstacles.The results of the LOO method testing are not perfect enough, Further exploration and optimization is still needed.EMG signal collection in amputees is compromised by poor muscle integrity, atrophy, and the prevalence of noise and signal artifacts, compounded by inconsistencies in electrode positioning [39].These issues are non-trivial as they directly affect the fidelity of muscle activation data, which is essential for accurate pattern recognition.Moreover, the infant state of EMG-based active prosthetic and exoskeleton technology development presents additional challenges.Current designs do not fully address the complex biomechanics of natural gait, often resulting in suboptimal user experience and adaptation.There is still a long way to go to help amputees, and we still need to continue working hard to solve these problems.

VI. CONCLUSION
This work presents three main contributions.First, We have built and open-sourced a very competitive well-structured and standardized benchmark dataset, JJ dataset for lower limb sEMG human motion recognition.Second, we also compared the advantages and disadvantages of various methods and found that time-domain processing was more effective than frequency-domain processing.The successful cross-subject experiment of the leave-one-out method further demonstrated the widespread practicality of our approach.We used data from thigh channels to simulate and explore the effect on patients with knee-below amputation.Our results showed that the sliding time window method improved the reaction speed of dynamic prosthetic knee, which could be adapted for use in the control of prosthetic knee and other assistive devices.Third, we adjust the ResNet architecture, which achieves an average accuracy rate of 95.34% for 10 gait phases on average and outperforms other sEMG-based classifiers.Our results suggests that ResNet could become a robust and effective method for lower limb human motion intention pattern recognition, opening exciting new avenues for research in this field.

Fig. 1 .
Fig. 1.A sample EMG data of channel 3 muscle from four different movement.Channel A represents the waveform corresponding to the back kick, Channel B represents the waveform corresponding to the squat, Channel C represents the waveform corresponding to the tip toe stand, Channel D represents the waveform corresponding to the sit leg extension.

Fig. 3 .
Fig. 3. 10 different gait phases, they are Stand, Sit, Sit leg extension, Stand with leg back, Stand with leg up, Front kick, Back kick, Side leg lift, Squat and Tiptoe stand.

Fig.
Fig.4.Sliding time window method in this experiment.The whole time domain sEMG data was divided into multiple samples, each consisting of 9 channels and a size of 2000ms.In this study, the sliding time window approach was applied to each sample using a window size of 200ms and a step of 50ms.Consequently, every sample was segmented into 37 parts, each with a size of 200ms and 9 channels.From each part, five time domain features were extracted, and the resulting data constituted the final dataset for analysis.
. However, up to now seldom papers use ResNet to do pattern recognition, classify the EMG signals.So this research will verify the accuracy of ResNet in EMG pattern recognition.Our contributions are summarized as follows: This is the pioneering first application of ResNet to the EMG signal domain for human lower limb motion intention recognition.This is the first experimental validation of improved ResNet in this field, and our experiments show that ResNet has achieved success and it shows the state-of-the-art classification accuracy on the JJ dataset, which is 95.34% for 10 gait phases on average.
1)We have built and open-sourced a very competitive wellstructured and standardized benchmark dataset with three major advantages, which is called JJ dataset.First, there are seldom publicly available datasets that record emg signals of human lower limb movements so far, it strongly fills the gap.Second, the amount of data in the JJ dataset is very sufficient, which covers 13350 clean EMG segments of 10 gait phases from 15 people.In addition the JJ dataset records the EMG data features of East Asian ethnicity, which is different from common public datasets.2) We discuss many foundational issues in the lower limb human motion intention pattern recognition area.the comparison of sliding time window method and other preprocessing methods, comparison of time-domain and frequency-domain signal processing effects, cross-subject motion recognition accuracy and the possibility of using thigh and calf muscles in amputees.Thus we can establish a better method of the lower limb human motion intention pattern recognition in the future work.3)

TABLE I PERFORMANCE
COMPARISON BY USING SLIDING TIME WINDOW METHOD Fig. 8. Thigh sEMG data performance testing.