A method for recognizing gait phase and velocity based on electromyography and deep learning

Abstract

This invention discloses a method for recognizing gait phase and speed based on electromyography (EMG) and deep learning. By comprehensively applying EMG signal data and plantar pressure data, this invention can accurately capture characteristic changes at different gait speeds. Through data preprocessing and PCA dimensionality reduction, it highlights the characteristic differences of EMG signal data at different gait speeds, providing a solid foundation for personalized gait analysis and facilitating a deeper understanding and analysis of each individual's gait characteristics at different gait speeds. The use of a CNN-LSTM model significantly improves the accuracy and precision of gait recognition, contributing to a more comprehensive understanding and assessment of muscle activity. This invention offers new possibilities for personalized rehabilitation treatment and exercise assessment. Through accurate identification of EMG signal data at different gait speeds, it can provide more personalized and precise data support for rehabilitation treatment and exercise assessment, promoting the personalization and refinement of rehabilitation treatment and exercise training.

Description

Technical Field

[0001] This invention relates to gait recognition technology, specifically to a method for recognizing walking gait phase and speed based on electromyography and deep learning. Background Technology

[0002] Surface electromyography (sEMG) signals, as biosignals, are widely used in gait recognition. Researchers have made continuous efforts in the field of gait analysis, utilizing sEMG signals to accurately identify gait phases and applying them to rehabilitation therapy and assistive exoskeletons. Currently, many related studies focus on the application of sEMG in gait phase recognition, but research on gait speed recognition is relatively limited. This restricts the effective integration of these two aspects, and some problems remain to be solved.

[0003] Traditional methods typically focus on gait phase identification, failing to fully integrate it with gait velocity identification. Gait phase and velocity are closely related in walking, but existing methods have not effectively utilized this correlation. This separate approach limits a complete understanding of overall gait characteristics. The lack of methods that simultaneously identify phase and velocity restricts the ability to identify real-world walking scenarios and fails to provide a comprehensive description of gait features.

[0004] Therefore, current research on gait phase and velocity recognition based on sEMG faces a significant challenge: how to effectively integrate the characteristics of electromyographic signals to achieve joint recognition of phase and velocity, thereby capturing and describing gait features more comprehensively. This topic has important theoretical and applied significance, and is of great value for developing more accurate and comprehensive gait recognition technology and for its application in rehabilitation and mobility aids. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method for recognizing gait phase and speed based on electromyography (EMG) and deep learning. By deeply analyzing the spatiotemporal characteristics of EMG signals, this invention achieves accurate recognition of gait phase and speed. Furthermore, it utilizes plantar pressure data provided by plantar pressure insoles for phase calibration, providing more accurate assistance for recognizing gait phase and speed from EMG signals.

[0006] The method for recognizing walking gait phase and velocity based on electromyography and deep learning of the present invention includes the following steps:

[0007] 1) A multi-channel electromyography (EMG) acquisition device was fixed to one leg of the subject, with each channel corresponding to the acquisition of EMG signals from one muscle; at the same time, a plantar pressure insole was fixed inside the shoe and worn on the foot on the same side as the foot with the multi-channel EMG acquisition device fixed. The plantar pressure insole was equipped with multiple force-sensitive resistors (FSR) sensors, each of which acquired plantar pressure at one location; the multi-channel EMG acquisition device and the multiple force-sensitive resistors were connected to the computer wirelessly.

[0008] 2) Subjects walked at a set gait speed. Each channel of the multi-channel electromyography (EMG) acquisition device collected EMG signals from the corresponding muscles and transmitted them to a computer to obtain raw EMG signal data. Simultaneously, multiple force-sensitive resistors collected plantar pressure data and transmitted it to a computer to obtain raw plantar pressure data. The multi-channel EMG acquisition device and multiple force-sensitive resistors were strictly synchronized in time under computer control. Each set of raw plantar pressure data corresponded to a set of raw EMG signal data that was completely synchronized in time, meaning that the synchronized raw plantar pressure data and the raw EMG signal data were in one-to-one correspondence. The raw EMG signal data were labeled with gait speed.

[0009] 3) The computer preprocesses the raw electromyography (EMG) signal data and raw plantar pressure data to obtain preprocessed EMG signal data and preprocessed plantar pressure data respectively; and uses a sliding window to perform segmented time-domain and frequency-domain feature extraction on the preprocessed EMG signal data to obtain EMG signal feature data, which correspond one-to-one with the group plantar pressure data.

[0010] 4) Plantar pressure data reflects the pressure distribution on the sole of the foot during gait. By preprocessing the plantar pressure data, a complete gait cycle is divided into multiple different gait phases, and the gait phase corresponding to each set of preprocessed plantar pressure data is obtained. Furthermore, by establishing a one-to-one correspondence between the synchronized plantar pressure data and the electromyographic (EMG) signal feature data, a corresponding gait phase label is attached to each set of corresponding EMG signal feature data. The EMG signal feature data obtains the corresponding gait phase, resulting in EMG signal feature data with gait speed and gait phase labels.

[0011] 5) Change the set gait speed and repeat steps 2) to 4) until all set gait speeds are completed, and obtain electromyographic signal feature data with gait speed labels and phase labels under all set gait speeds to form a dataset; divide the dataset into training set and test set.

[0012] 6) Construct a Convolutional Long Short-Term Memory (CNN-LSTM) model:

[0013] a) Construct a convolutional neural network (CNN) with electromyographic signal feature data as input and gait speed as output;

[0014] b) Divide the output of the convolutional neural network into different gait speed subsets according to the set gait speed. Each gait speed subset corresponds to the gait speed set in steps 2) to 5). The feature emphasis of muscle activity differs under different gait speeds. Perform principal component analysis (PCA) on each gait speed subset. The difference in feature emphasis is addressed by performing corresponding PCA dimensionality reduction on the data features under different gait speed subsets to obtain the dimensionality-reduced gait data features.

[0015] c) Construct a Long Short-Term Memory (LSTM) network, with the dimensionality-reduced gait data features as input and the gait phase as output;

[0016] d) Train the CNN-LSTM model using the training set and test the CNN-LSTM model using the test set to obtain the trained CNN-LSTM model;

[0017] 7) The multi-channel electromyography (EMG) acquisition device is fixed on the subject's leg on the same side. The multi-channel EMG acquisition device acquires EMG signals and transmits them to the computer to obtain raw EMG signal data. The computer preprocesses the raw EMG signal data and uses a sliding window to extract segmented time-domain and frequency-domain features to obtain EMG signal feature data. The EMG signal feature data is then input into the trained CNN-LSTM model to obtain gait speed and gait phase.

[0018] In step 1), the multi-channel electromyography (EMG) acquisition device has M channels, where M is a natural number ≥ 4; and N force-sensitive resistor sensors have N, where N is a natural number ≥ 3. The multi-channel EMG acquisition device acquires EMG signals from the muscles of a single leg, which include the tibialis anterior, medial gastrocnemius, vastus lateralis, vastus medialis, rectus femoris, biceps femoris, soleus, lateral gastrocnemius, and semitendinosus.

[0019] In step 2), each set of raw electromyographic signal data includes M raw electromyographic signal data, and each set of raw plantar pressure data includes N raw plantar pressure data.

[0020] In step 3), the preprocessing of the raw electromyography (EMG) signal data includes: first, applying high-pass and low-pass linear-phase finite-impulse response (FIR) filters to the raw EMG signal data, followed by full-wave correction and Butterworth low-pass filtering to obtain preprocessed EMG signal data; the preprocessing of the raw pressure data includes: sequentially performing denoising, filtering, and interpolation preprocessing operations on the raw plantar pressure data to obtain preprocessed plantar pressure data; data preprocessing ensures data quality and accuracy. A sliding window is used to extract segmented time-domain and frequency-domain features from the preprocessed EMG signal data. A fixed window size with data overlap is employed, which helps extract rich time-domain and frequency-domain features. Each sliding step is one frame.

[0021] In step 5), the ratio of the training set to the velocity set is 70-80%: 30-20%.

[0022] In step 7)a), the convolutional neural network includes: an input layer, a sequence folding layer, a convolutional layer, a rectified linear unit (ReLU) activation function, a fully connected layer, and a softmax function; wherein, the input layer receives electromyographic signal feature data; the sequence folding layer converts the input data into a sequence format; the convolutional layer uses convolution kernels to perform convolution operations on the sequence, and the output of the convolutional layer is input to the rectified linear unit activation function, which introduces nonlinearity; the fully connected layer processes the output, and then the softmax function is used to obtain the gait speed.

[0023] In step 7)b), different principal component features are extracted based on the electromyography data under different gait velocities. Only important features are retained, resulting in good recognition performance. The difference in feature emphasis is addressed by performing corresponding PCA dimensionality reduction on the data features under different gait velocities. The processed data is then used for gait phase recognition via LSTM.

[0024] In step 7)c), the Long Short-Term Memory network includes: a sequence defolding layer, a flattening layer, an LSTM layer, a fully connected layer, and a softmax function. The sequence defolding layer performs data format conversion operations, transforming the dimensionality-reduced gait data features into a format acceptable to the LSTM layer, resulting in defolded data. The flattening layer flattens the defolded data to obtain flattened data. The LSTM layer extracts time-series features from the flattened data to obtain temporal feature data. The fully connected layer integrates and synthesizes the temporal feature data output from the LSTM layer, and finally, the gait phase is obtained through the softmax function.

[0025] Advantages of this invention:

[0026] (1) By comprehensively applying electromyography and plantar pressure data, this invention can accurately capture feature changes under different gait phases and speeds; by using the CNN-LSTM structure, it can achieve accurate classification of gait phase and speed, providing strong support for the comprehensive recognition of different gait features;

[0027] (2) Through data preprocessing and PCA dimensionality reduction, the data were refined to highlight the characteristic differences of electromyographic signal data at different gait speeds; this difference analysis provides a solid foundation for personalized gait analysis and helps to understand and analyze the gait characteristics of each individual at different gait speeds more deeply.

[0028] (3) By adopting the CNN-LSTM model and combining the ability of convolutional neural networks to extract spatial features and long short-term memory networks to learn temporal information, the accuracy and precision of gait recognition are significantly improved, which helps to understand and evaluate muscle activity status more comprehensively. (4) It provides new possibilities for personalized rehabilitation treatment and exercise assessment. By accurately identifying electromyographic signal data under different chronotropic speeds, it can provide more personalized and accurate data support for the field of rehabilitation treatment and exercise assessment, and promote the personalization and refinement of rehabilitation treatment and exercise training. Attached Figure Description

[0029] Figure 1 The flowchart shows the walking gait phase and velocity recognition method based on electromyography and deep learning of the present invention.

[0030] Figure 2 This is a flowchart illustrating the construction of a CNN-LSTM model for the walking gait phase and velocity recognition method based on electromyography and deep learning according to the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0032] This embodiment presents a walking gait phase and velocity recognition method based on electromyography and deep learning, such as... Figure 1 As shown, it includes the following steps:

[0033] 1) An eight-channel electromyography (EMG) acquisition device was fixed to one leg of the subject. Each channel corresponds to the acquisition of EMG signals from one muscle. The muscles of the single leg include the vastus lateralis, vastus medialis, rectus femoris, biceps femoris, tibialis anterior, soleus, gastrocnemius medialis, and semitendinosus. At the same time, a plantar pressure insole was fixed inside the shoe and worn on the same foot as the foot on which the multi-channel EMG acquisition device was fixed. The plantar pressure insole was equipped with eight force-sensitive resistors (FSRs). Each force-sensitive resistor sensor acquired plantar pressure at one location. Three of the force-sensitive resistors were used, located at the toes, the first metatarsal head of the forefoot, and the heel, respectively. The eight-channel EMG acquisition device and the three force-sensitive resistors were connected to the computer wirelessly.

[0034] 2) Participants walked in a straight line on a treadmill at set gait speeds: low (2 km / h), medium (3.5 km / h), and high (5 km / h). Each channel of the multi-channel electromyography (EMG) acquisition device collected EMG signals from the corresponding muscles at a sampling rate of 1000 Hz and transmitted them to a computer to obtain raw EMG signal data. Simultaneously, multiple force-sensitive resistors collected plantar pressure data at a sampling rate of 1000 Hz and transmitted it to a computer to obtain raw plantar pressure data. The multi-channel EMG acquisition device and the multiple force-sensitive resistors were controlled by the computer to collect data in strict temporal synchronization. Each set of raw EMG signal data included eight raw EMG signal data points, and each set of raw plantar pressure data included three raw plantar pressure data points. Each set of raw plantar pressure data corresponded to a set of raw EMG signal data that was completely synchronized in time; that is, the synchronized raw plantar pressure data and the raw EMG signal data corresponded one-to-one. The raw EMG signal data were labeled with gait speed.

[0035] 3) The computer preprocesses the raw electromyography (EMG) signal data: First, the raw EMG signal data is processed by high-pass (cutoff frequency 20Hz) and low-pass (cutoff frequency 450Hz) linear-phase finite-impulse response (FIR) filters, then full-wave correction and Butterworth low-pass (cutoff frequency 5Hz) filters are applied to extract the envelope and obtain the filtered EMG signal data. Preprocessing of the raw pressure data includes: denoising, filtering, and interpolation preprocessing operations are performed sequentially on the raw plantar pressure data to obtain the preprocessed plantar pressure data. Data preprocessing ensures data quality and accuracy.

[0036] A sliding window method is used to extract time-domain and frequency-domain features from the data: the sliding step size is 1 frame, and the window length is set to 20. This means that the data is processed by sliding a window of length 20, moving one data point at a time. This method can extract time-domain and frequency-domain data features under different windows for subsequent analysis and processing to obtain electromyographic signal feature data. The electromyographic signal feature data is labeled with gait speed, and the electromyographic signal feature data corresponds one-to-one with the plantar pressure data. The features extracted in this embodiment are root mean square (RMS), mean absolute value (MAV), variance (Var), median frequency (MF), zero-crossing rate (ZC), power spectral density (PSD), waveform length (WL), and mean energy (ME).

[0037] 4) Plantar pressure data reflects the pressure distribution on the sole of the foot during gait. By preprocessing the plantar pressure data, a complete gait cycle is divided into multiple different gait phases, yielding the gait phase corresponding to each set of preprocessed plantar pressure data:

[0038] In this embodiment, the labels for electromyographic signal characteristic data are distinguished based on the changes in the three FSRs of the plantar pressure insole. According to the pressure change curve of each FSR in the gait cycle, three representative positions are selected: the big toe, the first metatarsal head of the forefoot, and the heel to set the FSRs to divide the gait phase. These three support points bear the body weight sequentially during walking, providing a basis for accurate division of gait phases. Their arrangement conforms to the physiological structure of the human foot and better reflects the plantar pressure distribution. The gait phase is divided according to the ground contact force (GCF) state of these three support points. When the GCF value is "0", it means that the FSR is not under pressure, the support point is off the ground, and the foot is in an OFF-GROUND state. When the GCF value is "1", it means that the FSR is under maximum pressure, the support point is on the ground, and the foot is in an ON-GROUND state. Based on the GCF state of these three positions, a complete gait cycle is divided into five gait phases: early stance (first phase), late stance (second phase), early swing (third phase), mid swing (fourth phase), and late swing (fifth phase).

[0039] Furthermore, by establishing a one-to-one correspondence between synchronized plantar pressure data and electromyographic signal feature data, a corresponding gait phase label is attached to each set of corresponding electromyographic signal feature data, and the corresponding gait phase is obtained from the electromyographic signal feature data, resulting in electromyographic signal feature data with gait speed label and gait phase label.

[0040] 5) Change the set gait speed and repeat steps 2) to 4) until all set low (2 km / h), medium (3.5 km / h), and high (5 km / h) gait speeds are completed. Obtain electromyographic signal feature data with gait speed and phase labels for all set gait speeds to form a dataset. For each subject's electromyographic signal feature data, take a single gait cycle as the smallest unit, extract 70% of the gait cycle data as the training set, and the remaining 30% as the test set. Divide the dataset into training set and test set.

[0041] 6) Construct a Convolutional Long Short-Term Memory (CNN-LSTM) model, such as... Figure 2 As shown:

[0042] a) Convolutional Neural Network (CNN) is constructed, with electromyography (EMG) signal feature data as input and gait speed as output; 8 features are extracted from each channel, and the final dataset has a sequence data dimension of 64; the CNN includes: input layer, sequence folding layer, convolutional layer, and Rectified Linear Unit (ReLU).

[0043] The system consists of an activation function, a fully connected layer, and a softmax function. The input layer receives electromyographic data; the sequence folding layer converts the input data into a sequence format; the convolutional layer performs convolution operations on the sequence using convolutional kernels, and the output of the convolutional layer is input to the ReLU activation function, which introduces nonlinearity; the fully connected layer processes the output, and then the gait speed is obtained through the softmax function.

[0044] b) Divide the output of the convolutional neural network into different gait speed subsets according to the set gait speed. Each gait speed subset corresponds to the low speed (2KM / H), medium speed (3.5KM / h), and high speed (5KM / H) set in steps 2) to 5). The feature emphasis of muscle activity differs under different gait speeds. Principal component analysis (PCA) is performed on each gait speed subset. The difference in feature emphasis is addressed by performing corresponding PCA dimensionality reduction processing on the data features under different gait speed subsets to obtain the dimensionality-reduced gait data features.

[0045] c) Construct a Long Short-Term Memory (LSTM) network, with the dimensionality-reduced gait data features as input and the gait phase as output;

[0046] Long Short-Term Memory (LSTM) networks consist of: a sequence defolding layer, a flattening layer, an LSTM layer, a fully connected layer, and a softmax function. The sequence defolding layer performs data format conversion, transforming the dimensionality-reduced gait data features into a format acceptable to the LSTM layer, resulting in defolded data. The flattening layer flattens the defolded data to obtain flattened data. The LSTM layer then extracts time-series features from the flattened data, yielding temporal feature data. The fully connected layer integrates and synthesizes the temporal feature data output from the LSTM layer, and finally, the softmax function is used to obtain the gait phase.

[0047] d) The CNN-LSTM model was trained using the training set and tested using the test set to obtain the trained CNN-LSTM model; the accuracy of gait phase classification was verified to be 96.86%, and the accuracy of gait speed recognition was 87.38%.

[0048] 7) The multi-channel electromyography (EMG) acquisition device is fixed on the subject's leg on the same side. The multi-channel EMG acquisition device acquires EMG signals and transmits them to the computer to obtain raw EMG signal data. The computer preprocesses the raw EMG signal data and uses a sliding window to perform segmented time-domain and frequency-domain feature extraction to obtain EMG signal feature data. The EMG signal feature data is then input into the trained CNN-LSTM model to obtain gait speed and gait phase.

[0049] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A method for recognizing walking gait phase and velocity based on electromyography and deep learning, characterized in that, The walking gait phase and velocity recognition method includes the following steps: 1) A multi-channel electromyography (EMG) acquisition device was fixed to one leg of the subject, with each channel corresponding to the acquisition of EMG signals from one muscle; at the same time, a plantar pressure insole was fixed inside the shoe and worn on the foot on the same side as the foot on which the multi-channel EMG acquisition device was fixed. The plantar pressure insole was equipped with multiple force-sensitive resistor sensors, each of which acquired plantar pressure at one location; the multi-channel EMG acquisition device and the multiple force-sensitive resistor sensors were connected to the computer wirelessly. 2) Subjects walked at a set gait speed. Each channel of the multi-channel electromyography (EMG) acquisition device collected EMG signals from the corresponding muscles and transmitted them to a computer to obtain raw EMG signal data. Simultaneously, multiple force-sensitive resistors collected plantar pressure data and transmitted it to a computer to obtain raw plantar pressure data. The multi-channel EMG acquisition device and multiple force-sensitive resistors were strictly synchronized in time under computer control. Each set of raw plantar pressure data corresponded to a set of raw EMG signal data that was completely synchronized in time, meaning that there was a one-to-one correspondence between the synchronized raw plantar pressure data and the raw EMG signal data. The raw EMG signal data was labeled with gait speed. 3) The computer preprocesses the raw electromyography (EMG) signal data and raw plantar pressure data to obtain preprocessed EMG signal data and preprocessed plantar pressure data respectively; and uses a sliding window to perform segmented time-domain and frequency-domain feature extraction on the preprocessed EMG signal data to obtain EMG signal feature data, which correspond one-to-one with the group plantar pressure data. 4) Plantar pressure data reflects the pressure distribution on the sole of the foot during gait. By preprocessing the plantar pressure data, a complete gait cycle is divided into multiple different gait phases, and the gait phase corresponding to each set of preprocessed plantar pressure data is obtained. Furthermore, by establishing a one-to-one correspondence between the synchronized plantar pressure data and the electromyographic (EMG) signal feature data, a corresponding gait phase label is attached to each set of corresponding EMG signal feature data. The EMG signal feature data obtains the corresponding gait phase, resulting in EMG signal feature data with gait speed and gait phase labels. 5) Change the set gait speed and repeat steps 2) to 4) until all set gait speeds are completed, and obtain electromyographic signal feature data with gait speed labels and phase labels under all set gait speeds to form a dataset; divide the dataset into training set and test set; 6) Construct a Convolutional Long Short-Term Memory (CNN-LSTM) model: a) Construct a convolutional neural network (CNN) with electromyographic signal feature data as input and gait speed as output; b) Divide the output of the convolutional neural network into different gait speed subsets according to the set gait speed. Each gait speed subset corresponds to the gait speed set in steps 2) to 5). The feature emphasis of muscle activity differs under different gait speeds. Perform principal component analysis for each gait speed subset. The feature emphasis difference is used to perform corresponding PCA dimensionality reduction processing on the data features under different gait speed subsets to obtain the dimensionality-reduced gait data features. c) Construct a Long Short-Term Memory (LSTM) network, with the dimensionality-reduced gait data features as input and the gait phase as output; d) Train the CNN-LSTM model using the training set and test the CNN-LSTM model using the test set to obtain the trained CNN-LSTM model; 7) The multi-channel electromyography (EMG) acquisition device is fixed on the subject's leg on the same side. The multi-channel EMG acquisition device acquires EMG signals and transmits them to the computer to obtain raw EMG signal data. The computer preprocesses the raw EMG signal data and uses a sliding window to extract segmented time-domain and frequency-domain features to obtain EMG signal feature data. The EMG signal feature data is then input into the trained CNN-LSTM model to obtain gait speed and gait phase.

2. The method for recognizing walking gait phase and velocity as described in claim 1, characterized in that, In step 1), the number of channels of the multi-channel electromyography (EMG) acquisition device is M, where M is a natural number ≥ 4; the number of force-sensitive resistor sensors is N, where N is a natural number ≥ 3.

3. The method for recognizing walking gait phase and velocity as described in claim 1, characterized in that, In step 1), a multi-channel electromyography (EMG) acquisition device acquires EMG signals of the muscles in a single leg. The muscles in a single leg include the tibialis anterior and medial gastrocnemius, as well as multiple muscles from the vastus lateralis, vastus medialis, rectus femoris, biceps femoris, soleus, lateral gastrocnemius, and semitendinosus.

4. The method for recognizing walking gait phase and velocity as described in claim 2, characterized in that, In step 2), each set of raw electromyographic signal data includes M raw electromyographic signal data, and each set of raw plantar pressure data includes N raw plantar pressure data.

5. The method for recognizing walking gait phase and speed as described in claim 1, characterized in that, In step 3), the preprocessing of the raw electromyographic signal data includes: first, processing the raw electromyographic signal data with high-pass and low-pass linear phase finite impulse response filters, and then performing full-wave correction and Butterworth low-pass filter processing to obtain the preprocessed electromyographic signal data.

6. The method for recognizing walking gait phase and velocity as described in claim 1, characterized in that, In step 3), the preprocessing of the original pressure data includes performing noise reduction, filtering and interpolation preprocessing operations on the original plantar pressure data in sequence to obtain the preprocessed plantar pressure data.

7. The method for recognizing walking gait phase and velocity as described in claim 1, characterized in that, In step 3), the preprocessed electromyographic signal data is segmented into time domain and frequency domain features using a sliding window method with a fixed window size and overlapping data.

8. The method for recognizing walking gait phase and velocity as described in claim 1, characterized in that, In step 5), the ratio of training set to test set is 70%:30%~80%:20%.

9. The method for recognizing walking gait phase and velocity as described in claim 1, characterized in that, In step 6)a), the convolutional neural network includes: an input layer, a sequence folding layer, a convolutional layer, a linear correction unit activation function, a fully connected layer, and a soft quantization function; wherein, the input layer receives electromyographic signal feature data; the sequence folding layer converts the input data into a sequence format; the convolutional layer uses convolution kernels to perform convolution operations on the sequence, and the output of the convolutional layer is input into the linear correction unit activation function, which introduces nonlinearity; the fully connected layer processes the output, and then the soft quantization function is used to obtain the gait speed.

10. The method for recognizing walking gait phase and velocity as described in claim 1, characterized in that, In step 6)c), the Long Short-Term Memory network includes: a sequence defolding layer, a flattening layer, an LSTM layer, a fully connected layer, and a soft quantization function. The sequence defolding layer performs a data format conversion operation, transforming the dimensionality-reduced gait data features into a format acceptable to the LSTM layer, resulting in defolded data. The flattening layer flattens the defolded data to obtain flattened data. The LSTM layer extracts time-series features from the flattened data to obtain temporal feature data. The fully connected layer integrates and synthesizes the temporal feature data output from the LSTM layer, and finally, the gait phase is obtained through the soft quantization function.

Images