A gait phase estimation method based on lightweight LSTM

Abstract

The application discloses a gait phase estimation method based on a light LSTM, and relates to the fields of artificial intelligence and rehabilitation robots. The method first carries out pretreatment on a hip joint angle signal collected by a sensor, and the pretreated data is used to detect a time when a heel touches the ground, generate a gait phase label, introduce feature engineering, extract 8-dimensional gait features, construct a light LSTM model based on an improved activation function, and finally output a gait phase. Compared with the prior art, the application reduces the dependence of a traditional method on a sensor, can improve model precision, reduce model complexity, and relieve the gradient vanishing or explosion problem.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and rehabilitation robots, and specifically to a gait phase estimation method based on lightweight LSTM. Background Technology

[0002] Hemiplegic gait is a prominent feature of stroke patients. Due to insufficient muscle strength on the affected side, the hip and knee joints exhibit flexion defects and phase lag. Traditional rehabilitation training suffers from drawbacks such as being labor-intensive and lacking repetitiveness. Exoskeleton-assisted rehabilitation holds promise for addressing these issues. Lower limb rehabilitation exoskeletons, based on bionic principles, provide support and assistance to the wearer through external fixation and drive mechanisms. They can effectively meet the needs of rehabilitation training, alleviate the shortage of professional nursing staff, and help stroke patients regain lower limb motor function.

[0003] Gait phase is a crucial state variable that helps exoskeletons provide assistance to patients. Exoskeleton systems need to determine the assistance time and calculate reference assistance torque based on the gait phase to achieve precise trajectory or torque tracking control. Finite state machines (FSMs) or oscillators are commonly used to estimate gait phase, but FSMs require manual calibration of gait events, and oscillators struggle to adapt to irregular gait variations. Deep convolutional neural networks (CNNs) and long short-term memory (LSTM) neural networks can accurately estimate the phase of irregular gait, but due to the large number of network layers, their training time is long, and they are prone to gradient vanishing and gradient exploding problems. Summary of the Invention

[0004] This invention proposes a gait phase estimation method based on a lightweight LSTM, aiming to improve the prediction accuracy of the model, reduce its complexity, and alleviate the gradient vanishing or exploding problem in deep neural networks. First, the hip joint angle signal acquired by the sensor is preprocessed. The preprocessed data is used to detect the heel strike time and generate gait phase labels. Then, feature engineering is introduced, and a lightweight LSTM model is constructed based on an improved activation function. Finally, the gait phase is output. The method is implemented through the following technical solutions:

[0005] Step 1: Collect the subject's hip joint angle.

[0006] Step 2: Preprocess the original hip joint angle.

[0007] Step 2.1: To address the jumps and missing information in the angle signal, median filtering is used to suppress impulse noise and fill in signal gaps.

[0008] Step 2.2: To reduce the impact of sensor baseline drift and local fluctuations within the gait cycle on the detection of heel strike, a moving average filter is used to smooth the signal.

[0009] Step 3: Generate gait phase labels.

[0010] Step 3.1: Mark the local minimum value of the hip joint angle as the heel position and detect the heel strike time.

[0011] Step 3.2: Mark the gait phase label. The gait phase corresponding to the first heel strike in a complete gait cycle is 0%, and the next heel strike corresponds to 100%.

[0012] Step 4: To make the model lightweight and improve its accuracy, feature engineering is introduced to extract 8-dimensional features of the hip joint angle: original angle angular velocity angular acceleration Position within the period Period length ratio Sliding window mean Sliding window standard deviation Low-frequency components of Fast Fourier Transform The specific expression is as follows:

[0013] ,

[0014] ,

[0015] ,

[0016] ,

[0017] ,

[0018] in, For the current moment, Indicates the first The moment of the first heel strike within a gait cycle It is the first The length of a gait cycle Represents the current cycle length. The average period length, Indicates the size of the sliding window. It is a discrete angle signal. Represents signal length. For the first The amplitude of each frequency component.

[0019] Step 5: Construct a lightweight LSTM model to output gait phase.

[0020] Step 5.1: The input layer of the model receives an 8-dimensional temporal feature sequence.

[0021] Step 5.2: The only LSTM layer updates the state through a gating mechanism to capture the temporal dependencies of the gait.

[0022] Step 5.3: The output of the LSTM layer is normalized and then input into the fully connected layer.

[0023] Step 5.4: The fully connected layer compresses the feature dimensions to focus on the core information and applies L2 regularization to prevent overfitting.

[0024] Step 5.5: Activation function layer alleviates gradient problem. The activation function is defined as:

[0025] ,

[0026] Among them, the sine term Introducing nonlinearity, Used to adjust the convergence speed of the gradient, residual term Stabilize the gradient.

[0027] Step 5.6: The output layer maps the features to gait phase.

[0028] Step 6: Use root mean square error and correlation coefficient to evaluate the accuracy of the model, and use floating-point operations to evaluate the computational complexity of the model.

[0029] This invention estimates gait phase using only the hip joint angle, reducing reliance on sensors. Feature engineering is introduced to cover the time, frequency, statistical, and kinematic features of gait, contributing to accurate gait phase estimation. The lightweight LSTM model significantly reduces model complexity and training time. The activation function designed in this invention effectively alleviates the vanishing and exploding gradient problems. Attached Figure Description

[0030] Figure 1 This is an overall flowchart of an embodiment of the present invention. Detailed Implementation

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

[0032] Figure 1 This is a flowchart of an embodiment. This embodiment provides a gait phase estimation method based on a lightweight LSTM. The specific process includes: collecting the hip joint angle of the subject; preprocessing the angle data; detecting the heel strike time and generating gait phase labels; introducing feature engineering to extract 8-dimensional features; constructing a lightweight LSTM model to output the gait phase; and evaluating the model performance using root mean square error, correlation coefficient, and floating-point computation.

[0033] The specific implementation steps of a gait phase estimation method based on lightweight LSTM are as follows:

[0034] Step 1: Collect the hip joint angle of the subject while walking on the treadmill.

[0035] Step 2: Preprocess the original hip joint angle.

[0036] Step 2.1: To address the jumps and missing information in the angle signal, median filtering is used to suppress impulse noise and fill in signal gaps.

[0037] Step 2.2: To reduce the impact of sensor baseline drift and local fluctuations within the gait cycle on the detection of heel strike, a moving average filter is used to smooth the signal.

[0038] Step 3: Generate gait phase labels.

[0039] Step 3.1: Mark the local minimum value of the hip joint angle as the heel position and detect the heel strike time.

[0040] Step 3.2: Mark the gait phase label. The gait phase corresponding to the first heel strike in a complete gait cycle is 0%, and the next heel strike corresponds to 100%.

[0041] Step 4: To make the model lightweight and improve its accuracy, feature engineering is introduced to extract 8-dimensional features of the hip joint angle: original angle angular velocity angular acceleration Position within the period Period length ratio Sliding window mean Sliding window standard deviation Low-frequency components of Fast Fourier Transform The specific expression is as follows:

[0042] ,

[0043] ,

[0044] ,

[0045] ,

[0046] ,

[0047] in, For the current moment, Indicates the first The moment of the first heel strike within a gait cycle It is the first The length of a gait cycle Represents the current cycle length. The average period length, Indicates the size of the sliding window. It is a discrete angle signal. Represents signal length. For the first The amplitude of each frequency component, .

[0048] Step 5: Construct a lightweight LSTM model to output gait phase.

[0049] Step 5.1: The input layer of the model receives an 8-dimensional temporal feature sequence.

[0050] Step 5.2: The only LSTM layer updates the state through a gating mechanism to capture the temporal dependencies of gait, and the number of hidden neurons is set to 96.

[0051] Step 5.3: The output of the LSTM layer is processed by batch normalization and layer normalization and then input into the fully connected layer.

[0052] Step 5.4: The fully connected layer compresses the feature dimensions to focus on the core information and applies L2 regularization to prevent overfitting.

[0053] Step 5.5: The activation function layer alleviates the vanishing and exploding gradient problems. The expression for the activation function is as follows:

[0054] ,

[0055] Among them, the sine term Introducing nonlinearity, Used to adjust the convergence speed of the gradient, residual term Stabilize the gradient.

[0056] Step 5.6: The output layer maps the features to gait phase.

[0057] Step 6: Use root mean square error Correlation coefficient To evaluate the prediction accuracy of the model, floating-point operations are used to assess its computational complexity. and The expression is as follows:

[0058] ,

[0059] ,

[0060] in, The total number of samples, Indicates the predicted phase. It is a true phase. , These are the average values ​​of the predicted phase and the actual phase, respectively.

Claims

1. A gait phase estimation method based on lightweight LSTM, characterized in that, Includes the following steps: Step 1: Collect the subject's hip joint angle; Step 2: Preprocess the original hip joint angle; Step 3: Generate gait phase tags; Step 4: Introduce feature engineering to extract 8-dimensional features of the hip joint angle: original angle angular velocity angular acceleration Position within the period Period length ratio Sliding window mean Sliding window standard deviation Low-frequency components of Fast Fourier Transform The specific expression is as follows: ， ， ， ， ， in, For the current moment, Indicates the first The moment of the first heel strike within a gait cycle It is the first The length of a gait cycle Represents the current cycle length. The average period length, Indicates the size of the sliding window. It is a discrete angle signal. Represents signal length. For the first The amplitude of each frequency component; Step 5: Construct a lightweight LSTM model and output gait phase; Step 6: Use root mean square error and correlation coefficient to evaluate the accuracy of the model, and use floating-point operations to evaluate the computational complexity of the model.

2. The gait phase estimation method based on lightweight LSTM according to claim 1, characterized in that, The lightweight LSTM model construction described in step 5 is implemented as follows: Step 5.1: The input layer of the model receives an 8-dimensional temporal feature sequence; Step 5.2: The unique LSTM layer updates its state through a gating mechanism; Step 5.3: The output of the LSTM layer is normalized and then input into the fully connected layer; Step 5.4: Fully connected layers compress feature dimensions to focus on core information; Step 5.5: Activation function layer alleviates gradient problem. The activation function is defined as: ， Among them, the sine term Introducing nonlinearity, Used to adjust the convergence speed of the gradient, residual term Stabilize the gradient; Step 5.6: The output layer maps the features to gait phase.

Images