Human lower limb motion intention recognition and exoskeleton robot angle predictive control method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2023-05-23
- Publication Date
- 2026-05-12
AI Technical Summary
Existing lower limb rehabilitation robots have problems with inaccurate classification of lower limb movements in terms of collaboration between exoskeleton robots and wearers, making it difficult to effectively combine the recognition of human lower limb movement intentions and angle prediction.
Preprocessing was performed using surface electromyography (EMG) signals, Butterworth bandpass filter, and discrete wavelet transform. Feature values of motion segments were extracted using the sliding window method and fast Fourier transform. A regression model was constructed using a backpropagation neural network for intention recognition, and joint angles were predicted using a NARX neural network.
This technology enables exoskeleton robots to synchronously and accurately predict the movement intentions and joint angles of the human lower limbs, improving the accuracy of system response and ensuring the accuracy of human-machine collaboration.
Smart Images
Figure CN116584961B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of human exoskeleton control technology, specifically to a method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot. Background Technology
[0002] Existing researchers have proposed many new methods for studying electromyography (EMG) signals and classifying lower limb movements. These include analyzing EMG signals using real-time feature extraction algorithms based on differential surface electromyography (sEMG); amplifying and filtering EMG signals using dual RC active notch filters; estimating torque magnitude through signal fusion to reflect human movement status; and using seven-jointed exoskeleton robots to assist human movement. Current research on lower limb rehabilitation robots can, to some extent, achieve simple control of rehabilitation institutions based on the recognition of human lower limb movement intentions. However, because exoskeleton robots operate by attaching themselves to the human body, the exoskeleton robot itself needs to work in close coordination with the wearer. Current research largely focuses on optimizing the exoskeleton hardware; therefore, effectively classifying lower limb movements is the primary issue in obtaining accurate human lower limb movement intentions. sEMG signals contain a wealth of biological information. By analyzing and processing this information, predictions of human behavior can be made. Therefore, research on lower limb movement classification based on surface EMG signals is indispensable for the development of rehabilitation robots.
[0003] Exoskeleton robots require a high degree of coordination with the wearer. Existing lower limb rehabilitation robots can, to some extent, recognize the movement intentions of the human lower limbs and achieve simple control of rehabilitation institutions. However, most current research focuses on optimizing the exoskeleton hardware. Further research is needed on how to effectively classify lower limb movements and obtain accurate human lower limb movement intentions. Summary of the Invention
[0004] This invention addresses the problem that existing human lower limb exoskeleton robots cannot effectively combine human lower limb movement intention recognition and angle prediction to assist human lower limb movement. It proposes a method for human lower limb movement intention recognition and exoskeleton robot angle prediction and control.
[0005] The technical solution adopted to achieve the purpose of this invention is as follows: a method for recognizing the movement intention of human lower limbs and predicting and controlling the angle of an exoskeleton robot, comprising the following steps:
[0006] 1) Collect electromyographic signals from the human body surface and convert the joint angle values collected by the articular angle meter into joint signals for recording.
[0007] 2) The acquired electromyographic and joint signals are preprocessed using a Butterworth bandpass filter to obtain preprocessed electromyographic and joint signals.
[0008] 3) Use discrete wavelet transform to perform noise reduction filtering on the preprocessed electromyography (EMG) and joint signals to obtain the noise-reduced EMG and joint signals.
[0009] 4) Use the sliding window method to extract the noise-reduced electromyographic signals to obtain the action segments in the electromyographic signals, and extract the feature values of the action segments.
[0010] 5) The captured action segments are converted into frequency domain signals by using Fast Fourier Transform, and frequency domain analysis is performed on the captured action segments to obtain the median frequency of the action segments.
[0011] 6) Using the extracted feature values and median frequency as indicators of accuracy, a regression model is constructed using a BP neural network, and the lower limb movement intention is identified through the regression model.
[0012] 7) Construct a NARX neural network and use the NARX neural network to predict the joint angle of the exoskeleton robot corresponding to the current lower limb movement intention.
[0013] Furthermore, the sites for collecting electromyographic signals on the human body surface include the rectus femoris, vastus medialis, biceps femoris, tibialis anterior, lateral gastrocnemius, and soleus.
[0014] Furthermore, the joint angles include hip joint angles and knee joint angles.
[0015] Furthermore, the movement intentions of the lower limbs include going up stairs, going uphill, walking on flat ground, going down stairs, and going downhill.
[0016] Furthermore, when the intended movement is walking on flat ground, the corresponding knee joint angle range is -1.3° to 71.55°, and the hip joint angle range is -3° to 13.18°.
[0017] When the intention of the movement is to climb stairs, the corresponding knee joint angle ranges from -92.52° to 1.98°, and the hip joint angle ranges from -0.36° to 19.57°.
[0018] When the intention of the movement is uphill, the corresponding range of knee joint angle is -60.34° to 13.59°, and the range of hip joint angle is -1.08° to 14.17°.
[0019] When the intention of movement is to go down stairs, the corresponding range of knee joint angle is -93.015° to 0.09°, and the range of hip joint angle is -1.35° to 18.49°.
[0020] When the intention of the movement is downhill, the corresponding range of knee joint angle is -63° to 1.125°, and the range of hip joint angle is -2.57° to 17.01°.
[0021] Furthermore, the step of using the sliding window method to extract the motion segments from the denoised electromyographic signal includes:
[0022] 4.1) Set the truncation threshold, sliding window width, and step value.
[0023] 4.2) Determine the starting position of the action segment.
[0024] 4.3) Determine the end position of the action segment.
[0025] 4.4) The action segment in the electromyographic signal is extracted by the beginning and end positions of the action segment.
[0026] Furthermore, the feature values of the action segment include: mean absolute value, root mean square, standard deviation, and zero crossing point.
[0027] Furthermore, the step of constructing a regression model using a BP neural network, with the extracted feature values and median frequency as indicators of accuracy, includes:
[0028] 6.1) Import the dataset containing the movement intentions of the lower limbs and the corresponding feature values and median frequencies.
[0029] 6.2) Divide the dataset into training and test sets.
[0030] 6.3) Normalize the training set and the test set to obtain the normalized training set and test set.
[0031] 6.4) Initialize the relevant parameters and construct the BP neural network.
[0032] 6.5) Train the BP neural network using the normalized training set to determine the optimal number of hidden layers and obtain the regression model.
[0033] 6.6) Test the regression model of the best hidden layer using the normalized test set, and perform inverse normalization and index error analysis on the test results. If the index error is greater than the preset error threshold, return to step 6.1). If the index error is less than or equal to the preset error threshold, end the training and output the regression model.
[0034] Furthermore, the steps for training the BP neural network using the normalized training set to determine the optimal number of hidden layers include:
[0035] 6.5.1) Calculate the number of hidden layers in the a-th iteration, i.e.:
[0036]
[0037] In the formula, hiddennum is the number of hidden layer nodes. m is the number of input layer nodes. n is the number of output layer nodes. a is initially 1. a is an integer.
[0038] 6.5.2) Using hiddennum as the number of hidden layers in the current BP neural network, train the BP neural network using the normalized training set, and calculate the mean square error of the BP neural network.
[0039] 6.5.3) Determine if a is less than k, where k is the maximum threshold for the number of layers. If yes, set a = a + 1 and return to step 6.5.1). If no, proceed to step 6.5.4.
[0040] 6.5.4) The number of hidden layers in the BP neural network is the number of layers corresponding to the minimum mean square error.
[0041] Furthermore, the NARX neural network is shown below:
[0042] y(t)=f{u(tD u ),…,u(t-1),u(t),y(tD y ),…,y(t-1)} (2)
[0043] In the formula, u(t) and y(t) are the input and output of the NARX neural network at time t, respectively. u D represents the maximum order of the input delay. y Let u(tD) be the maximum order of the output delay. u ),…,u(t-1) represent the historical input relative to time t. y(tD) y ),…,y(t-1)} represents the historical output relative to time t. f(·) is the nonlinear function obtained by fitting the NARX neural network.
[0044] The technical effects of this invention are undeniable. This invention enables exoskeleton robots to simultaneously realize the movement intention of the human lower limbs and the estimation and prediction of the joint angles of the lower limbs. The combination of the two serves as a double guarantee to ensure that the system response has higher accuracy, enabling the exoskeleton robot to accurately identify the wearer's movement intention and achieve a high degree of human-machine collaboration.
[0045] This invention combines the recognition of human lower limb movement intentions with joint angle prediction, enabling exoskeleton robots to assist human lower limb movements more accurately and efficiently, and with movements that are natural and conform to the human body. Attached Figure Description
[0046] Figure 1A flowchart of a method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot;
[0047] Figure 2 Flowchart for building a BP neural network;
[0048] Figure 3 This is a schematic diagram of the NARX neural network structure model. Detailed Implementation
[0049] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.
[0050] Example 1:
[0051] See Figures 1 to 3 A method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot includes the following steps:
[0052] 1) Collect electromyographic signals from the human body surface and convert the joint angle values collected by the articular angle meter into joint signals for recording.
[0053] 2) The acquired electromyographic and joint signals are preprocessed using a Butterworth bandpass filter to obtain preprocessed electromyographic and joint signals.
[0054] 3) Use discrete wavelet transform to perform noise reduction filtering on the preprocessed electromyography (EMG) and joint signals to obtain the noise-reduced EMG and joint signals.
[0055] 4) Use the sliding window method to extract the noise-reduced electromyographic signals to obtain the action segments in the electromyographic signals, and extract the feature values of the action segments.
[0056] 5) The captured action segments are converted into frequency domain signals by using Fast Fourier Transform, and frequency domain analysis is performed on the captured action segments to obtain the median frequency of the action segments.
[0057] 6) Using the extracted feature values and median frequency as indicators of accuracy, a regression model is constructed using a BP neural network, and the lower limb movement intention is identified through the regression model.
[0058] 7) Construct a NARX neural network and use the NARX neural network to predict the joint angle of the exoskeleton robot corresponding to the current lower limb movement intention.
[0059] Example 2:
[0060] The method for recognizing the movement intention of the human lower limbs and predicting and controlling the angle of the exoskeleton robot, the main steps of which are described in Example 1, further, the collection sites of the electromyographic signals on the human body surface include the rectus femoris, vastus medialis, biceps femoris, tibialis anterior, lateral gastrocnemius, and soleus.
[0061] Example 3:
[0062] The method for recognizing human lower limb movement intentions and predicting and controlling the angles of an exoskeleton robot, with the main steps described in either Example 1 or 2, further wherein the joint angles include the hip joint angle and the knee joint angle.
[0063] Example 4:
[0064] The method for recognizing the movement intention of the human lower limbs and predicting and controlling the angle of the exoskeleton robot, the main steps of which are described in Example 3, further, the movement intention of the lower limbs includes going up stairs, going uphill, walking on flat ground, going down stairs, and going downhill.
[0065] Example 5:
[0066] The method for recognizing the movement intention of the human lower limbs and predicting and controlling the angle of the exoskeleton robot, the main steps of which are described in Example 4, further, when the movement intention is to walk on flat ground, the corresponding knee joint angle range is -1.3° to 71.55°, and the hip joint angle range is -3° to 13.18°.
[0067] When the intention of the movement is to climb stairs, the corresponding knee joint angle ranges from -92.52° to 1.98°, and the hip joint angle ranges from -0.36° to 19.57°.
[0068] When the intention of the movement is uphill, the corresponding range of knee joint angle is -60.34° to 13.59°, and the range of hip joint angle is -1.08° to 14.17°.
[0069] When the intention of movement is to go down stairs, the corresponding range of knee joint angle is -93.015° to 0.09°, and the range of hip joint angle is -1.35° to 18.49°.
[0070] When the intention of the movement is downhill, the corresponding range of knee joint angle is -63° to 1.125°, and the range of hip joint angle is -2.57° to 17.01°.
[0071] Example 6:
[0072] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, with the main steps described in any one of Examples 1 to 5, further includes the step of using the sliding window method to extract the motion segment from the denoised electromyographic signal.
[0073] 4.1) Set the truncation threshold, sliding window width, and step value.
[0074] 4.2) Determine the starting position of the action segment.
[0075] 4.3) Determine the end position of the action segment.
[0076] 4.4) The action segment in the electromyographic signal is extracted by the beginning and end positions of the action segment.
[0077] Example 7:
[0078] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot includes the main steps described in any one of Examples 1 to 6. Furthermore, the feature values of the action segment include: mean absolute value, root mean square, standard deviation, and zero crossing point.
[0079] Example 8:
[0080] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, with the main steps described in any one of Examples 1 to 7, further includes the step of constructing a regression model using a BP neural network with the extracted feature values and median frequency as indicators of accuracy.
[0081] 6.1) Import the dataset containing the movement intentions of the lower limbs and the corresponding feature values and median frequencies.
[0082] 6.2) Divide the dataset into training and test sets.
[0083] 6.3) Normalize the training set and the test set to obtain the normalized training set and test set.
[0084] 6.4) Initialize the relevant parameters and construct the BP neural network.
[0085] 6.5) Train the BP neural network using the normalized training set to determine the optimal number of hidden layers and obtain the regression model.
[0086] 6.6) Test the regression model of the best hidden layer using the normalized test set, and perform inverse normalization and index error analysis on the test results. If the index error is greater than the preset error threshold, return to step 6.1). If the index error is less than or equal to the preset error threshold, end the training and output the regression model.
[0087] Example 9:
[0088] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, with the main steps described in Example 8, further includes the step of training the BP neural network using a normalized training set to determine the optimal number of hidden layers, which includes:
[0089] 6.5.1) Calculate the number of hidden layers in the a-th iteration, i.e.:
[0090]
[0091] In the formula, hiddennum is the number of hidden layer nodes. m is the number of input layer nodes. n is the number of output layer nodes. a is initially 1. a is an integer.
[0092] 6.5.2) Using hiddennum as the number of hidden layers in the current BP neural network, train the BP neural network using the normalized training set, and calculate the mean square error of the BP neural network.
[0093] 6.5.3) Determine if a is less than k, where k is the maximum threshold for the number of layers. If yes, set a = a + 1 and return to step 6.5.1). If no, proceed to step 6.5.4.
[0094] 6.5.4) The number of hidden layers in the BP neural network is the number of layers corresponding to the minimum mean square error.
[0095] Example 10:
[0096] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, with the main steps described in any one of Examples 1 to 9, further wherein the NARX neural network is as follows:
[0097] y(t)=f{u(tD u ),…,u(t-1),u(t),y(tD y ),…,y(t-1)} (2)
[0098] In the formula, u(t) and y(t) are the input and output of the NARX neural network at time t, respectively. u D represents the maximum order of the input delay. y Let u(tD) be the maximum order of the output delay. u ),…,u(t-1) represent the historical input relative to time t. y(tD) y ),…,y(t-1)} represents the historical output relative to time t. f(·) is the nonlinear function obtained by fitting the NARX neural network.
[0099] Example 11:
[0100] See Figures 1 to 3 A method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot includes the following steps:
[0101] 1) Collect electromyographic signals on the human body surface using the Biometrics Ltd wireless electromyography signal acquisition device, and simultaneously convert the joint angle values collected by the articular angle meter into joint signals for recording.
[0102] 2) The acquired electromyographic and joint signals are preprocessed using a Butterworth bandpass filter to obtain preprocessed electromyographic and joint signals.
[0103] 3) Use discrete wavelet transform to perform noise reduction filtering on the preprocessed electromyography (EMG) and joint signals to obtain the noise-reduced EMG and joint signals.
[0104] 4) Use the sliding window method to extract the noise-reduced electromyographic signals to obtain the action segments in the electromyographic signals, and extract the feature values of the action segments.
[0105] 5) The captured action segments are converted into frequency domain signals by using Fast Fourier Transform, and frequency domain analysis is performed on the captured action segments to obtain the median frequency of the action segments.
[0106] 6) Using the extracted feature values and median frequency as indicators of accuracy, a regression model is constructed using a BP neural network, and the lower limb movement intention is identified through the regression model.
[0107] 7) Construct a NARX neural network and use the NARX neural network to predict the joint angle of the exoskeleton robot corresponding to the current lower limb movement intention.
[0108] Users wear exoskeleton robots, which assist them in climbing stairs, going uphill, walking on flat ground, and going downhill.
[0109] Example 12:
[0110] The method for recognizing the movement intention of the human lower limbs and predicting and controlling the angle of the exoskeleton robot, the main steps of which are described in Example 11, further, the collection sites of the electromyographic signals on the human body surface include the rectus femoris, vastus medialis, biceps femoris, tibialis anterior, lateral gastrocnemius, and soleus.
[0111] Example 13:
[0112] The method for recognizing human lower limb movement intentions and predicting and controlling the angles of an exoskeleton robot includes the main steps described in either Example 11 or 12. Furthermore, the joint angles include the hip joint angles and knee joint angles of the left and right legs.
[0113] Example 14:
[0114] The method for recognizing the movement intention of the human lower limbs and predicting and controlling the angle of the exoskeleton robot has the main steps described in Example 13. Furthermore, the movement intention of the lower limbs includes going up stairs, going uphill, walking on flat ground, going down stairs, and going downhill.
[0115] Example 15:
[0116] The method for recognizing the movement intention of the human lower limbs and predicting and controlling the angle of the exoskeleton robot, the main steps of which are described in Example 14, further, when the movement intention is to walk on flat ground, the corresponding knee joint angle range is -1.3° to 71.55°, and the hip joint angle range is -3° to 13.18°.
[0117] When the intention of the movement is to climb stairs, the corresponding knee joint angle ranges from -92.52° to 1.98°, and the hip joint angle ranges from -0.36° to 19.57°.
[0118] When the intention of the movement is uphill, the corresponding range of knee joint angle is -60.34° to 13.59°, and the range of hip joint angle is -1.08° to 14.17°.
[0119] When the intention of movement is to go down stairs, the corresponding range of knee joint angle is -93.015° to 0.09°, and the range of hip joint angle is -1.35° to 18.49°.
[0120] When the intention of the movement is downhill, the corresponding range of knee joint angle is -63° to 1.125°, and the range of hip joint angle is -2.57° to 17.01°.
[0121] Example 16:
[0122] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, with the main steps described in any one of Examples 11 to 15, further includes the step of using the sliding window method to extract the motion segment from the denoised electromyographic signal.
[0123] 4.1) Set the truncation threshold, sliding window width, and step value.
[0124] 4.2) Determine the starting position of the action segment.
[0125] 4.3) Determine the end position of the action segment.
[0126] 4.4) The action segment in the electromyographic signal is extracted by the beginning and end positions of the action segment.
[0127] Example 17:
[0128] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot includes the main steps described in any one of Examples 11 to 16. Furthermore, the feature values of the action segment include: mean absolute value, root mean square, standard deviation, and zero crossing point.
[0129] Example 18:
[0130] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, with the main steps described in any one of Examples 11 to 17, further includes the step of constructing a regression model using a BP neural network with the extracted feature values and median frequency as indicators of accuracy.
[0131] 6.1) Import the dataset containing the movement intentions of the lower limbs and the corresponding feature values and median frequencies.
[0132] 6.2) Divide the dataset into training and test sets.
[0133] 6.3) Normalize the training set and the test set to obtain the normalized training set and test set.
[0134] 6.4) Initialize the relevant parameters and construct the BP neural network.
[0135] 6.5) Train the BP neural network using the normalized training set to determine the optimal number of hidden layers and obtain the regression model.
[0136] 6.6) Test the regression model of the best hidden layer using the normalized test set, and perform inverse normalization and index error analysis on the test results. If the index error is greater than the preset error threshold, return to step 6.1). If the index error is less than or equal to the preset error threshold, end the training and output the regression model.
[0137] Example 19:
[0138] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, with the main steps described in Example 18, further includes the following steps: determining the optimal number of hidden layers using a normalized training set and constructing a BP neural network with the optimal number of hidden layers:
[0139] 6.5.1) Calculate the number of hidden layers in the a-th iteration, i.e.:
[0140]
[0141] In the formula, hiddennum is the number of hidden layer nodes. m is the number of input layer nodes. n is the number of output layer nodes. a is initially 1. a is an integer.
[0142] 6.5.2) Using hiddennum as the number of hidden layers in the current BP neural network, train the BP neural network using the normalized training set, and calculate the mean square error of the BP neural network.
[0143] 6.5.3) Determine if a is less than k, where k is the maximum threshold for the number of layers, k = 10. If yes, set a = a + 1 and return to step 6.5.1). If no, proceed to step 6.5.4.
[0144] 6.5.4) The number of hidden layers in the BP neural network is the number of layers corresponding to the minimum mean square error.
[0145] Example 20:
[0146] The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, with main steps described in any one of Examples 11 to 19, further utilizes the NARX neural network (Based on the nonlinear autoregressive with exogeneous inputs neural network), a model for describing nonlinear discrete systems. The NARX neural network is shown below:
[0147] y(t)=f{u(tD u ),…,u(t-1),u(t),y(tD y ),…,y(t-1)} (2)
[0148] In the formula, u(t) and y(t) are the input and output of the NARX neural network at time t, respectively. u D represents the maximum order of the input delay. y Let u(tD) be the maximum order of the output delay. u ),…,u(t-1) represent the historical input relative to time t. y(tD) y ),…,y(t-1)} represents the historical output relative to time t. f(·) is the nonlinear function obtained by fitting the NARX neural network.
[0149] Example 21:
[0150] See Figures 1 to 3 The working principle of the method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot is as follows:
[0151] This embodiment extracts electromyographic (EMG) signals from six muscles in both the left and right legs: rectus femoris, vastus medialis, biceps femoris, tibialis anterior, gastrocnemius lateralis, and soleus. Simultaneously, the hip and knee joint angles of both legs are converted into signals for recording. The extracted signals are preprocessed using a Butterworth bandpass filter, followed by further noise reduction filtering using discrete wavelet transform. A sliding window method is used to perform time-domain analysis on the processed EMG signals, extracting four features: mean absolute value, root mean square, standard deviation, and zero-crossing points. A fast Fourier transform is then used to convert the time-domain signal into a frequency-domain signal, and frequency domain analysis is performed to obtain the median frequency. A backpropagation (BP) neural network is used to construct a regression model, using the extracted features as an accuracy metric to achieve lower limb movement intention recognition and joint angle prediction.
[0152] Example 22:
[0153] See Figures 1 to 3 The main contents of the method for recognizing the movement intention of the human lower limbs and predicting and controlling the angle of the exoskeleton robot are as follows:
[0154] Surface electromyography signal preprocessing and feature extraction
[0155] Preprocessing: The raw electromyographic signals were processed using Butterworth filtering and wavelet denoising.
[0156] This project uses the sliding window method to extract feature values of electromyography (EMG) signals from the human body surface. By setting a reasonable sliding window width and step value, and setting an appropriate threshold, motion segments in the EMG signals are extracted, thereby extracting the feature values of the motion segments.
[0157] 2. Using a backpropagation neural network for human motion intention recognition
[0158] The movement intention of this project includes the following 5 actions: climbing stairs, going uphill, walking on flat ground, going down stairs, and going downhill. The flowchart for building the BP neural network is as follows: Figure 2 .
[0159] The optimal number of hidden layers is determined by calculating the range of hidden layer nodes based on the number of input and output nodes. The range of hidden layer nodes is calculated using the following formula, where m represents the number of input layer nodes, n represents the number of output layer nodes, and a is an integer between 1 and 10. According to formula (1), 10 selectable hidden layer nodes are obtained. These 10 hidden layer nodes are then set sequentially for the BP network, and the hidden layer node with the smallest error is obtained by calculating the mean square error (MSE) of the training set. This is the optimal hidden layer node.
[0160]
[0161] 3. Angle Prediction NARX Neural Network
[0162] NARX neural networks (Based on the nonlinear autoregressive with exogeneous inputs neural network) are models used to describe nonlinear discrete systems. They are represented as:
[0163] y(t)=f{u(tD u ),…,u(t-1),u(t),y(tD y ),…,y(t-1)}
[0164] In the formula: u(t) and y(t) are the input and output of the network at time t, respectively; D u D is the maximum order of the input delay; y The maximum order of the output delay is given by u(tD); therefore, u(tD) u ),…,u(t-1) represent the historical input relative to time t; y(tD) y ),…,y(t-1) represent the historical outputs relative to time t; f is the nonlinear function obtained by network fitting.
[0165] NARX neural network structure model as follows Figure 3 As shown.
Claims
1. A method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot, characterized in that, Includes the following steps: 1) Collect electromyographic signals from the human body surface and simultaneously convert the joint angle values collected by the articular angle meter into joint signals for recording; The joint angles include the hip joint angle and the knee joint angle; 2) The acquired electromyographic and joint signals are preprocessed using a Butterworth bandpass filter to obtain preprocessed electromyographic and joint signals; 3) Use discrete wavelet transform to perform noise reduction filtering on the preprocessed electromyography (EMG) and joint signals to obtain the noise-reduced EMG and joint signals; 4) Use the sliding window method to extract the noise-reduced electromyographic signals to obtain the action segments in the electromyographic signals, and extract the feature values of the action segments; The step of extracting motion segments from the denoised electromyographic signal using the sliding window method includes: 4.1) Set the truncation threshold, sliding window width, and step value; 4.2) Determine the starting position of the action segment; 4.3) Determine the endpoint of the action segment; 4.4) By identifying the beginning and end positions of the action segment, the action segment in the electromyographic signal is extracted; 5) Convert the captured action segments into frequency domain signals using Fast Fourier Transform, perform frequency domain analysis on the captured action segments, and obtain the median frequency of the action segments. 6) Using the extracted feature values and median frequency as indicators of accuracy, a regression model is constructed using a BP neural network, and the regression model is used to identify lower limb movement intentions; The lower limb movement intentions include going up stairs, going uphill, walking on flat ground, going down stairs, and going downhill; When the intended movement is walking on flat ground, the corresponding knee joint angle range is -1.3° to 71.55°, and the hip joint angle range is -3° to 13.18°. When the intention of movement is to climb stairs, the corresponding range of knee joint angle is -92.52° to 1.98°, and the range of hip joint angle is -0.36° to 19.57°. When the intention of the movement is uphill, the corresponding range of knee joint angle is -60.34° to 13.59°, and the range of hip joint angle is -1.08° to 14.17°. When the intention of movement is to go down stairs, the corresponding range of knee joint angle is -93.015° to 0.09°, and the range of hip joint angle is -1.35° to 18.49°. When the intention of the movement is downhill, the corresponding range of knee joint angle is -63° to 1.125°, and the range of hip joint angle is -2.57° to 17.01°. 7) Construct a NARX neural network and use the NARX neural network to predict the joint angle of the exoskeleton robot corresponding to the current lower limb movement intention.
2. The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot according to claim 1, characterized in that, The sites for collecting electromyographic signals on the human body surface include the rectus femoris, vastus medialis, biceps femoris, tibialis anterior, lateral gastrocnemius, and soleus.
3. The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot according to claim 1, characterized in that, The feature values of the action segment include: mean absolute value, root mean square, standard deviation, and zero crossing point.
4. The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot according to claim 1, characterized in that, The steps for constructing a regression model using a BP neural network, with the extracted feature values and median frequency as indicators of accuracy, include: 6.1) Import the dataset containing the movement intentions of the lower limbs and the corresponding feature values and median frequencies; 6.2) Divide the dataset into a training set and a test set; 6.3) Normalize the training set and test set to obtain the normalized training set and test set; 6.4) Initialize the relevant parameters and construct the BP neural network; 6.5) Train the BP neural network using the normalized training set to determine the optimal number of hidden layers and obtain the regression model; 6.6) Test the regression model of the best hidden layer using the normalized test set, and perform inverse normalization and index error analysis on the test results. If the index error is greater than the preset error threshold, return to step 6.1). If the index error is less than or equal to the preset error threshold, end the training and output the regression model.
5. The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot according to claim 4, characterized in that, The steps for training a BP neural network using a normalized training set and determining the optimal number of hidden layers include: 6.5.1) Calculate the number of hidden layers in the a-th iteration, i.e.: In the formula, m is the number of hidden layer nodes; n is the number of input layer nodes; a is the number of output layer nodes; a is initially 1; a is an integer. 6.5.2) with The number of hidden layers in the current BP neural network is used as the basis for training the BP neural network using the normalized training set, and the mean square error of the BP neural network is calculated. 6.5.3) Determine if a is less than k, where k is the maximum threshold for the number of layers; if yes, let a = a + 1 and return to step 6.5.1); if no, proceed to step 6.5.
4. 6.5.4) The value corresponding to the minimum mean square error The number of hidden layers in a BP neural network.
6. The method for recognizing human lower limb movement intentions and predicting and controlling the angle of an exoskeleton robot according to claim 1, characterized in that, The NARX neural network is shown below: In the formula, , These are the input and output of the NARX neural network at time t, respectively. The maximum order of the input delay; This represents the maximum order of the output delay. This is the historical input relative to time t; This is the historical output relative to time t; The nonlinear function obtained by fitting the NARX neural network.