CN113520413B - Lower limb multi-joint angle estimation method based on surface electromyogram signals - Google Patents
Lower limb multi-joint angle estimation method based on surface electromyogram signals Download PDFInfo
- Publication number
- CN113520413B CN113520413B CN202110978485.1A CN202110978485A CN113520413B CN 113520413 B CN113520413 B CN 113520413B CN 202110978485 A CN202110978485 A CN 202110978485A CN 113520413 B CN113520413 B CN 113520413B
- Authority
- CN
- China
- Prior art keywords
- noise
- neural network
- return
- follows
- joint
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/389—Electromyography [EMG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/103—Detecting, measuring or recording devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
- A61B5/107—Measuring physical dimensions, e.g. size of the entire body or parts thereof
- A61B5/1071—Measuring physical dimensions, e.g. size of the entire body or parts thereof measuring angles, e.g. using goniometers
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/103—Detecting, measuring or recording devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
- A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor, mobility of a limb
- A61B5/1121—Determining geometric values, e.g. centre of rotation or angular range of movement
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/45—For evaluating or diagnosing the musculoskeletal system or teeth
- A61B5/4528—Joints
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/45—For evaluating or diagnosing the musculoskeletal system or teeth
- A61B5/4538—Evaluating a particular part of the muscoloskeletal system or a particular medical condition
- A61B5/4585—Evaluating the knee
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/725—Details of waveform analysis using specific filters therefor, e.g. Kalman or adaptive filters
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Veterinary Medicine (AREA)
- Public Health (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- Animal Behavior & Ethology (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Surgery (AREA)
- Artificial Intelligence (AREA)
- Physiology (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Signal Processing (AREA)
- Dentistry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Orthopedic Medicine & Surgery (AREA)
- Rheumatology (AREA)
- Geometry (AREA)
- Mathematical Physics (AREA)
- Fuzzy Systems (AREA)
- Evolutionary Computation (AREA)
- Physical Education & Sports Medicine (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The invention discloses a lower limb multi-joint angle estimation method based on surface electromyogram signals. The interference of the measured noise to the signal acquisition process is considered, an anti-noise type return-to-zero neural network model with the noise suppression capacity is designed, the problem that the recognition result is greatly deviated from the true value is effectively solved, the adverse effect of the noise on the regression model parameter solving can be overcome, the accurate recognition of the joint angle of the lower limb under the noise environment is realized, and the invention provides technical reference for avoiding the secondary damage possibly generated by the rehabilitation training of the affected limb.
Description
Technical Field
The invention relates to the field of human body active movement intention identification, in particular to a lower limb multi-joint angle estimation method based on surface electromyogram signals.
Background
In recent years, the number of people with motor dysfunction and even hemiplegia caused by motor central nerve injury or cerebral apoplexy is increasing. The traditional treatment means is to restore nerve tissues by means of rehabilitation medicine, and with the improvement of medical technology and the development of robot technology, an exoskeleton robot and a bionic artificial limb become a current rehabilitation training research hotspot. According to the rehabilitation medical theory, the rehabilitation training mode of recognizing the active movement intention of the human body is combined, so that the rehabilitation training device has a positive effect on the rehabilitation of patients. Therefore, how to accurately and reliably recognize the human motion intention is a key problem for the research of the robot system. Compared with the traditional human-computer interaction mode based on program control, the modern robot can actively understand the human motion mode, has independent adaptability, and realizes the robot to assist the limb disease patient to complete corresponding training actions. Biological signals are often used as a way to measure motion information, including electroencephalogram signals, electromyogram signals, and other peripheral nerve electrical signals. Because the surface electromyogram signal has the advantages of rich information, noninvasive acquisition and the like, the surface electromyogram signal is favored by a plurality of researchers.
Currently, the main research content on intent recognition is divided into two parts: one is classification of discrete motions and recognition of motion patterns, and the other is direct decoding of neural motor intention information, such as joint angles, angular velocities, moments, and the like. For the classification problem, the research method is mature, but only a few discrete limb actions can be predicted, when the rehabilitation robot is controlled by applying the prediction result, the rehabilitation robot cannot move freely like a human joint, and the condition that the human-computer motion continuous matching is guaranteed is the premise of realizing the safety control of the rehabilitation robot, so that the effect of estimating the human joint continuous motion through the surface electromyogram signal on the rehabilitation training of a patient is positive. According to the invention, the surface electromyographic signals are utilized to identify the multi-joint angles of the lower limbs of the human body, the estimation result can be used as reference input to control the robot, the self-adaptive capacity of the robot is improved, a safe and comfortable training environment can be provided for a patient, and the development of human-computer interaction is promoted.
Disclosure of Invention
The invention discloses a lower limb multi-joint angle estimation method based on surface electromyogram signals, which provides a lower limb active movement intention identification method combining a least square support vector machine and a return-to-zero neural network by utilizing original surface electromyogram signals of a tester and actual movement angles of knee joints and hip joints of lower limbs, and realizes accurate identification of continuous movement angles of the lower limb multi-joints. In the signal acquisition process, the disturbance generated by other electronic equipment in the surrounding environment is considered, for example, bluetooth equipment, or the corresponding sensor is translated when acquiring signals, and the situation can be regarded as measurement noise and can interfere with signal acquisition, so that the movement intention identification result is influenced. Therefore, the anti-noise return-to-zero neural network model with the noise suppression capability is established, the problem that the recognition result is greatly deviated from the true value due to noise is effectively solved, the recognition precision is ensured, the influence of the noise on the solving of model parameters is overcome, and technical reference is provided for avoiding secondary damage possibly caused by the rehabilitation training of the affected limb. The technical scheme of the invention is as follows by combining the attached drawings of the specification:
a lower limb multi-joint continuous motion estimation method based on surface electromyogram signals comprises the following specific steps:
s1: in order to estimate the active movement intention of the human body, the electromyographic signals of lower limb rectus femoris and vastus lateralis muscles of a tester and the movement angle signals of knee joints and hip joints are synchronously collected;
s2: filtering the acquired original electromyographic signals of the rectus femoris and the vastus lateralis, removing noise and reserving available signals;
s3: mapping the preprocessed surface electromyographic signals into muscle activity by adopting a nonlinear exponential function, and taking the muscle activity of the rectus femoris and the vastus lateralis muscle obtained by calculation as input signals of a regression model;
s4: and constructing a regression model of the return-to-zero neural network based on a least square support vector machine to complete the angle estimation of the knee joint and the hip joint of the lower limb of the human body. The input of the model is the acquired muscle activity, the output of the model is the measured actual joint angle, and the active movement intention of the tester is identified;
s5: a return-to-zero neural network with noise suppression capability is designed, and when the interference of measurement noise in the signal acquisition process, such as disturbance generated by other electronic equipment in the surrounding environment or translation of a corresponding sensor during signal acquisition, is considered, the influence of noise on solving model parameters can be effectively suppressed, so that the identification precision is ensured.
The specific process of the step S1 is as follows:
in the invention, a Biopac system is used for collecting the electromyographic signals of the lower limbs of the human body. Surface electromyogram signals of two muscles of a rectus femoris muscle and a lateral femoris muscle of a lower limb of a human body are collected through equipment, and actual joint angle signals of the knee joint and the hip joint moving in a sagittal plane are obtained through an inertia measurement unit. The specific process of the step S1 is as follows:
s101: wiping and cleaning the skin surface areas corresponding to the rectus femoris muscles and the vastus lateralis muscles to be collected by using alcohol;
s102: pasting electrode plates on the skin surfaces corresponding to the rectus femoris muscle and the vastus lateralis muscle to be acquired, connecting signal acquisition equipment with the electrode plates, and needing two signal acquisition channels; binding the angle sensors to the thigh and the small leg of the lower limb respectively; the electromyographic signal acquisition equipment is connected with a computer through a network cable port, and signal recording is carried out at the computer end by using software matched with the equipment.
S103: the lower limbs of a tester do periodic treadmill movement on a sagittal plane, and surface electromyographic signals of the rectus femoris and the vastus lateralis muscles and angle signals of the knee joints and the hip joints are acquired in real time by an electromyographic signal acquisition device and an angle sensor.
The specific process of the step S2 is as follows:
through the step S1, the primitive surface electromyogram signals and joint angular movement angle signals of the lower limb rectus femoris and vastus lateralis muscles of the subject are acquired. The electromyographic signal data is very weak and unstable, and can also be interfered by a large amount of noise, wherein the data comprises an electromyographic signal acquisition module, skin surface sweat, temperature and the like, so the electromyographic signal acquired in the step S1 needs to be subjected to filtering and denoising processing. In addition, the sampling frequency of the electromyographic signal acquisition system is different from the sampling frequency of the angle sensor, so the electromyographic signal needs to be further sub-sampled, so that the sampling frequency of the electromyographic signal is consistent with the sampling frequency of the angle sensor.
S201: designing a 500HZ high-pass filter to remove the interference of high-frequency signals;
s202: designing a 20HZ low-pass filter to remove the interference of low-frequency signals;
s203: designing a 50HZ notch filter to remove interference of power frequency signals;
s204: the collected electromyographic signals show strong randomness on the amplitude, the absolute value operation can convert the signal amplitude into a positive value, and the contraction strength of muscles can be intuitively reflected. Sub-sampling the electromyographic signals filtered in the steps S201, S202 and S203 to make the sampling frequency of the electromyographic signals consistent with the sampling frequency of the angle sensor, wherein the specific mathematical expression is as follows:
wherein, N represents the length of the time window, and the value of N is 20 in the invention because the sampling frequency of the electromyographic signal acquisition equipment is 2000Hz, the sampling frequency of the angle sensor is 100Hz, and the difference between the two is 20 times. sEMG s And (n) is represented as electromyographic signals which are subjected to sub-sampling processing, keep the same sampling frequency with the angle sensor, and finally are subjected to normalization processing.
The specific process of step S3 is:
the muscle activity is the size of the muscle reflected by the electrical nerve stimulation, reflects the autonomous contraction strength of the muscle, is more stable than the original surface electromyogram signal, and therefore, the muscle activity is used as an input signal of a regression model.
S301: in the invention, a common nonlinear exponential function expression is used for calculating the muscle activity, and the specific mathematical expression is as follows:
wherein m is j (n) muscle activity of the jth channel, C j Is a constant term coefficient which determines the degree of non-linearity between the electromyographic signals and the muscle activity. C j Is usually set in the range of-3 to 0 (linear), and the muscle activity and the joint angle show good correlation with the change of the coefficient, the invention will C j The value of (d) is set to 0.8.b j (n) is the electromyographic signal of the j channel processed by the filter, and the specific expression is as follows:
therein, max (semG) j (n)) and min (sEMG) j (n)) are respectively expressed as the maximum value and the minimum value of the surface electromyogram signal sequence of the jth channel.
The specific process of step S4 is:
and S3, calculating to obtain the muscle activity of the two channels, and establishing a least square support vector machine regression model according to the calculated muscle activity and the actual angle of the knee joint and the hip joint of the lower limb obtained in the step S1, so as to identify the active movement intention of the subject and predict the angle of the continuously moving joint. The input of the regression model is the muscle activity, the output is the joint angle, and the nonlinear mapping relationship between the muscle activity and the joint angle can be specifically expressed as follows:
wherein the content of the first and second substances,
the nonlinear regression method is a nonlinear function, and is characterized in that a nonlinear sample of an original low-dimensional input space is mapped into a high-dimensional feature space, so that an optimal hyperplane is constructed in the high-dimensional space by a model, and a nonlinear regression problem is converted into a linear regression problem.
S401: the least square support vector machine is evolved from a support vector machine, is a machine learning method, and has the core of solving a convex optimization problem. The least squares support vector machine can be described as the following quadratic programming problem with equality constraints, and the objective function introduces a quadratic term of an error factor, which can be expressed in a specific form:
wherein i =1 \ 8230n represents the length of the training data, α is the weight vector of the training sample, b is the offset, and e represents the error variable. ζ is a penalty factor representing a coefficient for balancing the maximum interval and the minimum deviation, the higher the value of ζ, the smaller the error that can be tolerated, but easily leads to overfitting.
S402: the optimization problem described in step S401 can be solved by using a lagrangian multiplier method, and the equality constraint optimization problem is converted into an unconstrained optimization problem, where a specific expression of a lagrangian function is in the following form:
where ρ is i Expressed as a Lagrange multiplier, and the derivative of the Lagrange function to the unknown variables alpha, b, epsilon and rho is equal to 0 according to the existence condition of the extremum of the multivariate function, so that the extremum of the unconstrained optimization problem is obtained, and the optimal solution of the problem is obtained. This process can be described as the following mathematical expression:
s403: and establishing a radial basis kernel function. As can be seen from step S402, when mapping to the high-dimensional feature space by non-linearity, the dot product needs to be calculated in the high-dimensional feature space, and therefore, a kernel function method is adopted instead of the dot product operation. The introduction of the kernel function replaces the inner product of a high-dimensional space, so that the calculation amount and complexity are greatly reduced. In the invention, a Gaussian kernel function containing a parameter sigma is used for calculation, and the following mathematical expression is specifically defined:
according to the above analysis, let
The quadratic programming problem described by the least squares support vector machine can then be transformed into a problem that solves a system of linear equations in the following specific form:
s404: and establishing a return-to-zero neural network. Through steps S401, S402, and S403, the original optimization problem with equality constraints has been transformed into a problem of solving a set of linear equations, and the zeroing neural network is a novel recurrent neural network, and is mainly applied to solving the problem of time-varying quadratic programming. The least square support vector machine describes a special case condition in time-varying quadratic programming, so that the return-to-zero neural network is designed to solve the equation set and further obtain the parameters of the regression model, and the specific design thought is as follows:
firstly, the linear equation set obtained in step S403 is recorded as the following equation set form with time-varying parameters, and the specific expression is:
A(t)x(t)＝B(t)
the return-to-zero neural network is an error-based dynamic solving method, and the core of the method is to make each element in an error function converge to 0, and define a matrix type error function as follows:
e(t)＝A(t)x(t)-B(t)
in order to make the error function e (t) tend to 0, a design formula of the dynamic return-to-zero neural network is obtained, and a specific mathematical expression is as follows:
where the parameter γ is a positive value, representing the convergence rate of the nulling neural network model. By combining an error function and a design formula, a differential power system model applied to solving the time-varying quadratic programming problem can be obtained, and the specific mathematical expression is as follows:
the least square support vector machine is constructed as a static quadratic programming problem, and the return-to-zero neural network aims at solving a time-varying system, so that two ends of the equation set obtained in step S403 need to be regarded as a constant matrix and a vector which vary with time. Finally, a model for solving the equation set in step S403 is obtained, and the specific mathematical expression is:
s405: according to the step S404, an equation set for solving the step S403 is designed, so as to obtain a lagrangian multiplier ρ and an offset b, and according to a lagrangian multiplier method, a weight vector α in the regression model can be calculated, so as to obtain a final regression model, where a specific mathematical expression is:
parameters in the regression model are trained through the training samples, and then the test data are substituted into the model, so that the motion angle of the lower limb joint is estimated.
The specific process of step S5 is:
step S4 describes that the lower limb joint angle is identified under the condition that there is no measurement noise interference in the solving process. In practice, noise caused by implementation errors or external interference in hardware implementation is difficult to avoid, so that the anti-noise return-to-zero neural network is designed, the condition that large deviation exists between an angle estimation value and an actual value due to noise is overcome, and accurate identification of joint angles of lower limbs of a human body in a noise environment is realized.
S501: and designing an anti-noise return-to-zero neural network model based on the improved return-to-zero neural network model design steps. According to step S404, the model with noise suppression capability is evolved from the original return-to-zero neurodynamic system, and the specific design formula is as follows:
wherein, gamma and lambda are adjustable parameters, which respectively control the convergence rate and the noise suppression degree of the model and can ensure that the index of the error function e (t) is converged to 0. The integral term is introduced mainly to suppress noise interference. And expanding the anti-noise type return-to-zero neural network dynamic model by combining the expression of the error function to obtain a hidden neural network model expression as follows:
based on the same-way analysis in step S404, a return-to-zero neural network model with noise suppression capability can be obtained, and is used for solving a linear equation system converted by a least squares support vector machine, so as to obtain regression model parameters, where a specific mathematical expression is:
the invention utilizes surface electromyographic signals of lower limb muscles to identify the active movement intention of a human body, and provides an active movement intention identification method based on a least square support vector machine and a return-to-zero neural network, so that the estimation of the multi-joint angle of the lower limb is realized, and a good prediction effect is achieved. In the signal acquisition process, the disturbance generated by other electronic equipment in the surrounding environment is considered, for example, bluetooth equipment or corresponding sensors translate during signal acquisition, the situation can be regarded as measurement noise and can interfere signal acquisition, and further the movement intention recognition result is influenced, so that the anti-noise return-to-zero neural network is introduced, the influence of the noise on the parameter process of the training model is overcome, the accuracy of final angle estimation is ensured, and technical reference is provided for avoiding secondary damage possibly generated by rehabilitation training of the affected limb.
Drawings
FIG. 1 is a flow chart of a method and a system for estimating a multi-joint angle of a lower limb based on a surface electromyogram signal according to the present invention;
FIG. 2 is a diagram of the original surface electromyography of the lateral thigh muscle and the rectus femoris of the lower limb obtained by the device system during the signal acquisition process of the present invention.
Fig. 3 is a diagram of the actual angle change of the knee joint and hip joint of the lower limb obtained by the angle sensor in the signal acquisition process.
FIG. 4 is an electromyogram of the lateral thigh muscle and the rectus femoris of the lower limb after the signals are preprocessed by filtering, denoising, sub-sampling and the like.
FIG. 5 shows the muscle activity of the lateral femoral muscle and the rectus femoris muscle of the lower limb according to the collected surface electromyographic signals.
FIG. 6 is a lower limb knee joint angle tracking diagram estimated based on a least squares support vector machine and a return-to-zero neural network method according to the present invention.
FIG. 7 is a lower limb hip joint angle tracking diagram estimated based on the least square support vector machine and the return-to-zero neural network method.
FIG. 8 is a lower limb knee joint angle tracking error map based on a least squares support vector machine and a return-to-zero neural network method according to the present invention.
FIG. 9 is a graph of lower extremity hip joint angle tracking error based on the least squares support vector machine and the return-to-zero neural network method of the present invention.
FIG. 10 is a comparison graph of the predicted lower limb knee joint angle tracking based on the return-to-zero neural network and the anti-noise return-to-zero neural network model under the condition of noise consideration.
FIG. 11 is a comparison graph of the predicted lower limb hip joint angle tracking based on the return-to-zero neural network and the anti-noise return-to-zero neural network model under the condition of noise consideration.
FIG. 12 is a diagram of the root mean square error of the knee joint angle of the lower limb obtained by adjusting the parameters of the anti-noise type return-to-zero neural network model under the condition of noise consideration.
FIG. 13 is a diagram of the root mean square error of the hip joint angle of the lower limb obtained by adjusting the parameters of the anti-noise type zeroing neural network model under the condition of noise consideration.
Detailed Description
In order to more clearly and completely describe the estimation method and technique and the specific processing and design process of data of the present invention, the present invention will be further described with reference to the accompanying drawings, and those skilled in the art can implement the present invention according to the content of the description:
the invention discloses a surface electromyogram signal-based lower limb multi-joint angle estimation method, a system flow chart is shown in figure 1, and the method comprises the following specific steps:
s1: in order to estimate the active movement intention of the human body, electromyographic signals of rectus femoris and vastus lateralis muscles of lower limbs of a tester and movement angle signals of knee joints and hip joints are synchronously acquired;
in step S1, the human body signal collecting device is composed of a Biopac system and an inertial measurement unit. The method comprises the following steps of acquiring surface electromyographic signals of two muscles of a lower limb of a human body by using electromyographic signal acquisition equipment, wherein the surface electromyographic signals comprise thigh rectus muscles and thigh lateral muscles, angle signals comprise actual joint angle signals of knee joints and hip joints moving in a sagittal plane, and the specific process of the step S1 is as follows:
s101: wiping and cleaning the skin surface areas corresponding to the rectus femoris muscle and the vastus lateralis muscle to be collected by alcohol;
s102: pasting electrode plates on the skin surfaces corresponding to the rectus femoris muscle and the vastus lateralis muscle to be acquired, and connecting the signal acquisition equipment with the electrode plates to share two signal acquisition channels; binding the angle sensors to the thigh and the small leg of the lower limb respectively; the electromyographic signal acquisition equipment is connected with a computer through a network cable port, and signal recording is carried out at the computer end by using software matched with the equipment.
S103: the lower limbs of a tester do periodic treadmill movement on a sagittal plane, and surface electromyographic signals of the rectus femoris and the vastus lateralis muscles and angle signals of the knee joints and the hip joints are acquired in real time by an electromyographic signal acquisition device and an angle sensor.
After the signal collection process of step S1, as shown in fig. 2, the original surface electromyograms of the lateral femoral muscle and the rectus femoris of the lower limb are collected in the signal collection process of the present invention. Fig. 3 shows the real angle change of the knee joint and hip joint of the lower limb.
S2: preprocessing an electromyographic signal;
the surface electromyogram signal data acquired in step S1 is weak and unstable, and thus cannot be used directly. In step S2, a high-pass filter, a low-pass filter, and a notch filter are used to perform filtering and denoising processing on the collected electromyographic signals, and meanwhile, sub-sampling processing needs to be performed on the preprocessed electromyographic signals, so that the sampling frequency of the electromyographic signal collecting device is consistent with the sampling frequency of the angle sensor, and the specific process is as follows:
s201: designing a 500HZ high-pass filter to remove the interference of high-frequency signals;
s202: designing a 20HZ low-pass filter to remove the interference of low-frequency signals;
s203: designing a 50HZ notch filter to remove interference of power frequency signals;
s204: the acquired electromyographic signals show strong randomness on the amplitude, the signal amplitude can be converted into a positive value through absolute value operation, and the contraction strength of muscles can be visually reflected. Sub-sampling the electromyographic signals filtered in the steps S201, S202 and S203 to make the sampling frequency of the electromyographic signals consistent with the sampling frequency of the angle sensor, wherein the specific mathematical expression is as follows:
wherein, N represents the length of the time window, and the value of N is 20 in the invention because the sampling frequency of the electromyographic signal acquisition equipment is 2000Hz, the sampling frequency of the angle sensor is 100Hz, and the difference between the two is 20 times. sEMG s (n) is expressed as the sum angle after the sub-sampling processThe degree sensor keeps electromyographic signals with the same sampling frequency, and finally, the signals are subjected to normalization processing. Myoelectric signals of the vastus lateralis and rectus femoris after pretreatment are shown in fig. 4.
S3: constructing a nonlinear exponential function, and calculating the muscle activity of the vastus lateralis and the vastus rectus based on the preprocessed surface electromyogram signals;
the muscle activity is the size of the muscle reflected by the electrical nerve stimulation, reflects the autonomous contraction strength of the muscle, is more stable than the original surface electromyographic signal, is associated with the angle change of the limb joint, and is more suitable for being used as the input of a regression model. The specific process is as follows:
s301: establishing a common nonlinear exponential function to calculate the muscle activity based on the collected surface electromyographic signals, wherein the specific mathematical expression is as follows:
wherein m is j (n) muscle activity of the jth channel, C j Determining the degree of non-linearity between the electromyographic signal and the muscle activity as a constant coefficient, C j Is typically set in the range of-3 to 0 (linear). With the change of the coefficient, the muscle activity and the joint angle show different correlations, and the invention combines the results of the calculation of the coefficient C j The value of (d) is set to 0.8.b is a mixture of j (n) is the electromyographic signal of the j channel processed by the filter, and the specific expression is as follows:
therein, max (semG) j (n)) and min (sEMG) j (n)) are respectively expressed as the maximum value and the minimum value of the surface electromyogram signal sequence of the jth channel. Fig. 5 shows a muscle activity diagram of the lateral and rectus femoris muscles calculated based on the surface electromyographic signals according to the present invention.
S4: constructing a regression model of a return-to-zero neural network based on a least square support vector machine, and identifying the active movement intention of the tester;
in step S4, a least squares support vector machine regression model is established based on the muscle activity of the vastus lateralis and rectus femoris calculated in step S3 and the actual joint angles of the knee joint and hip joint calculated in step S1, so as to identify the active movement intention of the subject and estimate the continuous movement joint angle. The input signal of the regression model is muscle activity, the output signal is joint angle, and the nonlinear mapping relation of the two can be specifically expressed as:
wherein the content of the first and second substances,
the method is a nonlinear function, and a nonlinear sample of an original low-dimensional input space is mapped into a high-dimensional feature space, so that a nonlinear regression problem is converted into a linear regression problem.
S401: the least square support vector machine is improved from a support vector machine, is a machine learning method, and has the core of solving a convex optimization problem. The least squares support vector machine can be described as the following quadratic programming problem with equality constraints, and the objective function introduces a quadratic term of an error factor, which can be expressed in a specific form:
wherein i =1 \ 8230, N represents the length of the training data, α is the weight vector of the training sample, b is the offset, and e represents the error variable. ζ is a penalty factor representing a coefficient for balancing the maximum interval and the minimum deviation, the higher the value of ζ, the smaller the error that can be tolerated, but easily leads to overfitting.
S402: the optimization problem described in step S401 can be solved by using a lagrangian multiplier method, and the equality constraint optimization problem is converted into an unconstrained optimization problem, where the specific expression of the lagrangian function is as follows:
where ρ is i Expressed as a Lagrange multiplier, and the derivative of the Lagrange function to the unknown variables alpha, b, epsilon and rho is equal to 0 according to the existence condition of the extremum of the multivariate function, so that the extremum of the unconstrained optimization problem is obtained, and the optimal solution of the problem is obtained. This process can be described as the following mathematical expression:
s403: and establishing a radial basis kernel function. As can be seen from step S402, when mapping to the high-dimensional feature space by non-linearity, the dot product needs to be calculated in the high-dimensional feature space, and therefore, a kernel function method is adopted instead of the dot product operation. The introduction of the kernel function replaces the inner product of a high-dimensional space, thereby greatly reducing the calculation amount and complexity. In the invention, a Gaussian kernel function containing a parameter sigma is used for calculation, and the following mathematical expression is specifically defined:
according to the above analysis, order
The quadratic programming problem described by the least squares support vector machine can then be transformed into a problem solving a system of linear equations in the following specific form:
s404: and establishing a return-to-zero neural network. According to the steps S401, S402 and S403, the original optimization problem with equality constraint is converted into the problem of solving a group of linear equations, and the return-to-zero neural network is a novel recurrent neural network and is mainly applied to solving the problem of time-varying quadratic programming. The least square support vector machine describes a special case condition in time-varying quadratic programming, so that the return-to-zero neural network is designed to solve the equation set and further obtain the parameters of the regression model, and the specific design thought is as follows:
firstly, the linear equation set in step S403 is expressed as the following equation set form with time-varying parameters, and the specific expression is:
A(t)x(t)＝B(t)
the return-to-zero neural network is an error-based dynamic solving method, and the core of the method is to make each element in an error function converge to 0, and define a matrix type error function as follows:
e(t)＝A(t)x(t)-B(t)
in order to make the error function e (t) tend to 0, a design formula of the dynamic return-to-zero neural network is obtained, and a specific mathematical expression is as follows:
wherein the parameter γ is a positive value representing the convergence rate of the zeroing neural network model. Combining an error function and a design formula, a differential power system model applied to solving the time-varying quadratic programming problem can be obtained, and the specific mathematical expression is as follows:
the least squares support vector machine is a static quadratic programming problem, and the return-to-zero neural network aims at solving a time-varying system, so that two ends of the equation set in step S403 need to be regarded as a constant matrix and a vector which vary with time. Finally, a model for solving the equation set in step S403 is obtained, and the specific mathematical expression is:
s405: according to the step S404, an equation set for solving the step S403 is designed, so as to obtain a lagrangian multiplier ρ and an offset b, and according to a lagrangian multiplier method, a weight vector α in the regression model can be calculated, so as to obtain a final regression model, where a specific mathematical expression is:
it should be noted that, in the joint angle estimation process, the lateral femoris and the rectus femoris are selected to estimate the knee joint and the hip joint respectively, and half of the whole data set is used as a training set and the other half is used as a testing set. And solving a linear equation set converted by the least square support vector machine by adopting the designed return-to-zero neural network model so as to obtain unknown parameters of the regression model, and substituting the test set into the model to estimate the motion angle of the lower limb multi-joint. Fig. 6 and 7 show joint angle tracking diagrams of knee joints and hip joints of lower limbs predicted based on a least squares support vector machine and a zero-return neural network method, and show good performance. Fig. 8 and 9 show joint angle tracking error maps of the lower limb knee joint and hip joint according to the prediction method of the present invention.
S5: designing an anti-noise return-to-zero neural network model with noise suppression capability;
step S4 describes that the lower limb joint angle is identified under the condition that there is no measurement noise interference in the solving process. In practice, noise introduced due to implementation errors or external interference in hardware implementation is difficult to avoid, so that the anti-noise return-to-zero neural network is designed, the situation that the angle estimation value and the actual value have large deviation due to noise is overcome, and accurate identification of the joint angle of the lower limb of the human body in a noise environment is realized.
S501: and designing an anti-noise return-to-zero neural network model based on the improved return-to-zero neural network model designing step. According to step S404, the model with noise suppression capability is evolved from the original return-to-zero neurodynamic system, and the specific design formula is as follows:
wherein, γ and λ are adjustable parameters, which respectively control the convergence rate and the noise suppression degree of the model, and can ensure that the index of the error function e (t) is converged to 0. The integral term is introduced mainly to suppress noise interference. And (3) expanding the anti-noise type return-to-zero neural network dynamic model by combining the expression of the error function to obtain a concealed neural network model expression as follows:
based on the same-theory analysis in step S404, an anti-noise type return-to-zero neural network model with noise suppression capability can be obtained, and is used to solve a linear equation set converted by a least squares support vector machine, so as to obtain regression model parameters, where a specific mathematical expression is:
fig. 10 and 11 are angle tracing graphs of knee joints and hip joints of lower limbs of different models under the interference of constant measurement noise. Because the return-to-zero neural network does not have the noise suppression capability, noise has a large influence on training data, and further influences the parameter for solving the regression model, so that the final angle estimation result has a large deviation from the true value. The anti-noise return-to-zero neural network designed by the invention has the capability of noise suppression, can effectively suppress the influence of noise on the parameters of the training model, further ensures that the anti-noise return-to-zero neural network can have good recognition accuracy even in a noise environment, and provides technical reference for avoiding secondary damage possibly generated by rehabilitation training of the affected limb. Fig. 12 and 13 show root mean square error plots of the knee joint and hip joint under noise. By adjusting parameters in the anti-noise type return-to-zero neural network, the prediction performance of the model is better.
Claims (1)
1. A lower limb multi-joint angle estimation method based on surface electromyogram signals is characterized by comprising the following steps:
s1: in order to estimate the active movement intention of the tester, the electromyographic signals of the lateral femoral muscle and the rectus femoris of the lower limb of the tester and the movement angle signals of the knee joint and the hip joint are synchronously collected;
s2: preprocessing the electromyographic signals;
s3: constructing a nonlinear exponential function, mapping the preprocessed surface electromyogram signals into muscle activity, and taking the muscle activity as an input signal of a regression model, wherein the specific process is as follows:
a common nonlinear exponential function expression is applied to calculate the muscle activity, and the specific mathematical expression is as follows:
wherein m is j (n) muscle activity of the jth channel, C j Is a constant coefficient which determines the degree of non-linearity between the electromyographic signals and the muscle activity, C j Is usually set in the range of-3 to 0 (linear), with good correlation between muscle activity and joint angle as the coefficient changes, this time-C j Is set to 0.8; b j (n) is the electromyographic signal of the j channel processed by the filter, and the specific expression is as follows:
therein, max (semG) j (n)) and min (sEMG) j (n)) respectively expressed as the maximum value and the minimum value of the surface electromyogram signal sequence of the jth channel;
s4: based on a least square support vector machine, a regression model of a return-to-zero neural network is constructed, and the active movement intention of a tester is identified, wherein the specific process is as follows:
based on the obtained surface electromyographic signals and the actual joint angles, a least square support vector machine regression model is established to estimate the human lower limb multi-joint continuous motion angles, the human active motion intention is accurately identified, the muscle activity is input, the joint angles are output, and the specific mathematical expression is as follows:
wherein the content of the first and second substances,
the method is a nonlinear function, and a nonlinear sample of an original low-dimensional input space is mapped into a high-dimensional feature space, so that a nonlinear regression problem is converted into a linear regression problem; the intention recognition technology is equivalent to solving an equality constraint optimization problem described by a least squares support vector machine, and the concrete form is expressed as follows:
wherein, i =1 \ 8230, N represents the length of the training sample, alpha is the weight vector of the training sample, b is the offset, epsilon represents the error variable, and xi is the penalty factor; the Lagrange multiplier method is utilized to convert the original quadratic programming problem into a problem solved by a linear equation system, and the specific form is as follows:
wherein, K (a) i ,a j )＝exp(-||a i -a j || 2 /2σ 2 ) Expressed as a gaussian kernel function containing the parameter σ; because the dot product needs to be calculated in the high-dimensional feature space, a kernel function method is adopted to replace the dot product operation, so that the calculation amount and the complexity are reduced;
in order to obtain model parameters, identify human motion intentions, establish a return-to-zero neural network to solve the equation set, firstly, a matrix type error function is defined, and a specific expression is as follows:
e(t)＝A(t)x(t)-B(t)
in order to make the error function e (t) tend to 0, a design formula of the dynamic return-to-zero neural network is obtained, and a specific expression is as follows:
wherein, the parameter gamma represents the convergence rate of the model, and a return-to-zero neural network model for solving a linear equation set is obtained according to an error function and a design formula, so as to obtain a regression model for intention identification, and the specific expression is as follows:
in the intention recognition process, an unknown parameter in a least square support vector machine is calculated by using a return-to-zero neural network model, so that a regression model about muscle activity and joint angle is obtained, a parameter training stage is completed, test data are substituted into the regression model, the continuous motion angle of the joint of the lower limb of the human body is estimated, and the active motion intention recognition technology of the human body based on the surface electromyogram signal is realized;
s5: designing an anti-noise return-to-zero neural network model, and inhibiting the interference of noise on an intention recognition result, wherein the specific process comprises the following steps:
when the interference of measurement noise in the signal acquisition process is considered, the intention identification result is influenced, including the disturbance generated by other electronic equipment in the surrounding environment or the translation of a corresponding sensor during signal acquisition, so that the anti-noise return-to-zero neural network is established to ensure the estimation precision, and a specific design formula is as follows:
gamma and lambda are adjustable parameters, the integral term is mainly used for resisting noise interference, an anti-noise return-to-zero neural network model for solving a linear equation set is obtained by combining an error function, the interference of noise on the solving of regression model parameters is effectively inhibited, the problem that an identification result deviates from a true value due to noise is solved, the accuracy of human motion intention identification is further ensured, and the accurate identification of the motion angle of the lower limb joint of the human body in a noise environment is realized.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110978485.1A CN113520413B (en) | 2021-08-25 | 2021-08-25 | Lower limb multi-joint angle estimation method based on surface electromyogram signals |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110978485.1A CN113520413B (en) | 2021-08-25 | 2021-08-25 | Lower limb multi-joint angle estimation method based on surface electromyogram signals |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113520413A CN113520413A (en) | 2021-10-22 |
| CN113520413B true CN113520413B (en) | 2022-10-21 |
Family
ID=78092910
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110978485.1A Active CN113520413B (en) | 2021-08-25 | 2021-08-25 | Lower limb multi-joint angle estimation method based on surface electromyogram signals |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113520413B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114326389A (en) * | 2021-12-07 | 2022-04-12 | 燕山大学 | Pneumatic artificial muscle parallel mechanism angle tracking control based on time delay compensator |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105615890A (en) * | 2015-12-24 | 2016-06-01 | 西安交通大学 | Angle and myoelectricity continuous decoding method for human body lower limb walking joint |
| CN107622260A (en) * | 2017-10-26 | 2018-01-23 | 杭州电子科技大学 | Lower limb gait phase identification method based on multi-source bio signal |
| CN113052271A (en) * | 2021-05-14 | 2021-06-29 | 江南大学 | Biological fermentation data prediction method based on deep neural network |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120123232A1 (en) * | 2008-12-16 | 2012-05-17 | Kayvan Najarian | Method and apparatus for determining heart rate variability using wavelet transformation |
| EP2827769A4 (en) * | 2012-03-19 | 2015-11-25 | Univ Florida | Methods and systems for brain function analysis |
| US11052011B2 (en) * | 2016-09-02 | 2021-07-06 | Panasonic Intellectual Property Management Co., Ltd. | Standing-up assistance apparatus, standing-up assistance method, and non-transitory computer-readable recording medium |
| CN106826830B (en) * | 2017-02-23 | 2019-11-08 | 重庆交通大学 | A kind of robot control method and device |
| CN107957730B (en) * | 2017-11-01 | 2020-02-18 | 华南理工大学 | Unmanned aerial vehicle stable flight control method |
| CN109657651A (en) * | 2019-01-16 | 2019-04-19 | 杭州电子科技大学 | A kind of continuous method for estimating of lower limb knee joint based on electromyography signal |
| CN110755070B (en) * | 2019-08-28 | 2022-07-05 | 北京精密机电控制设备研究所 | Multi-sensor fusion-based lower limb movement pose rapid prediction system and method |
| CN111241982B (en) * | 2020-01-07 | 2023-04-28 | 金陵科技学院 | Robot hand recognition method based on CAE-SVM |
-
2021
- 2021-08-25 CN CN202110978485.1A patent/CN113520413B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105615890A (en) * | 2015-12-24 | 2016-06-01 | 西安交通大学 | Angle and myoelectricity continuous decoding method for human body lower limb walking joint |
| CN107622260A (en) * | 2017-10-26 | 2018-01-23 | 杭州电子科技大学 | Lower limb gait phase identification method based on multi-source bio signal |
| CN113052271A (en) * | 2021-05-14 | 2021-06-29 | 江南大学 | Biological fermentation data prediction method based on deep neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113520413A (en) | 2021-10-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111317600B (en) | Artificial limb control method, device, system, equipment and storage medium | |
| Zhang et al. | Muscle strength assessment system using sEMG-based force prediction method for wrist joint | |
| Shi et al. | Feature extraction and classification of lower limb motion based on sEMG signals | |
| Koçer | Classification of EMG signals using neuro-fuzzy system and diagnosis of neuromuscular diseases | |
| Islam et al. | Application of min-max normalization on subject-invariant EMG pattern recognition | |
| Ma et al. | Design on intelligent perception system for lower limb rehabilitation exoskeleton robot | |
| Xiao et al. | Continuous estimation of elbow joint angle by multiple features of surface electromyographic using grey features weighted support vector machine | |
| Samuel et al. | A novel time-domain descriptor for improved prediction of upper limb movement intent in EMG-PR system | |
| Ai et al. | Advanced rehabilitative technology: neural interfaces and devices | |
| Zou et al. | A multimodal fusion model for estimating human hand force: Comparing surface electromyography and ultrasound signals | |
| CN113520413B (en) | Lower limb multi-joint angle estimation method based on surface electromyogram signals | |
| CN105796091B (en) | A kind of intelligent terminal for removing electrocardiosignal vehicle movement noise | |
| Wang et al. | Research on multimodal fusion recognition method of upper limb motion patterns | |
| Borelli et al. | Spectral-distance-based detection of EMG activity from capacitive measurements | |
| Li et al. | Joint torque closed-loop estimation using NARX neural network based on sEMG signals | |
| Su et al. | Ankle joint torque prediction based on surface Electromyographic and angular velocity signals | |
| Mamikoglu et al. | Elbow joint angle estimation by using integrated surface electromyography | |
| Kato et al. | Estimating wrist joint angle with limited skin deformation information | |
| Mascolini et al. | Non-linear optimized spatial filter for single-trial identification of movement related cortical potential | |
| Ubeda et al. | Single joint movement decoding from EEG in healthy and incomplete spinal cord injured subjects | |
| Keating | Relating forearm muscle electrical activity to finger forces | |
| Li et al. | sEMG based joint angle estimation of lower limbs using LS-SVM | |
| CN111258426B (en) | Joint motion estimation method based on myoelectric-muscle sound model and unscented particle filtering | |
| Delis et al. | Development of a myoelectric controller based on knee angle estimation | |
| Kocer | Classifying myopathy and neuropathy neuromuscular diseases using artificial neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |