Lower limb motion intention recognition method based on sEMG fusion features

Abstract

The invention relates to a lower limb motion intention recognition method based on surface electromyogram (sEMG) fusion features. An sEMG feature fusion model and a genetic radial neural network model are included. The sEMG is widely applied to gait analysis and robot control, and due to the fact that the accuracy rate of intention recognition is low and development of an exoskeleton robot is limited, a genetic radial neural network algorithm based on sEMG multi-feature fusion is provided to achieve accurate recognition of eight lower limb actions. An original sEMG signal is collected through a gait experiment, and time domain, frequency domain and wavelet decomposition signal characteristic parameters of the original sEMG signal are extracted and fused. And using a principal component analysis method to carry out dimension reduction on fusion features, and inputting sEMG fusion feature vectors into a radial basis function neural network optimized by a genetic algorithm to carry out training and identification. According to the method provided by the invention, the sEMG fusion features are utilized to identify the human body lower limb actions, the accurate identification of the lower limb actions is realized in combination with a high-performance neural network algorithm, and support is provided for the application of sEMG signals to an intelligent rehabilitation robot.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology and rehabilitation medicine, relates to a movement intention recognition system, and in particular to a lower limb movement intention recognition method based on sEMG fusion features. Background Art

[0002] At present, there are more than 7 million stroke patients in China, with a disability rate of more than 80%. The cause of disability is largely related to the lack of effective rehabilitation treatment in the early stage. It is estimated that the number of people with physical disabilities will reach 25 million in 2020, and the number of disabled people is still increasing. Among them, only less than 10% of the disabled people have received rehabilitation services to varying degrees.

[0003] Intelligent lower limb exoskeleton rehabilitation robots can not only help stroke patients with rehabilitation training, but also help lower limb disabled patients resume daily activities, making the lives of stroke patients and lower limb disabled people happier. As an intelligent rehabilitation medical device, intelligent rehabilitation equipment can partially or completely replace doctors to perform rehabilitation exercises or rehabilitation treatment for patients. At present, there are two modes of lower limb rehabilitation equipment: active training and passive training. Passive rehabilitation equipment relies on mechanical force to drive patients to train, and patients cannot actively participate in the exercise process. Its therapeutic effect is worse than that of active equipment. Therefore, the research and development of active lower limb rehabilitation equipment has great practical value, and the correct recognition of lower limb movement intention is the prerequisite for active rehabilitation training. The commonly used lower limb movement intention recognition technologies are recognition based on traditional angle and film pressure sensors and recognition based on biological signals. Recognition based on traditional sensors requires the collection of physical quantities related to the movement of human lower limb joints and plantar pressure, such as joint torque, joint angle, angular velocity, plantar pressure, etc., but these sensors are prone to displacement and deformation errors when in contact with the human body, resulting in inaccurate information collection. The action recognition technology based on biological signals recognizes actions by detecting the bioelectric signals generated by the human body. Typical biological signals include surface electromyography (sEMG). Electromyography is widely used in rehabilitation medicine, human biological signal diagnosis, human motion analysis, human-computer interaction and other fields due to its advantages of simple acquisition, obvious characteristics and good stability.

[0004] The most commonly used machine learning methods for lower limb movement recognition are: support vector machine, artificial neural network, deep convolutional neural network, k-nearest neighbor algorithm, etc. The features of sEMG are generally of the following types: time domain features (such as average absolute value and number of zero crossings), frequency domain features (such as average power frequency and short-time Fourier transform, etc.), time-frequency domain features (such as wavelet transform). The recognition model composed of these classifiers and features has good performance, but there are also some shortcomings, such as a small number of prediction categories, a large number of repetitions of training models, and a long time. The main difficulties in lower limb movement recognition are: there are large differences in individual gait, poor repeatability of movements, etc. In response to the above problems, sEMG of lower limb muscles is collected, and the methods of signal preprocessing and feature extraction are studied. A radial basis function neural network (RBF) neural network classification model optimized by genetic algorithm (GNA) is proposed to improve the accuracy of lower limb sEMG movement recognition and reduce the training and recognition time of the model.

[0005] Patent CN202210091942.X discloses a method and system for human motion intention recognition for lower limb exoskeleton, which obtains a signal set of lower limb electromyographic signals and inertial signals during human gait movement, performs noise reduction processing on the electromyographic signals in the signal set in turn through Butterworth filter and variational mode decomposition, performs noise reduction processing on the inertial signals in the signal set in turn through Butterworth filter and wavelet denoising, extracts the time-frequency information of each signal in the signal set after noise reduction through continuous wavelet transform, obtains a three-dimensional color map of the corresponding signal based on the time-frequency information, performs offline classification training on the two-stream convolutional neural network through the three-dimensional color map of the preset proportion, and verifies the human motion intention recognition on the three-dimensional color map of the remaining proportion through the two-stream convolutional neural network after training. In order to achieve truly humanized and personalized rehabilitation training, the existing technology needs to be improved to achieve faster and more accurate intention recognition.

[0006] Patent CN202410659338.1 discloses a method and system for recognizing human lower limb movement intention under non-ideal conditions. This method aims at the problems of low precision and poor quality in the recognition of continuous lower limb movement intention under non-ideal conditions, and designs a method for recognizing human lower limb movement intention based on an improved neural network. The method includes the following steps: a. Collecting physiological signals of lower limbs during continuous walking; b. Analyzing the muscle and body state of the subject and performing signal preprocessing; c. Establishing an improved neural network model according to demand; d. Inputting preprocessed signal training and finding the optimal hyperparameters of the improved neural network model; e. Inputting preprocessed sEMG features to predict the knee joint angle and verify the prediction quality and accuracy of the model. However, the existing technology still uses conventional gait parameters as a data set to train the intention recognition model, which makes it difficult to achieve fast and accurate intention recognition.

[0007] Through searching the published patent documents, no published patent documents identical to the present patent application were found. Summary of the invention

[0008] The purpose of the present invention is to supplement the deficiencies of the prior art by improving the speed and accuracy of intention recognition through a lower limb movement intention recognition method based on sEMG fusion features, so that a rehabilitation robot can more scientifically and effectively assist stroke patients in rehabilitation training.

[0009] The present invention solves the technical problem by the following technical solutions:

[0010] A method for lower limb movement intention recognition based on sEMG fusion features, which includes sEMG feature fusion model and genetic radial neural network model.

[0011] Preferably, the sEMG feature fusion model collects the original sEMG signals of the lower limb muscles through gait experiments, extracts and fuses the time domain, frequency domain and wavelet decomposition signal feature parameters of the original sEMG signals, and uses principal component analysis (PCA) to reduce the dimension of the sEMG fusion signal to achieve an optimal combination of multiple feature vectors.

[0012] Preferably, the lower limb muscle acquisition objects include (1) rectus femoris: its function is to control the extension movement of the knee joint; (2) vastus lateralis: its function is to achieve knee joint extension movement; (3) medial gastrocnemius: its function is to flex the ankle and knee; (4) sartorius: its function is thigh extension, calf flexion and internal and external rotation movement; (5) tibialis anterior: its function is to achieve various movements such as dorsiflexion, abduction and eversion of the foot.

[0013] Preferably, the genetic radial neural network model uses a genetic algorithm to optimize the initial weights of the RBF neural network, and proposes a dynamic adaptive learning rate method to accelerate the convergence speed of the classifier, and finally obtains a genetic radial neural network (GNA-RBF) classifier. The improved recognition method can achieve accurate recognition of 8 lower limb movements.

[0014] Preferably, the eight lower limb movements include walking on flat ground, climbing stairs, descending stairs, sitting down, standing up, turning left, turning right and squatting. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the lower limb movement intention recognition process based on sEMG fusion features in an embodiment of the present invention;

[0016] Figure 2 Schematic diagram of the distribution of lower limb muscle groups in the embodiment of the present invention;

[0017] Figure 3 Schematic diagram of surface electrode distribution in an embodiment of the present invention;

[0018] Figure 4 This is a schematic diagram of the principal component analysis process in an embodiment of the present invention;

[0019] Figure 5 Schematic diagram of the improved GNA-RBF process in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] This embodiment uses sEMG as the signal source for rehabilitation robot motion recognition, extracts features from sEMG and inputs them into the machine learning model for training and recognition, and controls the exoskeleton rehabilitation robot to train or provide assistance to the patient based on the recognition results, so as to achieve the purpose of active training or active movement. In view of the problems existing in the recognition of lower limb movement intentions, such as low recognition accuracy due to large differences in individual gait and poor repeatability of movements. This embodiment proposes a method for lower limb movement intention recognition based on sEMG feature fusion. By collecting sEMG of lower limb muscles, different features are extracted for fusion, and an improved GNA-RBF model is used to realize the recognition of 8 types of lower limb movements (walking on flat ground, going up stairs, going down stairs, sitting down, standing up, turning left, turning right and squatting). The recognition process is as follows: Figure 1 shown.

[0021] The experimental subjects included 15 young subjects, aged 24 to 26 years old, 175 to 185 cm in height, and 65 to 75 kg in weight. Before the experiment began, the subjects were informed of the 8 experimental movements and movement requirements. All subjects were familiar with the experimental movements and experimental procedures before the collection, and performed 3 to 5 minutes of warm-up and stretching exercises. None of the subjects had undergone lower limb muscle, knee, and ankle surgery before the test, and all expressed their willingness to participate in this experiment. 8 channels were selected to collect sEMG of 8 muscles in the left and right legs, respectively. Channels 1 to 8 correspond to the subjects' left gastrocnemius, right gastrocnemius, left rectus femoris, right rectus femoris, left semitendinosus, right semitendinosus, left vastus lateralis, and right vastus lateralis, respectively. Figure 2 Attach the dedicated electrode sheet to the muscle belly, and the direction of the electromyographic sheet should be the same as the direction of the muscle fibers. Figure 3 shown.

[0022] 15 subjects completed 8 actions: walking on flat ground, climbing stairs, descending stairs, turning left, turning right, standing up, sitting down and jogging. Each action was repeated 10 times in a group, and each action was performed 6 times. After completing each action, they rested for 5 minutes. In order to set up a comparative experiment, 3 of the subjects were subjected to the same experiment and data was collected one week later.

[0023] sEMG is a typical random non-stationary signal, and it is sensitive and easily disturbed by external noise, electrodes and power lines. These pollution sources will affect the classification results of the model. The sEMG signal is denoised using wavelet transform technology (DWT). Wavelet transform technology can effectively decompose the time and frequency components, decompose the signal into different frequency domains, and filter the signal using the wavelet frequency domain energy threshold method.

[0024] DWT is a method of analysis in the time and frequency domains, which overcomes the disadvantage of Fourier transform losing time information while retaining time and frequency domain information. Its expression is:

[0025] Where: S(t) represents the analyzed signal, * represents the complex conjugate, Ψ represents the wavelet function, and WT(a,b) represents the decomposed wavelet coefficients

[0026] Due to the complexity of sEMG, the correct selection of signal features is a key step in classification. Time domain features do not require complex calculations and are highly real-time. Traditional frequency domain features are based on the assumption of stationary signals, while sEMG is a non-stationary time-varying signal. Traditional frequency features will lose information about the signal time. Wavelet decomposition signals have information in both the time domain and the frequency domain, and the information they show is more comprehensive and rich. However, a large amount of information will affect the recognition efficiency. This study uses principal component analysis (PCA) to quantitatively analyze data components, replacing complex original variables with fewer main components, and improving recognition speed without reducing recognition accuracy.

[0027] PCA is a multivariate dimensionality reduction method, which is often used to study the correlation between multiple variables. This paper uses principal component analysis to perform linear analysis on the time domain and time-frequency domain feature variables of sEMG, and selects feature variables with weak linear correlation to replace complex original features. The calculation steps of this method are as follows:

[0028] Step 1: Assume that all data has m variables and each feature has n data variables. Calculate the correlation coefficient matrix of the feature set, as shown in formula (2):

[0029] Among them, r ij is the feature variable X i With X j The correlation coefficient is as follows:

[0030] Step 2: Calculate the eigenvalues ​​and eigenvectors according to formula (4). |λI-R|=0 Formula (4)

[0031] Among them, I is the unit vector and R is the correlation coefficient matrix.

[0032] According to formula (4), the eigenvalue λ is solved l , principal component y i The calculation of is shown in formula (5):

[0033] Step 3: Calculate the principal component contribution rate.

[0034] According to equations (6) and (7), the individual contribution rates and cumulative contribution rates are calculated, and the first n principal components with cumulative contribution rates greater than 90% are selected to replace the original features.

[0035] Step 4: Process the features using principal component analysis. Calculate the contribution of the time domain and wavelet coefficient features respectively, further obtain the principal components, and finally determine the features after dimensionality reduction. The PCA flowchart is as follows: Figure 4 shown.

[0036] RBF neural network is a feedforward neural network consisting of input layer, hidden layer and output layer. The transformation from input node to hidden layer node is nonlinear, while the transformation from hidden layer to output is linear, and the activation function is Gaussian kernel function. Compared with BP neural network, RBF neural network does not need to set a specific number of neurons, but can automatically increase the number of neurons to meet the accuracy, with fast learning efficiency and high recognition accuracy. However, the number of neurons in the hidden layer is difficult to determine, and the learning ability of weights from hidden layer to output layer is low. These defects easily lead to a decrease in learning ability and recognition accuracy. To solve this problem, this study proposes a new optimized RBF neural network algorithm based on genetic algorithm (GNA-RBF), which uses genetic algorithm to optimize the weights and structure of RBF neural network.

[0037] The basic steps of the GNA-RBF neural network algorithm are described as follows:

[0038] Step 1: Set the RBF neural network according to the maximum number of neurons in the hidden layer, use the k-clustering algorithm to obtain the basis function center, and use formula (8) to calculate the width of the center.

[0039] Among them, d is the distance between each neuron, and s is the maximum number of hidden neurons in the RBF neural network.

[0040] Step 2: Set the genetic algorithm parameters, population size, crossover rate, mutation rate, selection mechanism, crossover operator and mutation operator, objective function error, and maximum number of iterations.

[0041] Step 3: Randomly initialize the population P, the population size is N, and the network corresponding to each individual is encoded by formula (9). c 1 c 2 …c s ω 11 ω 21 …ω s1 ω 12 ω 22 …ω s2 …ω 1m ω 2m …ω sm θ 1 θ 2 …θ m (9)

[0042] Among them, the number of hidden layer neurons is encoded using binary coding method, expressed as c i , when c i When it is 1, it means that this neuron exists, c i A value of 0 means that the neuron does not exist. ij is the connection weight from the i-th output neuron to the j-th hidden neuron, θ j is the threshold of the jth output neuron, and m is the number of output neurons.

[0043] Step 4: Use training samples D to train the initially constructed RBF neural network, the number of which is N, and use formula (10) to calculate the network training error E.

[0044] Among them, t k is the true value, y k is the output value of the neural network. The optimal neural network is obtained by iterating equation (10).

[0045] Step 5: According to the training error E and the number of hidden layer neurons s, use formula (11) to calculate the fitness of the chromosome corresponding to each network. The fitness function is:

[0046] Where C is a constant, s max It is the upper limit of the number of hidden layer neurons. The smaller the network scale (the fewer the number of hidden layer neurons), the smaller the training error and the higher the fitness of the chromosome.

[0047] Step 6: Sort the chromosomes according to their fitness values; select the best fitness of the population, defined as F b , if E <E min , go to step 9, otherwise go to step 7;

[0048] Step 7: Select several best individuals to generate a new next generation population, defined as NewP;

[0049] Step 8: Select a pair of chromosomes for single-point crossover to generate two new individuals as members of the next generation. Repeat this process until the new generation reaches the maximum population size P. s ;

[0050] Step 9: Make the new generation of population mutate, use different mutation strategies for the binary coding part and the real number coding part, then generate a new population, take P = NewP, and return to step 4;

[0051] Step 10: Get the optimal neural network structure and terminate the iteration of the genetic algorithm, that is, stop the optimization.

[0052] The advantages and beneficial effects of the present invention are:

[0053] Two advantages and beneficial effects of the present invention are as follows:

[0054] 1. Improve the accuracy of lower limb movement intention recognition: By adopting the genetic radial neural network (GNA-RBF) algorithm based on sEMG multi-feature fusion, the present invention can effectively identify eight different lower limb movements. This algorithm combines multiple feature parameters such as time domain, frequency domain and wavelet decomposition signals, and performs feature dimension reduction through principal component analysis (PCA), thereby improving the accuracy of recognition. This is particularly important for the development of exoskeleton robots, because the improvement of accuracy can significantly enhance the practicality and reliability of exoskeleton robots, thereby promoting their widespread application in fields such as rehabilitation medicine.

[0055] 2. Supporting the application of intelligent rehabilitation robots: The method proposed in this invention uses sEMG fusion features to identify human lower limb movements, and combines high-performance neural network algorithms to achieve accurate recognition of lower limb movements. This plays an important supporting role in the design and development of intelligent rehabilitation robots, because it provides an effective technical means to achieve human-computer interaction, so that rehabilitation robots can respond to patients' movement intentions more accurately, thereby providing more personalized and effective rehabilitation training.

Claims

1. A method for lower limb movement intention recognition based on sEMG fusion features, characterized in that: The original sEMG signals of lower limb muscles in different movements are collected, and the signal features in time domain, frequency domain and wavelet decomposition are extracted. The principal component analysis method is used to reduce the dimension of the sEMG fusion signal to achieve the optimal combination of multiple feature vectors. A new genetic radial neural network model is proposed to achieve accurate recognition of eight lower limb movements.

2. The method for identifying lower limb movement intention based on sEMG fusion features according to claim 1, characterized in that: The sEMG feature fusion model described above conducts sEMG signal acquisition experiments and uses wavelet basis functions to denoise the signal. The sEMG signal time domain, frequency domain and wavelet decomposition signal features are extracted. To address the problem of low accuracy in action intention recognition using a single feature, the principal component analysis method is used to reduce the dimension of the feature and select components with high contribution rates as recognition features.

3. The method for identifying lower limb movement intention based on sEMG fusion features according to claim 1, characterized in that: The genetic radial neural network algorithm model uses a genetic algorithm to optimize the initial weights of the radial basis neural network to address the shortcomings of the radial basis neural network, such as long training time and easy falling into local optimal solutions, and proposes a dynamic adaptive learning rate method to accelerate the convergence speed of the classifier, ultimately obtaining a genetic radial neural network algorithm classifier.

Images