Human-robot collaborative control method based on human skill learning and imitation

Abstract

The present invention provides a human-robot collaborative control method based on human skill learning and imitation, comprising the following steps: a human tutor drags a robot to move freely, and collects electromyographic signals and joint angle information of the human tutor's arm muscle groups and real arm strength in the process as training samples of a three-dimensional arm strength estimation model; extracts amplitude information and tremor information in the electromyographic signals; performs data fusion on the amplitude information and joint angle information to obtain force-related information in the three-dimensional motion direction; constructs a three-dimensional arm strength estimation model with the help of a parallel long short-term memory neural network to which an electromyographic signal correction unit and an input-output control unit are added; establishes a regression relationship with arm strength as input and motion speed as output to complete the construction of a human collaborative skill imitation model; converts the three-dimensional arm strength into a speed adjustment amount of the robot, thereby controlling the robot to cooperate with humans to complete collaborative tasks, so that the collaborative system has more flexible interactive performance.

Description

Technical Field

[0001] The present invention belongs to the field of human-machine interaction and artificial intelligence technology, and specifically relates to a human-robot collaborative control method based on human skill learning and imitation. Background Art

[0002] In the past few decades, robotics has played an increasingly important role in industrial development. However, with the user's demand for personalized and flexible manufacturing, human-robot collaborative operation has received widespread attention. In some collaborative scenarios, such as collaborative assembly, there are often small gaps and unstructured features, so machine vision technology cannot be fully utilized. The interactive force information generated by the human body during the collaborative process has a two-way feedback capability, which is very suitable as a medium for human-robot interaction. At present, the arm strength of the human body is often measured by installing a force sensor at the end of the robot. However, this method is not completely suitable for human-robot collaborative scenarios. For example, in the process of human-robot collaboration, the gravity of the object being moved or the contact force between the assembly parts will change the value of the force sensor, so the real human arm strength information is difficult to distinguish.

[0003] Researchers have found that it is convenient and efficient to obtain arm strength by collecting muscle activity levels through electromyographic sensors, and this method can ensure the safety and smoothness of human-robot interaction. However, how to accurately estimate muscle strength has always been the core issue of this research. With the development of computer technology, a large number of non-parametric models based on machine learning algorithms have been applied to research in this field, such as FOS model, PCI model, long short-term memory (LSTM) model, convolutional neural network (CNN) model, etc. However, most of these models currently focus on estimating one-dimensional arm strength. Limited by the arm strength model, the current human-robot collaborative system based on electromyographic signals is only suitable for performing some simple tasks, such as sawing and screwing. Therefore, it is of great research value to extend the estimation of arm strength to three-dimensional motion scenes and improve the versatility of human-robot collaborative systems.

[0004] Clinical experiments have found that human output force is inaccurate and unstable, and the output force always has a certain degree of involuntary tremor. The tremor effect of this muscle will be reflected in the amplitude information of the electromyographic signal. The interactive information between humans and robots should be stable and valuable. The non-stationarity of electromyographic signals is not conducive to the collaborative completion of precise tasks between humans and robots. Therefore, in order to improve the accuracy of arm force estimation, it is essential to eliminate the volatility of the amplitude information of the electrical signal, especially the signal fluctuation caused by the tremor effect.

[0005] In addition, with the development of human-robot collaboration technology, researchers have realized that human power information carries rich motion information and operation skills. Transferring human skills to robots will help improve the efficiency of human-robot collaboration systems. Peternel et al. proposed a multimodal robot teaching framework to learn the movement amplitude and frequency of human mentors during sawing, allowing robots to autonomously complete collaborative sawing tasks with humans (L. Peternel, T. Petric, E. Oztop, J. Babic, Teaching robots to cooperate with humans in dynamic manipulation tasks based on multi-modal human-in-the-loop approach, Auton. Robot. 36 (2014) 123-136.). Dong et al. used dynamic primitive language to transmit the coupling information of the desired trajectory and stiffness to the robot through the teaching demonstration of the human tutor, realizing the transfer and reproduction of human-robot impedance adaptive skills (JLDong, WYSi, CGYang, ADMP-based Online Adaptive StiffnessAdjustment Method, in: IECON 2021–47th Annual Conference of the IEEE Industrial Electronics Society, 2021, pp. 1-6). However, this type of method only focuses on imitating human operating skills, and does not fully consider the direct interaction between humans and robots, lacking the flexibility and autonomy of robot control. The combination of skill imitation and direct interaction helps to closely connect the operator and the robot, allowing the robot to better understand and execute human movement intentions. Summary of the invention

[0006] The purpose of the present invention is to provide a human-robot collaborative control method based on human skill learning and imitation, aiming to transfer the human skill of adjusting the following speed by sensing the change of external force to the robot, so that the robot has a more flexible collaborative ability.

[0007] The present invention is achieved by at least one of the following technical solutions.

[0008] A human-robot collaborative control method based on human skill learning and imitation includes signal acquisition and processing, construction of a three-dimensional arm strength estimation model, construction of a human collaborative skill imitation model, and human-robot collaboration, which correspond to the following steps respectively:

[0009] Step 1: The human instructor drags the robot to move freely by applying arm forces of different magnitudes and directions, and collects the arm movement signals (including the electromyographic signals of the arm muscle groups and the joint angle information) and the real arm strength of the human instructor in the process as training samples for the three-dimensional arm strength estimation model; and uses a root mean square filter to extract the amplitude information in the electromyographic signal, and uses a fast Fourier transform to extract the tremor information in the electromyographic signal, thereby completing signal acquisition and processing;

[0010] Step 2: Using the fast orthogonal search (FOS) method, the obtained amplitude information and joint angle information are fused to obtain force-related information in the three-dimensional motion direction; using the obtained amplitude information, tremor information and force-related information as input, and the real arm strength as output, a three-dimensional arm strength estimation model is constructed with the help of a parallel long short-term memory (LSTM) neural network with an electromyographic signal correction unit and an input-output control unit added;

[0011] Step 3: Using a three-dimensional arm strength estimation model and an angle sensor to obtain arm strength information and speed information during the human collaborative demonstration process, a regression relationship is established with arm strength as input and movement speed as output, thereby completing the construction of a human collaborative skill imitation model;

[0012] Step 4: Based on the constructed human collaborative skill imitation model, the estimated three-dimensional arm force is directly converted into the speed adjustment of the robot, thereby controlling the robot to cooperate with humans to complete collaborative tasks.

[0013] Preferably, in step 1, the electromyographic signals of the arm muscle group include electromyographic signals of the front end of the deltoid muscle, the rear end of the deltoid muscle, the biceps brachii, the triceps brachii, the pectoralis major and the infraspinatus. Among them, the front end of the deltoid muscle and the rear end of the deltoid muscle are a pair of antagonistic muscles, responsible for the arm strength estimation in the X-axis direction; the biceps brachii and the triceps brachii are a pair of antagonistic muscles, responsible for the arm strength estimation in the Y-axis direction; the pectoralis major and the infraspinatus are a pair of antagonistic muscles, responsible for the arm strength estimation in the Z-axis direction. In the experiment, the human instructor faces the YZ plane of the robot base coordinate system, and the division of arm strength and movement direction is based on the robot base coordinate system.

[0014] Preferably, in step 1, the joint angle information includes elbow joint angle information and shoulder joint angle information.

[0015] Preferably, in step 1, the tremor information is the average amplitude of the electromyographic signal of 4-12 Hz, which is expressed as:

[0016]

[0017] In the formula, E FFT For tremor information, FE rawis the bilateral power density spectrum of the time domain signal, which is obtained by fast Fourier transform; N is the number of sample data; n min and n max They represent the double-sided power density spectrum serial numbers corresponding to frequencies of 4 Hz and 12 Hz respectively.

[0018] Preferably, in step 2, the FOS method uses the correlation coefficient between the extracted signal and the actual output arm force as an iteration standard to extract the force-related information of the electromyographic signal without the influence of joint rotation. The expression of the iteration standard is:

[0019]

[0020] In the formula, Cov(F y ,E y ) is the covariance of the force-related information Ey and the measured arm force Fy; Var[F y ] is the arm strength F y Variance of Var[E y ]For relevant information y The variance of .

[0021] Preferably, in step 2, the electromyographic signal correction unit adopts a neural network structure based on the principle of a low-pass discrete filter.

[0022] E ES [n] = E ES [n-1]+σ(E FFT W E +b E )(E RMS [n]-E ES [n-1]);

[0023] In the formula, E ES [n] is the electromyographic signal after the nth data correction; W E and b E is the weight and bias of the EMG signal correction unit; σ(·) is the Sigmoid function; E RMS [n] is the amplitude information of the nth electromyographic signal processed by RMS filtering.

[0024] Preferably, the input-output control unit uses a naive Bayes algorithm to obtain the direction of human movement intention, which is expressed as follows:

[0025]

[0026] In the formula, E RMS1 ,E RMS2 ,……,E RMS6 are samples to be classified, which are the amplitude information of electromyographic signals after RMS filtering of six muscle groups; kIt is the direction of human movement intention, which is divided into four characteristic attributes, namely no output force, X-axis force, Y-axis force and Z-axis force. When the forces in the three directions are all less than 4N, it is considered as no output force; P(E RMSi |y k ) is the probability of each classification sample when the characteristic attribute occurs; P(y k ) is the conditional probability of each characteristic attribute occurring; P(E RMSi ) is the probability of occurrence of classified samples; yt is the direction of human movement intention at the current moment.

[0027] Preferably, in step 2, the parallel LSTM neural network uses three sub-LSTM neural networks with parallel structures to estimate the three-dimensional arm strength respectively.

[0028] Preferably, in step 3, the human collaborative skill imitation model adopts a multi-model Gaussian process regression algorithm based on K-means clustering to obtain the regression relationship between arm strength and speed adjustment amount (that is, the arm strength and movement speed are automatically classified using the K-means clustering method, and multiple Gaussian process regression models are used to fit the classification results to obtain the regression relationship between arm strength and speed adjustment amount).

[0029] Preferably, in step 3, the human collaborative skill imitation model is the speed adjustment skill of humans when facing changes in force information.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The present invention uses the human arm strength obtained by human motion information such as electromyographic signals as a medium for human-robot interaction. By learning the speed adjustment skills of humans in the face of changes in external force information, the robot is controlled to complete collaborative tasks. A parallel LSTM neural network with an electromyographic signal correction unit based on tremor information and an input-output control unit based on Naive Bayes is used to estimate the three-dimensional arm strength, thereby reducing the impact of electromyographic signal fluctuations on the accuracy and stability of interactive information. A multi-model Gaussian process regression algorithm is used to capture collaborative skills in different scenarios from inaccurate human demonstration samples in a probabilistic estimation manner, so that the collaborative system has more flexible interactive performance.

[0032] The present invention can directly convert the myoelectric signal of the human body into the speed adjustment of the robot by using the human skill imitation model, effectively avoiding the difficulty of control parameter selection in the traditional control model and showing good coordination ability between precise tracking and comfortable cooperation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The invention provides a framework for a human-robot collaborative control method based on human skill learning and imitation.

[0034] Figure 2 Schematic diagram of a signal acquisition platform according to an embodiment of the present invention.

[0035] Figure 3 Schematic diagram of the structure of a three-dimensional arm strength estimation model according to an embodiment of the present invention.

[0036] Figure 4 A schematic diagram of a human collaboration demonstration according to an embodiment of the present invention.

[0037] Figure 5 This is a control flow chart of a human-robot collaborative control method based on human skill learning and imitation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] Humans have super perception and decision-making abilities. In human collaboration, followers can adjust the speed and position of following by sensing changes in external forces. Transferring this kind of human collaborative skills to robots will help human-robot collaborative technology develop in the direction of intelligence. However, due to the limitations of the arm force estimation model and human-robot collaborative control methods, the current human-robot collaborative system based on electromyographic signals is only suitable for performing some simple tasks, such as sawing and screwing.

[0039] The present invention adds an electromyographic signal correction unit based on tremor information and an input-output control unit based on naive Bayes to the original long short-term memory neural network, and uses a parallel network structure to estimate the three-dimensional arm strength. A multi-model Gaussian process regression algorithm is used to capture the collaborative skills of different scenarios from inaccurate human demonstration samples in a probabilistic estimation manner, so that human-robot collaborative control can take into account both accuracy and speed. The human-robot collaborative control method proposed in the present invention is an artificial intelligence control method that can directly convert the arm movement information of the human instructor into the speed adjustment amount of the robot, avoiding the difficulty of control model selection in traditional control. In addition, the present invention realizes the accurate estimation of three-dimensional arm strength, and more human-robot collaborative tasks can be achieved with the help of this force estimation model.

[0040] In order to better understand the present invention, the present invention is further described below in conjunction with the accompanying drawings.

[0041] Example 1

[0042] like Figure 1 As shown, the human-robot collaborative control method based on human skill learning and imitation includes the following steps:

[0043] 1) Signal acquisition and processing

[0044] The human instructor drags the robot to move freely by exerting arm forces of different magnitudes and directions, and collects the arm movement signals of the human instructor (including the electromyographic signals of the arm muscle groups and the joint angle information) and the actual arm strength in the process as training samples for the three-dimensional arm strength estimation model.

[0045] 1-1) If Figure 2 As shown in the figure, the goniometer is placed on the outside of the elbow joint to collect one-dimensional elbow angle information; the inertial measurement unit (IMU) is placed on the inner lower end of the upper arm to collect and measure the three-dimensional shoulder angle information; six electromyographic sensors are placed on the front end of the deltoid muscle, the rear end of the deltoid muscle, the biceps brachii, the triceps brachii, the pectoralis major and the infraspinatus muscle to collect the electromyographic signals of each muscle group. The six-dimensional force sensor is used to collect the real arm strength. In the experiment, the human instructor faces the YZ plane of the robot base coordinate system, and the division of arm strength and movement direction is based on the robot base coordinate system. The front end of the deltoid muscle and the rear end of the deltoid muscle are a pair of antagonistic muscles, responsible for the estimation of arm strength in the X-axis direction; the biceps brachii and the triceps brachii are a pair of antagonistic muscles, responsible for the estimation of arm strength in the Y-axis direction; the pectoralis major and the infraspinatus are a pair of antagonistic muscles, responsible for the estimation of arm strength in the Z-axis direction.

[0046] 1-2) The original electromyographic signal cannot be used directly to estimate arm strength, because the electromyographic signal is time-varying and nonlinear. It is necessary to extract the features related to arm strength from the time-varying electromyographic signal. The amplitude information of the electromyographic signal is extracted using a root mean square filter, and its expression is:

[0047]

[0048] In the formula, E raw [m] is the mth original electromyographic signal; E RMS [i] is the amplitude information of the i-th electromyographic signal processed by RMS filtering; N1 is the sliding window length.

[0049] 1-3) The tremor of the EMG signal is related to its frequency, especially the power spectrum of 4-12 Hz. Therefore, the power spectrum of the EMG signal in the original EMG signal was extracted by fast Fourier transform, and the average amplitude of the 4-12 Hz EMG signal was calculated.

[0050] Parseval's theorem states that the energy of a signal in the time domain and in the frequency domain is equal, and the relationship between the two is as follows:

[0051]

[0052] Where FE raw [k] is the kth value of the bilateral power density spectrum of the time domain signal, which is obtained by fast Fourier transform; N is the amount of sample data.

[0053] According to formula (2), the average amplitude of 4-12 Hz can be derived, that is, the tremor information E FFT for,

[0054]

[0055] Where n min and n max They represent the double-sided power density spectrum serial numbers corresponding to the frequencies of 4 Hz and 12 Hz in the power density spectrum respectively.

[0056] 2) Estimation of three-dimensional arm strength

[0057] In the process of three-dimensional arm strength analysis, the electromyographic signals of each muscle group are complicated. If the prediction of strength and direction is completely based on the neural network, the complexity of the network model will undoubtedly increase, and the amount of training data required is also huge. Therefore, the present invention first determines the direction of human intention based on the characteristic information of the electromyographic signal, obtains the corresponding intention direction, and then estimates the strength value based on the electromyographic signal in the direction related to the direction.

[0058] 2-1) Prediction of human intention direction based on Naive Bayes

[0059] The amplitude information of the electromyographic signals of the six muscle groups after RMS filtering is used as samples to be classified. At this time, the sample set to be classified is E = {E RMS1， E RMS2， …,E RMS6 The main output force direction (the main output force direction is the direction of the maximum output force, and in the present invention, the maximum output force direction is regarded as the direction of human intention) is divided into four characteristic attributes, namely no output force, X-axis force, Y-axis force and Z-axis force. When the forces in the three directions are all less than the minimum resolution threshold of the three-dimensional arm strength estimation model, it is regarded as no output force.

[0060] Since the electromyographic signals of each muscle group are collected independently and do not interfere with each other, the Bayesian theorem is satisfied. At this time, the human intention direction prediction model is:

[0061]

[0062] In the formula, E RMS1 ,E RMS2 ,……,E RMS6 are the amplitude information of the electromyographic signals after RMS filtering of the six muscle groups, which are used as samples to be classified; i is the direction of human movement intention, which is divided into four characteristic attributes: no output force, X-axis force, Y-axis force and Z-axis force; P(E RMSi |y i) is the probability of each classification sample when the characteristic attribute occurs; P(y i ) is the conditional probability of each characteristic attribute occurring; P(E RMSi ) is the probability of occurrence of classified samples; yt is the predicted direction of human movement intention at the current moment.

[0063] 2-2) Force-related information model

[0064] The FOS algorithm is used to obtain the force-related information model in each direction. The advantage of this method is that it can remove the influence of joint angle on arm force estimation and reduce the fluctuation of electromyographic signals through deep learning algorithms.

[0065] At this time, taking the arm strength in the Y direction as an example, the Y direction force related information model based on the FOS algorithm is:

[0066] E y (E R [n],θ[n])=(E R1 [n]-E R2 [n])-E c (E RMS [n],θ[n]); (5)

[0067] In the formula, E y (E R [n],θ[n]) is the force related information model; E c (E R [n],θ[n]) is the joint rotation compensation model; E R [n] is the amplitude information of the nth group of electromyographic signals after RMS filtering. The muscle groups involved in the Y direction include the pectoralis major and infraspinatus muscles; E R1 [n] and E R2 [n] is the amplitude information of the electromyographic signal of the nth pectoralis major and infraspinatus muscle after RMS filtering; θ[n] is the value of the nth group of joint rotation angles, including the shoulder joint angle and the elbow joint angle.

[0068] Gram-Schmidt is used to orthogonalize the joint rotation compensation model.

[0069]

[0070] Where gm is the orthogonal basis function q m (n) is the coefficient of the polynomial; M is the number of terms in the polynomial.

[0071] The correlation coefficient between the extracted force-related information and the actual output arm force is used as the standard, and the quasi-Newton method is used to search for the orthogonal basis function q m (n), and calculate the optimal coefficient gm, thereby completing the construction of the force-related information model.

[0072] The iteration criterion can be expressed as,

[0073]

[0074] In the formula, Cov(F y ,E y ) is the covariance of the force-related information Ey and the measured arm force Fy; Var[F y ] is the arm strength F y Variance of Var[E y ]For relevant information y The variance of .

[0075] 2-3) Construction of a three-dimensional arm strength estimation model based on parallel LSTM neural network

[0076] (1) EMG signal correction unit

[0077] The average amplitude of the EMG signal at 4-12 Hz is closely related to the fluctuation of the EMG signal. FFT As a regulating parameter to reduce the fluctuation of electromyographic signal, a discrete low-pass filter with a neural network unit structure is proposed to adjust the amplitude of the electromyographic signal E after the root mean square filtering. RMS [n] is corrected. The structure of the electromyographic signal correction unit is as follows:

[0078] E ES [n] = E ES [n-1]+σ(E FFT W E +b E )(E RMS [n]-E ES [n-1]); (8)

[0079] In the formula, E ES [n] is the electromyographic signal after the nth data correction; W E and b E is the weight and bias of the data correction unit; σ(·) is the Sigmoid function; E RMS [n] is the amplitude information of the nth electromyographic signal processed by RMS filtering.

[0080] (2) Input and output control unit

[0081] like Figure 3 As shown, the input-output control unit is responsible for controlling the data input and output of the activation main motion direction and selecting the training model. Its activation direction is the direction of human intention predicted by naive Bayes.

[0082] (3) Sub-LSTM neural network

[0083] By using three parallel sub-LSTM neural networks to train the data model separately, the model structure is simplified and the stability of its operation will be improved. In the improved neural network, in addition to adding the electromyographic signal correction unit and the input and output control unit, the sub-LSTM neural network structure still uses the traditional LSTM neural network structure.

[0084] 3) Construction of human collaborative skills imitation model

[0085] Taking human-robot collaborative shaft-hole assembly as an example, the construction method of the human collaborative skill imitation model in the present invention is introduced. Figure 4 As shown, during the human collaborative assembly process, the leader leads the collaborators to complete the shaft-hole assembly, and the follower only moves in the direction of the greatest force felt, collecting the arm strength and movement speed information of the human mentor during the collaborative process.

[0086] Two local optimal models (fast model and slow model) are used to fit human collaboration skills, taking into account both "fast" and "accurate" performance. The collaboration information is divided into two clusters using the K-means clustering algorithm, and the center point of each cluster is found through iterative learning.

[0087] According to the clustering results, the clustered cooperation information and speed adjustment information are stored in different sample spaces respectively, and a multi-model Gaussian process regression model is constructed. Suppose the training sample of a clustering space model is The Gaussian process regression algorithm is used to train the samples Fitting is performed to obtain the speed adjustment skill f(F i ,dF i ), where F i is the arm strength, dF i is the differential of arm strength. Since the sample data after clustering processing obeys multi-dimensional Gaussian distribution, the speed regulation skill model can be expressed as:

[0088] f(F f ,dF f )~GP(μ(F f ,dF f ),k([F f ,dF f ] i ,[F f ,dF f ] λ )); (9)

[0089] In the formula, μ(F f ,dF) is the mean function; k([F i ,dF i ],[Fj ,dF j ]) is the sample point [F i ,dF i ],[F j ,dF j ]’s covariance.

[0090] According to the properties of the Gaussian process, for the predicted sample point [F t+1 ,dF t+1 ], there exists a joint Gaussian distribution as,

[0091]

[0092] In the formula, v 1:t represents the observed human motion speed set {v1, v2, …, vt}; f t+1 is the predicted human body speed value. K represents the covariance matrix of the known samples; k represents the predicted sample point [F t+1 ,dF t+1 ] and the covariance matrix of the remaining samples.

[0093] At this time, f t+1 The posterior probability of is,

[0094]

[0095] μ(F t+1 ,dF t+1 ) is the robot adjustment speed obtained from the human mentor's power information.

[0096] 4) Human-Robot Collaboration

[0097] like Figure 5 As shown, in the speed adjustment mode, the robot can convert the acquired three-dimensional arm force into the robot speed adjustment amount according to the constructed human collaborative skill imitation model, and control the robot to cooperate with humans to complete collaborative tasks.

[0098] Example 2

[0099] Compared with Example 1, the difference between this embodiment and Example 1 is that the input-output control unit uses a naive Bayesian algorithm with a threshold to identify the movement direction of the human instructor. b , reduce the prediction errors caused by the noise of the samples to be classified, when the maximum posterior probability is greater than the probability threshold p b The predicted direction of human motion intention is updated only when . The expression of the naive Bayes algorithm with threshold is:

[0100]

[0101] In the formula, pb is the probability threshold; y t is the predicted direction of human movement intention at the current moment; t-1 It is the direction of human movement intention at the previous moment.

[0102] Example 3

[0103] Compared with Example 1, the present embodiment is different in that the input-output control unit uses an RBF neural network to identify the movement intention direction of the human instructor, and the structure of the RBF neural network is:

[0104]

[0105] In the formula, E={E RMS1 , E RMS2 ,…,E RMS6} is the sample set to be classified, E RMSi The amplitude information of the electromyographic signal after RMS filtering of each muscle group; Ec i is the cluster center vector of the electromyographic signal of each muscle group; σ i is the width vector of the hidden layer neurons; R i (E) is the basis function of the hidden layer nodes; W ik and d k is the output layer weight and threshold; m is the number of arm muscle groups; p(y k ) is the probability of each motion direction.

[0106] Example 4

[0107] Compared with Example 1, this embodiment is different in that the relative Euclidean distance is improved for the above-mentioned K-means clustering, thereby improving the ability of K-means to handle the clustering of discrete boundary points. The ratio of the standard Euclidean distance to the maximum Euclidean distance in the cluster is defined as the relative Euclidean distance R. The improved k-means clustering expression is,

[0108]

[0109] In the formula, k is the category to be selected, K is the category to be selected, and x f The sample to be classified is composed of arm strength F f and its differential Composition, C k is the sample space of fast and slow speed. k is the maximum Euclidean distance in the kth cluster.

[0110] Using the cross-distance method, we find the sample points with the maximum cross-distance. The calculation method is as follows:

[0111]

[0112] Where i1 is the number of the sample point with the maximum Euclidean distance from cluster center 2, and i2 is the number of the sample point with the maximum Euclidean distance from cluster center 1.

[0113] At this time, the maximum Euclidean distance from the sample points to their respective cluster centers can be approximately expressed as:

[0114]

[0115] In the formula, x i1 and x i2 is the sample point with the maximum cross-Euclidean distance, d1 is the maximum Euclidean distance of cluster center 1, and d2 is the maximum Euclidean distance of cluster center 2.

[0116] The optimization goal is to minimize the sum of squared errors of the Euclidean distance from the sample point to the cluster center, and find the cluster center that is most closely related to the cooperation information. This optimization goal can be expressed as:

[0117]

[0118] Iterate and update the cluster center according to formula (14-17) until the cluster center does not change.

[0119] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A human-robot collaborative control method based on human skill learning and imitation, characterized in that: The method includes signal acquisition and processing, construction of a three-dimensional arm strength estimation model, construction of a human collaborative skill imitation model, and human-robot collaboration, which correspond to the following steps respectively: Step 1: The human instructor drags the robot to move freely by applying arm forces of different magnitudes and directions, and collects the electromyographic signals and joint angle information of the human instructor's arm muscle groups and the real arm strength as training samples for the three-dimensional arm strength estimation model; The root mean square filter is used to extract the amplitude information in the electromyographic signal, and the fast Fourier transform is used to extract the tremor information in the electromyographic signal, thereby completing the signal acquisition and processing; Step 2: Using a fast orthogonal search method, the obtained amplitude information and joint angle information are fused to obtain force-related information in the three-dimensional motion direction; The acquired amplitude information, tremor information and force-related information are used as input, and the real arm strength is used as output. A three-dimensional arm strength estimation model for estimating the three-dimensional arm strength is constructed by means of a parallel long short-term memory neural network with an electromyographic signal correction unit and an input-output control unit added. Step 3: Using the three-dimensional arm strength estimation model and angle sensor to obtain the arm strength information and speed information of the human collaborative demonstration process, establish a regression relationship with arm strength as input and movement speed as output, and complete the construction of the human collaborative skill imitation model; Step 4: Based on the constructed human collaborative skill imitation model, the estimated three-dimensional arm force is directly converted into the speed adjustment of the robot, so as to control the robot to cooperate with humans to complete the collaborative task; The electromyographic signal correction unit adopts a neural network structure based on the principle of low-pass discrete filter. E ES [n]mE ES [n-1]+σ(E FFT W E +b E )(E RMS [n]-E ES [n-1]) In the formula, E ES [n] is the electromyographic signal after correction of the nth data; W E and b E is the weight and bias of the EMG signal correction unit; σ(·) is the Sigmoid function; E RMS [n] is the amplitude information of the nth electromyographic signal processed by RMS filtering; The input-output control unit uses the naive Bayes algorithm to obtain the direction of human movement intention, which is expressed as: In the formula, E RMS1 ,E RMS2 ,……,E RMS6 are samples to be classified, which are the amplitude information of electromyographic signals after RMS filtering of six muscle groups; k It is the direction of human movement intention, which is divided into four characteristic attributes, namely no output force, X-axis force, Y-axis force and Z-axis force. When the forces in the three directions are all less than 4N, it is considered as no output force; P(E RMSi |y k ) is the probability of each classification sample when the characteristic attribute occurs; P(y k ) is the conditional probability of each characteristic attribute occurring; P(E RMS i) is the probability of occurrence of classified samples; y t The direction of human movement intention at the current moment; The tremor information is the average amplitude of the electromyographic signal of 4-12 Hz, and its expression is: In the formula, E FFT For tremor information, FE raw [k] is the kth value of the bilateral power density spectrum of the time domain signal, which is obtained by fast Fourier transform; N is the sample data size; n min and n max They represent the double-sided power density spectrum serial numbers corresponding to frequencies of 4 Hz and 12 Hz respectively.

2. The human-robot collaborative control method based on human skill learning and imitation according to claim 1 is characterized in that: The myoelectric signals of the arm muscle group include the myoelectric signals of the anterior end of the deltoid muscle, the posterior end of the deltoid muscle, the biceps brachii, the triceps brachii, the pectoralis major and the infraspinatus muscle.

3. The human-robot collaborative control method based on human skill learning and imitation according to claim 1 is characterized in that: The joint rotation angle information includes elbow joint rotation angle information and shoulder joint rotation angle information.

4. The human-robot collaborative control method based on human skill learning and imitation according to claim 2 is characterized in that: The fast orthogonal search method uses the correlation coefficient between the extracted signal and the actual output arm force as the iteration standard to extract the force-related electromyographic signal amplitude information without the influence of joint rotation. The expression of the iteration standard is: In the formula, Cov(F y ,E y ) is the covariance of the force-related information Ey and the measured arm force Fy; Var[F y ] is the arm strength F y Variance of Var[E y ]For relevant information y The variance of .

5. The human-robot collaborative control method based on human skill learning and imitation according to claim 1 is characterized in that: The parallel long short-term memory neural network uses three parallel-structured sub-LSTM neural networks to estimate the three-dimensional arm strength respectively.

6. The human-robot collaborative control method based on human skill learning and imitation according to claim 1 is characterized in that: The human collaborative skill imitation model adopts a multi-model Gaussian process regression algorithm based on K-means clustering to obtain the regression relationship between arm strength and speed adjustment.

7. The human-robot collaborative control method based on human skill learning and imitation according to claim 1 is characterized in that: The human collaboration skill imitation model is the speed adjustment skill of humans facing changes in force information.

Images