A Human-Machine Collaboration Method Based on Dynamic Arm Force Estimation

Abstract

The present invention discloses a human-machine collaboration method based on dynamic arm force estimation. The method of the present invention includes the steps of: collecting surface electromyography signals sEMG at multiple positions on the arm; performing adaptive amplitude extraction on multiple sEMG; then performing adaptive smoothing processing on the sEMG amplitudes of multiple channels; passing the smoothed sEMG amplitudes of multiple channels through a fusion neural network to perform delay compensation, channel fusion, and timing capture on the sEMG amplitudes of multiple channels in sequence, obtaining fusion features corresponding to the sEMG amplitudes of multiple channels, and then inputting the fusion features into an estimation neural network to obtain the mapping relationship between the fusion features and the operator's arm force; performing collaborative control according to the mapping relationship to drive the collaborative robotic arm to perform collaborative actions. Compared with the prior art, the present invention can simultaneously solve the three bottlenecks of human-machine collaboration, namely, the delay of surface electromyography signals, multi-channel fusion, and feature timing relationship capture, and achieve the human-machine collaboration effect with low delay and high precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent control of human-machine collaborative equipment, and specifically relates to a human-machine collaborative method based on dynamic arm force estimation. Background Art

[0002] For physical human-computer interaction scenarios in industrial applications, intention recognition based on surface electromyography (sEMG) has the advantages of not being restricted by application scenarios, high flexibility, high adaptability, and low system complexity compared to other intention recognition methods such as vision and force sensors. Therefore, it has attracted widespread attention from researchers.

[0003] At present, in order to establish the relationship between sEMG and the dynamic arm force of the robot, a model of sEMG and output force of each channel is established through system identification, a combination of angle-based EMG amplitude calibration and parallel cascade identification is used to estimate the force generated by the elbow joint at the wrist during dynamic contraction, and a mapping relationship between the anterior and posterior deltoid muscles and horizontal force is established using a long short-term memory network (LSTM).

[0004] However, current methods only focus on estimating the force in a single direction or a single joint, and there are many challenges in applying them to complex human-machine collaboration scenarios. The bottleneck issues affecting the effect of force estimation are the stability of the sEMG amplitude estimation method, the sEMG delay problem in the arm force estimation process, the fusion problem of multiple channel sEMG, and the capture of the temporal relationship of sEMG features.

[0005] At present, although there are three types of technical solutions, including using parallel cascade models to capture electromyographic delay, using generalized canonical correlation analysis to establish a real-time feature extraction and fusion model, and using linear regression methods to weight multiple channel sEMG signals and using recursive neural networks to capture the time information between fused features, so far, no solution can simultaneously solve the problems of sEMG delay, multi-channel sEMG fusion, and capturing the temporal relationship of sEMG features.

[0006] In addition, the existing sEMG amplitude estimation method only uses empirical methods to determine the time window length of the amplitude estimation, and the experience-based method is usually difficult to ensure the reproducibility of the technical solution; on the other hand, the currently standardized amplitude estimation method is an amplitude estimation method based on non-causal iterative learning. Non-causal iterative learning means using future electromyographic data to estimate the amplitude at the current moment. Therefore, it is only suitable for offline analysis and cannot adapt to online real-time human-computer collaboration processes. Summary of the invention

[0007] To overcome one or more defects and deficiencies existing in the prior art, the object of the present invention is to provide a human-machine collaboration method based on dynamic arm force estimation, which is used to simultaneously solve the problems of sEMG delay, multi-channel sEMG fusion, and capturing the temporal relationship of sEMG features under real-time human-machine collaboration.

[0008] To achieve the above object, the present invention adopts the following technical solutions.

[0009] A human-machine collaboration method based on dynamic arm force estimation includes the following steps:

[0010] Collect the surface electromyogram signals sEMG, angular signals, and force signals applied to the collaborative robotic arm at multiple positions on the operator's arm when performing a set action during human-machine collaboration, and input the sEMG, angular signals, and force signals into a data signal processing system for subsequent processing;

[0011] Perform adaptive amplitude extraction on multiple sEMGs to obtain the sEMG amplitudes of corresponding multiple channels; then perform adaptive smoothing processing on the sEMG amplitudes of multiple channels to obtain the smoothed sEMG amplitudes of multiple channels;

[0012] Take the smoothed sEMG amplitudes of multiple channels as inputs, and perform delay compensation, channel fusion, and temporal capture on the sEMG amplitudes of multiple channels in sequence through a fusion neural network to obtain the fusion features corresponding to the sEMG amplitudes of multiple channels; then input the fusion features into an estimation neural network to obtain the mapping relationship between the fusion features and the operator's arm force, and realize the estimation of the operator's arm force from the sEMG amplitudes of multiple channels;

[0013] The data signal processing system performs collaborative control according to the estimation of the operator's arm force, obtains a collaboration that matches the operator's set action, and then drives the collaborative robotic arm to perform a collaborative action.

[0014] Preferably, the steps of performing adaptive amplitude extraction on multiple sEMGs are as follows:

[0015] Construct an extraction window for sEMG amplitude to obtain an iterative formula for the length of the extraction window; according to the iterative convergence criterion based on information entropy, iterate on the length of the extraction window, and when the iterative convergence criterion is satisfied, the length of the extraction window is optimal;

[0016] After obtaining the optimal length of the extraction window, use the optimal extraction window to perform adaptive extraction on the sEMG amplitude to obtain the adaptive sEMG amplitudes of multiple channels.

[0017] Preferably, the iterative formula for the length of the extraction window is:

[0018]

[0019] Among them, M k represents the length of the extraction window, α is the rectifier coefficient, v is the re-linearization coefficient, ω k is the adaptive sEMG amplitude at time k, A k is the first derivative of the sEMG amplitude, B k is the second derivative of the sEMG amplitude. Γ(·) represents the Euler gamma function.

[0020] Furthermore, the iterative convergence criterion based on information entropy is as follows:

[0021]

[0022] Among them, H(S k ) represents the iterative convergence criterion based on information entropy. S k represents the average value of the sEMG amplitude s k extracted within the extraction window length M at the current iteration number, p(S k ) represents the probability density function χ k of the chi-square distribution with M k degrees of freedom at the iteration number. 2 (·).

[0023] Preferably, the steps for adaptive smoothing processing are as follows:

[0024] Use the fractal dimension D to evaluate the degree of fluctuation of the envelopes of the sEMG amplitudes of multiple channels after adaptive amplitude extraction;

[0025] Construct the relationship between the exponential decay function ε and the fractal dimension D, and obtain the optimal decay factor A in the exponential decay function ε according to the particle swarm optimization algorithm;

[0026] Use the exponential decay function ε containing the optimal decay factor A to smooth the envelopes of the sEMG amplitudes of multiple channels after adaptive amplitude extraction.

[0027] Furthermore, the calculation of the fractal dimension D is shown in the following formula:

[0028]

[0029] And:

[0030]

[0031]

[0032] Among them, N(Δ) and N(2Δ) respectively represent the number of grids required to cover the sEMG amplitude envelope with grids of width Δ and 2Δ, l represents the sample size of the sEMG signal collected within the set sampling period, and ω i represents the sEMG amplitude envelope at the i-th moment within the set sampling period.

[0033] Furthermore, the relationship between the exponential decay function ε and the fractal dimension D is shown in the following formula:

[0034] ε = e -AD

[0035] The envelope of the sEMG amplitude is smoothed as shown in the following formula:

[0036]

[0037] Among them, is the sEMG amplitude feature at the k-th moment.

[0038] Preferably, the fusion neural network includes a delay compensation module, a channel fusion module, and a timing capture module connected in sequence; the fusion neural network takes the sEMG amplitudes of multiple channels after smoothing as inputs, and the output is the corresponding fusion feature;

[0039] The delay compensation module is used to perform delay compensation on the sEMG amplitudes of multiple channels after adaptive smoothing;

[0040] The channel fusion module is used to fuse the sEMG amplitudes of multiple channels after delay compensation to obtain the corresponding fusion features;

[0041] The timing capture module is used to analyze the timing relationship between the fusion features and output the fusion features with timing relationship.

[0042] Furthermore, the estimation neural network includes a long short-term memory network and an attention mechanism network connected to each other;

[0043] The long short-term memory network takes the fusion features with timing relationship as inputs, and the long short-term memory network outputs the fusion features with enhanced timing relationship;

[0044] The attention mechanism network takes the fusion features with enhanced timing relationship as inputs, and the output is the estimated arm force, thereby mapping the sEMG amplitudes of multiple channels to the operator's arm force to achieve arm force estimation.

[0045] Preferably, the cooperative control specifically calculates the displacement using two methods: PD speed control and segmented speed control. The PD speed control and segmented speed control are shown in the following formulas respectively:

[0046]

[0047]

[0048] Among them, F x and F y are respectively the horizontal force and vertical force of the arm strength. represents the derivative of the horizontal force, and k P and k D are respectively the proportional adjustment parameter and differential adjustment parameter of PD control. Δx and Δy are respectively the displacements of the robot in the horizontal direction and vertical direction.

[0049] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0050] Compared with the existing dynamic arm strength estimation and cooperation method, the present invention realizes the standardized extraction of features based on the characteristics of the signal itself through the combination of adaptive amplitude extraction and adaptive smoothing based on fractal dimension, ensuring the standardization and stability of the extraction; taking into account the difficulties of delay, multi-channel fusion, and timing relationship capture in the arm strength estimation process, using intelligent algorithms to improve the accuracy of dynamic arm strength estimation, using long short-term memory network and attention mechanism to realize the mapping from sEMG amplitude to dynamic arm strength, establishing a customized cooperation control method, and converting the estimated dynamic arm strength into the cooperation behavior of the robot, which has the advantages of small delay and high accuracy of arm strength estimation. Description of the Drawings

[0051] Figure 1 is a schematic flow chart of a human-machine cooperation method based on dynamic arm strength estimation according to one embodiment of the present invention;

[0052] Figure 2 is Figure 1 a schematic diagram of the platform facilities used for data collection in Detailed Embodiments

[0053] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and their embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0054] Embodiment

[0055] As Figure 1 shown, a human-machine cooperation method based on dynamic arm strength estimation in this embodiment includes the following steps:

[0056] S1. Combine Figure 2 the platform shown for data collection. Figure 2 The platform shown can be regarded as a data signal processing system as a whole; the process includes:

[0057] S11. The operator attaches a bipolar electromyography sensor to the activated muscles to collect surface electromyography signals. The activated muscles include the anterior deltoid, posterior deltoid, biceps brachii, and triceps brachii of the arm. A gyroscope is attached to the shoulder joint to collect the angular signal of the shoulder joint rotation, and a six-axis force sensor at the end of the cooperative robotic arm is held to collect the force signal output by the arm.

[0058] The operator performs preset related actions. The surface electromyography signal and the angular signal are transmitted to the industrial control computer and the real-time control system through Bluetooth communication to obtain the sEMG signal and the angular signal. After the amplitude of the force signal is amplified by the amplitude doubling module, data interaction is carried out with the industrial control computer and the real-time control system through the control cabinet. In other embodiments, the specific muscle positions collected and the number of signal channels can be flexibly set.

[0059] S12. The sEMG signal, the angular signal, and the force signal are centralized in the industrial control computer and converted into data for storage for subsequent processing, and the subsequent processing process is also completed within the industrial control computer.

[0060] S2. Adaptive amplitude extraction is performed on the collected sEMG signals respectively to obtain the sEMG amplitudes of the four channels corresponding to the anterior deltoid, posterior deltoid, biceps brachii, and triceps brachii respectively. The process includes:

[0061] S21. Construct an extraction window for the sEMG amplitude and establish the corresponding relationship between the sEMG amplitude and the extraction window. Specifically:

[0062] The sEMG amplitude features include the mean absolute value (MAV) and the root mean square value (RMS). For the causal form of the sEMG amplitude features, their calculation is shown in Equation (1):

[0063]

[0064] Among them, M k represents the length of the extraction window, ω k-j is the sEMG amplitude at the k - j moment, α is the rectifier coefficient when collecting the signal, v is the re-linearization coefficient when collecting the signal. In particular, when α = 1 and v = 1, is the MAV estimated amplitude at the k moment. When α = 1 and v = 2, is the RMS estimated amplitude at the k moment.

[0065] It can be seen from Equation (1) that the estimated amplitude of the sEMG depends on the length M k of the extraction window;

[0066] Perform a second-order Taylor expansion of ω k-j in Equation (1) in the neighborhood of j = k, as shown in Equation (2):

[0067]

[0068] Among them, d k represents the first derivative of the amplitude curve at time k, and d′ k represents the second derivative of the amplitude curve at time k;

[0069] As shown in Equation (3), the Newton backward differentiation formula is used for approximate calculation to obtain d k , d′ k :

[0070]

[0071] The approximate formula as shown in Equation (4):

[0072]

[0073] Using Equation (4) to simplify Equation (2), the simplified Equation (2) can be rewritten as Equation (5):

[0074]

[0075] Using Equation (4) again to simplify Equation (5) and ignoring the high-order terms therein to obtain Equation (6):

[0076]

[0077] According to Equation (6), the deviation term can be expressed as:

[0078]

[0079] Thus, the first derivative A k of the amplitude envelope of sEMG as shown in Equation (8) and the second derivative B k of the amplitude envelope of sEMG as shown in Equation (9) are defined:

[0080]

[0081]

[0082] At this time, Equation (7) can be further expressed as:

[0083] b k = A k M k + B k M k 2 (10)

[0084] When estimating the amplitude, the corresponding variance term is shown in Equation (11):

[0085]

[0086] where f(α,v) is expressed as:

[0087]

[0088] where Γ(·) represents the Euler gamma function;

[0089] According to the derivation from Equation (1) to Equation (12), it can be seen that the square of the amplitude estimation error of sEMG can be expressed as Equation (13):

[0090]

[0091] From Equation (13), it can be obtained that the square e of the amplitude estimation error of sEMG k 2 with respect to the window length M k has a derivative as shown in Equation (14):

[0092]

[0093] Let to obtain the optimal length M of the extraction window k ;

[0094] Adopt the two specific cases where the first derivative A k of the amplitude envelope of sEMG dominates (Equation (8)) and the second derivative B k of the amplitude envelope of sEMG dominates (Equation (9)) to obtain a simplified approximate solution. At this time, the iterative formula for the optimal length M k of the extraction window is shown in Equation (15):

[0095]

[0096] S21. Set the iterative convergence criterion based on information entropy, and then iterate on the length M k of the extraction window in Equation (15) until the information entropy satisfies the iterative convergence criterion. At this time, the length M k of the extraction window is optimal; specifically:

[0097] The iterative convergence criterion based on the information entropy H(S k ) is shown in Equation (16):

[0098]

[0099] S in Equation (16) k , p(Sk ) are calculated as shown in formula (17) and formula (18):

[0100]

[0101]

[0102] Among them, S k Represents the collected sEMG signal s k The length of the current extraction window is M k The average value under p(S k ) indicates that the current k The probability density function of the chi-square distribution with the order degrees of freedom is 2 (·);

[0103] The length of the extraction window M k Iterate by accumulating from a set starting value, and then judge the information entropy H(S k ) continues to increase; when the information entropy H(S k ) no longer increases, it is determined that the optimal extraction window length M has been obtained. k , and the iteration ends; when the information entropy H(S k ) is still increasing, continue iterating;

[0104] S23. Obtaining the optimal extraction window length M in step S22 k Then, M k Substitute into equation (1) of step S21 to adaptively extract the sEMG amplitude, and then output the extracted sEMG amplitudes of the four channels;

[0105] S3, adaptively smoothing the sEMG amplitudes of the four channels after the adaptive amplitude extraction in step S2, improving the signal-to-noise ratio of the sEMG amplitude, and realizing adaptive sEMG amplitude estimation; the process includes:

[0106] The fractal dimension is used to measure the degree of jitter of the extracted sEMG amplitude envelope, and the exponential decay function relationship of the smoothing coefficient with respect to the fractal dimension is established, and the particle swarm algorithm is used to obtain the optimal smoothing relationship.

[0107] S31, using fractal dimension to evaluate the degree of fluctuation of the sEMG amplitude envelope extracted in step S2; the larger the fractal dimension, the more violent the fluctuation of the sEMG amplitude envelope, and the smaller the corresponding smoothing coefficient. The calculation of the fractal dimension D is shown in formula (19):

[0108]

[0109] Among them, N(Δ) and N(2Δ) are respectively shown in Formula (20) and Formula (21):

[0110]

[0111]

[0112] Among them, N(Δ) and N(2Δ) respectively represent the number of grids required to cover the sEMG amplitude envelope with grids of width Δ and 2Δ, l represents the sample size of the sEMG signal collected within the set sampling period, and ω i represents the sEMG amplitude envelope at the i-th moment within the set sampling period;

[0113] S32. After obtaining the fractal dimension D in step S31, construct the relationship between the exponential decay function ε and the fractal dimension D, obtain the optimal decay factor A in the exponential decay function ε according to the particle swarm optimization algorithm, and then use the exponential decay function ε containing the optimal decay factor A to adaptively smooth the sEMG amplitude envelope extracted in step S2; specifically:

[0114] The relationship between the exponential decay function ε and the fractal dimension D is shown in Formula (22):

[0115] ε = e -AD (22)

[0116] Among them, A represents the decay factor;

[0117] Use the particle swarm optimization algorithm to find the global optimal decay factor A. In this embodiment, it is preferably to limit the optimization range of the decay factor A to [2, 6]; the objective function corresponding to the decay factor A is set such that when the operator rotates the shoulder joint (that is, the muscle only maintains the joint rotation without outputting other forces), the correlation coefficient between the rotation angle θ of the operator's shoulder joint and the fusion envelope ω s after the sEMG amplitudes of the anterior deltoid muscle and the posterior deltoid muscle are fused is the largest, and the specific calculation method is shown in Formula (23):

[0118]

[0119] Among them, cov(ω s , θ) represents the covariance between the fusion envelope ω s and the shoulder joint rotation angle θ; σ θ respectively represent the standard deviations of the fusion envelope ω s and the shoulder joint rotation angle θ; after the particle swarm optimization algorithm obtains the minimum value in the first line of Formula (23), the corresponding decay factor A in the second line of the constraint condition takes the optimal value at this time;

[0120] Substitute the optimal value of the attenuation factor A into Equation (23) to obtain the optimal exponential decay function ε, and then smooth the sEMG amplitude envelope, as specifically shown in Equation (24):

[0121]

[0122] S33. After completing the smoothing of the sEMG amplitude envelope, output the sEMG amplitudes of the four channels after corresponding smoothing;

[0123] S4. Respectively construct a fusion neural network and an estimation neural network, and use the neural network learning method to determine the mapping relationship between the sEMG amplitudes of the four channels and the arm force, so as to realize the estimation of the arm force. The fusion neural network and the estimation neural network are interconnected; the process includes:

[0124] S41. Construct a fusion neural network to convert the sEMG amplitudes of the four channels into corresponding fusion features, which is used to solve the bottleneck problem that it is difficult to balance the three aspects of delay, fusion of sEMG amplitudes of the four channels, and capture of timing relationship in arm force estimation;

[0125] The fusion neural network includes a delay compensation module, a channel fusion module, and a timing capture module connected in sequence; the fusion neural network takes the sEMG amplitudes of the four channels after smoothing as input, inputs them into the delay compensation module, and then outputs fusion features with low delay and strong timing relationship after being processed by the channel fusion module and the timing capture module in sequence. The specific construction is as follows:

[0126] Set the length of the processing window for the sEMG amplitudes of the four channels after smoothing. The fusion neural network only processes the sEMG amplitudes within the processing window;

[0127] The delay compensation module takes the sEMG amplitudes of the four channels within the processing window as input, and the output is the sEMG amplitudes of the four channels within the processing window after delay compensation; since the sEMG amplitudes of the four channels will produce a time delay effect during the amplitude estimation process, and the delay cannot be eliminated by moving the time axis forward during the real-time arm force estimation process, the delay compensation module adopts an indirect way to achieve delay compensation. The principle of indirectly achieving delay compensation is to add or subtract a delay deviation to the sEMG amplitudes of the four channels respectively; the magnitude and sign of the delay deviation are respectively related to the time characteristics of the sEMG amplitudes within the processing window, such as amplitude slope, amplitude direction change, etc.; a one-dimensional convolutional network is constructed in the delay compensation module to analyze the time characteristics of the sEMG amplitudes of each channel. The one-dimensional convolutional network uses its own self-learning ability to establish the mapping relationship between the time characteristics of each channel's sEMG and the delay deviation, so as to indirectly achieve delay compensation;

[0128] The channel fusion module takes the sEMG amplitudes of four channels after delay compensation within the processing window as input; a one-dimensional convolutional network is built within the channel fusion module, and the one-dimensional convolutional network is used to analyze the activation degree among the sEMG amplitudes of the four channels at each moment within the processing window, and determines the fusion weight coefficients corresponding to the sEMG amplitudes of each channel by using its own self-learning ability, so as to fuse the sEMG amplitudes of the four channels and obtain the corresponding fusion features;

[0129] The timing capture module takes the fusion features after delay compensation and channel fusion within the processing window as input, and is used to analyze the timing relationship among the fusion features within the processing window, so as to output the fusion features with timing relationship; Timing capture is achieved by analyzing the similarity relationship between the fusion features at the current moment and the long short-term memory components at the past moment, determining the contribution degree of the fusion features at the current moment to the update of the long short-term memory, and then completing the update of the long short-term memory components, so as to achieve the capture of the timing relationship; The timing capture module consists of a one-dimensional convolutional network and a simplified long short-term memory update model; The one-dimensional convolutional network is used to analyze the similarity relationship between the fusion features at the current moment and the long short-term memory components at the past moment, and determine the contribution degree of the fusion features at the current moment to the update of the long short-term memory; The simplified long short-term memory update model is used to realize the update of the long short-term memory components, so as to achieve the capture of the timing relationship;

[0130] The one-dimensional convolutional network takes the fusion features at the current moment and the long short-term memory components at the past moment as input, and outputs the weight coefficients for long short-term memory update by analyzing the similarity relationship between the fusion features at the current moment and the long short-term memory components at the past moment. The weight coefficients include weight α and weight β;

[0131] The simplified long short-term memory update model is shown in Equation (25):

[0132]

[0133] where, ⊙ represents the dot product operation, c t-1 , h t-1 respectively represent the long-term memory and short-term memory of the previous moment; Weight α is used to select the important part of the fusion feature x t at the current moment and control the forgetting degree of the short-term memory component h t-1 at the past moment, and then obtain the intermediate memory component m; On the one hand, weight β is used to balance the long-term memory component c t-1 at the past moment and the intermediate memory component m to realize the update of the short-term memory component h t at the current moment, and on the other hand, it is used to add the important memory in the intermediate memory component to the long-term memory at the current moment to realize the update of the long-term memory c t at the current moment;

[0134] S42. After obtaining the fused features and their temporal relationships in step S41, an estimation neural network is constructed to map the fused features to the arm strength of the operator;

[0135] The estimation neural network takes the fused features with low latency and strong temporal relationships obtained in step S41 as input. Its structure adopts a neural network combining a long short-term memory network and an attention mechanism network, which is used to establish the mapping relationship between the fused features and the arm strength. The specific construction is as follows:

[0136] The long short-term memory network takes the fused features output by the fusion neural network as input and outputs fused features with enhanced temporal relationships. The long short-term memory network uses an input gate i, a forget gate f, and an output gate o to capture the temporal relationship of the input fused feature x″ t at time t. The outputs f t , i t , o t of the forget gate, input gate, and output gate are calculated by equations (26), (27), and (28) respectively:

[0137] f t = σ(W xf x″ t + W hf h′ t-1 + b f ) (26)

[0138] i t = σ(W xi x″ t + W hi h′ t-1 + b i ) (27)

[0139] o t = σ(W xo x″ t + W ho h′ t-1 + b o ) (28)

[0140] where σ is the non-linear activation function sigmoid, h′ t-1 is the hidden state of the long short-term memory network unit at the (t - 1)th moment; W xf , W hf , W xi , W hi , W xo , W ho are learnable weight matrices respectively, and b f , b i , b o are bias terms;

[0141] After being processed by three gates, the hidden state h' of the long short-term memory network at time t t The calculation formula is derived from the following three equations:

[0142]

[0143]

[0144] h' t = tanh(c' t ) ⊙ o t (31)

[0145] Among them, x″ t represents the fused feature of the input, represents the intermediate variable, c' t is the unit state of the long short-term memory network, representing the long-term memory component, W xc 、W hc are learnable weight matrices, b c is the bias term, f t 、i t 、o t are the outputs of the forget gate, input gate, and output gate respectively;

[0146] Through the above formulas (32), (33), (34), the fused feature h' with enhanced temporal relationship is obtained t ;

[0147] The attention mechanism network takes the fused feature h' with enhanced temporal relationship t as the input, and its output is the estimated arm strength. The attention mechanism network evaluates the importance of the fused feature using formulas (35), (36) and calculates the importance coefficient ζ i :

[0148] s i = tanh(W a h' i + b a ) (35)

[0149]

[0150] Among them, h' i represents the fused feature with enhanced temporal relationship at the i-th moment, W a ,W α are learnable weight matrices, b a is the bias term, s i represents the intermediate variable, and the superscript T represents the transpose of s i ;

[0151] The attention mechanism network calculates the estimated arm force according to equations (37) and (38):

[0152]

[0153]

[0154] where h′ i represents the fused feature with enhanced temporal relationship at the i-th moment, ζ i represents the importance coefficient, x″′ represents the fused feature corrected by the importance coefficient, F(·) represents the non-linear activation function; W f represents the learnable weight matrix, b f represents the bias term, represents the arm force estimated by the attention mechanism network;

[0155] S43. Train the constructed fused neural network and estimation neural network in an end-to-end manner. By calculating and minimizing the error between the estimated arm force and the true arm force, the parameter learning of the fused neural network and the estimation neural network is respectively achieved; the trained fused neural network and estimation neural network are used to complete the operation of the mapping relationship from the sEMG amplitudes of the four channels within the processing window to the arm force; the structures of the fused neural network and the estimation neural network are shown in Table 1:

[0156] Table 1 Network structures of the fused neural network and the estimation neural network

[0157]

[0158] S5. After mapping the sEMG amplitudes to the arm force in step S4 , the industrial control computer, according to the mapping relationship, converts the estimated arm force into the collaborative actions of the collaborative robotic arm through collaborative control, and enables the real-time control system to send control commands to the control cabinet, which then gives signals to the collaborative robotic arm to execute the corresponding collaborative actions; the collaborative actions are the mechanical movements of the collaborative robotic arm that cooperate with the preset actions of the operator; the process of collaborative control includes:

[0159] Calculate the displacement using both PD speed control and segmented speed control methods simultaneously, as shown in equations (39) and (40) respectively:

[0160]

[0161]

[0162] where F x and F y are the arm forces The horizontal force and vertical force represents the derivative of the horizontal force, k P and k D are respectively the proportional adjustment parameter and the differential adjustment parameter of the PD control law, and Δx and Δy are respectively the displacements of the robot in the horizontal direction and the vertical direction;

[0163] The cooperative manipulator performs corresponding cooperative actions according to the results of Equations (36) and (37), thereby realizing the corresponding human - machine cooperation process.

[0164] Compared with the prior art, the human - machine cooperation method based on dynamic arm force estimation in this embodiment has the beneficial effects that:

[0165] Through the combination of adaptive amplitude extraction based on iterative learning and adaptive smoothing based on fractal dimension, the standardized extraction of features based on the characteristics of the signal itself is realized, ensuring the standardization and stability of the extraction; taking into account the difficulties of delay, multi - channel fusion, and time - series relationship capture in the arm force estimation process, the fusion neural network and the estimation neural network are used to improve the accuracy of dynamic arm force estimation, and the long - short - term memory network and the attention mechanism are used to realize the mapping from sEMG amplitude to dynamic arm force; a cooperative control method of PD speed control and segmented speed control is established, converting the estimated dynamic arm force into the cooperative behavior of the robot, which has the advantages of small delay and high arm force estimation accuracy.

[0166] In an alternative implementation manner of this embodiment, the cooperative control in step S5 can also adopt the constant - force control method; the process of constant - force control is as follows:

[0167] Denote the arm force obtained in step S4 as F d , and denote the actual arm force collected by the six - dimensional force sensor at the end of the cooperative manipulator as F t , and the error e between the two is as shown in Equation (41):

[0168] e = F d - F t (41)

[0169] Calculate the compensation amount ΔF according to Equation (42):

[0170]

[0171] where K p and K d represent the proportional adjustment parameter and the differential adjustment parameter of the PD control law respectively, is the first - order derivative of the error e;

[0172] Execute Equation (43) according to the compensation amount ΔF:

[0173]

[0174] Among them, m represents the inertia coefficient, b represents the damping coefficient, k represents the stiffness coefficient, Δx(t), respectively represent the displacement compensation amount, the first derivative of the displacement compensation amount, and the second derivative of the displacement compensation amount;

[0175] Only considering the stiffness term, the position compensation of the collaborative robotic arm after simplification is shown in Equation (44):

[0176]

[0177] The collaborative robotic arm performs collaborative behaviors according to the obtained position compensation, thereby realizing human-robot collaboration.

[0178] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A human - machine collaboration method based on dynamic arm force estimation, characterized in that, it includes the following steps: Collect the surface electromyogram signal sEMG, angle signal, and force signal applied to the collaborative robotic arm at multiple positions on the operator's arm during the execution of a set action in human - machine collaboration, and input the sEMG, angle signal, and force signal into the data signal processing system for subsequent processing; Perform adaptive amplitude extraction on multiple sEMGs to obtain the sEMG amplitudes of corresponding multiple channels; then perform adaptive smoothing processing on the sEMG amplitudes of multiple channels to obtain the smoothed sEMG amplitudes of multiple channels; The steps for performing adaptive amplitude extraction on multiple sEMGs are as follows: Construct an extraction window for sEMG amplitude to obtain an iterative formula for the length of the extraction window; according to the iterative convergence criterion based on information entropy, iterate on the length of the extraction window, and when the iterative convergence criterion is met, the length of the extraction window is optimal; After obtaining the optimal length of the extraction window, use the optimal extraction window to perform adaptive extraction on the sEMG amplitude to obtain the adaptive sEMG amplitudes of multiple channels; The steps for performing adaptive smoothing processing are as follows: Use the fractal dimension D to evaluate the degree of fluctuation of the envelope of the sEMG amplitudes of multiple channels after adaptive amplitude extraction; Construct the relationship between the exponential decay function ε and the fractal dimension D, and obtain the optimal decay factor A in the exponential decay function ε according to the particle swarm optimization algorithm; Use the exponential decay function ε containing the optimal decay factor A to smooth the envelope of the sEMG amplitudes of multiple channels after adaptive amplitude extraction; Take the smoothed sEMG amplitudes of multiple channels as the input, and perform time - delay compensation, channel fusion, and time - series capture on the sEMG amplitudes of multiple channels in sequence through a fusion neural network to obtain the fusion features corresponding to the sEMG amplitudes of multiple channels; then input the fusion features into the estimation neural network to obtain the mapping relationship between the fusion features and the operator's arm force, and realize the estimation of the operator's arm force from the sEMG amplitudes of multiple channels; The fusion neural network includes a time - delay compensation module, a channel fusion module, and a time - series capture module connected in sequence; the fusion neural network takes the smoothed sEMG amplitudes of multiple channels as the input and outputs the corresponding fusion features; The time - delay compensation module is used to perform time - delay compensation on the sEMG amplitudes of multiple channels after adaptive smoothing processing; The channel fusion module is used to fuse the sEMG amplitudes of multiple channels after time - delay compensation to obtain the corresponding fusion features; The time - series capture module is used to analyze the time - series relationship between the fusion features and output the fusion features with time - series relationship; The data signal processing system performs collaborative control according to the estimation of the operator's arm force to obtain a collaboration that matches the operator's set action, and then drives the collaborative robotic arm to perform collaborative actions.

2. The human - machine collaboration method based on dynamic arm force estimation according to claim 1, characterized in that, the iterative formula for the length of the extraction window is: Among them, M k represents the length of the extraction window, α is the rectifier coefficient, v is the re-linearization coefficient, ω k is the adaptive sEMG amplitude at time k, A k is the first derivative of the sEMG amplitude, B k is the second derivative of the sEMG amplitude, Γ(·) represents the Euler gamma function.

3. The human - machine collaboration method based on dynamic arm force estimation according to claim 2, characterized in that, the iterative convergence criterion based on information entropy is: Among them, H(S k ) represents the iterative convergence criterion based on information entropy, S k represents the sEMG amplitude s k extracts the average value of the window length M k at the current iteration number, p(S k ) represents the chi-square distribution probability density function χ k with M 2 (·) at the iteration number.

4. The human-machine collaboration method based on dynamic arm force estimation according to claim 1, characterized in that, the calculation of the fractal dimension D is shown as follows: where: Among them, N(Δ) and N(2Δ) respectively represent the number of grids required to cover the sEMG amplitude envelope with grids of width Δ and 2Δ, l represents the sample size of the sEMG signal collected within the set sampling period, and ω i represents the sEMG amplitude envelope at the i-th moment within the set sampling period.

5. The human-machine collaboration method based on dynamic arm force estimation according to claim 4, characterized in that, the relationship between the exponential decay function ε and the fractal dimension D is shown as follows: ε = e -AD The envelope of the sEMG amplitude is smoothed as follows: Among them, is the sEMG amplitude feature at the k-th moment.

6. The human-machine collaboration method based on dynamic arm force estimation according to claim 1, characterized in that, the estimation neural network includes a long short-term memory network and an attention mechanism network connected to each other; the long short-term memory network takes the fusion features with a temporal relationship as input, and the long short-term memory network outputs the fusion features with a strengthened temporal relationship; the attention mechanism network takes the fusion features with a strengthened temporal relationship as input, and the output is the estimated arm force, so as to map the sEMG amplitudes of multiple channels to the arm force of the operator, realizing arm force estimation.

7. The human-machine collaboration method based on dynamic arm force estimation according to claim 1, characterized in that, the collaborative control specifically calculates the displacement in two ways: PD speed control and segmented speed control, and the PD speed control and segmented speed control are shown as follows respectively: Among them, F x and F y are respectively the horizontal force and the vertical force of the arm strength. represents the derivative of the horizontal force, and k P and k D are respectively the proportional adjustment parameter and the differential adjustment parameter of PD control. Δx and Δy are respectively the displacements of the robot in the horizontal direction and the vertical direction.

Images