Arm force estimation fatigue error compensation method and system based on electromyographic signals

Abstract

The invention discloses an arm strength estimation fatigue error compensation method and system based on an electromyographic signal, and the method comprises the following steps: dividing fatigue samples into a plurality of groups according to a time sequence, carrying out the arm strength fitting training of each group of fatigue samples through an arm output force estimation model, and carrying out the arm strength fitting training according to the verification error of the arm output force estimation model. Carrying out fatigue grade segmentation on the fatigue sample by adopting an optimized segmentation algorithm; inputting the actually measured arm output force into a fatigue grade identification model, and identifying the fatigue grade; and improving a fatigue level identification model, and selecting an arm output force estimation model corresponding to the fatigue level by using a self-iteration reverse reasoning mechanism of the improved fatigue level identification model to realize dynamic compensation of fatigue errors. According to the method, the influence of muscle fatigue on the dynamic arm strength estimation model can be eliminated, and the model error problem of assuming that fatigue linearly develops along with time in the traditional fatigue compensation process is solved.

Description

Technical Field

[0001] The invention belongs to the field of human-computer interaction and muscle fatigue detection, and in particular relates to a fatigue error compensation method and system for arm force estimation based on electromyographic signals. Background Art

[0002] With the development of biomedical field, myoelectric signals are gradually applied to human-robot collaborative assembly, collaborative handling and sports rehabilitation because they can obtain human active participation. Compared with the working principle of force sensor to obtain contact force through strain information, myoelectric signals come from human muscles themselves, avoiding the interference of external environment contact, and are more suitable for characterizing human instinctive response and movement intention. Although a large number of excellent methods of arm force estimation models based on muscle parameters and parameter-free arm force estimation models have been proposed, these models are only applicable to limited experimental conditions. Usually, in the process of constructing these models, the subjects rest for a long time to ensure that they are in a fatigue-free state, so the effect of muscle fatigue on arm force estimation is ignored. However, muscle fatigue is one of the most common muscle states. After humans participate in sports, physiological parameters will change more or less due to muscle fatigue. Appropriate muscle fatigue is accompanied by the production activities or sports rehabilitation activities of the operator. Therefore, considering the operator's muscle fatigue in the arm force estimation model and making corresponding compensation is of great significance for the development and application of human-robot interaction technology.

[0003] Among the many muscle fatigue prediction methods, electromyographic signals (EMG) can record and reflect the state of muscles during exercise, and are of great value for quantifying the degree of muscle fatigue. Studies have shown that during static contraction (i.e., the length and tension of the muscle remain unchanged), the muscle conduction velocity decreases as the degree of muscle fatigue increases, and the decrease in muscle conduction velocity will lead to a decrease in the average frequency (or median frequency) of the electromyographic signal. Therefore, the average frequency (median frequency) of the electromyographic signal is called the gold standard for predicting static muscle fatigue. In addition, under periodic cyclic motion conditions, due to the similarity and periodicity of biomechanics, some static fatigue standards can also be verified. However, during dynamic contraction (i.e., the length and tension of the muscle are not constrained), the surface electromyographic signal characteristics change faster than static contraction due to the influence of output force, joint rotation, etc., and the non-stationarity is stronger, and the fatigue indicators under static conditions are no longer applicable. Therefore, the research results of muscle fatigue based on electromyographic signals are mostly concentrated under static conditions or periodic cyclic motion conditions. This loose restriction also greatly limits the applicability of the evaluation scheme, because many clinical and industrial scenarios cannot avoid dynamic contractions involving unpredictable changes in muscle length and / or tension.

[0004] Most current studies have ignored the correlation between the amplitude and median fatigue of the electromyographic signal and the muscle output force. Studies have shown that muscle contraction level, muscle contraction speed, etc. can affect the physiological characteristics of muscle fatigue. Therefore, ignoring the change in output force and only considering the characteristics of the electromyographic signal parameters themselves will inevitably affect the practicality of muscle dynamic contraction fatigue tracking. However, in arm strength estimation, the muscle output force is unknown. At this time, the muscle output force will be used as both the input item of muscle fatigue identification and the output item of arm strength estimation. Therefore, how to coordinate the contradiction between the input and output of the model and establish a self-iterative correction model is of great significance for fatigue error compensation in the arm strength estimation model. Summary of the invention

[0005] This application is to overcome the existing technology and provide a method and system for compensating arm strength estimation fatigue errors based on electromyographic signals. The method and system can eliminate the influence of muscle fatigue on the dynamic arm strength estimation model and overcome the model error problem of assuming that fatigue develops linearly over time in the traditional fatigue compensation process.

[0006] The fatigue error compensation method of arm force estimation based on electromyographic signals includes the following steps:

[0007] S1. Fatigue samples are divided into multiple groups according to the time sequence. The arm output force estimation model is used to perform arm strength fitting training on fatigue samples in each group. Based on the verification error of the arm output force estimation model, the fatigue level of fatigue samples is segmented by using the optimized segmentation algorithm.

[0008] S2. Considering the influence of dynamic output force on fatigue characteristics of electromyographic signal, the measured arm output force is input into the fatigue level recognition model to identify the fatigue level;

[0009] S3. Considering that the arm output force in fatigue level identification is an unknown parameter, the fatigue level identification model is improved. The self-iterative reverse reasoning mechanism of the improved fatigue level identification model is utilized to select the arm output force estimation model corresponding to the fatigue level to realize dynamic compensation of fatigue error.

[0010] Optionally, in step S1, the fatigue sample includes electromyographic signals of arm muscle groups, shoulder joint angles and measured arm output forces; the electromyographic signals include amplitude information and median frequency of the electromyographic signals characterizing fatigue characteristics.

[0011] Optionally, in step S1, the arm output force estimation model adopts a long short-term memory neural network, the model input is the amplitude information of the electromyographic signal of the arm muscle group and the shoulder joint angle information, and the model output is the arm output force. The arm muscle group is one or more pairs of antagonist muscles and agonist muscles that are most closely related to the arm output force.

[0012] Optionally, in step S1, the optimization segmentation algorithm is an improved Bayesian optimal recursive optimization algorithm, and the specific process of using the optimization segmentation algorithm to segment the fatigue level of fatigue samples is as follows: the fatigue level division model of the N-1 segmentation points that have been obtained is used as the basic model, and the verification error of the arm output force estimation model of each group of fatigue samples is minimized as the optimization criterion of the Bayesian optimization model, and the best Nth segmentation point is found to obtain a new segmentation model, and the first N-1 segmentation points of the basic model are re-optimized in turn in a recursive manner until the fatigue level division of the N segmentation points is completed. This scheme uses an improved Bayesian optimal recursive optimization algorithm to solve the ordered sample segmentation problem, and optimizes the muscle fatigue level based on the minimization of the verification error of the arm output force estimation model.

[0013] Optionally, in step S2, the fatigue level recognition model is an intra-channel data feature extraction model, an inter-channel data feature extraction model or a spatiotemporal feature fusion neural network, and the spatiotemporal feature fusion neural network is obtained by fusing the intra-channel data feature extraction model and the inter-channel data feature extraction model through spatiotemporal features; a convolution module is added to each of the intra-channel data feature extraction model and the inter-channel data feature extraction model for downsampling, and the downsampled data nodes are further converted through a fully connected network to achieve fatigue level recognition.

[0014] Optionally, in step S2, the fatigue level recognition model adopts an intra-channel data feature extraction model, which is a feature extraction model that combines a one-dimensional dilated convolutional neural network with multi-scale attention. The intra-channel data feature extraction model uses convolution kernels of different scales to extract feature information of the input signal in the time dimension, thereby fully mining the multi-level information of data features within each channel of the input information; the input information is multi-channel time series data composed of joint angles, arm output force, and fatigue characteristics of electromyography signals.

[0015] Optionally, in step S2, the fatigue level recognition model adopts an inter-channel data feature extraction model, and the inter-channel recognition model is a feature extraction model that combines a long short-term memory neural network with a multi-head attention mechanism. The inter-channel data feature extraction model uses a long short-term memory neural network to extract feature information of the input information in the spatial dimension, and uses the multi-head attention mechanism to mine the data feature coupling relationship between each channel. The input information is multi-channel moment data composed of joint angles, arm output force, and fatigue characteristics of electromyography signals.

[0016] Preferably, in step S2, the fatigue level recognition model uses an intra-channel data feature extraction model and an inter-channel data feature extraction model to perform spatiotemporal feature fusion to obtain a spatiotemporal feature fusion neural network; the feature data extracted by the intra-channel data feature extraction model and the inter-channel data feature extraction model are further processed by a convolution module, and each branch is downsampled using a global average pooling layer.

[0017] Optionally, in step S3, the improved fatigue level recognition model adopts an improved spatiotemporal feature fusion network; the fatigue error compensation process is realized by using the self-iterative reverse reasoning mechanism of the spatiotemporal feature fusion neural network: based on the amplitude information and joint angle information of the currently acquired electromyographic signal, the arm strength estimation models of different fatigue levels are used to obtain the estimated arm strength corresponding to different fatigue levels in turn, and the prediction probabilities of the same conditions as the assumed fatigue state are obtained respectively with the help of the improved spatiotemporal feature fusion network, the fatigue level corresponding to the maximum prediction probability is regarded as the fatigue level of the current muscle, and the arm strength estimated under this fatigue level is regarded as the arm output force after fatigue compensation. The improved spatiotemporal feature fusion neural network model adopts a fast fine-tuning algorithm to fine-tune the weight parameters of the fully connected layer in the original spatiotemporal feature fusion neural network model, thereby optimizing the neural network parameters, and further ensuring that after the arm output force in the input information changes from the measured value to the estimated value, the fatigue level with the maximum probability in the output value is the same as the actual fatigue level.

[0018] In the above steps S2-S3, the influence of dynamic output force on the fatigue characteristics of electromyographic signals is considered, the output force is used as the input of the fatigue level recognition model, and a spatiotemporal feature fusion neural network is proposed to improve the accuracy of fatigue level recognition in dynamic muscle contraction by fusing spatiotemporal features; then, considering that the arm output force in fatigue level recognition is an unknown parameter, the self-iterative reverse reasoning ability of the spatiotemporal feature fusion neural network is utilized to select the arm force estimation model corresponding to the fatigue level, so as to realize dynamic compensation of fatigue error.

[0019] An arm force estimation fatigue error compensation system based on electromyographic signals is used to implement the arm force estimation fatigue error compensation method, the system comprising: a fatigue level optimization segmentation module, a fatigue level identification module and an arm output force fatigue compensation module;

[0020] The fatigue level optimization segmentation module is configured to construct an arm output force estimation model using fatigue samples, and perform fatigue level segmentation on the fatigue samples according to the verification error of the arm output force estimation model;

[0021] The fatigue level identification module is configured to input the arm output force into the fatigue level identification model to complete the fatigue level identification;

[0022] The arm output force fatigue compensation module is configured to select an arm output force estimation model corresponding to the fatigue level to complete dynamic compensation of fatigue error.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] In the past, muscle fatigue prediction was mostly used in static contraction scenarios. During dynamic contraction, the surface electromyographic signal characteristics change faster and more non-stationarily than static contraction due to the influence of output force and joint rotation. Fatigue indicators under static conditions are no longer applicable. In addition, different individuals may have subjective definitions of fatigue feelings, which makes the assessment of muscle fatigue unreliable. Most current fatigue classifications are based on the assumption that fatigue progresses linearly over time, but in fact, this assumption does not conform to the actual situation of muscle fatigue progression.

[0025] In view of this, in order to overcome the model error of the traditional fatigue compensation process that assumes that fatigue develops linearly with time, the arm force estimation fatigue error compensation method and system based on electromyographic signals proposed in this application aims to eliminate the influence of muscle fatigue on the dynamic arm force estimation model. This method reasonably divides the fatigue level and establishes an effective connection between the arm output force estimation model and muscle fatigue, which is of great value for fatigue compensation in dynamic output force estimation.

[0026] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments: BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart of a fatigue error compensation method for arm force estimation based on electromyographic signals of the present application;

[0028] Figure 2 It is a flow chart of the arm output force fatigue compensation method based on fatigue level optimization segmentation according to an embodiment of the present application;

[0029] Figure 3 A schematic diagram of signal acquisition in an embodiment of the present application;

[0030] Figure 4 This is a fatigue level optimization segmentation flow chart of an embodiment of the present application;

[0031] Figure 5 This is a structural diagram of the in-channel data feature extraction model of an embodiment of the present application;

[0032] Figure 6 This is a structural diagram of the inter-channel data feature extraction model of an embodiment of the present application;

[0033] Figure 7 This is a diagram of the structure of a spatiotemporal feature fusion neural network according to an embodiment of the present application;

[0034] Figure 8 Schematic diagram of a fatigue error dynamic compensation method according to an embodiment of the present application. DETAILED DESCRIPTION

[0035] The following is a detailed description of the embodiments of the technical solution of the present application in conjunction with the accompanying drawings. Unless otherwise specified, the technical terms or scientific terms used in the present application should have the common meanings understood by those skilled in the art to which the present application belongs.

[0036] Combination Figure 1 It is shown that the human active output force obtained based on electromyographic signals is increasingly being used in the field of human-robot interaction. However, the accuracy of the arm force estimation model caused by the fatigue state of the operator is reduced, which will seriously cause operational accidents. To solve this problem, the following embodiment proposes a fatigue error compensation method for arm force estimation based on electromyographic signals, aiming to eliminate the influence of muscle fatigue on the dynamic arm force estimation model. In order to overcome the model error of assuming that fatigue develops linearly over time in the traditional fatigue compensation process, an improved Bayesian optimization algorithm is proposed, and the muscle fatigue level is optimized based on the minimization of the verification error of the arm output force estimation model; then, the influence of dynamic output force on the fatigue characteristics of electromyographic signals is considered, and the output force is used as the input of the fatigue level recognition model, and a spatiotemporal feature fusion neural network is proposed. The accuracy of fatigue level recognition in dynamic muscle contraction is improved by fusing spatiotemporal features; then, considering that the arm output force in fatigue level recognition is an unknown parameter, the spatiotemporal feature fusion neural network is improved, and dynamic compensation of fatigue error is realized by model self-iteration.

[0037] The basic scheme of the arm force estimation fatigue error compensation method based on electromyographic signal of this embodiment includes the following steps:

[0038] S1. Fatigue samples are divided into multiple groups according to the time sequence. The arm output force estimation model is used to perform arm strength fitting training on fatigue samples in each group. Based on the verification error of the arm output force estimation model, the fatigue level of fatigue samples is segmented by using the optimized segmentation algorithm.

[0039] S2, inputting the measured arm output force into the fatigue level recognition model to identify the fatigue level;

[0040] S3. Improve the fatigue level identification model, use the self-iterative reverse reasoning mechanism of the improved fatigue level identification model, select the arm output force estimation model corresponding to the fatigue level, and realize dynamic compensation of fatigue error.

[0041] Based on the above concept, the following is further described with embodiments:

[0042] Example 1

[0043] The arm force estimation fatigue error compensation method based on electromyographic signal in this embodiment is as follows: Figure 2 As shown, the following steps are included:

[0044] 1) Fatigue level optimization segmentation

[0045] 1-1) Data Collection

[0046] During data collection, several volunteers were invited to ensure that the volunteers' upper limb muscles did not have strenuous exercise and had enough rest within 48 hours before the experimental collection. The volunteers dragged the assembly pins at the end of the robot back and forth with a maximum force of about 50N. During this period, the shoulder joint angle was -20-30 degrees, the arm's natural vertical posture was 0 degrees, the extension direction was positive, the dragging time was 60 minutes, and the dragging speed was about 5 times per minute. The electromyographic signals of the arm muscle groups, shoulder joint angles, and arm output forces during this process were collected as fatigue samples, such as Figure 3 As shown in the figure, in the experiment, the human subject faced the YZ plane of the robot base coordinate system, and the arm strength and movement direction were divided based on the robot base coordinate system. In the data acquisition process, the electromyographic signal sensor was used to collect the electromyographic signals of the six muscle groups, namely 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. 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. The goniometer was placed on the subject's shoulder to collect the shoulder joint angle. The force sensor was placed at the end of the robot to collect the actual arm output force of the human subject.

[0047] 1-2) Construction of arm output force estimation model

[0048] A long short-term memory neural network with circulation and memory capabilities is used as the muscle force estimation model. The input of the arm output force estimation model is the amplitude information of the electromyographic signals of one or more pairs of antagonist muscles and agonist muscles that are most closely related to the arm output force and the shoulder joint angle information, and the output is the arm output force.

[0049] 1-3) Fatigue time series segmentation method based on improved Bayesian optimization algorithm

[0050] Bayesian optimization can achieve global optimization of parameters by using a probabilistic proxy model to evaluate complex objective functions. However, in fatigue level segmentation, the segmentation points need to be obtained in chronological order, and traditional Bayesian optimization cannot meet the requirement that the optimization parameters are output in chronological order. In addition, the samples between the two segmentation points are the same set of training data, so the segmentation points to be optimized do not meet the basic principle of Bayesian optimization that the optimization parameters are independent of each other.

[0051] Therefore, the traditional Bayesian optimization method is improved. The optimal number of segmentation points and segmentation positions are determined by recursive optimization. The optimization method is as follows: Figure 4 As shown. The fatigue level division model with N-1 segmentation points is used as the basic model, and the verification error of the arm output force estimation model of each group of fatigue samples is minimized as the optimization criterion of the Bayesian optimization model to find the best Nth segmentation point and obtain a new segmentation model. Considering that the original segmentation points of the basic model are not suitable for the new segmentation model, the first N-1 segmentation points of the basic model are re-optimized in turn in a recursive manner.

[0052] 2) Fatigue level identification

[0053] 2-1) In-channel data feature extraction model

[0054] The intra-channel data feature extraction model is a feature extraction model that combines a one-dimensional dilated convolutional neural network with multi-scale attention. The model uses convolution kernels of different scales to extract the feature information of the input signal in the time dimension, thereby fully mining the multi-level information of the data features within each channel of the input information. The input information of the intra-channel data feature extraction model is multi-channel time series data composed of joint angles, arm output force, and fatigue features of electromyographic signals, and the fatigue features of electromyographic signals include amplitude information and median frequency of electromyographic signals. Figure 5 This is the structural diagram of the intra-channel data feature extraction model, the specific form is as follows;

[0055] The one-dimensional dilated convolutional neural network adopts a two-layer multi-scale dilated convolution structure with dilation rates of 1, 2, and 3 respectively. The convolution operation of each layer can be defined as:

[0056]

[0057] In the formula, K MS is the convolution kernel, q is the eigenvalue after dilated convolution, p is the eigenvalue after pooling, ReLU(·) is the ReLU activation function, and X refers to the time series data of a channel in the input information.

[0058] After double-layer one-dimensional dilated convolution, each channel will have 9 branches. The features extracted from each branch are used as the input of the multi-scale attention mechanism to obtain weights of different scales, thereby determining the importance of each branch. First, the feature p of each branch is pooled by the global average pooling (GAP) operation. i Fusion into a data g i .

[0059]

[0060] Next, g iPerform full connection and normalization to calculate the weight coefficient a of each branch c ,

[0061] a c =Softmax(W 2 ReLU(W 1 ·g+b 1 )+b 2 )

[0062] In the formula, g is all g i The matrix composed of g = [g 1 ,g 2 ,…,g N ]; W 1 and W 2 is the weight matrix of the fully connected layer; b 1 and b 2 is the bias term; the Softmax function can transform a set of input real numbers into a set of real numbers ranging from 0 to 1, and the sum of these real numbers is 1. The definition of the softmax function is as follows:

[0063]

[0064] Where z = [z 1 ,z 2 ,…,z N ] is a set of real numbers assumed to be input, softmax(z) i is the i-th component of the output.

[0065] Then, the obtained weight coefficient is used to adjust the feature p of each branch i After weighting, we get the weighted eigenvalue

[0066]

[0067] Finally, the weighted eigenvalues ​​of all branches are stacked in the channel dimension to complete the feature extraction of the data in the channel.

[0068]

[0069] Where P CMSA It is the data feature extracted by the in-channel data feature extraction model; Concat represents the connection operation in the channel dimension.

[0070] 2-1) Inter-channel data feature extraction model

[0071] The data feature extraction model between channels is a feature extraction model that combines the long short-term memory neural network with the multi-head attention mechanism. The model uses the long short-term memory neural network to extract the feature information of the input information in the spatial dimension, and uses the multi-head attention mechanism to fully explore the data feature coupling relationship between each channel. The input information of the multi-head attention sequential network model is multi-channel moment data composed of joint angles, arm output force, and fatigue characteristics of electromyographic signals.

[0072] Figure 6 This is the structural diagram of the inter-channel data feature extraction model, the specific form is as follows:

[0073] First, the input time data is processed by long short-term memory neural network to extract features.

[0074] (h[t],c[t])=LSTM(I[t],h[t-1],c[t-1])

[0075] In the formula, LSTM represents a universal long short-term memory neural network, h and c represent the hidden layer and state unit values ​​of the long short-term memory neural network respectively.

[0076] Collect the values ​​of each hidden layer to obtain the hidden layer data set H = [h[1], h[2], Kh[t]], and perform a linear transformation on the hidden layer data set to obtain the query vector Q of the multi-head attention mechanism {h} , key vector K {h} Sum value vector V {h} ,

[0077]

[0078] In the formula, and Q {h} , K {h} and V {h} The transformation matrix.

[0079] Finally, linear mapping is used to concatenate different attention mechanism heads to complete the feature extraction of inter-channel data.

[0080] P RMHA =concat(A(Q {1} ,K {1} ) {1} ,A(Q {2} ,K {2} ) {2} ,K,A(Q {h} ,K {h} ) {h} )W O

[0081] Where W O is the weight coefficient of the multi-head attention mechanism splicing, P RMHA It is the data features extracted by the inter-channel data feature extraction model.

[0082] 2-3) Spatial-temporal feature fusion neural network

[0083] In order to more comprehensively obtain the feature information in the sample data, the intra-channel data feature extraction model and the inter-channel data feature extraction model in this embodiment perform spatiotemporal feature fusion, and the spatiotemporal feature fusion neural network is called spatiotemporal feature fusion neural network. Figure 7 The amount of data processed by the intra-channel data feature extraction model and the inter-channel data feature extraction model is huge. Therefore, the convolution module is used to further process the extracted feature data, and the global average pooling layer is used to downsample each branch.

[0084] For the downsampled data node P STFF A fully connected network is used for further feature transformation.

[0085] z=σ(W l ·P STFF +b l )

[0086] Where W l is the weight of the fully connected layer; b l is the bias term; σ(·) represents the activation function, which is the Sigmoid function here.

[0087] Next, the Softmax function is introduced to realize the probability conversion of fatigue level, and fatigue level recognition is realized according to the size of probability. The expression is as follows:

[0088]

[0089] In the formula, y i * is the fatigue level of identification, and P(·) is the probability of the current attribute.

[0090] The arm force estimation fatigue invalid compensation method based on spatiotemporal feature fusion neural network is as follows: Step 1, the fatigue samples are divided into multiple groups according to the time sequence, and the arm output force estimation model is used to perform arm strength fitting training on the fatigue samples of each group, and the fatigue level is segmented by the optimized segmentation algorithm according to the verification error of the arm output force estimation model;

[0091] Step 2: Considering the influence of dynamic output force on fatigue characteristics of electromyographic signals, the output force is used as one of the inputs of the fatigue level recognition model, and a spatiotemporal feature fusion neural network is used to accurately identify the fatigue level in dynamic muscle contraction;

[0092] Step 3: Considering that the arm output force in fatigue level identification is an unknown parameter, the spatiotemporal feature fusion network is improved, and the self-iterative reverse reasoning ability of the spatiotemporal feature fusion neural network is utilized to select the arm output force estimation model corresponding to the fatigue level to realize dynamic compensation of fatigue error.

[0093] 3) Fatigue compensation of arm output force

[0094] In the actual application of fatigue level recognition model, the true value of the arm output force is unknown, which also makes the input information of the spatiotemporal feature fusion neural network model incomplete. In order to solve the contradiction caused by the muscle output force being used as both the input item of muscle fatigue recognition and the output item of arm strength estimation. This application proposes a fatigue error dynamic compensation method based on the reverse reasoning mechanism of spatiotemporal feature fusion neural network, such as Figure 8 As shown. Based on the amplitude information and joint angle information of the currently acquired electromyographic signal, the arm strength estimation models of different fatigue levels are used to obtain the estimated arm strength corresponding to different fatigue levels in turn, and the prediction probabilities of the same conditions as the assumed fatigue state are obtained respectively with the help of the spatiotemporal feature fusion network. The fatigue level corresponding to the maximum prediction probability is regarded as the fatigue level of the current muscle, and the arm strength estimated at this fatigue level is regarded as the arm output force after fatigue compensation. Considering the changes in the task requirements of the spatiotemporal feature fusion neural network model, it is necessary to fine-tune the weight parameters of the fully connected layer in the spatiotemporal feature fusion neural network model, so as to optimize the neural network parameters, and further ensure that the fatigue level with the maximum probability in the output value is the same as the actual fatigue level, which is called the improved spatiotemporal feature fusion neural network model.

[0095] This embodiment involves Figure 5-Figure 8 The following are some of the English meanings that appear in the text:

[0096] Padding=True Stride=1Kernel 1@1×3MaxPooling=1×2 is the basic structural information of the convolutional neural network.

[0097] Padding = True: means padding at the edge of the input data to ensure that after the convolution operation, the output size is the same as the input size.

[0098] Stride = 1: means that the convolution kernel slides 1 step on the input data each time.

[0099] Kernel 1@1×3: indicates that a one-dimensional convolution kernel is used, whose height is 1 and width is 3.

[0100] MaxPooling=1×2: indicates that the maximum pooling operation is used, and the height of the pooling window is 1 and the width is 2.

[0101] Example 2

[0102] The arm force estimation fatigue error compensation method based on electromyographic signals in this embodiment is used for a human-robot collaborative control based on a human dynamic arm force estimation model. Compared with Embodiment 1, the difference is that the optimization segmentation algorithm is a random search method, which optimizes the fatigue level segmentation of fatigue samples. The process of the random search method is as follows: S1, define the time length of the fatigue sample as the parameter space of the segmentation point to be optimized; S2, randomly generate a set of segmentation points from the parameter space;

[0103] S3. Group the fatigue samples according to the segmentation points and calculate the verification error of the arm output force estimation model of each group; S4. Record the best segmentation point currently found; S5. Repeat steps S2-S4 until the desired verification error or number of iterations is met.

[0104] Example 3

[0105] This embodiment is an arm force estimation fatigue error compensation method based on electromyographic signals, which is used for human-robot collaborative control based on a human dynamic arm force estimation model. Compared with Example 1, it is different in that the inter-channel data feature extraction model is a combination of a recurrent neural network and a multi-head attention mechanism.

[0106] Example 4

[0107] An arm force estimation fatigue error compensation system based on electromyographic signals is used to implement the arm force estimation fatigue error compensation method, and the system comprises:

[0108] Fatigue level optimization segmentation module, fatigue level identification module and arm output force fatigue compensation module;

[0109] The fatigue level optimization segmentation module is configured to construct an arm output force estimation model using fatigue samples, and perform fatigue level segmentation on the fatigue samples according to the verification error of the arm output force estimation model;

[0110] The fatigue level identification module is configured to input the arm output force into the fatigue level identification model to complete the fatigue level identification;

[0111] The arm output force fatigue compensation module is configured to select an arm output force estimation model corresponding to the fatigue level to complete dynamic compensation of fatigue error.

[0112] The present application has been disclosed as above with preferred implementation cases, but it is not used to limit the present invention. Any technician familiar with this profession can make slight changes or modifications to equivalent implementation cases with equivalent changes by using the above-disclosed structures and technical contents without departing from the scope of the technical solution of the present application, which still fall within the scope of the technical solution of the present application.

Claims

1. A fatigue error compensation method for arm force estimation based on electromyographic signals, characterized in that: The method comprises the following steps: S1. The fatigue samples are divided into multiple groups according to the time sequence. The arm output force estimation model is used to perform arm strength fitting training on the fatigue samples of each group. According to the verification error of the arm output force estimation model, the fatigue level of the fatigue samples is segmented by using the optimized segmentation algorithm. S2, inputting the measured arm output force into the fatigue level recognition model to identify the fatigue level; S3. Improve the fatigue level identification model, use the self-iterative reverse reasoning mechanism of the improved fatigue level identification model, select the arm output force estimation model corresponding to the fatigue level, and realize dynamic compensation of fatigue error.

2. The arm force estimation fatigue error compensation method based on electromyographic signal according to claim 1 is characterized in that: In step S1, the fatigue sample includes electromyographic signals of arm muscle groups, shoulder joint angles and measured arm output forces; the electromyographic signals include amplitude information and median frequency of the electromyographic signals characterizing fatigue characteristics.

3. The arm force estimation fatigue error compensation method based on electromyographic signal according to claim 1 is characterized in that: In step S1, the arm output force estimation model adopts a long short-term memory neural network, the model input is the amplitude information of the electromyographic signal of the arm muscle group and the shoulder joint angle information, and the model output is the arm output force.

4. The arm force estimation fatigue error compensation method based on electromyographic signal according to claim 1, characterized in that: In step S1, the optimization segmentation algorithm is an improved Bayesian optimization algorithm. The specific process of using the optimization segmentation algorithm to perform fatigue level segmentation on fatigue samples is: using the fatigue level division model of the N-1 segmentation points that have been obtained as the basic model, and taking the minimum verification error of the arm output force estimation model of each group of fatigue samples as the optimization criterion of the Bayesian optimization model, finding the best Nth segmentation point, obtaining a new segmentation model, and recursively reoptimizing the first N-1 segmentation points of the basic model in turn until the fatigue level division of the N segmentation points is completed.

5. The arm force estimation fatigue error compensation method based on electromyographic signal according to claim 2 is characterized in that: In step S2, the fatigue level recognition model is an intra-channel data feature extraction model, an inter-channel data feature extraction model or a spatiotemporal feature fusion neural network, and the spatiotemporal feature fusion neural network is obtained by fusing the intra-channel data feature extraction model and the inter-channel data feature extraction model through spatiotemporal features; a convolution module is added to each of the intra-channel data feature extraction model and the inter-channel data feature extraction model for downsampling, and the downsampled data nodes are further converted through a fully connected network to realize fatigue level recognition.

6. The arm force estimation fatigue error compensation method based on electromyographic signal according to claim 2 is characterized in that: In step S2, the fatigue level recognition model adopts an intra-channel data feature extraction model, which is a feature extraction model combining a one-dimensional dilated convolutional neural network with multi-scale attention. The intra-channel data feature extraction model uses convolution kernels of different scales to extract feature information of the input signal in the time dimension, thereby fully mining the multi-level information of the data features within each channel of the input information; the input information is multi-channel time series data composed of fatigue features of joint angles, arm output forces, and electromyographic signals; the specific form of the intra-channel data feature extraction model is as follows: The one-dimensional dilated convolutional neural network adopts a two-layer multi-scale dilated convolution structure with dilation rates of 1, 2, and 3 respectively. The convolution operation of each layer can be defined as: In the formula, K MS is the convolution kernel, q is the eigenvalue after dilated convolution, p is the eigenvalue after pooling, ReLU(·) is the ReLU activation function, and X refers to the time series data of a channel in the input information; After double-layer one-dimensional dilated convolution, each channel will have 9 branches. The features extracted from each branch are used as the input of the multi-scale attention mechanism to obtain weights of different scales, thereby determining the importance of each branch. First, the feature p of each branch is pooled by the global average pooling (GAP) operation. i Fusion into a data g i ; Next, g i Perform full connection and normalization to calculate the weight coefficient a of each branch c , <h2 style=";text-align:left;direction:ltr">a<h2 style=";text-align:left;direction:ltr"> c <h2 style=";text-align:left;direction:ltr"> = Softmax(W2·ReLU(W1·g+b1)+b2) In the formula, g is all g i The matrix composed of g = [g1, g2, ..., g N ]; W1 and W2 are weight matrices of the fully connected layer; b1 and b2 are bias terms; The Softmax function can transform a set of input real numbers into a set of real numbers ranging from 0 to 1, and the sum of these real numbers is 1. The definition of the softmax function is as follows: where z=[z1,z2,…,z N ] is a set of real numbers assumed to be input, softmax(z) i is the i-th component of the output; Then, the obtained weight coefficient is used to adjust the feature p of each branch i After weighting, we get the weighted eigenvalue Finally, the weighted eigenvalues ​​of all branches are stacked in the channel dimension to complete the feature extraction of the data in the channel. Where P CMSA The data features extracted by the in-channel data feature extraction model; Concat represents a concatenation operation on the channel dimension.

7. The arm force estimation fatigue error compensation method based on electromyographic signal according to claim 2 is characterized in that: In step S2, the fatigue level recognition model adopts an inter-channel data feature extraction model, which is a feature extraction model combining a long short-term memory neural network with a multi-head attention mechanism. The inter-channel data feature extraction model uses a long short-term memory neural network to extract feature information of the input information in the spatial dimension, and uses the data feature coupling relationship between each channel mined by the multi-head attention mechanism. The input information is multi-channel time data composed of fatigue features of joint angles, arm output forces, and electromyographic signals; the specific form of the inter-channel data feature extraction model is as follows: First, the input time data is processed by long short-term memory neural network to extract features. (h[t],c[t])=LSTM(I[t],h[t-1],c[t-1]) In the formula, LSTM represents a universal long short-term memory neural network, h and c represent the hidden layer and state unit values ​​of the long short-term memory neural network respectively, and t represents the current adoption time; Collect the values ​​of each hidden layer to obtain the hidden layer data set H = [h[1], h[2], Kh[t]], and perform a linear transformation on the hidden layer data set to obtain the query vector Q of the multi-head attention mechanism {h} , key vector K {h} Sum value vector V {h} , In the formula, and Q {h} , K {h} and V {h} The transformation matrix of The dot product attention is scaled, and the attention weight is expressed as, Finally, linear mapping is used to concatenate different attention mechanism heads to complete the feature extraction of inter-channel data. P RMHA =concat(A(Q {1} ,K {1} )·V {1} ,A(Q {2} ,K {2} )·V {2} ,,...,A(Q {h} ,K {h} )·V {h} )W O ; Where W O is the weight coefficient of the multi-head attention mechanism splicing, P RMHA It is the data features extracted by the inter-channel data feature extraction model.

8. The arm force estimation fatigue error compensation method based on electromyographic signal according to claim 2 is characterized in that: In step S2, the fatigue level recognition model uses the intra-channel data feature extraction model and the inter-channel data feature extraction model to perform spatiotemporal feature fusion to obtain a spatiotemporal feature fusion neural network; The feature data extracted by the intra-channel data feature extraction model and the inter-channel data feature extraction model are further processed by the convolution module, and each branch is downsampled using the global average pooling layer; For the downsampled data node P STFF A fully connected network is used for further feature transformation. z=σ(W l ·P STFF +b l ) Where W l is the weight of the fully connected layer; b l is the bias term; σ(·) represents the activation function, which is the Sigmoid function here; Next, the Softmax function is introduced to realize the probability conversion of fatigue level, and fatigue level recognition is realized according to the size of probability. The expression is as follows: In the formula, y i * is the fatigue level of identification, and P(·) is the probability of the current attribute.

9. The arm force estimation fatigue error compensation method based on electromyographic signals according to claim 8, characterized in that: In step S3, the improved fatigue level recognition model adopts an improved spatiotemporal feature fusion network; the fatigue error compensation process is realized by using the self-iterative reverse reasoning mechanism of the spatiotemporal feature fusion neural network: based on the amplitude information and joint angle information of the currently acquired electromyographic signal, arm strength estimation models of different fatigue levels are used to obtain the estimated arm strength corresponding to different fatigue levels in turn, and with the help of the improved spatiotemporal feature fusion network, the predicted probabilities under the same conditions as the assumed fatigue state are obtained respectively, the fatigue level corresponding to the maximum value of the predicted probability is regarded as the fatigue level of the current muscle, and the arm strength estimated under this fatigue level is regarded as the arm output force after fatigue compensation.

10. An arm force estimation fatigue error compensation system based on electromyographic signals, used to implement the arm force estimation fatigue error compensation method as claimed in claim 1, characterized in that: Contains: fatigue level optimization segmentation module, fatigue level identification module and arm output force fatigue compensation module; The fatigue level optimization segmentation module is configured to construct an arm output force estimation model using fatigue samples, and perform fatigue level segmentation on the fatigue samples according to the verification error of the arm output force estimation model; The fatigue level identification module is configured to input the arm output force into the fatigue level identification model to complete the fatigue level identification; The arm output force fatigue compensation module is configured to select an arm output force estimation model corresponding to the fatigue level to complete dynamic compensation of fatigue error.

Images