CN112232161B - Complex motion continuous estimation method based on electromyography mapping model switching - Google Patents

Abstract

The invention discloses a complex motion continuous estimation method based on myoelectric mapping model switching, which belongs to the field of human-computer interaction. First, according to the joint angle to be estimated and the surface muscles corresponding to the joint movement, joint angle signals and multi-channel myoelectric signals are collected. According to the characteristics of complex tasks, the task is divided into multiple subtasks, and then the collected EMG signals and joint angle signals are preprocessed respectively, and the LSTM deep learning network is used to establish the mapping sub-relationship between EMG signals and joint angle signals, and the estimator is obtained. Model, then, use the RF integrated learning algorithm to establish a mapping relationship between the EMG signal and each estimated sub-model, obtain the RF switching model, use the RF switching model to classify the EMG signal features, and determine the estimated sub-model to which the current motion belongs, Finally, the EMG features are input into the corresponding estimation sub-model to obtain the joint angle estimation. The invention effectively improves the estimation accuracy of each joint angle under complex tasks.

Description

A Continuous Estimation Method for Complex Motion Based on EMG Model Switching

technical field

The invention belongs to the field of human-computer interaction, and more specifically relates to a complex motion continuous estimation method based on myoelectric mapping model switching.

Background technique

Currently, the EMG interface is the only clinically feasible form of neural interface. It uses the surface muscle electromyography signal as a signal source to estimate the movement of each joint of the human body, which has become a clinically feasible solution, but what is really put into application is only the motion pattern recognition based on the electromyography signal. The continuous motion estimation and control of external devices (such as manipulators, manipulators, and exoskeletons) based on electromyographic signals have been invested in a lot of research and achieved certain results, but there are still many problems in the actual clinical application. One of them is The main problem is that the functions of external devices controlled by myoelectricity cannot adapt to complex tasks in real life-for example, we can easily deploy various joints of the upper limbs to complete arm extension (hand reaches the position of the water cup), grasp (finger grasps the water cup), flexion, etc. A series of actions such as raising the arm (taking back the water cup), raising the arm and turning the wrist (drinking water), etc., complete the task of drinking water. However, this series of actions involves the precise coordination and scheduling of multiple joints in time and space. For another example, in daily life, we sometimes need to grasp soft and light objects such as paper cups, and sometimes we need to grasp relatively hard and heavy objects such as pot handles. We will adaptively use different sizes of force to grasp according to the different grasping objects grip.

Changes in factors such as force, posture, and speed involved in the above tasks will make the task more complicated, especially for the continuous estimation of the task using EMG signals. Different muscle groups are activated in time or space to drive different joint movements, and the mapping relationship between EMG signals and joint movements at each stage is different. It can be seen that using a single estimation model to establish the mapping relationship between the EMG signal and the joint angles of the above-mentioned complex tasks often fails to obtain good results, and the model is not robust enough to successfully complete the accurate estimation of the entire task.

Therefore, it is necessary to develop a new complex motion continuous estimation method to successfully complete the accurate estimation of complex tasks.

Contents of the invention

In view of the defects and improvement needs of the prior art, the purpose of the present invention is to provide a complex motion continuous estimation method based on EMG model switching, which aims to solve the technical problem in the prior art that complex tasks cannot be continuously estimated.

In order to achieve the above object, the present invention provides a complex motion continuous estimation method based on myoelectric mapping model switching, which includes the following steps:

S1: According to the joint angle to be estimated and the superficial muscles that drive the corresponding joint movement, collect joint angle signals and multi-channel EMG signals, divide the task into multiple subtasks according to the characteristics of the complex task, and combine multiple subtasks into the entire complex task.

S2: Preprocessing the collected EMG signals and joint angle signals respectively,

S3: Using the preprocessed myoelectric signal and joint angle signal, the LSTM deep learning network is used to establish the mapping sub-relationship between the myoelectric signal and the joint angle signal, and the estimation sub-model is obtained.

S4: Use the RF integrated learning algorithm to establish a mapping relationship between the EMG signal and each estimated sub-model to obtain the RF switching model,

S5: Use the RF switching model to classify the EMG signal features, determine the estimation sub-model to which the current motion belongs, and then input the EMG features into the corresponding estimation sub-model to obtain the estimated value of the joint angle at this time.

Further, in step S2, the preprocessing of the electromyographic signal includes filtering, feature extraction and normalization of the electromyographic signal, specifically, performing notch filter processing on the electromyographic signal to eliminate power frequency interference, and further processing the electromyographic signal The signal is subjected to 20Hz-460Hz band-pass filter processing to eliminate low-frequency noise and high-frequency noise, and feature extraction is performed on the EMG signal after noise reduction processing, and the sliding window method is used to obtain the eigenvalue of the EMG signal, and after the EMG feature is obtained The myoelectric characteristics of each channel are normalized according to the maximum value and the minimum value of the characteristic value of each channel, and the normalized myoelectric characteristics of each channel are obtained.

Further, in step S2, preprocessing the joint angle signal refers to performing filtering processing on the measured joint angle signal and the estimated joint angle signal respectively, so that the measured joint angle and the estimated joint angle are input and output to the corresponding joint angle at a relatively smooth angle. Model.

Further, in step S3, the estimated sub-model obtained is the LSTM estimated sub-model, and the specific process of obtaining the LSTM estimated sub-model is as follows: first, initialize the parameters of the LSTM deep learning network, and set the weight matrix W _C and bias b of the candidate state _C. The weight matrix W _i and bias b _i of the input gate, the weight matrix W _f and bias b f of the forget gate, the weight matrix W _o and the bias _{b o of} the output gate are initialized to random numbers between 0 and 1 , set the number of neurons in the input layer of the LSTM deep learning network to M, the number of layers to i, and the output of each layer is used as the input of the next layer. After the initialization of the LSTM network is completed, the LSTM network is trained, and the preprocessed The myoelectric signal and joint angle signal of the input LSTM model, using the backpropagation algorithm to optimize the above multiple weight matrices and multiple offsets, to obtain the LSTM estimation sub-model.

Further, in step S5, the RF integrated learning algorithm is used to establish a mapping relationship between the EMG signal and each LSTM estimation sub-model to obtain the RF switching model, specifically: the preprocessed EMG signal and the corresponding LSTM model The label of the LSTM model is used as input, and the RF switching model is obtained through training. The output of the RF switching model is the label of the LSTM estimation sub-model. The label of the LSTM model is the number 1~n, and a total of n LSTM estimation sub-models are generated.

Further, the joints described in the joint angle signal are compound joints of the human body such as various joints of the human hand, shoulder and elbow joints, etc., and the multi-channel electromyographic signal is the electromyographic signal generated by the relevant superficial muscles that drive the movement of the compound joint. The number of channels of the electromyographic signal is Select Target determination based on the compound joint angles that need to be estimated.

Further, step S6 is also included, and step S6 is: using the Pearson correlation coefficient and the root mean square error as the evaluation indexes of the measured joint angle and the estimated joint angle, so as to evaluate the estimation accuracy of the estimated joint angle.

Generally speaking, compared with the prior art, the above technical solution conceived by the present invention has the following beneficial effects:

The invention utilizes the advantage of the LSTM deep learning network for the strong long-term memory of the electromyographic signal, and improves the estimation accuracy of a single estimation sub-model. The model switching mechanism is adopted to solve the defect that a single model cannot be accurately estimated in complex tasks, and realize accurate estimation under complex tasks. Experimental results show that the present invention effectively improves the estimation accuracy of each joint angle under complex tasks.

Description of drawings

Fig. 1 is the realization flowchart of the complex motion continuous estimation method based on myoelectric mapping model switching of the present invention;

Fig. 2 is the principle schematic diagram of the LSTM deep learning network that the present invention adopts;

Fig. 3 is the schematic diagram of the principle of the RF switching model adopted by the present invention;

Fig. 4 is the estimation accuracy curve chart that LSTM sub-model of the present invention is applied to five fingers;

Fig. 5 is a comparison graph of the estimation accuracy between the LSTM hybrid estimation model combined with the model switching mechanism and the single LSTM sub-model in complex grasping tasks according to the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The invention relates to a compound motion continuous estimation method under complex tasks based on an electromyographic signal combined with a model switching mechanism and an LSTM deep learning network. A LSTM hybrid estimation model combined with a model switching mechanism is proposed to realize continuous estimation of human joint compound motion under complex tasks. LSTM has great advantages in dealing with time series problems. The introduction of cell state to save long-term memory improves the defect that RNN cyclic neural network can only save short-term memory. Using LSTM deep learning network to build a mapping model between EMG signals and human joint motion can improve the estimation accuracy compared with other machine learning.

Specifically, for different task scenarios in real life, the task is divided into multiple sub-tasks according to the complex characteristics of the task, such as decomposing into n tasks, and different sub-tasks are trained using the LSTM deep learning network to obtain multiple LSTM estimation sub-models. The random forest (RF) ensemble learning algorithm is used to train the EMG signal and its LSTM estimation sub-model to obtain the switching model.

The technical scheme that the present invention takes comprises the steps:

S1: Collect joint angle signals and multi-channel EMG signals according to the joint angles to be estimated and the superficial muscles that drive the corresponding joint movements. According to the characteristics of the complex task, the task is divided into multiple subtasks, and the multiple subtasks are combined into the whole complex task. According to the requirements of n subtasks decomposed from the complex task, the multi-degree-of-freedom human joint angle is selected as the estimation target, and the multi-channel EMG signal is collected as the system input.

S2: Preprocessing the EMG signal, including filtering, feature extraction, and normalization of the EMG signal. Specifically, because the surface electromyography signal is too weak to be interfered by the surrounding electromagnetic field, in the process of actually measuring the electromyography signal, it is easy to be interfered by the power frequency of the joint angle measurement equipment and other electrical equipment, and the interference frequency distribution is 50Hz and its integer multiples frequency. Therefore, it is necessary to eliminate power frequency interference on the EMG signal. Bandpass filtering is performed according to the main frequency of the EMG signal to eliminate motion artifact noise (low frequency) and other high frequency noise. Feature extraction is performed on the noise-reduced EMG signal, and different single features or combined features are selected according to different tasks, such as root mean square (RMS), waveform length (WL), absolute value integral average (IAV), Mean absolute value (MAV), zero-crossing coefficient (ZC), slope sign change (SSC) and average amplitude change rate (AAC), etc. The EMG features of each channel were normalized according to the maximum and minimum eigenvalues of each channel, and the normalized EMG features of each channel were obtained as the input signals of the RF switching model and the LSTM model.

The joint angle signal is preprocessed, and the measured joint angle signal and the estimated joint angle signal are filtered, so that the joint angle is input and output to the model at a relatively smooth angle.

S3: The EMG features of each subtask and the corresponding joint angle signals are used as input to train the LSTM estimation submodel respectively.

Initialize the parameters of the LSTM deep learning network: the weight matrix W _C and bias b _C of the candidate state, the weight matrix W _i and bias b _i of the input gate, the weight matrix W _f and bias b _f of the forget gate, and the output gate The weight matrix W _o and bias b _o are initialized as random numbers between 0 and 1. The number of neurons in the input layer of the LSTM deep learning network is set to M, the number of layers is set to i, and the output of each layer is used as the input of the next layer.

After the LSTM network initialization is completed, the LSTM network is trained, and the preprocessed EMG signal and joint angle signal are input into the LSTM model, and the backpropagation algorithm is used to optimize the above weight matrix and bias to obtain the LSTM estimation sub-model. The LSTM model is also the LSTM network.

Among them, set the number of training iterations as epoch, the loss function as loss, the optimizer as optimizer, and the number of samples selected for one training as batch size.

S4: Training the RF switching model, the preprocessed EMG signal and the label of the corresponding LSTM model are used as model input to train the RF switching model, and the output of the RF switching model is the label of the LSTM estimation sub-model. Perform bootstrap sampling on the input samples to obtain k sample sets, respectively use the decision tree model to train the k sample sets to obtain k tree classifiers, in which the labels of the LSTM model are numbers 1 to n, and a total of n LSTM estimations are generated In the sub-model, the maximum depth of each tree classifier is set to m, and the gini function calculation method is used to judge whether the node continues to split. The minimum number of samples required for node splitting is set to f, and the maximum number of leaf nodes is not limited. Participate in node splitting The maximum number of features judged is equal to the number of all features.

S5: After training the RF switching model and each LSTM estimation sub-model, the entire LSTM hybrid estimation model combined with the model switching mechanism is trained. In the test phase, this hybrid model is used to perform continuous motion under complex tasks on the new input EMG data. estimate. In particular, in the estimation phase, the maximum and minimum values of the EMG features of each channel in the training phase are used to normalize the EMG features of each channel, and the normalized EMG features are input into the RF switching model, and the LSTM estimation corresponding to the EMG signal is output. sub-model, and then the normalized EMG features are input to the corresponding LSTM estimation sub-model for estimation, and the estimated value of each joint angle is output.

In actual engineering practice, the compound motion continuous estimation effect of the LSTM hybrid estimation model combined with the model switching mechanism proposed in the present invention will be evaluated under complex tasks. When evaluating, first, evaluate the performance of a single LSTM model in each estimation sub-model Estimate the accuracy to verify the estimation accuracy of each estimation sub-model, and then compare the estimation effect of the single LSTM model and the LSTM hybrid estimation model combined with the model switching mechanism under complex tasks to verify whether the LSTM hybrid estimation model combined with the model switching mechanism can It better adapts to changes in task scenarios, and has higher estimation accuracy and robustness for complex task estimation. The evaluation indicators are Pearson's correlation coefficient (CC) and root mean square error (RMSE). The results show that the method of the invention can effectively improve the estimation accuracy of each joint angle under complex tasks.

The whole flow chart of the present invention when estimating is shown in Figure 1. The myoelectric signal is first preprocessed to obtain the myoelectric feature, and the myoelectric feature is input into the RF switching model to obtain the classification result, and the label of the LSTM estimation sub-model is output, and then the myoelectric feature Enter the corresponding LSTM estimation sub-model to output the estimated compound motion angle. Due to the need to compare the results, the eigenvalues obtained after the EMG signal is preprocessed will also directly pass through a single LSTM model to directly obtain the compound motion angle output. The angle outputs obtained by the two models will be compared in the two dimensions of CC and RMSE to highlight the good estimation effect of the LSTM hybrid estimation model combined with the model switching mechanism under complex tasks.

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments. In particular, the complex task included in the specific embodiment is that people need to use different forces to grasp different objects in daily life, for example, when grasping a paper cup, it is necessary to ensure that a small force is used when grasping, and for grasping a pot handle So when you pick up a cooking pot, you need to use a lot of force, but you only need moderate force to hold a bottle of water in daily life.

(1) According to the characteristics of the complex task, the task is divided into multiple subtasks, and the multiple subtasks are combined into the whole complex task. According to the characteristics that people will use different forces to grasp different objects in the above-mentioned daily life situations, the five-degree-of-freedom human hand joint angle is selected as the estimation target, and the six-channel EMG signal is used as the system input. The five-degree-of-freedom human hand joints are the metacarpophalangeal joint of the thumb, the metacarpophalangeal joint of the index finger, the metacarpophalangeal joint of the middle finger, the metacarpophalangeal joint of the ring finger, and the metacarpophalangeal joint of the little finger. The selected six-channel muscle signals are flexor hallucis longus, flexor digitorum superficiale, flexor digitorum deep, extensor digitorum, extensor digitorum longus, and extensor hallucis longus. According to the joint angle to be estimated and the superficial muscles that drive the corresponding joint movement, the joint angle signal and the multi-channel EMG signal are collected.

(2) Preprocessing the EMG signal, including filtering, feature extraction, and normalization of the EMG signal.

(2a) Filtering: Since the surface electromyography signal is too weak to be interfered by the surrounding electromagnetic field, in the process of actually measuring the electromyography signal, it is subjected to power frequency interference from finger angle measurement equipment and other electrical equipment, and the interference frequency distribution is 50Hz and above. Integer multiple frequency. Therefore, a notch filter (Comb-Notching Filter) was performed on the EMG signal to eliminate power frequency interference. Since the main frequency of EMG is concentrated at 20-460 Hz, the EMG signal is processed by band-pass filtering at 20-460 Hz to eliminate motion artifact noise (low frequency) and other high-frequency noise.

(2b) Feature extraction is performed on the noise-reduced EMG signal. The main change factor in complex situational tasks is force. The average absolute value feature of the EMG signal represents the activity level of the muscle, which can reflect the grip force to a certain extent. size, so the mean absolute value (MAV) is chosen as the feature. The sliding window method is used to obtain the eigenvalues of the EMG signal. The size of the sliding window is set to 200ms, and each time slides backward for 50ms. There is 75% overlapping data between adjacent windows.

(2c) After obtaining the myoelectric features, normalize the myoelectric features of each channel according to the maximum and minimum values of the feature values of each channel, and obtain the normalized myoelectric features of each channel as the subsequent RF switching model and LSTM model (the next two models) input signal.

The five-degree-of-freedom finger metacarpophalangeal joint angle signal is preprocessed, and the measured angle signal is subjected to Butterworth low-pass filtering (cutoff frequency 2Hz, 4th order), so that the joint angle can be input into the LSTM estimation model at a relatively smooth angle.

输入门的输出值i_t，遗忘门的输出值f_t和输出门的输出值o_t，其中：(3) Divide the task into 3 subtasks according to the complexity of the task, task 1: grasp the object with a small force, task 2: grasp the object with a moderate force, task 3: grasp the object with a large force . The EMG features of each subtask and the corresponding joint angle signals are used as input to train the LSTM estimation sub-model respectively, that is, three LSTM estimation sub-models corresponding to three different levels of force are trained. The principle diagram of a single LSTM deep learning network is shown in Figure 2. When training each model, the preprocessed EMG signal and joint angle signal are input to the input gate, forget gate, output gate and candidate gate in the LSTM deep learning network. In the state, the output value of the candidate state of the LSTM deep learning network at time t is obtained

The output value _it of the input gate, the output value f _t of the forget gate and the output value o _t of the output gate, where:

i _t =σ(W _i ·[h _t-1 ,x _t ]+b _i )

f _t ＝σ(W _f ·[h _t-1 ,x _t ]+b _f )

o _t ＝σ(W _o ·[h _t-1 ,x _t ]+b _o )

h _t ＝o _t *tanh(C _t )

分别表示候选态在t、t-1时刻的输出，h_t、h_t-1分别表示LSTM深度学习网络在t、t-1时刻的输出，σ(·)表示sigmoid激活函数，tanh(·)表示tanh激活函数。Among them, x _t is the EMG signal after preprocessing, which represents the input of the model. W _C and b _C represent the weight matrix and bias of the candidate state, respectively, W _i and b _i represent the weight matrix and bias of the input gate, respectively, W _f and b _f represent the weight matrix and bias of the forget gate, respectively, W _o and b _o denote the weight matrix and bias of the output gate respectively, C _t and C _t-1 denote the output of the cell state storing the long-term memory at time t and t-1 respectively,

Represent the output of the candidate state at time t and t-1 respectively, h _t and h _t-1 represent the output of the LSTM deep learning network at time t and t-1 respectively, σ(·) represents the sigmoid activation function, tanh(·) Represents the tanh activation function.

Initialize the parameters of the LSTM deep learning network: the weight matrix W _C and bias b _C of the candidate state, the weight matrix W _i and bias b _i of the input gate, the weight matrix W _f and bias b _f of the forget gate, and the output gate The weight matrix W _o and bias b _o are initialized as random numbers between 0 and 1. The number of neurons in the input layer of the LSTM deep learning network is set to M, the number of layers is set to i, and the output of each layer is used as the input of the next layer.

After the LSTM network initialization is completed, train the LSTM network: input the preprocessed EMG signal and joint angle signal into the LSTM model, and use the backpropagation algorithm to optimize the above weight matrix and bias to obtain the LSTM training model, three LSTM The model performs the same operation to obtain three different LSTM estimation sub-models. Among them, set the number of training iterations as epoch, the loss function as loss, the optimizer as optimizer, and the number of samples selected for one training as batch size. In this embodiment, M=256, i=2, epoch=150, loss=mae (mean absolute error), optimizer=Adam (Adam optimizer), batch size=16.

(4) Training RF switching model: the preprocessed EMG signal and the label of the corresponding LSTM model are used as model input to train the RF switching model, and the output of the RF switching model is used as the label of the LSTM estimation sub-model. The schematic diagram of the RF switching model is shown in Figure 3. The input samples are sampled by bootstrap to obtain k sample sets, and the k sample sets are trained using the decision tree model to obtain k tree classifiers. Among them, the LSTM sub-model The labels of each tree are numbered from 1 to n (a total of n LSTM estimation sub-models are generated), the maximum depth of each tree classifier is set to m, the calculation method of the gini function is used to judge whether the node continues to split, and the minimum number of samples required for splitting is set For f, there is no limit to the maximum number of leaf nodes, and the maximum number of features involved in judging when a node splits is equal to the number of all features. In this embodiment, the sample set is divided into 10 sample sets according to bootstrap sampling, that is, k=10, there are three LSTM estimation sub-models, so n=3, the maximum depth of each tree classifier is set to 2, that is, m=2, and the split The minimum number of samples required is set to 2, ie f=2.

(5) After training the RF switching model and each LSTM estimation sub-model, the LSTM hybrid estimation model combined with the model switching mechanism is trained, and this model is used to continuously estimate compound motion under complex tasks for new input EMG data. In the estimation phase, the maximum value and minimum value of the EMG features of each channel in the training phase are used to normalize the EMG features of each channel, and the normalized EMG features are input into the RF switching model, and the LSTM estimation sub-model corresponding to the EMG features is output. Then the normalized EMG feature is input into the corresponding LSTM estimation sub-model for estimation, and the estimated value of each joint angle is output.

The LSTM hybrid estimation model with the ensemble model switching mechanism proposed by the present invention is evaluated on the continuous estimation effect of myoelectric compound motion under complex tasks.

Evaluate the estimation accuracy of a single LSTM model in each estimation sub-model, and the evaluation indicators are Pearson correlation coefficient (CC) and root mean square error (RMSE). The following describes the estimation accuracy of the LSTM deep learning network through a single LSTM deep learning network constructed under a moderate force situation in this embodiment: average CC=0.944, average RMSE=11.63°. CC>0.8 indicates that there is a strong correlation between the estimated angle and the measured angle. As shown in Figure 4, RMSE indicates that the root mean square error of the five-finger angle is 11.63°.

Comparing the estimation effects of the single LSTM model and the LSTM hybrid estimation model combined with the model switching mechanism under complex tasks, it can be seen that the results show that the single LSTM deep learning network cannot adapt to accurate estimation under complex tasks, while the LSTM combined with the model switching mechanism The hybrid estimation model can adapt to the scene change to achieve accurate estimation under complex tasks. The estimation effect of the single LSTM and the LSTM hybrid estimation model combined with the model switching mechanism under the complex task of force change in this embodiment will be further explained below.

Use the single LSTM sub-model trained under different force levels to perform cross-estimation of the grasping motion under different force levels, and use the LSTM hybrid estimation model combined with the model switching mechanism to estimate the grasping motion under different force levels. Continuous estimation is compared, and the estimation effects of each model are shown in Table 1 and Table 2 below, where the L, M, and H models represent the LSTM sub-models trained using low-level force, medium-level force, and high-level force, respectively, and the S model represents LSTM Hybrid Estimation Model with Model Switching Mechanism. L, M, and H forces respectively represent the low-level force, middle-level force, and high-level force input during estimation. Among them, the evaluation index of the estimated results in Table 1 is the Pearson correlation coefficient (CC), and the evaluation index of the estimated results in Table 2 is the root mean square error (RMSE).

Table 1 The evaluation index of the estimated results is the result of the Pearson correlation coefficient

Table 2 The evaluation index of the estimated results is the result of the root mean square error

Table 1 and Table 2 show the estimation effect of each model for different levels of force. It can be seen from the table that the single LSTM model under each force level condition has a good estimation effect only for the force level condition under this model. For Estimated effects become worse for other levels of force. The LSTM hybrid estimation model combined with the model switching mechanism has better estimation results for different levels of force conditions. The comparison results show that the LSTM hybrid estimation model combined with the model switching mechanism has better robustness and estimation accuracy for complex tasks. Figure 5 shows the comparison between the estimation model trained under low-level force conditions and the LSTM hybrid estimation model combined with the model switching mechanism for estimating the grip under three force level conditions. It can be seen from Figure 5 that the angle curve estimated by the LSTM hybrid estimation model combined with the model switching mechanism in the case of three kinds of force levels mixed fits the measured angle curve well, while the single LSTM estimated by training under low-level force conditions The angle curves estimated by the model under the conditions of medium force and high force are seriously inconsistent with the measured angle curves. Therefore, a single LSTM deep learning network cannot adapt to accurate estimation under complex tasks, while the LSTM hybrid estimation model combined with the model switching mechanism can adapt to scene changes to achieve accurate estimation under complex tasks.

The method of the present invention can automatically switch under different scenarios, and use the LSTM estimation sub-model to perform accurate estimation after the scenario is determined, to ensure the robustness of the algorithm under complex tasks, and to realize the continuous compound movement under complex tasks based on electromyographic signals estimate.

In the present invention, in order to accurately distinguish between the training phase, the estimation phase and different estimation models, the LSTM deep learning network, LSTM model, and LSTM sub-model are used in the invention. Use the LSTM deep learning network before getting an accurate estimation model, that is, before training all the parameters of the model. The LSTM model refers to the estimation model after all the parameters in the LSTM deep learning network have been trained. The LSTM sub-model refers to the specific The estimation models corresponding to different scenarios in the same complex task involved in the invention, the RF switching model refers to the estimation model that uses the RF integrated learning algorithm to accurately identify the myoelectric feature as a certain scenario in a complex task, LSTM is LongShort-TermMemory Acronym. RF is an acronym for RandomForest, which means Random Forest Classifier.

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, All should be included within the protection scope of the present invention.

Claims (6)

1. A complex motion continuous estimation method based on electromyography model switching is characterized by comprising the following steps:

s1: collecting joint angle signals and multi-channel electromyographic signals according to the joint angle to be estimated and surface muscles driving corresponding joints to move, dividing a task into a plurality of subtasks according to the characteristics of a complex task, combining the plurality of subtasks into the whole complex task,

s2: respectively preprocessing the collected electromyographic signals and joint angle signals,

s3: utilizing the preprocessed electromyographic signals and joint angle signals, adopting an LSTM deep learning network to establish a mapping sub-relation of the electromyographic signals and the joint angle signals to obtain an estimation sub-model,

s4: establishing a mapping relation between the electromyographic signals and each estimation submodel by using an RF ensemble learning algorithm to obtain an RF switching model,

s5: classifying the electromyographic signal characteristics by using an RF switching model, judging an estimation submodel to which the current movement belongs, inputting the electromyographic characteristics into the corresponding estimation submodel to obtain a joint angle estimation value at the moment,

the method comprises the following steps of establishing a mapping relation between an electromyographic signal and each LSTM estimation submodel by using an RF ensemble learning algorithm to obtain an RF switching model, and specifically comprises the following steps: and taking the preprocessed electromyographic signals and the corresponding labels of the LSTM models as input, training to obtain an RF switching model, wherein the output of the RF switching model is the label of the LSTM estimation sub-model, the labels of the LSTM models are numbers 1-n, and n LSTM estimation sub-models are formed jointly.

2. The method of claim 1, wherein the step S2 of pre-processing the electromyographic signals comprises filtering, feature extraction and normalization of the electromyographic signals,

in particular, the electromyographic signals are subjected to notch filtering processing to eliminate power frequency interference, band-pass filtering processing of 20 Hz-460 Hz is also carried out on the electromyographic signals to eliminate low-frequency noise and high-frequency noise,

extracting the feature of the electromyographic signal after noise reduction, adopting a sliding window method to extract the feature value of the electromyographic signal,

and after the electromyographic features are obtained, normalizing the electromyographic features of the channels according to the maximum value and the minimum value of the characteristic values of the channels respectively to obtain the electromyographic features after normalization of the channels.

3. The method as claimed in claim 2, wherein the step S2 of preprocessing the joint angle signal is to filter the measured joint angle signal and the estimated joint angle signal, respectively, so that the measured joint angle and the estimated joint angle are input and output to the corresponding model at a relatively smooth angle.

4. The method for continuous estimation of complex motion based on electromyography model switching as claimed in claim 3, wherein the estimation submodel obtained in step S3 is an LSTM estimation submodel, and the specific process for obtaining the LSTM estimation submodel is as follows:

firstly, initializing parameters of the LSTM deep learning network, and converting the weight matrix W of the candidate state _C And bias b _C Weight matrix W of input gates _i And bias b _i Weight matrix W of forgetting gate _f And bias b _f Weight matrix W of output gates _o And bias b _o Initializing to random number between 0 and 1, setting the neuron number of the input layer of the LSTM deep learning network as M, setting the layer number as i, using the output of each layer as the input of the next layer,

after the initialization of the LSTM deep learning network is completed, the LSTM deep learning network is trained, the preprocessed electromyographic signals and the preprocessed joint angle signals are input into the LSTM deep learning network, and the plurality of weight matrixes and the plurality of biases are optimized by adopting a back propagation algorithm to obtain an LSTM estimation sub-model.

5. The method for continuously estimating complex motion based on electromyography model switching as claimed in claim 4, wherein the joints in the joint angle signals are human body compound joints including joints of human hands and joints of shoulders and elbows, and the multichannel electromyography signals are electromyography signals generated by relevant surface muscles driving the compound joints to move.

6. The method for continuous estimation of complex motion based on electromyography model switching according to claim 5, further comprising step S6, wherein step S6 is: and adopting the Pearson correlation coefficient and the root mean square error as evaluation indexes of the actually measured joint angle and the estimated joint angle so as to evaluate the estimation precision of the estimated joint angle.

Images