CN114647314A

CN114647314A - Wearable limb movement intelligent sensing system based on myoelectricity

Info

Publication number: CN114647314A
Application number: CN202210279960.0A
Authority: CN
Inventors: 高忠科; 王洋阳; 马超
Original assignee: Tianjin University
Current assignee: Tianjin University
Priority date: 2022-03-21
Filing date: 2022-03-21
Publication date: 2022-06-21

Abstract

The invention provides a wearable limb movement intelligent sensing system based on myoelectricity, which comprises wearable myoelectricity acquisition equipment, a hand myoelectricity active training interface, a physiological state analysis module and a remote unmanned aerial vehicle module. Wearing formula flesh electricity collection equipment including: the system comprises an acquisition part connected with a human body and acquisition equipment, a bioelectricity signal acquisition module used for amplifying and converting an electromyographic signal, an STM32 microprocessor used for controlling the acquisition of the electromyographic signal, a physiological state analysis module and transmitting an sEMG electromyographic signal. The hand myoelectricity active training interface comprises a mobile phone end display screen, a WIFI module and a mobile phone end processor and is used for training hand actions. The physiological state analysis module identifies the current intention action of the user and sends the instruction to the unmanned aerial vehicle module. The unmanned aerial vehicle module sends a signal to control the remote unmanned aerial vehicle module through the wireless transmission circuit according to the result output by the physiological state analysis module. The collection, classification, and feedback of muscle signals to train a machine to understand the intent of a wearer's gesture.

Description

Wearable limb movement intelligent sensing system based on myoelectricity

Technical Field

The invention relates to a system for intelligently sensing limb actions. In particular to a wearable limb movement intelligent perception system based on myoelectricity.

Background

In limb movement, hand movement is an indispensable limb language in life, is applied to daily communication in a large number, and with the development of scientific technology, intelligent recognition and application of hand movement gradually enter daily life, and becomes an embodiment of intelligent life. The human hand muscle electric signals are collected, classified identification is carried out, human gesture intentions are understood, relevant responses are made, and the human hand muscle electric signals gradually become an important trend of intelligent life. Before the hand movement, corresponding hand muscle electric signals are generated.

Whenever cells in the hand are stimulated by an electric shock or are activated by nerves, the body generates corresponding muscle potentials. Therefore, the electrical signals of the human muscle are measured at this time, and data analysis is performed, so that the corresponding muscle movement of the human body can be detected and analyzed, or the level of the stimulated neuron can be known. Therefore, when the muscle contracts or expands, the transmitted electric signals can reflect the movement conditions of the nerve and the muscle to a certain extent, and then the training classification is carried out by utilizing an algorithm.

The technology of collecting human body electric signals is a new way of establishing an information channel between a human body and a computer. The human body electric signal acquisition technology is used for acquiring and analyzing hand electromyographic signals of a testee, extracting rich characteristics contained in the electromyographic signals, further judging the hand action state of the testee, and possibly being used for prosthesis motion control, clinical hand disease detection, clinical diagnosis of motion injury, athlete training, medical rehabilitation and improvement of daily life activities, even in some game and entertainment fields.

However, most of the existing multi-channel myoelectricity acquisition devices have the inconveniences of high price, heavy volume, complex operation, incapability of real-time classification and the like, and the wearable myoelectricity acquisition devices generally have the defects of insufficient precision, fewer channels and the like. Therefore, it is necessary to design and develop a set of high-precision and multi-channel wearable myoelectric acquisition equipment and use the equipment in the field of human body electric signal acquisition.

An embedded single chip microcomputer (STM32) microprocessor with high performance, low cost and low power consumption is widely applied to the application fields of industrial control, consumer electronics, Internet of things, communication equipment, medical service, security monitoring and the like. Wearing formula flesh electricity collection equipment based on STM32 stability and the high performance conceptual design of operation has solved the passageway less, wireless transmission scheduling problem. Therefore, the STM32 is used as a main control chip of the wearable myoelectricity acquisition equipment and matched with the high-precision bioelectricity signal acquisition module, and the acquisition precision of the equipment can be met.

Electromyographic signals are bioelectric signals that originate from any tissue or organ, typically as a function of time and a series of amplitudes, frequencies, and waveforms. The electromyographic signals are bioelectric signals generated along with muscle contraction, and the collected electromyographic signals on the skin surface are called surface electromyographic signals sEMG. sEMG electromyographic signals are bioelectric currents generated by contraction of muscles on the surface of the human body. The nervous system controls the activity of the muscles (contraction or relaxation) and the different muscle fiber motor units in the surface skin generate mutually different signals at the same time. Therefore, the electromyographic signals have the characteristics of nonlinearity, unsteadiness, serious noise interference and the like.

With the rise of deep learning and the improvement of computing power, the classification judgment according to the algorithm and the real-time judgment of human body movement become possible. In recent years, deep learning has shown its powerful potential in the fields of object detection, speech recognition, and natural language processing. The deep convolutional neural network is a representative of successful applications in deep learning, and can effectively extract features in the grid-like data. The models of the self-coding neural network and the deep convolution neural network of the multi-layer neurons can exert the corresponding advantages of the respective components, extract the time domain and frequency domain characteristics of the corresponding electromyographic signals, and further realize accurate identification of the hand motions of the testee by combining an attention mechanism.

Attention mechanism has become an important component of neural network architecture and has a number of applications in the fields of natural language processing, statistical learning, speech and computers, etc. Note that the model aims to mitigate the fixed length-induced information loss challenge by allowing the decoder to access the entire encoded input sequence. The core idea is to introduce attention weights on the input sequence to prioritize the set of locations where relevant information exists to generate the next output. The attention model may be more effective in increasing the accuracy of the identification.

Disclosure of Invention

The invention aims to solve the technical problem of providing a wearable limb movement intelligent sensing system based on myoelectricity, which can carry out quick and accurate primary identification on the hand movement of a testee so as to control a four-rotor aircraft.

The technical scheme adopted by the invention is as follows: wearable equipment based on ADS of low-power consumption gathers chip design is used for gathering human surface muscle signal of telecommunication (sEMG), and at the acquisition process, sEMG signal is through filtering, enlargies, and is digital, transmits to the cell-phone end through wireless transmission equipment again, in the dedicated APP of cell-phone, carries out digital filtering, and the application deep learning algorithm trains and classifies. This research has set up and has developed the cell-phone end software platform that can carry out data analysis, can be fast, accurate classify to multiple gesture, and the classification effect can reach higher level to in the APP of cell-phone end, can look over the real-time waveform of sEMG flesh signal. After the signals are collected, the predicted signals can be transmitted to the four-rotor aircraft for the next action. The higher accuracy rate indicates that the set of system functions is feasible. In the invention, a detailed algorithm structure and a hand recognition action based on neural network training combined with an attention mechanism are provided.

Drawings

FIG. 1 is a block diagram of the wearable limb movement intelligent sensing system based on myoelectricity;

FIG. 2 is a block diagram of the wearable myoelectricity collecting device according to the present invention;

FIG. 3 is a block diagram of a wireless transmission module according to the present invention;

FIG. 4 is a block diagram of the parallel CNN algorithm of the present invention incorporating the attention mechanism;

FIG. 5 is a block diagram of a first CNN algorithm incorporating attention mechanism in the present invention;

fig. 6 is a block diagram of a second CNN algorithm incorporating attention mechanism according to the present invention.

Detailed Description

The wearable limb movement intelligent perception system based on myoelectricity is described in detail with reference to examples and the accompanying drawings.

As shown in fig. 1, the wearable limb movement intelligent sensing system based on myoelectricity comprises: wearable myoelectricity collection equipment (1), hand myoelectricity initiative training interface (2), physiological state analysis module (3) and long-range four rotor control module (4). The hand myoelectric active training device is characterized in that a wearer collects sEMG myoelectric signals from the hand of the wearer by applying the wearable hand myoelectric collecting device (1) through the hand myoelectric active training interface (2).

The physiological state analysis module (3) receives the sEMG electromyographic signals and intelligently identifies the hand movements of the wearer by combining a convolutional neural network algorithm (CNN) of an attention mechanism. In the invention, a mobile phone end software platform capable of carrying out data analysis is built and developed, namely, the function of the physiological state analysis module (3) can be realized at a mobile phone end, and the waveform of each channel sEMG electromyography acquisition device can be checked. After intelligent identification is carried out on the mobile phone end by utilizing a deep learning convolutional neural network algorithm (CNN), the quad-rotor unmanned aerial vehicle equipment (4) is remotely controlled through wireless remote equipment.

Remote four rotor control module (4) regard physiological status analysis module (3) as the basis, through wireless remote equipment, control remote four rotor unmanned aerial vehicle equipment (4) make with the corresponding motion of hand action, the gyroscope module also passes through wireless remote equipment in addition, sends data, combines together with physiological status analysis module (3), controls remote four rotor unmanned aerial vehicle equipment (4) jointly.

As shown in fig. 2, the wearable limb movement intelligent sensing system based on myoelectricity is characterized in that the wearable myoelectricity collection device (1) comprises: the electrode paste and the lead wire thereof (11) which are connected in sequence are used for collecting sEMG electromyographic signals, a bioelectric signal collecting module (12) used for amplifying and converting the electromyographic signals, an STM32 microprocessor (13) and a WIFI wireless data transmission circuit (14) used for controlling the collection of the electromyographic signals and transmitting the sEMG electromyographic signals to a physiological state analyzing module (2), a gyroscope circuit (15) is used for measuring the acceleration, the angular velocity and the angle of the wearable electromyographic signal collecting device (1), and is respectively connected with a system power supply circuit (16) of the bioelectric signal collecting module (12) and the STM32 microprocessor (13) and an NRF24L01 wireless remote circuit (17), wherein the electrode paste and the electrode paste in the lead wire thereof (11) are used for collecting the sEMG electromyographic signals of different muscles of hands and are connected with the bioelectric signal collecting module (12) through the lead wire and a PJ313B interface, the device is used for collecting and transmitting bioelectricity signals; the electrode paste is pasted on the forearm of the hand of a testee, and the muscle measured by the myoelectricity acquisition equipment is as follows: extensor carpi ulnaris, extensor digitorum, extensor radialis brachialis, extensor radialis longus, brachioradialis, circumflex, flexor radialis, flexor palmaris longus, and flexor ulnaris, and myoelectrical signals of 8 channels of the hand can be obtained.

The bioelectrical signal acquisition module (12) is composed of a plurality of bioelectrical signal acquisition chips which are integrated with a high common mode rejection ratio analog input AD module for receiving human body surface muscle voltage sEMG signals acquired by the electrode patches, a low-noise programmable gain amplifier for measuring the electromyogram voltage sEMG and amplifying the electromyogram signals and a high-resolution synchronous sampling analog-to-digital converter for converting the analog signals into digital signals.

The STM32 microprocessor (13) is used for adjusting the acquisition mode of the bioelectrical signal acquisition module, adjusting and controlling the WIFI wireless data transmission module (14) and the NRF24L01 wireless remote circuit (17) to output sEMG electromyographic signals, and sending the sEMG electromyographic signals to a mobile phone end so as to be used for the physiological state analysis module (3) to analyze data. And carrying out data analysis on the acceleration, the angular velocity and the angle signal measured by the gyroscope circuit (15), and finally sending the signals to the remote quad-rotor unmanned aerial vehicle equipment (4).

The WIFI wireless data transmission module (14) works in an AP mode, the highest transmission rate is 4Mbps, and under the control of an STM32 microprocessor (13), collected sEMG electromyographic signals are periodically output to a hand electromyographic active training interface (2) and a physiological state analysis module (3) at a mobile phone end in a data packet mode through the WIFI wireless data transmission module (14).

The NRF24L01 wireless remote circuit (17) has the highest transmission rate of 2Mbps, under the control of an STM32 microprocessor (13), the collected sEMG electromyographic signals are periodically output to a hand electromyographic active training interface (2) and a physiological state analysis module (3) through the NRF24L01 wireless remote circuit (17) in the form of data packets, and in addition, the NRF24L01 wireless remote circuit (17) can send the data measured by the gyroscope circuit (15) to a remote quadrotor unmanned aerial vehicle device after simple analysis of the STM32 microprocessor (13), as shown in FIG. 3;

the gyroscope circuit (15) is mainly applied to a six-axis attitude angle sensor, and a corresponding circuit is designed. The gyroscope circuit (15) adopts a high-precision gyroscope accelerometer MPU6050, reads the measurement data of the MPU6050 through an STM32 microprocessor (13), then outputs through a serial port, integrates an attitude resolver in the module, and can accurately output the acceleration, the angular velocity and the angle which are required to correspond under the dynamic environment by matching with a dynamic Kalman filtering algorithm, the current attitude of the module, the attitude measurement precision is 0.05 degree, and the stability is extremely high. The data output frequency is 100Hz (baud rate 115200) or 20Hz (baud rate 9600), and the requirements on the gyroscope circuit in the invention can be met.

The input voltage of the system power supply circuit (16) is 3.7V, the lithium battery (17) supplies power, and the working voltages of different chips of the system provided by the voltage conversion module are-2.5V, 2.5V and 3.3V respectively.

The wearable limb movement intelligent sensing system based on myoelectricity is characterized in that a remote four-rotor control module (4) is a four-axis unmanned aerial vehicle, and the four-axis unmanned aerial vehicle mainly comprises a motor, an electric controller, a battery, blades, a rack, a remote controller and a flight controller. The built-in attitude calculation of the UAV (unmanned aerial vehicle), namely a controller reads self sensor data, and calculates the attitude angle of the UAV in real time. The controller on the long-range four rotor control module (4) calculates the output quantity of 4 motors according to these information, makes the aircraft keep balanced stable or keep certain angle of inclination to the four shaft unmanned aerial vehicle aircraft of setting for the direction flight can realize the flight of deciding the altitude, fixed point flight, 4D air-over, key take-off and key landing. The attitude solution applied by the remote four-rotor control module (4) in the invention is a complementary filtering algorithm to calculate roll angle (roll), pitch angle (pitch) and yaw angle (yaw) so as to control and maintain the stability and reliability of the aircraft in flight. After the complementary filtering algorithm is applied, the four-axis unmanned aerial vehicle can realize eight kinds of movement in the air, namely vertical ascending, vertical descending, forward movement, backward movement, leftward movement, backward movement, clockwise course changing and anticlockwise course changing.

The wearable limb movement intelligent sensing system based on myoelectricity is characterized in that the physiological state analysis module (3) has an online real-time classification function, and the real-time online classification function can classify 20 hand movements.

(1) The 20 hand movements are: the gesture comprises a thumb erecting gesture, a fist holding gesture, a palm outward swinging gesture, a palm inward swinging gesture, a palm upward swinging gesture, a thumb and forefinger pinching gesture, a thumb and middle finger pinching gesture, a thumb and ring finger pinching gesture, a thumb and little finger pinching gesture, a five-finger tennis ball grasping gesture, a thumb and forefinger water bottle pinching gesture, a cup grasping gesture, a palm cup grasping gesture, a thumb and forefinger pen holding gesture to keep the pen vertical, a palm tennis ball grasping gesture, a brush pen holding gesture, a pen container holding gesture, a book holding gesture and a stick pointing to the front.

Firstly, sEMG electromyographic signals of a testee are collected through wearable electromyography collecting equipment, the collected sEMG electromyography signals are periodically transmitted to a hand electromyography active training interface (2) through a WIFI wireless data transmission module (14), then the actual voltage values of the sEMG electromyography signals are analyzed from A/D conversion results through a conversion algorithm, a deep learning parallel Convolutional Neural Network (CNN) model is trained, after training is completed, the sEMG electromyography signals are transmitted to a physiological state analysis module (3) through an NRF24L01 wireless remote circuit (17), wherein the actual voltage values of the sEMG electromyography signals are analyzed from the A/D conversion results through the conversion algorithm again, and finally the sEMG electromyography signals are transmitted to the physiological state analysis module (3) for online real-time classification.

The wearable limb movement intelligent sensing system based on electromyography is characterized in that the actual voltage value of the sEMG electromyography signal is analyzed from the A/D conversion result through a conversion algorithm, and the system comprises the following steps:

1) determining a reference voltage V of a bioelectrical signal acquisition module (12)_REFAnd the amplification factor G of the programmable gain amplifier_PGA；

2) Converting the original A/D conversion result V of each channel₁₆Converted into decimal A/D conversion result V₁₀；

3) Calculating the actual voltage value V of the sEMG electromyographic signal according to the following formula_IN：

Wherein

The wearable limb movement intelligent sensing system based on myoelectricity is characterized in that the convolutional neural network parallel CNN algorithm based on myoelectricity time domain frequency domain feature fusion comprises the following steps:

1) obtaining raw sEMG electromyographic signals

The method comprises the steps that N is the number of channels of original sEMG electromyographic signals, L is the data length of the original sEMG electromyographic signals of each channel, and represents the g-th numerical value of the original sEMG electromyographic signals collected by the c-th electrode in the original sEMG electromyographic signals;

2) carrying out digital band-pass filtering on the original sEMG electromyographic signals, carrying out 50Hz notch filtering, removing power frequency interference, and obtaining processed sEMG electromyographic signals

Wherein, X_c,gRepresenting the g-th numerical value in the sEMG electromyographic signals corresponding to the c-th electrode after filtering;

3) sEMG (sEMG electromyography) signal based on digital filtering

After filtering, a plurality of sample sets are constructed.

4) And entering a parameter fine adjustment stage, sequentially sending the sEMG electromyographic signals after digital filtering of each wearer into a parallel convolution neural network model with an initial depth, and performing training and gradient correction.

The wearable limb movement intelligent sensing system based on electromyography is characterized in that the step 2) of digitally filtering the original sEMG electromyography signals adopts a band-pass filter, and the first stop band frequency F of the band-pass filter_stop10.001Hz, first passband frequency F_pass110Hz, second pass band frequency F_pass230Hz, second stop band frequency F_stop2The first stopband attenuation rate is 5dB, and the second stopband attenuation percentage is 5dB at 40 Hz.

The wearable limb movement intelligent sensing system based on the electromyography is characterized in that a digital filtered sEMG electromyography signal sample set collected by each wearer is sequentially sent into a convolutional neural network model with an initial depth. Carrying out full supervision training on the deep convolutional neural network model, setting the initial learning rate of the model to be 0.004, attenuating the learning rate in an exponential form to prevent the fixed learning rate from obtaining the optimal model, carrying out 500-period cyclic training totally, setting the Batchsize to be 128, and setting an early stopping mechanism of Earlystopping. The best model for training the model in all cycles of training is obtained. Model parameter fine tuning is not necessary on a pre-training basis.

In the invention, when a deep convolutional neural network model is designed, myoelectric data of a plurality of people are collected for training, so that the deep learning model has the cross-tested generalization capability. And the output of the deep convolutional neural network model is the result which is output by the physiological state analysis module (2) and corresponds to the four-rotor motion. Compared with many traditional machine learning algorithms, the CNN uses a multilayer structure to improve the generalization performance and the abstraction performance of the recognition model, and multiple experiments show that the CNN algorithm applied in the invention is an efficient method for hand motion pattern recognition with respect to multi-channel sEMG signal processing.

An attention mechanism is added to the feature fusion Convolutional Neural Network (CNN). The attention mechanism, i.e., the attention model, is mainly divided into a local attention model and a global attention model. The global attention model is similar to the soft attention model. The key idea is to first detect a point of attention or a position in the input sequence and then select a window around the position to create a local soft attention model. The positions in the input sequence can be set (monotonic alignment) or learned by a prediction function (predictive alignment). Therefore, the advantage of local focus is the parameter tradeoff between soft focus, hard focus, computational efficiency, and intra-window differentiability. Note that force mechanisms can be used to assign important weights to these different representations, which can determine the most relevant aspects while ignoring noise and redundancy in the input. Attention mechanism was applied to globalaveragepoolic 2D, globalaxpoling 2D, multiply, Permute and Lambda.

In the deep convolutional neural network model, forward propagation and backward propagation are mainly utilized. Firstly, analyzing original sEMG electromyographic signal samples through all CNN layers, and forward propagating sEMG electromyographic signal data sets to obtain output values. The error between the output value and the desired value is then calculated to determine the accuracy of the output. Next, the weight values are modified using an error back-propagation process. These two processes are repeatedly performed by the iterative operating system until the loss value of the network is minimized. The weighting values are then modified using a gradient descent algorithm. In order to prevent the final model from being not the optimal model due to excessive modification of the weight, an early stopping mechanism of Earlystopping is arranged in the invention for reducing excessive fitting, and the model with the minimum value of the test loss function is set as the optimal model for training at present.

The deep convolutional neural network model comprises two branches, and the input of each branch is a digital filtered sEMG electromyographic signal; the two branches are different convolutional neural networks CNN, the two convolutional neural networks run in parallel, the two deep convolutional neural networks respectively emphasize and extract one-dimensional data in a time domain and a frequency domain, the dimensionality of time domain and frequency domain features is reduced in a Maxpooling layer, action prediction results are finally output, and the prediction results of the two deep convolutional neural network models are superposed to obtain final prediction results. The convolution layers and the pooling layers of the two deep convolution neural network models are deep enough and stable, and the time-frequency domain characteristics of the sEMG electromyographic signals can be effectively extracted. Therefore, there is a complementary effect on the prediction results.

In both deep convolutional neural network models, the same initialization method is used to initialize the convolution kernel weights, trying to make the output and input obey the same probability distribution as much as possible. L2 regularization is applied to the entire connection layer, both in the two deep convolutional neural network models; a complexity index model is added to the loss function to improve the model's ability to recognize random noise. The model adopts an Adam optimizer, and the loss reduction rule adopts a gradient reduction algorithm.

The first convolutional neural network comprises in sequence:

(1) the data input layer is used for inputting data which are the sEMG electromyographic signal samples after digital filtering;

(2) the first convolution layer has 32 convolution kernels, the size of which is 1 × 4, and the step size is (1, 1). The convolution kernel selects ReLU as the activation function,

(3) the first maximum pooling layer has the pooling core size of 2 multiplied by 1 and the step length of (1,1), and extracts the maximum value of the elements of the input data covered by the current pooling core as output;

(4) in the second convolutional layer, the number of convolutional kernels is 64, the convolutional kernel size is 4 × 1, and the step size is (1, 1). The convolution kernel selects ReLU as the activation function,

(6) in the third convolutional layer, the number of convolution kernels is 64, the convolution kernel size is 1 × 4, and the step size is (1, 1). The convolution kernel selects ReLU as the activation function,

(7) the second maximum pooling layer, the size of the pooling core is 2 multiplied by 1, the step length is (1,1), and the maximum value of the elements of the input data covered by the current pooling core is extracted and used as output;

(8) in the fourth convolution layer, the number of convolution kernels is 128, the convolution kernel size is 4 × 1, and the step size is (1, 1). The convolution kernel selects ReLU as the activation function,

(9) attention mechanism for assigning important weights, ignoring noise and redundancy in the input;

(10) a Flatten layer for one-dimensionalizing the multi-dimensional input;

(11) the first Dropout layer is used for preventing overfitting from improving the generalization capability of the model, selecting the neuron on the upper layer randomly to enable the neuron not to output, and the selection probability is 0.5;

(12) the node number of the first full-connection layer is 128, the ReLU is selected as an activation function, the L2 norm is selected as a regularization term, and the L2 norm is set to be 0.004;

(13) the second Dropout layer prevents overfitting from improving the generalization capability of the model, randomly selects the neuron on the upper layer to ensure that the neuron does not output, and the selection probability is 0.5;

(14) the second full-link layer outputs the classification result, the number of the selected nodes is 4, Softmax is selected as an activation function, and the Softmax function is a normalized exponential function in nature and is defined as

Wherein e is a natural logarithm value, z_hFor the output of the h neuron, in the formulaThe denominator acts as a regularization term, such that

The second convolutional neural network comprises in sequence:

(1) the data input layer is used for inputting data into the digital filtered sEMG electromyographic signals;

(2) the first convolution layer has 16 convolution kernels, the size of the convolution kernels is 1 × 40, and the step size is (1, 1). The convolution kernel selects ReLU as the activation function,

(3) the first maximum pooling layer, the size of the pooling core is 1 multiplied by 8, the step length is (1,8), and the maximum value of the elements of the input data covered by the current pooling core is extracted and used as output;

(4) in the second convolutional layer, the number of convolutional kernels is 32, the size of convolutional kernels is 1 × 3, and the step size is (1, 3). The convolution kernel selects ReLU as the activation function,

(5) the second maximum pooling layer, the size of the pooling kernel is 25 multiplied by 1, the step length is (25,1), and the maximum value of the elements of the input data covered by the current pooling kernel is extracted and used as output;

(6) attention mechanism for assigning important weights, ignoring noise and redundancy in the input;

(7) a Flatten layer for one-dimensionalizing the multi-dimensional input;

(8) the first Dropout layer is used for preventing overfitting from improving the generalization capability of the model, selecting the neuron on the upper layer randomly to enable the neuron not to output, and the selection probability is 0.3;

(9) the node number of the first full-connection layer is 128, the ReLU is selected as an activation function, the L2 norm is selected as a regularization term, and the L2 norm is set to be 0.004;

(10) the second Dropout layer prevents overfitting from improving the generalization capability of the model, randomly selects the neuron on the upper layer to ensure that the neuron does not output, and the selection probability is 0.3;

(11) the second full-link layer outputs the classification result, the number of the selected nodes is 4, Softmax is selected as an activation function, and the Softmax function is a normalized exponential function in nature and is defined as

Wherein e is a natural logarithm value, z_hFor the output of the h-th neuron, the denominator in the equation acts as a regularization term, such that

The outputs of the full-link layers of the two branches are connected in parallel to form a parallel layer, and the parallel layer integrates the characteristics of a deep convolutional neural network. The output after parallel connection is the output of the physiological state analysis module (2), as shown in fig. 4.

The wearable limb movement intelligent sensing system based on electromyography is characterized in that collected sEMG electromyographic signals of 8 channels are taken as a sample 250 times before entering two deep learning algorithms, so that the deep learning algorithms can better extract the time domain and frequency domain characteristics of the sEMG electromyographic signals and send the characteristics into the deep learning algorithms.

The wearable limb movement intelligent sensing system based on the electromyography is characterized in that in the sEMG electromyography signals, the electromyography data of each channel can be represented as waveform data of one measured muscle channel. Therefore, in the invention, when the deep learning model is used for semantic segmentation, the conventional spatial convolution pooling of the regular shape NxN is not adopted, but the long and narrow convolution and the strip pooling are adopted. The square sized convolutional and pooling layers limit the correlation and flexibility of upper and lower data information per channel to capture sEMG electromyographic signals in real-time hand movements.

In order to more effectively capture the correlation and flexibility of the sEMG electromyographic signals to the upper and lower data information of each channel, the receiving range of a first deep learning model is expanded by utilizing long and narrow convolution and stripe pooling, and the correlation characteristics of the upper and lower data of the sEMG electromyographic signals are collected. In the first deep learning model, in the first convolution layer, in order to extract the frequency characteristics of the sEMG electromyogram signal, a long and narrow convolution kernel of 1 × 4 size is used to extract the frequency domain characteristics of the sEMG, in addition, in the second convolution layer, the time domain characteristics of the sEMG electromyogram signal are extracted by long and narrow convolution of 4 × 1 size, and in the third and fourth convolution layers, the frequency domain characteristics and the time domain characteristics of the sEMG electromyogram signal are sequentially and repeatedly extracted.

The wearable limb movement intelligent sensing system based on myoelectricity is characterized in that the used calculation formula of the long and narrow convolution is as follows:

in the first deep learning convolutional neural network CNN, according to the frequency characteristics of the sEMG electromyographic signals, the electromyographic frequencies are mainly distributed at 0-500Hz, and after digital filtering, the frequencies of the sEMG electromyographic signals are all within the frequency range of 0-500Hz, but in each action of the collected sEMG electromyographic signals, the frequency distribution characteristics are more obvious, so in terms of the size of an electromyographic convolution kernel, the frequency domain characteristics of the electromyographic signals are firstly extracted, the extracted frequency domain characteristics can effectively represent the multivariate time sequence of the sEMG electromyographic signals, the input sEMG electromyographic signals are subjected to horizontal and vertical long and narrow convolution kernel convolution to calculate H × 1 and 1 × W, the input sEMG electromyographic signals are converted into H × 1 and 1 × W after long and narrow pooling, the element values in the pooling kernel are maximized, and the values are used as pooling output values.

Attention is a mechanism, and the attention mechanism has the characteristics of few parameters, high speed and good effect. The attention mechanism applies to both the channel attention mechanism and the spatial attention mechanism. The channel attention mechanism is as follows: when dimension compression is carried out on input features, Max pooling is additionally introduced as a supplement, and after two pooling functions, two one-dimensional vectors can be obtained in total. Global average potential has feedback to each pixel point on the characteristic diagram, and Global max potential has gradient feedback only in the position with the maximum response in the characteristic diagram when performing gradient back propagation calculation, and can be used as a supplement of GAP. The spatial attention mechanism is that the input feature graph is compressed by using Average and Max Pooling, the compression is changed into the compression on the channel level, and mean and Max operations are respectively carried out on the input feature graph on the channel dimension. And finally, two-dimensional features are obtained, the two-dimensional features are spliced together according to the channel dimension to obtain a feature map with the channel number being 2, then a hidden layer containing a single convolution kernel is used for carrying out convolution operation on the feature map, and the finally obtained features are ensured to be consistent with the input feature map in the spatial dimension.

Therefore, in the present invention, the algorithm is used as the first deep learning model, and the convolutional neural network CNN algorithm is configured as a block diagram as shown in fig. 5.

The wearable limb movement intelligent sensing system based on electromyography is characterized in that in a second deep learning model, 1 x 40 long and narrow convolution is applied firstly, frequency domain characteristics of sEMG electromyography signals are extracted, then 1 x 8 maximum value banding pooling is carried out, multiple times of characteristic fusion and dimension reduction are carried out on frequency distribution of each action in the system, and after the banding pooling layer is applied, Maxpooling has local invariability, can extract obvious characteristics and simultaneously reduce parameters of the model, so that overfitting of the model is reduced. The Maxpooling only extracts the significant features of the sEMG electromyographic signals and discards insignificant information, the extracted time-frequency domain features can effectively represent the multivariate time sequence of the sEMG electromyographic signals, and the generation of overfitting can be relieved to a certain extent due to the fact that the parameters of the model are reduced. In the present invention, the algorithm is used as a second deep learning model, and a block diagram of the second deep learning convolutional neural network CNN algorithm is shown in fig. 6.

The invention relates to an application of a wearable limb movement intelligent sensing system based on myoelectricity, a remote four-rotor control module is characterized by comprising the following components:

(1) the result output by the physiological state analysis module (2) of claim 7, is sent by the NRF24L01 remote module to control the remote quad-rotor control module (4).

The invention discloses an application of a wearable limb movement intelligent perception system based on myoelectricity, which is characterized by comprising the following steps:

(1) in order to ensure that the electrode paste is completely contacted with the skin; before measuring the electromyographic signal sEMG, wiping the skin of the tested position of the dominant hand of the tested person by alcohol cotton, wherein the tested person is required to keep sitting still, annotating a screen on eyes and completing the action required by a hand electromyographic active training interface; after a system power supply circuit (16) is ensured to be normal, the system is started, sEMG electromyographic signals are collected through wearable electromyographic collection equipment and are transmitted to a physiological state analysis module, and electromyographic collection work is completed;

(2) the physiological state analysis module automatically carries out digital filtering on the collected sEMG electromyographic signals;

(3) the classification and identification system performs classification and identification according to the digitized filtered sEMG electromyographic signals based on a pre-trained deep convolutional neural network model;

(4) the classification result is signaled to a remote quad-rotor control module (4) by an NRF24L01 remote module.

Claims

1. The utility model provides a wearing formula limb movement intelligence perception system based on flesh electricity, including connecting gradually: the system comprises wearable myoelectricity acquisition equipment (1), a hand myoelectricity active training interface (2), a physiological state analysis module (3) and a remote unmanned aerial vehicle module (4); the hand myoelectricity active training system is characterized in that a user collects sEMG (surface EMG) myoelectricity signals from the hand of the user by applying the wearable hand myoelectricity collecting equipment (1) through the hand myoelectricity active training interface (2); the physiological state analysis module (3) receives the sEMG electromyographic signals and intelligently identifies hand motions selected by a user by combining an attention-based convolutional neural network algorithm (CNN); remote unmanned aerial vehicle module (4) regard physiological state analysis module (3) as the basis, through wireless remote equipment, control remote unmanned aerial vehicle equipment (4) and make the motion corresponding with the hand action, the gyroscope module also passes through wireless remote equipment in addition, sends data, combines together with physiological state analysis module (3), controls remote unmanned aerial vehicle equipment (4) jointly.

2. The myoelectricity-based wearable limb movement intelligent sensing system of claim 1, wherein the wearable myoelectricity acquisition device (1) comprises: an electrode patch and a lead wire (11) thereof which are connected in sequence and used for collecting sEMG electromyographic signals, a bioelectrical signal acquisition module (12) used for electromyographic signal amplification and conversion, an STM32 microprocessor (13) and a WIFI wireless data transmission circuit (14) used for controlling the acquisition of the electromyographic signal and transmitting the sEMG electromyographic signal to the physiological state analysis module (2), a gyroscope circuit (15) used for measuring the acceleration, the angular velocity and the angle of the wearable electromyographic signal acquisition equipment (1), a system power supply circuit (16) respectively connected with the bioelectrical signal acquisition module (12) and the STM32 microprocessor (13) and an NRF24L01 wireless remote circuit (17), wherein the electrode patches and the electrode patches in the lead lines (11) thereof are used for collecting sEMG (surface acoustic EMG) electromyographic signals of different muscles of the hand, is connected with the bioelectricity signal acquisition module (12) through a lead wire and a PJ313B interface and is used for acquiring and transmitting bioelectricity signals; the electrode paste is pasted on the hand of a testee, and the muscle collected by the myoelectricity collection module is as follows: extensor carpi ulnaris, extensor digitorum, extensor radialis brevis, extensor radialis longus, brachioradialis, circumflex, flexor radialis, flexor palmaris longus, and flexor radioulnaris, and can obtain the electromyographic signals of 8 channels of the hand;

the bioelectrical signal acquisition module (12) consists of a plurality of bioelectrical signal acquisition chips which are integrated with an analog input AD module with high common mode rejection ratio for receiving muscle voltage acquired by an electrode patch and sEMG (small electric field) electromyographic signals, a low-noise programmable gain amplifier for measuring the sEMG and amplifying the EMG, and a high-resolution synchronous sampling analog-to-digital converter for converting the analog signals into digital signals;

the STM32 microprocessor (13) is used for adjusting the acquisition mode of the bioelectrical signal acquisition module, adjusting and controlling the WIFI wireless data transmission module (14) and the NRF24L01 wireless remote circuit (17) to output sEMG electromyographic signals, and sending the sEMG electromyographic signals to a mobile phone end for the physiological state analysis module (3) to analyze data; the acceleration, the angular velocity and the angle signals measured by the gyroscope circuit (15) are subjected to data analysis and finally sent to the remote quad-rotor unmanned aerial vehicle (4);

the WIFI wireless data transmission module (14) works in an AP mode, the highest transmission rate is 4Mbps, and under the control of an STM32 microprocessor (13), collected sEMG electromyographic signals are periodically output to a hand electromyographic active training interface (2) and a physiological state analysis module (3) at a mobile phone end in the form of data packets through the WIFI wireless data transmission module (14);

the NRF24L01 wireless remote circuit (17) has the highest transmission rate of 2Mbps, under the control of an STM32 microprocessor (13), the collected sEMG electromyographic signals are periodically output to a hand electromyographic active training interface (2) and a physiological state analysis module (3) through the NRF24L01 wireless remote circuit (17) in the form of data packets, and in addition, the NRF24L01 wireless remote circuit (17) can send data measured by the gyroscope circuit (15) to a remote quadrotor unmanned aerial vehicle device after simple analysis of the STM32 microprocessor (13);

the gyroscope circuit (15) is mainly applied to a six-axis attitude angle sensor and is provided with a corresponding circuit; the gyroscope circuit (15) adopts a high-precision gyroscope accelerometer MPU6050, reads the measurement data of the MPU6050 through an STM32 microprocessor (13), then outputs the measurement data through a serial port, integrates an attitude resolver in the module, and can accurately output the acceleration, the angular velocity and the angle which are required to correspond under a dynamic environment by matching with a dynamic Kalman filtering algorithm, the current attitude of the module and the attitude measurement precision are 0.05 degrees, and the stability is extremely high; the data output frequency is 100Hz (baud rate 115200) or 20Hz (baud rate 9600), and the requirements on the gyroscope circuit in the invention can be met;

the input voltage of the system power supply circuit (16) is 3.7V, the system power supply circuit is powered by a lithium battery (17), and the working voltages of different chips of the system, which are provided by the voltage conversion module, are-2.5V, 2.5V and 3.3V respectively;

the wearable limb movement intelligent sensing system based on myoelectricity is characterized in that a remote four-rotor control module (4) is a four-axis aircraft, and the four-axis aircraft mainly comprises a motor, an electric controller, a battery, blades, a rack, a remote controller and a flight controller; the built-in attitude calculation of the four-axis aircraft is realized, namely, a controller reads the data of a sensor of the four-axis aircraft, and the attitude angle of the four-axis aircraft is calculated in real time; the controller on the remote four-rotor control module (4) calculates the output quantities of the 4 motors according to the information, so that the four-axis aircraft can realize fixed-height flight, fixed-point flight, 4D air overturning, one-key takeoff and one-key landing by keeping the aircraft balanced and stable or keeping a certain inclination angle and flying towards a set direction; the attitude solution applied by the remote four-rotor control module (4) in the invention is a complementary filtering algorithm to calculate roll angle (roll), pitch angle (pitch) and yaw angle (yaw) so as to control and maintain the stability and reliability of the aircraft in flight; after the complementary filtering algorithm is applied, the four-axis aircraft can realize eight kinds of movement in the air, namely vertical ascending, vertical descending, forward movement, backward movement, leftward movement, backward movement, clockwise course changing and anticlockwise course changing.

3. The myoelectricity-based wearable limb-movement intelligent sensing system of claim 1, wherein the physiological status analysis module (3) has an online real-time classification function, and the real-time online classification mode comprises 20 hand movements;

the 20 hand movements are: a vertical thumb gesture, a fist holding gesture, a palm outward swinging gesture, a palm inward swinging gesture, a palm upward swinging gesture, a thumb and forefinger pinching gesture, a thumb and middle finger pinching gesture, a thumb and ring finger pinching gesture, a thumb and little finger pinching gesture, a five-finger tennis ball gripping gesture, a thumb and forefinger water bottle pinching gesture, a cup gripping gesture, a palm cup gripping gesture, a thumb and forefinger pen holding gesture to keep the pen vertical, a palm tennis ball gripping gesture, a brush holding gesture, a pen container holding gesture, a book holding gesture to keep the book vertical, a book holding gesture, and a wooden stick pointing to the front;

firstly, sEMG electromyographic signals of a testee are collected through wearable electromyography collecting equipment, the collected sEMG electromyography signals are periodically transmitted to a hand electromyography active training interface (2) through a WIFI wireless data transmission module (14), then actual voltage values of the sEMG electromyography signals are analyzed from A/D conversion results through a conversion algorithm, a attention system parallel Convolution Neural Network (CNN) model is trained, after training is completed, the sEMG electromyography signals are transmitted to a physiological state analyzing module (3) through an NRF24L01 wireless remote circuit (17), wherein the actual voltage values of the sEMG electromyography signals are analyzed from the A/D conversion results through the conversion algorithm again, and finally the sEMG electromyography signals are transmitted to the physiological state analyzing module (3) for online real-time classification.

4. The wearable electromyography-based intelligent sensing system for limb movement of claim 3, wherein the analyzing of the actual voltage value of the sEMG electromyography signal from the A/D conversion result by the conversion algorithm comprises the following steps:

2) Converting the original A/D conversion result V of each channel₁₆Converted into a decimal A/D conversion result V₁₀；

Wherein

5. The wearable electromyography-based limb movement intelligent perception system of claim 3, wherein the parallel CNN algorithm based on the fusion of the electromyography time domain and frequency domain features with the convolutional neural network comprises the following steps:

1) obtaining raw sEMG electromyographic signals

Wherein N is the number of channels of the original sEMG electromyographic signals, and L is the data length of the original sEMG electromyographic signals of each channelDegree, E_c,gThe g-th numerical value in the original sEMG electromyographic signals collected by the c-th electrode in the original sEMG electromyographic signals is represented;

3) sEMG electromyographic signal based on digital filtering

Constructing a plurality of sample sets after filtering;

4) and entering a parameter fine adjustment stage, and sequentially sending the sEMG electromyographic signals after digital filtering of each user into a convolution neural network model with an initial depth for training and gradient correction.

6. The wearable electromyography-based intelligent sensing system for limb movement of claim 5, wherein the step 2) of digitally bandpass filtering the raw sEMG electromyography signals employs a Chebyshev I-type bandpass filter having a first stop band frequency F_stop10.001Hz, first passband frequency F_pass110Hz, second pass band frequency F_pass230Hz, second stop band frequency F_stop2The first stopband attenuation rate is 5dB, and the second stopband attenuation percentage is 5dB at 40 Hz.

7. The wearable electromyography-based intelligent sensing system for limb movement of claim 5, wherein the digitally filtered sEMG electromyography signal sample sets collected by each wearer of step 4) are sequentially fed into a convolutional neural network model of initial depth; carrying out full supervision training on the deep convolutional neural network model, setting the initial learning rate of the model to be 0.004, attenuating the learning rate in an exponential form to prevent the fixed learning rate from obtaining the optimal model, carrying out 500-period cyclic training totally, setting the Batchsize to be 128, and setting an early stopping mechanism of Earlystopping; the best model for training the model in all periods can be obtained; model parameter fine adjustment is not needed on the basis of pre-training;

when the deep convolutional neural network model is designed, myoelectric data of a plurality of people are collected for training, so that the deep learning model has the cross-tested generalization capability; the output of the deep convolutional neural network model is a result which is output by the physiological state analysis module (2) and corresponds to the four-rotor motion; compared with many traditional machine learning algorithms, the CNN uses a multilayer structure to improve the generalization performance and the abstract performance of the recognition model, and multiple experiments show that the CNN algorithm applied in the invention is an efficient method for recognizing hand motion modes with respect to multi-channel sEMG signal processing;

adding an attention mechanism into a feature fusion Convolutional Neural Network (CNN); an attention mechanism, namely an attention model, is mainly divided into a local attention model and a global attention model; the global attention model is similar to the soft attention model; the key idea is to first detect the point of attention or position in the input sequence and then select a window around that position to create a local soft attention model; the positions in the input sequence can be set (monotonic alignment) or learned by a prediction function (predictive alignment); therefore, the local focus is advantageous in the parameter trade-off between soft focus, hard focus, computational efficiency, within-window differentiability; note that force mechanisms can be used to assign important weights to these different representations, which can determine the most relevant aspects, while ignoring noise and redundancy in the input; attention mechanism was applied to globalaveragepoiling 2D, globalmaxploling 2D, multiply, Permute and Lambda;

in the deep convolutional neural network model, forward propagation and backward propagation are mainly utilized; firstly, analyzing original sEMG electromyographic signal samples and forward-propagated sEMG electromyographic signal data sets through all CNN layers to obtain output values; then calculating the error between the output value and the expected value to determine the accuracy of the output; next, the weight values are modified using an error back-propagation process; the two processes are repeatedly executed by the iterative operating system until the loss value of the network is minimum; then modifying the weighted value by using a gradient descent algorithm; in order to prevent the final model from being not the optimal model due to excessive modification of the weight, therefore, in the invention, an early stopping mechanism of Earlystopping is arranged for reducing excessive fitting, and the model with the minimum value of the test loss function is set as the optimal model for the current training;

the deep convolutional neural network model comprises two branches, and the input of each branch is the digital filtered sEMG electromyographic signal; the two branches are different convolutional neural networks CNN, the two convolutional neural networks run in parallel, the two deep convolutional neural networks respectively emphasize and extract one-dimensional data in a time domain and a frequency domain, the dimensionality of characteristics of the time domain and the frequency domain is reduced in a Maxpoling layer, and the weight is balanced by combining an attention mechanism; finally outputting an action prediction result, and overlapping the prediction results of the two deep convolutional neural network models to obtain a final prediction result; the convolution layer and the pooling layer of the two deep convolutional neural network models are deep enough and stable enough, and can effectively extract the time-frequency domain characteristics of the sEMG electromyographic signals; therefore, there is a complementary effect on the prediction results;

in two deep convolution neural network models, the same initialization method is used for initializing convolution kernel weights, and the output and the input are subjected to the same probability distribution as much as possible; l2 regularization is applied to the entire connection layer, both in the two deep convolutional neural network models; the complexity index model is added into the loss function to improve the random noise identification capability of the model; the model adopts an Adam optimizer, and the loss reduction rule adopts a gradient reduction algorithm;

the first convolutional neural network comprises in sequence:

(2) a first convolution layer, the number of convolution kernels is 32, the size is 1 multiplied by 4, and the step length is (1, 1); the convolution kernel selects ReLU as the activation function,

(3) a first maximum pooling layer, wherein the size of the pooling core is 2 multiplied by 1, the step length is (1,1), and the maximum value of the elements of the input data covered by the current pooling core is extracted as output;

(4) a second convolution layer, the number of convolution kernels is 64, the size of the convolution kernels is 4 multiplied by 1, and the step size is (1, 1); the convolution kernel selects ReLU as the activation function,

(6) a third convolution layer, the number of convolution kernels is 64, the size of the convolution kernels is 1 × 4, and the step size is (1, 1); the convolution kernel selects ReLU as the activation function,

(8) a fourth convolution layer, the number of convolution kernels is 128, the size of the convolution kernels is 4 x 1, and the step size is (1, 1); the convolution kernel selects ReLU as the activation function,

(10) a Flatten layer for one-dimensionalizing the multi-dimensional input;

The second convolutional neural network comprises in sequence:

(2) a first convolution layer, the number of convolution kernels is 16, the size is 1 x 40, and the step length is (1, 1); the convolution kernel selects ReLU as the activation function,

(4) a second convolution layer, the number of convolution kernels is 32, the size of the convolution kernels is 1 multiplied by 3, and the step length is (1, 3); the convolution kernel selects ReLU as the activation function,

(7) a Flatten layer for one-dimensionalizing the multi-dimensional input;

(9) selecting 128 nodes in the first full connection layer, selecting ReLU as an activation function, selecting an L2 norm as a regularization term, and setting an L2 norm to be 0.004;

(11) the second full-link layer outputs the classification result, the number of the selected nodes is 4, Softmax is selected as an activation function, and the Softmax function is essentially a normalized exponential function and is defined as

The outputs of the full connection layers of the two branches are connected in parallel to form a parallel connection layer, and the output after parallel connection is the output of the physiological state analysis module (2);

the wearable limb movement intelligent sensing system based on electromyography is characterized in that collected sEMG electromyographic signals of 8 channels are taken as a sample 250 times before entering two deep learning algorithms, so that the deep learning algorithms better extract the time domain and frequency domain characteristics of the sEMG electromyographic signals and send the characteristics into the deep learning algorithms;

the wearable limb movement intelligent sensing system based on the electromyography is characterized in that in the sEMG electromyography signals, the electromyography data of each channel can be represented as waveform data of one measured muscle channel; therefore, when the deep learning model is used for semantic segmentation, the regular spatial convolution pooling of the regular shape NxN is not adopted, but the long and narrow convolution and the strip pooling are adopted; the square convolution layer and the pooling layer limit the relevance and flexibility of upper and lower data information of each channel for capturing sEMG (semG) electromyographic signals in real-time hand movements;

in order to more effectively capture the correlation and flexibility of the sEMG electromyographic signals to the upper and lower data information of each channel, the receiving range of a first deep learning model is expanded by utilizing long and narrow convolution and stripe pooling, and the correlation characteristics of the upper and lower data of the sEMG electromyographic signals are collected; in the first deep learning model, in the first convolution layer, in order to extract the frequency characteristics of the sEMG electromyogram signals, 1 × 4 long and narrow convolution kernels are used to extract the frequency domain characteristics of the sEMG electromyogram signals, in addition, in the second convolution layer, the time domain characteristics of the sEMG electromyogram signals are extracted through 4 × 1 long and narrow convolution, and in the third and fourth convolution layers, the frequency domain characteristics and the time domain characteristics of the sEMG electromyogram signals are sequentially and repeatedly extracted;

in a first attention-based convolutional neural network CNN, according to the frequency characteristics of sEMG (electrical magnetic stimulation) electromyographic signals, electromyographic frequencies are mainly distributed in the range of 0-500Hz, and after digital filtering, the frequencies of the sEMG electromyographic signals are all in the frequency range of 0-500Hz, but the frequency distribution characteristics are more obvious in each action of the collected sEMG electromyographic signals; therefore, in the aspect of the size of the electromyographic convolution kernel, firstly, extracting frequency domain characteristics of the electromyographic signals, wherein the extracted frequency domain characteristics can effectively represent a multivariate time sequence of the sEMG electromyographic signals, the input sEMG electromyographic signals are subjected to horizontal and vertical long and narrow convolution kernel convolution to calculate H multiplied by 1 and 1 multiplied by W, the input sEMG electromyographic signals are changed into H multiplied by 1 and 1 multiplied by W after long and narrow pooling is carried out, the maximum value of element values in the pooling kernel is obtained, and the maximum value is used as a pooling output value;

attention is a mechanism, and the attention mechanism has the characteristics of few parameters, high speed and good effect; the attention mechanism is applied to a channel attention mechanism and a space attention mechanism; the channel attention mechanism is as follows: when dimension compression is carried out on input features, Max pooling is additionally introduced as supplement, and two one-dimensional vectors can be obtained in total after two pooling functions are carried out; the Global average potential has feedback on each pixel point on the characteristic diagram, and the Global max potential has gradient feedback only in the position with the maximum response in the characteristic diagram when performing gradient back propagation calculation and can be used as a supplement of GAP; performing compression operation on the input feature graph by using Average and Max Pooling, wherein the compression operation is changed into compression on a channel level, and mean and Max operations are respectively performed on the input features on channel dimensions; finally, two-dimensional features are obtained, the two-dimensional features are spliced together according to the channel dimension to obtain a feature map with the channel number being 2, then a hidden layer containing a single convolution kernel is used for carrying out convolution operation on the feature map, and the finally obtained features are ensured to be consistent with the input feature map in the spatial dimension;

the wearable limb movement intelligent sensing system based on electromyography is characterized in that in a second deep learning model, 1 x 40 long and narrow convolution is applied firstly, frequency domain characteristics of sEMG electromyography signals are extracted, then 1 x 8 maximum value banding pooling is carried out, multiple times of characteristic fusion and dimension reduction are carried out on frequency distribution of each action in the system, and after the banding pooling layer is applied, Maxpooling has local invariability, can extract obvious characteristics and simultaneously reduce parameters of the model, so that overfitting of the model is reduced; the Maxpooling only extracts the significant features of the sEMG electromyographic signals and discards insignificant information, the extracted time-frequency domain features can effectively represent the multivariate time sequence of the sEMG electromyographic signals, and the generation of overfitting can be relieved to a certain extent due to the fact that the parameters of the model are reduced; in the present invention, this algorithm is used as a second deep learning model, and a block diagram of the CNN algorithm in the second attention mechanism convolutional neural network is shown in fig. 6.

8. Use of a wearable myoelectric-based limb movement intelligent sensing system of claim 1, a remote drone control module, comprising:

(1) the result output by the physiological state analysis module (2) of claim 7, the remote drone module (4) is controlled by sending a signal through the NRF24L01 remote module.

9. Use of the wearable electromyography-based intelligent sensing system of limb movement of claim 1, comprising:

(1) in order to ensure that the electrode paste is completely contacted with the skin; before the electrode paste is pasted, firstly, wiping the skin to be tested of a testee, which is used by a hand, by alcohol cotton, wherein the testee is required to keep sitting still, annotating a screen on eyes, and completing the action required by a hand myoelectricity active training interface; after a system power supply circuit (16) is ensured to be normal, the system is started, sEMG electromyographic signals are collected through wearable electromyographic signal collection equipment and are transmitted to a physiological state analysis module, and electromyographic signal collection is completed;

(4) the remote module sends a signal to control the remote drone module (4) through the NRF24L 01.