CN111360792A

CN111360792A - Wearable portable real-time control action recognition system for bionic manipulator

Info

Publication number: CN111360792A
Application number: CN202010238380.8A
Authority: CN
Inventors: 曹天傲; 刘丹; 王启松; 孙金玮
Original assignee: Harbin Institute of Technology
Current assignee: Harbin Institute of Technology
Priority date: 2020-03-30
Filing date: 2020-03-30
Publication date: 2020-07-03

Abstract

Wearable portable real-time control action recognition system towards bionical manipulator relates to a bionical manipulator real-time control technique, in order to solve current manipulator with high costs, function singleness, real-time poor and gather with the precision problem on the low side of discerning. The invention determines the number and the placement positions of the muscle areas and the electrodes to be identified and collected by analyzing the generation mechanism and the characteristics of the surface electromyographic signals. The method comprises the steps of collecting 4-channel forearm surface electromyographic signals by using a laboratory independent research and development collecting device, carrying out interference removal processing, carrying out data windowing, extracting characteristic values, carrying out gesture classification, and finally realizing control over the bionic manipulator. The intelligent collecting and identifying system has the advantages of easiness and convenience in use, low cost, rich functions, good real-time performance and high collecting and identifying precision.

Description

Wearable portable real-time control action recognition system for bionic manipulator

Technical Field

The invention relates to a real-time control technology of a bionic manipulator.

Background

The dexterity and the coordination of the hands play an important role in the daily life and interpersonal interaction of people; however, a part of the limbs of the human body are disabled to different degrees due to natural disasters, accidents, congenital or acquired diseases, and the like; the data of the national health committee show that the proportion of the number of the physical disabilities in China currently accounts for about 30 percent of the total number of the physical disabilities, the number of the physical disabilities continuously rises, huge burden is brought to the society and the family, and meanwhile, the physical disabilities also cause damage to the mind of the people to different degrees; the artificial limb is an important means for enabling the disabled with limbs to normally carry out daily life, and for the hand amputated patient, the mechanical arm can be controlled according to the intention of the hand amputated patient, so that the requirements of basic gripping and interpersonal communication gestures in daily life are met, and the life quality of the patient can be improved to a great extent.

At present, common commercial manipulators are mainly divided into a decorative manipulator, a cable-controlled manipulator and a myoelectric manipulator; the decoration manipulator only helps a patient to restore the natural appearance and the body balance from the appearance, and the action function of a human hand cannot be realized; the cable-controlled manipulator is mainly controlled by pulling a cable by a patient through a residual limb, and does not conform to a normal human nerve control path; the myoelectric manipulator is controlled by an electric signal generated when muscles of a human body contract independently, has various controllable modes, can realize basic gestures of a human hand, and is closer to the human hand. The surface electromyogram signal sEMG is an electric signal generated by corresponding muscle contraction when limbs move, and is an important method for non-invasively detecting muscle activity on the body surface; compared with the muscle electrical signals acquired in an implanted way, the surface electromyographic signals are widely applied to the field of rehabilitation medicine because of no surgical trauma and close connection with muscle movement; some patients lose arms but do not damage the nervous system, the intention of the brain is accurately identified by using information contained in surface electromyographic signals through phantom limb imagination, and patients with congenital disabilities can also generate corresponding electromyographic signals through training so as to control the movement of the artificial limb. At present, many researches are carried out on multifunctional humanoid myoelectric manipulators by some companies and research institutions, so that the manipulators can be close to human hands in function. Many commercial manipulators employ myoelectric signal controlled prostheses to enable the patient to control the manipulator with his or her own intent. But it is expensive, complex to deploy, and currently unacceptable to most patient populations.

Although the application of the surface electromyogram signal and the research of the manipulator have made a certain progress at present, a plurality of problems to be solved still exist, such as low and unstable gesture overall recognition rate, limited recognizable actions, no complete matched acquisition equipment on a software training platform, and the like, so that the research of the bionic manipulator control based on the surface electromyogram signal is of great significance.

Disclosure of Invention

The invention aims to solve the problems of high cost, single function, poor real-time performance and low acquisition and identification precision of the conventional manipulator, and provides a wearable portable real-time control action identification system for a bionic manipulator.

The wearable portable real-time control action recognition system facing the bionic manipulator is used for collecting myoelectric signals of the surface of the forearm of an experimenter, extracting effective characteristics of the myoelectric signals of the surface of the forearm of the experimenter and finally realizing control of the bionic manipulator;

the bionic manipulator comprises a steering engine and a finger linkage device; the steering engine is used for inputting action signals and controlling the finger linkage device to act according to the action signals;

the identification system comprises a USR-C322 chip, a processing module, a collecting electrode and a power supply module;

the USR-C322 chip comprises a single chip microcomputer, a WIFI module, an analog-to-digital converter and a PGA amplifier, and a JTAG port is arranged at the edge of the USR-C322 chip;

the power supply module is used for supplying power to the singlechip;

the collecting electrode is arranged on the forearm of the experimenter and is used for collecting the electromyographic signals of the surface of the forearm of the experimenter;

the JTAG port is used for connecting the acquisition electrode and the PGA amplifier;

the PGA amplifier is used for amplifying the electromyographic signals collected by the collecting electrode;

the analog-to-digital converter is used for converting the amplified myoelectric signals, and the converted amplified myoelectric signals are wirelessly transmitted to the processing module through the WIFI module under the control of the single chip microcomputer;

the processing module comprises a data windowing module, a preprocessing module, a feature extraction module, a classifier and an action label module;

the data windowing module is used for windowing the converted amplified electromyographic signals transmitted by the WIFI module;

the preprocessing module is used for preprocessing the windowed data and removing noise and other interference;

the feature extraction module is used for extracting features of the preprocessed data;

the classifier is used for carrying out action classification on the preprocessed data and distinguishing 5 actions to be recognized;

the action tag module is used for outputting the action classification result to the steering engine in the form of an action tag, and the steering engine controls the finger linkage device to act according to the action tag.

The working principle of the invention is as follows: attaching the collecting electrode to the surface of the forearm of an experimenter, and connecting the collecting electrode to a JTAG port through a collecting line so as to carry out non-invasive collection on the surface myoelectric signal; the electromyographic signals are input in a differential mode, are simply amplified by a PGA amplifier of the USR-C322 chip, are subjected to high-precision analog-to-digital conversion by an analog-to-digital converter, are input to the single chip microcomputer and are wirelessly transmitted to the processing module by the WIFI module; in the processing module, a data windowing module performs data segment windowing operation, a preprocessing module performs signal preprocessing to remove noise and other interference, a feature extraction module extracts required features, and a classifier performs action classification; and finally, the obtained classification result is in the form of an action label, and is converted into a corresponding control command to directly control a steering engine in the bionic manipulator, and the steering engine drives the finger linkage device to enable the bionic manipulator to make corresponding action.

The bionic manipulator real-time waveform receiving and storing system has the advantages that real-time waveforms of uploaded data and drawn data can be accurately and effectively received through the processing module, operations such as feature extraction, classification training, feature access, training network and the like can be performed on the data, functions such as data activity segment detection, network training, serial port communication and the like are integrated, and real-time control over the bionic manipulator is achieved; the system is easy and convenient to use, low in cost, rich in functions, good in real-time performance and high in acquisition and identification precision.

Drawings

Fig. 1 is a schematic block diagram of a wearable portable real-time control action recognition system facing a bionic manipulator according to a first embodiment;

FIG. 2 is a block diagram of an experimental verification-based action recognition system according to a first embodiment;

FIG. 3 is a flowchart of a complete real-time control of the bionic manipulator according to the experimental verification in the first embodiment;

FIG. 4 is a graph showing a statistical polygonal line of the mean accuracy of the experimental verification according to the first embodiment over time.

Detailed Description

The first embodiment is as follows: the embodiment is described with reference to fig. 1 to 4, and the wearable portable real-time control action recognition system facing the bionic manipulator in the embodiment is used for acquiring myoelectric signals of the forearm surface of an experimenter, extracting effective characteristics of the myoelectric signals of the forearm surface of the experimenter, and finally realizing control of the bionic manipulator 5;

the bionic manipulator 5 comprises a steering engine 5-1 and a finger linkage device 5-2; the steering engine 5-1 is used for inputting action signals and controlling the finger linkage device 5-2 to act according to the action signals; the finger linkage device 5-2 is a 6-degree-of-freedom manipulator, and the 6 degrees of freedom are five finger movements and wrist rotation;

the identification system comprises a USR-C322 chip 1, a processing module 2, a collecting electrode 3 and a power supply module 6;

the USR-C322 chip 1 comprises a single chip microcomputer 1-1, a WIFI module 1-2, an analog-to-digital converter 1-3 and a PGA amplifier 1-4, and JTAG ports 1-5 are arranged at the edge of the USR-C322 chip 1; JTAG ports 1-5 are standard 20pin-JTAG ports, thus improving compatibility and universality; the model of the singlechip 1-1 is CC 3200; the model of the analog-to-digital converter 1-3 is ADS1299, the ADS1299 analog-to-digital converter 1-3 can meet the requirement of multi-channel electromyographic signal acquisition, the ADS1299 analog-to-digital converter 1-3 is an AD chip specially used for acquiring bioelectric signals, has the advantages of 24-bit resolution, high precision and low noise, can perform multi-channel synchronous acquisition and has strong functions, the analog-to-digital converter is in a delta-integral form, a crystal oscillator and a reference voltage are both designed in a built-in mode, and a programmable gain amplifier is built in the analog-to-digital converter, so that the acquisition precision, the reliability and the stability are all ensured; the identification system enables all functional parts to work in sequence under the control of the USR-C322 chip 1, and the WIFI module 1-2 is used for communicating with the processing module 2; the singlechip 1-1 is a singlechip based on a Cortex-M4 inner core, and the highest operating frequency is 80 MHz; a WIFI module 1-2 in the USR-C322 chip 1 is an industrial-grade WIFI module, and is a low-power-consumption wireless communication chip specially used for realizing embedded system design;

the power supply module 6 is used for supplying power to the singlechip 1-1; the acquisition electrode 3 is arranged on the forearm of the experimenter and used for acquiring myoelectric signals of the surface of the forearm of the experimenter so as to control the bionic manipulator 5; JTAG port 1-5 is used for connecting collecting electrode 3 and PGA amplifier 1-4; the PGA amplifier 1-4 is used for amplifying the electromyographic signals collected by the collecting electrode 3; the analog-to-digital converter 1-3 is used for converting the amplified myoelectric signal, and the converted amplified myoelectric signal is wirelessly transmitted to the processing module 2 through the WIFI module 1-2 under the control of the singlechip 1-1; the acquisition signal output end of the acquisition electrode 3 is connected with the acquisition signal input ends of PGA amplifiers 1-4 through JTAG ports 1-5, the amplification signal output end of the PGA amplifiers 1-4 is connected with the amplification signal input end of an analog-to-digital converter 1-3, the analog-to-digital signal output end of the analog-to-digital converter 1-3 is connected with the analog-to-digital signal input end of a single chip microcomputer 1-1, and the data signal output end of the single chip microcomputer 1-1 is connected with the data signal input end of a WIFI module 1-2; the collecting electrode 3 is used for collecting electromyographic signals on the surface of the bionic manipulator 5, and the collecting electrode 3 is connected with JTAG ports 1-5 in a four-channel differential input mode; the 20pin-JTAG ports 1-5 are used as acquisition signal input ends, the acquisition input is 8 channels, in order to ensure the stability and the flexibility of signals, a differential input mode is adopted, 16 ports in total are used for accessing an acquisition electrode wire, a power supply and a grounding port are removed, and the remaining two ports are respectively a reference end and a right leg driving input end;

the processing module 2 comprises a data windowing module 2-1, a preprocessing module 2-2, a feature extraction module 2-3, a classifier 2-4 and an action label module 2-5; the classifiers 2 to 4 are BP neural network five classifiers; the processing module 2 is realized by an upper computer, the matlab is used in the upper computer to perform signal preprocessing such as band elimination, band-pass filtering, median filtering and the like, required features are extracted, action classification is performed in a classifier, and finally obtained classification results are used for controlling the bionic manipulator 5; the data windowing module 2-1 is used for windowing the converted amplified electromyographic signals transmitted by the WIFI module 1-2; the preprocessing module 2-2 is used for preprocessing the windowed data and removing noise and other interference; the feature extraction module 2-3 is used for extracting features of the preprocessed data; the classifier 2-4 is used for classifying the actions of the preprocessed data and distinguishing 5 actions to be recognized; the action tag module 2-5 is used for outputting the action classification result to the steering engine 5-1 in the form of an action tag, and the steering engine 5-1 controls the finger linkage device 5-2 to make an action according to the action tag; the wireless signal output end of the WIFI module 1-2 is connected with the wireless signal input end of the data windowing module 2-1; the window division signal output end of the data window division module 2-1 is connected with the window division signal input end of the preprocessing module 2-2, the preprocessing signal output end of the preprocessing module 2-2 is connected with the preprocessing signal input end of the feature extraction module 2-3, the feature signal output end of the feature extraction module 2-3 is connected with the feature signal input end of the classifier 2-4, the classification signal of the classifier 2-4 is connected with the classification signal input end of the action tag module 2-5, and the action signal output end of the action tag module 2-5 is connected with the action signal input end of the steering engine 5-1.

In the present embodiment, the identification system further includes an SD card 4;

the SD card 4 is used for storing the amplified electromyographic signals converted by the analog-to-digital converters 1-3 into the SD card; the acquisition data signal input end of the SD card 4 is connected with the acquisition data signal output ends of the analog-to-digital converters 1-3; a Micro-SD card slot is also arranged on the board where the USR-C322 chip 1 is arranged, so that collected data can be stored in the SD card 4, and later or off-line analysis is facilitated; the analog-to-digital converters 1 to 3 and the SD card 4 communicate through an SPI bus to transmit collected data.

In this embodiment, the processing module 2 further includes a training network access module 2-6;

the training network access module 2-6 is used for storing 5 actions to be recognized which are output by the classifier 2-4; the storage signal input end of the training network access module 2-6 is connected with the storage signal output end of the classifier 2-4; the training network access module 2-6 is used for storing 5 actions to be recognized which are output by the classifier 2-4, so that the classification of new data in the future is facilitated.

In the present embodiment, the power module 6 includes a battery 6-1 and a power management model 6-2;

the power supply output end of the battery 6-1 is connected with the power supply input end of the power supply management model 6-2, and the port output end of the power supply management model 6-2 is connected with the port input end of the singlechip 1-1;

the battery 6-1 is a Li-ion battery with 430mAh/3.7V, and has the advantages of low internal resistance, high capacity, rapid charge and discharge, low self-discharge, high stability and the like;

the power management model 6-2 is divided into a charging mode, a normal mode and a debugging mode, a micro-USB port is arranged on a board where the USR-C322 chip 1 is located, and the micro-USB port is used as a charging and program debugging port of a circuit board and is used for charging a Li-ion battery and downloading a program into the USR-C322 chip 1; meanwhile, in order to collect clean and high-quality electromyographic signals, a positive power supply and negative power supply symmetrical design mode is adopted, and the power supply design not only improves the stability of the analog-digital converters 1-3 and the accuracy of reference voltage, but also simplifies the difficulty of electrostatic shielding protection design; the power management model 6-2 is a battery monitoring chip and is of the type LTC2942, and the power management model 6-2 is communicated with the single chip microcomputer 1-1 in an I2C bus mode.

In the present embodiment, the indexes of the electromyographic signals collected by the collecting electrode 3 are shown in table 1, and the indexes to be analyzed include: the electromyographic signals are acquired with precision, system stability and cruising ability; in the aspect of endurance, the battery can work for 13 hours under the condition of maximum power, can be as long as 24 hours under the normal mode, and the selection of the battery 6-1 can be changed at any time.

TABLE 1 technical indexes of electromyographic signal acquisition device

Under the condition of maximum wireless power, setting the sampling rate to be 1kHz, setting the amplification factor of a PGA amplifier to be 1-4 to be 6, and short-circuiting the electrodes of the analog positive and negative input ends, wherein the signal collected by the equipment is an interference signal, the average peak-to-peak value of the interference noise is about 0.68 muV, even under the condition of strong interference, the maximum peak-to-peak value is only about 4.29 muV, and the interference noise can be almost ignored after the myoelectric signal is amplified by 6 times; the system noise of the video-free voltage input is 5 mu V, the amplitude of the electromyographic signal is 5mV, and the SNR (signal to noise ratio) can be calculated to be 60 dB.

In order to ensure that the data transmission process is stable and reliable and reduce the power consumption of the acquisition equipment, the data communication format needs to be set, and the length and the format of the electromyographic data acquired by the device are shown in table 2.

TABLE 2 electromyographic signal data format table

In view of the problem of the user packet transparent transmission protocol, the packet format of the whole acquisition system is shown in table 3.

TABLE 3 data packet Format

The data packet needs to have 2Byte length start bits for unpacking processing; the packet number unsigned type of 2Byte, so the maximum value is 65536, if the sampling rate is 1kHz, the cycle will be once in about half an hour; the state of the 3byte is used for judging the state of the equipment, and the connection, communication, disconnection and battery residual capacity are judged through the state bit; the length of the test data is selected to be less than 220 bytes to ensure that the packet protocol is still valid when the sampling setting is changed, and to allow for 10 bytes space for future extended functions, and finally the transmission is terminated by a 2Byte end bit. It can be seen that the length of the data packet is ultimately controlled by the upper computer software.

The stability of communication means the time how long electromyographic signal collection process can keep, after 5 experiments with collection equipment work in the collection state, finds that experimental facilities all normally worked at every turn and the battery power is zero, and the operating position is 5m apart from the host computer, and operating duration is about 13 hours, can prove that collection equipment has good stability.

The wearable collecting electrode 3 is of a patch type, and for the purposes of portability and easy controllability and simplifying the scale of a hardware circuit as much as possible, the signal collecting equipment is a rectangular box with the size of 6.4cm x 3.7cm x 1.7cm, and a signal collecting plate and a 3.7VLi-ion battery are arranged in the signal collecting equipment; the circuit is highly integrated, and the physiological signal can be conveniently acquired.

After the USR-C322 chip 1 is connected with an acquisition platform of an upper computer, acquired physiological signals can be displayed in software of the upper computer in real time by waveforms; after the signals can be effectively collected, the gesture actions are collected according to the length of a preset sample, the collected sample is subjected to feature extraction, a feature vector set is formed and is input into a network for classification training; the action classification results of the classifier are set to be 1,2, 3, 4 and 5, which respectively correspond to a fist, a hand, a wrist bending and a wrist stretching, and the number six, and the static action independently corresponds to 0, and the static action is not calculated in the classification action because the static action is used for judging whether the action is generated or not by calculating a threshold value.

Selecting a window length L as 11 to perform median filtering, and removing baseline drift of an original signal; and (3) removing power frequency interference by adopting a Butterworth wave trap: stop band frequency of 49-51Hz, pass band attenuation r _p1 dB; stopband attenuation r_s50 dB; and a Chebyshev filter is adopted to remove high-frequency interference, the passband frequency is 10-300Hz, and the filter order is 256.

Since the signal lasts for a long time in the acquisition process, if the long time is applied to the real-time control stage, great delay is necessarily generated on the action change, so that the signal is divided into a series of small windows, and each window is processed independently. In addition, when the manipulator is controlled by gesture recognition through the sEMG signals, if the recognition system receives continuous sEMG signals for a long time, the processing speed is reduced, and an experimenter can feel obvious action delay. In the design, an overlapping sliding window method is adopted to process signals. The average absolute value of a window is calculated and is recorded as the energy value Q. If the number of times that the value is larger than a certain set threshold value A exceeds N times, the action is considered to be in the initial stage, and if Q is smaller than A in the later acquisition process, the action is considered to be terminated. The window size is set to 128ms, the sliding step is 50ms, and the threshold a is set to 1.5 times the Q value. After windowing, the program calculates a threshold value a, which is stored and used to determine the start and end of the action in the subsequent real-time control.

After signal preprocessing, feature extraction is required. The surface electromyographic signal is a non-stationary random one-dimensional bioelectric signal, but can be processed as a stationary random signal in a short time. The method selects three characteristic values of Mean Absolute Value (MAV), Root Mean Square (RMS) and Waveform Length (WL) as a characteristic extraction method of collected surface electromyographic signals to carry out classification training of gesture actions; each channel extracts these three features.

Assuming that { x (i) | i ═ 1, 2., N } is a single-channel surface electromyographic signal acquired by the acquisition device, the feature extraction formula of the one-dimensional signal is as follows. In the formula, N is the number of collected surface electromyographic signal data points, x_iIs the ith data in the signal data sequence.

(1) Mean Absolute Value (MAV)

(2) Root Mean Square (RMS)

(3) Wave Length (WL)

The invention adopts BP neural network to carry out pattern recognition on the action characteristics, and selects three-layer network to carry out classification training, namely, the hidden layer is determined as 1 layer. When the number of the hidden layer nodes is 9, the training speed is fastest, and the recognition rate is highest. The complete network is configured as a 12-9-5 structure. The input node corresponds to 12-dimensional features of 4 channels, and the output node corresponds to 5 actions to be identified.

And (3) experimental verification:

in order to verify that the identification system can realize the control of the bionic manipulator;

the software in the upper computer selects a matlab platform, and the matlab platform is utilized to integrate the design of algorithms such as equipment communication, signal acquisition and processing, real-time control and the like, so that the integrated operation is facilitated. The acquisition aspect comprises the functions of establishing connection between the upper computer and the equipment through a UDP protocol, changing preset parameters, displaying real-time waveforms, storing and the like; the algorithm design comprises a model training part and a real-time control part; the device sends the acquired data to acquisition software of an upper computer through networking, the software performs operations such as median filtering, notch filtering, band-pass filtering, feature extraction, network training and the like on the acquired data, and stores a trained network. The training network is ready, the collected data can be directly input into the network and compared with the trained data to generate classification labels of gesture actions, the labels are converted into corresponding manipulator action code sequences, the manipulator action code sequences are sent to the bionic manipulator 5 through a serial port, and then the bionic manipulator 5 can make corresponding actions, and the overall design is as shown in fig. 2.

And after the sampling rate is selected and the parameter setting of the SD card 4 is started or not, establishing a collection platform on the upper computer to be connected with the USR-C322 chip 1. The acquisition electrode and an upper computer are ensured to be connected to the same local area network; the algorithm judges whether motion occurs according to a threshold value of a static state, so that a static motion signal needs to be acquired independently. And after the arm keeps still, formally collecting gesture motion signals. Then, performing an overlapping sliding window analysis method on the action signals, wherein when the action signals are greater than a static state threshold value, the number of the action windows is increased by 1, and if the number of the action windows is greater than or equal to the number of the sliding windows, the acquisition meets the requirement of the capacity of the training data sample, and meanwhile, the acquisition is prompted to be completed; if the number of the action windows is less than that of the sliding windows, unqualified acquisition signals are prompted, and acquisition needs to be carried out again. The number of acquisition channels is set to be 4, and the number of acquisition actions is set to be 5. The software also designs the functions of storing and reading characteristic values, reading off-line data and the like. This is to ensure that pattern recognition can be continued using the stored feature data as long as the position of the electrode pad placed on the arm is not changed after the software is turned off. After electromyographic signals are collected, the program extracts the characteristics of each action window, extracts the signal characteristics of corresponding channels, creates a matrix of a training set and a test set, puts the extracted characteristics into the training set and the test set, and randomly scrambles the sequence. The part is also added with a storing and reading function, the weight and the threshold matrix of the trained BP neural network can be stored, and the training network can be used for repeatedly controlling peripheral equipment such as a manipulator and the like on the premise that the position of the patch electrode is not changed.

When the bionic manipulator 5 is controlled, the baud rate of the serial port is set to 115200; the bionic manipulator 5 controls a linkage device 5-2 of fingers by respective independent steering engines 5-1, the fingers are driven by the linkage device to change an opening and closing angle, and the opening and closing angle is determined by the rotation angle of the steering engines 5-1; in order to control the bionic manipulator 5 to perform gesture actions the same as those of a human hand, firstly, a steering engine angle instruction corresponding to the action of the bionic manipulator 5 needs to be stored, then, a label of each gesture action is set for the output of a BP (back propagation) neural network, when the current gesture is recognized by the network, a label of the corresponding action is output, then, a program sends an action instruction statement of the bionic manipulator 5 corresponding to the label to the bionic manipulator 5 through a serial port, and the corresponding action can be controlled; in the actual training, a plurality of times of training can be carried out to achieve better results.

In the experiment, the electromyographic signals of the four muscle positions in the forearm of a human body are collected through the collecting electrode 3, the button type electrode is adopted, and the collecting electrode wire and the collecting device which are designed by a user are matched to obtain electromyographic signal data, so that four-channel signal input is formed. The method comprises the steps of setting a 1000Hz sampling rate to collect four-channel electromyographic signals, converting analog signals into digital signals through a high-precision ADS1299 on a collecting device, transmitting the digital signals to an upper computer, and filtering interference noise by performing median filtering, notch filtering and band-pass filtering on the digital electromyographic signals in an upper computer program. Then, carrying out sliding windowing on the signal data to obtain the number of windows, calculating the mean absolute value MAV of each window, judging the window as an action if the sum of the mean absolute values of one window of the four-channel electromyographic signals is larger than the mean absolute value of one window in a static state, namely acquiring a threshold value calculated after the static action, and extracting the characteristics in the window; if the value is less than the threshold value, the program directly judges that the movement is kept static, and a static signal is output to the manipulator. When action is generated, after characteristics of action window data are extracted, a 12-dimensional characteristic vector matrix is formed by the four channels and the three characteristics together, and the matrix is sent into a trained BP neural network for online classification. And then processing the action tag output by the BP neural network, and if the output action tag is 1, sending an action code stored in the tag 1 to the manipulator through a serial port: making a fist; if the output action tag is 2, the action code stored in the tag 2 is sent to the manipulator through the serial port: stretching the hand; if the output action tag is 3, the action code stored in the tag 3 is sent to the manipulator through the serial port: flexing the wrist; if the output action tag is 4, the action code stored in the tag 4 is sent to the manipulator through the serial port: extending the wrist; if the output action tag is 5, the action code stored in the tag 5 is transmitted to the manipulator through the serial port: six gestures; if the robot does not do any work and keeps the static motion, the robot directly sends a static motion instruction to the manipulator through the serial port. The complete real-time control robot flow is shown in fig. 3.

Selecting 4 ginseng in total and performing real-time control experiments, wherein the age interval is 21-27, three men and one woman are convenient to operate an upper computer for experiments, the collected arms are all performed by selecting the left arm, and the experiment participants are all healthy without any diseases in the aspect of muscle. Before collection, firstly ensuring that collected muscle parts of participants are kept relaxed for a period of time to ensure that the muscles are not tired, then cleaning hair on the arm skin surfaces of the participants, and then cleaning the hair by using cleaning solution or alcohol wet tissue to remove redundant cuticles and reduce impedance; then coating conductive adhesive on the patch electrode to enhance conductivity; and finally, placing the collecting electrode at a required position, and according to the corresponding relation between the positions of the relevant muscles and the actions, identifying the actions of the fingers needs to collect the electromyographic signals of the extensor hallucis longus, the extensor digitorum longus and the longimanus longus. Because the invention adopts the four-channel differential input design, 8 patch electrodes are needed for muscle signal positions, and meanwhile, the elbow which is not influenced by other muscles and has less muscle distribution is selected as a reference electrode, and 9 acquisition electrodes are used for experiments in total. The experimenter needs to collect a 12s static motion signal firstly, and then collects 5 motions according to the sequence of fist making, hand opening, wrist bending, wrist stretching and six numbers. In the collection process, the forearm of the testee is flatly placed on a table, the action force is moderate, and the consistency among the same action modes is kept as much as possible. The experimenter needs to keep sitting posture correct, and makes actions in sequence according to the prompt, wherein each action needs to be collected for 6 times, each time is 12s, and one action needs 72 s. If there is no pause in the middle, the total time for the entire procedure is 6 minutes and 12 seconds. Of course, if the influence of muscle fatigue on signal quality and strength is considered, the acquisition can be stopped every time, and the start of the acquisition can be decided by experimenters. During the experiment, the change condition of the surface myoelectric signal is monitored in real time through matlab software. After electromyographic data is acquired, the electromyographic data is subjected to operations such as preprocessing, feature extraction and the like by a program. The number of sliding windows of the length of the collected sample is 1039, 50 windows are selected as test samples, the rest are training samples for classification, the single accuracy and the average accuracy of the 5 motion pattern recognition of 4 testers are tabulated and counted, the experiment is carried out for three days, namely, the collection experiment of a complete flow is carried out once a day, and the result is shown in table 4.

TABLE 43 day Pattern recognition accuracy for experimenter #1

As can be seen from table 4, although the accuracy of the partial motion is higher or lower with the increase of the training days, the overall accuracy is improved, and the recognition rate of the fist making motion is always 100%, so that the motion is most easily distinguished; the average accuracy data of four experimenters are plotted by a broken line statistical chart by taking the experimental days as an abscissa, and the ordinate is the average accuracy, and the result is shown in figure 4. As can be seen from the figure, the average accuracy of the motion classification also shows a rising trend after the training days of 4 experimenters are increased, so that the accuracy can be improved by a plurality of experiments, the number of experimental days and the like. The average accuracy over the number of experimental days for 4 experimenters is shown in table 5.

TABLE 5 average accuracy of experimenters over experimental days

Therefore, the requirements of the running time are considered by the sliding windowing method and the real-time control calculation, the electromyographic signals are received by the whole control process from the upper computer, the average absolute value is calculated to judge whether the electromyographic signals are larger than the threshold value, the time for the serial port of the feature extraction online classification output action label to send the control command is less than 200ms and less than 300ms of the delay time accepted by people, and therefore the requirements of the delay time are completely met. Therefore, the effect is better in the experiment, the mechanical arm can quickly act and is almost synchronous with the hand of an experimenter, and the acquisition control system designed by the invention is effective and reasonable.

Claims

1. The wearable portable real-time control action recognition system is used for collecting myoelectric signals of the forearm surface of an experimenter, extracting effective characteristics of the myoelectric signals of the forearm surface of the experimenter and finally realizing control over the bionic manipulator (5);

the bionic manipulator (5) comprises a steering engine (5-1) and a finger linkage device (5-2); the steering engine (5-1) is used for inputting action signals and controlling the finger linkage device (5-2) to act according to the action signals;

the identification system is characterized by comprising a USR-C322 chip (1), a processing module (2), a collecting electrode (3) and a power supply module (6);

the USR-C322 chip (1) comprises a single chip microcomputer (1-1), a WIFI module (1-2), an analog-to-digital converter (1-3) and a PGA amplifier (1-4), and JTAG ports (1-5) are arranged at the edge of the USR-C322 chip (1);

the power supply module (6) is used for supplying power to the singlechip (1-1);

the collecting electrode (3) is arranged on the forearm of the experimenter and is used for collecting the electromyographic signals of the surface of the forearm of the experimenter;

the JTAG port (1-5) is used for connecting the acquisition electrode (3) and the PGA amplifier (1-4);

the PGA amplifier (1-4) is used for amplifying the electromyographic signals collected by the collecting electrode (3);

the analog-to-digital converter (1-3) is used for converting the amplified myoelectric signals, and the converted amplified myoelectric signals are controlled by the singlechip (1-1) and are wirelessly transmitted to the processing module (2) through the WIFI module (1-2);

the processing module (2) comprises a data windowing module (2-1), a preprocessing module (2-2), a feature extraction module (2-3), a classifier (2-4) and an action label module (2-5);

the data windowing module (2-1) is used for windowing the converted amplified electromyographic signals transmitted by the WIFI module (1-2);

the preprocessing module (2-2) is used for preprocessing the windowed data and removing noise and other interference;

the feature extraction module (2-3) is used for extracting features of the preprocessed data;

the classifier (2-4) is used for classifying the actions of the preprocessed data and distinguishing 5 actions to be recognized;

the action tag module (2-5) is used for outputting the action classification result to the steering engine (5-1) in the form of an action tag, and the steering engine (5-1) controls the finger linkage device (5-2) to act according to the action tag.

2. The wearable portable real-time control motion recognition system for bionic manipulators as claimed in claim 1, characterized in that the recognition system further comprises an SD card (4);

the SD card (4) is used for storing the amplified electromyographic signals converted by the analog-to-digital converters (1-3) into the SD card.

3. The wearable portable real-time control motion recognition system for bionic manipulators as claimed in claim 1, characterized in that the processing module (2) further comprises a training network access module (2-6);

the training network access module (2-6) is used for storing 5 actions which are output by the classifier (2-4) and are used for distinguishing to-be-recognized actions.

4. The wearable portable real-time control motion recognition system facing a bionic manipulator according to claim 1, wherein the power supply module (6) comprises a battery (6-1) and a power supply management model (6-2);

the power supply output end of the battery (6-1) is connected with the power supply input end of the power supply management model (6-2), and the port output end of the power supply management model (6-2) is connected with the port input end of the singlechip (1-1);

the battery (6-1) is a Li-ion battery with 430 mAh/3.7V;

the power management model (6-2) is divided into a charging mode, a normal mode and a debugging mode.