3D-based skeleton control system

Abstract

The invention discloses a bone control system based on 3D. Comprising a sensor system, an actuator system and a control unit, the sensor system is used for collecting real-time data of a human exoskeleton, wherein the real-time data comprises a physiological state and a motion posture of a wearer; the actuator system is used for converting the control signal into a physical action so as to control and adjust the exoskeleton of the human body; the control unit is responsible for data processing and instruction output, data information from the sensor system is fused and processed in real time, and a motion instruction is generated through an algorithm; physiological states are collected through the electromyogram sensor and the brain signal sensor, brain signals are divided into the X axis, the Y axis and the Z axis, motion of the exoskeleton on the X axis, the Y axis and the Z axis is analyzed and processed, and finally the motion intention of a wearer is obtained through a classic machine learning method or a deep learning method.

Description

Technical Field

[0001] The present invention belongs to the technical field of exoskeleton control, and more specifically, particularly relates to a 3D-based skeleton control system. Background Art

[0002] Exoskeleton robot technology is a comprehensive technology that integrates sensing, control, information, fusion, and mobile computing to provide a wearable mechanical mechanism for operators. Exoskeleton robots refer to robots that are worn outside the human body, also known as "wearable robots."

[0003] As a close-fitting device, exoskeleton robots are regarded as an extension of the human body, so ensuring the comfort of the wearer is also a necessary condition. Although the research on wearable exoskeleton equipment has been relatively mature, there has not been enough attention paid to civilian comfort.

[0004] Exoskeleton robot technology is a comprehensive technology that integrates sensing, control, information, fusion, and mobile computing to provide a wearable mechanical mechanism for operators. It refers to a robot that is worn outside the human body, also known as a "wearable robot."

[0005] Exoskeleton power-assisted robots can be used in three major areas, namely military, civilian and medical. It can be said that the essence of power-assisted machines is to amplify human strength and movement speed several times or even thousands of times. Combining the current research progress and human survival needs, looking forward to the future development of technology, the power-assisted equipment will not only not hinder human movements, but also amplify human target movements to the required target value according to human brain consciousness, just like the clothes we wear.

[0006] Brain-controlled exoskeleton system, this exoskeleton can realize the interaction between bones, muscles and nervous system. All bones and muscles are directly controlled by the brain. The brain-controlled exoskeleton system was designed by the Neuromechanics Laboratory of the University of Michigan.

[0007] The existing technology has the problem that when the exoskeleton robot is in use, the control effect is inaccurate, that is, the signal transmission is slow, which easily causes delays in data information analysis, making it impossible for the exoskeleton to assist the human body. When the brain signal is in use, it is impossible to perform three-dimensional analysis of the brain and it is impossible to control and adjust according to the results of the three-dimensional analysis. Therefore, we propose a 3D-based skeletal control system. Summary of the invention

[0008] In view of the problems existing in the prior art, the purpose of the present invention is to provide a 3D-based skeletal control system for controlling the exoskeleton through brain signals, effectively improving the signal processing efficiency, and performing three-dimensional analysis of brain signals to facilitate precise control and adjustment, so that the exoskeleton can effectively assist the human body in movement.

[0009] To achieve the above-mentioned object, the present invention provides the following technical solutions: a 3D-based skeleton control system, comprising a sensor system, an actuator system and a control unit; The sensor system is used to collect real-time data of the human exoskeleton, including the wearer's physiological state and movement posture; the physiological state is collected in real time through an electromyography sensor and a brain signal sensor, and the movement posture is collected in real time through an inertial measurement unit, a pressure sensor and a position sensor; The electromyography sensor is used to detect the wearer's muscle activity, which is used to infer the wearer's intention and control the movement of the exoskeleton; The brain signal sensor controls the movement of the exoskeleton by directly acquiring brain signals; The inertial measurement unit is used to detect the wearer's posture, movement and acceleration; The pressure sensor is used to monitor the force feedback of the exoskeleton on the wearer to ensure safety; The position sensors are used to determine the angles and positions of various parts of the exoskeleton; The actuator system converts the control signal into physical action to realize the control and adjustment of the human exoskeleton. The actuators in the actuator system include electric motors, hydraulic drives, pneumatic drives and servo motors. The control unit is responsible for data processing and command output, integrating and processing the data information from the sensor system in real time, and generating motion commands through algorithms; The control unit preprocesses the collected data information to remove noise from the data information and enhance the data information, and then extracts features from the data information, and decomposes the extracted features on the X-axis, Y-axis and Z-axis based on the extracted features, and performs pattern recognition and decoding on the extracted features, and obtains the wearer's movement intention through classical machine learning methods and deep learning methods, and then transmits control instructions to the actuator system, and controls and adjusts the human exoskeleton through the electric motor, the hydraulic drive, the pneumatic drive and the servo motor.

[0010] Specifically, the preprocessing uses a bandpass filter to remove noise from the data information collected from the physiological state, limiting the signal to a specific frequency range and removing low-frequency and high-frequency noise: The calculation formula of the bandpass filter is as follows: , in: is the low frequency cutoff frequency of the filter; is the high frequency cutoff frequency of the filter; is the frequency of the input data signal, is the frequency of the filtered data information; When the signal frequency Falling and When the signal is between , the signal will not be attenuated, otherwise the signal will be completely attenuated; The bandpass filter has a filtering range of 13-100 Hz for brain signals and is used to control beta waves and gamma waves in brain signals to pass through the bandpass filter; The preprocessing further uses a second-order IIR low-pass filter to denoise the data information collected from the motion posture. The calculation formula of the second-order IIR low-pass filter is as follows: , To output data, is the input data, a and b are the coefficients of the filter, n is the number of filtering times, k=0 is the data information processed this time, k=1 is the data information processed next time, and nk is the position of the data information.

[0011] Specifically, the preprocessing also includes independent component analysis, which is used to remove artifact signals of eye movements and electromyography from brain signals and decompose the observed signals into statistically independent components; , in, is the observation signal matrix, is the mixing matrix, It is an independent signal source; And from the observed signal Recover independent source signals from , that is, find an anti-mixing matrix , so that: .

[0012] Specifically, the enhancement in the preprocessing is performed on the data information, and a practical AGC algorithm is used. The gain value calculation formula of the practical AGC algorithm is: , Wherein, A(n+1) is the gain value after adjustment for data signal transmission; A(n) is the gain value before adjustment; is the signal amplitude value; is the expected amplitude value; is a constant; The value is different, and the AGC adjustment speed is also different. According to the actual requirements of the practical AGC algorithm, when the data signal output amplitude is less than the upper limit of the expected value, , the gain adjustment will become slower. , the gain adjustment speed will be faster at this time.

[0013] Specifically, the feature extraction adopts multi-scale feature extraction, which includes all-round and multi-scale feature extraction of time domain features, frequency domain features, and cepstrum domain features, and then performs dimensionality reduction on the extracted features to facilitate screening out effective features; The calculation formula of time domain characteristics is as follows: , Where: The signal amplitude of the nth point of the audio in the i-th frame after framing; L represents the frame length; n represents the point number of each frame; sgn represents the sign function; Among them, the sgn symbol function is defined as: ; Frequency domain features are used to convert time domain signals into frequency domain. They are defined as the signal power within a unit frequency band and intuitively show the functional relationship between signal variance and frequency.

[0014] Specifically, the cepstral domain feature is a Mel cepstral coefficient feature, and a group of H triangular bandpass filter groups are arranged from low frequency to high frequency in the Mel scale frequency domain according to the size of the critical bandwidth from dense to sparse; the number of filter groups and the number of critical bands must be close, generally 22-26; The input data signal is Fourier transformed, then passed through a triangular bandpass filter group, and the logarithmic energy of each filter group output is calculated, and then the MFCC coefficients can be obtained by discrete cosine transform; MFCC feature calculation formula: , Where: S represents the Mel filter energy; subscript i represents the frame number; h represents the Mel filter number; n represents the spectrum line after discrete cosine transform.

[0015] Specifically, the extracted features are decomposed into the X-axis, Y-axis and Z-axis as follows: For spatial decomposition, brain signals at different spatial locations are displayed through scalp potential maps, which are obtained by interpolation or spatial filtering; Potential map calculation: Assume that Brain signal matrix recorded by electrodes , then the scalp potential map is obtained by the following method: , in: is the spatial coordinate The potential diagram value on It is with The weight of the electrodes, It is Brain signals from electrodes The potential map of electrodes at different positions is reconstructed through spatial interpolation technology to obtain the movement trend of the wearer on the X-axis, Y-axis and Z-axis.

[0016] Specifically, the pattern recognition and decoding are mainly processed by computational analysis using deep learning methods; The deep learning method performs analysis and processing through a long short-term memory network; The calculation formula of the long short-term memory network is as follows: Forget Gate: , Input Gate: , Candidate memory cells: , Memory unit update: , Output Gate: , Hide status updates: , in: , , , is the weight matrix of the input gate, forget gate, candidate memory unit and output gate; , , , is the weight matrix from the hidden layer to each gate; , , , is the bias term of each gate; is the state of the memory unit; The LSTM network can effectively capture long-term dependencies in a sequence through its gating mechanism and predict resource requirements within a certain period of time in the future.

[0017] Specifically, the long short-term memory network updates the algorithm through an adaptive mechanism. The adaptive mechanism combines the control of the forget gate and the input gate, so that the state of the memory unit is adaptively updated with the input data and task requirements, and can automatically learn and memorize sequence information at different time steps, and adaptively adjust the state and information flow of its memory unit according to different input and context information; And the calculation formula of the adaptive mechanism is as follows: Memory unit update: ; That is, the data information collected by the inertial measurement unit, the pressure sensor and the position sensor is reversely input into the long short-term memory network, and the data information is updated by updating the memory unit, so as to adaptively adjust the state and information flow of its memory unit.

[0018] Specifically, the control instruction is generated by outputting the operation intention predicted by the long short-term memory network in combination with the model predictive control, so as to facilitate the control and regulation of the electric motor, the hydraulic drive, the pneumatic drive and the servo motor; The calculation process of the model predictive control is as follows: Assume the dynamic equation of the system is: ; in, is the system status, is the control input, i.e., the operation intention predicted by the LSTM network, A and B are the dynamic matrices of the system; Minimize a cost function, which is the weighted sum of errors over a period of time in the future: , in: and is the weighting matrix, which determines the weights of the error and control input, is the reference trajectory, The exoskeleton is controlled and regulated by predicting future outputs and determining current control inputs by solving an optimization problem.

[0019] Technical effects and advantages of the present invention: The sensor system of the present invention includes an electromyogram sensor and a brain signal sensor, which are used to collect the physiological state of the person wearing the exoskeleton, so as to realize more accurate and efficient control of the operation of the exoskeleton, and when performing operation control, the data information collected by the inertial measurement unit, pressure sensor and position sensor in the motion posture is used for real-time reverse adjustment control, that is, the collected data information is reversely input into the long short-term memory network, and the data information is updated by updating the memory unit, so as to facilitate adaptive adjustment of the state and information flow of the memory unit, so as to realize more accurate control and adjustment of the exoskeleton; The control unit removes noise and enhances the collected data information to improve the accuracy of the data information, facilitates subsequent data processing, and extracts features from the data information, especially the features of brain signals, so as to divide the data information into X-axis, Y-axis and Z-axis according to the fluctuation of brain signals, so as to analyze and process the movement of the exoskeleton on the X-axis, Y-axis and Z-axis, and finally obtains the wearer's movement intention through classical machine learning methods or deep learning methods, and then transmits the control instructions to the actuator system to control and adjust the human exoskeleton.

[0020] Further features and advantages of the present invention will become apparent from the following detailed description of exemplary embodiments of the present invention with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the system structure provided by the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, a 3D-based skeleton control system provided by an embodiment of the present invention includes a sensor system, an actuator system and a control unit; The sensor system is used to collect real-time data of the human exoskeleton, including the wearer's physiological state and movement posture; the physiological state is collected in real time through an electromyography sensor and a brain signal sensor, and the movement posture is collected in real time through an inertial measurement unit, a pressure sensor and a position sensor; The electromyography sensor is used to detect the wearer's muscle activity, which is used to infer the wearer's intention and control the movement of the exoskeleton; The brain signal sensor controls the movement of the exoskeleton by directly acquiring brain signals; The inertial measurement unit is used to detect the wearer's posture, movement and acceleration; The pressure sensor is used to monitor the force feedback of the exoskeleton on the wearer to ensure safety; The position sensors are used to determine the angles and positions of various parts of the exoskeleton; The actuator system converts the control signal into physical action to realize the control and adjustment of the human exoskeleton. The actuators in the actuator system include electric motors, hydraulic drives, pneumatic drives and servo motors. The control unit is responsible for data processing and command output, integrating and processing the data information from the sensor system in real time, and generating motion commands through algorithms; The control unit preprocesses the collected data information to remove noise from the data information and enhance the data information, and then extracts features from the data information, and decomposes the extracted features on the X-axis, Y-axis and Z-axis based on the extracted features, and performs pattern recognition and decoding on the extracted features, and obtains the wearer's movement intention through classical machine learning methods and deep learning methods, and then transmits control instructions to the actuator system, and controls and adjusts the human exoskeleton through the electric motor, the hydraulic drive, the pneumatic drive and the servo motor.

[0024] In this embodiment, preferably, the preprocessing uses a bandpass filter to remove noise from the data information collected from the physiological state, limiting the signal to a specific frequency range and removing low-frequency and high-frequency noise: The calculation formula of the bandpass filter is as follows: , in: is the low frequency cutoff frequency of the filter; is the high frequency cutoff frequency of the filter; is the frequency of the input data signal, is the frequency of the filtered data information; When the signal frequency Falling and When the signal is between , the signal will not be attenuated, otherwise the signal will be completely attenuated; The bandpass filter has a filtering range of 13-100 Hz for brain signals and is used to control beta waves and gamma waves in brain signals to pass through the bandpass filter; The preprocessing further uses a second-order IIR low-pass filter to denoise the data information collected from the motion posture. The calculation formula of the second-order IIR low-pass filter is as follows: , To output data, is the input data, a and b are the coefficients of the filter, n is the number of filtering times, k=0 is the data information processed this time, k=1 is the data information processed next time, and nk is the location of the data information; It should be noted that the physiological state data information is filtered through a bandpass filter to improve the accuracy of the data information, and the motion posture data information is processed through a second-order IIR low-pass filter to facilitate subsequent processing of the data information.

[0025] In this embodiment, preferably, the preprocessing further includes independent component analysis, which is used to remove artifact signals of eye movements and electromyography from brain signals and decompose the observed signals into statistically independent components; , in, is the observation signal matrix, is the mixing matrix, It is an independent signal source; And from the observed signal Recover independent source signals from , that is, find an anti-mixing matrix , so that: ; It should be noted that the use of independent component analysis to detect eye movement and electromyography artifact signals in brain signals can effectively improve the accuracy of brain signal data information and facilitate subsequent accurate analysis and processing.

[0026] In this embodiment, preferably, the enhancement processing of the enhancement realization data information in the preprocessing adopts a practical AGC algorithm, and the gain value calculation formula of the practical AGC algorithm is: , Wherein, A(n+1) is the gain value after adjustment for data signal transmission; A(n) is the gain value before adjustment; is the signal amplitude value; is the expected amplitude value; is a constant; The value is different, and the AGC adjustment speed is also different. According to the actual requirements of the practical AGC algorithm, when the data signal output amplitude is less than the upper limit of the expected value, , the gain adjustment will become slower. , the gain adjustment speed will be faster at this time; It should be noted that by using a practical AGC algorithm to enhance brain signal processing, it is possible to prevent the loss of weak brain signals and improve the integrity of brain signals.

[0027] In this embodiment, preferably, the feature extraction adopts multi-scale feature extraction, and the multi-scale feature extraction includes time domain features, frequency domain features, and cepstrum domain features to perform comprehensive and multi-scale feature extraction, conduct comprehensive analysis, and then perform dimensionality reduction on the extracted features to facilitate screening out effective features; The calculation formula of time domain characteristics is as follows: , Where: The signal amplitude of the nth point of the audio in the i-th frame after framing; L represents the frame length; n represents the point number of each frame; sgn represents the sign function; Among them, the sgn symbol function is defined as: ; Frequency domain features are used to convert time domain signals to frequency domain. They are defined as the signal power within a unit frequency band and intuitively show the functional relationship between signal variance and frequency. It should be noted that comprehensive, multi-scale feature extraction through time domain features, frequency domain features, and cepstrum domain features can improve the accuracy of feature extraction from brain signals, making it easier to perform subsequent movement intention analysis based on the extracted features.

[0028] In this embodiment, preferably, the cepstral domain feature is a Mel cepstral coefficient feature, and a group of H triangular bandpass filter groups are arranged from low frequency to high frequency in the Mel scale frequency domain according to the size of the critical bandwidth from dense to sparse; the number of filter groups and the number of critical bands need to be close, generally 22-26; The input data signal is Fourier transformed, then passed through a triangular bandpass filter group, and the logarithmic energy of each filter group output is calculated, and then the MFCC coefficients can be obtained by discrete cosine transform; MFCC feature calculation formula: , Where: S represents the Mel filter energy; subscript i represents the frame number; h represents the Mel filter number; n represents the spectrum line after discrete cosine transform; It should be noted that by performing cepstrum domain feature processing on the features and performing Fourier transform on the features to calculate the energy carried by the features, it is convenient to calculate the motion intention on the X-axis, Y-axis and Z-axis.

[0029] In this embodiment, preferably, the extracted features are decomposed into the X-axis, Y-axis and Z-axis as follows: For spatial decomposition, brain signals at different spatial locations are displayed through scalp potential maps, which are obtained by interpolation or spatial filtering; Potential map calculation: Assume that Brain signal matrix recorded by electrodes , then the scalp potential map is obtained by the following method: , in: is the spatial coordinate The potential diagram value on It is with The weight of the electrodes, It is Brain signals from 10 electrodes; The potential map of electrodes at different positions is reconstructed through spatial interpolation technology to obtain the movement trend of the wearer on the X-axis, Y-axis and Z-axis; It should be noted that the trend of movement on the X-axis, Y-axis and Z-axis can be analyzed and calculated through the potential map of the cerebral cortex, so as to obtain the characteristic information of the X-axis, Y-axis and Z-axis.

[0030] In this embodiment, preferably, the pattern recognition and decoding are mainly processed by deep learning methods for computational analysis; The deep learning method performs analysis and processing through a long short-term memory network; The calculation formula of the long short-term memory network is as follows: Forget Gate: , Input Gate: , Candidate memory cells: , Memory unit update: , Output Gate: , Hide status updates: , in: , , , is the weight matrix of the input gate, forget gate, candidate memory unit and output gate; , , , is the weight matrix from the hidden layer to each gate; , , , is the bias term of each gate; is the state of the memory unit; The LSTM network can effectively capture the long-term dependencies in the sequence through its gating mechanism and predict the resource demand within a certain period of time in the future; It should be noted that the features are analyzed and processed through the long short-term memory network, and the wearer's intention is predicted based on the results of the feature analysis and processing, which facilitates the operation of the exoskeleton and assists the wearer's movement.

[0031] In this embodiment, preferably, the long short-term memory network updates the algorithm through an adaptive mechanism, and the adaptive mechanism combines the control of the forget gate and the input gate, so that the state of the memory unit is adaptively updated with the input data and task requirements, and can automatically learn and memorize sequence information of different time steps, and adaptively adjust the state and information flow of its memory unit according to different input and context information; And the calculation formula of the adaptive mechanism is as follows: Memory unit update: ; That is, the data information collected by the inertial measurement unit, the pressure sensor and the position sensor is reversely input into the long short-term memory network, and the data information is updated by updating the memory unit, so as to adaptively adjust the state and information flow of its memory unit; It should be noted that by introducing an adaptive mechanism, that is, the data information collected by the inertial measurement unit, pressure sensor and position sensor is reversely input into the long short-term memory network, and the data information is updated by updating the memory unit, so as to facilitate the adaptive adjustment of the state and information flow of the memory unit.

[0032] In this embodiment, preferably, the control instruction is generated by outputting the operation intention predicted by the long short-term memory network in combination with model predictive control, so as to facilitate control and regulation of the electric motor, the hydraulic drive, the pneumatic drive and the servo motor; The calculation process of the model predictive control is as follows: Assume the dynamic equation of the system is: ; in, is the system status, is the control input, i.e., the operation intention predicted by the LSTM network, A and B are the dynamic matrices of the system; Minimize a cost function, which is the weighted sum of errors over a period of time in the future: , in: and is the weighting matrix, which determines the weights of the error and control input, is the reference trajectory, The exoskeleton is controlled and regulated by predicting future outputs and determining current control inputs by solving an optimization problem.

[0033] The specific operation process of this application: When in use, the wearer wears the exoskeleton, and the electromyography sensor and brain signal sensor are respectively attached to the wearer's muscle skin and brain to facilitate acquisition of electromyography signals and brain signals; Then, when the wearer has a tendency to move, the electromyography sensor is used to detect the wearer's muscle activity, which is used to infer the wearer's intention and control the movement of the exoskeleton. The brain signal sensor controls the movement of the exoskeleton by directly acquiring brain signals; the control unit pre-processes the collected data information, that is, denoises the data information through a bandpass filter and a second-order IIR low-pass filter, and removes the artifact signals of eye movement and electromyography from the brain signal through independent component analysis, removes the noise in the data information, and enhances the data information through a practical AGC algorithm, and then extracts features from the data information, and uses time domain features, The frequency domain features and the cepstral domain features are used for comprehensive and multi-scale feature extraction, and the extracted features are decomposed into X-axis, Y-axis and Z-axis according to the extracted features, and the extracted features are subjected to pattern recognition and decoding, the wearer's movement intention is acquired through the long short-term memory network, and the data information of the inertial measurement unit, the pressure sensor and the position sensor are analyzed and processed through the adaptive mechanism, and then the control command is transmitted to the actuator system, the actuator system performs calculation processing through model predictive control, and then the human exoskeleton is controlled and adjusted through the electric motor, the hydraulic drive, the pneumatic drive and the servo motor.

[0034] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A 3D-based skeleton control system, characterized in that: It includes sensor system, actuator system and control unit; The sensor system is used to collect real-time data of the human exoskeleton, including the wearer's physiological state and movement posture; the physiological state is collected in real time through an electromyography sensor and a brain signal sensor, and the movement posture is collected in real time through an inertial measurement unit, a pressure sensor and a position sensor; The electromyography sensor is used to detect the wearer's muscle activity, which is used to infer the wearer's intention and control the movement of the exoskeleton; The brain signal sensor controls the movement of the exoskeleton by directly acquiring brain signals; The inertial measurement unit is used to detect the wearer's posture, movement and acceleration; The pressure sensor is used to monitor the force feedback of the exoskeleton on the wearer to ensure safety; The position sensors are used to determine the angles and positions of various parts of the exoskeleton; The actuator system converts the control signal into physical action to realize the control and adjustment of the human exoskeleton. The actuators in the actuator system include electric motors, hydraulic drives, pneumatic drives and servo motors. The control unit is responsible for data processing and command output, integrating and processing the data information from the sensor system in real time, and generating motion commands through algorithms; The control unit preprocesses the collected data information to remove noise from the data information and enhance the data information, and then extracts features from the data information, and decomposes the extracted features on the X-axis, Y-axis and Z-axis based on the extracted features, and performs pattern recognition and decoding on the extracted features, and obtains the wearer's movement intention through classical machine learning methods and deep learning methods, and then transmits control instructions to the actuator system, and controls and adjusts the human exoskeleton through the electric motor, the hydraulic drive, the pneumatic drive and the servo motor.

2. A 3D-based skeleton control system according to claim 1, characterized in that: The preprocessing uses a bandpass filter to remove noise from the data information collected from the physiological state, limiting the signal to a specific frequency range and removing low-frequency and high-frequency noise: The calculation formula of the bandpass filter is as follows: ， in: is the low frequency cutoff frequency of the filter; is the high frequency cutoff frequency of the filter; is the frequency of the input data signal, is the frequency of the filtered data information; When the signal frequency Falling and When the signal is between , the signal will not be attenuated, otherwise the signal will be completely attenuated; The bandpass filter has a filtering range of 13-100 Hz for brain signals and is used to control beta waves and gamma waves in brain signals to pass through the bandpass filter; The preprocessing further uses a second-order IIR low-pass filter to denoise the data information collected from the motion posture. The calculation formula of the second-order IIR low-pass filter is as follows: ， To output data, is the input data, a and b are the coefficients of the filter, n is the number of filtering times, k=0 is the data information processed this time, k=1 is the data information processed next time, and nk is the position of the data information.

3. A 3D-based skeleton control system according to claim 1, characterized in that: The preprocessing also includes independent component analysis, which is used to remove artifact signals of eye movement and electromyography from brain signals and decompose the observed signals into statistically independent components; ， in, is the observation signal matrix, is the mixing matrix, It is an independent signal source; And from the observed signal Recover independent source signals from , that is, find an anti-mixing matrix , so that: 。 4. A 3D-based skeleton control system according to claim 1, characterized in that: The enhancement in the preprocessing is to perform enhancement processing on the data information, using a practical AGC algorithm, and the gain value calculation formula of the practical AGC algorithm is: ， Wherein, A(n+1) is the gain value after adjustment for data signal transmission; A(n) is the gain value before adjustment; is the signal amplitude value; is the expected amplitude value; is a constant; The value is different, and the AGC adjustment speed is also different. According to the actual requirements of the practical AGC algorithm, when the data signal output amplitude is less than the upper limit of the expected value, , the gain adjustment will become slower. , the gain adjustment speed will be faster at this time.

5. A 3D-based skeleton control system according to claim 1, characterized in that: The feature extraction adopts multi-scale feature extraction, which includes time domain features, frequency domain features, and cepstrum domain features for comprehensive and multi-scale feature extraction, comprehensive analysis, and then dimensionality reduction of the extracted features to facilitate screening out effective features; The calculation formula of time domain characteristics is as follows: ， Where: The signal amplitude of the nth point of the audio in the i-th frame after framing; L represents the frame length; n represents the point number of each frame; sgn represents the sign function; Among them, the sgn symbol function is defined as: ； Frequency domain features are used to convert time domain signals into frequency domain. They are defined as the signal power within a unit frequency band and intuitively show the functional relationship between signal variance and frequency.

6. A 3D-based skeleton control system according to claim 5, characterized in that: The cepstral domain feature is a Mel cepstral coefficient feature. In the Mel scale frequency domain, a group of H triangular bandpass filter groups are arranged from low frequency to high frequency according to the size of the critical bandwidth from dense to sparse; the number of filter groups and the number of critical bands must be close, generally 22-26; The input data signal is Fourier transformed, then passed through a triangular bandpass filter group, and the logarithmic energy of each filter group output is calculated, and then the MFCC coefficients can be obtained by discrete cosine transform; MFCC feature calculation formula: ， Where: S represents the Mel filter energy; subscript i represents the frame number; h represents the Mel filter number; n represents the spectrum line after discrete cosine transform.

7. A 3D-based skeleton control system according to claim 1, characterized in that: The extracted features are decomposed into X-axis, Y-axis and Z-axis as follows: For spatial decomposition, brain signals at different spatial locations are displayed through scalp potential maps, which are obtained by interpolation or spatial filtering; Potential map calculation: Assume that Brain signal matrix recorded by electrodes , then the scalp potential map is obtained by the following method: ， in: is the spatial coordinate The potential diagram value on It is with The weight of the electrodes, It is Brain signals from 10 electrodes; The potential map of electrodes at different positions is reconstructed through spatial interpolation technology to obtain the movement trend of the wearer on the X-axis, Y-axis and Z-axis.

8. A 3D-based skeleton control system according to claim 1, characterized in that: The pattern recognition and decoding are mainly processed by computational analysis through deep learning methods; The deep learning method performs analysis and processing through a long short-term memory network; The calculation formula of the long short-term memory network is as follows: Forget Gate: , Input Gate: , Candidate memory cells: , Memory unit update: , Output Gate: , Hide status updates: , in: , , , is the weight matrix of the input gate, forget gate, candidate memory unit and output gate; , , , is the weight matrix from the hidden layer to each gate; , , , is the bias term of each gate; is the state of the memory unit; The LSTM network can effectively capture long-term dependencies in a sequence through its gating mechanism and predict resource requirements within a certain period of time in the future.

9. A 3D-based skeleton control system according to claim 8, characterized in that: The LSTM network updates the algorithm through an adaptive mechanism. The adaptive mechanism combines the control of the forget gate and the input gate so that the state of the memory unit is adaptively updated with the input data and task requirements, and can automatically learn and memorize sequence information at different time steps, and adaptively adjust the state and information flow of its memory unit according to different input and context information; And the calculation formula of the adaptive mechanism is as follows: Memory unit update: ; That is, the data information collected by the inertial measurement unit, the pressure sensor and the position sensor is reversely input into the long short-term memory network, and the data information is updated by updating the memory unit, so as to adaptively adjust the state and information flow of its memory unit.

10. A 3D-based skeleton control system according to claim 9, characterized in that: The control instructions are generated by outputting the operation intention predicted by the long short-term memory network in combination with model predictive control, so as to facilitate control and regulation of the electric motor, the hydraulic drive, the pneumatic drive and the servo motor; The calculation process of the model predictive control is as follows: Assume the dynamic equation of the system is: ； in, is the system status, is the control input, i.e., the operation intention predicted by the LSTM network, A and B are the dynamic matrices of the system; Minimize a cost function, which is the weighted sum of errors over a period of time in the future: ， in: and is the weighting matrix, which determines the weights of the error and control input, is the reference trajectory, The exoskeleton is controlled and regulated by predicting future outputs and determining current control inputs by solving an optimization problem.

Images