CN113426007B - A closed-loop epidural electrical stimulation system for upper limb function recovery - Google Patents

Abstract

The present invention discloses a closed-loop epidural electrical stimulation system for upper limb function recovery. A motion capture device acquires multi-angle human posture images in real time and sends them to a control device; an electroencephalogram acquisition device acquires electroencephalogram signals in real time, and the electroencephalogram signals are amplified and filtered and sent to a control device; a control device receives human posture images and electroencephalogram signals to obtain upper limb posture data and electroencephalogram signal characteristics, generates electrical stimulation instructions and sends them to a spinal cord electrical stimulation device; a spinal cord electrical stimulation device receives instructions and generates electrical stimulation pulses to the human spinal cord. The system of the present invention can actively modulate the neural activities that control the upper limb functions of the human body based on the human body's own will, more effectively assist rehabilitation training, and promote the functional recovery of neural circuits related to the human upper limbs.

Description

A closed-loop epidural electrical stimulation system for upper limb function recovery

Technical Field

The present invention relates to a spinal cord electrical stimulation system in the medical field, and in particular to a closed-loop epidural electrical stimulation system for recovering upper limb function after spinal cord injury or stroke.

Background technique

Damage to the central nervous system caused by stroke or spinal cord injury can cause the human body to be unable to normally generate central nervous system commands to control movement or be unable to transmit these commands to skeletal muscles, causing upper limb motor dysfunction. Currently, a commonly used solution is functional electrical stimulation, which fixes one or more sets of electrodes on the surface of the skin of the upper limbs of the human body to electrically stimulate the relevant muscles to help the human body move the arms and palms, but the disadvantage is that it is easy to cause muscle fatigue and cannot work for a long time. Epidural electrical stimulation can apply a stimulating current to the dorsal dura mater of the spinal cord, stimulate the central pattern generating network located on the human spinal cord, enhance the excitability of the spinal cord circuit, and thus mobilize skeletal muscles to contract. It can effectively reduce muscle fatigue and promote the remodeling of damaged neural circuits. However, its temporal and spatial resolution is limited in actual applications, and it cannot achieve precise control of the upper limbs like functional electrical stimulation.

Summary of the invention

The purpose of the present invention is to develop a closed-loop epidural electrical stimulation technology to address the problems that existing upper limb functional electrical stimulation easily causes muscle fatigue and the upper limb epidural electrical stimulation technology has insufficient temporal and spatial resolution. The technology changes the electrical stimulation strategy in real time based on the real-time EEG signals and movement state of the human body, induces the activity of the corresponding neural circuits, and compensates for the movement of the target skeletal muscles.

To achieve the above object, the technical solution adopted by the present invention is:

The closed-loop epidural electrical stimulation system includes a motion capture device, an electroencephalogram acquisition device, a control device, and a spinal cord electrical stimulation device, wherein:

The motion capture device is used to acquire multi-angle human posture images in real time and send the human posture images to the control device;

The EEG acquisition device is used to acquire the EEG signals of the human body in real time, and the EEG signals are sent to the control device after being amplified and filtered;

The control device receives the human posture image from the motion capture device and the EEG signal from the EEG acquisition device, and then processes and analyzes the upper limb posture data in the human posture image and the EEG signal features in the EEG signal in real time, generates an instruction containing electrical stimulation parameters and sends it to the spinal cord electrical stimulation device;

The spinal cord electrical stimulation device receives instructions containing electrical stimulation parameters from the control device and sends electrical stimulation pulses to the human spinal cord.

The motion capture device is mainly composed of at least two cameras, which collect images of human upper limbs from multiple viewing angles as human posture images.

The EEG acquisition device includes at least multiple acquisition channels, a reference channel and an EEG electrode array. The EEG electrode array includes multiple electrodes arranged on the human brain. Each acquisition channel is connected to its own electrode, and the reference channel is connected to an electrode and grounded. The electrodes of the acquisition channel are arranged on different brain regions to acquire EEG changes in the brain regions.

The EEG acquisition device includes a preamplifier, a differential amplifier and an analog filter for processing the original electrical signals collected by each channel. The original electrical signals collected by each channel are sequentially processed by preamplification, differential amplification and filtering and then output to the control device.

The control device receives the real-time posture data and original EEG signals of the human body and performs preprocessing to generate processed upper limb posture data and EEG signal features, then obtains the difference between the current movement state and the target state of the human body based on the upper limb posture data and EEG signal features, generates electrical stimulation instructions, and sends them to the spinal cord electrical stimulation device.

The control device stores a human posture estimation program based on deep learning, which can mark the key points of the upper limbs of the human body according to the multi-view human posture images transmitted by the motion capture device through the deep learning model, and three-dimensionally reconstruct the spatial coordinates of each key point to generate kinematic data of the upper limbs of the human body as the upper limb posture data;

The control device stores an EEG signal processing program, which performs band-pass filtering and spatial filtering on the original EEG signal transmitted by the EEG acquisition device to obtain EEG signal characteristics;

The control device stores a motion state classification program based on a neural network structure, and the motion state classification program determines the current motion state of the human body according to kinematic data and EEG signal characteristics, and then obtains the difference between the motion state and the target state;

The control device stores an electrical stimulation control program, which has a predefined anatomical mapping built in according to the kinematic functions of the muscles of the upper limbs and the distribution of the motor neurons that control the muscles of the upper limbs between the second cervical segment and the first thoracic segment of the spinal cord. The electrical stimulation control program first determines the electrical stimulation site according to the predefined anatomical mapping, and then obtains specific electrical stimulation parameters according to the difference between the current motion state and the target state, including parameters such as the frequency, amplitude, width and number of electrical stimulation pulses. Finally, the electrical stimulation site and electrical stimulation parameters are jointly encoded into an electrical stimulation instruction and sent to the spinal cord electrical stimulation device.

The spinal cord electrical stimulation device receives electrical stimulation instructions from the control device, and applies an electrical stimulation pulse sequence with corresponding parameters to the electrical stimulation site on the dura mater of the spinal cord in the electrical stimulation instruction. The motor neurons located at the electrical stimulation site are activated, which in turn control the muscles to contract, causing the upper limbs to make corresponding movements. The movements are then collected by the motion capture device and fed back to the control device, and the closed-loop control is continuously repeated to perform electrical stimulation.

The beneficial effects of the present invention are:

The system of the present invention can actively modulate the neural activities that control the functions of the upper limbs of the human body based on the human body's own will, more effectively assist rehabilitation training, and promote the functional recovery of the neural circuits related to the upper limbs of the human body.

The present invention can not only avoid rapid muscle fatigue caused by electrical stimulation, but also combine the real-time EEG signals and movement state of the human body to apply electrical stimulation to the corresponding area of the spinal cord to mobilize the corresponding skeletal muscles, thereby improving the spatiotemporal resolution of the electrical stimulation strategy.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention is further described below in conjunction with the accompanying drawings and embodiments;

FIG1 is a schematic structural diagram of a closed-loop epidural electrical stimulation system according to an embodiment of the present invention.

FIG2 is a flow chart of a control method of a closed-loop epidural electrical stimulation system according to an embodiment of the present invention.

FIG. 3 is a data processing flow chart of a posture estimation system according to an embodiment of the present invention.

FIG. 4 is a schematic diagram of a key point structure of a human limb according to an embodiment of the present invention.

FIG5 is a schematic diagram of a neural network structure for feature extraction and fusion and motion state classification according to an embodiment of the present invention.

FIG6 is a schematic diagram of EEG signal characteristics and posture characteristics when a person performs a grasping action according to an embodiment of the present invention.

FIG. 7 is an electrical stimulation parameter corresponding to FIG. 6 of an electrical stimulation control program when a person performs a grasping action according to an embodiment of the present invention.

FIG. 8 is a predefined anatomical map according to an embodiment of the present invention.

FIG. 9 is a schematic diagram of closed-loop control logic according to an embodiment of the present invention.

In the figure: a pair of cameras 1, EEG electrode array 2, computer 3, upper limb posture data 4, EEG signal features 5, neural network 6, spinal cord stimulation device 7; face center 8, neck center 9, shoulder joint 10, elbow joint 11, wrist joint 12, palm joint collection 13.

Detailed ways

The following description is made with reference to the accompanying drawings to introduce an implementation example of the present invention, and the present invention is further described in detail, but the scope of the present invention is not limited.

The contents in the drawings are not drawn according to the actual scale, and some details are omitted or enlarged for the purpose of structural clarity.

As shown in Figure 1, the specific implementation system includes:

The closed-loop epidural electrical stimulation system includes a motion capture device, an electroencephalogram acquisition device, a control device, and a spinal cord electrical stimulation device, wherein:

A motion capture device, used to obtain human posture images from multiple angles in real time and send the human posture images to a control device;

The EEG acquisition device is used to obtain the human body's EEG signals (EEG, electroencephalo-graph) in real time. The EEG signals are sent to the control device after pre-amplification and filtering;

A control device receives a human posture image from a motion capture device and an EEG signal from an EEG acquisition device, and then processes and analyzes upper limb posture data 4 in the human posture image and EEG signal features 5 in the EEG signal in real time, calculates electrical stimulation parameters, generates instructions containing the electrical stimulation parameters, and sends the instructions to a spinal cord electrical stimulation device;

The spinal cord electrical stimulation device 7 receives instructions containing electrical stimulation parameters from the control device, and sends electrical stimulation pulses to the dura mater of the human spinal cord to achieve upper limb function recovery.

The upper limb function recovery of the present invention is upper limb function recovery after spinal cord injury or cerebral stroke.

The motion capture device is mainly composed of at least two cameras 1. The cameras 1 collect images of human upper limbs from multiple perspectives as human posture images, and restore the spatial coordinates of various joints of the human upper limbs based on the human posture images.

The EEG acquisition device includes at least a plurality of acquisition channels, a reference channel and an EEG electrode array 2. The EEG electrode array 2 includes a plurality of electrodes arranged on the human head and brain. The number of electrodes is the same as the total number of acquisition channels and reference channels. Each acquisition channel is connected to an electrode of its own. The reference channel is connected to an electrode and grounded. The electrodes of the acquisition channels are arranged on different brain regions to acquire EEG changes in the brain regions. In a specific implementation, eight acquisition channels are set, which can be composed of one or more analgesic electrodes currently commonly used in clinical practice.

The EEG acquisition device includes a preamplifier, a differential amplifier and an analog filter for processing the original electrical signals collected by each channel. The original electrical signals collected by each channel are sequentially processed by preamplification, differential amplification and filtering and then output to the control device.

The EEG acquisition device acquires original head electrophysiological signals as EEG signals through the electrode array of the EEG electrode array 2.

The control device can adopt a computer 3, which receives the real-time posture data and original EEG signals of the human body and performs pre-processing to generate processed upper limb posture data 4 and EEG signal characteristics 5 with kinematic data, and then obtains the difference between the current motion state and the target state of the human body based on the upper limb posture data 4 and the EEG signal characteristics 5, generates an electrical stimulation instruction, and sends it to the spinal cord electrical stimulation device 7.

Upper limb posture data refers to the three-dimensional spatial coordinates of each joint of the human upper limb, and kinematic data refers to the physical quantities such as angle, angular velocity and angular acceleration of each joint of the human upper limb calculated based on the three-dimensional spatial coordinates. Motion state refers to the predefined state of the human upper limb performing various actions, such as arm extension, hand clenching, etc. A series of continuous motion states constitute a motion state sequence. Patients undergoing rehabilitation training are in a motion state in the sequence at every moment, which is defined as the current motion state, and the next motion state in the sequence is defined as the target motion state.

The control device stores a human posture estimation program based on deep learning, which can mark the key points of the upper limbs of the human body according to the multi-view human posture images transmitted by the motion capture device through the deep learning model, and three-dimensionally reconstruct the spatial coordinates of each key point to generate kinematic data of the upper limbs of the human body as the upper limb posture data 4;

The control device stores an EEG signal processing program, which performs bandpass filtering and spatial filtering on the original EEG signal transmitted by the EEG acquisition device to obtain EEG signal feature 5; specifically, the original head electrophysiological signal is collected for digital bandpass filtering, the electromyography signal is removed, and the spatial filter is used for preprocessing.

The control device stores a motion state classification program constructed based on a neural network 6. The motion state classification program determines the current motion state of the human body according to kinematic data and EEG signal characteristics 5, and then obtains the difference between the motion state and the target state; the target state is a pre-set human motion state.

The control device stores an electrical stimulation control program, which has a built-in anatomical mapping predefined according to the kinematic functions of the muscles of the upper limbs and the distribution of the motor neurons that control the muscles of the upper limbs between the second cervical segment and the first thoracic segment of the spinal cord. The electrical stimulation control program first determines the electrical stimulation site based on the predefined anatomical mapping. The electrical stimulation site is on the dura mater of the spinal cord. Then, based on the difference between the current motion state and the target state, specific electrical stimulation parameters are obtained, including parameters such as the frequency, amplitude, width and number of electrical stimulation pulses. Finally, the electrical stimulation site and the electrical stimulation parameters are encoded together as an electrical stimulation instruction and sent to the spinal cord electrical stimulation device.

Specifically, in the electrical stimulation control program, the difference information between the current motion state and the target state is simultaneously input into the PID controller and the dynamic inverse dynamics model, and the output results of the PID controller and the dynamic inverse dynamics model are added and then limited to obtain the setting parameters of the electrical stimulation pulse, which are then input into the spinal cord electrical stimulation device 7. The electrical stimulation control program continuously repeats the above process, updates the current motion state and the target state, until the training action is completed by assisting the human body.

The spinal cord electrical stimulation device 7 receives the electrical stimulation instruction from the control device, and applies an electrical stimulation pulse sequence with corresponding parameters to the electrical stimulation site on the dura mater of the spinal cord in the electrical stimulation instruction. The motor neurons located at the electrical stimulation site are activated, which in turn control the muscles to contract, causing the upper limbs to make corresponding movements. The movements are then collected by the motion capture device and fed back to the control device. The new posture data and the original EEG signal are monitored, and the closed-loop control is continuously repeated to perform electrical stimulation.

For example:

A) When the current movement state is forward flexion and the target state is arm extension, an electrical stimulation command is generated, specifically on the outside of C5 on the spinal cord, with a frequency of 20 Hz, an amplitude of 0.5 mA, and a width of 0.2 ms.

B) When the current movement state is arm extension and the target state is palm opening, an electrical stimulation command is generated, specifically at the center and outside of C5 on the spinal cord, with a frequency of 20 Hz, an amplitude of 0.5 mA, and a width of 0.2 ms.

C) When the pre-exercise state is an open palm and the target state is a clenched palm, an electrical stimulation command is generated, specifically to the outside of C5 and C6 on the spinal cord, with a frequency of 30 Hz, an amplitude of 0.3 mA, and a width of 0.2 ms.

D) When the current movement state is clenched hand and the target state is retracted arm, an electrical stimulation command is generated, specifically on the outside of C5 and C8 on the spinal cord, with a frequency of 20 Hz, an amplitude of 0.3 mA, and a width of 0.2 ms.

E) When the current movement state is arm adduction and the target state is arm relaxation, an electrical stimulation command is generated, specifically at the center and lateral sides of C2 and the lateral sides of C8 on the spinal cord, with a frequency of 20 Hz, an amplitude of 0.5 mA, and a width of 0.2 ms.

In a specific implementation, for example, the human body can perform tasks such as grasping an object during human rehabilitation training, but the present invention is not limited thereto.

In a specific implementation, the upper limbs of the human body are divided into a facial center 8, a neck center 9, a shoulder joint 10, an elbow joint 11, a wrist joint 12 and a palm joint set 13.

In a specific implementation, multiple sites are set in the spinal cord, and multiple arrayed sites are set on the back of the human body between the spine joints C2 to T1. The column direction of the site array is parallel to the spine direction, and the row direction of the site array is perpendicular to the spine direction. In a specific implementation, 18 sites are set in three columns and six rows, and the sites in the middle column are located on the back where the center line of the spine is located.

In a specific implementation, the closed-loop electrical stimulation system is composed of a camera 1, an EEG acquisition device, a computer 3 and a spinal cord electrical stimulation device 7. The camera 1 communicates with the computer 3 through the network, the EEG acquisition device communicates with the computer 3 through Bluetooth, the computer 3 has a human posture estimation program, an EEG signal processing program, an electrical stimulation control program and a human-computer interaction interface, and the current parameters of each channel of the spinal cord electrical stimulation device 7 are controlled by the control module of the spinal cord electrical stimulation device 7, and the control module of the spinal cord electrical stimulation device 7 communicates with the computer 3 through Bluetooth. There are a variety of wireless communication methods to choose from, such as Zig-Bee, Bluetooth or wifi. The lithium battery of the spinal cord stimulation device 7 is charged by a close-range wireless charger.

Camera 1 is composed of 1-3 cameras, which transmit real-time human motion images to the computer through the local network. Two ordinary 720P 30FPS office cameras can be selected to communicate with the computer through the local area network.

The human posture estimation program installed in the computer 3 can load the pre-trained deep learning model, analyze the human posture image in real time, determine the parameters of electrical stimulation, including the area of electrical stimulation, the intensity and waveform of the current, and send them to the spinal cord electrical stimulation device 7 through the Bluetooth communication device. The human posture estimation program adopts a convolutional neural network model, uses a public multi-view human posture data set for training and saves the trained model.

The human-computer interaction interface on the computer 3 provides a visual monitoring window for the current posture data and motion state of the human body, and has a control interface for adjusting the electrical stimulation strategy.

The spinal cord stimulation device 7 is surgically implanted below the cervical vertebral lamina and above the spinal dura mater of the human body and is powered by a built-in lithium battery.

As shown in FIG2 , when the system is working, the patient autonomously controls the arm to make small movements through the remaining sensory motor functions. Accordingly, the EEG signal related to the movement intention is sent to the computer 3. At the same time, the patient's body shape is captured by the camera 1 and the image is sent to the computer 3. The body posture estimation program on the computer 3 analyzes the current posture information of the human body, and calculates the coordinates of the elbow joint and each finger joint of the upper limb of the human body, and the angular velocity and angular acceleration are transmitted to the motion state classification program.

The motion state classification program determines the current motion state and its target state according to the EEG signal and the joint kinematics data, including forward flexion and raising, arm extension, open palm, clenched palm, arm adduction and arm relaxation. The electrical stimulation control program formulates an electrical stimulation strategy according to the current motion state and the target state, including the stimulation area, the waveform, frequency and amplitude of the current. The communication program encodes the electrical stimulation strategy into instructions and sends them to the spinal cord stimulation device 7 to perform the corresponding electrical stimulation.

As shown in Figure 3, in the preparation stage of rehabilitation training, the camera should be calibrated first to determine the camera's internal and external parameters, including the camera's relative position, focal length, etc. At the same time, a public multi-view human posture dataset is used to train a posture estimation program based on a convolutional neural network, and its effectiveness is verified when the human body performs grasping training movements.

During the rehabilitation training stage, the camera transmits the patient's body image data stream to the computer in real time through the network. The posture estimation program marks the key points of the upper body and uses multi-view images for three-dimensional reconstruction based on the internal and external parameters of the camera. The three-dimensional coordinates of the upper limb joints are calculated and transmitted to the electrical stimulation control program.

As shown in FIG4 , a common human body key point set is mainly composed of basic anatomical joint positions. This system mainly uses the following key points: face center 8 , neck center 9 , shoulder joint 10 , elbow joint 11 , wrist joint 12 , and palm joint set 13 .

Optionally, the palm joint set can be resized according to the level of detail of the motion state. For example, when the motion state of the palm is set to focus only on the opening and closing of the palm, the palm joint set can only include the joints at the center of the palm and the ends of the fingers; when the motion state of the palm is set to focus on the closed shape of the palm, the palm joint set should include all the joints of the human fingers.

As shown in Figure 5, it is a schematic diagram of the network structure of the motion state classification program based on deep learning. EEG signals and kinematic data are used as inputs of the neural network, and after preliminary feature extraction and fusion, they are passed to the classifier based on the fully connected network, and finally the current motion state and its target state are output. Based on the end-to-end learning framework, the collected EEG signals and kinematic data are preprocessed and respectively input into two one-dimensional convolutional neural networks for feature extraction. The feature vectors extracted from the two signals are then merged into a compatible vector space through a single-layer neural network, and transmitted through a three-layer fully connected network. Finally, a motion classifier is obtained through training data, which can determine and output the motion state and the corresponding target state according to the corresponding input data.

As shown in Figure 6, the subject is performing a grasping action, and the EEG signal features are represented by spatial hot spots with specific positions. When the subject is in the S1 state, the EEG signal features indicate that the subject has a movement intention, and the body posture features are not obvious at this time; when the subject is in the S2 state, the subject has a movement intention to extend the forearm and the corresponding posture features; when the subject is in the S3 state, the subject has a movement intention to open the palm and the corresponding posture features; when the subject is in the S4 state, the subject has a movement intention to grasp and the corresponding posture features; when the subject is in the S5 state, the subject has a movement intention to retract the forearm and the corresponding posture features; when the subject is in the S6 state, there is no movement intention and posture features about grasping. Make multiple subjects repeat such grasping actions, collect corresponding EEG signals and human posture data, and train the neural network shown in Figure 5, which can be used to judge the current movement state and target state of the human body according to the EEG signal features and posture features of the human body in rehabilitation training as the basis for formulating electrical stimulation strategies.

As shown in Figure 7, when the subjects were performing the six states of the grasping action shown in Figure 6, the electrical stimulation control program formulated electrical stimulation strategies S1, S2, S3, S4, S5, and S6 corresponding to the six states. The electrical stimulation sites were represented by shaded rectangles with specific positions in the electrode array in the figure, and the waveform, pulse width, amplitude, and frequency of the electrical stimulation were represented by the discounted and frequency values at the top of the figure.

In the figure, C2, C3, C5, C6, C7, C8, and T1 represent the corresponding segments of the spinal cord. The motor neurons that control the muscles of the forelimbs are concentrated in different segments. When the corresponding segments are electrically stimulated, the motor neurons in this area are activated to control the movement of the corresponding skeletal muscles, driving the forearms and palms to make corresponding movements.

As shown in FIG8 , an anatomical map is predefined according to the functions of the muscles of the upper limbs of a person and the distribution of the motor neurons that control these muscles between the second cervical segment and the first thoracic segment of the spinal cord, and the anatomical map is built into the electrical stimulation control program.

The electrical stimulation strategy is determined by a set of electrical stimulation parameters, including site, waveform, pulse width, frequency, and amplitude. The waveform and pulse width are determined by known clinical studies and pre-experiments, the site is determined by the anatomical mapping shown in FIG8 , and the frequency, amplitude, and time are calculated and updated by the closed-loop control loop shown in FIG9 . The feature fusion module in the figure is implemented by the motion state classification program shown in FIG5 . According to the current motion state and target motion state of the human body determined by the EEG and kinematic features, the computer 3 performs closed-loop spinal cord electrical stimulation control. .

The duration of the electrical stimulation is related to the duration of the motion state; the electrical stimulation frequency is related to the target action joint angle, and can be calibrated according to the human body condition in the first few rounds of each rehabilitation training. The calibrator is preferably a clinician or a physical therapist. In this embodiment, the electrical stimulation frequency and amplitude are selected as closed-loop control self-tuning parameters, and a control method combining PID and an inverse dynamics model is adopted. The actual motion state is compared with the target motion state judged according to the EEG and kinematic characteristics to obtain the motion state deviation. The electrical stimulation frequency and amplitude deviation are calculated by PID as the output error of the inverse dynamics model, and the real-time kinematic data such as the angular velocity of the joint motion are used as negative feedback input. Then, as shown in the dotted line in the figure, the error propagation method is used to update the connection weight coefficient of the inverse dynamics model, and finally the sum of the inverse dynamics model output and the PID output is limited and output to the spinal cord stimulation device. During rehabilitation training, after several rounds of model updating, the connection weight coefficient of the inverse dynamics model will reach a stable state. Among them, the dynamic inverse dynamics model can be built using a neural network, which establishes the relationship between the electrical stimulation frequency and amplitude and the target motion state.

For the parameters of the dynamic inverse dynamics model, the system simulation experiments are performed by setting different values and comparing the experimental results. In the PID feedback control part, the PID parameters can be determined by the Cohen-Coon method and the CHR method.

Claims (5)

1. A closed-loop epidural electrical stimulation system for upper limb function recovery, characterized in that: The closed-loop epidural electrical stimulation system includes a motion capture device, an electroencephalogram acquisition device, a control device, and a spinal cord electrical stimulation device, wherein: The motion capture device is used to acquire multi-angle human posture images in real time and send the human posture images to the control device; The EEG acquisition device is used to acquire the EEG signals of the human body in real time, and the EEG signals are sent to the control device after being amplified and filtered; The control device receives the human posture image from the motion capture device and the EEG signal from the EEG acquisition device, and then processes and analyzes the upper limb posture data (4) in the human posture image and the EEG signal features (5) in the EEG signal in real time, generates an instruction containing electrical stimulation parameters and sends it to the spinal cord electrical stimulation device; The spinal cord electrical stimulation device (7) receives instructions containing electrical stimulation parameters from the control device and sends electrical stimulation pulses to the human spinal cord; The control device receives the real-time posture data and the original EEG signal of the human body and performs preprocessing to generate processed upper limb posture data (4) and EEG signal characteristics (5), then obtains the difference between the current movement state and the target state of the human body based on the upper limb posture data (4) and the EEG signal characteristics (5), generates an electrical stimulation instruction, and sends it to the spinal cord electrical stimulation device (7); The control device stores a human posture estimation program based on deep learning, which can mark the key points of the upper limbs of the human body according to the multi-view human posture images transmitted by the motion capture device through the deep learning model, and three-dimensionally reconstruct the spatial coordinates of each key point to generate kinematic data of the upper limbs of the human body as the upper limb posture data (4); The control device stores an EEG signal processing program, which performs bandpass filtering and spatial filtering on the original EEG signal transmitted by the EEG acquisition device to obtain EEG signal features (5); The control device stores a motion state classification program constructed based on a neural network (6), wherein the motion state classification program determines the current motion state of the human body according to kinematic data and EEG signal characteristics (5), and then obtains the difference between the motion state and the target state; The control device stores an electrical stimulation control program, which has a predefined anatomical mapping built in according to the kinematic functions of the muscles of the upper limbs and the distribution of the motor neurons that control the muscles of the upper limbs between the second cervical segment and the first thoracic segment of the spinal cord. The electrical stimulation control program first determines the electrical stimulation site according to the predefined anatomical mapping, and then obtains specific electrical stimulation parameters according to the difference between the current motion state and the target state, including the frequency, amplitude, width and number parameters of the electrical stimulation pulses. Finally, the electrical stimulation site and electrical stimulation parameters are jointly encoded into an electrical stimulation instruction and sent to the spinal cord electrical stimulation device. 2. A closed-loop epidural electrical stimulation system for upper limb function recovery according to claim 1, characterized in that: the motion capture device is mainly composed of at least two cameras (1), and the cameras (1) collect images of human upper limbs from multiple perspectives as human posture images. 3. A closed-loop epidural electrical stimulation system for upper limb function recovery according to claim 1, characterized in that: the EEG acquisition device comprises at least a plurality of acquisition channels, a reference channel and an EEG electrode array (2), the EEG electrode array (2) comprises a plurality of electrodes arranged on the human brain, each acquisition channel is connected to a respective electrode, the reference channel is connected to an electrode and is grounded, and the electrodes of the acquisition channel are arranged on different brain regions to acquire EEG changes in the brain regions. 4. A closed-loop epidural electrical stimulation system for upper limb function recovery according to claim 3, characterized in that the EEG acquisition device includes a preamplifier, a differential amplifier and an analog filter for processing the original electrical signals collected by each channel, and the original electrical signals collected by each channel are sequentially output to the control device after preamplification, differential amplification and filtering. 5. A closed-loop epidural electrical stimulation system for upper limb function recovery according to claim 1, characterized in that: the spinal cord electrical stimulation device (7) receives the electrical stimulation instruction from the control device, applies an electrical stimulation pulse sequence with corresponding parameters to the electrical stimulation site on the dura mater of the spinal cord in the electrical stimulation instruction, and the motor neurons located at the electrical stimulation site are activated, thereby controlling the muscles to contract, causing the upper limbs to make corresponding movements, which are then collected by the motion capture device and fed back to the control device, and the closed-loop control is continuously repeated to perform electrical stimulation.