Active induction-based rehabilitation system and device for mi-bci

Abstract

The application discloses a kind of active induction rehabilitation system and equipment based on MI-BCI, belong to brain-computer interface technical field.The rehabilitation system includes MI-EEG signal acquisition and preprocessing module, data feature extraction and classification algorithm module and FNS active induction module;MI-EEG signal acquisition and preprocessing module uses the signal acquisition mode of MI in combination with EEG to collect original EEG brain electrical signal, and carries out data preprocessing;Data feature extraction and classification algorithm module is extracted to the ERD / ERS feature of electroencephalogram alpha band and beta band rhythm by one-to-many common space mode OVR-CSP method, then using VGG-LSTMnet network is classified to feature;FNS active induction module is used to complete active induction according to the result of data feature extraction and classification algorithm module classification.This application not only has the advantage of realizing accurate rehabilitation under non-invasive condition, but also can realize active induction on the basis of neuromuscular feedback pathway, to improve the dysfunction caused by neuromuscular pathway obstruction of patient.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, and in particular to an active induction rehabilitation system and device based on MI-BCI. Background Technology

[0002] Brain-computer interface (BCI) technology is a novel rehabilitation therapy technology controlled by neural activity and capable of real-time feedback output. It can establish a communication and control channel between the brain and the external environment by decoding an individual's psychological intentions, which is independent of peripheral nerve and muscle conduction. It transforms neural activity into internal or external command signals, enabling direct interaction between the brain and the external environment.

[0003] Electroencephalography (EEG) is convenient to acquire and has the advantages of being non-invasive, repeatable, and having high spatiotemporal resolution. Therefore, using EEG as a non-invasive method for acquiring BCI signals can avoid the safety issues such as risks and immune reactions associated with invasive BCI surgery, and has unique application advantages and value in the field of rehabilitation medicine.

[0004] Motor imagery (MI), as a way to induce EEG signals, can induce event-related desynchronization of μ and β rhythms in the brain's primary motor cortex without external stimuli or obvious motor output. By analyzing the MI tasks corresponding to different characteristic changes, it is possible to understand the user's true motor intentions and spontaneously execute specific real-world interactive tasks. Therefore, the use of MI-EEG offers possibilities for the safety and ease of promotion of BCI.

[0005] Functional neuromuscular stimulation (FNS) outputs electrical currents to stimulate peripheral muscles. Through a pathway of "stimulated muscle - nerve controlling the muscle - central nervous system - efferent nerve controlling the muscle - back to the stimulated muscle," it promotes neural remodeling, thereby strengthening central nervous system control over the periphery and promoting the recovery of neural function. Therefore, FNS technology has been applied clinically; however, its application primarily focuses on peripheral neuromuscular stimulation with limited central nervous system involvement. It cannot directly establish pathways for central nervous system control of active movement. Therefore, MI-BCI (Molecular Induction-Body Integration) can be used as a "switch" for active induction, constructing an active induction rehabilitation system based on MI-BCI, where active induction is achieved through FNS.

[0006] However, MI-EEG signals are weak, and in clinical applications, small sample sizes are typically used. Signal quality is directly affected by the subject's level of focus during MI and the excitability of motor cortex neurons. Furthermore, signal acquisition often uses simple symbols or images as prompts, lacking a connection to the subject, making it difficult to guarantee the intensity and duration of the subject's active MI, thus leading to decreased signal quality. In addition, interference from electrooculography (EOG), electromyography (EMG), and volume conduction effects further exacerbate signal distortion and reduce the MI-EEG signal-to-noise ratio (SNR). Although artifact removal methods such as wavelet transform and blind source separation can improve the SNR to some extent, they cannot fundamentally improve the MI-EEG signal quality. Summary of the Invention

[0007] To address the aforementioned problems, this invention aims to provide an active induction rehabilitation system and device based on MI-BCI, which not only has the advantage of achieving precise rehabilitation under non-invasive conditions, but also enables active induction based on neuromuscular feedback pathways to improve functional impairments caused by neuromuscular pathway blockages, and further effectively alleviate problems such as insufficient therapists and uneven distribution of medical resources.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] The MI-BCI-based active induction rehabilitation system includes an MI-EEG signal acquisition and preprocessing module, a data feature extraction and classification algorithm module, and an FNS active induction module.

[0010] The MI-EEG signal acquisition and preprocessing module uses a combination of MI and EEG signal acquisition to acquire raw EEG brain signals and performs data preprocessing.

[0011] The data feature extraction and classification algorithm module is used to extract and classify features from preprocessed EEG signals;

[0012] The FNS active guidance module is used to complete active guidance based on the results of data feature extraction and classification algorithm modules.

[0013] Furthermore, the method for data preprocessing of the raw EEG electroencephalogram (EEG) signals by the MI-EEG signal acquisition and preprocessing module includes the following steps:

[0014] (1): Notch filtering was applied to the original EEG brainwave signal to remove power frequency interference;

[0015] (2): Low-frequency drift and high-frequency noise are removed by bandpass filtering from 0.5Hz to 50Hz;

[0016] (3): Use the FastICA algorithm to eliminate electrooculography artifacts;

[0017] (4) Use the Fast Fourier Transform method to refine the frequency to 1Hz and extract the time window of the EEG signal from the 14 channels;

[0018] (5): Calculate the power spectral density of the EEG signal as the primary feature of the input data feature extraction and classification algorithm module.

[0019] Furthermore, the method for calculating the power spectral density of EEG signals is as follows:

[0020]

[0021] In the formula, P represents the power spectral density; w k The discrete frequency component is represented by N, the signal length is represented by y(t), the time-domain signal is represented by i, and the imaginary unit is represented by i.

[0022] Furthermore, the specific operations of the data feature extraction and classification algorithm module for feature extraction of the preprocessed EEG signals include the following steps: using the one-to-many co-space mode OVR-CSP method, ERD / ERS features of the EEG signal power spectral density data input by the MI-EEG signal acquisition and preprocessing module are extracted from the α-band and β-band rhythms of the EEG signals.

[0023] Furthermore, the specific operations for extracting ERD / ERS features from the alpha and beta band rhythms of EEG signals using the OVR-CSP method include the following steps:

[0024] Step 1: Only three channels C3, C2, and C4 are considered for the preprocessed EEG signals;

[0025] Step 2: Calculate brain region maps for the frequency band (8-14Hz) and the frequency band (14-30Hz), mapping ERD to red and ERS to blue.

[0026] Furthermore, the data feature extraction and classification algorithm module adopts a hybrid LSTM and VGG network structure, VGG-LSTMnet, to classify the ERD / ERS features of the extracted EEG signal α-band and β-band rhythms.

[0027] Furthermore, the VGG-LSTMnet network model includes three convolutional layers and one 1-D Conv layer. The first convolutional layer is a two-layer stacked structure with 32 kernels [3×3], the second convolutional layer is a two-layer stacked structure with 64 kernels [3×3], and the third convolutional layer is a single-layer structure with 128 kernels [3×3]. Each convolutional layer corresponds to one LSTM time-level unit structure. The output of the 1-D Conv layer is fused with the LSTM output in parallel and then input into the fully connected layer. The fully connected layer still adopts a two-layer stacked structure. The first layer is a flat structure, which is transformed into a 1×512 vector through a fully connected layer to complete the final classification. Based on the number of categories required by the sample label, it finally enters the Softmax layer to complete the final solution of the multi-class problem.

[0028] Furthermore, the FNS active induction module relies on PyCharm Community Edition software to complete active induction. It adjusts the EEG threshold according to the patient's test results and the current intensity according to the patient's tolerance. While collecting EEG signals, it plays a guiding video for the patient and prompts the patient to perform the MI task autonomously through visual stimulation and voice prompts. When the EEG device collects MI information and reaches the active induction threshold, it will trigger FNS and stimulate the corresponding neuromuscular through electrode conduction to complete active induction.

[0029] Furthermore, the present invention also includes an active-induced rehabilitation device based on MI-BCI, comprising at least one processor and a memory communicatively connected to the processor, wherein the memory stores the active-induced rehabilitation system based on MI-BCI as described in any one of claims 1-8, such that the processor is capable of executing instructions of the active-induced rehabilitation system based on MI-BCI.

[0030] The beneficial effects of this invention are:

[0031] 1. The active induction rehabilitation system based on MI-BCI in this invention realizes a highly robust and accurate feature extraction method of MI-BCI, which can meet the requirements of general data dimensionality reduction and accurate classification methods for different types of functional impairments; and applies MI-BCI to control FNS to actively induce stimulation of neuromuscular function to improve the patient's functional impairment, and generates feedback by stimulating the peripheral nervous system to reshape brain function, thereby strengthening the central regulation of the periphery.

[0032] 2. The active induction rehabilitation system based on MI-BCI in this invention can not only achieve precise rehabilitation under non-invasive conditions, but also achieve active induction based on neuromuscular feedback pathways to improve the functional impairment caused by the obstruction of neuromuscular pathways, so as to maximize the rehabilitation benefits of patients, and at the same time further alleviate the problems of insufficient therapists and uneven distribution of medical resources.

[0033] 3. The feature extraction and classification methods in this invention are simple and robust, and are suitable for datasets with few subjects and small data volume (small sample size). Attached Figure Description

[0034] Figure 1 This is a brain region map showing ERD and ERS phenomena displayed by channels C3, C2, and C4 when imagining hand movements in this invention;

[0035] Figure 2 This is a schematic diagram of the VGG-LSTMnet network model structure in this invention. Detailed Implementation

[0036] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0037] Example 1:

[0038] Example 1 provides an active induction rehabilitation system based on MI-BCI, including an MI-EEG signal acquisition and preprocessing module, a data feature extraction and classification algorithm module, and an FNS active induction module.

[0039] The MI-EEG signal acquisition and preprocessing module uses a signal acquisition method combining MI and EEG to acquire raw EEG brain signals;

[0040] The preprocessing operations for the acquired raw EEG signals include: (1) applying a 50Hz notch filter to the raw EEG signals to remove power frequency interference; (2) applying a bandpass filter from 0.5Hz to 50Hz to remove low-frequency drift and high-frequency noise; (3) eliminating electrooculogram artifacts using the FastICA algorithm; (4) refining the frequency of the EEG signals to 1Hz using the Fast Fourier Transform method and extracting the time window of the EEG signals from the 14 channels; and (5) calculating the power spectral density (PSD) of the EEG signals, which is then used as a primary feature input to the data feature extraction and classification algorithm module for further feature extraction and classification. The power spectral density extracted in this invention is the average power per unit frequency resolution, and its calculation formula is as follows:

[0041]

[0042] In the formula, P represents the power spectral density; w k The discrete frequency component is represented by N, the signal length is represented by y(t), the time-domain signal is represented by i, and the imaginary unit is represented by i.

[0043] Furthermore, the data feature extraction and classification algorithm module is used to extract and classify features from the EEG electroencephalogram signals preprocessed by the MI-EEG signal acquisition and preprocessing module.

[0044] Specifically, data feature extraction involves using the one-versus-rest commonspatial pattern (OVR-CSP) method to extract ERD / ERS features of the α-band and β-band rhythms of the EEG signal power spectral density data input from the MI-EEG signal acquisition and preprocessing module.

[0045] Data classification is achieved by introducing generative adversarial networks (GANs) into the structure of a convolutional neural network (CNN) and optimizing the network weights to build a highly robust semi-supervised learning model, thus enabling accurate classification of MI-EEG electroencephalogram (EEG) signals. The feature extraction and classification methods in this invention are simple and robust, making them suitable for datasets with a small number of subjects and limited data volume (small sample size).

[0046] The FNS active induction module is used to perform active induction. After the raw MI-EEG signal is processed by the MI-EEG signal acquisition and preprocessing module, and the data feature extraction and classification algorithm module, it is transformed into motor imagery classification data that can be recognized by the BCI device control terminal. Then, the BCI controls functional neuromuscular electrical stimulation (FNS) to generate impulses that act on peripheral nerves and muscles, achieving the purpose of active induction. It should be noted that the FNS active induction module is existing technology and is already in clinical use. The functional neuromuscular electrical stimulation (FNS) output current stimulates peripheral muscles. Through the mode of "stimulated muscle - nerve controlling the muscle - nerve center - efferent nerve controlling the muscle - back to the stimulated muscle," it promotes neural remodeling, thereby strengthening the central nervous system's control over the periphery and promoting the recovery of nerve function.

[0047] The active induction rehabilitation system based on MI-BCI in this invention relies on PyCharm Community Edition software. It adjusts the EEG threshold according to the patient's test results and the current intensity according to the patient's tolerance. It collects EEG signals and performs corresponding data processing. At the same time, it plays guidance videos for the patient. The FNS active induction module converts the patient's EEG signals into motor imagery classification data that can be recognized by the BCI device control terminal. Through visual stimulation and voice prompts, the patient performs MI tasks autonomously. When the EEG device collects MI information and reaches the active induction threshold, it will trigger FNS, which will stimulate the corresponding neuromuscular signals through electrode conduction to complete the active induction.

[0048] Example 2:

[0049] Example 2 provides a specific implementation of a data feature extraction and classification algorithm module based on Example 1.

[0050] The specific steps for extracting ERD / ERS features from the preprocessed EEG power spectral density data using the OVR-CSP method for the α and β band rhythms of the EEG signals include the following: Analyzing the features extracted by CSP, all subjects exhibited corresponding ERD and ERS phenomena in the cerebral cortex during two types of motor imagery. For example, for two different motor imagery tasks in Dataset 1, namely imagining hand and foot lifting movements, the goal was to generate ERDS mappings for each of the two tasks. First, 5s of data were imported. The dataset contains multiple channels; only three channels, C3, Cz, and C4, were considered. Then, brain region maps were calculated for the frequency bands (8–14 Hz) and (14–30 Hz), mapping ERD to red and ERS to blue. The ERD and ERS phenomena displayed in channels C3, Cz, and C4 during hand imagery are shown in the attached figure. Figure 1 As shown. Finally, a calibration test was performed for multiple comparisons within the channel to estimate significant ERDS values.

[0051] Furthermore, in this embodiment, by introducing generative adversarial networks (GANs) into the structure of a convolutional neural network (CNN) and optimizing the network weights, a highly robust semi-supervised learning model is built to achieve accurate MI-EEG brainwave signals; specifically, the accuracy of MI-EEG brainwave signals is achieved through the VGG-LSTMnet neural network structure.

[0052] The advantage of the VGG network architecture lies in its implicit regularization term, which decomposes 5×5 or 7×7 convolutions into a stack of multiple 3×3 convolutions. Fully connected layers use 1×1 convolutional kernels instead. By flexibly modifying the network structure and pre-initializing some layers, model training can be faster and more accurate than typical networks. Adjusting the number of convolutional layers and kernels in the VGG network structure yields a convolutional neural network model with fewer layers, named VGG-LSTMnet. The model framework of the VGG network structure is shown in Table 1 below.

[0053] Table 1 VGG Network Model Framework

[0054]

[0055] By adjusting the number of convolutional layers and kernels in the VGG network structure, a convolutional neural network model with a smaller number of layers was obtained, named VGG-LSTMnet, and its model framework is shown in Table 2 below.

[0056] Table 2 VGG-LSTMnet Network Model Framework

[0057]

[0058] The VGG-LSTMnet neural network structure of this invention is shown in the appendix. Figure 2As shown, the model consists of three convolutional layers. The first layer is a two-layer stacked structure with 32 convolutional kernels [3×3], denoted as C1; the second layer is a two-layer stacked structure with 64 convolutional kernels [3×3], denoted as C2; and the third layer is a single-layer structure with 128 convolutional kernels [3×3], denoted as C3. The Long Short-Term Memory (LSTM) network needs to capture different temporal pattern features across multiple frames, corresponding to the three LSTM time-stamp unit structures. The inputs at times t-1, t, and t+1 come from features in the shallow, deeper, and more advanced layers, respectively (the transformation between spatial and temporal features, mapping high-dimensional feature vectors to low-dimensional space). After max pooling, each layer is expanded into vectors by a reshape layer, serving as the input structure for the LSTM network, achieving serial fusion. Ultimately, the LSTM layer and layers with different numbers of storage units, each composed of 128 cells, achieve the best results. It is important to note that after the output in this model, a 1-D Conv layer, i.e., a 1×1 convolutional layer with 64 convolutional kernels, is added. A 1×1 convolution essentially performs a linear combination of different channels on each pixel, filtering adjacent elements of the one-dimensional input while preserving the original planar structure of the feature map. Adjusting the depth also helps reduce the dimensionality of the feature map. Next, the output of the 1-D Conv layer is fused in parallel with the LSTM output. A Reshape layer changes the dimension of the input data, excluding the dimension of the number of samples (batch size), while keeping the content unchanged. The fused data is then fed into a fully connected layer, which still uses a two-layer stacked structure. The first layer is a flat structure, which is transformed into a 1×512 vector by the fully connected layer to complete the final classification. Based on the number of classes required for the sample labels, the vector finally enters the Softmax layer to solve the multi-class classification problem.

[0059] In this invention, ERD / ERS features of alpha and beta frequency bands of EEG signals are input into a VGG-LSTMnet neural network model. The network model decomposes 5×5 or 7×7 convolutions into a stack of multiple 3×3 convolutions, and replaces fully connected layers with 1×1 convolution kernels. This allows for flexible modification of the network structure and pre-initialization of network layers, resulting in faster model training, excellent generalization ability, and higher accuracy compared to general networks. Experimental comparisons show a classification accuracy of up to 90.38%, making it suitable for high-performance multi-task classification and recognition systems for motor imagery. This interface achieves accurate MI-EEG classification of limbs, tongue, and eyes while ensuring small sample sizes and high robustness. The classification results reach the active induction threshold, inducing patients' voluntary movement and enabling active rehabilitation treatment for patients with motor, visual, language, and cognitive impairments.

[0060] The classification performance of the VGG-LSTMnet neural network model in this embodiment was verified. The average classification performance parameters were obtained through 5-fold cross-validation experiments. The average accuracy of different categories obtained from different classification algorithms and experiments with each subject is shown in Table 3. The experimental results prove the effectiveness of the VGG-LSTMnet neural network model in this invention.

[0061] Table 3 Classification performance of Dataset 1

[0062]

[0063] Application examples:

[0064] The active induced rehabilitation system based on MI-BCI in this invention is created using Python and built upon the Keras+Tensorlow deep learning framework of Theano. The environment consists of an 8th generation Intel processor (6 cores, 5.2GHz), 32.0GB DDR4 memory, a 500GB SSD + 4TB SATA hard drive, and a GeForce RTX 2080 GPU with 16GB of video memory. The dataset uses the internationally recognized GDF format and includes left (right) hand (foot) movements, eye movements (four directions), and tongue MI EEG signals. The sampling frequency is 125Hz, and the sampling channel uses 16 leads. Electrode placement is performed using electrode caps according to the international 10-20 standard. The signals from channels C3, C4, and Cz are closely related to brain motor imagery and serve as the main reference electrodes for the neural network classifier, supplemented by three midline electrodes (Fz, Oz, Pz) and two ear electrodes (A1, A2).

[0065] The raw EEG signal was subjected to a 50Hz notch filter to remove power line interference; low-frequency drift and high-frequency noise were removed using a bandpass filter from 0.5Hz to 50Hz. A fixed-point iterative FastICA algorithm was introduced to eliminate electrooculogram artifacts and determine the non-Gaussian maximum value of the WTX (wavelength to frequency transformation). The Fast Fourier Transform method was used to refine the frequency to 1Hz accuracy, and time windows of the EEG signal were extracted from the 14 channels, with bandpass filtering used to select the desired frequency bands.

[0066] The original dataset was preprocessed to obtain EEG signals after noise removal. Then, the ERD / ERS features of the α and β frequency bands of the EEG signals were extracted using the OVR-CSP method. Finally, GANs were introduced into the CNN structure and the network weights were optimized to build a highly robust semi-supervised learning model to complete the accurate classification of MI-EEG.

[0067] The process of using this rehabilitation system for active-induced rehabilitation training is as follows: Determine the patient's position, identify the target nerves and muscles for active induction, fully expose the skin, attach electrode pads to the target points, set FNS stimulation parameters, and select appropriate stimulation modes, intensity, frequency, pulse width, treatment time, and rehabilitation cycle based on the patient's function and tolerance. The patient wears an EEG device and watches MI task-guided videos, achieving active-induced MI-BCI training through visual, auditory, and FNS multi-sensory input interaction. The specific implementation process includes the following steps:

[0068] (1) Comprehensively assess the patient's functional status in terms of movement, sensation, cognition, etc., identify the target neuromuscular pathways for active induction, and then determine the rehabilitation treatment prescription. The prescription should include, but is not limited to: stimulation mode, stimulation intensity, frequency, pulse width, treatment time, and rehabilitation treatment cycle.

[0069] (2) Select a quiet room as the treatment room to ensure the patient's concentration. The patient should be seated with arms relaxed, forearms fully exposed. Attach electrodes to the radial end of the wrist and three finger-widths proximal to the ulnar styloid process, and three finger-widths distal to the lateral epicondyle of the wrist extensor muscles. Place the EEG cap over the patient's head, laying it flat. Electrode placement follows the 10-20 international standard lead system, with a total of 16 leads: FP1, FP2, F3, F4, F7, F8, C3, C4, T3, T4, T5, T6, P3, P4, O1, and O2. The EEG power spectrum shows that the power density of MI is mainly concentrated in the 0-30Hz range. The frequency range of oscillating cortical activity corresponding to MI EEG signals is mainly reflected in two frequency bands: alpha waves (8-13Hz) and beta waves (13-30Hz). The required frequency band can be intercepted by bandpass filtering, and the EEG threshold can be calculated based on the ratio of alpha waves to beta waves.

[0070] (3) Open the PyCharm Community Edition software. Adjust the EEG threshold according to the patient's test results and the current intensity according to the patient's tolerance, aiming to induce wrist dorsiflexion without causing pain. The frequency is 40Hz, the pulse width is 200-400μs, the treatment lasts for 15 minutes, 3 times a week, for 4 weeks. Right-click and select "Run 'BCI-hub'", select "COM5" for the serial port, and then click "Start Acquisition". While acquiring EEG signals, play a guiding video for the patient. Through visual stimulation and verbal prompts, the patient will perform a motor imagery task of wrist dorsiflexion. When the EEG device acquires MI information, it will trigger electrical stimulation, thereby inducing the patient's wrist dorsiflexion. If the patient does not perform the MI task or does not reach the EEG threshold, active induction of FNS cannot be induced. The EEG threshold is a test performed before treatment. Have the patient perform the MI task and observe the threshold at which they can induce FES. Set this threshold as the training threshold. This threshold can be adjusted according to the patient's treatment results.

[0071] (4) Ask the patient to watch the video and imagine the action they are about to perform based on the video guidance. No actual movement is required. After successful MI, the electrode pads at the target location will be electrically stimulated. If any discomfort occurs during the stimulation process, please inform the therapist immediately. The therapist will then immediately click the off button on the electrical stimulator to avoid causing discomfort to the patient.

[0072] Example 3:

[0073] Example 3 provides an active induced rehabilitation device based on MI-BCI, including at least one processor and a memory communicatively connected to the processor. The memory stores the active induced rehabilitation system based on MI-BCI described in Example 1, enabling the processor to execute the instructions of the active induced rehabilitation system based on MI-BCI.

[0074] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An active induced rehabilitation system based on MI-BCI, characterized in that: The system comprises an MI-EEG signal acquisition and preprocessing module, a data feature extraction and classification algorithm module, and an FNS active induction module. The MI-EEG signal acquisition and preprocessing module acquires original EEG signals in an MI-EEG signal acquisition mode and performs data preprocessing. The data feature extraction and classification algorithm module is used for feature extraction and classification of the preprocessed EEG signals. The FNS active induction module is used for active induction according to the classification result of the data feature extraction and classification algorithm module. The method for data preprocessing of the original EEG signals by the MI-EEG signal acquisition and preprocessing module comprises the following steps: (1) notch filtering of the original EEG signals to remove power frequency interference; (2) band-pass filtering of 0.5 Hz to 50 Hz to remove low-frequency drift and high-frequency noise; (3) elimination of eye movement artifacts by using FastICA algorithm; (4) frequency accuracy of 1 Hz by using fast Fourier transform method, and time window extraction of the EEG signals from 14 channels; (5) calculation of the power spectral density of the EEG signals as the input data feature extraction and classification algorithm module primary feature; The specific operation of the data feature extraction and classification algorithm module for feature extraction of the preprocessed EEG signals comprises the following steps: ERD / ERS feature extraction of the EEG signal power spectral density data input by the MI-EEG signal acquisition and preprocessing module through the one-to-many common spatial pattern OVR-CSP method; classification of the extracted EEG signal α band and β band rhythm ERD / ERS features by using the mixed VGG-LSTMnet network structure; The specific operation of the OVR-CSP method for ERD / ERS feature extraction of the EEG signal α band and β band rhythm comprises the following steps: Step 1: only three channels C3, Cz and C4 are considered for the preprocessed EEG signals; Step 2: calculation of the frequency band (8-14 Hz) and frequency band (14-30 Hz) brain region map, mapping ERD as red and ERS as blue; The VGG-LSTMnet network model comprises three convolutional layers and a 1-D Conv layer, the first convolutional layer is a two-layer stacked structure with 32 convolutional kernels of size [3x3], the second convolutional layer is a two-layer stacked structure with 64 convolutional kernels of size [3x3], and the third convolutional layer is a single-layer structure with 128 convolutional kernels of size [3x3]; each convolutional layer corresponds to an LSTM time unit structure; the output result of the 1-D Conv layer is parallelly fused with the LSTM output and then input to a fully connected layer, the fully connected layer also adopts a two-layer stacked structure, the first layer is a flat structure, which is changed into a 1x512 vector through a fully connected layer to complete the final classification, according to the number of categories required by the sample label, finally entering the Softmax layer to complete the solution of the final multi-classification problem. The FNS active induction module is completed by relying on the Pycharm Community Edition software, adjusts the EEG threshold according to the test situation of the patient, adjusts the current intensity according to the tolerance degree of the patient, collects the EEG signal and plays a guide video for the patient at the same time, prompts the patient to perform the MI task by visual stimulation and voice and text, when the EEG device collects the MI information and reaches the active induction threshold, the FNS will be triggered, the corresponding nerve and muscle are stimulated through the electrode conduction, and the active induction is completed; When the patient does not perform the MI task or does not reach the EEG threshold, the FNS active induction cannot be triggered; the EEG threshold is tested before treatment, the patient performs the MI task, the threshold at which the FES can be excited is observed, and the threshold is set as the training threshold, and the threshold can be adjusted according to the treatment situation of the patient.

2. The MI-BCI based active induced rehabilitation system according to claim 1, characterized in that: The power spectral density calculation method of the EEG signal is In the formula, P represents the power spectral density; w k represents a discrete frequency component, N represents a signal length, y(t) represents a time domain signal, and i represents an imaginary unit.

3. A MI-BCI based active induced rehabilitation device, characterized in that: The system comprises at least one processor, and a memory connected with the processor in communication, and the memory has the active induction rehabilitation system based on the MI-BCI in any one of claims 1-2 stored therein, so that the processor can execute the instructions of the active induction rehabilitation system based on the MI-BCI.

Images