Intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus

Abstract

The invention discloses an intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus, and relates to the technical field of electrical stimulation rehabilitation therapeutic apparatuses, the intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus comprises an electroencephalogram signal acquisition module, an electromyographic signal acquisition module and a mixed signal processing module, the mixed signal processing module is connected with the electroencephalogram signal acquisition module and the electromyographic signal acquisition module; the method is used for preprocessing electroencephalogram and electromyogram signals, fusing multi-modal features and classifying motion intentions. By installing the mixed signal processing module, preprocessing, multi-modal feature fusion and motion intention classification of electroencephalogram and electromyogram signals are achieved, a preprocessing unit can effectively remove ocular artifacts and power frequency noise, meanwhile, independent component analysis and a dynamic baseline calibration strategy are adopted, the signal quality is further improved, and the accuracy of motion intention classification is improved. According to the multi-mode signal processing mode, the limitation of single signal processing in the prior art is overcome, and the accuracy and effectiveness of rehabilitation treatment are improved.

Description

Technical Field

[0001] The present invention relates to the field of electro - stimulation rehabilitation treatment instruments, and specifically to an intelligent brain - controlled muscle electro - stimulation rehabilitation treatment instrument. Background Art

[0002] In the field of modern rehabilitation treatment, with the continuous development of neuroscience and biomedical engineering technologies, the demand for rehabilitation treatment for motor dysfunction caused by nervous system diseases is increasing day by day. However, traditional muscle electro - stimulation devices mostly adopt a unidirectional stimulation mode, relying on therapists' experience to preset fixed parameters, and it is difficult to dynamically adapt to the real - time physiological state of patients.

[0003] Patent CN105233406B discloses a stroke functional electro - stimulation rehabilitation system based on a synergistic muscle group stimulation mode. The above - mentioned patent realizes the purpose of assisting patients in motor function rehabilitation when the patients are undergoing rehabilitation training.

[0004] The above - mentioned patent has real - time and accurate motion signal acquisition. The signal transmission adopts Bluetooth technology, with low latency and low power consumption. The signal receiver is connected to the computer through USB, the motion signal analysis module communicates with the main controller through USB, and the main controller communicates with each electro - stimulator through the CAN bus. When the patient is undergoing rehabilitation training, by detecting the patient's real - time motion signal, a corresponding electro - stimulation plan can be constructed in the motion signal analysis module according to the synergistic muscle group control theory, and then coordinated by the main controller to control the multi - channel electro - stimulator to stimulate the corresponding muscles of the patient, so as to achieve the purpose of assisting the patient in motor function rehabilitation. However, there is still room for optimization in the aspect of dynamically adjusting by accurately combining the patient's motion intention and muscle state.

[0005] Therefore, this application proposes an intelligent brain - controlled muscle electro - stimulation rehabilitation treatment instrument that can accurately combine the patient's motion intention and muscle state for dynamic adjustment. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent brain - controlled muscle electro - stimulation rehabilitation treatment instrument to solve the technical problem of not being able to accurately combine the patient's motion intention and muscle state for dynamic adjustment as mentioned in the above background art.

[0007] To achieve the above - mentioned purpose, the present invention provides the following technical solution: An intelligent brain - controlled muscle electro - stimulation rehabilitation treatment instrument, including an electroencephalogram signal acquisition module, an electromyogram signal acquisition module, and a mixed signal processing module. The mixed signal processing module is connected to the electroencephalogram signal acquisition module and the electromyogram signal acquisition module, and is used for pre - processing electroencephalogram and electromyogram signals, multi - modal feature fusion, and motion intention classification;

[0008] The hybrid signal processing module includes: a preprocessing unit, a lightweight variational autoencoder, a cross-modal temporal attention fusion unit, and a multi-task learning strategy;

[0009] The preprocessing unit performs 4 - 50Hz band-pass filtering, electrooculogram artifact removal, and independent component analysis on electroencephalogram signals, 30 - 150Hz band-pass filtering and power frequency noise suppression on electromyogram signals, and adopts a dynamic baseline calibration strategy to eliminate individual impedance differences, ensuring high signal-to-noise ratio input of multi-modal biological signals;

[0010] The lightweight variational autoencoder extracts deep spatio-temporal features of electroencephalogram and electromyogram respectively through unsupervised learning;

[0011] The cross-modal temporal attention fusion unit dynamically fuses multi-modal features through a weight allocation mechanism;

[0012] The multi-task learning strategy synchronously executes motion intention classification, signal reconstruction, and physiological consistency constraint tasks.

[0013] Preferably, the electroencephalogram signal acquisition module can accurately capture weak electrical signals emitted by the brain, providing a reliable data basis for subsequent signal processing and analysis, while the electromyogram signal acquisition module accurately monitors the electrical activity of muscles to understand the contraction and relaxation states of muscles in real time. The two work together to provide strong support for the formulation and implementation of rehabilitation treatment plans;

[0014] The electroencephalogram signal acquisition module is used to acquire electroencephalogram signals of patients, including a 32-channel electrode array with a sampling frequency of 500Hz;

[0015] The electromyogram signal acquisition module is used to acquire surface electromyogram signals of patients, including a 4-channel electrode array with a sampling frequency of 1000Hz.

[0016] Preferably, the lightweight variational autoencoder is used to compress the spatio-temporal dimensions of electroencephalogram and electromyogram signals through depthwise separable convolution, construct low-dimensional feature representations in the latent variable space, and ensure feature interpretability through KL divergence constraints, realizing efficient feature extraction on embedded platforms;

[0017] The lightweight variational autoencoder includes: an encoder network, a decoder network, and a latent variable space constraint;

[0018] The encoder network has 3 layers of depthwise separable convolution, with each convolution kernel size of 3×3 and output channel numbers of 128, 256, and 512 respectively;

[0019] The decoder network has 3 layers of transposed depthwise separable convolution, and fuses low-level features through skip connections;

[0020] The latent variable space constraint adopts the KL divergence loss function, and the dimension of the latent variable is 64.

[0021] Preferably, the cross-modal temporal attention fusion unit dynamically aligns the spatio-temporal features of electroencephalogram and electromyogram signals through a bidirectional cross-attention mechanism, calculates the complementary weights between brain-muscle electrical signals, and uses residual connections to retain the original modal information, thereby solving the problem of asynchronous signal fusion and improving classification robustness.

[0022] The cross-modal temporal attention fusion unit includes: a time alignment layer, bidirectional cross-attention calculation, and residual connection.

[0023] The time alignment layer performs upsampling the electroencephalogram signal to 1000Hz through linear interpolation to synchronize the sampling rate with the electromyogram signal.

[0024] Bidirectional cross-attention calculation:

[0025] ;

[0026] Among them, Z eeg is the depth feature matrix of the electroencephalogram signal, with dimensions of R T×d , where T is the time step and d is the feature dimension. Z emg is the depth feature matrix of the electromyogram signal, with the same dimensions as Z eeg , W q , W k and W v are learnable projection weight matrices, with dimensions of R d×d , Q eeg is the query matrix generated by projecting the electroencephalogram features through W q , with dimensions of R T×d , K emg is the key matrix generated by projecting the electromyogram features through W k , with dimensions of R T×d , V emg is the value matrix generated by projecting the electromyogram features through W q , with dimensions of R T×d , K T emg is the transpose matrix of K emg . The Attention mechanism is a computational model that mimics the human attention distribution, and the Softmax function is a normalization method that converts an arbitrary real-valued vector into a probability distribution.

[0027] The residual connection adds the fused features to the original features to prevent the vanishing gradient.

[0028] Preferably, the multi-task learning strategy is used to synchronously optimize the tasks of motion intention classification, electroencephalogram and electromyogram signal reconstruction, and physiological consistency constraint. By means of a joint loss function, feature space alignment and biomechanical law joint modeling are realized, and the generalization recognition ability of the model for the motion intention of stroke patients is improved;

[0029] The multi-task learning strategy includes: a main task, an auxiliary task, and a joint loss function;

[0030] The main task is motion intention classification, which outputs the probabilities of left and right hand motor imagery and rest state. The loss function is Focal Loss, with γ = 2;

[0031] The auxiliary tasks include electroencephalogram signal reconstruction and physiological consistency constraint:

[0032] Electroencephalogram signal reconstruction: The mean square error loss constrains the output of the decoder;

[0033] Physiological consistency constraint: Through a pre-trained OpenSim muscle activation model, the cosine similarity between electromyogram features and biomechanical simulation data is restricted to be > 0.85;

[0034] The joint loss function is the weighted sum of the motion intention classification loss, the signal reconstruction loss, and the physiological consistency constraint loss. The weight ratio of the three is 5:2:3. Among them, the classification task has the highest weight to prioritize ensuring the intention recognition accuracy, the physiological consistency constraint comes second to ensure the safety of electrical stimulation, and the signal reconstruction weight is the lowest to prevent the model from overfitting.

[0035] Preferably, an electrical stimulation module is further provided in the therapeutic apparatus. The electrical stimulation module is used to generate a multi-channel adaptive electrical pulse sequence according to the motion intention classification result, dynamically adjust the stimulation intensity to match the muscle activation threshold, and combine impedance feedback to optimize the current distribution in real time to ensure accurate neuromuscular electrical stimulation of the targeted muscle group;

[0036] The electrical stimulation module includes: a multi-channel stimulation electrode array, an adaptive parameter adjustment unit, and a safety protection circuit;

[0037] The multi-channel stimulation electrode array is arranged at the motor points of the patient's targeted muscles;

[0038] The adaptive parameter adjustment unit dynamically adjusts the amplitude, frequency, and duty cycle of the stimulation pulses according to the classification confidence;

[0039] The safety protection circuit monitors the impedance change in real time and triggers overload protection.

[0040] Preferably, a virtual reality interaction module is also provided in the therapeutic apparatus. The virtual reality interaction module is used to map the motion intention classification result into limb movements in a virtual scene in real time, and enhance the patient's active participation and training immersion in neural remodeling through mirror training tasks and visual-proprioceptive feedback loops;

[0041] The virtual reality interaction module has an LCD screen integrated outside the virtual reality glasses unit for presenting rehabilitation training scenarios, motion intention mapping results, electrical stimulation parameters, and alarms.

[0042] Preferably, the preprocessing unit is used to perform 4 - 50 Hz band-pass filtering and independent component analysis on the original electroencephalogram signal to eliminate electrooculogram artifacts, and at the same time perform 30 - 150 Hz band-pass filtering and adaptive power frequency noise suppression on the electromyogram signal, and adopt a dynamic baseline calibration strategy to eliminate individual impedance differences, ensuring a high signal-to-noise ratio input of multimodal biological signals;

[0043] The preprocessing unit supports a multi-band parallel processing mode, including:

[0044] Theta band, 4 - 8 Hz feature extraction channels;

[0045] Alpha band, 8 - 13 Hz event-related desynchronization analysis channels;

[0046] Beta band, 13 - 30 Hz motion intention decoding channels;

[0047] Gamma band, 30 - 50 Hz muscle co-activation detection channels.

[0048] Preferably, a feedback control unit is also designed in the therapeutic apparatus. The feedback control unit is used to optimize the neuromuscular electrical stimulation parameters in real time;

[0049] The feedback control unit includes: a fatigue state detection algorithm and an adaptive stimulation strategy;

[0050] The fatigue state detection algorithm evaluates the muscle fatigue level by real-time monitoring of the average power frequency decline rate of the electromyogram signal;

[0051] When the average power frequency decline rate ≥ 15% is detected, the adaptive stimulation strategy automatically switches to the intermittent pulse mode, the stimulation / rest cycle ratio is 1:2, and the stimulation intensity is reduced to the safety threshold, while triggering a visual alarm to prompt the patient to adjust the training posture.

[0052] Preferably, the therapeutic apparatus supports the following clinical application modes:

[0053] The active rehabilitation mode triggers synchronous electrical stimulation based on motion imagination classification;

[0054] The passive induction mode induces muscle contraction through a pre-programmed electrical stimulation sequence;

[0055] The hybrid enhancement mode is a two-way closed-loop control that combines electroencephalogram intention detection and electromyogram feedback.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] 1. By installing a hybrid signal processing module, the present invention realizes the preprocessing of electroencephalogram and electromyogram signals, multi-modal feature fusion, and motion intention classification. The preprocessing unit can effectively remove electrooculogram artifacts and power frequency noise, and at the same time adopts independent component analysis and dynamic baseline calibration strategies to further improve the signal quality. This multi-modal signal processing method overcomes the limitations of single-signal processing in the prior art and improves the accuracy and effectiveness of rehabilitation treatment;

[0058] 2. By installing an electrical stimulation module, the present invention can generate a multi-channel adaptive electrical pulse sequence according to the motion intention classification result, dynamically adjust the stimulation intensity to match the muscle activation threshold, and combine impedance feedback to optimize the current distribution in real time to ensure precise neuromuscular electrical stimulation of the targeted muscle group, solving the problem that the electrical stimulation parameters in the prior art are fixed and cannot be dynamically adjusted according to the real-time state of the patient;

[0059] 3. By installing a virtual reality interaction module, the present invention maps the motion intention classification result into limb movements in a virtual scene in real time, and through mirror training tasks and visual-proprioceptive feedback loops, enhances the patient's active participation and training immersion in nerve remodeling. This immersive training method significantly improves the patient's rehabilitation enthusiasm and treatment effect, overcoming the deficiencies of single rehabilitation training methods and low patient participation in the prior art;

[0060] 4. By installing a feedback control unit, the present invention can real-time monitor the average power frequency decline rate of electromyogram signals, evaluate the muscle fatigue level, and automatically adjust the electrical stimulation parameters according to the fatigue state. This dynamic feedback mechanism ensures the safety and effectiveness of electrical stimulation, solving the problem of the lack of real-time feedback and dynamic adjustment functions in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a schematic diagram of the control work flow of the present invention;

[0062] Figure 2 is a schematic diagram of the hybrid signal processing module of the present invention;

[0063] Figure 3 is a schematic diagram of the electrical stimulation module of the present invention;

[0064] Figure 4 is a schematic diagram of the feedback control unit of the present invention. Detailed implementation manners

[0065] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0066] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "upper", "lower", "inner", "outer", "front end", "rear end", "both ends", "one end", "the other end", etc. is based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0067] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "installed", "provided with", "connected", etc. should be understood in a broad sense. For example, "connected" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0068] Please refer to Figure 1 and Figure 2 , an embodiment provided by the present invention: An intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus, including an electroencephalogram signal acquisition module, an electromyogram signal acquisition module, and a mixed signal processing module, where the mixed signal processing module is connected to the electroencephalogram signal acquisition module and the electromyogram signal acquisition module, and is used for preprocessing, multi-modal feature fusion, and motion intention classification of electroencephalogram (EEG) and electromyogram (EMG) signals;

[0069] The mixed signal processing module includes: a preprocessing unit, a lightweight variational autoencoder, a cross-modal temporal attention fusion unit, and a multi-task learning strategy;

[0070] The preprocessing unit is used for performing 4 - 50 Hz band-pass filtering, electrooculogram artifact removal, and independent component analysis on electroencephalogram signals, performing 30 - 150 Hz band-pass filtering and power frequency noise suppression on electromyogram signals, and adopting a dynamic baseline calibration strategy to eliminate individual impedance differences, so as to ensure a high signal-to-noise ratio input of multi-modal biological signals;

[0071] The lightweight variational autoencoder extracts the deep spatio-temporal features of electroencephalogram (EEG) and electromyogram (EMG) through unsupervised learning respectively;

[0072] The cross-modal temporal attention fusion unit dynamically fuses multi-modal features through a weight allocation mechanism;

[0073] The multi-task learning strategy synchronously executes tasks of motion intention classification, signal reconstruction, and physiological consistency constraint;

[0074] Furthermore, in the preprocessing stage, the EEG signal first undergoes 4 - 50 Hz band-pass filtering, and the electrooculogram (EOG) artifact components are separated by combining with the independent component analysis algorithm, such as the high-amplitude frontal fluctuations caused by the corneal-retinal potential difference. Meanwhile, the signal drift caused by individual skin impedance differences is eliminated through dynamic baseline calibration. For the EMG signal, 30 - 150 Hz band-pass filtering is used to suppress environmental high-frequency noise, and an adaptive power frequency notch filter with a center frequency of 50 Hz and an attenuation depth of ≥95.8% is used to eliminate the mains interference, ensuring that the signal-to-noise ratio of the input signal > 30 dB. This stage not only retains the key EEG frequency band features such as the Theta band, Alpha band, and Beta band, but also realizes muscle co-activation detection through parallel channels, that is, the Gamma band. The lightweight variational autoencoder adopts a depthwise separable convolution architecture, and the spatio-temporal dimensions of the EEG and EMG signals are compressed layer by layer through 3×3 convolution kernels. The number of encoder channels is 128 → 256 → 512, and a 64-dimensional low-dimensional feature representation is constructed in the latent variable space. The model introduces KL divergence constraint with a weight coefficient of 0.01 to force the latent variable distribution to approximate the standard normal distribution, which not only avoids overfitting but also enhances the feature interpretability. For the EEG signal, the lightweight variational autoencoder focuses on extracting event-related desynchronization and synchronization features, and for the EMG signal, it focuses on the time-frequency energy distribution of the motor unit action potential sequence, such as the mean power frequency. Through the unsupervised pre-training and fine-tuning mechanism, the model can achieve real-time feature extraction with a delay of ≤50 ms on the ARM embedded platform;

[0075] Meanwhile, the cross-modal temporal attention fusion unit uses bidirectional cross-attention calculation. First, the electroencephalogram (EEG) signal is upsampled to 1000 Hz through linear interpolation to synchronize the sampling rate with the electromyogram (EMG) signal. Subsequently, the dynamic interaction between the query matrix and the key-value matrix is constructed. This calculation retains the original modal information through residual connections. In the movement intention decoding of the Beta band, the weight ratio of EEG features can reach 60%. For the muscle fatigue state of the EMG, such as the average power frequency decline rate ≥ 15%, it is fed back to the central signal processing layer through the reverse attention path to form a closed-loop regulation. The multi-task learning strategy realizes the collaborative modeling of movement intention recognition and biomechanical laws through the joint loss function. The main task uses the Focal Loss with γ = 2 to classify the movement imagination and resting state of the left and right hands. In the auxiliary task, the EEG signal reconstruction constrains the waveform similarity of the decoder output through the mean square error, while the physiological consistency constraint introduces the OpenSim muscle activation model to force the cosine similarity between the EMG features and the biomechanical simulation data to be > 0.85, ensuring that the electrical stimulation parameters conform to the dynamic characteristics of muscle contraction.

[0076] Please refer to Figure 2 and Figure 4 An embodiment provided by the present invention: An intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus. The lightweight variational autoencoder is used to compress the spatio-temporal dimensions of EEG and EMG signals through depthwise separable convolution, construct a low-dimensional feature representation in the latent variable space, and ensure the interpretability of the features through KL divergence constraint, realizing efficient feature extraction on the embedded platform;

[0077] The lightweight variational autoencoder includes: an encoder network, a decoder network, and a latent variable space constraint;

[0078] The encoder network has 3 layers of depthwise separable convolution, with the convolution kernel size of each layer being 3×3, and the output channel numbers being 128, 256, and 512 respectively;

[0079] The decoder network has 3 layers of transposed depthwise separable convolution, and fuses the low-level features through skip connections;

[0080] The latent variable space constraint uses the KL divergence loss function, and the latent variable dimension is 64;

[0081] The cross-modal temporal attention fusion unit dynamically aligns the spatio-temporal features of EEG and EMG signals through the bidirectional cross-attention mechanism, calculates the complementary weights between brain-EMG signals, and retains the original modal information using residual connections, thus solving the problem of asynchronous signal fusion and improving the classification robustness;

[0082] The cross-modal temporal attention fusion unit includes: a time alignment layer, bidirectional cross-attention calculation, and a residual connection;

[0083] The time alignment layer performs upsampling the electroencephalogram (EEG) signal to 1000 Hz through linear interpolation to synchronize the sampling rate with the electromyogram (EMG) signal;

[0084] Bidirectional cross-attention calculation:

[0085] ;

[0086] Among them, Z eeg is the depth feature matrix of the EEG signal, with dimensions of R T×d , where T is the time step and d is the feature dimension. Z emg is the depth feature matrix of the EMG signal, with the same dimensions as Z eeg , W q , W k and W v are learnable projection weight matrices, with dimensions of R d×d , Q eeg is the query matrix generated by projecting the EEG features through W q , with dimensions of R T×d , K emg is the key matrix generated by projecting the EMG features through W k , with dimensions of R T×d , V emg is the value matrix generated by projecting the EMG features through W q , with dimensions of R T×d , K T emg is the transpose matrix of K emg . The Attention mechanism is a computational model that mimics the human attention allocation, and the Softmax function is a normalization method that converts any real-valued vector into a probability distribution;

[0087] The residual connection adds the fused features and the original features to prevent the vanishing gradient;

[0088] The therapeutic device is also designed with a feedback control unit, which is used to optimize the neuromuscular electrical stimulation parameters in real time;

[0089] The feedback control unit includes: a fatigue state detection algorithm and an adaptive stimulation strategy;

[0090] The fatigue state detection algorithm evaluates the muscle fatigue level by real-time monitoring of the average power frequency decline rate of the EMG signal;

[0091] When the average power frequency decline rate ≥ 15% is detected, the adaptive stimulation strategy automatically switches to the intermittent pulse mode, with a stimulation and rest cycle ratio of 1:2, reduces the stimulation intensity to the safety threshold, and simultaneously triggers a visual alarm to prompt the patient to adjust the training posture;

[0092] Furthermore, through the collaborative processing of the lightweight variational autoencoder and the cross-modal temporal attention fusion unit, the dynamic analysis and closed-loop regulation of central-peripheral nerve signals are achieved. The lightweight variational autoencoder first gradually compresses the spatio-temporal dimensions of electroencephalogram (EEG) and electromyogram (EMG) signals through three layers of depthwise separable convolution in the encoder network. During the encoding process, the time-frequency features of EEG signals and the energy distribution of muscle fiber action potentials in EMG signals are extracted as 64-dimensional latent variables, and the KL divergence loss function is used to constrain the latent variable distribution to approximate the standard normal distribution, which not only avoids overfitting but also enhances feature interpretability. The decoder network uses three layers of transposed depthwise separable convolution to fuse low-level features through skip connections, realizing signal reconstruction and retaining biomechanical properties;

[0093] In the cross-modal temporal attention fusion unit, the time alignment layer upsamples the EEG signal from 500 Hz to 1000 Hz through linear interpolation to synchronize the sampling rate with the EMG signal, ensuring time axis alignment. Subsequently, the bidirectional cross-attention mechanism dynamically calculates the complementary weights of EEG and EMG. This mechanism assigns 60% weight to EEG features in the Beta band, strengthening the dominant role of the movement intention. At the same time, the muscle fatigue state of EMG is fed back to the central processing layer through the reverse attention path to form a closed-loop regulation. The fused features are superimposed on the original signals through residual connections, which not only retains modality-specific information, such as the ERD feature of EEG and the time-domain energy distribution of EMG, but also prevents gradient disappearance, and finally outputs a more robust multi-modal representation for movement intention classification and dynamic optimization of electrical stimulation parameters.

[0094] Please refer to Figure 3 and Figure 4 For an embodiment provided by the present invention: An intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus, the therapeutic apparatus further includes an electrical stimulation module, and the electrical stimulation module is used to generate a multi-channel adaptive electrical pulse sequence according to the movement intention classification result, dynamically adjust the stimulation intensity to match the muscle activation threshold, and combine impedance feedback to optimize the current distribution in real time to ensure precise neuromuscular electrical stimulation of the targeted muscle group;

[0095] The electrical stimulation module includes: a multi-channel stimulation electrode array, an adaptive parameter adjustment unit, and a safety protection circuit;

[0096] The multi-channel stimulation electrode array is arranged at the motor points of the patient's targeted muscles;

[0097] The adaptive parameter adjustment unit dynamically adjusts the amplitude, frequency, and duty cycle of the stimulation pulses according to the classification confidence;

[0098] The safety protection circuit monitors the impedance change in real time and triggers overload protection;

[0099] The therapeutic instrument is also designed with a feedback control unit, which is used to optimize neuromuscular electrical stimulation parameters in real time;

[0100] The feedback control unit includes: a fatigue state detection algorithm and an adaptive stimulation strategy;

[0101] The fatigue state detection algorithm evaluates the muscle fatigue level by real-time monitoring of the average power frequency decline rate of the electromyogram signal;

[0102] When the average power frequency decline rate ≥ 15% is detected, the adaptive stimulation strategy automatically switches to the intermittent pulse mode, with a stimulation and rest cycle ratio of 1:2, reduces the stimulation intensity to the safety threshold, and triggers a visual alarm to prompt the patient to adjust the training posture;

[0103] Furthermore, through the closed-loop collaborative mechanism between the electrical stimulation module and the feedback control unit, precise neuromuscular regulation and dynamic optimization are achieved. The electrical stimulation module first generates a multi-channel adaptive electrical pulse sequence according to the result of motion intention classification, such as left and right hand motor imagery or resting state. The multi-channel stimulation electrode array is arranged according to the distribution of the motor points of the targeted muscle. Through the adaptive parameter adjustment unit, the system dynamically adjusts the pulse amplitude, frequency, and duty cycle to ensure that the stimulation intensity precisely matches the muscle activation threshold. For example, when the motion intention classification confidence ≥ 80%, the electrical stimulation intensity increases at a gradient of ±5 mA / second. At the same time, the safety protection circuit monitors the electrode-skin contact impedance in real time to avoid the risk of unbalanced current distribution or local overheating caused by impedance mutation. The feedback control unit analyzes the change in the average power frequency of the surface electromyogram signal in real time through the fatigue state detection algorithm. When the average power frequency decline rate ≥ 15% is detected, the adaptive stimulation strategy is immediately activated, the electrical stimulation mode switches to intermittent pulse, and the stimulation intensity is limited to the safety threshold. At the same time, the LCD interface triggers a yellow visual alarm to prompt the patient to adjust the training posture to avoid muscle over-fatigue. This process combines the impedance feedback optimization algorithm to calibrate the distribution of current in the targeted muscle group in real time, ensuring that the electrical stimulation energy focuses on the target muscle fiber bundle and improving the neuromuscular activation efficiency.

[0104] Please refer to Figure 1 In an embodiment provided by the present invention: an intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic instrument, the therapeutic instrument is also provided with a virtual reality interaction module, which is used to map the result of motion intention classification into limb movements in a virtual scene in real time, and enhance the patient's active participation and training immersion in neuromuscular remodeling through mirror training tasks and visual-proprioceptive feedback loops;

[0105] The virtual reality interaction module has an LCD screen integrated on the outside of the virtual reality glasses unit, which is used to present the rehabilitation training scene, the result of motion intention mapping, electrical stimulation parameters, and alarms;

[0106] The therapeutic device supports the following clinical application modes:

[0107] The active rehabilitation mode triggers synchronous electrical stimulation based on motor imagery classification;

[0108] The passive induction mode induces muscle contraction through a pre-programmed electrical stimulation sequence;

[0109] The hybrid enhancement mode is a two-way closed-loop control that combines electroencephalogram (EEG) intention detection and electromyogram (EMG) feedback;

[0110] Furthermore, the virtual reality interaction module first maps the motor intention classification results output by the hybrid signal processing module, such as the motor imagery probabilities of the left and right hands being ≥ 75%, into three-dimensional limb movements in the virtual scene in real time, and induces the patient to establish a "mind-action-visual feedback" neural remodeling loop through mirror training tasks. In the clinical application mode, the active rehabilitation mode triggers synchronous electrical stimulation based on the EEG motor imagery classification results. When the patient successfully activates the motor characteristics, the electrical stimulation module immediately generates multi-channel adaptive pulses to precisely activate the targeted muscle groups. The passive induction mode induces muscle contraction through a pre-programmed electrical stimulation sequence, such as 3 seconds of stimulation / 5 seconds of intermittent, and is applicable to patients with severe motor dysfunction. Its stimulation parameters are dynamically adjusted according to the EMG signals. The hybrid enhancement mode integrates EEG intention detection and EMG biofeedback. When the patient's active movement intention signal is synchronized with the EMG muscle activation state, the system starts two-way closed-loop control, dynamically matching the electrical stimulation intensity with the movement amplitude in the virtual scene, forming a rehabilitation path of "brain-muscle-virtual environment" collaborative enhancement, and can also unlock new levels through task-oriented design, such as the success rate of virtual object grasping being ≥ 80%, significantly improving the training compliance.

[0111] Please refer to Figure 1 and Figure 2 , an embodiment provided by the present invention: an intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic device. The EEG signal acquisition module can accurately capture the weak electrical signals emitted by the brain, providing a reliable data basis for subsequent signal processing and analysis. The EMG signal acquisition module precisely monitors the electrical activity of the muscles to understand the contraction and relaxation states of the muscles in real time. The two work together to provide strong support for the formulation and implementation of the rehabilitation treatment plan;

[0112] The EEG signal acquisition module is used to collect the electroencephalogram signals of the patient, including a 32-channel electrode array with a sampling frequency of 500 Hz;

[0113] The EMG signal acquisition module is used to collect the surface electromyogram signals of the patient, including a 4-channel electrode array with a sampling frequency of 1000 Hz;

[0114] The preprocessing unit is used to perform 4 - 50Hz band - pass filtering and independent component analysis on the original electroencephalogram (EEG) signals to eliminate electro - oculogram (EOG) artifacts. At the same time, it performs 30 - 150Hz band - pass filtering and adaptive power - frequency noise suppression on the electromyogram (EMG) signals, and adopts a dynamic baseline calibration strategy to eliminate individual impedance differences, ensuring a high signal - to - noise ratio input for multi - modal biological signals;

[0115] The preprocessing unit supports a multi - band parallel processing mode, including:

[0116] Theta band, 4 - 8Hz feature extraction channels;

[0117] Alpha band, 8 - 13Hz event - related desynchronization analysis channels;

[0118] Beta band, 13 - 30Hz movement intention decoding channels;

[0119] Gamma band, 30 - 50Hz muscle co - activation detection channels;

[0120] Furthermore, weak electrical signals of the patient's cerebral cortex are collected through a 32 - channel EEG electrode array, covering the brain regions related to motor imagery. At the same time, the real - time electrical activities of the targeted muscle groups are monitored using a 4 - channel surface EMG electrode array. The preprocessing unit adopts a multi - modal signal sub - band parallel processing strategy. In the Beta band, the movement intention is analyzed through the event - related desynchronization algorithm to trigger brain - controlled electrical stimulation pulses. In the Gamma band, the stimulation parameters are dynamically adjusted in combination with the muscle co - activation matrix to achieve a two - way matching of central intention and peripheral muscle state. The high - signal - to - noise ratio signals after preprocessing are transmitted to the LCD display interface, which highlights the EMG dynamic range and EEG characteristic spectrum in logarithmic coordinates, and synchronously visualizes the electrical stimulation parameters and safety alerts.

[0121] Working principle: First, the therapeutic instrument captures the weak electrical signals related to motor imagery of the patient's cerebral cortex in real - time through the EEG signal acquisition module. The preprocessing unit performs 4 - 50Hz band - pass filtering on the original EEG signals to eliminate EOG artifacts and low - frequency drift, and separates the characteristic frequency bands related to movement intention through independent component analysis;

[0122] Next, the therapeutic instrument synchronously monitors the electrical activities of the targeted muscle groups through the surface EMG acquisition module. After the EMG signals are filtered by 30 - 150Hz band - pass filtering and adaptive power - frequency noise suppression, the root - mean - square of the sliding window is used to quantify the muscle activation intensity. The co - activation detection channels in the Gamma band analyze the temporal coupling pattern of muscle contraction through the covariance matrix and dynamically adjust the electrical stimulation parameters. The electrical stimulation uses multi - channel adaptive pulses to accurately activate the target muscle groups. At the same time, combined with the visual feedback of the virtual reality scene, a closed - loop path of "intention - action - stimulation" is formed;

[0123] Finally, the system integrates EEG and EMG data through a mixed-signal processing module to construct a bidirectional regulation loop between the central and peripheral nerves. For example, when the patient's active movement intention is synchronized with the EMG muscle activation state, the intensity of electrical stimulation is dynamically matched with the amplitude of the action in the virtual scene. The LCD interface optimizes the visualization of EMG signals and EEG features using logarithmic coordinates and displays the electrical stimulation parameters and safety alerts in real time. The virtual reality module enhances proprioceptive feedback through mirror tasks to improve the patient's active participation.

[0124] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced by the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. An intelligent brain-controlled muscle electrical stimulation rehabilitation treatment instrument, characterized in that: It includes an electroencephalogram (EEG) signal acquisition module, an electromyogram (EMG) signal acquisition module, and a hybrid signal processing module. The hybrid signal processing module is connected to the EEG signal acquisition module and the EMG signal acquisition module, and is used for preprocessing EEG and EMG signals, multi-modal feature fusion, and motion intention classification; The hybrid signal processing module includes: a preprocessing unit, a lightweight variational autoencoder, a cross-modal temporal attention fusion unit, and a multi-task learning strategy; The preprocessing unit is used to perform 4 - 50 Hz band-pass filtering, electrooculogram (EOG) artifact removal, and independent component analysis on EEG signals, perform 30 - 150 Hz band-pass filtering and power frequency noise suppression on EMG signals, and adopt a dynamic baseline calibration strategy to eliminate individual impedance differences, ensuring high signal-to-noise ratio input of multi-modal biological signals; The lightweight variational autoencoder extracts the deep spatio-temporal features of EEG and EMG signals through unsupervised learning respectively; The cross-modal temporal attention fusion unit dynamically fuses multi-modal features through a weight assignment mechanism; The multi-task learning strategy synchronously executes motion intention classification, signal reconstruction, and physiological consistency constraint tasks.

2. The intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus according to claim 1, wherein: The EEG signal acquisition module can accurately capture the weak electrical signals emitted by the brain, providing a reliable data basis for subsequent signal processing and analysis. The EMG signal acquisition module accurately monitors the electrical activity of muscles to understand the contraction and relaxation states of muscles in real time. The two work together to provide strong support for the formulation and implementation of rehabilitation treatment plans; The EEG signal acquisition module is used to acquire the EEG signals of patients, including a 32-channel electrode array, and the sampling frequency is 500 Hz; The EMG signal acquisition module is used to acquire the surface EMG signals of patients, including a 4-channel electrode array, and the sampling frequency is 1000 Hz.

3. An intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus according to claim 1, characterized in that: The lightweight variational autoencoder is used to compress the spatio-temporal dimensions of EEG and EMG signals through depthwise separable convolution, construct a low-dimensional feature representation in the latent variable space, and ensure feature interpretability through KL divergence constraint, realizing efficient feature extraction on an embedded platform; The lightweight variational autoencoder includes: an encoder network, a decoder network, and a latent variable space constraint; The encoder network has 3 layers of depthwise separable convolution, the convolution kernel size of each layer is 3×3, and the output channel numbers are 128, 256, and 512 respectively; The decoder network has 3 layers of transposed depthwise separable convolution, and fuses low-level features through skip connections; The latent variable space constraint adopts a KL divergence loss function, and the latent variable dimension is 64.

4. An intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus according to claim 1, characterized in that: The cross-modal temporal attention fusion unit dynamically aligns the spatio-temporal features of EEG and EMG signals through a bidirectional cross-attention mechanism, calculates the complementary weights between brain-EMG signals, and uses residual connections to retain the original modal information, thus solving the problem of asynchronous signal fusion and improving classification robustness; The cross-modal temporal attention fusion unit includes: a time alignment layer, a bidirectional cross-attention calculation, and a residual connection; The time alignment layer performs upsampling the EEG signal to 1000 Hz through linear interpolation to synchronize the sampling rate with the EMG signal; Bidirectional cross-attention calculation: ； Among them, Z eeg is the depth feature matrix of the electroencephalogram signal, with a dimension of R T×d , where T is the time step, d is the feature dimension, and Z emg is the depth feature matrix of the electromyogram signal, with the same dimension as Z eeg , W q , W k and W v are learnable projection weight matrices, with a dimension of R d×d , Q eeg is the query matrix generated by projecting the electroencephalogram features through W q , with a dimension of R T×d , K emg is the key matrix generated by projecting the electromyogram features through W k , with a dimension of R T×d , V emg is the value matrix generated by projecting the electromyogram features through W q , with a dimension of R T×d , K T emg is the transpose matrix of K emg The Attention mechanism is a computational model that mimics the human attention allocation, and the Softmax function is a normalization method that converts any real number vector into a probability distribution; The residual connection adds the fused features and the original features to prevent the vanishing gradient.

5. An intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus according to claim 1, characterized in that: The multi-task learning strategy is used to synchronously optimize the motion intention classification, electroencephalogram (EEG) and electromyogram (EMG) signal reconstruction, and physiological consistency constraint tasks. It realizes the joint modeling of feature space alignment and biomechanical laws through a joint loss function, improving the model's generalization ability to recognize the motion intention of stroke patients. The multi-task learning strategy includes: the main task, the auxiliary tasks, and the joint loss function. The main task is motion intention classification, which outputs the probabilities of left-hand and right-hand motor imagery and the resting state. The loss function is FocalLoss, with γ = 2. The auxiliary tasks include EEG signal reconstruction and physiological consistency constraint: EEG signal reconstruction: The mean square error loss constrains the decoder output. Physiological consistency constraint: Through a pre-trained OpenSim muscle activation model, the cosine similarity between the EMG features and the biomechanical simulation data is restricted to be > 0.

85. The joint loss function is the weighted sum of the motion intention classification loss, the signal reconstruction loss, and the physiological consistency constraint loss. The weight ratio of the three is 5:2:

3. Among them, the classification task has the highest weight to prioritize ensuring the intention recognition accuracy, the physiological consistency constraint comes second to ensure the safety of electrical stimulation, and the signal reconstruction weight is the lowest to prevent the model from overfitting.

6. The intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus according to claim 1, wherein: The therapeutic device also has an electrical stimulation module. The electrical stimulation module is used to generate a multi-channel adaptive electrical pulse sequence according to the motion intention classification result, dynamically adjust the stimulation intensity to match the muscle activation threshold, and combine impedance feedback to optimize the current distribution in real time to ensure precise neuromuscular electrical stimulation of the targeted muscle group. The electrical stimulation module includes: a multi-channel stimulation electrode array, an adaptive parameter adjustment unit, and a safety protection circuit. The multi-channel stimulation electrode array is arranged at the motor points of the patient's targeted muscles. The adaptive parameter adjustment unit dynamically adjusts the amplitude, frequency, and duty cycle of the stimulation pulses according to the classification confidence. The safety protection circuit monitors the impedance change in real time and triggers overload protection.

7. An intelligent brain-controlled muscle electrical stimulation rehabilitation treatment instrument according to claim 1, characterized in that: The therapeutic device also has a virtual reality interaction module. The virtual reality interaction module is used to map the motion intention classification result into the limb movements in the virtual scene in real time, and through the mirror training task and the visual-proprioceptive feedback loop, enhance the patient's active participation and training immersion in neural remodeling. The virtual reality interaction module has an LCD screen integrated on the outer side of the virtual reality glasses unit, which is used to present the rehabilitation training scene, the motion intention mapping result, the electrical stimulation parameters, and the alarm.

8. An intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus according to claim 1, characterized in that: The preprocessing unit is used to perform 4 - 50Hz band-pass filtering and independent component analysis on the original EEG signal to eliminate electrooculogram artifacts, and at the same time perform 30 - 150Hz band-pass filtering and adaptive power frequency noise suppression on the EMG signal, and adopt a dynamic baseline calibration strategy to eliminate individual impedance differences, ensuring a high signal-to-noise ratio input of multi-modal biological signals. The preprocessing unit supports a multi-band parallel processing mode, including: Theta band, 4 - 8Hz feature extraction channels; Alpha band, 8 - 13Hz event-related desynchronization analysis channels; Beta band, 13 - 30Hz motion intention decoding channels; Gamma band, 30 - 50Hz muscle co-activation detection channels.

9. An intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus according to claim 1, characterized in that: A feedback control unit is also designed inside the therapeutic instrument, and the feedback control unit is used to optimize the neuromuscular electrical stimulation parameters in real time; The feedback control unit includes: a fatigue state detection algorithm and an adaptive stimulation strategy; The fatigue state detection algorithm evaluates the muscle fatigue level by monitoring the average power frequency decline rate of the electromyogram signal in real time; When the MPF decline rate ≥ 15% is detected, the adaptive stimulation strategy automatically switches to the intermittent pulse mode, the stimulation and rest cycle ratio is 1:2, and the stimulation intensity is reduced to the safety threshold, while triggering a visual alarm to prompt the patient to adjust the training posture.

10. The intelligent brain-controlled muscle electrical stimulation rehabilitation therapeutic apparatus according to claim 1, wherein: The therapeutic instrument supports the following clinical application modes: The active rehabilitation mode triggers synchronous electrical stimulation based on motor imagery classification; The passive induction mode induces muscle contraction through a pre-programmed electrical stimulation sequence; The hybrid enhancement mode is a two-way closed-loop control that combines electroencephalogram intention detection and electromyogram feedback.

Images