Brain-controlled rehabilitation robot real-time interaction method and system based on brain-computer interface

Abstract

This invention provides a real-time interactive method and system for a brain-controlled rehabilitation robot based on a brain-computer interface. It synchronously acquires the user's electroencephalogram (EEG) signals and records the steady-state visual evoked potential (SSVEP) stimulation frequencies associated with different rehabilitation tasks. The system then processes the EEG signals in parallel to extract features: extracting motor imagery-related frequency band signals, obtaining tangent space feature vectors by calculating the Riemann covariance matrix, and calculating the spectral Shannon entropy; calculating response features for each SSVEP stimulation frequency; constructing a composite feature vector, inputting the vector into an intention classification model, and outputting preliminary decoding results including candidate motor intentions and confidence levels; setting a rejection threshold; if the highest confidence level is below the threshold, it is determined as an invalid instruction and feature extraction is repeated; if it is above the threshold, the corresponding candidate intention is determined as a control instruction; inputting the control instruction confidence level, SSVEP response features, and spectral Shannon entropy into a dynamic mapping model to calculate motion parameters and drive the rehabilitation robot to perform actions.

Description

Technical Field

[0001] This application belongs to the field of real-time interaction, and in particular relates to a real-time interaction method and system for a brain-controlled rehabilitation robot based on a brain-computer interface. Background Technology

[0002] Brain-computer interface (BCI) technology, particularly rehabilitation robots based on BCIs, can assist patients in active rehabilitation training based on their motor intentions. Non-invasive BCI paradigms include motor imagery and steady-state visual evoked potentials (SSVEP). Motor imagery controls robot movement by decoding the rhythmic electrical activity changes in the cerebral cortex when the user imagines specific limb movements. However, motor imagery EEG signals have a low signal-to-noise ratio, significant individual variability, and classification accuracy is easily affected by user fatigue, lack of concentration, and other factors, leading to low stability and reliability.

[0003] SSVEP (Simultaneous Visual Stabilization Electronic Response) induces EEG responses at the same and harmonic frequencies as the stimulation by having the user focus on flashing visual stimuli of different frequencies, thus recognizing the user's selection intent. Hybrid MI-SSVEP uses SSVEP as an auxiliary channel for switching tasks or selecting targets, while motor imagery is used to trigger specific actions. However, at the feature fusion level, simple feature splicing or linear weighting fails to adjust the contribution of different paradigm features according to the user's real-time cognitive state. The lack of a command rejection mechanism easily leads to misinterpreting the user's unintentional state or vague intentions as commands, causing the robot to perform unexpected actions. Brain-controlled commands are usually discrete, making it difficult to regulate the movement process of rehabilitation robots, resulting in a less smooth and natural human-computer interaction experience. Therefore, how to plan a method that can integrate multimodal EEG information, intelligently identify and filter invalid commands, and achieve real-time interaction for robot control is a pressing technical challenge in the field of brain-controlled rehabilitation robots. Summary of the Invention

[0004] This invention proposes a real-time interaction method for brain-controlled rehabilitation robots based on brain-computer interfaces, addressing the problem that existing technologies fail to adjust the contribution of different paradigm features according to the user's real-time cognitive state, easily misinterpreting the user's unintentional state or vague intentions as commands. The method includes: The system synchronously acquires the user's EEG signals and records the steady-state visual evoked potential stimulation frequencies of multiple visual stimuli associated with different rehabilitation tasks. The system then processes the EEG signals in parallel to extract features: on one hand, it extracts signals from the motor imagery-related frequency bands, obtains the signal tangent space feature vector by calculating the Riemann covariance matrix, and calculates the spectral Shannon entropy of the frequency bands; on the other hand, it calculates the corresponding response features representing the user's attention selection for each steady-state visual evoked potential stimulation frequency. The tangent space feature vector of the motion imagery is weighted using the spectral Shannon entropy and fused with the steady-state visual evoked potential response features to construct a composite feature vector that integrates motion imagery and attention information. The composite feature vector is then input into a preset intention classification model to output a preliminary decoding result containing multiple candidate motion intentions and their respective confidence levels. A rejection threshold is jointly set based on the discrete coefficients of the decoding result sequence within the past time window and the Shannon entropy of the spectrum; if the highest confidence level in the preliminary decoding result is lower than the rejection threshold, it is determined to be an invalid instruction and feature extraction is performed on the subsequent EEG signals. If the highest confidence level is not lower than the rejection threshold, the corresponding candidate motion intention is determined as a control command; and the confidence level of the control command, the steady-state visual evoked potential response characteristics corresponding to the control command, and the spectral Shannon entropy are jointly input into a preset dynamic mapping model to calculate the motion parameters of the control rehabilitation robot and drive the robot to perform corresponding rehabilitation actions.

[0005] Furthermore, this invention also relates to a real-time interactive system for a brain-controlled rehabilitation robot based on a brain-computer interface, comprising the following modules: The calculation module is used to synchronously acquire the user's EEG signals and record the steady-state visual evoked potential stimulation frequencies of multiple visual stimuli associated with different rehabilitation tasks; it processes the EEG signals in parallel to extract features: on the one hand, it extracts signals from the frequency bands related to motor imagery, obtains the tangent space feature vector of the signal by calculating the Riemann covariance matrix, and calculates the spectral Shannon entropy of the frequency band; on the other hand, it calculates the corresponding response features representing the user's attention selection for each of the steady-state visual evoked potential stimulation frequencies. The output module is used to weight the tangent space feature vector of the motion imagery using the spectral Shannon entropy, and fuse it with the steady-state visual evoked potential response features to construct a composite feature vector that integrates motion imagery and attention information; the composite feature vector is input into a preset intention classification model, and the output is a preliminary decoding result containing multiple candidate motion intentions and their respective confidence levels; The extraction module is used to jointly set a rejection threshold based on the discrete coefficients of the decoding result sequence within the past time window and the Shannon entropy of the spectrum; if the highest confidence level in the preliminary decoding result is lower than the rejection threshold, it is determined to be an invalid instruction and feature extraction is performed on the subsequent EEG signals. The driving module is used to determine the corresponding candidate motion intention as a control command if the highest confidence level is not lower than the rejection threshold; and to input the confidence level of the control command, the steady-state visual evoked potential response characteristics corresponding to the control command, and the spectral Shannon entropy into a preset dynamic mapping model to calculate the motion parameters of the control rehabilitation robot and drive the robot to perform corresponding rehabilitation actions.

[0006] This invention constructs a rich composite feature vector by using the spectral Shannon entropy, which reflects the user's cognitive state, to weight motor imagery features and deeply fusing it with steady-state visual evoked potential features. A command rejection mechanism filters out false commands generated by the user in an unintentional or ambiguous state, preventing accidental triggering of the rehabilitation robot. Through a dynamic mapping model, the confirmed command confidence, attention-related features, and cognitive state information are combined to transform the robot's specific motion parameters, enabling continuous adjustment of the rehabilitation robot's speed and amplitude details. This results in a smooth and natural human-computer interaction process, improving the experience and effectiveness of rehabilitation training. Attached Figure Description

[0007] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of feature weighting and fusion. Figure 3 This is a diagram illustrating intent classification. Figure 4 This is a schematic diagram of dynamic mapping. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0009] The term "multiple" in this application refers to two or more. Furthermore, it should be understood that the terms "first," "second," etc., used in the description of this application are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order.

[0010] In the first embodiment, the present invention proposes a real-time interaction method for a brain-controlled rehabilitation robot based on a brain-computer interface, such as... Figure 1 ,include: S1. Simultaneously acquire the user's EEG signals and record the steady-state visual evoked potential stimulation frequencies of multiple visual stimuli associated with different rehabilitation tasks; process the EEG signals in parallel to extract features: on the one hand, extract signals from the frequency bands related to motor imagery, obtain the tangent space feature vector of the signal by calculating the Riemann covariance matrix, and calculate the spectral Shannon entropy of the frequency band; on the other hand, for each steady-state visual evoked potential stimulation frequency, calculate the corresponding response features representing the user's attention selection. Brain-computer interfaces (BCIs) include non-invasive, invasive, and semi-invasive types. This invention preferably uses a non-invasive approach; however, those skilled in the art will know that invasive and semi-invasive methods can also be applied, differing only in the method of acquiring EEG signals. An EEG cap with 64 leads is used, positioned according to the international 10-20 standard, to acquire whole-brain EEG signals at a sampling rate of 1000Hz. Notch filtering at 50Hz and bandpass filtering from 0.5Hz to 40Hz are then applied to remove power line interference and baseline drift. Simultaneously, multiple rehabilitation task options are presented to the user on the display screen, such as left-hand grasping, right-hand extension, and rest. Each option corresponds to a graphic block that flashes at a specific frequency; for example, left-hand grasping corresponds to 8Hz, and right-hand extension corresponds to 10Hz. The correspondence between these frequencies and tasks is pre-stored.

[0011] For the motor imagery component, the preprocessed EEG signal is passed through an 8-30Hz bandpass filter to extract the motion-related μ and β rhythms. For a signal within a 2-second time window, the multi-channel covariance matrix of the signal is calculated. This covariance matrix is ​​projected onto the tangent space of a Riemannian manifold with the geometric mean of the covariance matrices of all training samples as the reference point, thus obtaining a tangent space feature vector in Euclidean space. In another embodiment, for the multi-channel EEG signal after filtering the motor imagery-related frequency bands, the covariance matrix between each channel is first calculated within a fixed-length time window, and this covariance matrix is ​​considered as a point on a symmetric positive definite matrix manifold. A reference point composed of the covariance matrices of the training data's mean is selected, and the covariance matrix to be processed is mapped to the tangent space of the reference point using a Riemannian logarithm, obtaining a symmetric matrix form of the tangent space matrix. The upper triangular elements of this matrix are flattened row-wise into a one-dimensional vector, thus obtaining the tangent space feature vector used to characterize the motor imagery characteristics of that time window.

[0012] Simultaneously, a Fast Fourier Transform (FFT) is performed on the signal in the 8-30Hz frequency band to obtain the power spectral density. After normalization, the spectral Shannon entropy is calculated, and this entropy value is used to represent the signal complexity. For the steady-state visual evoked potentials, canonical correlation analysis is used to calculate the correlation between the EEG signal and preset stimulus frequencies, such as 8Hz, 10Hz, and harmonic sine and cosine reference signals. The maximum correlation coefficient value corresponding to each stimulus frequency is obtained, and this value is the response characteristic, the magnitude of which reflects the user's attention to the corresponding visual stimulus.

[0013] In an optional embodiment, calculating the spectral Shannon entropy of the frequency band includes: The acquired EEG signals are bandpass filtered from 8 to 30 Hz to extract the μ and β rhythm signals related to motor imagery. The power spectral density is calculated by applying Fast Fourier Transform to the signals. The power spectral density is normalized and used as a probability distribution, and then the Shannon entropy of the spectrum is calculated.

[0014] Specifically, a raw EEG signal data stream is acquired from an EEG acquisition device. A digital bandpass filter is applied to this data segment, with the passband frequency range set to 8Hz to 30Hz. Low-frequency noise and high-frequency interference unrelated to the motor imagery task are filtered out, retaining only signal components containing μ rhythms (8Hz to 12Hz) and β rhythms (13Hz to 30Hz), the energy changes of which are closely related to the motor imagery activity.

[0015] The filtered signal segment is input into the Fast Fourier Transform (FFT) algorithm module. This module calculates the power spectral density of the signal in the frequency range of 8Hz to 30Hz, obtaining a series of values, each representing the signal power at a specific frequency point. For example, the power is 5 at 10Hz, 8 at 11Hz, and so on. The power spectral density values ​​of all frequency points are summed to obtain a total. The power value of each frequency point is divided by this total to normalize the power spectral density data so that the sum is 1, forming a probability distribution p(i), where p(i) represents the probability that the energy appears at the i-th frequency point. This probability distribution is then substituted into the Shannon entropy calculation formula to calculate the spectral Shannon entropy value H. A lower entropy value indicates that the energy is concentrated at a few frequency points, reflecting a more focused user attention; conversely, a higher entropy value indicates that the energy is widely distributed, and the user attention is scattered.

[0016] In an optional embodiment, the calculation of the corresponding response features represents the user's attention selection, including: Using task-related component analysis, a spatial filtering template is constructed for each steady-state visual evoked potential stimulation frequency by averaging historical training data corresponding to the frequency. The real-time acquired EEG signal is filtered through the spatial filtering template, and the Pearson correlation coefficient between the filtered signal and the sinusoidal reference signal of the corresponding stimulation frequency is calculated. The coefficient value is used as the response feature.

[0017] During the calibration phase, multiple visual stimuli are presented to the user, flashing at different frequencies; for example, target A flashes at 10Hz, and target B flashes at 12Hz. While the user focuses on target A, multiple data segments of multi-lead EEG signals are recorded. The data segments recorded while focusing on target A are averaged to obtain a multi-channel weight vector for the 10Hz stimulus; this vector serves as the spatial filtering template for the 10Hz stimulus. Similarly, a dedicated spatial filtering template is generated for the 12Hz stimulus of target B. These templates detect the spatial distribution of the strongest response pattern of the brain to stimuli at specific frequencies on the scalp.

[0018] To determine which target the user is currently focusing on, the newly acquired EEG signal is filtered using pre-constructed 10Hz and 12Hz spatial filtering templates. For example, when using the 10Hz template, each channel of the new signal is multiplied by its corresponding weight in the template and then summed to obtain a one-dimensional time-series signal that enhances the steady-state visual evoked potential component at 10Hz. A standard 10Hz sine and cosine wave are generated as reference signals. The Pearson correlation coefficient between the 10Hz template-filtered signal and the two reference signals is calculated. The coefficient value, for example, 0.85, indicates the degree of synchronization between the current EEG signal and the 10Hz stimulus. The same operation is performed on the 12Hz template to obtain another correlation coefficient value, for example, 0.2. These two values, 0.85 and 0.2, constitute the current response characteristics, indicating that the user may be focusing on target A.

[0019] S2, the tangent space feature vector of the motion imagery is weighted using the spectral Shannon entropy and fused with the steady-state visual evoked potential response features to construct a composite feature vector that integrates motion imagery and attention information; the composite feature vector is input into a preset intention classification model to output a preliminary decoding result containing multiple candidate motion intentions and their respective confidence levels; The calculated spectral Shannon entropy is multiplied by each element of the tangent space feature vector of motor imagery to scale the vector. If a user has a high spectral Shannon entropy, indicating potential fatigue or inattention, this weighting operation reduces the weight of motor imagery features in subsequent classification. The weighted tangent space feature vector is then concatenated with a vector composed of all steady-state visual evoked potential response features to form a high-dimensional, information-rich composite feature vector, such as... Figure 2 .

[0020] A pre-trained Support Vector Machine (SVM) classifier based on historical data is used as the intent classification model. The generated composite feature vector is used as the input to this SVM. The model's output is a probability vector. For example, for the three tasks of left hand, right hand, and rest, the output might be 0.8, 0.1, and 0.1, indicating that the model determines the user's intent as having an 80% confidence level for left-hand grasping, a 10% confidence level for right-hand extension, and a 10% confidence level for rest. The left-hand grasping intent is the candidate motion intent corresponding to the highest confidence level. Figure 3 .

[0021] In an optional embodiment, the step of weighting the tangent space feature vector of the motion imagery using the spectral Shannon entropy and fusing it with the steady-state visual evoked potential response features includes: The Shannon entropy values ​​of the spectrum are normalized to the range of 0-1 to obtain the normalized Shannon entropy of the spectrum. Weighting coefficients Multiply each element of the tangent space feature vector of the motion imagination to obtain a weighted feature vector; concatenate the weighted feature vector with the steady-state visual evoked potential response feature vector to form the composite feature vector.

[0022] Specifically, obtain the calculated spectral Shannon entropy value, assuming it to be 1.8. Normalize it based on the historical range of this user's Shannon entropy value recorded during the calibration phase or long-term operation, for example, a minimum of 1.0 and a maximum of 2.0. In this example, the normalized spectral Shannon entropy... That is, 0.8. The value of 0.8 indicates that the user's current attention is relatively unfocused.

[0023] The weighting coefficient w is calculated using the formula, which is 0.2. This weighting coefficient reflects the level of trust in the motor imagery features; the less focused the attention, the lower the level of trust. Assume the extracted motor imagery tangent space feature vector is a vector containing 10 elements, for example, [1.2, -0.5, ..., 2.1]. Multiply each element of this vector by the weighting coefficient 0.2 to obtain the weighted feature vector [0.24, -0.1, ..., 0.42]. Obtain the steady-state visual evoked potential response feature vector at this moment, assumed to be [0.85, 0.2]. Concatenate the weighted motor imagery feature vector and the steady-state visual evoked potential response feature vector sequentially to form a composite feature vector [0.24, -0.1, ..., 0.42, 0.85, 0.2]. This composite vector contains information about both the action the user wants to perform and the target the user has selected, with the former adjusted according to the user's attention level.

[0024] In an optional embodiment, inputting the composite feature vector into a preset intent classification model includes: The composite feature vector is input into a support vector machine classifier using radial basis function kernels, and a one-to-many strategy is used to output the classification confidence of each candidate motion intention.

[0025] Assume three intentions are supported: controlling the robot arm to move left, controlling the arm to move right, and keeping it stationary. Three independent classifiers are trained using a one-to-many strategy: Classifier 1 determines whether the intention is left, Classifier 2 determines whether it is right, and Classifier 3 determines whether it is stationary. All classifiers use radial basis function kernels to handle separable data. The training process is performed using composite feature vector data with known intention labels.

[0026] The real-time generated composite feature vector, such as [0.24, -0.1, ..., 0.42, 0.85, 0.2], is simultaneously input into the three pre-trained support vector machine classifiers. Each classifier calculates the distance from the input vector to the vector decision boundary; this distance reflects the probability that the input sample belongs to the class represented by that classifier. To obtain an intuitive confidence level, the distance value is typically mapped to a value between 0 and 1 using a probability transformation function, such as Platt scaling or the Sigmoid function. For example, the output of classifier 1 might be 0.8, indicating an 80% confidence level for a leftward intention; the output of classifier 2 might be 0.15, indicating a 15% confidence level for a rightward intention; and the output of classifier 3 might be 0.05, indicating a 5% confidence level for remaining still. This set of confidence scores [0.8, 0.15, 0.05] represents the output of the classification model. The intention with the highest confidence level, i.e., leftward, is selected as the current decoding result.

[0027] S3, a rejection threshold is jointly set based on the discrete coefficients of the decoding result sequence within the past time window and the Shannon entropy of the spectrum; if the highest confidence level in the preliminary decoding result is lower than the rejection threshold, it is determined to be an invalid instruction and feature extraction is performed on the subsequent EEG signals. Maintain a queue containing the highest confidence scores from the past five consecutive decoding results. Calculate the ratio of the standard deviation to the mean of the five confidence values ​​in this queue, i.e., the coefficient of variation. Simultaneously, obtain the spectral Shannon entropy at the current moment. The rejection threshold is determined by subtracting a term proportional to the spectral Shannon entropy and an inversely proportional term to the coefficient of variation from a base threshold, such as 0.7. When user inattention leads to an increase in entropy, or when unstable decoding results lead to an increase in the coefficient of variation, the threshold is lowered accordingly, making the acceptance criteria for instructions more stringent. Compare the highest confidence score in the current preliminary decoding result, such as 0.8, with the calculated rejection threshold; if it is lower than the threshold, ignore this result.

[0028] In an optional embodiment, setting the rejection threshold based on the discrete coefficients of the decoding result sequence within the past time window and the spectral Shannon entropy includes: Obtain the sequence of highest confidence scores within the past N decoding time windows; calculate the standard deviation and mean of the sequence, and use the ratio of the two as the coefficient of variation. The formula for calculating the rejection threshold T is as follows: ,in is the normalized Shannon entropy of the spectrum, and a, b, and c are constant weights preset based on experimental data.

[0029] Specifically, a first-in, first-out (FIFO) data queue is maintained to store the highest confidence scores output by the classifier in the most recent N decoding cycles. Assuming N is set to 10, at the current moment, this queue stores the highest confidence scores from the past 10 decoding cycles, for example, the sequence [0.8, 0.75, 0.82, 0.78, 0.6, 0.55, 0.79, 0.81, 0.77, 0.83]. The mean of this sequence is approximately 0.75, and the standard deviation is approximately 0.09. Dividing the standard deviation by the mean yields the coefficients of variation. The value is 0.12. This value reflects the stability of the decoding result; the smaller the value, the more stable and consistent the result.

[0030] Assume that, according to the pre-set experimental parameters, the constant weights a is 2.0, b is 0.5, and c is 0.3. Meanwhile, the current normalized spectral Shannon entropy is known from previous calculations. The value is 0.8. Substituting this value into the formula, the threshold T is 0.94. This threshold is variable; if the decoding result is unstable... Large or user attention is not focused The higher the confidence level, the higher the threshold. The highest confidence level obtained in the current decoding cycle, for example, 0.8, is compared with the calculated threshold of 0.94. Because 0.8 is less than 0.94, the decoding result will be rejected, and no control command will be generated, thus avoiding erroneous operations when the user is in a poor state or the intent is unclear.

[0031] S4. If the highest confidence level is not lower than the rejection threshold, the corresponding candidate motion intention is determined as a control command; and the confidence level of the control command, the steady-state visual evoked potential response characteristics corresponding to the control command, and the spectral Shannon entropy are jointly input into a preset dynamic mapping model to calculate the motion parameters of the control rehabilitation robot and drive the robot to perform corresponding rehabilitation actions.

[0032] If the highest confidence level of 0.8 is higher than or equal to the rejection threshold, the left-hand grasp is confirmed as a command. The confidence level of this command (0.8), the corresponding steady-state visual evoked potential response characteristics (i.e., the canonical correlation coefficient value under 8Hz stimulation), and the current spectral Shannon entropy are input into a pre-trained multiple linear regression model. Based on these three inputs, the model outputs a set of continuous motion parameters, such as a target grasping angle of 90° and an angular velocity of 30 degrees per second for the robotic hand. These parameters are sent to the underlying controller of the rehabilitation robot to drive the robotic hand to complete the grasping action at the calculated speed. In another embodiment, the dynamic mapping model is an MLP with 2–3 hidden layers (e.g., 128→64→32 units), ReLU activation, BatchNorm, and Dropout 0.2, using a composite feature vector as input, and outputting the mean angular velocity μ and logarithmic variance s of each robot joint.

[0033] In an optional embodiment, calculating the motion parameters for controlling the rehabilitation robot includes: The confidence level of the control command, the corresponding steady-state visual evoked potential response feature value, and the normalized spectral Shannon entropy are used as input layer data and input to a multilayer perceptron neural network with at least one hidden layer. The target motion angular velocity and motion duration of the rehabilitation robot joint are output by the network output layer as motion parameters.

[0034] Once a control command passes the rejection threshold, an input vector is prepared to be fed into a pre-trained multilayer perceptron neural network. The input vector consists of three values: the confidence level of the command, for example, 0.95; the steady-state visual evoked potential response characteristic associated with the command, for example, the response characteristic of the user's gaze on target A is 0.85; and the current normalized spectral Shannon entropy. For example, 0.2. Therefore, the vector input to the neural network is [0.95, 0.85, 0.2].

[0035] The network contains at least one hidden layer, such as one with eight neurons, and an output layer. Input data propagates forward through the network, is weighted and summed by the connection weights between neurons in each layer, and is then transformed by activation functions such as ReLU or the sigmoid function. This process allows the neural network to map the input EEG decoding information into specific robot movement commands based on patterns learned during training.

[0036] Assume the output layer has two neurons. The first neuron outputs the target motion angular velocity, and the second neuron outputs the motion duration. After network calculation, the value of the first output neuron might be 20.5, representing that a certain joint of the robot should move at an angular velocity of 20.5 degrees per second; the value of the second output neuron might be 1.5, representing that the motion should last for 1.5 seconds. These two values, 20.5 and 1.5, are the generated motion parameters, which are sent to the underlying motion controller of the rehabilitation robot to execute the corresponding actions, such as... Figure 4 The intensity of the user's intent, the clarity of their attention target, and the degree of their concentration can collectively regulate the speed and amplitude of robot-assisted movement.

[0037] In the second embodiment, the present invention also proposes a real-time interactive system for a brain-controlled rehabilitation robot based on a brain-computer interface, comprising the following modules: The calculation module is used to synchronously acquire the user's EEG signals and record the steady-state visual evoked potential stimulation frequencies of multiple visual stimuli associated with different rehabilitation tasks; it processes the EEG signals in parallel to extract features: on the one hand, it extracts signals from the frequency bands related to motor imagery, obtains the tangent space feature vector of the signal by calculating the Riemann covariance matrix, and calculates the spectral Shannon entropy of the frequency band; on the other hand, it calculates the corresponding response features representing the user's attention selection for each of the steady-state visual evoked potential stimulation frequencies. The output module is used to weight the tangent space feature vector of the motion imagery using the spectral Shannon entropy, and fuse it with the steady-state visual evoked potential response features to construct a composite feature vector that integrates motion imagery and attention information; the composite feature vector is input into a preset intention classification model, and the output is a preliminary decoding result containing multiple candidate motion intentions and their respective confidence levels; The extraction module is used to jointly set a rejection threshold based on the discrete coefficients of the decoding result sequence within the past time window and the Shannon entropy of the spectrum; if the highest confidence level in the preliminary decoding result is lower than the rejection threshold, it is determined to be an invalid instruction and feature extraction is performed on the subsequent EEG signals. The driving module is used to determine the corresponding candidate motion intention as a control command if the highest confidence level is not lower than the rejection threshold; and to input the confidence level of the control command, the steady-state visual evoked potential response characteristics corresponding to the control command, and the spectral Shannon entropy into a preset dynamic mapping model to calculate the motion parameters of the control rehabilitation robot and drive the robot to perform corresponding rehabilitation actions.

[0038] In an optional embodiment, calculating the spectral Shannon entropy of the frequency band includes: The acquired EEG signals are bandpass filtered from 8 to 30 Hz to extract the μ and β rhythm signals related to motor imagery. The power spectral density is calculated by applying Fast Fourier Transform to the signals. The power spectral density is normalized and used as a probability distribution, and then the Shannon entropy of the spectrum is calculated.

[0039] In an optional embodiment, the calculation of the corresponding response features represents the user's attention selection, including: Using task-related component analysis, a spatial filtering template is constructed for each steady-state visual evoked potential stimulation frequency by averaging historical training data corresponding to the frequency. The real-time acquired EEG signal is filtered through the spatial filtering template, and the Pearson correlation coefficient between the filtered signal and the sinusoidal reference signal of the corresponding stimulation frequency is calculated. The coefficient value is used as the response feature.

[0040] In an optional embodiment, the step of weighting the tangent space feature vector of the motion imagery using the spectral Shannon entropy and fusing it with the steady-state visual evoked potential response features includes: The Shannon entropy values ​​of the spectrum are normalized to the range of 0-1 to obtain the normalized Shannon entropy of the spectrum. Weighting coefficients Multiply each element of the tangent space feature vector of the motion imagination to obtain a weighted feature vector; concatenate the weighted feature vector with the steady-state visual evoked potential response feature vector to form the composite feature vector.

[0041] In an optional embodiment, inputting the composite feature vector into a preset intent classification model includes: The composite feature vector is input into a support vector machine classifier using radial basis function kernels, and a one-to-many strategy is used to output the classification confidence of each candidate motion intention.

[0042] In an optional embodiment, setting the rejection threshold based on the discrete coefficients of the decoding result sequence within the past time window and the spectral Shannon entropy includes: Obtain the sequence of highest confidence scores within the past N decoding time windows; calculate the standard deviation and mean of the sequence, and use the ratio of the two as the coefficient of variation. The formula for calculating the rejection threshold T is as follows: ,in is the normalized Shannon entropy of the spectrum, and a, b, and c are constant weights preset based on experimental data.

[0043] In an optional embodiment, calculating the motion parameters for controlling the rehabilitation robot includes: The confidence level of the control command, the corresponding steady-state visual evoked potential response feature value, and the normalized spectral Shannon entropy are used as input layer data and input to a multilayer perceptron neural network with at least one hidden layer. The target motion angular velocity and motion duration of the rehabilitation robot joint are output by the network output layer as motion parameters.

[0044] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0045] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0046] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A real-time interaction method for a brain-controlled rehabilitation robot based on a brain-computer interface, characterized in that, Includes the following steps: The system synchronously acquires the user's EEG signals and records the steady-state visual evoked potential stimulation frequencies of multiple visual stimuli associated with different rehabilitation tasks. The system then processes the EEG signals in parallel to extract features: on one hand, it extracts signals from the motor imagery-related frequency bands, obtains the signal tangent space feature vector by calculating the Riemann covariance matrix, and calculates the spectral Shannon entropy of the frequency bands; on the other hand, it calculates the corresponding response features representing the user's attention selection for each steady-state visual evoked potential stimulation frequency. The tangent space feature vector of the motion imagery is weighted using the spectral Shannon entropy and fused with the steady-state visual evoked potential response features to construct a composite feature vector that integrates motion imagery and attention information. The composite feature vector is then input into a preset intention classification model to output a preliminary decoding result containing multiple candidate motion intentions and their respective confidence levels. A rejection threshold is jointly set based on the discrete coefficients of the decoding result sequence within the past time window and the Shannon entropy of the spectrum; if the highest confidence level in the preliminary decoding result is lower than the rejection threshold, it is determined to be an invalid instruction and feature extraction is performed on the subsequent EEG signals. If the highest confidence level is not lower than the rejection threshold, the corresponding candidate motion intention is determined as a control command; and the confidence level of the control command, the steady-state visual evoked potential response characteristics corresponding to the control command, and the spectral Shannon entropy are jointly input into a preset dynamic mapping model to calculate the motion parameters of the control rehabilitation robot and drive the robot to perform corresponding rehabilitation actions.

2. The method according to claim 1, characterized in that, The calculation of the spectral Shannon entropy of the frequency band includes: The acquired EEG signals are bandpass filtered from 8 to 30 Hz to extract the μ and β rhythm signals related to motor imagery. The power spectral density is calculated by applying Fast Fourier Transform to the signals. The power spectral density is normalized and used as a probability distribution, and then the Shannon entropy of the spectrum is calculated.

3. The method according to claim 1, characterized in that, The response features corresponding to the calculation represent the user's attention selection, including: Using task-related component analysis, a spatial filtering template is constructed for each steady-state visual evoked potential stimulation frequency by averaging historical training data corresponding to the frequency. The real-time acquired EEG signal is filtered through the spatial filtering template, and the Pearson correlation coefficient between the filtered signal and the sinusoidal reference signal of the corresponding stimulation frequency is calculated. The coefficient value is used as the response feature.

4. The method according to claim 1, characterized in that, The step of weighting the tangent space feature vector of the motion imagery using the spectral Shannon entropy and fusing it with the steady-state visual evoked potential response features includes: The Shannon entropy values ​​of the spectrum are normalized to the range of 0-1 to obtain the normalized Shannon entropy of the spectrum. Weighting coefficients Multiply each element of the tangent space feature vector of the motion imagination to obtain a weighted feature vector; concatenate the weighted feature vector with the steady-state visual evoked potential response feature vector to form the composite feature vector.

5. The method according to claim 1, characterized in that, The step of inputting the composite feature vector into a preset intent classification model includes: The composite feature vector is input into a support vector machine classifier using radial basis function kernels, and a one-to-many strategy is used to output the classification confidence of each candidate motion intention.

6. The method according to claim 1, characterized in that, The step of setting a rejection threshold based on the discrete coefficients of the decoding result sequence within the past time window and the spectral Shannon entropy includes: Obtain the sequence of highest confidence scores within the past N decoding time windows; calculate the standard deviation and mean of the sequence, and use the ratio of the two as the coefficient of variation. The formula for calculating the rejection threshold T is as follows: ,in is the normalized Shannon entropy of the spectrum, and a, b, and c are constant weights preset based on experimental data.

7. The method according to claim 1, characterized in that, The calculation of motion parameters for controlling the rehabilitation robot includes: The confidence level of the control command, the corresponding steady-state visual evoked potential response feature value, and the normalized spectral Shannon entropy are used as input layer data and input to a multilayer perceptron neural network with at least one hidden layer. The target motion angular velocity and motion duration of the rehabilitation robot joint are output by the network output layer as motion parameters.

8. A real-time interactive system for a brain-controlled rehabilitation robot based on a brain-computer interface, characterized in that, Includes the following modules: The calculation module is used to synchronously acquire the user's EEG signals and record the steady-state visual evoked potential stimulation frequencies of multiple visual stimuli associated with different rehabilitation tasks; it processes the EEG signals in parallel to extract features: on the one hand, it extracts signals from the frequency bands related to motor imagery, obtains the tangent space feature vector of the signal by calculating the Riemann covariance matrix, and calculates the spectral Shannon entropy of the frequency band; on the other hand, it calculates the corresponding response features representing the user's attention selection for each of the steady-state visual evoked potential stimulation frequencies. The output module is used to weight the tangent space feature vector of the motion imagery using the spectral Shannon entropy, and fuse it with the steady-state visual evoked potential response features to construct a composite feature vector that integrates motion imagery and attention information; the composite feature vector is input into a preset intention classification model, and the output is a preliminary decoding result containing multiple candidate motion intentions and their respective confidence levels; The extraction module is used to jointly set a rejection threshold based on the discrete coefficients of the decoding result sequence within the past time window and the Shannon entropy of the spectrum; If the highest confidence level in the preliminary decoding result is lower than the rejection threshold, it is determined to be an invalid instruction and feature extraction is repeated on subsequent EEG signals; The driving module is used to determine the corresponding candidate motion intention as a control command if the highest confidence level is not lower than the rejection threshold; and to input the confidence level of the control command, the steady-state visual evoked potential response characteristics corresponding to the control command, and the spectral Shannon entropy into a preset dynamic mapping model to calculate the motion parameters of the control rehabilitation robot and drive the robot to perform corresponding rehabilitation actions.

9. The system according to claim 8, characterized in that, The calculation of the spectral Shannon entropy of the frequency band includes: The acquired EEG signals are bandpass filtered from 8 to 30 Hz to extract the μ and β rhythm signals related to motor imagery. The power spectral density is calculated by applying Fast Fourier Transform to the signals. The power spectral density is normalized and used as a probability distribution, and then the Shannon entropy of the spectrum is calculated.

10. The system according to claim 8, characterized in that, The response features corresponding to the calculation represent the user's attention selection, including: Using task-related component analysis, a spatial filtering template is constructed for each steady-state visual evoked potential stimulation frequency by averaging historical training data corresponding to the frequency. The real-time acquired EEG signal is filtered through the spatial filtering template, and the Pearson correlation coefficient between the filtered signal and the sinusoidal reference signal of the corresponding stimulation frequency is calculated. The coefficient value is used as the response feature.

Images