Intelligent electroencephalogram characteristic analysis method based on motor imagery

Abstract

The invention discloses an intelligent electroencephalogram characteristic analysis method based on motor imagery, which is characterized in that an optimal normal form adaptive to an individual is screened through a normal form selection method based on sensitivity, and an improved Riemann minimum mean distance classifier is combined, so that the task instruction recognition rate of an EEG blind group is remarkably improved, and the classification accuracy is close to the group average level; a pulse neural network adopting an STDP learning mechanism and a side suppression mechanism is designed, the robustness of a low signal-to-noise ratio signal is enhanced, the classification precision is remarkably improved, a classification result is monitored in real time through a safety stop signal based on a recall rate and a confusion matrix, the accuracy of emergency instruction triggering is ensured, the misoperation risk is remarkably reduced, and the accuracy of emergency instruction triggering is improved. Through a multi-mode control strategy combining motor imagery and steady-state visual evoked potential, the delay of instruction response is optimized, the reliability of the system is improved, multi-task cooperation is realized through a parallel channel, the instruction recognition rate is greatly improved, and multi-scene application of medical rehabilitation robots, intelligent manufacturing and the like can be supported.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology and relates to a method for analyzing intelligent EEG characteristics, in particular to a method for analyzing intelligent EEG characteristics based on motor imagery. Background Art

[0002] Brain-computer interface (BCI) technology enables direct interaction between the human brain and external devices by decoding electroencephalogram (EEG) signals, showing broad application prospects in medical rehabilitation, intelligent manufacturing, and other fields. However, existing non-invasive EEG-based BCI technologies face the following key bottlenecks in large-scale group testing:

[0003] The problem of "EEG blindness" is prominent: the EEG signals of some subjects have significant specific characteristics, which causes the classification accuracy of traditional analysis algorithms to be far lower than the group average. This phenomenon seriously limits the universality of brain-computer interfaces, especially in medical rehabilitation scenarios, and cannot meet personalized needs.

[0004] Low efficiency in motor imagery (MI) signal processing: The core challenge of MI-BCI is to extract effective features from MI-EEG signals with low signal-to-noise ratio. Existing methods, such as common spatial pattern (CSP), have poor feature stability in small samples or complex noise environments, large fluctuations in classification accuracy, and are difficult to adapt to individual differences.

[0005] The coordinated control of multimodal signals is complex: When existing systems simultaneously process multiple types of EEG signals such as MI and SSVEP, there is interference between signals, which leads to a decrease in the command recognition rate of parallel channels. In addition, traditional classifiers have insufficient generalization capabilities in multi-task scenarios, and the diversity of control commands is limited.

[0006] The application of spiking neural networks (SNNs) is limited: Although SNNs have natural advantages in spatiotemporal signal processing, the traditional fully connected SNN network structure is rigid and cannot dynamically adapt to individual EEG feature differences. High-density connections can easily lead to a significant increase in computing power consumption. In addition, the classification accuracy of SNN training algorithms under low signal-to-noise ratio EEG signals is obviously insufficient.

[0007] Current research in brain-computer interface technology focuses on addressing these issues through brain-inspired intelligent models and dynamic adaptation algorithms. However, balancing individual differences and system versatility in large groups, and achieving low-latency, high-precision, multimodal command collaborative control, remain key challenges that require urgent breakthroughs. Summary of the Invention

[0008] In order to solve the technical problems existing in the background technology, the present invention proposes an analysis method of intelligent EEG characteristics based on motor imagery, which significantly improves the performance and applicability of the brain-computer interface system.

[0009] The purpose of the present invention can be achieved through the following technical solutions:

[0010] An intelligent EEG characteristic analysis method based on motor imagery, comprising:

[0011] Step S1, obtaining the subject's motor imagery EEG signal through a three-lead non-invasive EEG acquisition device, preprocessing the original signal using wavelet transform combined with Kalman filtering algorithm to remove eye artifacts and extract motor imagery-related signals in a preset frequency band;

[0012] Step S2, a paradigm selection method based on sensitivity indicators, generating multiple paradigm combinations through the combination number formula, calculating the signal-to-noise ratio and classification stability of each combination, and screening the optimal paradigm suitable for the subject;

[0013] Step S3, constructing an improved generalized Riemann minimum mean distance classifier, introducing a dynamic weight factor between the log-Euclidean metric and the Riemann divergence, and optimizing the weight parameters through cross-validation to adapt to individual differences;

[0014] Step S4, designing a fully connected spiking neural network, using the STDP learning mechanism and LIF neuron model, converting the preprocessed EEG signals into spike trains for training, and optimizing the salient weight parameters to improve classification accuracy;

[0015] Step S5, evaluating the recall rate of the classification results based on the confusion matrix, selecting the instruction with the highest recall rate as the safe stop signal, and generating control instructions including forward, backward, left turn, right turn and stop;

[0016] In step S6, a multimodal brain-computer interface system is constructed by combining motor imagery and steady-state visual evoked potential paradigms to achieve real-time control command output through parallel channels.

[0017] Furthermore, the Kalman filtering algorithm described in step S1 includes:

[0018] Establish the AR model of electrooculogram artifacts, and use the state equation x k =A k x k-1 +w k-1 and the observation equation y k =C k x k +v k Estimating artifact signals;

[0019] Among them, x k is the state vector at the kth moment, representing the estimated value of the electrooculogram artifact, A k is the state transfer matrix, describing the change of system state from time k-1 to time k, w k-1is process noise, which obeys zero-mean white noise distribution and has a covariance matrix of Q k ,y k is the observation vector at the kth moment, that is, the actual collected mixed signal, C k is the observation matrix, which maps the state vector to the observation space, v k is the observation noise, which obeys the zero-mean white noise distribution and the covariance matrix is ​​R k ;

[0020] Adopting output error correction strategy Remove artifacts, where H k is the Kalman gain matrix.

[0021] Furthermore, the calculation of the sensitivity index in step S2 includes:

[0022] The power spectrum characteristics of EEG signals are extracted through wavelet time-frequency analysis;

[0023] Calculate the signal-to-noise ratio and classification variance of the features under different paradigm combinations, and select the combination with the smallest variance and the highest signal-to-noise ratio.

[0024] Furthermore, the optimization of the dynamic weight factor in step S3 includes:

[0025] Leave-one-out cross-validation was used to determine the initial weights;

[0026] For "EEG blind" subjects, the weight factors are dynamically adjusted according to their specialized EEG features, so that the classifier's balance between Euclidean space and Riemannian manifold can adapt to individual differences.

[0027] Furthermore, the training of the spiking neural network in step S4 includes:

[0028] AER coding is used to achieve spatiotemporal sparse transmission of pulse events;

[0029] The synaptic weight is adjusted by the STDP rule. If the presynaptic neuron pulse is earlier than the postsynaptic neuron pulse, the synaptic connection is strengthened, otherwise it is weakened.

[0030] Optimizing network sparsity using lateral inhibition mechanism.

[0031] Furthermore, the determination of the safety stop signal in step S5 includes:

[0032] Construct a multi-category confusion matrix and calculate the recall rate of each instruction;

[0033] Set a recall rate threshold and select instructions exceeding the threshold as emergency stop signals.

[0034] Furthermore, the design of the multimodal system in step S6 includes:

[0035] The motor imagery paradigm was used for directional control;

[0036] The steady-state visual evoked potential paradigm was used to trigger specific action instructions;

[0037] The international 10-20 system was used to arrange electrodes, and motor imagery signals and steady-state visual evoked potential signals were collected synchronously.

[0038] A brain-computer interface system based on the above-mentioned analytical method comprises:

[0039] The EEG signal acquisition module is equipped with three-lead non-invasive dry electrodes, and the electrode positions are arranged according to the international 10-20 system to collect the subject's motor imagery EEG signals and steady-state visual evoked potential signals in real time;

[0040] The signal preprocessing module integrates the wavelet transform algorithm to extract the motor imagery related signals in the preset frequency band and uses the Kalman filter algorithm to remove the electrooculogram artifacts;

[0041] The spiking neural network module builds a fully connected spiking neural network, adopts the LIF neuron model and STDP learning mechanism, realizes the spatiotemporal sparse transmission of spiking events through AER encoding, and optimizes the network sparsity using the lateral inhibition mechanism;

[0042] The pulse classification output module integrates an improved generalized Riemann minimum mean distance classifier, balances the log-Euclidean metric and Riemannstein divergence through a dynamic weight factor, calculates the recall rate of each instruction based on the confusion matrix, and selects the classification result with the highest recall rate;

[0043] The pulse signal decoding module decodes the classification results into control instructions including forward, backward, left turn, right turn and stop;

[0044] The control command output module transmits the decoded commands to the rehabilitation robot in real time through a parallel channel, driving it to perform corresponding actions.

[0045] Furthermore, the system is optimized for stroke rehabilitation including:

[0046] Adapt patient-specific EEG characteristics through sensitivity index method;

[0047] A simplified brain-like model is used to reversely generate an equivalent spiking neural network and delete silent neuron synapses.

[0048] A computer-readable storage medium stores a computer program, which implements the steps of the above-mentioned analysis method when executed by a processor.

[0049] Beneficial effects of the present invention: The intelligent EEG characteristic analysis method provided in this application screens the optimal paradigm adapted to the individual through a sensitivity-based paradigm selection method. Combined with an improved Riemann minimum mean distance classifier, the task command recognition rate of the "EEG blind" group is significantly improved, and the classification accuracy is close to the group average level. The design adopts the STDP learning mechanism and the lateral inhibition mechanism of the spiking neural network to enhance the robustness to low signal-to-noise ratio signals and significantly improve the classification accuracy. By designing a safe stop signal based on the recall rate and monitoring the classification results in real time through the confusion matrix, the accuracy of emergency command triggering is ensured and the risk of misoperation is significantly reduced. By combining the multimodal control strategy of motor imagery and steady-state visual evoked potential, MI is used for direction control and SSVEP is used to trigger operation commands, the delay of command response is optimized and the reliability of the system is improved. In addition, the control commands are expanded to five categories: forward, backward, left turn, right turn, and stop. Multi-task collaboration is achieved through parallel channels, which greatly improves the command recognition rate, optimizes the visual stimulation interface, reduces the cognitive load of the subjects, and can support multiple scenarios such as medical rehabilitation robots and intelligent manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Flow chart of the method of the present invention.

[0051] Figure 2 Schematic diagram of the structure of the pulse neural network of the present invention.

[0052] Figure 3 Schematic diagram of the system of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0054] like Figure 1 As shown, the present invention provides an intelligent EEG characteristic analysis method based on motor imagery, which specifically includes the following steps:

[0055] In step S1, the motor imagery EEG signal (MI-EEG) of the subject is acquired through a three-lead non-invasive EEG acquisition device. The original signal is preprocessed using wavelet transform combined with Kalman filtering algorithm to remove eye artifacts and extract motor imagery-related signals in the 8-30 Hz frequency band.

[0056] The Kalman filtering algorithm described in step S1 includes:

[0057] Establish the AR model of electrooculogram artifacts, and use the state equation x k =A k x k-1 +w k-1 and the observation equation y k =C k x k +v k Estimating artifact signals;

[0058] Among them, x k is the state vector at the kth moment, representing the estimated value of the electrooculogram artifact, A k is the state transfer matrix, describing the change of system state from time k-1 to time k, w k-1 is process noise, which obeys zero-mean white noise distribution and has a covariance matrix of Q k ,y k is the observation vector at the kth moment, that is, the actual collected mixed signal, C k is the observation matrix, which maps the state vector to the observation space, v k is the observation noise, which obeys the zero-mean white noise distribution and the covariance matrix is ​​R k ;

[0059] Adopting output error correction strategy Remove artifacts, where H k is the Kalman gain matrix.

[0060] Step S2, based on the paradigm selection method of the sensitivity index (SPS), generates multiple paradigm combinations through the combination number formula, calculates the signal-to-noise ratio and classification stability of each combination, and screens the optimal paradigm suitable for the subject.

[0061] Specifically, the calculation of the sensitivity index in step S2 includes:

[0062] The power spectral density characteristics of EEG signals were extracted through wavelet time-frequency analysis, and the average power of EEG signals in the 8-30 Hz frequency band under different paradigm combinations was calculated;

[0063] The signal-to-noise ratio and classification stability were calculated under different paradigm combinations. The signal-to-noise ratio (SNR) was defined as the ratio of signal power to noise power. The noise power was calculated using the task-free resting signal in the frequency band above 30 Hz. The classification stability was evaluated by the variance of the classification accuracy of multiple experiments under the same paradigm. The paradigm combination with a variance less than 0.05 and a signal-to-noise ratio greater than 10 dB was selected.

[0064] In step S3, an improved generalized Riemann minimum mean distance classifier is constructed, a dynamic weight factor is introduced between the log-Euclidean metric and the Riemann divergence, and the weight parameters are optimized through cross-validation to adapt to individual differences.

[0065] Specifically, the optimization of the dynamic weight factor in step S3 includes:

[0066] The leave-one-out cross-validation method is used to determine the initial weights. The training set is divided into (N-1) subsets to train the classifier, and the remaining subset is used for validation. After repeating N times, the average classification accuracy is calculated. The initial value of the weight factor λ is set to 0.5, and then the value of λ is adjusted according to the accuracy of the validation set, with an adjustment step of 0.1.

[0067] For "EEG-blind" subjects, the weight factors are dynamically adjusted according to their specialized EEG characteristics, and an individual offset Δλ is introduced into the classifier parameters. The offset range is set to ±0.2, so that the classifier can adapt to individual differences in the balance between Euclidean space and Riemannian manifold.

[0068] Step S4, design a fully connected spiking neural network (SNN), such as Figure 2 As shown in the figure, the STDP learning mechanism and LIF neuron model are used to convert the preprocessed EEG signals into pulse trains for training, and the salient weight parameters are optimized to improve the classification accuracy.

[0069] Specifically, the training of the spiking neural network in step S4 includes:

[0070] AER coding is used to achieve spatiotemporal sparse transmission of spike events, where each spike event contains the neuron address and firing timestamp;

[0071] The synaptic weight is adjusted by the STDP rule, wherein the specific implementation of the STDP rule is: if the presynaptic neuron pulse emission time t pre Earlier than the postsynaptic neuron pulse firing time t post , the synaptic weight increases On the contrary, the synaptic weight decreases Among them, α=0.01, β=0.05, τ=20ms;

[0072] A lateral inhibition mechanism (Winner-Takes-All) is used to optimize network sparsity. The lateral inhibition mechanism is implemented through a 1:1 connection between excitatory neurons and inhibitory neurons. The pulse output of each excitatory neuron inhibits the membrane potential of other excitatory neurons in the same layer, and the inhibition strength is 0.3 times the current membrane potential value.

[0073] Step S5: Evaluate the recall rate of the classification results based on the confusion matrix, select the instruction with the highest recall rate as the safe stop signal, and generate control instructions including forward, backward, left turn, right turn and stop.

[0074] Specifically, the determination of the safety stop signal in step S5 includes:

[0075] Construct a multi-class confusion matrix, where the rows represent the true categories, the columns represent the predicted categories, and the element values ​​are the number of classification results;

[0076] Calculate the recall rate of each instruction, recall rate R i =TP i / (TP i +FN i ), where TP i is the number of true positives for category i, FN i is the number of false negatives;

[0077] The recall rate threshold is set to 95%, and the first instruction exceeding the threshold is selected as the emergency stop signal.

[0078] In step S6, a multimodal brain-computer interface system is constructed by combining motor imagery (MI) and steady-state visual evoked potential (SSVEP) paradigms to achieve real-time control command output through parallel channels.

[0079] Specifically, the design of the multimodal system in step S6 includes:

[0080] The motor imagery (MI) paradigm was used to control the movement direction of the rehabilitation robot. The left / right arrows in the visual stimulation interface prompted the subjects to perform the corresponding hand movement imagery;

[0081] The steady-state visual evoked potential (SSVEP) paradigm was used to trigger the “forward” and “backward” commands, with the stimulation frequencies set at 12 Hz and 15 Hz, corresponding to the two operational commands, respectively;

[0082] The electrodes were arranged at C3, C4, and Pz points in the international 10-20 system to collect motor imagery (MI) signals, and at Oz point to collect steady-state visual evoked potential (SSVEP) signals.

[0083] like Figure 3 As shown, the present application also provides a brain-computer interface system based on an intelligent EEG characteristic analysis method of motor imagery, including: an EEG signal acquisition module, a signal preprocessing module, a pulse neural network module, a pulse classification output module, a pulse signal decoding module, and a control instruction output module.

[0084] Among them, the EEG signal acquisition module is equipped with three-lead non-invasive dry electrodes, and the electrode positions are arranged according to the international 10-20 system, which is used to collect the subject's motor imagery EEG signals (MI-EEG) and steady-state visual evoked potential (SSVEP) signals in real time.

[0085] The signal preprocessing module integrates a wavelet transform algorithm to extract motor imagery-related signals in a preset frequency band, and uses a Kalman filter algorithm to remove electrooculogram artifacts.

[0086] The spiking neural network module constructs a fully connected spiking neural network (SNN), adopts the LIF neuron model and STDP learning mechanism, realizes the spatiotemporal sparse transmission of spike events through AER encoding, and uses the lateral inhibition mechanism (Winner-Takes-All) to optimize the network sparsity.

[0087] The pulse classification output module integrates an improved generalized Riemann minimum mean distance classifier, balances the log-Euclidean metric and Riemannstein divergence through a dynamic weight factor, calculates the recall rate of each instruction based on the confusion matrix, and selects the classification result with the highest recall rate.

[0088] The pulse signal decoding module decodes the classification results into control instructions including forward, backward, turn left, turn right and stop.

[0089] The control instruction output module transmits the decoded instructions to the rehabilitation robot in real time through a parallel channel, driving it to perform corresponding actions.

[0090] Among them, the optimization of brain-computer interface system for stroke rehabilitation includes:

[0091] Adapt patient-specific EEG characteristics through sensitivity index method;

[0092] A simplified brain-like model is used to reversely generate an equivalent spiking neural network and delete silent neuron synapses.

[0093] The present application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-mentioned intelligent EEG characteristic analysis method based on motor imagery.

[0094] The above content is merely an example and explanation of the structure of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the scope of protection of the present invention.

Claims

1. An intelligent EEG characteristic analysis method based on motor imagery, characterized in that: include: Step S1, obtaining the subject's motor imagery EEG signal through a three-lead non-invasive EEG acquisition device, preprocessing the original signal using wavelet transform combined with Kalman filtering algorithm to remove eye artifacts and extract motor imagery-related signals in a preset frequency band; Step S2, a paradigm selection method based on sensitivity indicators, generating multiple paradigm combinations through the combination number formula, calculating the signal-to-noise ratio and classification stability of each combination, and screening the optimal paradigm suitable for the subject; Step S3, constructing a generalized Riemann minimum mean distance classifier, introducing a dynamic weight factor between the log-Euclidean metric and the Riemann divergence, and optimizing the weight parameters through cross-validation to adapt to individual differences; Step S4, designing a fully connected spiking neural network, using the STDP learning mechanism and LIF neuron model, converting the preprocessed EEG signals into spike trains for training, and optimizing the salient weight parameters to improve classification accuracy; Step S5, evaluating the recall rate of the classification results based on the confusion matrix, selecting the instruction with the highest recall rate as the safe stop signal, and generating control instructions including forward, backward, left turn, right turn and stop; In step S6, a multimodal brain-computer interface system is constructed by combining motor imagery and steady-state visual evoked potential paradigms to achieve real-time control command output through parallel channels.

2. The analysis method according to claim 1, wherein: The Kalman filtering algorithm described in step S1 includes: Establish the AR model of electrooculogram artifacts, and use the state equation x k =A k x k-1 +w k-1 and the observation equation y k =C k x k +v k Estimating artifact signals; Among them, x k is the state vector at the kth moment, representing the estimated value of the electrooculogram artifact, A k is the state transfer matrix, describing the change of system state from time k-1 to time k, w k-1 is process noise, which obeys zero-mean white noise distribution and has a covariance matrix of Q k ,y k is the observation vector at the kth moment, that is, the mixed signal actually collected, C k is the observation matrix, which maps the state vector to the observation space, v k is the observation noise, which obeys the zero-mean white noise distribution and the covariance matrix is ​​R k ; Adopting output error correction strategy Remove artifacts, where H k is the Kalman gain matrix.

3. The analysis method according to claim 1, wherein: The calculation of the sensitivity index in step S2 includes: The power spectrum characteristics of EEG signals are extracted through wavelet time-frequency analysis; Calculate the signal-to-noise ratio and classification variance of the features under different paradigm combinations, and select the combination with the smallest variance and the highest signal-to-noise ratio.

4. The analysis method according to claim 1, wherein: The optimization of the dynamic weight factor in step S3 includes: Leave-one-out cross-validation was used to determine the initial weights; For "EEG-blind" subjects, the weight factors are dynamically adjusted according to their specialized EEG features, so that the classifier's balance between Euclidean space and Riemannian manifold can adapt to individual differences.

5. The analysis method according to claim 1, wherein: The training of the spiking neural network in step S4 includes: AER coding is used to achieve spatiotemporal sparse transmission of pulse events; The synaptic weight is adjusted by the STDP rule. If the presynaptic neuron pulse is earlier than the postsynaptic neuron pulse, the synaptic connection is strengthened, otherwise it is weakened. Optimizing network sparsity using lateral inhibition mechanism.

6. The analysis method according to claim 1, wherein: The determination of the safety stop signal in step S5 includes: Construct a multi-category confusion matrix and calculate the recall rate of each instruction; Set a recall rate threshold and select instructions exceeding the threshold as emergency stop signals.

7. The analysis method according to claim 1, wherein: The design of the multimodal system in step S6 includes: The motor imagery paradigm was used for directional control; The steady-state visual evoked potential paradigm was used to trigger specific action instructions; The international 10-20 system was used to arrange electrodes, and motor imagery signals and steady-state visual evoked potential signals were collected synchronously.

8. A brain-computer interface system based on the analysis method according to any one of claims 1 to 7, characterized in that: include: The EEG signal acquisition module is equipped with three-lead non-invasive dry electrodes, and the electrode positions are arranged according to the international 10-20 system to collect the subject's motor imagery EEG signals and steady-state visual evoked potential signals in real time; The signal preprocessing module integrates the wavelet transform algorithm to extract the motor imagery related signals in the preset frequency band and uses the Kalman filter algorithm to remove the electrooculogram artifacts; The spiking neural network module builds a fully connected spiking neural network, adopts the LIF neuron model and STDP learning mechanism, realizes the spatiotemporal sparse transmission of spiking events through AER encoding, and optimizes the network sparsity using the lateral inhibition mechanism; The pulse classification output module integrates an improved generalized Riemann minimum mean distance classifier, balances the log-Euclidean metric and Riemannstein divergence through a dynamic weight factor, calculates the recall rate of each instruction based on the confusion matrix, and selects the classification result with the highest recall rate; The pulse signal decoding module decodes the classification results into control instructions including forward, backward, left turn, right turn and stop; The control command output module transmits the decoded commands to the rehabilitation robot in real time through a parallel channel, driving it to perform corresponding actions.

9. The brain-computer interface system according to claim 8, characterized in that: The system's optimization for stroke rehabilitation includes: Adapt patient-specific EEG characteristics through sensitivity index method; A simplified brain-like model is used to reversely generate an equivalent spiking neural network and delete silent neuron synapses.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the parsing method according to any one of claims 1 to 7 are implemented.

Images