Electroencephalogram-driven behavior recognition method based on dynamic evolution of coherent network features

Abstract

The invention belongs to the technical field of signal processing and analysis, and particularly relates to an electroencephalogram driving behavior recognition method based on coherent network characteristic dynamic evolution, which comprises the following steps of: 1, performing standard preprocessing on an original electroencephalogram signal, selecting a beta-wave frequency band as an analysis frequency band, calculating a coherent value under the analysis frequency band, and calculating a coherent value under the analysis frequency band; the synchronous intensity of the two electroencephalogram signals is measured to serve as a network connection weight, and a weighted coherent network is constructed; and 2, for each time point, extracting the degree and clustering coefficient of each node in the weighted coherent network, for each pair of nodes, calculating the instantaneous phase difference of each pair of nodes, and meanwhile, introducing the space geometric distance between two electrodes to obtain a behavior characteristic response index. And step 3, obtaining power spectrum parameters of all brain region signals of the target behavior under the analysis frequency band, comparing the target behavior identification value with the behavior threshold value, and judging whether the target behavior is activated or not. The accuracy, the real-time performance and the physiological interpretation performance of behavior recognition are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal analysis and processing, and particularly relates to an electroencephalogram (EEG)-driven behavior recognition method based on the dynamic evolution of coherent network features. Background Art

[0002] As an electrophysiological manifestation of neural activities, electroencephalogram (EEG) signals are widely used in many fields such as brain-computer interfaces, neuroengineering, clinical auxiliary diagnosis, and motor function monitoring. With the increasing application requirements of neuromodulation, neurofeedback, motor intention recognition, and rehabilitation control, how to accurately extract behavior-related neural activity patterns from EEG signals and construct a high-precision recognition method for specific motor or cognitive behaviors has become one of the research focuses in this field in recent years.

[0003] Traditional EEG-driven behavior recognition methods generally follow the following technical paths: One is the event-related desynchronization / synchronization (ERD / ERS) analysis method. This type of method analyzes the power changes of specific frequency bands (such as α, β, μ waves) before and after the execution of the target behavior. For example, by comparing the power suppression of the β frequency band before and after movement, movement behavior recognition or regional activation localization can be achieved. However, this type of method usually requires a large number of trials for averaging to stably recognize, has poor anti-interference ability, and can only reflect the local response of a single brain region, making it difficult to characterize the dynamic cooperation between multiple brain regions. The second is the independent component analysis (ICA) and source separation methods. Such methods attempt to separate independent neural source signals related to specific behaviors from the mixed EEG signals, and perform feature extraction and classification based on the source signals. This method solves the signal aliasing problem to a certain extent, but it has a large dependence on computing resources, insufficient algorithm robustness, and the actually separated components lack behavioral semantic interpretability, and the results are not universal.

[0004] The third is the coherence analysis and functional connectivity map modeling method. In recent years, EEG signals, as an expression carrier of functional connectivity networks, have received increasing attention. Related research constructs functional connectivity networks based on the coherence value, phase locking value, mutual information, etc. between two channels, and then combines graph theory indicators such as degree, clustering coefficient, and efficiency for recognition analysis. For example, existing research has constructed a whole-brain coherence network in the β band and used node centrality or motif structure analysis for classification. However, such methods have significant limitations: (1) Most networks are static networks, and only calculate the connection values on average within the entire behavior cycle, ignoring the real-time evolution characteristics of the network induced by the behavior; (2) The connection values only consider the strength, such as the coherence value, but ignore the dynamic changes of the connection structure over time, such as the connection growth rate and the local structure change trend; (3) The connection indicators are not integrated with the frequency domain power parameters, resulting in the decoupling of frequency domain information and network structure; (4) The behavior recognition values are usually not uniformly modeled, lacking a comprehensive recognition function. Summary of the Invention

[0005] In view of this, the main object of the present invention is to provide an electroencephalogram (EEG)-driven behavior recognition method based on the dynamic evolution of coherent network features. This method can simultaneously capture the dynamic reconstruction process of the neural network structure and the behavior-induced changes in frequency domain features, overcoming the problems of insufficient dynamics, structural spectrum fragmentation, and weak feature fusion in traditional methods.

[0006] The technical solution adopted by the present invention is as follows: An EEG-driven behavior recognition method based on the dynamic evolution of coherent network features, the method comprising: Step 1: Perform standard preprocessing on the original EEG signal, select the β-wave frequency band as the analysis frequency band, and use short-time wavelet transform to extract the instantaneous frequency components of the analysis frequency band; for each pair of electrodes, within a specified time window, calculate the coherence value in the analysis frequency band to measure the synchronization strength of the two EEG signals, and use it as the network connection weight to construct a weighted coherence network; Step 2: For each time point, extract the degree and clustering coefficient of each node in the weighted coherence network, and calculate the dynamic evolution index; combine the network connection weight with the dynamic evolution index to construct a unified EEG-driven behavior feature mapping; at each time point, for each pair of nodes, calculate their instantaneous phase difference, and at the same time introduce the spatial geometric distance between the two electrodes to obtain the behavior feature response index; Step 3: Obtain the power spectrum parameters of all brain region signals of the target behavior in the analysis frequency band, combine them with the feature response index, calculate the target behavior recognition value, set the behavior threshold of the target behavior in the analysis frequency band, compare the target behavior recognition value with the behavior threshold, and determine whether the target behavior is activated.

[0007] Further, after selecting the β-wave frequency band as the analysis frequency band, use a band-pass filter to process the original EEG signal, and the frequency range of the band-pass filter is 13 Hz to 30 Hz; the target behavior movement-related behaviors include at least: eye movement behavior, manual behavior, and foot movement behavior.

[0008] Further, the power spectrum parameters include: the average power spectrum value in the resting state, the standard deviation of the power spectrum in the resting state, the entropy value of the power spectrum, and the average value of the entropy value of the power spectrum.

[0009] Further, the network connection weight is: ; wherein, represents the network connection weight between the EEG signal of electrode and the EEG signal of electrode at time ; is the time window; is the time integration variable; is the electrode under the β-wave frequency band The electroencephalogram (EEG) signal and the electrode The coherence value of the EEG signal at time ; Is the spectral entropy of the EEG signal of the electrode in the β-wave frequency band; Is the spectral entropy of the EEG signal of the electrode in the β-wave frequency band; Is the voltage amplitude of the EEG signal of the electrode ; Is the voltage amplitude of the EEG signal of the electrode ; Is the time The central frequency of the EEG signal of the electrode in the β-wave frequency band at is the power spectral density at; Is the time The central frequency of the EEG signal of the electrode in the β-wave frequency band at is the power spectral density at; Is the phase of the EEG signal of the electrode ; Is the phase of the EEG signal of the electrode ; Is an indicator function, with a value of 1 when the condition is met and 0 otherwise; Is the signal-to-noise ratio of the EEG signal of the electrode ; Is the set signal-to-noise ratio threshold; Is the signal-to-noise ratio of the EEG signal of the electrode ;

[0010] Furthermore, the dynamic evolution index at time is: ; ; Among them, is the clustering coefficient of the node corresponding to the th electrode at time ; is the degree of the node corresponding to the th electrode at time ;

[0011] Furthermore, the characteristic response index at time is : ; Among them, is the electrode at time ​The phase and time of the electroencephalogram (EEG) signal The electrode at time The phase difference of the phase of the EEG signal; Is the electrode And the electrode The spatial Euclidean distance; Is the spatial attenuation scale coefficient. When The unit of is centimeter, the value ranges from 2 to 5; Is the total number of electrodes.

[0012] Furthermore, at time The clustering coefficient of the node corresponding to the th electrode is As follows: ; Wherein, Is the th actual number of edges existing between the neighbors of the node corresponding to the electrode.

[0013] Furthermore, at time The target behavior recognition value Is as follows: ; Wherein, Is the power spectral density integral value of the electrode At time In the Wave frequency band, , Is the frequency integration variable; Is the average power spectrum of the electrode 𝑖 in the resting state; Is the electrode The standard deviation of the power spectrum in the resting state; Is the spectral entropy of the EEG signal of the electrode At time In the β wave frequency band; Is the entropy value of the power spectrum of the electrode At time ; Is the average value of the entropy values of the power spectra of all electrodes in the resting state.

[0014] Furthermore, when the target behavior is an eye movement behavior, the recognition threshold Ranges from 0.20 to 0.30; when the target behavior is a manual behavior, the recognition threshold Ranges from 0.35 to 0.45; when the target behavior is a foot behavior, the recognition threshold Ranges from 0.50 to 0.60; if Is greater than the recognition threshold , it is determined that the target behavior is in an active state.

[0015] With the above technical solutions, the present invention has the following beneficial effects: significantly improving the recognition accuracy and dynamic discrimination ability of EEG signals in complex motor behavior states. Traditional EEG recognition methods generally have technical bottlenecks such as only being able to capture local signal features, lacking the ability of multi-region collaborative modeling, and being insensitive to temporal behavior changes. In response to these problems, the present invention proposes to extract multi-dimensional signal features such as the synchrony, connection strength, phase relationship, and spectral reconstruction between neural activities from time-continuous and multi-channel EEG signals, and construct a dynamic evolution network based on these features to real-time depict the topological evolution process of the EEG network structure as the behavior state changes. This network model not only considers the coherent connection weights between brain regions, but also introduces the connection growth rate, local clustering changes, spatial structure relationships, and network fluctuation trends, so as to achieve high-precision modeling of the brain region collaborative response mechanism. On this basis, the present invention further introduces spectral power changes, spectral entropy reconstruction, and baseline statistical features into the recognition index function, and constructs a unified recognition value expression form, making the recognition model not only have a high dynamic response ability, but also have cross-subject adaptability and behavior specificity. In addition, the present invention introduces a signal-to-noise ratio constraint mechanism in the aspect of neural signal quality regulation, effectively suppressing the interference of low-quality signals on the recognition results, thereby improving the stability and reliability of the overall recognition system. Compared with traditional recognition methods based on static thresholds or single-frequency features, the present invention can continuously discriminate multiple types of behavior states such as eye movement, hand movement, and foot movement within an extremely short time window, and is particularly suitable for application scenarios with high requirements for response speed and recognition accuracy such as real-time brain-computer interface control, movement intention recognition, and neurorehabilitation feedback. By unifying and integrating structural features and frequency domain indicators in the same modeling framework, the present invention realizes the complex modeling and dynamic analysis of neural behavior signals while maintaining computational controllability, and has high engineering adaptability and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flowchart of the EEG-driven behavior recognition method based on the dynamic evolution of coherent network features provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] All features disclosed in this specification, or all steps in the disclosed methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

[0018] Any feature disclosed in this specification (including any additional claims, abstract) can be replaced by other equivalent or similar-purpose alternative features unless specifically stated. That is, unless specifically stated, each feature is only an example of a series of equivalent or similar features.

[0019] refer to Figure 1 :A method for identifying EEG-driven behaviors based on the dynamic evolution of coherent network features, the method comprising: Step 1: Perform standard preprocessing on the original EEG signal, select the β wave frequency band as the analysis frequency band, and use short-time wavelet transform to extract the instantaneous frequency component of the analysis frequency band; for each pair of electrodes, calculate the coherence value under the analysis frequency band within the specified time window to measure the synchronization strength of the two EEG signals, which is used as the network connection weight to construct a weighted coherence network; As a high-noise, low-amplitude, non-stationary bioelectric signal, EEG signals are often mixed with a variety of interference components in their original form, including eye movement artifacts, myoelectric noise, power frequency interference, etc. Therefore, a multi-channel synchronous cleaning strategy is adopted in the preprocessing stage, using a high-pass filter to remove DC drift, a low-pass filter to suppress high-frequency myoelectric interference, and a 50Hz or 60Hz notch process to remove power frequency noise. On this basis, the inter-channel reference settings are standardized according to the configuration of the experimental equipment to ensure that all channels have a unified reference benchmark. In addition, it is necessary to eliminate possible outliers, use statistical distribution analysis methods to identify outlier data, and use interpolation or noise reduction reconstruction technology to repair missing or abnormal signal segments to ensure the continuity and physical consistency of the signal.

[0020] After the initial cleaning, this method explicitly selects the β wave frequency band as the core area for analysis. This frequency band generally corresponds to a frequency range between 13Hz and 30Hz, and is widely believed to be closely related to brain activities such as motor preparation, attention regulation, and behavioral activation. In order to accurately extract the dynamic frequency characteristics of the β band, the short-time wavelet transform technology is introduced. This technology has good time-frequency localization capabilities and can achieve an optimal balance between time accuracy and frequency resolution. It is particularly suitable for processing non-stationary EEG signals. During the short-time wavelet transform processing, the original signal is decomposed into multiple local oscillation components with different scales, so that the instantaneous frequency, instantaneous amplitude, and instantaneous phase information in the target frequency band can be effectively extracted. After performing this transformation on each channel signal, a β band energy representation distributed in the two-dimensional time-frequency domain can be obtained for subsequent coherence calculations and network weight assignments.

[0021] When constructing a coherence network, the core idea is to quantitatively measure the degree of synchronous activity in the beta band between different brain regions, and then use this as the connection weight between nodes in the network. To this end, in a sliding time window, the coherence analysis of the beta band components of any two-channel signals is carried out. By calculating the coherence values of the signals, a metric index reflecting the resonance degree of the two signals in the specified frequency band can be obtained. Different from traditional static connectivity analysis, this process has obvious time dependence, that is, each time window corresponds to a new coherence value matrix, reflecting the dynamic evolution characteristics of the brain network. This construction method can not only reflect the short-term cooperative relationship between different brain regions, but also retain the information on the change of coherence patterns before and after the initiation of behavior.

[0022] The calculation results of the coherence values are used in the construction of the weighted network. In this network model, each EEG channel is regarded as a network node, and the edge weight between the nodes is determined by their coherence values. The higher the edge weight value, the stronger the functional coupling between the two brain regions in the beta band. This weighted network not only contains the spatial distribution information, but also incorporates the dynamic changes in the time dimension, which is an extension of the traditional static functional connectivity network. By advancing the time axis in a sliding window manner, the system can continuously construct multiple coherence network frames, forming a network evolution process with a time series structure, and then providing continuous input for the extraction of dynamic features and behavior discrimination.

[0023] In addition, during the construction of the coherence network, it is also necessary to combine the topological arrangement structure between EEG channels to ensure the consistency between the network connections and the actual electrode layout. In practical applications, EEG channels are arranged according to the international standard 10-20 system, and the geometric distribution between different electrodes affects the physical interpretability of their coherence values. Therefore, when calculating the connection strength, not only the frequency domain coherence is considered, but also the physical distance between channels is processed to avoid the interference of non-physiological strong connections on the overall structure of the network. At the same time, to maintain the sparsity and robustness of the network structure, the connection edges below a certain threshold can be weakened or even deleted, so as to construct a more distinguishable EEG coherence network model.

[0024] Finally, after calculating the coherence values between all channel pairs, a multi-frame time series weighted network cluster can be generated, providing basic data for subsequent analysis. These network frames, as the core input for the extraction and recognition of behavioral features in the subsequent steps, show the evolution trend of the connection patterns in time, which is the key feature that this method attempts to capture. Overall, this step not only realizes the noise suppression and frequency band focusing of the original EEG data, but also effectively transforms the multi-channel biological signals into an EEG network form with physical meaning and time series structure through the combination of time-frequency analysis and network modeling, providing rich and structured input features for the behavior recognition algorithm.

[0025] Step 2: For each time point, extract the degree and clustering coefficient of each node in the weighted coherence network, and calculate the dynamic evolution index; combine the network connection weight with the dynamic evolution index to construct a unified electroencephalogram-driven behavior feature mapping; at each time point, for each pair of nodes, calculate their instantaneous phase difference, and at the same time introduce the spatial geometric distance between two electrodes to obtain the behavior feature response index. Specifically, this step first performs a structural analysis on the coherence network at each time point, and extracts the network degree and clustering coefficient of each node as basic topological indicators. The degree of a node reflects the activity level of this electrode in the functional connection network, that is, the number of significant functional connections it has with other brain regions within the current time window. The change in degree can reflect the transfer of the electroencephalogram activity center and the dynamic change in the participation degree of brain regions. Especially during the process of movement preparation or execution behavior, the functional connections of certain specific brain regions may show a phenomenon of rapid enhancement. The clustering coefficient is used to measure the tightness of the local sub-network around a certain node, reflecting whether the brain region where this node is located is in a highly integrated state at the current moment. For the initiation of cognitive behavior, the change in the aggregation of the local network is of great significance, and obvious structural adjustments often occur along with the occurrence of specific behaviors.

[0026] After obtaining the time series of degree and clustering coefficient, the method further introduces a dynamic evolution analysis strategy to perform time difference processing on these indicators to extract the first-order change and second-order acceleration information in the time dimension. This analysis not only captures the absolute change of the indicator value, but also emphasizes its change rate and trend, and is used to reflect the dynamic transformation ability of the electroencephalogram network structure during the behavior initiation process. This dynamic evolution index has the characteristics of strong burst response and high time sensitivity, and can effectively characterize the functional activity change process of behavior-related brain regions. Especially in the instantaneous stage when the behavior changes from preparation to execution, this index usually shows violent fluctuations, providing a clear boundary signal for behavior recognition.

[0027] On this basis, to enhance the discriminative ability of features, the method also jointly fuses the connection weight of the coherence network with the above dynamic evolution index to construct a unified feature mapping expression. This fusion not only retains the connection strength information in the coherence network, but also introduces the time evolution attribute of the network topology, thus forming a multi-dimensional and multi-scale feature representation system. The essence of the feature mapping is a tensor structure that combines the network weight matrix and the topological time derivative matrix, and is used to simultaneously encode signal synchrony, connection stability and structural change trend. This mapping can be directly input into the subsequent behavior recognition module and has strong generalization ability and temporal reasoning ability.

[0028] In addition, this step further examines the distribution characteristics of the instantaneous phase difference between electrode pairs at each time point. As a basic component of the signal after wavelet transform, the instantaneous phase can reveal the relative synchronous delay of signals between brain regions without changing the amplitude energy. In a functional coupling network, the phase difference stability between two brain regions is often closely related to their degree of cooperation. The method calculates the instantaneous phase difference between all electrode pairs through each time window, not only capturing the network coupling state but also providing a basis for identifying phase reconstruction patterns related to specific behaviors.

[0029] To enhance the interpretability of this feature in the spatial domain, this method simultaneously introduces the spatial geometric distance between electrodes as a modulation factor and incorporates the distance factor into the calculation of the response of behavioral features. In terms of physiological mechanisms, the functional connection strength between different brain regions is significantly affected by their spatial distance. Strong synchronous signals are more likely to occur between closely spaced electrodes, but strong connections between distant brain regions are more behaviorally significant. Therefore, when calculating the behavioral response, the combined result of the phase difference and network weight is modulated by the spatial distance to enhance the behavioral relevance expressed by remote synchronous activation.

[0030] Step 3: Obtain the power spectrum parameters of the signals of all brain regions in the analysis frequency band for the target behavior, combine the feature response index, calculate the target behavior recognition value, set the behavior threshold for the target behavior in the analysis frequency band, compare the target behavior recognition value with the behavior threshold, and determine whether the target behavior is activated; During the power spectrum extraction process, this method continues the result of the short-time wavelet transform, integrates the energy of the signal in each time window within the β-wave frequency band to obtain the total power of each electrode at that moment in the frequency band. Further, by calculating the center frequency of the power spectrum, the center of gravity position of the current frequency band energy distribution can be determined, and modeling the statistical mean and standard deviation of the center frequency over a continuous time period helps to evaluate the stability of the frequency distribution and the spectrum reconstruction pattern under behavioral guidance. In addition, this method also introduces the power spectrum entropy as an index to measure the flatness and information complexity of the spectrum. The larger the power spectrum entropy, the more uniform the signal frequency distribution and the more discrete the brain region activity state; while a smaller power spectrum entropy usually represents a concentrated burst of energy release in a certain frequency band, which is often directly related to specific motor behaviors or cognitive events. These power spectrum parameters constitute the frequency domain representation of behavioral features, providing a stable and interpretable input basis for subsequent recognition.

[0031] Meanwhile, to achieve the fusion of frequency-domain features and coherent network evolution features, this method combines the frequency-domain power spectrum features with the behavioral feature response index obtained in the previous stage to construct the target behavior recognition value. The behavioral feature response index comprehensively considers the connection strength, phase difference, spatial topology, and dynamic evolution trend between brain regions, and is a quantitative expression of the comprehensive response ability of the entire brain network during the behavior triggering process. By combining this response index with the spectral energy change, the accuracy and robustness of behavior recognition can be significantly improved. Especially during the behavior activation process where the overall synchronization of the neural network is enhanced, the local clustering and global synchronization of the coherent network structure will bring a significant increase in spectral energy, and this change is accurately captured by the recognition value function, manifested as a rapid increase in the recognition value.

[0032] To convert the recognition value into a clear behavior activation criterion, this method constructs a behavior recognition threshold model for the target behavior in the β-wave frequency band. The threshold setting process is based on the statistical analysis of a large amount of experimental data, comprehensively considering the differences in the recognition value distributions of different behavior categories to ensure that the recognition value intervals of each behavior category are independent and non-overlapping. During the threshold determination process, the system compares the current recognition value with the threshold corresponding to the target behavior type in real time to determine whether the behavior is in an active state. For example, when the recognition value exceeds the activation threshold of the eye movement behavior but does not reach the determination interval of the manual behavior, the system can determine that the current is the active state of the eye movement behavior. Similarly, when the recognition value significantly increases and crosses the high threshold interval corresponding to the manual or foot movement behavior, the system can timely identify a higher-level or larger-scale behavior execution state. The recognition threshold not only plays a classification role but also provides a quantitative scale for the behavior intensity, enabling the system to not only identify whether a behavior occurs but also preliminarily judge the degree of the behavior response and the active range of the participating brain regions.

[0033] Furthermore, after selecting the β-wave frequency band as the analysis frequency band, a band-pass filter is used to process the original electroencephalogram (EEG) signal, and the frequency range of the band-pass filter is from 13 Hz to 30 Hz; the target behavior, i.e., the movement-related behavior, at least includes: eye movement behavior, manual behavior, and foot movement behavior.

[0034] The main purpose of using the band-pass filter is to suppress the interference components of other frequency bands unrelated to the β-wave while retaining the effective signal components within the β frequency band. Signal components higher than 30 Hz, such as γ waves and electromyogram interference, will be effectively weakened, and components lower than 13 Hz, such as δ, θ, or α waves, will also be filtered out, thereby ensuring the frequency-domain purity of the signal input and avoiding misjudgment caused by the mixing of non-target frequency band information into the behavior features. In addition, the filter design also needs to meet the phase linearity characteristic to avoid the phase distortion caused by filtering affecting the subsequent instantaneous phase difference calculation and coherence analysis. Usually, a finite impulse response filter (FIR) or a zero-phase IIR filter is used to maintain the balance between the time-domain structure and the frequency-domain accuracy of the signal.

[0035] After completing the frequency band limitation, the present invention confines the target behaviors to the category of typical motion-related behaviors, which at least includes three basic behavior types: eye movement behavior, manual behavior, and foot movement behavior. These three types of behaviors have highly specific spatial activation patterns and spectral response rules in neuroelectrophysiological activities, especially showing significant phenomena of enhanced frequency energy, enhanced network synchronization, and reconstructed coherence structure in the β-wave frequency band. For example, eye movement behavior often activates the low-intensity connection between the frontal lobe and the occipital lobe, showing relatively localized β-wave enhancement; while manual behavior is more involved in the motor cortex area near the central sulcus, and its β-wave activity shows stronger power increase and enhanced network coupling; foot movement behavior shows more obvious cross-hemisphere long-distance synchronization characteristics. These behaviors have quantifiable differences in the electroencephalogram network topology, power spectrum characteristics, and phase response patterns, so the accurate recognition of multiple types of behaviors can be achieved through a unified frequency band analysis framework.

[0036] Furthermore, the power spectrum parameters include: the average value of the power spectrum in the resting state, the standard deviation of the power spectrum in the resting state, the entropy value of the power spectrum, and the average value of the entropy value of the power spectrum.

[0037] Specifically, the average value of the power spectrum in the resting state is obtained by performing time-domain averaging on the power spectral density of each electrode in the β-wave frequency band during the non-task period, that is, when the subject is in a state of not performing any target behavior. This parameter represents the energy baseline of the brain region in the natural relaxation state of an individual and is an important reference for judging whether a certain brain region has behavior-induced activation. When a behavior occurs, if the power spectral density is significantly higher than this baseline value, it can usually be regarded that this brain region participates in a certain functional processing process. The standard deviation of the power spectrum reflects the degree of dispersion of power changes in the resting state and reveals the fluctuation amplitude of the spectral energy during the resting period. The smaller the standard deviation, the more stable the power spectrum during the resting period, and the easier it is to identify the change during behavior activation as a significant response; if the standard deviation is large, a stronger power change is required to distinguish it from the resting state.

[0038] In order to further capture the complexity of the power spectrum in the frequency distribution, this method introduces the power spectrum entropy as a measurement index. The power spectrum entropy is the entropy value calculated based on the normalized probability density of the power distribution in the β frequency band, representing the dispersion degree and complexity of the signal frequency domain structure. The higher the entropy value, the more uniform the energy distribution in the frequency band, and the signal shows multi-band cooperative participation; the lower the entropy value, it means that the energy is concentrated in a relatively narrow frequency band, which may correspond to the mandatory neural activation induced by specific behaviors. In behavior recognition, the power spectrum entropy can provide a more detailed discrimination ability than the total energy, especially suitable for identifying those behavior processes in which the frequency domain pattern is reconstructed while the total energy change is not significant.

[0039] In addition, to normalize and standardize the changes in power spectral entropy, this method also introduces the average value of power spectral entropy in the resting state as a reference benchmark. When the subject is not performing a task, the system records the power spectral entropy of each electrode for a long time, and averages the entropy values of all sampling points during this period to obtain the baseline level of entropy. In actual recognition, the power spectral entropy during the behavior period will be compared with this baseline value to evaluate whether the current signal frequency complexity exceeds the normal physiological fluctuation range. This comparison mechanism helps to avoid recognition errors caused by inter-individual differences in neuromodulation and improve the stability of behavior recognition across subjects.

[0040] Furthermore, the network connection weight is: ; Where represents the network connection weight between the EEG signal of electrode and the EEG signal of electrode at time ; is the time window; is the time integration variable; is the coherence value between the EEG signal of electrode in the β-wave frequency band and the EEG signal of electrode at time ; is the spectral entropy of the EEG signal of electrode in the β-wave frequency band; is the spectral entropy of the EEG signal of electrode in the β-wave frequency band; is the voltage amplitude of the EEG signal of electrode ; is the voltage amplitude of the EEG signal of electrode ; is the power spectral density of the EEG signal of electrode in the β-wave frequency band at the center frequency of the β-wave frequency band at time ; is the power spectral density of the EEG signal of electrode in the β-wave frequency band at the center frequency of the β-wave frequency band at time ; is the phase of the EEG signal of electrode ; is the phase of the EEG signal of electrode ; is the indicator function, with a value of 1 when the condition is met and 0 otherwise; is the signal-to-noise ratio of the EEG signal of electrode ; is the set signal-to-noise ratio threshold; is the electrode signal-to-noise ratio of the electroencephalogram signal.

[0041] The entire weight is defined as the integral average of the function terms at each moment within the sliding time window The core idea is to collect the coherence relationship and multi-dimensional neuroelectrical characteristics within a period of time before the analysis time moment, and estimate the functional connection strength between brain regions within this local time window. This method is manifested in engineering implementation as the distribution weighted average of the samples within the sliding window, so as to obtain a stable and dynamically responsive connection value, avoiding the instability of recognition caused by the violent fluctuation of instantaneous values. At the same time, the integral form reflects the time-domain filtering characteristics, has a high response sensitivity to the enhancement of connections related to slow-varying behaviors, and shows strong robustness to sudden artifact perturbations. In the integral kernel, the first factor to appear is , which is the electrode and the electrode at time coherence value in the β-wave frequency band. This coherence value is the most basic calculation element in the functional connection network, reflecting whether there is a stable phase relationship and frequency matching between two neural signals within a specific frequency band, that is, the frequency-domain cooperation degree. The enhancement of coherence in the β-wave frequency band is usually closely related to the activation of the motor-related functional network. Therefore, in the process of identifying eye movement, hand movement, and foot movement behaviors in the present invention, the coherence value is a key index for the collaborative participation of brain regions. Different from simple correlation calculations, the coherence value is modeled in the frequency domain space, emphasizing cross-temporal consistency and functional network characteristics, and has higher behavioral specificity and physiological interpretability.

[0042] However, the coherence value itself cannot fully reflect the effectiveness and behavioral relevance of the connections between brain regions. Therefore, the present invention introduces multiple modulation factors to perform multi-dimensional weight adjustment on the coherence value based on the physiological signal attributes. Among them, the second product term is , where and respectively represent the spectral entropy of the electrode and the electrode in the β frequency band. The larger the spectral entropy, the more uniform the energy distribution, the more complex the spectrum, and the lack of obvious dominant frequency components, which may mean that the current brain region is in a non-specific activation state or random fluctuation state. On the contrary, the smaller the spectral entropy, the more concentrated the energy, usually corresponding to the neural activation of behavioral relevance at a specific frequency. This term is modulated in the form of average entropy. The smaller the entropy value, the larger the connection value is amplified, and the larger the entropy value, the more the connection is suppressed, reflecting the enhancement effect of the "focus" of the brain region on the connection quality. The third product term involves the joint ratio of voltage amplitude and power spectral density. The formula form is , where is the instantaneous voltage amplitude, which is an intuitive physical quantity reflecting the intensity of neural activity. The power spectral density in the denominator is the energy distribution intensity of the channel at the center frequency of the beta wave. By combining the time-domain amplitude with the frequency-domain energy structure, this ratio emphasizes the responsiveness of high-amplitude, low-spectral-density signals to behavior. A large voltage amplitude indicates strong neural activity; a large power spectral density indicates that the signal is too concentrated and may lack flexibility in frequency regulation. This term makes the system more inclined to identify connection states with amplitude activity but not overly concentrated spectra, which is highly representative in motor behavior.

[0043] Next introduced is , which is the cosine term of the instantaneous phase difference between electrode and electrode . It measures the instantaneous synchrony of neural signals on a microscopic time scale. In neural activity, phase synchrony is a more sensitive connectivity metric than power, especially suitable for describing long-distance, low-latency co-activation phenomena across brain regions in neural networks. When the phases of two signals are exactly the same, the cosine value is 1, indicating the maximum connection strength; when the phase difference is 90 degrees, this term is 0; when the phases are opposite, it is -1. This term thus provides a direction-independent measure of enhanced instantaneous synchrony, helping the system identify short-term, strongly connected events induced by behavior. The last two factors are and , which are important components of the signal quality control mechanism. SNR, that is, the signal-to-noise ratio, is a key metric for measuring the proportion of effective neural activity in a signal. The threshold is a preset constant used to determine whether the current signal is of analytical value. If the signal-to-noise ratio of a certain channel is lower than this threshold, the connections constructed by it will be regarded as unreliable and thus excluded or assigned a value of 0. This control measure can effectively exclude false connection enhancements caused by electromyogram artifacts, eye movement interference, and environmental noise in practical applications, thus ensuring the physiological credibility and behavioral interpretability of the connection network structure.

[0044] The entire integral expression constructs It is a dynamic network connection weight with the ability of multi-source information fusion, which not only has smoothness and continuity on the time axis, but also integrates multiple dimensions such as frequency-domain coupling, phase consistency, amplitude intensity, spectral concentration, and signal quality control at the physical quantity level. The design of this connection weight follows the multi-scale coupling mechanism of neural network connections. Instead of relying on a single feature for EEG recognition, it uses a dynamic feature combination method to model behavioral responses. By using this weight definition, the weighted coherence network constructed by the present invention can accurately reflect the brain functional connection pattern in the current behavioral state, and provide a high-resolution structural basis for subsequent network topology index extraction, dynamic evolution trend modeling, and behavioral recognition feature generation. Looking at the entire recognition process, this weight formula performs connection modeling on the cleaned beta wave signals after band selection and signal filtering in the preprocessing stage, and continuously updates the connection value sequence in a sliding window manner during the dynamic network construction stage to form a time-varying coherence network structure sequence. This structure sequence becomes an important input source for behavioral feature mapping, and fuses with modules such as power spectrum parameters and evolution indicators to form a unified recognition feature vector. Finally, with the support of the recognition threshold mechanism, stable recognition of movement behaviors such as eye movement, hand movement, and foot movement in a complex EEG background is achieved.

[0045] Further, the dynamic evolution index at time is: : ; where is the clustering coefficient of the node corresponding to the th electrode at time ; is the degree of the node corresponding to the th electrode at time ;

[0046] This expression indexes all nodes in the network (i.e., EEG channels), summarizes the relationship between the degree of topological change and signal quality of each node at the moment of time , so as to form a macroscopic dynamic estimation of the entire network. Specifically, this index consists of two parts: the first half measures the rate of change of the network structure over time, and the second half regulates the uncertainty impact brought by signal noise. First, focus on in the expression, which represents the degree of the node corresponding to electrode at time . Node degree is a basic quantity in network topology analysis, indicating how many other nodes the node has functional connections with at the current time. In the EEG coherence network, a higher degree usually indicates that the brain region is in a relatively central position and participates in multiple functional paths. Take the first-order derivative of with respect to time and take the absolute value, that is , representing the rate of change of the number of connections of this node over time. If this derivative is large, it indicates that the brain region connection structure has changed drastically, which may mean the activation or deactivation of the brain regions related to the current behavior, and has a high behavioral relevance. The absolute value processing is adopted to unify the change range when the degree increases and decreases, so as to capture the sudden structural evolution, rather than the linear trend in a single direction.

[0047] Then look at , this variable represents the node 's clustering coefficient, which is used to describe the tightness of the local subgraph where this node is located in the current network structure. A high clustering coefficient indicates that its adjacent nodes are interconnected, forming a compact local network cluster. Different from the node degree, the clustering coefficient reflects more local integration and stability. During the behavioral task, some brain regions will not only increase the number of their connections, but also form a highly organized modular structure through local reconstruction. For Take the second derivative with respect to time and take the absolute value, that is , used to quantify the "acceleration" of the change in the local organizational structure of this node. Compared with the first derivative, the second derivative can better reflect the degree of mutation of the structural change. Especially at the moment of behavior initiation, the behavior transition period, and the cognitive load regulation stage, this index shows a significant response, which helps to distinguish short-term behavior induction from long-term trend changes. After adding the above two items, it constitutes an overall assessment of the topological change trend of each node at time : It includes both the change speed of the number of connections and the degree of violent fluctuations in the local network structure. In theory, the larger this weighted term, the more likely it indicates that this node is in a dynamically active state, its connection network is undergoing large structural changes, and it is very likely to be in the behavior-related activity chain.

[0048] The structural dynamic response of each above-mentioned node needs to be combined with its signal quality in order to comprehensively evaluate its effectiveness. For this reason, the method introduces a signal-to-noise ratio correction term , used to penalize those nodes with too low signal-to-noise ratio. The function makes the correction value smaller when the signal-to-noise ratio is high, and the contribution weight to the overall dynamic index is larger; while for nodes with low signal-to-noise ratio, the value of this item is larger, resulting in a decrease in the weight of its topological change. This mechanism avoids misjudging the structural changes of nodes with unreliable sources as real behavior-induced responses, thereby improving the accuracy and robustness of the overall recognition. Introducing a logarithmic function instead of linear scaling also helps to control the degree of suppression of extremely large SNR values and enhance numerical stability. Finally, by adding the structural change rate of the nodes corresponding to all electrodes to their contributions after signal-to-noise ratio correction, the dynamic evolution index of the entire coherence network at time The larger the value of this indicator, the more likely it indicates that the network is currently in a period of intense restructuring, which is very likely to be the time node for the initiation or conversion of behavior. Conversely, if the value remains at a low level, it indicates that the current network structure is stable and the behavior has not occurred or is in a non-activated period.

[0049] Furthermore, at time the characteristic response index is : ; Among them, is the phase difference between the phase of the electroencephalogram signal of electrode at time and the phase of the electroencephalogram signal of electrode at time ; is the spatial Euclidean distance between electrode and electrode ; is the spatial attenuation scale coefficient. When is in centimeters, its value ranges from 2 to 5; is the total number of electrodes.

[0050] The external normalization term is to ensure that this index does not lose comparability due to changes in the number of electrodes. In a fully connected network, any electrode channels can form undirected connection pairs. After summing over all connection pairs and performing normalization averaging, it is ensured that has consistent physical meanings in different subjects and different electroencephalogram systems. The innermost summation term is a combined calculation for all electrode pairs , where means excluding self-connections. This summation process comprehensively incorporates the connection structure in the entire electroencephalogram coherence network into the response modeling perspective, thereby constructing an overall whole-brain response index.

[0051] In the summation term, the first part is , which represents the network connection weight between electrode and at time . This weight already incorporates coherence value, spectral entropy, amplitude, power density, phase consistency, and signal-to-noise ratio regulation factors, reflecting the strength and reliability of the connection. It is a prerequisite for whether the connection pair can have an actual effect on the behavioral response. Introducing this term in the characteristic response index can ensure that only when there is an actual neural coupling relationship, the corresponding electrode pair contributes weight to the network behavioral response. The second term is , which represents the connection between electrode and At time The instantaneous phase difference reflects the relative timing differences in neural activities between brain regions. In neuroscience, the phase difference is an important indicator for measuring whether neuron clusters are in a synchronous or asynchronous state. When the phase difference between two brain regions is small, it means that their signal oscillations tend to be synchronous, indicating that they may be in a co-activated state. When the phase difference increases significantly, the synchrony decreases and the cooperation may weaken. Therefore, the degree of network cooperation is introduced through the absolute value of the phase difference, enhancing the sensitivity to synchronous responses induced by behavior. It should be noted that the cosine term is not used, but the absolute value is retained, emphasizing the amplitude rather than the direction of the phase difference, which is suitable for cooperative modeling in the non-linear state.

[0052] The third term is the spatial modulation factor , where represents the electrode and The Euclidean distance between them, in centimeters. is the spatial decay scale coefficient, whose value ranges from 2 to 5. This term reflects the spatial propagation delay and coupling decay mechanism between neural activities. In the cerebral cortex, it is easier to establish synchronous activation patterns between adjacent regions, while for regions far apart, even if there is coherence, the synchronous ability is often weakened due to structural limitations. Therefore, an exponential decay function is introduced to reasonably suppress the response contribution of long-range connections. This term ensures that the model has a preferential response when identifying behavior-related local synchronous clusters, while retaining the ability to handle certain long-range coupling patterns, but only adjusting the contribution intensity exponentially. The last multiplicative factor is , that is, the dynamic evolution index at time . This index has been defined in detail in the previous section. It reflects the rate of structural change and instability of the entire EEG network at this moment. Introducing this term in this formula is to emphasize that only when the brain network shows a strong evolutionary trend in the macroscopic structure, local connections and phase responses have practical behavioral significance. If the network as a whole is in a static state, even if some local connection indicators are high, they may just be background noise or stable fluctuations, so their contributions must be adjusted according to the global dynamics. By introducing this term, becomes a composite response index that integrates network structure information, phase cooperation relationship, spatial topology, and macroscopic dynamic trend. Physically, the magnitude of reflects whether the current brain as a whole is in a high-response state: when the connection strength is high, the phase difference is significant, the network evolution is intense, and the spatial coupling structure is active, this value will increase significantly. Conversely, when the connection network is stable, the signal phase change is weak, or the network structure is stable, this value tends to be lower. In particular, for behavior recognition tasks, can be used to determine whether the behavior is in an activated state, or further for calculating the behavior recognition value and judging the classification boundary.

[0053] Furthermore, at time the clustering coefficient of the node corresponding to the th electrode is : ; wherein, is the actual number of edges existing between the neighbors of the node corresponding to the th electrode.

[0054] In this expression, represents the degree of the node corresponding to the th electrode at time , that is, the number of other electrodes with which this electrode has non-zero connection weights in the EEG coherence network; represents the actual number of edges existing between these adjacent nodes. That is to say, for all possible node pairs in the neighbor set of node , if there is an effective connection (i.e., non-zero weight) between node and node , then it is counted in . Therefore, is an index representing the tightness of the connection relationship between neighbors, while the denominator is the maximum possible number of edges that could exist between neighbors in theory. Thus, the given by this formula is a normalized ratio, with a value range between 0 and 1.

[0055] When , it indicates that there are mutual connections between all the neighbor nodes of this node, forming a complete clique structure, that is, each node connected to node is also connected to each other, forming a closed local functional cluster. This usually means that this brain region is in a highly integrated state at this moment and may be the central area for behavior-related neural collaborative processing. And when , it indicates that there are no connections between the neighbors of node , and the local structure is a radially loose connection, which may be just an information pathway rather than an actual processing center.

[0056] In the present invention, the introduction of the clustering coefficient has multiple functional significances. First, in the definition of the dynamic evolution index , the second derivative It is used as a measure of local network structure mutation. That is to say, when the connection relationship among the neighbors of a certain node changes drastically in a short period of time, especially when it suddenly changes from a loose state to a highly clustered state, or vice versa, this item will rise rapidly and significantly affect the dynamic index of the overall network. Such local mutations are often typical features when neural activity transitions from the preparatory period to the execution period, or when external stimuli are introduced, and have clear physiological behavior correlations.

[0057] Secondly, the clustering coefficient is also involved in the spatial topological regulation of the characteristic response index of the entire network. The coherence between brain regions depends not only on the weight and phase relationship between each pair of them, but also highly depends on the local aggregation pattern of the network structure. For example, eye movement behavior often activates short-path clusters between the prefrontal lobe and the occipital lobe, manual behavior activates high-density cluster structures around the central sulcus, and foot movement behavior involves long-distance connections across hemispheres but with a lower degree of clustering. As a local modular connection feature, the clustering coefficient is an important geometric topological quantity for characterizing the differences in specific connection maps of these behaviors.

[0058] Furthermore, in the process of network visualization and graph theory modeling, the clustering coefficient can also be used as an important basis for selecting the connection sparsification threshold and adjusting the feature boundary. In practical applications, since the EEG connection network is naturally a sparse graph, the connections between its nodes are not evenly distributed, but there are a small number of dense connection centers and a large number of edge nodes. By analyzing the changing trend of the clustering coefficient of nodes, key modular regions under behavior activation can be identified, and then used for neuroengineering tasks such as electrode selection, spatial channel reconstruction, or regional function assignment.

[0059] It should be noted that to avoid the situation where the denominator is zero in actual calculations, the system will skip the clustering coefficient calculation or assign a zero value to nodes with degree during the calculation process. This is because when the node degree is 0 or 1, there is no possibility of connection between its neighbors, and the theoretical maximum number of edges is 0, so the formula no longer has mathematical meaning. Therefore, the clustering coefficient is mainly applicable to nodes with a degree greater than or equal to 2.

[0060] In engineering implementation, the calculation of the clustering coefficient depends on the binarization or sparsification of the network connection weights. Although the original is a continuous value, in order to achieve edge counting, it is often necessary to set a connection strength threshold to judge all connections above a certain set value as "existing connections", and the rest as 0. This threshold can be adaptively set by statistical methods, such as retaining the top ten percent of the maximum values in the connection matrix, or can also be set as a tuning parameter during the training stage of the recognition model. In actual operation, the combined analysis of connection strength and neighbor topological density can significantly improve the expressiveness of recognition features.

[0061] Furthermore, time The target behavior recognition value at time is: ; where is time at which the electrode in the band frequency range, the integrated value of the power spectral density, , is the frequency integration variable; is the average value of the power spectrum of electrode 𝑖 in the resting state; is the electrode in the resting state, the standard deviation of the power spectrum; is time at which the spectral entropy of the electroencephalogram signal of electrode in the β band frequency range; is time at which the entropy value of the power spectrum of electrode ; is the average value of the entropy values of the power spectra of all electrodes in the resting state.

[0062] In the electroencephalogram-driven behavior recognition method based on the dynamic evolution of coherent network features, the target behavior recognition value is the core judgment index running through the entire recognition system. Its purpose is to use the coupling relationship between the structural network response parameters extracted from the electroencephalogram signal and the frequency-domain dynamic characteristics to judge whether the brain is in an activated state corresponding to a specific target behavior at a certain time point . The design logic of this index integrates features from multiple dimensions such as neural network connection strength, phase synchronization, spectral energy change, and spectral complexity reorganization to achieve precise modeling and quantitative judgment of behavior states in a complex physiological signal environment.

[0063] The entire formula consists of two parts: the pre-factor represents the characteristic response intensity of the brain functional connection network at the current moment and is an index to measure the overall brain activity level from the structural dynamic level; the post-factor is the weighted average of the spectral responses of multiple-channel electrodes, reflecting the activation state of brain regions depicted from the frequency domain. The design of this expression not only realizes the unified integration of time domain, frequency domain, spatial domain, and network topology information but also has a clear physiological interpretation path and engineering implementation logic.

[0064] The pre-factor has been derived and defined in detail in the previous section. It is an integration of the network connection strength , phase difference , spatial distance weight , and network evolution index The high-order coupling index, whose value reflects the global responsiveness level of the brain network at the current time point. If this value is close to 0, it means that the current brain network as a whole is in a resting or low-active state. Even if there is a local power increase in some electrodes, it does not have sufficient behavioral explanatory power. On the contrary, when this value rises, it indicates that the brain as a whole is in a state of linkage, high synchronization, and high evolution speed, and it is appropriate to determine that the neural activity related to the target behavior is ongoing. Therefore, among the entire recognition values, plays a role in dynamic gating and response regulation.

[0065] The latter factor part is the frequency-domain composite response intensity averaged across electrodes, and inside is the product of the three spectral feature factors of the electrode at the current moment. The first factor is: ; This expression subtracts the total power of the electrode in the wave band at the current time point from the average value of its power spectrum in the resting state , and normalizes it with the standard deviation in the resting state. The in the formula is defined as: : ; Among them is the power spectral density of the electrode at time and frequency . The integration interval is the β wave band (13 - 30 Hz). The physical meaning of this normalization term is to measure whether the power of the current electrode significantly deviates from its resting baseline state. If the current power spectral density is significantly higher than the resting mean, this term is positive, indicating that the electrode may be in a behavior-induced activation state; if it is lower than the baseline, it is negative, indicating that there is no significant activity sign in this brain region; the standard deviation is used to eliminate the intensity difference of individual EEG signals and achieve a consistent evaluation across subjects and electrode channels.

[0066] The second factor is: ; Among them, represents the electrode at time , The spectral entropy in the wave frequency band is a measure of the complexity of the energy distribution of a signal in the frequency domain. The larger the spectral entropy, the more uniform the power distribution and the more complex the frequency components, indicating that the signal may not have a frequency-based behavioral feature focus; conversely, the smaller the entropy value, the more concentrated the frequency distribution, which may correspond to enhanced neural discharges related to the behavior in a certain frequency band. This factor uses a non-linear transformation method to enhance low to medium entropy values and suppress high entropy values, thereby strengthening the contribution of electrodes with obvious frequency focus characteristics. Adding is to avoid logarithmic divergence or division by zero problems in the denominator when the entropy value approaches 0, which is a typical numerical stability processing method.

[0067] The third factor is: ; where represents the entropy value of the power spectrum of electrode at the current moment, while is the average value of the power spectrum entropy of all electrodes in the resting state. This factor uses the Sigmoid function to map the change in power spectrum entropy to a continuous value between 0 and 1. When the spectral entropy of electrode is higher than the average level in the resting state, this factor approaches 1, indicating that a significant change has occurred in the spectral structure, which may be related to behavior execution; when it is lower than the resting entropy level, the Sigmoid value approaches 0, representing that the current spectral entropy has not changed significantly and should not be regarded as a behavior activation feature. The Sigmoid function provides a flexible boundary and non-linear discrimination ability, which has better stability and differentiability compared to hard thresholds, and helps the subsequent application of behavior recognition values in regression or classification models.

[0068] Multiply the above three factors, that is: ; This product constitutes the comprehensive frequency domain activation feature value of electrode in response to the target behavior. Its physical meaning is: when the power of a certain electrode rises significantly at the current moment, the spectrum shows a structural concentration (low entropy) or entropy value reorganization (higher than the benchmark), then this electrode has a higher contribution value of behavioral characteristics at that moment.

[0069] Take the average of the characteristic contribution values of all electrodes, that is: ; This operation reflects that behavior activation is a systematic phenomenon involving multi-region cooperation, rather than the isolated activation behavior of individual brain regions. By averaging across electrodes, local false activations can be reduced to mislead the overall recognition, while also retaining the characteristics of extensive responses in behavior-related brain regions.

[0070] Finally, compare this average result with the structural response index Multiply to form a complete behavior recognition value This value has the following characteristics: the larger its value, the stronger the brain structure connection, the higher the phase coupling, and the more active the network evolution at the current moment. At the same time, there is an increase in the power of multiple electrodes, spectral focusing, or spectral entropy reorganization in the frequency domain, and it is very likely to be in the target behavior activation state. On the contrary, if the structural response is low or the spectral characteristics are not significant, even if a certain parameter is large, it cannot significantly increase the recognition value, thus maintaining the stability and noise resistance of the recognition system.

[0071] Recognition value Can be used for real-time comparison with the behavior activation threshold For online behavior judgment. Its continuous output characteristic also allows it to be used as a regression variable to measure the behavior response intensity or to construct a behavior boundary detection model based on the time series of the recognition value.

[0072] Furthermore, when the target behavior is an eye movement behavior, the value range of the recognition threshold Is from 0.20 to 0.30; when the target behavior is a manual behavior, the value range of the recognition threshold Is from 0.35 to 0.45; when the target behavior is a foot behavior, the value range of the recognition threshold Is from 0.50 to 0.60; if Is greater than the recognition threshold Then it is judged that the target behavior is in the activation state.

[0073] When the target behavior is an eye movement behavior, the system sets the value range of its recognition threshold To be from 0.30 to 0.45. This is based on the analysis results of the electroencephalogram signal characteristics caused by eye movement behavior during execution. Eye movement usually causes a more localized network response in the prefrontal to occipital region. Although the characteristic response index Increases during this process, its spectral energy change and spectral entropy reconstruction degree are lower than those of manual or foot movement behaviors. Therefore, the overall recognition value The distribution center of gravity under eye movement behavior is usually lower than that of other behaviors. According to experimental statistics, when the recognition value is below 0.30, it is mostly in the non-eye movement state, and there is a high probability that it is in the eye movement activation area between 0.30 and 0.45. Therefore, this range is set as the recognition threshold interval for eye movement behavior. Finally, if The system judges that the brain is in the eye movement behavior activation state at this time.

[0074] When the target behavior is a foot movement behavior, its recognition threshold The value range of is from 0.50 to 0.60. Foot movement behavior, due to the involvement of subcortical motor areas and more extensive cross-hemispheric functional area coordination, is usually accompanied by higher characteristic response indices and enhanced power spectrum features. Especially in the central and parietal regions, the degree of spectral energy and spectral entropy reconstruction is significantly enhanced. Therefore the mean and variance of

[0075] are both higher than those of eye movement or hand movement behavior. According to the interval analysis of the statistical learning model, the distribution of the recognition values corresponding to foot movement behavior usually concentrates between 0.52 and 0.58. Based on this, the system sets 0.50 as the minimum judgment threshold and 0.60 as the upper limit stable interval boundary to ensure recognition accuracy and stability. Therefore, during the recognition process, if the current recognition value If the target behavior is eye movement behavior, set the threshold When it is determined that the eye movement behavior is in an active state; If the target behavior is foot movement behavior, set the threshold When it is determined that the foot movement behavior is in an active state.

[0076] It should be noted that the recognition threshold corresponding to each behavior is not a fixed single value, but is selected from a reasonable interval verified by physiological experiments and can be further optimized according to the subject individual parameters, signal sampling conditions or recognition model training results. In system implementation, the threshold selection can also be dynamically adaptively adjusted through cross-validation to meet the specific discrimination requirements under different subjects or task types. Finally, this behavior activation determination mechanism based on discriminative behavior recognition values enables this method to have a clear and continuous behavior recognition logic and can form non-overlapping recognition boundaries among multiple behavior types, thus ensuring high-precision and mutual exclusivity of behavior determination and providing a stable threshold control strategy for the multi-class motor EEG recognition system.

[0077] The following provides a complete implementation example of an EEG-driven behavior recognition method based on the dynamic evolution of coherent network features, including raw EEG data parameters, processing procedures at each stage, formula applications, and final behavior recognition results. This example takes eye movement behavior as the recognition target and selects the wave frequency band (13–30 Hz) for analysis.

[0078] Number of electrodes: , taken from Fp1, Fp2, C3, C4, P3, P4, O1, O2 of the standard 10-20 system. Sampling rate: 256 Hz; Filter bandpass range: 13 Hz to 30 Hz (β wave); Time window , sliding step size: 0.25 s; current analysis time point: ; signal-to-noise ratio threshold: .

[0079] Taking electrode , as an example, calculate the coherence value in the frequency band: ; Spectral entropy (normalized) of electrodes C3 and C4: ; Instantaneous voltage amplitude (unit: ): ; Power spectral density at the center frequency: ; Phase difference: ; Signal-to-noise ratio: ; Spatial distance (scalp electrode): , take ; ; For electrodes, construct a network, and calculate a total of pairs, take partial calculations: set the average connection weight ; average phase difference ; spatial distance , take the average value of 5; dynamic evolution index ; ; ; ; Measured band total power: ; ; ; Product of the three: ; Set after averaging over 8 electrodes: ; ; The current recognition target is eye movement behavior; Threshold setting range: ; Current , meeting the recognition activation condition.

[0080] Therefore, the system judges that the eye movement behavior has been activated.

[0081] Although the specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are only illustrative. Without departing from the principles and essence of the present invention, those skilled in the art can make various omissions, substitutions, and changes to the details of the above methods and systems. For example, combining the above method steps so as to perform substantially the same function in a substantially the same way to achieve substantially the same result falls within the scope of the present invention. Therefore, the scope of the present invention is only defined by the appended claims.

Claims

1. An electroencephalogram-driven behavior recognition method based on the dynamic evolution of coherent network features, characterized in that The method includes: Step 1: Perform standard preprocessing on the original EEG signals, select the β-wave frequency band as the analysis frequency band, and use short-time wavelet transform to extract the instantaneous frequency components of the analysis frequency band; for each pair of electrodes, within a specified time window, calculate the coherence value in the analysis frequency band to measure the synchronization strength of the two EEG signals, which is used as the network connection weight, and construct a weighted coherence network; Step 2: For each time point, extract the degree and clustering coefficient of each node in the weighted coherence network, and calculate the dynamic evolution index; combine the network connection weight with the dynamic evolution index to construct a unified EEG-driven behavior feature map; at each time point, for each pair of nodes, calculate their instantaneous phase difference, and at the same time introduce the spatial geometric distance between the two electrodes to obtain the behavior feature response index; Step 3: Obtain the power spectrum parameters of all brain region signals of the target behavior in the analysis frequency band, combine them with the feature response index, calculate the target behavior recognition value, set the behavior threshold of the target behavior in the analysis frequency band, compare the target behavior recognition value with the behavior threshold, and determine whether the target behavior is activated.

2. The method for identifying EEG-driven behaviors based on the dynamic evolution of coherent network features according to claim 1, characterized in that, After selecting the β-wave frequency band as the analysis frequency band, use a band-pass filter to process the original EEG signals, and the frequency range of the band-pass filter is from 13 Hz to 30 Hz; the target behavior is movement-related behavior, including at least: eye movement behavior, manual behavior, and foot movement behavior.

3. The electroencephalogram-driven behavior recognition method based on the dynamic evolution of coherent network features according to claim 2, wherein The power spectrum parameters include: the average power spectrum value in the resting state, the standard deviation of the power spectrum in the resting state, the entropy value of the power spectrum, and the average value of the entropy value of the power spectrum.

4. The electroencephalogram-driven behavior recognition method based on the dynamic evolution of coherent network features according to claim 3, characterized in that The network connection weight is: ； Among them, represents the network connection weight between the electroencephalogram (EEG) signal of electrode and the EEG signal of electrode at time ; is the time window; is the time integration variable; is the coherence value between the EEG signal of electrode in the β-wave frequency band and the EEG signal of electrode at time ; is the spectral entropy of the EEG signal of electrode in the β-wave frequency band; is the spectral entropy of the EEG signal of electrode in the β-wave frequency band; is the voltage amplitude of the EEG signal of electrode ; is the voltage amplitude of the EEG signal of electrode ; is the power spectral density of the EEG signal of electrode in the β-wave frequency band at the center frequency of the β-wave frequency band at time ; is the power spectral density of the EEG signal of electrode in the β-wave frequency band at the center frequency of the β-wave frequency band at time ; is the phase of the EEG signal of electrode ; is the phase of the EEG signal of electrode ; is the indicator function, which has a value of 1 when the condition is met and 0 otherwise; is the signal-to-noise ratio (SNR) of the EEG signal of electrode ; is the set SNR threshold; is the signal-to-noise ratio of the EEG signal of electrode .

5. The method for electroencephalogram-driven behavior recognition based on the dynamic evolution of coherent network features according to claim 4, wherein Time Dynamic evolution index at is as follows: ； wherein, is the time at which the clustering coefficient of the node corresponding to the is the time at which the degree of the node corresponding to the 6. The electroencephalogram-driven behavior recognition method based on the dynamic evolution of coherent network features as claimed in claim 5, wherein Time The characteristic response index at is ； Among them, is the time when the phase of the electroencephalogram signal of the electrode and the phase difference between the phase of the electroencephalogram signal of the electrode at time ; is the spatial Euclidean distance between the electrode and the electrode ; is the spatial attenuation scale coefficient. When is in centimeters, its value ranges from 2 to 5; is the total number of electrodes.

7. The electroencephalogram-driven behavior recognition method based on the dynamic evolution of coherent network features according to claim 6, characterized in that Time At the clustering coefficient of the node corresponding to the electrode is: ； Among them, is the actual number of edges existing between neighbors of the node corresponding to the th electrode.

8. The method for identifying EEG-driven behavior based on the dynamic evolution of coherent network features according to claim 7, characterized in that, Time Target behavior recognition value at is as follows: ； Among them, is the time when the power spectral density integral value of the electrode at the wave frequency band, , and is the average power spectrum of electrode 𝑖 in the resting state; is the standard deviation of the power spectrum of electrode in the resting state; is the time when the spectral entropy of the electroencephalogram signal of electrode at the β wave frequency band; is the time when the entropy value of the power spectrum of electrode ; is the average value of the entropy values of the power spectra of all electrodes in the resting state.

9. The electroencephalogram-driven behavior recognition method based on the dynamic evolution of coherent network features as claimed in claim 8, wherein When the target behavior is an eye movement behavior, the recognition threshold ranges from 0.20 to 0.30; when the target behavior is a manual behavior, the recognition threshold ranges from 0.35 to 0.45; when the target behavior is a foot behavior, the recognition threshold ranges from 0.50 to 0.60; if is greater than the recognition threshold , it is determined that the target behavior is in an active state.

Images