Intention recognition feedback interaction control method, system and device based on electroencephalogram signals

Abstract

The application discloses an intention recognition feedback interaction control method, system and device based on electroencephalogram signals, and relates to the technical field of brain-computer interaction control.The application acquires real-time continuous electroencephalogram signals, performs segmented denoising based on a sliding time window, and extracts a current multi-dimensional feature vector; the feature vector is input into a classification model to obtain a candidate classification state, and a historical time window intention classification state sequence is combined to perform timing smoothing and state transition analysis, so that a stable target intention is output and a historical state cache is updated; the stable target intention is converted into a control instruction to drive an external device, real-time running physical parameters in a motion execution process are acquired; the real-time running physical parameters are converted into an interaction feedback signal containing a motion deviation prompt to be presented, and a closed-loop physical stimulation is formed.The application effectively filters out classification decision jumps in an intention conversion boundary region, and significantly improves the timing consistency of continuous interaction control instructions and the robustness of a man-machine control closed loop.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface control technology, specifically to an intention recognition feedback interactive control method, system, and device based on electroencephalogram (EEG) signals. Background Technology

[0002] Brain-computer interfaces (BCIs), as a cutting-edge human-computer interaction paradigm, can bypass peripheral nerves and body tissues to directly establish a communication and control link between the human brain and external electromechanical devices. Electroencephalography (EEG) is the mainstream non-invasive brainwave acquisition technology. With the iteration of hardware acquisition equipment and the optimization of artificial intelligence decoding algorithms, EEG-based motion intention recognition technology has completed multi-scenario verification in the laboratory and has the potential for industrial transformation in fields such as disability rehabilitation, smart homes, special robot control, and drone flight assistance guidance.

[0003] The standardized workflow of existing conventional EEG control systems is as follows: Weak EEG signals from the scalp are collected using wearable EEG acquisition devices (EEG caps, lightweight wireless headbands, etc.), transmitted to a backend processing unit for time-series segmentation, noise suppression, feature extraction, and machine learning pattern classification. Finally, the single-window recognition result is mapped into control commands, which are then sent down to execution terminals such as androids, robotic arms, and drones. However, existing EEG signal intention recognition methods typically use fixed time windows to preprocess and extract features from continuously acquired EEG signals, classifying the intention based on the EEG features within the current time window. However, in practical applications, there is a transition phase between the generation and stable expression of user intention, especially during intention switching. Adjacent time windows often simultaneously contain residual features from the previous intention and initial features of the new intention. Since most existing feature classification models rely solely on features extracted within the current time window for independent discrimination, they lack effective utilization of the continuity of intention evolution and historical recognition states. This leads to EEG features in the intention transition phase easily falling into the boundary regions of different feature categories.

[0004] Currently, EEG intent recognition algorithms generally use fixed-length sliding windows for signal preprocessing and feature extraction, relying solely on EEG features within a single window to independently determine intent. However, in real-world interaction scenarios, there is a transition period from the emergence of a user's intent to its stable output; within the time window of the intent switching phase, residual features of the previous intent and initial features of the new intent are usually retained simultaneously. Traditional classification models only make independent distinctions based on single-window features, failing to explore the temporal evolution of intent and historical recognition states, and are prone to classifying mixed features from the transition phase into the multi-class decision boundary.

[0005] In this situation, even if the classifier output meets the preset classification conditions, the recognition results may frequently change within a continuous time window. For example, if a user is initially expressing the intention to "move left," the classification result may fluctuate repeatedly between "move left," "stop," and "invalid state" when preparing to switch to the intention to "stop." Or, if the user maintains the same intention continuously, minor fluctuations in the EEG signal may cause the classification result to switch incorrectly within a short period of time. This phenomenon will reduce the stability of the subsequently generated control commands.

[0006] Therefore, this invention proposes an intention recognition feedback interactive control method, system, and device based on electroencephalogram (EEG) signals. Summary of the Invention

[0007] The purpose of this invention is to provide an intention recognition feedback interactive control method, system, and device based on electroencephalogram (EEG) signals to solve the problems mentioned in the background art.

[0008] According to a first aspect of the present invention, in order to achieve the above-mentioned objective, the present invention provides the following technical solution: a method for intention recognition feedback interactive control based on electroencephalogram (EEG) signals, comprising the following steps: The raw continuous EEG signal is acquired, and the raw continuous EEG signal is segmented and denoised based on a preset sliding time window to obtain a high signal-to-noise ratio EEG signal sequence within the current time window. Feature extraction is performed on the high signal-to-noise ratio EEG signal sequence to obtain a current multidimensional feature vector representing the current EEG state; The current multidimensional feature vector is input into the pre-constructed intent classification model to obtain the current candidate classification state. The historical time window intent classification state sequence recorded in the local cache is extracted. The current candidate classification state and the historical time window intent classification state sequence are subjected to joint temporal smoothing and state transition analysis to obtain a stable target intent after temporal continuity calibration. The stable target intent is written into the local cache to update the historical time window intent classification state sequence. The stable target intent is converted into control commands according to the preset mapping rules, and the control commands are sent to external devices to drive them to perform actions. The real-time physical parameters of the external devices during the action execution process are obtained as real-time running status data. The real-time operating status data is converted into an interactive feedback signal that includes indications of device action deviations, which is used to correct subsequent determinations of stable target intent, thereby forming a closed-loop feedback interactive control.

[0009] Furthermore, the original continuous EEG signal is segmented and preprocessed for noise reduction based on a preset sliding time window to obtain a high signal-to-noise ratio EEG signal sequence within the current time window, specifically including: Based on the preset sliding time window length and sliding step size, the original continuous EEG signal is truncated and segmented to obtain the original EEG data matrix within the current time window; The original EEG data matrix is ​​preprocessed by temporal denoising using a finite impulse response filter with a preset passband range, and the temporal filtered EEG data matrix is ​​output. Using the time-domain filtered EEG data matrix as input, a common average reference algorithm is used for spatial denoising preprocessing. The high signal-to-noise ratio EEG signal sequence within the current time window is obtained by calculating the average potential of all channels and subtracting it from the potential of the corresponding target channel.

[0010] Furthermore, feature extraction is performed on the high signal-to-noise ratio EEG signal sequence to obtain a current multidimensional feature vector representing the current EEG state, specifically including: The high signal-to-noise ratio EEG signal sequence is input into a pre-trained spatial filter for spatial feature projection to obtain a spatial filter signal matrix. Calculate the signal energy variance of each spatial filtering channel in the spatial filtering signal matrix within the current time window to obtain the spatial energy feature vector of the current time window; The spatial energy feature vector is logarithmically normalized to obtain the current multidimensional feature vector representing the current EEG state.

[0011] Further, the current multidimensional feature vector is input into a preset intent classification model to obtain the current candidate classification state; the historical time window intent classification state sequence recorded in the local cache is extracted, and the current candidate classification state and the historical time window intent classification state sequence are subjected to joint temporal smoothing and state transition analysis to obtain a stable target intent after temporal continuity calibration, specifically including: The current multidimensional feature vector is input into a preset intent classification model, and the output confidence of each intent category corresponding to the current time window is calculated to obtain the current candidate classification state vector. Extract the historical smooth probability vector written in the previous time window from the local cache as the historical time window intent classification state sequence, and extract the preset intent state transition matrix; based on the preset temporal smoothing coefficient, fuse the current candidate classification state vector, the historical smooth probability vector and the intent state transition matrix to obtain the continuity calibration probability vector of the current time window. The maximum value of each calibration probability in the continuous calibration probability vector is optimized, and the intent category corresponding to the maximum calibration probability is determined as the stationary target intent after time-series continuous calibration. After extracting the stable target intent, the continuity calibration probability vector of the current time window is written into the local cache, overwriting the current historical smoothing probability vector.

[0012] Furthermore, the construction and training process of the intent classification model is as follows: Multiple sets of offline raw continuous EEG signals and corresponding real intention labels of users under preset offline interactive tasks are collected, and the offline raw continuous EEG signals are segmented, denoised, preprocessed and feature extracted to construct a training set containing multiple offline samples. The basic network structure of the intent classification model is defined as a single-layer fully connected Softmax mapping network. The loss function of the intent classification model is constructed based on the cross-entropy function and by introducing a regularization penalty term. The gradient descent optimization algorithm is used to minimize the loss function. The feature mapping weight matrix and decision bias vector of the intent classification model are iteratively updated within a preset number of iterations until the preset convergence condition is met. The converged network parameters are then output and configured into the intent classification model.

[0013] Furthermore, the intended smooth target is converted into control commands according to a preset mapping rule, and the control commands are sent to external devices to drive them to perform actions. The real-time physical parameters of the external devices during the action execution process are obtained as real-time operating status data, specifically including: Based on the stated stable target intent, the corresponding baseline control command vector is extracted from the preset control command mapping dictionary; The reference control command vector is smoothed using a first-order low-pass dynamic filtering algorithm to generate a smoothed control command vector for the current time window, and the smoothed control command vector is sent to an external device. The real-time physical parameter vector of the external device during the action execution process is collected in real time. The difference analysis between the real-time physical parameter vector and the smooth control command vector is performed to calculate the action execution deviation vector. The real-time running physical parameter vector and the action execution deviation vector are concatenated to construct the real-time running status data for the current time window.

[0014] Furthermore, the real-time operating status data is converted into an interactive feedback signal that includes indications of device action deviations, specifically including: The real-time running status data is analyzed, the action execution deviation vector is extracted, and the action execution deviation vector is evaluated and calculated using a weighted Euclidean norm based on feature sensitivity to obtain a global deviation index. The global deviation index is input into a preset nonlinear activation function to calculate the dynamic feedback adjustment factor. The real-time running physical parameter vector is extracted and synthesized with the dynamic feedback adjustment factor and the preset action deviation warning feature vector to obtain the interactive feedback signal vector of the current time window. The interactive feedback signal vector is then presented through the information feedback module of the external device.

[0015] According to a second aspect of the present invention, the present invention provides an intention recognition feedback interactive control system based on electroencephalogram (EEG) signals, for implementing the intention recognition feedback interactive control method based on EEG signals described in the first aspect, comprising: The EEG signal preprocessing module is used to acquire raw continuous EEG signals, and perform segmentation and noise reduction preprocessing on the raw continuous EEG signals based on a preset sliding time window to obtain a high signal-to-noise ratio EEG signal sequence within the current time window. The feature extraction module is used to extract features from the high signal-to-noise ratio EEG signal sequence to obtain a current multidimensional feature vector representing the current EEG state. The feature classification module is used to input the current multidimensional feature vector into a pre-built intent classification model to obtain the current candidate classification state, extract the historical time window intent classification state sequence recorded in the local cache, perform joint temporal smoothing and state transition analysis on the current candidate classification state and the historical time window intent classification state sequence to obtain a stable target intent after temporal continuity calibration, and write the stable target intent into the local cache to update the historical time window intent classification state sequence; The control command generation module is used to convert the stable target intent into control commands according to a preset mapping rule, and send the control commands to external devices to drive them to perform actions, and obtain the real-time physical parameters of the external devices during the action execution process as real-time running status data. The closed-loop control module is used to convert the real-time operating status data into an interactive feedback signal that includes equipment action deviation prompts, which is used to correct subsequent stable target intent judgments, thereby forming a closed-loop feedback interactive control.

[0016] According to a third aspect of the present invention, a terminal device is provided, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor, and when the processor loads and executes the computer program, it employs the intention recognition feedback interactive control method based on electroencephalogram (EEG) signals described in the first aspect.

[0017] According to a fourth aspect of the present invention, the present invention provides a storage medium comprising computer-executable instructions, wherein the computer-executable instructions, when executed by a computer processor, are used to perform an intention recognition feedback interactive control method based on electroencephalogram (EEG) signals as described in the first aspect.

[0018] The present invention has at least the following beneficial effects: 1. To address the problem of frequent jumps in classification results when facing intent transitions in existing technologies, this invention introduces a historical time window intent classification state sequence and performs joint temporal smoothing and state transition analysis on the current candidate classification state and historical prior state. This mechanism fully utilizes the temporal continuity of human cognitive intent evolution. When EEG features within a continuous time window are in the fuzzy boundary region of different intent categories or are disturbed by small random fluctuations, the system can generate continuous inertia based on the historical context state, effectively filtering out misjudgments and high-frequency decision jitter caused by isolated time windows. This accurately distinguishes between the user's "real intent switching" and "random fluctuations in EEG signals," ensuring the smooth output of continuous interactive control commands from the algorithm's underlying layer.

[0019] 2. In the process of mapping EEG intentions to device actions, this invention not only outputs discrete intention categories, but also further introduces a first-order low-pass dynamic filtering algorithm to smooth and mitigate the baseline control commands. This mechanism can effectively eliminate command step phenomena caused by sudden switching of user intentions or instantaneous changes in signals, prevent abrupt mechanical shocks in the drive commands issued to external devices, and ensure the smooth and rounded physical execution trajectory of external devices.

[0020] 3. This invention combines finite impulse response temporal filtering with common average reference spatial denoising. While avoiding EEG rhythm phase generation group delay distortion, it eliminates global common-mode interference artifacts and high-frequency environmental noise to the greatest extent. In addition, the spatial energy features are logarithmically normalized during the feature extraction stage, which effectively compresses the dynamic range of the feature scale and reduces the negative impact of individual user differences, fatigue levels, and EEG signal errors, making the feature distribution input to the classification model more stable.

[0021] 4. In the offline training of the intent classification model, this invention constructs a loss function based on the cross-entropy function and introduces a regularization penalty term. This mechanism applies structural constraints to the network weights during the gradient descent optimization process, effectively suppressing the phenomenon of the model being too offline that is very likely to occur during the training of small sample EEG data, and improving the generalization prediction ability of the cognitive classification model for unknown EEG data.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the process flow of the method described in this invention; Figure 2 This is a schematic diagram of the framework principle of the method described in this invention. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Example 1: Please see Figure 1 This invention provides a technical solution: an intention recognition feedback interactive control method based on electroencephalogram (EEG) signals, comprising the following steps: S1. Acquire raw continuous EEG signals, and perform segmentation and noise reduction preprocessing on the raw continuous EEG signals based on a preset sliding time window to obtain a high signal-to-noise ratio EEG signal sequence within the current time window, as follows: S1.1. Real-time acquisition via EEG acquisition equipment The raw, continuous EEG signals from each channel are represented as discrete time series. ,in For channel index and , This represents the total number of EEG acquisition channels. For continuous sampling time point indexing; based on a preset sliding time window length. and sliding step size The original continuous EEG signal was truncated and segmented to obtain the first... Raw EEG data matrix within each time window Among them, the raw EEG data matrix The formula for extracting matrix elements is: in, Represents the raw EEG data matrix The Middle The first passage is inside the window. The signal amplitude at each sampling point The time series index of the sliding time window and , Index of sampling points within the window and ; By adopting a segmentation method with sliding step size truncation, the refresh rate of the instruction output for system intent recognition can be significantly improved while ensuring that the amount of data within a single time window is sufficient to support feature extraction, thus meeting the real-time requirements of continuous control of external devices. S1.2. A finite impulse response filter with a preset passband range is used to process the raw EEG data matrix. Temporal denoising preprocessing is performed to filter out low-frequency baseline drift and high-frequency environmental noise, outputting a temporally filtered EEG data matrix. The calculation formula for temporal denoising preprocessing is as follows: in, Represents the EEG data matrix after time-domain filtering The Middle The first passage is inside the window. The filtered output amplitude at each sampling point The first finite impulse response filter represents the... Tap coefficient Let the order of the finite impulse response filter be . For tap coefficient index and ;when When this happens, a zero-fill strategy is adopted, that is, take... ; The finite impulse response filter is selected for time-domain filtering. Its function is to provide strict linear phase characteristics, which can not only completely filter out low-frequency breathing artifacts and high-frequency mains interference, but also avoid phase group delay distortion of EEG rhythms of different frequency components, thus preserving the true time-domain form of the brain power signal. S1.3. Filter the EEG data matrix in the time domain. As input, a common average reference algorithm is used for spatial domain denoising preprocessing. By calculating the average potential of all channels at the same sampling point and subtracting it from the potential of the corresponding target channel, global common-mode interference artifacts are eliminated, ultimately yielding a high signal-to-noise ratio EEG signal sequence within the current time window. The calculation formula for spatial domain denoising preprocessing is as follows: in, High signal-to-noise ratio EEG signal sequences The Middle The first passage is inside the window. The amplitude of the denoised signal at each sampling point Represents the EEG data matrix after time-domain filtering The Middle The first passage is inside the window. The filtered output amplitude at each sampling point The channel traversal index is used to calculate the average potential of all channels and ; It should be noted that the common average reference algorithm calculates and removes the common-mode components of all electrodes on the scalp surface, which can greatly suppress common-mode interference artifacts caused by reference electrode impedance fluctuations or widespread background noise in the whole brain, thereby highlighting the specific activation signals of local brain regions. S2. Feature extraction is performed on the high signal-to-noise ratio EEG signal sequence to obtain the current multidimensional feature vector representing the current EEG state, as follows: S2.1. The high signal-to-noise ratio EEG signal sequence output in step S1.3 Spatial feature projection is performed on the input pre-trained spatial filter to obtain the spatial filtered signal matrix. This is used to extract the spatial distribution features of EEG signals; the formula for calculating the spatial feature projection is: in, Represents the spatial filtering signal matrix The Middle The spatial filter channel is within the window. The signal amplitude at each sampling point; This represents the preset total number of spatial filtering channels. For spatial filter channel index and ; Represents the pre-trained spatial filter weight matrix The Middle Line 1 The weight coefficient of the column is used to characterize the weight coefficient of the first column. The EEG acquisition channel in the first Projection weights in each spatial filtering channel; For channel index and , This represents the total number of EEG acquisition channels. High signal-to-noise ratio EEG signal sequences The Middle The first passage is inside the window. The amplitude of the denoised signal at each sampling point; For the time series index of the sliding time window, Index of sampling points within the window and , The preset sliding time window length; The spatial weight matrix linearly combines multi-channel signals, which projects the signals from the original acquisition channels onto mutually orthogonal subspaces. The technical effect is to maximize the variance difference between EEG signals corresponding to different intentional tasks, thereby enhancing the separability between categories at the spatial distribution level. S2.2. Based on the spatial filtering signal matrix output in step S2.1 Calculate the signal energy variance of each spatial filter channel within the current time window to obtain the spatial energy feature vector of the current time window. Spatial energy eigenvector The Middle Energy variance corresponding to each spatial filter channel The extraction formula is: Among them, based on the prior time-domain and spatial-domain denoising processing, the spatial filtering signal matrix The average value of each channel signal is zero; S2.3. The spatial energy eigenvector output from step S2.2 Logarithmic scaling is performed to reduce the impact of individual differences and non-stationary perturbations on the feature scale, ultimately yielding the current multidimensional feature vector representing the current EEG state. Current multidimensional feature vector The Middle Feature components of each dimension The calculation formula is: in, The traversal index used to calculate the sum of energy for all spatial filter channels and , Represents the spatial energy eigenvector The Middle The energy variance corresponding to each spatial filtering channel; the current multidimensional feature vector output format is: ; Log-proportional normalization of the energy variance is used to compress the dynamic range of the feature scale, which makes the extracted feature distribution closer to the Gaussian normal distribution, satisfies the assumptions of subsequent linear or nonlinear classifiers about the data distribution characteristics, and significantly reduces the non-stationary drift of the feature scale caused by microwaves in the user's physiological state in a single experiment. S3. Input the current multidimensional feature vector into the pre-built intent classification model to obtain the current candidate classification state. Extract the historical time window intent classification state sequence recorded in the local cache. Perform joint temporal smoothing and state transition analysis on the current candidate classification state and the historical time window intent classification state sequence to obtain the stationary target intent after temporal continuity calibration. Write the stationary target intent into the local cache to update the historical time window intent classification state sequence. The specific steps are as follows: S3.1. The current multidimensional feature vector output in step S2.3 Input a pre-defined intent classification model, and use the Softmax regression function to calculate the output confidence score for each intent category in the current time window, thus obtaining the current candidate classification state vector. Current candidate classification state vector The Middle Candidate probabilities corresponding to each intent category The calculation formula is: in, The total number of intent categories preset for the system; Indexed by intent category and ; The first in the intention classification model The feature mapping weight row vectors corresponding to each intent category; Indicates the first Decision biases corresponding to each intent category; The category traversal index is used to calculate the sum of probabilities for all intent categories, and ; S3.2. Extract the historical time window intent classification state sequence written in the previous time window from the local cache, and represent it as a historical smooth probability vector. Extract the preset intention state transition matrix. The current candidate classification state vector output in step S3.1 With historical smooth probability vector By performing joint temporal smoothing and state transition analysis, the continuity calibration probability vector of the current time window is obtained. Among them, the continuity calibration probability vector The Middle Calibration probability of each intent category The calculation formula is: in, The preset timing smoothing coefficient and , used to adjust the fusion weight between the current candidate probability and the historical prior state; Indexed for historical intent categories and ; Represents the historical smoothing probability vector The Middle Historical calibration probabilities for each intent category; Represents the intention state transition matrix From the middle Switch the intent category to the first one. The prior transition probabilities of each intent category; in particular, when the temporal index of the sliding time window... At that time, initialize the historical smoothing probability vector. All elements are ; This step is the core algorithm mechanism for solving the problem of continuous control decision jitter in this embodiment. By introducing Hidden Markov state transition prior knowledge, the judgment at the current moment no longer absolutely depends on isolated single-window features. In continuous interaction scenarios, when the user's EEG intention is in the transition boundary area, the continuity inertia of the temporal context can be used to effectively filter out high-frequency decision jumps caused by weak random fluctuations in EEG, and minimize the probability of incorrect classification results, thereby greatly improving the temporal consistency of the system outputting control commands to the next level. S3.3. Continuity calibration probability vector output from step S3.2 The maximum value of each calibration probability is optimized, and the intent category corresponding to the maximum calibration probability is determined as a stationary target intent after time-series continuous calibration. Stable goals and intentions The extraction formula is: Extracting the intention of a stable target Then, the continuity calibration probability vector of the current time window is... Write it to the local cache as the time-series state input for the next time window to cover the current historical smoothing probability vector. This completes the closed-loop update of the historical time window intention classification state sequence. This closed-loop writing action can ensure that the latest generation of confidence information is always maintained in the historical state buffer queue, thus guaranteeing the coherence and self-consistency of the Markov chain time sequence deduction. It should be further explained that the specific training process for the intent classification model is as follows: Collect multiple sets of offline raw continuous EEG signals and corresponding real intention labels from users under preset offline interactive tasks, and construct a system containing... Training set of offline samples ,in, Indicates the first The training feature vector corresponding to each training sample has a data dimension that is different from the current multidimensional feature vector. completely consistent; Indicates the first The true intent label corresponding to each training sample; The total number of training samples. For training sample index and By maintaining complete isomorphism between offline feature extraction and online processing, it is possible to ensure the consistency of the distribution of trained model parameters in real-world application scenarios, and prevent performance degradation caused by the offset between online and offline data domains. An intent classification model is constructed using a single-layer fully connected Softmax mapping network as its basic architecture. The optimization parameters to be trained include the feature mapping weight matrix. With decision bias vector And based on the cross-entropy function and introducing Regularization penalty term, constructing the loss function of the intent classification model for: in, For the indicator function, when the true intent label The value equals the intent category index. hour, ,otherwise ; Represents the feature mapping weight matrix The Row vectors; Represents the decision bias vector The One component; This is the category traversal index used to calculate the sum of probabilities for all intent categories; Represents a vector beg The square value after the norm; Preset Regularization coefficients are used to suppress overfitting of model parameters; The gradient descent optimization algorithm is used to minimize the loss function. To achieve the goal, the feature mapping weight matrix is ​​adjusted within a preset number of iterations. With decision bias vector Perform iterative updates; when the loss function Training stops when the preset convergence condition is met or the maximum number of iterations is reached, and the converged feature map weight row vector is output. With decision bias This configuration is then embedded into the intent classification model for forward computation of intent categories during the online control phase. S4. The intended target is converted into control commands according to preset mapping rules, and these commands are sent to external devices to drive them to perform actions. The real-time physical parameters of the external devices during the action execution process are obtained as real-time operational status data, as detailed below: S4.1. Extract the preset control command mapping dictionary ,in Indicates the relationship with the first The baseline control command vector corresponding to each intent category Indexed by intent category and , The total number of intent categories, The preset physical space dimension for control commands; based on the smooth target intent output in step S3.3. From the control instruction mapping dictionary The corresponding baseline control command vector is extracted by matching. A first-order low-pass dynamic filtering algorithm is used to filter the reference control command vector. Smoothing is performed to eliminate potential mechanical abrupt changes in equipment that may occur during the intended switch, generating a smooth control command vector for the current time window. The calculation formula for the dynamic filtering algorithm is: in, The preset mechanical damping smoothing coefficient and It is used to adjust the response sensitivity and mechanical stability of control commands; This represents the smooth control command vector output from the previous time window; specifically, when the timing index of the sliding time window... During initialization It is a zero vector; By employing a first-order low-pass dynamic filter to inertially mitigate discrete and abrupt intent commands, it is possible to prevent sudden intent changes from causing hard impacts on external devices such as robotic arms or motors. This ensures smooth and rounded physical execution trajectories of the controlled external devices and extends the mechanical lifespan of the hardware. S4.2. The smooth control command vector output in step S4.1 The data is transmitted to external devices via the communication bus, driving the underlying motion controllers of the external devices to execute corresponding mechanical or logical actions. At the end of the current sliding time window, the external devices' built-in physical sensors collect real-time physical parameter vectors of the external devices during the action execution process. Among them, the real-time running physical parameter vector Dimensional distribution and smooth control command vector Maintaining alignment is used to characterize the current real-space pose or real-velocity response characteristics of external devices. S4.3. The real-time running physical parameter vector collected in step S4.2 With the smooth control command vector issued in step S4.1 Perform differential analysis to calculate the action execution deviation vector. To quantify the execution error caused by external equipment due to physical environmental interference or mechanical lag; action execution deviation vector The calculation formula is: Subsequently, the physical parameter vector will be run in real time. Deviation vector from action execution Vector features are concatenated to construct real-time running status data for the current time window. Real-time operating status data The expression is: Real-time running status data As a data source that fully characterizes the current physical response of the device, it directly serves as the input benchmark for generating interactive feedback signals containing action deviation prompts in the next process; S5. The real-time operating status data is converted into an interactive feedback signal that includes indications of equipment action deviations. This signal is used to correct subsequent judgments of the intended stable target, thereby forming a closed-loop feedback interactive control, as detailed below: S5.1. Parse the real-time running status data constructed in step S4.3 Extract the action execution deviation vector that characterizes the physical environment interference or mechanical lag of the external equipment. The action execution bias vector is calculated using a weighted Euclidean norm based on feature sensitivity. A dimensionality reduction assessment is performed to calculate a global deviation index used to quantify the degree of deviation in the current device's actions. Global Deviation Index The calculation formula is: in, This represents the global deviation index corresponding to the current time window; Represents the action execution deviation vector The Middle Each dimension of action execution deviation components; Indexed by physical space dimension and , The preset physical space dimension is the same as that of the control command in step S4.1; Indicates the pre-configured corresponding number The deviation sensitivity weights of each dimension are used to evaluate the impact of physical deviations in different degrees of freedom on the overall safety of interactive control tasks. By introducing independent weight parameters for each physical dimension, the degree of danger of deviation can be assessed differently according to specific task scenarios (such as the fault tolerance rate of the height dimension being lower than that of the horizontal dimension when an aircraft avoids obstacles), thus avoiding the problem of over- or under-feedback caused by a single Euclidean distance. S5.2. The global deviation index output in step S5.1 Input a preset nonlinear activation function to calculate the dynamic feedback adjustment factor. This can smoothly map an unbounded physical deviation index to a bounded interactive cue intensity; dynamic feedback adjustment factor The calculation formula is: in, The preset feedback gain coefficient is used to control the growth slope of the motion deviation prompt intensity; The preset lower bound threshold for deviation tolerance is used when the global deviation index... When the value is below this threshold, the device is determined to be in a safe redundancy state and no additional action deviation warning intensity is triggered. This is a function to find the maximum value. The purpose of nonlinear activation mapping is to set a dead zone and smoothly limit the upper limit of feedback gain. When the deviation is within a safe threshold, invalid interference is shielded. When the deviation exceeds the limit, the feedback intensity increases naturally without abrupt change, ensuring the comfort of the user's visual or tactile load when receiving deviation stimuli. S5.3. Extract the real-time running physical parameter vector acquired in step S4.2. This is compared with the dynamic feedback adjustment factor output in step S5.2. Signal synthesis is performed to obtain the interactive feedback signal vector for the current time window. ; the interactive feedback signal vector The information feedback module output to the external device is rendered to generate closed-loop physical stimuli; interactive feedback signal vector The synthesis formula is: in, This represents the control vector used to directly drive the physical intensity of the output of the information feedback module (such as a visual display or vibrating haptic array of an external device); The preset total number of interactive feedback signal output channels; This represents the preset state rendering projection matrix, used to represent the real-time running physical parameter vectors that characterize the device's true spatial pose or velocity. Convert to the basic prompt signal in normal operating mode; This represents a preset motion deviation warning feature vector, used to superimpose differentiated strong visual or strong tactile warning features onto the basic prompt signal; the information feedback module presents the interactive feedback signal vector. The system then outputs a real-time closed-loop physical stimulus, inducing the user to generate an error-related potential corresponding to the current control deviation. This induced next-state potential fluctuation is directly used as the next sliding time window (i.e., the time sequence index is...). The time window) is the data input source of the original continuous EEG signal in step S1.1, thereby completing the intention recognition feedback control closed loop in the temporal evolution.

[0027] Specifically, the abstract underlying errors are transformed into sensory stimulation signals (visual alerts or tactile pulses) that users can directly perceive. These stimulation signals can be rendered and presented in advance by the information feedback module of external devices at high frequency, inducing the user's brain to spontaneously adjust its EEG rhythm when cognitive conflict occurs (generating negative feedback regulatory signals such as ErrP error-related potentials). The new round of physiological fluctuations induced by this is immediately and seamlessly collected by the EEG cap and entered into the next sliding window (i+1). This forcibly breaks the one-way deadlock of the traditional intention recognition "open-loop direct input" and establishes a dynamic, fine-tuned, adaptive, and strong real-time communication link between the user's brain nerve center and the external actuator, ultimately achieving highly robust and high-precision closed-loop intention interaction control.

[0028] Example 2: This invention provides an intention recognition feedback interactive control system based on electroencephalogram (EEG) signals, used to implement the intention recognition feedback interactive control method based on EEG signals described in Embodiment 1, comprising: The EEG signal preprocessing module is used to acquire raw continuous EEG signals, and to perform segmentation and noise reduction preprocessing on the raw continuous EEG signals based on a preset sliding time window to obtain a high signal-to-noise ratio EEG signal sequence within the current time window. The feature extraction module is used to extract features from high signal-to-noise ratio EEG signal sequences to obtain a current multidimensional feature vector representing the current EEG state. The feature classification module is used to input the current multidimensional feature vector into the pre-built intent classification model to obtain the current candidate classification state, extract the historical time window intent classification state sequence recorded in the local cache, perform joint temporal smoothing and state transition analysis on the current candidate classification state and the historical time window intent classification state sequence to obtain the stable target intent after temporal continuity calibration, and write the stable target intent into the local cache to update the historical time window intent classification state sequence. The control command generation module is used to convert the intended target into control commands according to the preset mapping rules, and send the control commands to external devices to drive them to perform actions, and obtain the real-time physical parameters of the external devices during the action execution process as real-time running status data. The closed-loop control module is used to convert real-time operating status data into interactive feedback signals that include equipment action deviation prompts, which are used to correct subsequent stable target intent judgments, thereby forming closed-loop feedback interactive control.

[0029] Example 3: The present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it employs the intention recognition feedback interactive control method based on electroencephalogram (EEG) signals described in Embodiment 1.

[0030] It should be noted that the terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server, and the terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0031] Furthermore, the processor can be a central processing unit (CPU). Of course, depending on the actual use, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be used. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it in this regard.

[0032] Example 4: The present invention provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the intention recognition feedback interactive control method based on electroencephalogram (EEG) signals described in Embodiment 1.

[0033] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.

[0034] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0035] The foregoing has shown and described the basic principles, main features, and advantages of this application. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this application. Various changes and modifications can be made to this application without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of the claims of this application.

Claims

1. A method for intention recognition feedback interactive control based on electroencephalogram (EEG) signals, characterized in that, Includes the following steps: The raw continuous EEG signal is acquired, and the raw continuous EEG signal is segmented and denoised based on a preset sliding time window to obtain a high signal-to-noise ratio EEG signal sequence within the current time window. Feature extraction is performed on the high signal-to-noise ratio EEG signal sequence to obtain a current multidimensional feature vector representing the current EEG state; The current multidimensional feature vector is input into the pre-constructed intent classification model to obtain the current candidate classification state. The historical time window intent classification state sequence recorded in the local cache is extracted. The current candidate classification state and the historical time window intent classification state sequence are subjected to joint temporal smoothing and state transition analysis to obtain a stable target intent after temporal continuity calibration. The stable target intent is written into the local cache to update the historical time window intent classification state sequence. The stable target intent is converted into control commands according to the preset mapping rules, and the control commands are sent to external devices to drive them to perform actions. The real-time physical parameters of the external devices during the action execution process are obtained as real-time running status data. The real-time operating status data is converted into an interactive feedback signal that includes indications of device action deviations, which is used to correct subsequent determinations of stable target intent, thereby forming a closed-loop feedback interactive control.

2. The intention recognition feedback interactive control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that: The original continuous EEG signal is segmented and preprocessed for noise reduction based on a preset sliding time window to obtain a high signal-to-noise ratio EEG signal sequence within the current time window, specifically including: Based on the preset sliding time window length and sliding step size, the original continuous EEG signal is truncated and segmented to obtain the original EEG data matrix within the current time window; The original EEG data matrix is ​​preprocessed by temporal denoising using a finite impulse response filter with a preset passband range, and the temporal filtered EEG data matrix is ​​output. Using the time-domain filtered EEG data matrix as input, a common average reference algorithm is used for spatial denoising preprocessing. The high signal-to-noise ratio EEG signal sequence within the current time window is obtained by calculating the average potential of all channels and subtracting it from the potential of the corresponding target channel.

3. The intention recognition feedback interactive control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that: Feature extraction is performed on the high signal-to-noise ratio EEG signal sequence to obtain a current multidimensional feature vector representing the current EEG state, specifically including: The high signal-to-noise ratio EEG signal sequence is input into a pre-trained spatial filter for spatial feature projection to obtain a spatial filter signal matrix. Calculate the signal energy variance of each spatial filtering channel in the spatial filtering signal matrix within the current time window to obtain the spatial energy feature vector of the current time window; The spatial energy feature vector is logarithmically normalized to obtain the current multidimensional feature vector representing the current EEG state.

4. The intention recognition feedback interactive control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that: The current multidimensional feature vector is input into a preset intent classification model to obtain the current candidate classification state; the historical time window intent classification state sequence recorded in the local cache is extracted, and the current candidate classification state and the historical time window intent classification state sequence are subjected to joint temporal smoothing and state transition analysis to obtain a stable target intent after temporal continuity calibration, specifically including: The current multidimensional feature vector is input into a preset intent classification model, and the output confidence of each intent category corresponding to the current time window is calculated to obtain the current candidate classification state vector. Extract the historical smooth probability vector written in the previous time window from the local cache as the historical time window intent classification state sequence, and extract the preset intent state transition matrix; based on the preset temporal smoothing coefficient, fuse the current candidate classification state vector, the historical smooth probability vector and the intent state transition matrix to obtain the continuity calibration probability vector of the current time window. The maximum value of each calibration probability in the continuous calibration probability vector is optimized, and the intent category corresponding to the maximum calibration probability is determined as the stationary target intent after time-series continuous calibration. After extracting the stable target intent, the continuity calibration probability vector of the current time window is written into the local cache, overwriting the current historical smoothing probability vector.

5. The intention recognition feedback interactive control method based on electroencephalogram (EEG) signals according to claim 4, characterized in that: The construction and training process of the intent classification model is as follows: Multiple sets of offline raw continuous EEG signals and corresponding real intention labels of users under preset offline interactive tasks are collected, and the offline raw continuous EEG signals are segmented, denoised, and feature extracted to construct a training set containing multiple offline samples. The basic network structure of the intent classification model is defined as a single-layer fully connected Softmax mapping network. The loss function of the intent classification model is constructed based on the cross-entropy function and by introducing a regularization penalty term. The gradient descent optimization algorithm is used to minimize the loss function. The feature mapping weight matrix and decision bias vector of the intent classification model are iteratively updated within a preset number of iterations until the preset convergence condition is met. The converged network parameters are then output and configured into the intent classification model.

6. The intention recognition feedback interactive control method based on electroencephalogram (EEG) signals according to claim 1, characterized in that: The intended smooth target is converted into control commands according to a preset mapping rule, and the control commands are sent to external devices to drive them to perform actions. The real-time physical parameters of the external devices during the action execution process are acquired as real-time operating status data, specifically including: Based on the stated stable target intent, the corresponding baseline control command vector is extracted from the preset control command mapping dictionary; The reference control command vector is smoothed using a first-order low-pass dynamic filtering algorithm to generate a smoothed control command vector for the current time window, and the smoothed control command vector is sent to an external device. The real-time physical parameter vector of the external device during the action execution process is collected in real time. The difference analysis between the real-time physical parameter vector and the smooth control command vector is performed to calculate the action execution deviation vector. The real-time running physical parameter vector and the action execution deviation vector are concatenated to construct the real-time running status data for the current time window.

7. The intention recognition feedback interactive control method based on electroencephalogram (EEG) signals according to claim 6, characterized in that: The real-time operating status data is converted into an interactive feedback signal that includes indications of device action deviations, specifically including: The real-time running status data is analyzed, the action execution deviation vector is extracted, and the action execution deviation vector is evaluated and calculated using a weighted Euclidean norm based on feature sensitivity to obtain a global deviation index. The global deviation index is input into a preset nonlinear activation function to calculate the dynamic feedback adjustment factor. The real-time running physical parameter vector is extracted and synthesized with the dynamic feedback adjustment factor and the preset action deviation warning feature vector to obtain the interactive feedback signal vector of the current time window. The interactive feedback signal vector is then presented through the information feedback module of the external device.

8. A brainwave signal-based intention recognition feedback interactive control system, used to implement the brainwave signal-based intention recognition feedback interactive control method according to any one of claims 1 to 7, characterized in that, include: The EEG signal preprocessing module is used to acquire raw continuous EEG signals, and perform segmentation and noise reduction preprocessing on the raw continuous EEG signals based on a preset sliding time window to obtain a high signal-to-noise ratio EEG signal sequence within the current time window. The feature extraction module is used to extract features from the high signal-to-noise ratio EEG signal sequence to obtain a current multidimensional feature vector representing the current EEG state. The feature classification module is used to input the current multidimensional feature vector into a pre-built intent classification model to obtain the current candidate classification state, extract the historical time window intent classification state sequence recorded in the local cache, perform joint temporal smoothing and state transition analysis on the current candidate classification state and the historical time window intent classification state sequence to obtain a stable target intent after temporal continuity calibration, and write the stable target intent into the local cache to update the historical time window intent classification state sequence; The control command generation module is used to convert the stable target intent into control commands according to a preset mapping rule, and send the control commands to external devices to drive them to perform actions, and obtain the real-time physical parameters of the external devices during the action execution process as real-time running status data. The closed-loop control module is used to convert the real-time operating status data into an interactive feedback signal that includes equipment action deviation prompts, which is used to correct subsequent stable target intent judgments, thereby forming a closed-loop feedback interactive control.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs the intention recognition feedback interactive control method based on EEG signals as described in any one of claims 1 to 7.

10. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the intention recognition feedback interactive control method based on EEG signals as described in any one of claims 1 to 7.

Images