Real-time intention decoding method and system of earphone brain-computer interface

Abstract

The invention relates to the technical field of real-time intention decoding methods, in particular to a real-time intention decoding method and system for an earphone brain-computer interface, and the method comprises the following steps: obtaining an original electroencephalogram signal collected by an earphone; preprocessing the original electroencephalogram signal to obtain a preprocessed signal; extracting a dynamic feature based on the preprocessed signal to obtain a feature vector; fusing the feature vector with other modal features to obtain a fused feature; performing adaptive intention decoding based on the fusion feature to obtain an intention representation; outputting an intention according to the intention representation and obtaining feedback information; on-line learning and model updating are carried out based on fusion features, intention representation and feedback information; through an innovative NTFJA algorithm and a DMDRL network, the identification accuracy of 95% or above can still be kept in a complex environment, and the reliability of the system and the user experience are greatly improved; and the response time is reduced to be within 150ms, and real real-time control is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of real-time intention decoding methods, and more specifically, to a real-time intention decoding method for a headphone brain-computer interface and a system thereof. Background Art

[0002] As a direct human-computer interaction method, brain-computer interface technology has attracted widespread attention in the fields of smart home, assisted medical care, gaming and entertainment in recent years. Traditional brain-computer interface systems usually use head-mounted electroencephalogram caps for signal acquisition. Although this method has good signal quality, it has problems such as poor portability and low comfort in daily use. With the development of technology, headphone-type brain-computer interface has gradually become a research hotspot due to its portability and comfort. However, headphone-type brain-computer interface still faces many challenges in real-time intent decoding.

[0003] The main problems of existing technologies include:

[0004] 1. Unstable signal quality: The EEG signals collected by headphones are often subject to more environmental noise and electromyographic interference, resulting in unstable signal quality. Traditional filtering methods such as wavelet denoising or independent component analysis (ICA) have limited effect in processing such complex noise.

[0005] 2. Low feature extraction efficiency: Commonly used feature extraction methods such as power spectral density (PSD) or wavelet transform often fail to effectively capture key features related to intent when processing EEG signals collected by headphones, resulting in a decrease in the accuracy of subsequent classification.

[0006] 3. Difficulty in capturing the dynamic nature of intent: User intent may change rapidly over time, especially in multi-task scenarios. Traditional static classification methods such as support vector machines (SVM) or linear discriminant analysis (LDA) are difficult to adapt to such dynamic changes.

[0007] 4. Poor adaptability to individual differences: There are significant differences in the EEG patterns of different users, and existing transfer learning methods such as domain adaptation often require a large amount of target domain data, making it difficult to quickly adapt to new users.

[0008] 5. Limited scalability of intent categories: As application scenarios expand, new intent categories continue to emerge. Traditional methods often require retraining the entire model when dealing with new intent categories, which is not flexible enough.

[0009] In view of the above problems, there is an urgent need for a real-time intention decoding method for headphone brain-computer interface that can effectively solve these technical problems. Summary of the invention

[0010] The purpose of the present invention is to provide a real-time intention decoding method and system for a headphone brain-computer interface, so as to solve the problems in the prior art such as unstable quality of EEG signals collected by headphones, low feature extraction efficiency, difficulty in capturing the dynamics of intentions, poor adaptability to individual differences, and limited scalability of intention categories.

[0011] The present invention provides a real-time intention decoding method for a headphone brain-computer interface, comprising the following steps:

[0012] Get the original EEG signal collected by the headset;

[0013] Preprocessing the original EEG signal to obtain a preprocessed signal;

[0014] Extracting dynamic features based on the preprocessed signal to obtain a feature vector;

[0015] Fusing the feature vector with other modal features to obtain a fused feature;

[0016] Performing adaptive intent decoding based on the fused features to obtain intent representation;

[0017] Outputting an intention according to the intention expression and obtaining feedback information;

[0018] Online learning and model updating are performed based on the fused features, the intent representation and the feedback information.

[0019] Specifically, the step of preprocessing the original EEG signal includes:

[0020] An adaptive hybrid noise reduction algorithm AHDA is applied, where:

[0021] AHDA(x)=α*EEMD(x)+β*WNICA(x)+γ*AICA(x)

[0022] Among them, x is the original EEG signal, EEMD is the integrated empirical mode decomposition, WNICA is the combination of wavelet transform and independent component analysis, AICA is the adaptive independent component analysis; α, β, γ are adaptive weight coefficients, which are dynamically adjusted by the following formula:

[0023] α=sigmoid(SNR(x)-θ1)

[0024] β=sigmoid(Kurtosis(x)-θ2)

[0025] γ=1-α-β

[0026] Among them, SNR is the signal-to-noise ratio function, Kurtosis is the kurtosis function, and θ1 and θ2 are learnable threshold parameters.

[0027] Specifically, the step of extracting dynamic features based on the preprocessed signal includes:

[0028] The nonlinear time-frequency joint analysis algorithm NTFJA is applied, where:

[0029]

[0030] F f (x) = {P δ (x),P θ (x),P α (x),P β (x),P γ (x)}

[0031]

[0032] Where x is the preprocessed signal, is the time domain feature set, is the frequency domain feature set, is the time-frequency joint feature, μ and σ are the mean and standard deviation respectively, P * is the power of each EEG frequency band, STFT is short-time Fourier transform, CWT is continuous wavelet transform, It is the tensor outer product operation.

[0033] Specifically, the step of fusing the feature vector with other modal features includes: applying a quantum-inspired multimodal feature fusion algorithm QIMFF, wherein:

[0034] IMFF(X1,X2,...,X n )=Q_entangle(Q_superposition(X1)

[0035] Q_superposition(X2),...,Q_superposition(X2)

[0036] Q s upper position (X) = ∑ i α i |ψ i >

[0037]

[0038] Among them, X i is the characteristic of the i-th mode, |ψ i > is the quantum state representation, α i is the complex amplitude, is the tensor product and N is the normalization factor.

[0039] Specifically, the steps of performing adaptive intent decoding based on the fusion features include:

[0040] Apply the dynamic multi-task deep reinforcement learning network DMDRL, which includes the following sub-steps:

[0041] a) Use the fractal task decomposition algorithm FTD for task decomposition;

[0042] b) Use the quantum walk attention mechanism QWAM for feature weighting;

[0043] c) Use the fractal meta-learning algorithm FML for fast adaptation;

[0044] d) Use the fractal policy optimization algorithm FPO to optimize the decoding policy.

[0045] Specifically, the fractal task decomposition algorithm FTD is defined as follows:

[0046] FTD(T) = {T1, T2,..., T n}

[0047] T i = FTD r ecursive(t i , d m ax)

[0048]

[0049] where T is the total task, T i is the sub-task set, t i is a single task, d m ax is the maximum decomposition depth, is a tomic is a function to judge whether the task can be further decomposed, and split is the task splitting function.

[0050] Specifically, the quantum walk attention mechanism QWAM is defined as follows:

[0051]

[0052] U t = exp(-iHt)

[0053] H = Σ i jA i j|i><j|

[0054] where H is the input feature matrix, U t is the quantum walk evolution operator, A i j is the attention weight, |i><j| is the outer product operator, and i is the imaginary unit.

[0055] Specifically, the step of outputting the intention according to the intention representation and obtaining feedback information includes:

[0056] A fuzzy cognitive mapping feedback mechanism FCMF is applied, where:

[0057] FCMF(I,F)=FCM u pdate(FCM i nit,I,F)

[0058]

[0059] Among them, I is the intention output vector, F is the feedback vector, and FCM i nit is the initial fuzzy cognitive map, W is the concept weight matrix, η is the learning rate, It is the outer product operation.

[0060] Specifically, the steps of performing online learning and model updating based on fusion features, intent representation and feedback information include:

[0061] The fractal online learning algorithm FOL is applied, where:

[0062] FOL(M,D)=FOL r ecursive(M,D,d m ax)

[0063]

[0064] Among them, M is the current model, D is the newly collected data, and d m ax is the maximum recursive depth, threshold is the data size threshold, SGD is the stochastic gradient descent algorithm, partition is the data partition function, parallel_map is the parallel mapping function, and ensemble is the model integration function.

[0065] A real-time intention decoding system for a headphone brain-computer interface, comprising:

[0066] The signal acquisition module is used to obtain the original EEG signal collected by the headphones;

[0067] A preprocessing module, used for preprocessing the original EEG signal to obtain a preprocessed signal;

[0068] A feature extraction module, used for extracting dynamic features based on the preprocessed signal to obtain a feature vector;

[0069] A feature fusion module, used for fusing the feature vector with other modal features to obtain a fused feature;

[0070] An intent decoding module, used for performing adaptive intent decoding based on the fused features to obtain intent representation;

[0071] An output feedback module, used to output an intention according to the intention expression and obtain feedback information;

[0072] An online learning module is used to perform online learning and model updating based on the fusion features, the intent representation and the feedback information.

[0073] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0074] 1. Improve the accuracy of intent recognition: Through the innovative NTFJA algorithm and DMDRL network, the present invention can still maintain a recognition accuracy of more than 95% in complex environments, greatly improving the reliability of the system and user experience.

[0075] 2. Shorten the response time: With the help of the QIMFF algorithm and the FPO algorithm, the present invention reduces the response time to less than 150ms, realizing true real-time control, so that users can experience a more natural and smooth interactive experience.

[0076] 3. Accelerate the adaptation of new users: Thanks to the FML algorithm and the FOL algorithm, new users can become proficient in using the system in about 15 minutes, which greatly reduces the learning cost and improves the universality of the system.

[0077] 4. Enhanced anti-interference capability: Through the AHDA algorithm and QWAM mechanism, the present invention can still maintain stable performance in a high-noise environment, with the average accuracy rate dropping by no more than 10%, significantly expanding the application scenarios of the system.

[0078] 5. Improve multi-task processing capabilities: With the help of the multi-task learning architecture of the FTD algorithm and the DMDRL network, the present invention can still maintain an accuracy of more than 90% when processing multiple intent categories at the same time, enabling the system to cope with more complex control scenarios.

[0079] 6. Achieve continuous learning and optimization: Through the FCMF mechanism and FOL algorithm, the system can continuously learn and adapt to user usage habits and environmental changes to achieve long-term performance improvement.

[0080] In summary, the present invention effectively solves the main challenges faced in the prior art through a series of innovative algorithms and carefully designed processing flows, provides reliable technical support for the practical application of headphone-type brain-computer interfaces, and is expected to promote the widespread application of brain-computer interface technology in daily life. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 This is a flow chart of a real-time intent decoding method for a headphone brain-computer interface of the present invention.

[0082] Figure 2 This is a schematic diagram of the framework of a real-time intent decoding system for a headphone brain-computer interface of the present invention. DETAILED DESCRIPTION

[0083] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0084] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0085] Please refer to Figure 1 , the present invention provides a real-time intention decoding method for a headphone brain-computer interface, comprising the following steps:

[0086] 1. Signal acquisition steps: obtain the original EEG signal collected by the headphones;

[0087] 2. Preprocessing step: preprocessing the original EEG signal to obtain a preprocessed signal;

[0088] 3. Dynamic feature extraction step: extracting dynamic features based on the preprocessed signal to obtain a feature vector;

[0089] 4. Multimodal feature fusion step: fusing the feature vector with other modal features to obtain fused features;

[0090] 5. Adaptive intent decoding step: performing adaptive intent decoding based on the fusion features to obtain intent representation;

[0091] 6. Intention output and feedback step: outputting the intention according to the intention expression and obtaining feedback information;

[0092] 7. Online learning and model updating step: performing online learning and model updating based on the fusion features, the intent representation and the feedback information.

[0093] The following describes in detail the various steps of the present invention and its innovative algorithm in conjunction with specific application scenarios.

[0094] Suppose a user is using our headphone BCI system for smart home control. The user wants to adjust the light brightness, temperature, and music playback in the room by thinking. This scenario involves multiple intent categories (such as "adjust the light", "adjust the temperature", "control the music") and different intent intensities (such as the specific adjustment degree of the light brightness).

[0095] 1. Signal acquisition step, in this step, we use the high-precision, low-noise bioelectric signal acquisition chip integrated in the headset (such as TI's ADS1299) to collect the user's EEG signals. The sampling frequency is set to 1024Hz to ensure that high-frequency EEG activity is captured. This sampling frequency is sufficient to capture the EEG bands related to intent, including beta waves (13-30Hz) and gamma waves (30-100Hz).

[0096] 2. Preprocessing step: After obtaining the original EEG signal, we need to preprocess it to improve the signal quality. This step uses the innovative Adaptive Hybrid Denoising Algorithm (AHDA). The mathematical expression of the AHDA algorithm is as follows:

[0097] AHDA(x)=α*EEMD(x)+β*WNICA(x)+γ*AICA(x)

[0098] Among them, x is the original EEG signal, EEMD is the integrated empirical mode decomposition, WNICA is the combination of wavelet transform and independent component analysis, AICA is the adaptive independent component analysis; α, β, γ are adaptive weight coefficients, which are dynamically adjusted by the following formula:

[0099] α=sigmoid(SNR(x)-θ1)

[0100] β=sigmoid(Kurtosis(x)-θ2)

[0101] γ=1-α-β

[0102] Among them, SNR is the signal-to-noise ratio function, Kurtosis is the kurtosis function, and θ1 and θ2 are learnable threshold parameters.

[0103] In practice, we found that the initial values ​​of θ1 and θ2 set to 5 and 3 worked well, but these parameters will be continuously optimized during system operation. The advantage of the AHDA algorithm is that it can adaptively select the noise reduction method that best suits the current signal characteristics. For example, when the user is walking, more electromyographic interference may be generated. At this time, the weight of WNICA will automatically increase to better remove this type of noise.

[0104] 3. Dynamic feature extraction step. In this step, we apply the nonlinear time-frequency joint analysis algorithm (NTFJA) to extract dynamic features. The mathematical expression of NTFJA is as follows:

[0105] The nonlinear time-frequency joint analysis algorithm NTFJA is applied, where:

[0106] NTFJA(x)={F t (x),F f (x),F tf (x)}

[0107] in:

[0108]

[0109] F f (x) = {P δ (x),P θ (x),P α (x),P β (x),P γ (x)}

[0110]

[0111] Here, x is the preprocessed signal, is the time domain feature set, is the frequency domain feature set, is the time-frequency joint feature, μ and σ are the mean and standard deviation respectively, P * is the power of each EEG frequency band, STFT is short-time Fourier transform, CWT is continuous wavelet transform, It is the tensor outer product operation.

[0112] In our smart home control scenario, the NTFJA algorithm can effectively capture the characteristics of different user intentions. For example, when a user wants to adjust the lights, there may be stronger activity in the alpha band (8-13Hz); while when controlling music, there may be more obvious changes in the beta band. The NTFJA algorithm can comprehensively capture these subtle differences by simultaneously analyzing the characteristics of the time domain, frequency domain, and time-frequency domain.

[0113] 4. Multimodal feature fusion step, in order to improve the accuracy of intent decoding, we not only use EEG signals, but also fuse information from other modalities, such as accelerometer and gyroscope data in the headset. We use the quantum-inspired multimodal feature fusion algorithm (Quantum-Inspired Multimodal Feature Fusion, QIMFF) to achieve this goal. The mathematical expression of QIMFF is as follows:

[0114] IMFF(X1,X2,...,X n )=Q_entangle(Q_superposition(X1)

[0115] Q_superposition(X2),...,Q_superposition(X2)

[0116] Q s upper position(X) = ∑ i α i |ψ i >

[0117]

[0118] Among them, X i is the characteristic of the i-th mode, |ψ i > is the quantum state representation, α i is the complex amplitude, is the tensor product and N is the normalization factor.

[0119] In our smart home control scenario, the QIMFF algorithm can effectively fuse EEG signal features (X1), accelerometer data (X2), and gyroscope data (X3). For example, when a user wants to adjust the temperature, they may unconsciously make small head movements, which can be captured by the accelerometer and gyroscope. The QIMFF algorithm uses quantum-inspired methods to capture the complex nonlinear relationship between these modalities, thereby improving the accuracy of intent decoding. In practical applications, we found that the α i The initial value is set to (where n is the number of modes) has a good effect, but these parameters will be continuously optimized during the operation of the system.

[0120] 5. Adaptive intent decoding step, which is the core step of the present invention, we use Dynamic Multi-task Deep Rein forcement Learning Network (DMDRL) for adaptive intent decoding. DMDRL contains the following four innovative sub-modules:

[0121] a) Fractal Task Decomposition (FTD)

[0122] The mathematical expression of FTD is as follows:

[0123] FTFTD(T) = {T1, T2,..., T n}

[0124] T i = FTD r ecursive(t i , d m ax)

[0125]

[0126] Here, T is the total task, T i is the subtask set, t i is a single task, d m ax is the maximum decomposition depth, is a tomic is a function to judge whether the task can be further decomposed, and split is the task splitting function.

[0127] In our smart home control scenario, the FTD algorithm can decompose complex intent decoding tasks into finer-grained subtasks. For example, the total task of "adjusting the lights" can be decomposed into subtasks such as "identifying the light adjustment intent", "estimating the adjustment direction (increase / decrease brightness)", and "estimating the adjustment amplitude". We found that setting d max to 3 usually achieves a good balance, neither over-decomposing the task nor fully utilizing the hierarchical structure of the task.

[0128] b) Quantum Walk Attention Mechanism (QWAM), and the mathematical expression of QWAM is as follows:

[0129]

[0130] U t = exp(-iHt)

[0131] H = ∑ i jA i j|i><j|

[0132] where H is the input feature matrix, U t is the quantum walk evolution operator, A i j is the attention weight, |i><j| is the outer product operator, and i is the imaginary unit.

[0133] QWAM can dynamically adjust the attention to different features and time periods. In our scenario, when the user is adjusting the light, QWAM automatically increases the attention to the brain signals related to vision. We found that by setting the time parameter t of quantum walk between 0.1 and 1, the system can quickly switch attention between different tasks while maintaining sufficient stability. c) Fractal Meta-Learning (FML), the mathematical expression of FML is as follows:

[0134] FML(θ)=FML r ecursive(θ,d m ax)

[0135]

[0136] Here, θ represents the model parameters, d m ax represents the maximum recursive depth, and MAML represents the model-agnostic meta-learning algorithm.

[0137] The FML algorithm allows our system to quickly adapt to new users or new intent categories. For example, when the system first encounters the user's intent to control smart curtains, FML can quickly learn and adapt to this new task based on other control tasks that have been learned (such as adjusting lights). We found that setting d_max to 2 or 3 usually achieves a good balance between computational complexity and adaptability.

[0138] d) Fractal Policy Optimization (FPO), the mathematical expression of FPO is as follows:

[0139] FPO(π)=FPO r ecursive(π,d m ax)

[0140]

[0141] Here, π represents the policy function, d m ax represents the maximum recursive depth, DDPG represents the deep deterministic policy gradient algorithm, split p policy represents the policy decomposition function, and mergepolicy represents the policy merging function.

[0142] The FPO algorithm is able to find the best balance between decoding speed and accuracy. In our smart home control scenario, different control tasks may require different response speeds. For example, adjusting the light may require a faster response, while adjusting the temperature can be a little slower. The FPO algorithm is able to adaptively optimize the decoding strategy for each subtask. We found thatm Setting ax to 2 usually achieves a good balance between optimization effect and computational complexity.

[0143] These four innovative sub-modules work together to enable the DMDRL network to decode the user's intention efficiently and accurately. In our smart home control scenario, DMDRL is able to handle multiple control tasks simultaneously (such as adjusting lights, temperature, and music) and can quickly adapt to the user's personalized needs and emerging control requirements.

[0144] 6. Intention output and feedback step, in this step, we use the fuzzy cognitive map feedback mechanism (FCMF) to process the system output intention and user feedback. The mathematical expression of FCMF is as follows:

[0145] FCMF(I,F)=FCM u pdate(FCM i nit,I,F)

[0146]

[0147] Among them, I is the intention output vector, F is the feedback vector, and FCM i nit is the initial fuzzy cognitive map, W is the concept weight matrix, η is the learning rate, It is the outer product operation.

[0148] In our smart home control scenario, the FCMF mechanism can effectively handle complex feedback information. For example, when the system dims the lights based on the user's EEG signals, but the user then manually brightens the lights, FCMF will capture this feedback and update the fuzzy cognitive map accordingly. This allows the system to gradually learn the user's preferences and improve the accuracy of future decoding.

[0149] We found that setting the learning rate η between 0.01 and 0.1 works well, which can ensure the system's sensitivity to feedback while avoiding excessive fluctuations. i NITs can be pre-set based on domain knowledge, for example, the initial weight between the concepts of "adjusting lighting" and "visual comfort" is set high.

[0150] 7. Online learning and model updating steps. Finally, we use the fractal online learning algorithm (Fractal Online Learning, FOL) to continuously update and optimize the entire system. The mathematical expression of FOL is as follows:

[0151] FOL(M,D)=FOL r ecursive(M,D,dm ax)

[0152]

[0153] Here, M is the current model, D is the newly collected data, and d m ax is the maximum recursive depth, threshold is the data size threshold, SGD is the stochastic gradient descent algorithm, partition is the data partition function, parallelmap is the parallel mapping function, and ensemble is the model integration function.

[0154] Here, M represents the current model, D represents the newly collected data, d_max represents the maximum recursive depth, threshold represents the data size threshold, SGD represents the stochastic gradient descent algorithm, partition represents the data partition function, parallel_map represents the parallel mapping function, and ensemble represents the model integration function.

[0155] In our smart home control scenario, the FOL algorithm can continuously optimize system performance. For example, as the user spends more time using the system, FOL will gradually adjust the parameters of each submodule to make the system better adapt to the user's usage habits and environmental changes. We found that setting d_max to 3 and threshold to 1000 sample points can usually achieve a good balance between learning effect and computational efficiency.

[0156] These seven steps form a complete closed-loop system that can continuously learn and improve, providing users with an increasingly accurate brain-computer interface experience. In actual applications, we found that the system can achieve good personalization effects after about 30 minutes of use, and the user's intention recognition accuracy can reach more than 90%.

[0157] Please refer to Figure 2 Based on the above method, the present invention also provides a real-time intention decoding system for a headphone brain-computer interface, comprising the following modules:

[0158] The signal acquisition module 1 is used to obtain the original EEG signal collected by the earphone. The module includes a high-precision, low-noise bioelectric signal acquisition chip (such as ADS1299 from TI) integrated in the earphone.

[0159] The preprocessing module 2 is used to preprocess the original EEG signal to obtain a preprocessed signal. This module includes a processing unit that implements the AHDA algorithm.

[0160] The feature extraction module 3 is used to extract dynamic features based on the preprocessed signal to obtain a feature vector. This module includes a processing unit that implements the NTFJA algorithm.

[0161] The feature fusion module 4 is used to fuse the feature vector with other modal features to obtain a fusion feature. This module includes a processing unit that implements the QIMFF algorithm.

[0162] The intention decoding module 5 is used for adaptively decoding the intention based on the fusion features to obtain the intention representation. The module includes a processing unit for implementing the DMDRL network, which includes four sub-units: FTD, QWAM, FML and FPO.

[0163] The output feedback module 6 is used to output the intention according to the intention expression and obtain feedback information. This module includes a processing unit that implements the FCMF mechanism.

[0164] The online learning module 7 is used for online learning and model updating based on the fusion features, intent representation and feedback information. This module includes a processing unit for implementing the FOL algorithm.

[0165] These modules can be implemented on a general-purpose processor through software or through a dedicated hardware accelerator (such as FPGA or ASIC) to meet the needs of real-time processing. In our prototype system, we used the Xilinx Zynq UltraScale + MPSoC ZCU104 evaluation board, which contains an ARM Cortex-A53 processor and programmable logic resources, which can meet the needs of flexibility and real-time performance.

[0166] The headphone brain-computer interface real-time intention decoding method and system of the present invention effectively solves the main challenges faced in the prior art through a series of innovative algorithms and carefully designed processing procedures. The system can adapt to changes in signal quality, efficiently extract features, capture the dynamics of intent, quickly adapt to individual differences, and can flexibly expand to new intent categories. This enables users to control smart home devices more naturally and accurately through their thoughts, greatly improving the user experience and the practicality of the system.

[0167] To verify the superiority of the present invention, we conducted a group of comparative experiments, including 1 embodiment and 3 comparative examples.

[0168] Embodiment 1: Real-time intention decoding method of the headset brain-computer interface of the present invention

[0169] Innovative algorithms such as AHDA, NTFJA, QIMFF, DMDRL (including FTD, QWAM, FML, FPO), FCMF and FOL are adopted.

[0170] Comparative Example 1: Traditional EEG feature extraction + SVM classification method

[0171] Wavelet transform was used for feature extraction and support vector machine (SVM) was used for classification.

[0172] Comparative Example 2: Deep Learning Method

[0173] The raw EEG signals are directly processed and classified using a convolutional neural network (CNN).

[0174] Comparative Example 3: Traditional Brain-Computer Interface Method + Transfer Learning

[0175] Common spatial patterns (CSP) are used for feature extraction, linear discriminant analysis (LDA) is used for classification, and transfer learning is adopted to adapt to new users.

[0176] Test indicators and methods

[0177] 1. Intent recognition accuracy: Use 5-fold cross validation to calculate the percentage of correctly recognized intents.

[0178] 2. Response time: The average time (milliseconds) from receiving the EEG signal to outputting the intention.

[0179] 3. New user adaptation time: The training time (minutes) required for a new user to reach 85% accuracy.

[0180] 4. Anti-interference ability: the average accuracy reduction percentage under different noise levels (SNR from 20dB to 0dB).

[0181] 5. Multitasking capability: Average accuracy when processing multiple intent categories (2-5) simultaneously.

[0182] The test results are shown in the following table

[0183] index Example 1 Comparative Example 1 Comparative Example 2 Comparative Example 3 Intent recognition accuracy 95.2％ 82.5％ 88.7％ 85.3％ Response time 150ms 280ms 200ms 250ms New user adaptation time 15 minutes 60 minutes 45 minutes 30 minutes Anti-interference ability 8.5％ 25.3％ 18.7％ 22.1％ Multitasking capabilities 91.8％ 75.6％ 83.2％ 79.5％

[0184] According to the test results in the above table, Example 1 (the method of the present invention) is significantly better than other comparative methods in all indicators, and can therefore be considered the best embodiment. The following is an in-depth analysis of each indicator:

[0185] 1. Intent recognition accuracy: The accuracy of the method of the present invention reaches 95.2%, which is much higher than other methods. This is mainly due to the accurate extraction of dynamic features by the NTFJA algorithm and the adaptive intent decoding capability of the DMDRL network. The high accuracy means that users can more reliably control devices through their thoughts, greatly improving the user experience.

[0186] 2. Response time: The response time of the proposed method is only 150ms, which is 30-130ms faster than other methods. This is mainly due to the efficient feature fusion of the QIMFF algorithm and the strategy optimization of the FPO algorithm. Low latency is crucial for real-time control and allows users to experience a more natural and smooth interactive experience.

[0187] 3. New user adaptation time: The method of the present invention can enable new users to reach a high level of proficiency in use in just 15 minutes, which is 2-4 times faster than other methods. This is due to the rapid adaptation ability of the FML algorithm and the continuous optimization of the FOL algorithm. The shorter adaptation time greatly reduces the user's learning cost and improves the universality of the system.

[0188] 4. Anti-interference ability: Under different noise levels, the accuracy of the proposed method drops by only 8.5% on average, while the other methods drop by 18.7%-25.3%. This is mainly due to the adaptive noise reduction capability of the AHDA algorithm and the dynamic attention mechanism of QWAM. The strong anti-interference ability enables the system to work stably in various complex environments, greatly expanding the application scenarios.

[0189] 5. Multi-task processing capability: When processing multiple intent categories at the same time, the method of the present invention maintains a high accuracy of 91.8%, while the performance of other methods drops significantly. This is mainly due to the task decomposition capability of the FTD algorithm and the multi-task learning architecture of the DMDRL network. The excellent multi-task processing capability enables the system to cope with more complex control scenarios, such as adjusting multiple smart home devices at the same time.

[0190] The real-time intention decoding method of the headset brain-computer interface of the present invention shows significant superiority in all key indicators. These results fully demonstrate the effectiveness and advancement of the innovative algorithm proposed by the present invention. In particular, the outstanding performance in accuracy, response speed and adaptability makes this method very suitable for practical application scenarios, such as smart home control.

[0191] In addition, the advantages of the method of the present invention in terms of anti-interference ability and multi-tasking ability further expand its application scope, enabling it to work stably in more complex and changeable environments. This not only improves the user experience, but also paves the way for the widespread application of brain-computer interface technology in daily life.

[0192] It should be pointed out that although the method of the present invention performs well in all indicators, in practical applications, further optimization and adjustment may be required according to specific scenarios and user needs. For example, for a specific user group or a specific control task, it may be necessary to fine-tune the algorithm parameters to obtain the best performance.

[0193] In general, these test results fully verify the innovation and practicality of the present invention in the field of real-time intent decoding of headphone brain-computer interface, and provide the possibility for more natural and intelligent human-computer interaction in the future.

[0194] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A real-time intention decoding method for a headphone brain-computer interface, characterized in that: The following steps are involved: Get the original EEG signal collected by the headset; Preprocessing the original EEG signal to obtain a preprocessed signal; Extracting dynamic features based on the preprocessed signal to obtain a feature vector; Fusing the feature vector with other modal features to obtain a fused feature; Performing adaptive intent decoding based on the fused features to obtain intent representation; Outputting an intention according to the intention expression and obtaining feedback information; Online learning and model updating are performed based on the fused features, the intent representation and the feedback information.

2. The method according to claim 1, characterized in that The step of preprocessing the original EEG signal comprises: An adaptive hybrid noise reduction algorithm AHDA is applied, where: AHDA(x)=α*EEMD(x)+β*WNICA(x)+γ*AICA(x) Among them, x is the original EEG signal, EEMD is the integrated empirical mode decomposition, WNICA is the combination of wavelet transform and independent component analysis, AICA is the adaptive independent component analysis; α, β, γ are adaptive weight coefficients, which are dynamically adjusted by the following formula: α=sigmoid(SNR(x)-θ1) β=sigmoid(Kurtosis(x)-θ2) γ=1-α-β Among them, SNR is the signal-to-noise ratio function, Kurtosis is the kurtosis function, and θ1 and θ2 are learnable threshold parameters.

3. The method according to claim 1, characterized in that The step of extracting dynamic features based on the preprocessed signal comprises: The nonlinear time-frequency joint analysis algorithm NTFJA is applied, where: F f (x)={P δ (x),P θ (x),P α (x),P β (x),P γ (x)} Where x is the preprocessed signal, is the time domain feature set, is the frequency domain feature set, is the time-frequency joint feature, μ and σ are the mean and standard deviation respectively, P * is the power of each EEG frequency band, STFT is short-time Fourier transform, CWT is continuous wavelet transform, It is the tensor outer product operation.

4. The method according to claim 1, characterized in that: The step of fusing the feature vector with other modal features includes: Apply the quantum-inspired multimodal feature fusion algorithm QIMFF, where: IMFF(X1,X2,...,X n )=Q_entangle(Q_superposition(X1) Q_superposition(X2),...,Q_superposition(X2) Q s superposition(X)=∑ i a i |ψ i > Among them, X i is the characteristic of the i-th mode, |ψ i > is the quantum state representation, α i is the complex amplitude, is the tensor product and N is the normalization factor.

5. The method according to claim 1, characterized in that The step of performing adaptive intent decoding based on fusion features comprises: Apply the dynamic multi-task deep reinforcement learning network DMDRL, which includes the following sub-steps: a) Use the fractal task decomposition algorithm FTD to decompose tasks; b) Use quantum walk attention mechanism QWAM for feature weighting; c) Use the fractal meta-learning algorithm FML for fast adaptation; d) Use the fractal strategy optimization algorithm FPO to optimize the decoding strategy.

6. The method according to claim 5, characterized in that The fractal task decomposition algorithm FTD is defined as follows: FTD(T)={T1,T2,...,T n } T i =FTD r ecursive(t i ,d m ax) Among them, T is the total task, T i is the subtask set, t i For a single task, d m ax is the maximum decomposition depth, is a tomic is a function for determining whether a task can be further decomposed, and split is a task splitting function.

7. The method according to claim 5, characterized in that The quantum walk attention mechanism QWAM is defined as follows: U t =exp(-iHt) H=∑ i and i j|i> <j| Among them, H is the input feature matrix, and U t is the quantum walk evolution operator, A i j is the attention weight, |i><j| is the outer product operator, and i is the imaginary unit.

8. The method according to claim 1, characterized in that The step of outputting the intention according to the intention representation and obtaining feedback information comprises: A fuzzy cognitive mapping feedback mechanism FCMF is applied, where: FCMF(I,F)=FCM u pdate(FCM i nit,I,F) Among them, I is the intention output vector, F is the feedback vector, and FCM i nit is the initial fuzzy cognitive map, W is the concept weight matrix, η is the learning rate, It is the outer product operation.

9. The method according to claim 1, characterized in that: The steps of performing online learning and model updating based on fusion features, intent representation and feedback information include: The fractal online learning algorithm FOL is applied, where: FOL(M,D)=FOL r ecursive(M,D,d m ax) Among them, M is the current model, D is the newly collected data, and d m ax is the maximum recursive depth, threshold is the data size threshold, SGD is the stochastic gradient descent algorithm, partition is the data partition function, parallel_map is the parallel mapping function, and ensemble is the model integration function.

10. A real-time intention decoding system for a headphone brain-computer interface, characterized in that: include: The signal acquisition module is used to obtain the original EEG signal collected by the headphones; A preprocessing module, used for preprocessing the original EEG signal to obtain a preprocessed signal; A feature extraction module, used for extracting dynamic features based on the preprocessed signal to obtain a feature vector; A feature fusion module, used for fusing the feature vector with other modal features to obtain a fused feature; An intent decoding module, used for performing adaptive intent decoding based on the fused features to obtain intent representation; An output feedback module, used to output an intention according to the intention expression and obtain feedback information; An online learning module is used to perform online learning and model updating based on the fusion features, the intent representation and the feedback information.

Images