Cross-day drift suppression adaptive update method and system for long-term access of brain-computer interfaces
By using multimodal data modeling, confidence-driven sample selection mechanism, and incremental learning, the problem of signal drift across days in brain-computer interface systems during long-term use was solved, achieving stable adaptive updates and an efficient user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2026-05-08
- Publication Date
- 2026-06-05
AI Technical Summary
Existing brain-computer interface systems suffer from decreased recognition accuracy due to signal drift over time during long-term use. Existing methods increase the burden on users, are not suitable for long-term continuous use, and are prone to catastrophic amnesia.
We employ multimodal data modeling, a confidence-driven sample selection mechanism, and an incremental learning strategy. We train the basic model using group multimodal data and fine-tune it using individual multimodal data to construct individual models and adaptively update them using a historical memory database.
It effectively suppresses cross-day signal drift, reduces user calibration burden, improves system stability and robustness, maintains memory of historical distribution, reduces noise interference, and improves user experience.
Smart Images
Figure CN122153469A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of brain-computer interfaces, specifically relating to a cross-day drift suppression adaptive update method and system for long-term brain-computer interface access. Background Technology
[0002] Brain-computer interface (BCI) systems enable information interaction between users and external devices by collecting users' electroencephalogram (EEG) signals or other physiological signals. BCI technology has been widely applied in fields such as intelligent assistive devices and human-computer interaction.
[0003] However, in practical applications, brain-computer interface (BCI) systems typically require long-term online use. Due to factors such as changes in electrode placement, fluctuations in skin condition, changes in the user's physiological and psychological state, and environmental interference, EEG signals and other physiological signals often exhibit significant distribution differences between different days of use, a phenomenon known as cross-day signal drift. This drift leads to a decrease in the recognition accuracy of BCI systems trained on fixed models, affecting the system's stability and practicality. Existing technologies often employ frequent recalibration or complete model retraining to mitigate cross-day drift, but these methods require users to invest significant time in annotation, increasing their burden and are unsuitable for long-term continuous use. Some online adaptive methods can update model parameters to some extent, but are prone to catastrophic forgetting, meaning the model loses its ability to recognize historical distributions while adapting to new distributions. Therefore, there is an urgent need for a BCI method that can effectively suppress cross-day signal drift and reduce calibration requirements under long-term use conditions, while simultaneously ensuring model stability and adaptability. Summary of the Invention
[0004] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a cross-day drift suppression adaptive update method and system for long-term brain-computer interface access. Through multimodal data modeling, confidence-driven sample selection mechanism, and incremental learning strategy combined with historical playback, the brain-computer interface model can achieve stable adaptive updates during long-term use, thereby suppressing cross-day signal drift and alleviating catastrophic forgetting problems, while reducing the user calibration burden and improving the system's practicality and robustness.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] One aspect of the present invention provides a cross-day drift suppression adaptive update method for long-term brain-computer interface access, comprising the following steps:
[0007] Individual multimodal data from several users are collected as group multimodal data; the individual multimodal data includes electroencephalogram (EEG) signals and physiological parameter signals;
[0008] A basic model based on a deep neural network structure is selected, trained using group multimodal data, and fine-tuned using calibrated individual multimodal data to obtain an initial individual model, and a historical memory database corresponding to the individual model is constructed.
[0009] During online use, individual models are used to make real-time predictions on newly collected individual multimodal data;
[0010] Set up confidence assessment indicators to assess the confidence level of the real-time prediction results;
[0011] Based on the confidence assessment results, individual multimodal data with confidence scores higher than a preset threshold and their corresponding real feedback labels are selected and stored in the historical memory database;
[0012] Incremental learning is used to jointly optimize the model parameters of individual models based on current online data and historical memory databases.
[0013] As a preferred technical solution, the physiological parameter signal includes at least one of near-infrared signal, electromyography signal, electrooculography signal, electrocardiogram signal, electrodermal signal, and respiratory signal.
[0014] As a preferred technical solution, the individual multimodal data undergoes at least one of the following processes after acquisition: time synchronization, filtering and denoising, artifact removal, normalization, and feature extraction.
[0015] As a preferred technical solution, the basic model based on the deep neural network structure adopts one or a combination of the following structures: convolutional neural network, recurrent neural network and Transformer network.
[0016] As a preferred technical solution, during training using multimodal population data, the objective function is set as follows:
[0017] ;
[0018] in, The loss function for supervised learning; For group multimodal data; The model parameters are: The basic model for input data The prediction results; For input data Corresponding real feedback tags; For mathematical expectation operators; For parameter optimization operators, used to solve for the model parameters that minimize the objective function. ; These are the optimal model parameters obtained through training.
[0019] As a preferred technical solution, the confidence evaluation index includes one or more of the following: prediction probability, class interval, distribution uncertainty, latent feature space similarity, and signal quality index;
[0020] The predicted probability is specifically: the maximum probability value corresponding to the predicted category calculated based on the category probability distribution output by the individual model;
[0021] The category interval is calculated based on the probability difference between the category with the highest predicted probability and the category with the second highest predicted probability;
[0022] The distribution uncertainty specifically refers to the uncertainty of the prediction result calculated by performing multiple random inferences, random deactivation inferences, or model ensemble inferences on the same individual multimodal data.
[0023] The latent feature space similarity is specifically calculated as follows: the similarity between the current individual multimodal data in the latent feature space of the individual model and the features in the historical memory database is calculated;
[0024] The signal quality index is calculated based on the signal-to-noise ratio, artifact level, or channel stability of the collected individual multimodal data.
[0025] As a preferred technical solution, the confidence level of the prediction results is evaluated, specifically as follows:
[0026] The overall confidence level is obtained by weighting and fusing one or more confidence assessment indicators:
[0027] ;
[0028] in, For evaluation Individual multimodal data at time 1 Scores on each confidence assessment metric To assess the overall confidence level, This represents the total number of confidence assessment indicators. For the first Each confidence assessment indicator corresponds to a preset normalized weight coefficient, and .
[0029] As a preferred technical solution, the joint optimization of the model parameters of the individual models specifically involves:
[0030] A predetermined proportion of data is extracted from the historical memory database and used in conjunction with current online data to train individual models; the loss function is set as follows:
[0031] ;
[0032] in, Represents the total loss item. This represents the loss term of the currently online data. This represents the loss term of the data extracted from the historical memory database. These are the weighting coefficients.
[0033] As a preferred technical solution, the joint optimization is performed periodically, or it is set to be triggered when the amount of data, time span, or distribution change of the data stored in the historical memory database reaches a preset condition.
[0034] Another aspect of the present invention provides a cross-day drift suppression adaptive update system for long-term brain-computer interface access, applied to the above-mentioned cross-day drift suppression adaptive update method for long-term brain-computer interface access, including a data acquisition module, a model training module, an online prediction module, a confidence assessment module, and a joint optimization module;
[0035] The data acquisition module is used to collect individual multimodal data from several users as group multimodal data; the individual multimodal data includes electroencephalogram (EEG) signals and physiological parameter signals;
[0036] The model training module is used to select a basic model based on the Transformer architecture, train it according to the group multimodal data, and fine-tune it according to the calibrated individual multimodal data to obtain an initial individual model and construct a historical memory database corresponding to the individual model.
[0037] The online prediction module is used to make real-time predictions on newly collected individual multimodal data based on the individual model during online use.
[0038] The confidence assessment module is used to assess the confidence of the real-time prediction results; based on the confidence assessment results, individual multimodal data with confidence scores higher than a preset threshold and their corresponding real feedback labels are selected and stored in the historical memory database.
[0039] The joint optimization module uses incremental learning to jointly optimize the model parameters of individual models based on current online data and historical memory database.
[0040] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0041] (1) This invention improves the initial performance of the model and reduces the need for individual calibration by combining pre-training of group data with individual fine-tuning;
[0042] (2) This invention uses multi-dimensional confidence assessment and sample screening to utilize only high-quality samples in model updates, effectively reducing the interference of noisy samples on model performance;
[0043] (3) This invention, through historical memory and historical playback mechanisms, maintains the memory of historical distribution while adapting to new data distribution, effectively alleviating the problem of catastrophic forgetting;
[0044] (4) This invention can achieve stability and adaptive updates of brain-computer interface system under long-term use without frequent recalibration, thereby improving system robustness and user experience. Attached Figure Description
[0045] Figure 1 This is a schematic diagram of the overall process of the cross-day drift suppression adaptive update method for long-term access to brain-computer interfaces according to an embodiment of the present invention;
[0046] Figure 2 This is a schematic diagram of the historical memory database and model update process in an embodiment of the present invention. Detailed Implementation
[0047] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.
[0048] Example 1:
[0049] like Figure 1 , Figure 2 As shown, this embodiment provides a cross-day drift suppression adaptive update method for long-term brain-computer interface access, including the following steps:
[0050] S1. Data Acquisition and Preprocessing.
[0051] Collect individual multimodal data of users, including electroencephalogram (EEG) signals and at least one other physiological parameter signal. The individual multimodal data of one user is regarded as a sample, and multiple samples constitute the group multimodal data.
[0052] Among them, the electroencephalogram (EEG) signal can be acquired by a multi-channel EEG acquisition device, and the sampling frequency is preferably 250Hz to 1000Hz; other physiological parameter signals include one or more of the following: near-infrared signal, electromyography (EMG) signal, electrooculography (EOG) signal, electrocardiogram (ECG) signal, electrodermal signal, or respiratory signal.
[0053] In a specific implementation, at Individual multimodal data collected at various times are represented as follows:
[0054] ;
[0055] in, Indicates brain electrical signals, They represent the first Other physiological parameter signals of the modality, and .
[0056] Preferably, before inputting the individual multimodal data into the model, time synchronization processing is performed on each modal signal, and at least one of the following processes is performed respectively: bandpass filtering, artifact removal, normalization processing, and feature extraction processing to improve the stability of subsequent model prediction.
[0057] S2. Model pre-training and individual initialization: Constructing the initial individual model of the brain-computer interface system.
[0058] We selected a basic model based on a deep neural network structure and used a multi-modal dataset of multiple users. Conduct training.
[0059] The optimization objective function during training is set as follows:
[0060] ;
[0061] in, The cross-entropy loss function or other supervised learning loss function is used to quantize the predicted output. With real feedback tags The error; The model parameters are: The basic model for input data The prediction results; For input data Corresponding real feedback tags; The mathematical expectation operator is used for population multimodal datasets. The loss of all samples is averaged. For parameter optimization operators, used to solve for the model parameters that minimize the objective function. ; The optimal model parameters obtained through training are represented as follows: .
[0062] Using a small initial calibration dataset of the target users The trained base model is fine-tuned to obtain the initial individual model. .
[0063] Simultaneously, an individual-level historical memory database is initialized for this user, designed for a single user, to store high-confidence samples selected during subsequent long-term use. The data stored in the historical memory database includes at least: individual multimodal data, prediction results of the individual model, real feedback labels, and corresponding time information.
[0064] Preferably, the basic model based on the deep neural network structure adopts one or more combinations of convolutional neural networks, recurrent neural networks, and Transformer networks to extract discriminative features from multimodal signals and output classification results.
[0065] Preferably, the initial individual model is obtained by fine-tuning some or all of the model parameters of the trained base model.
[0066] S3. Online prediction and latent feature extraction: Real-time prediction of newly acquired individual multimodal data during the long-term online operation of the brain-computer interface system.
[0067] Individual model for the current moment Collected individual multimodal data Perform inference and output the predicted probability vector:
[0068] ;
[0069] in, The class probability distribution output by the individual model. For the number of task categories, These represent the samples being predicted as the first. The probability value of the class; The model parameters are: Individual model.
[0070] Preferably, while the model is making predictions, latent feature representations of samples are extracted from the intermediate layers of the model. This is used to describe the distribution location of the sample in the feature space, for subsequent confidence assessment and distribution consistency judgment.
[0071] S4. Confidence Assessment and Sample Quality Determination: A multi-dimensional confidence assessment is performed on the prediction results of the individual model. The confidence assessment is based on at least one of the following confidence assessment indicators:
[0072] 1. Predicted Probability. Calculate the maximum probability value corresponding to the predicted class based on the class probability distribution output by the individual model, as shown in the following formula:
[0073] ;
[0074] in, express The time sample is predicted by the individual model as the first time. The probability value of the class. This represents the category index (category number) output by the individual model. express The score of the predicted probability of a sample at time step is the maximum probability value that the sample is predicted to be of a certain class, which is used to measure the confidence of the model in the classification result of the sample.
[0075] 2. Category Interval. The category interval is calculated based on the probability difference between the category with the highest predicted probability and the category with the second highest predicted probability, as shown in the following formula:
[0076] ;
[0077] in, express The score of the class interval of the sample at time step, that is, the probability difference between the class with the highest predicted probability and the second highest predicted probability, is used to measure the discriminative power of the model classification results; and These represent the categories with the highest and second-highest predicted probabilities, respectively.
[0078] 3. Distribution uncertainty. By performing multiple random inferences, random inactivation inferences, or model ensemble inferences on the same sample, the variance of the prediction results (i.e., the uncertainty index) is calculated and defined as follows:
[0079] ;
[0080] in, express The score for the uncertainty of the distribution of the sample at time point is used to quantify the certainty of the model's prediction result for that sample (the higher the value, the more stable the model prediction and the lower the uncertainty). This represents the variance calculation operator, used to calculate the dispersion of prediction results obtained from multiple inferences, and measures the uncertainty of model predictions. express The set of all predicted probabilities obtained after multiple random inferences on the time sample.
[0081] 4. Latent Feature Space Similarity. This is achieved by calculating the latent feature representation of the current sample. (In the latent feature space of the individual model) and sample features in the historical memory database The cosine similarity between samples is used to assess the consistency between the samples and historical distributions, as shown in the following formula:
[0082] ;
[0083] in, express The score of the latent feature space similarity of the current sample and the sample features in the historical memory database is used to evaluate the consistency between the current sample and the historical data distribution. This represents the historical memory database corresponding to the individual model; express The feature vector of the sample at time step in the model's latent feature space; express The Middle The latent feature vector of a historical sample Represents the norm.
[0084] 5. Signal quality indicators are calculated based on the signal-to-noise ratio, artifact level, or channel stability of EEG signals or other physiological parameter signals.
[0085] In this embodiment, one or more of the above-mentioned confidence assessment indicators are weighted and fused according to preset weights to obtain the comprehensive confidence level of the sample. :
[0086] ;
[0087] in, For evaluation Individual multimodal data at time 1 Scores on each confidence assessment metric This represents the total number of confidence assessment metrics (i.e., dimensions). For the first Each confidence assessment dimension corresponds to a preset normalized weight coefficient, and .
[0088] S5. Select high-quality samples based on the confidence assessment results.
[0089] In this embodiment, when the overall confidence level of a sample meets the following conditions, the sample and its true feedback label are stored in the historical memory database:
[0090] ;
[0091] in, For a preset threshold, a fixed threshold or an adaptive threshold can be used. The adaptive threshold is dynamically adjusted based on the statistical characteristics of historical samples, the confidence distribution of historical samples, the model prediction performance, or the statistical characteristics of online data.
[0092] When the confidence level of a sample is low, it will not participate in the model update to avoid introducing noisy samples into the model and causing performance degradation.
[0093] The real feedback label is provided by the user through active confirmation, automatic feedback of task execution results, or external system.
[0094] S6. Model update and cross-day drift suppression.
[0095] like Figure 2 As shown, the individual model is updated when preset conditions are met.
[0096] During the model update phase, based on the current online samples (denoted as...) And extract a predetermined proportion of data (i.e., historical samples, denoted as) from the historical memory database. By setting different training weights, learning rates, or loss function constraints, individual models are trained together; where the loss function is set as follows:
[0097] ;
[0098] in, Represents the total loss item. Indicates the current online sample The loss item, Representing historical samples The loss item, These are the weighting coefficients.
[0099] By employing the above methods, individual models can adapt to new data distributions while retaining their memory of historical distributions, thereby effectively suppressing cross-day signal drift and mitigating catastrophic forgetting.
[0100] Preferably, step S6 is executed periodically, or it is set to be triggered when the number of samples stored in the historical memory database, the time span, or the degree of change in distribution reaches a preset condition.
[0101] Example 2:
[0102] In this embodiment, a long-term brain-computer interface access cross-day drift suppression adaptive update system is provided. The system includes a data acquisition module, a model training module, an online prediction module, a confidence assessment module, and a joint optimization module.
[0103] The data acquisition module is used to collect individual multimodal data from several users as group multimodal data; the individual multimodal data includes electroencephalogram (EEG) signals and physiological parameter signals;
[0104] The model training module is used to select a basic model based on a deep neural network structure, train it according to the group multimodal data, and fine-tune it according to the calibrated individual multimodal data to obtain an initial individual model and construct a historical memory database corresponding to the individual model.
[0105] The online prediction module is used to make real-time predictions on newly collected individual multimodal data based on the individual model during online use.
[0106] The confidence assessment module is used to assess the confidence of the prediction results. The confidence assessment indicators include one or more of the following: prediction probability, class interval, distribution uncertainty, latent feature space similarity, and signal quality indicators. Based on the confidence assessment results, individual multimodal data with confidence scores higher than a preset threshold and their corresponding real feedback labels are selected and stored in the historical memory database.
[0107] The joint optimization module uses incremental learning to jointly optimize the model parameters of individual models based on current online data and historical memory database, thereby suppressing cross-day signal distribution drift.
[0108] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above. This system is the cross-day drift suppression adaptive update method for long-term access to brain-computer interfaces in the above embodiments.
[0109] Example 3:
[0110] In this embodiment, a storage medium is also provided, storing a program. When the program is executed by a processor, it implements the cross-day drift suppression adaptive update method for long-term brain-computer interface access described in the above embodiment, specifically:
[0111] Individual multimodal data from several users are collected as group multimodal data; the individual multimodal data includes electroencephalogram (EEG) signals and physiological parameter signals;
[0112] A basic model based on a deep neural network structure is selected, trained using group multimodal data, and fine-tuned using calibrated individual multimodal data to obtain an initial individual model, and a historical memory database corresponding to the individual model is constructed.
[0113] During online use, individual models are used to make real-time predictions on newly collected individual multimodal data;
[0114] Set confidence evaluation indicators to evaluate the confidence of the prediction results. The confidence evaluation indicators include one or more of the following: prediction probability, class interval, distribution uncertainty, latent feature space similarity, and signal quality indicators.
[0115] Based on the confidence assessment results, individual multimodal data with confidence scores higher than a preset threshold and their corresponding real feedback labels are selected and stored in the historical memory database;
[0116] Incremental learning is used to jointly optimize the model parameters of individual models based on current online data and historical memory databases, thereby suppressing cross-day signal distribution drift.
[0117] Example 4:
[0118] This embodiment uses a brain-computer interface system based on a motor imagery task as a specific application scenario to illustrate the practical application process of the cross-day drift suppression adaptive update method and system for long-term brain-computer interface access proposed in this invention.
[0119] (a) Application scenario description.
[0120] In this embodiment, the brain-computer interface system is used to identify the user's motor imagination intentions, such as left-hand motor imagination, right-hand motor imagination, two-foot motor imagination, or resting state, and uses the recognition results to control external devices, such as rehabilitation training robots, wheelchairs, robotic arms, or virtual interactive interfaces.
[0121] Because the distribution of EEG signals varies significantly between different days of use, factors such as changes in electrode placement, skin impedance, and fatigue levels cause the accuracy of traditional fixed models to gradually decline over long-term use. This embodiment introduces the long-term adaptive update method proposed in this invention to achieve stable operation of the system in multi-day continuous use scenarios.
[0122] (ii) Data collection and task paradigm design.
[0123] In this embodiment, the EEG acquisition electrodes are arranged using the international 10-20 system, focusing on acquiring channel signals located in the motor cortex region, such as channels C3, C4, and Cz. The sampling frequency is preferably 250Hz, 500Hz, or 1000Hz.
[0124] The paradigm for the motion visualization task is set as follows:
[0125] The system presents visual cues to the user, guiding them to perform left-hand or right-hand motor imagery tasks. Each task lasts 3 to 5 seconds, with a 2 to 4-second interval between tasks. The system simultaneously records EEG signal data and labels it with the corresponding task category.
[0126] In addition to EEG signals, near-infrared signals or electrooculogram (EOG) signals can be acquired simultaneously to increase modal information or remove eye movement artifacts and improve the accuracy of the output.
[0127] (III) Model initialization and individual model construction.
[0128] When using the system for the first time, a small amount of initial individual calibration data is collected, and 10 to 30 trial data are collected for each type of task to fine-tune the basic model based on the deep neural network structure.
[0129] A base model is trained using EEG data from multiple users' group motor imagery. The trained base model is then fine-tuned to obtain an initial individual model suitable for that user. Simultaneously, the user's historical memory database is initialized to store high-confidence samples selected in subsequent online phases.
[0130] (iv) Online prediction and confidence-driven sample selection.
[0131] During the long-term online operation of the system, users continuously perform motor imagery tasks, and the system collects EEG signals in real time and inputs them into individual models for prediction.
[0132] The model outputs the corresponding class probability distribution and extracts latent feature representations from the intermediate layers of the network.
[0133] The system calculates a comprehensive confidence score based on predicted probability, class interval, uncertainty assessment results, feature similarity, and signal quality indicators.
[0134] When the confidence level of a sample is higher than a preset threshold, the sample and its true feedback label are stored in the historical memory database.
[0135] When the confidence level is low, it will not participate in the model update to avoid introducing noisy samples into the model and causing performance degradation.
[0136] (v) Cross-day model update.
[0137] During cross-day usage, the system executes model update strategies periodically or triggered by specific events.
[0138] During the update phase, training data is extracted from the currently collected online data and high-confidence samples in historical memory, and the model parameters of individual models are optimized by using incremental learning combined with historical replay mechanism.
[0139] Experimental results show that, under continuous multi-day use, the brain-computer interface system in this embodiment can maintain a stable recognition accuracy, which is significantly better than the comparison method without adaptive updates, while effectively suppressing cross-day signal drift and mitigating catastrophic forgetting.
[0140] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0141] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A cross-day drift suppression adaptive update method for long-term brain-computer interface access, characterized in that, Includes the following steps: Individual multimodal data from several users are collected as group multimodal data; the individual multimodal data includes electroencephalogram (EEG) signals and physiological parameter signals; A basic model based on a deep neural network structure is selected, trained using group multimodal data, and fine-tuned using calibrated individual multimodal data to obtain an initial individual model, and a historical memory database corresponding to the individual model is constructed. During online use, individual models are used to make real-time predictions on newly collected individual multimodal data; Set up confidence assessment indicators to assess the confidence level of the real-time prediction results; Based on the confidence assessment results, individual multimodal data with confidence scores higher than a preset threshold and their corresponding real feedback labels are selected and stored in the historical memory database; Incremental learning is used to jointly optimize the model parameters of individual models based on current online data and historical memory databases.
2. The cross-day drift suppression adaptive update method for long-term brain-computer interface access according to claim 1, characterized in that, The physiological parameter signals include at least one of near-infrared signals, electromyography signals, electrooculography signals, electrocardiogram signals, electrodermal signals, and respiratory signals.
3. The cross-day drift suppression adaptive update method for long-term brain-computer interface access according to claim 1, characterized in that, The individual multimodal data undergoes at least one of the following processes after acquisition: time synchronization, filtering and denoising, artifact removal, normalization, and feature extraction.
4. The cross-day drift suppression adaptive update method for long-term brain-computer interface access according to claim 1, characterized in that, The underlying model based on deep neural network architecture adopts one or a combination of the following structures: convolutional neural network, recurrent neural network, and Transformer network.
5. The cross-day drift suppression adaptive update method for long-term brain-computer interface access according to claim 1, characterized in that, During training using multimodal population data, the objective function is set as follows: ； in, The loss function for supervised learning; For group multimodal data; The model parameters are: The basic model for input data The prediction results; For input data Corresponding real feedback tags; For mathematical expectation operators; For parameter optimization operators, used to solve for the model parameters that minimize the objective function. ; These are the optimal model parameters obtained through training.
6. The cross-day drift suppression adaptive update method for long-term brain-computer interface access according to claim 1, characterized in that, The confidence assessment metrics include one or more of the following: prediction probability, class interval, distribution uncertainty, latent feature space similarity, and signal quality metrics. The predicted probability is specifically: the maximum probability value corresponding to the predicted category calculated based on the category probability distribution output by the individual model; The category interval is calculated based on the probability difference between the category with the highest predicted probability and the category with the second highest predicted probability; The distribution uncertainty specifically refers to the uncertainty of the prediction result calculated by performing multiple random inferences, random deactivation inferences, or model ensemble inferences on the same individual multimodal data. The latent feature space similarity specifically refers to: calculating the similarity between the current individual multimodal data in the latent feature space of the individual model and the features in the historical memory database; The signal quality index is calculated based on the signal-to-noise ratio, artifact level, or channel stability of the collected individual multimodal data.
7. The cross-day drift suppression adaptive update method for long-term brain-computer interface access according to claim 6, characterized in that, The confidence level of the prediction results is assessed, specifically as follows: The overall confidence level is obtained by weighting and fusing one or more confidence assessment indicators: ； in, For evaluation Individual multimodal data at time 1 Scores on each confidence assessment metric To assess the overall confidence level, This represents the total number of confidence assessment indicators. For the first Each confidence assessment indicator corresponds to a preset normalized weight coefficient, and .
8. The cross-day drift suppression adaptive update method for long-term brain-computer interface access according to claim 1, characterized in that, The joint optimization of the model parameters of the individual models specifically involves: A predetermined proportion of data is extracted from the historical memory database and used in conjunction with current online data to train individual models; the loss function is set as follows: ； in, Represents the total loss item. This represents the loss term of the currently online data. This represents the loss term of data extracted from the historical memory database. These are the weighting coefficients.
9. The cross-day drift suppression adaptive update method for long-term brain-computer interface access according to claim 1, characterized in that, The joint optimization is performed periodically, or it is set to be triggered when the amount of data, time span, or distribution change of the data stored in the historical memory database reaches a preset condition.
10. A long-term brain-computer interface (BCI) system for cross-day drift suppression and adaptive update, characterized in that: The cross-day drift suppression adaptive update method for long-term access to brain-computer interfaces according to any one of claims 1-9 includes a data acquisition module, a model training module, an online prediction module, a confidence assessment module, and a joint optimization module; The data acquisition module is used to collect individual multimodal data from several users as group multimodal data; the individual multimodal data includes electroencephalogram (EEG) signals and physiological parameter signals; The model training module is used to select a basic model based on a deep neural network structure, train it according to the group multimodal data, and fine-tune it according to the calibrated individual multimodal data to obtain an initial individual model and construct a historical memory database corresponding to the individual model. The online prediction module is used to make real-time predictions on newly collected individual multimodal data based on the individual model during online use. The confidence assessment module is used to assess the confidence of the real-time prediction results; based on the confidence assessment results, individual multimodal data with confidence scores higher than a preset threshold and their corresponding real feedback labels are selected and stored in the historical memory database. The joint optimization module uses incremental learning to jointly optimize the model parameters of individual models based on current online data and historical memory database.