A personalized brain-computer interface system

Abstract

The application discloses a kind of individualized brain-computer interface systems, including electroencephalogram data engineering module, individualized paradigm library module, individualized paradigm selection module, model matching and selection module, offline individualized module and online decoding and adaptive module, each module cooperates, realizes the whole process individualization from paradigm to decoding model, effectively considers the user experience, individualized depth and long-term decoding performance of BCI system.

Description

Technical Field

[0001] This invention belongs to the technical field of brain-computer interfaces, specifically relating to a personalized brain-computer interface system. Background Technology

[0002] Brain-computer interface (BCI) systems enable information interaction between users and external devices by decoding brain activity signals. Among these, EEG-based BCIs have shown great application potential in fields such as medical rehabilitation, neural engineering, and human-computer interaction due to their non-invasiveness, portability, and relatively low cost. However, current EEG-based BCI technology still faces many challenges in practical applications.

[0003] First, due to significant individual differences in EEG patterns, different users exhibit marked variations in their EEG response patterns to the same experimental paradigm. Furthermore, the EEG signals of the same user at different times and under different physiological states also show non-stationarity. This substantial intra- and inter-individual variability leads to low decoding accuracy and a poor user experience for new users in traditional BCI systems with fixed paradigms and universal models. Second, most existing BCI systems adopt a "one-size-fits-all" strategy, providing all users with the same experimental paradigm and decoding model. While some studies have attempted to fine-tune the model using calibration data, a personalized solution for the entire process from paradigm selection to model decoding is lacking. Moreover, traditional methods require users to undergo lengthy and tedious calibration training to obtain a usable decoding model, placing a heavy burden on users, easily causing fatigue, and hindering the widespread adoption of the system. Finally, most online decoding systems are fixed in use after offline training and cannot adapt to changes in EEG signals over time or caused by user learning effects.

[0004] Therefore, there is an urgent need for a new type of BCI system that can achieve deep personalization from source paradigm selection to terminal model parameters, and can be efficiently initialized and stably adapted online. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a personalized brain-computer interface system. Through user-selected paradigms, intelligent matching of decoding models by the system, and incremental learning for online adaptation, the system achieves full-process personalization from paradigm to decoding model, effectively balancing the user experience, personalization depth, and long-term decoding performance of the BCI system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a personalized brain-computer interface system, comprising:

[0008] The EEG data engineering module is used to collect the user's EEG signals and preprocess them.

[0009] The Personalized Paradigm Library module is used to store pre-validated brain-computer interface experimental paradigms. The paradigms use relational data storage for paradigm metadata. Each paradigm includes a paradigm ID, paradigm type, paradigm description, pre-validation metric, and recommendation index.

[0010] The personalized paradigm selection module is used to display brain-computer interface experimental paradigms and select brain-computer interface experimental paradigms based on quantitative indicators and user feedback.

[0011] The model matching and selection module is used to determine at least one candidate decoding model suitable for the target paradigm according to the preset paradigm-model correspondence, and to determine an initial decoding model from the candidate decoding models based on the user's calibrated EEG data.

[0012] The offline personalization module is used to set the initial personalization configuration parameters and train the initial decoding model using reinforcement learning to obtain the personalized initial decoding model.

[0013] The online decoding and adaptive module is used to decode the user's real-time EEG signals using a personalized initial decoding model, and to continuously optimize the personalized initial decoding model using an incremental learning approach based on the new EEG data and decoding results generated during the user's online use.

[0014] As a preferred technical solution, the EEG data engineering module acquires multi-channel raw EEG signal sequences at a set time and performs preprocessing on the multi-channel raw EEG signal sequences, including bandpass filtering, notch filtering, artifact removal, baseline correction, rereference, and downsampling.

[0015] As a preferred technical solution, the paradigm ID is a unique identifier;

[0016] The paradigm types include motion imagination paradigm, visual motion imagination paradigm, spatial navigation imagination paradigm, SSVEP paradigm, P300 paradigm, voice imagination paradigm, music imagination paradigm, and image imagination paradigm.

[0017] The paradigm description includes textual and graphical explanations that illustrate how the paradigm works and the mental tasks that users need to perform.

[0018] The pre-verification metrics include decoding accuracy and average information transmission rate;

[0019] The recommendation index is a score calculated based on historical user selection feedback.

[0020] As a preferred technical solution, the personalized paradigm library module collects subjective comfort scores, calculates paradigm scores according to the decoding accuracy and average information transmission rate, and stores the corresponding pre-validated brain-computer interface experimental paradigms based on the paradigm scores.

[0021] As a preferred technical solution, the quantitative indicators include decoding performance indicators and user subjective experience indicators;

[0022] The system performs a weighted comprehensive score on the quantitative indicators, as shown in the following formula:

[0023] ,

[0024] ,

[0025] in, For comprehensive evaluation and scoring, For decoding performance metrics, This refers to subjective user experience metrics. As the weight of decoding performance metrics, Weighting of user subjective experience metrics;

[0026] The decoding performance metrics are as follows:

[0027] ,

[0028] in, Let be the classification accuracy of the i-th classification accuracy in the classification accuracy sequence. Let i be the i-th information transmission rate in the information transmission rate sequence. The maximum value in the classification accuracy sequence. This represents the maximum value in the information transmission rate sequence. The weights for classification accuracy, Weights for information transmission rate;

[0029] The user subjective experience indicators are specifically as follows:

[0030] ,

[0031] in, Score the subjective difficulty level. Rate the degree of fatigue. For focus self-assessment, For subjective difficulty rating weights, As the weight of the fatigue level score, Weighting for focus level self-assessment.

[0032] As a preferred technical solution, the model matching and selection module maps the selected brain-computer interface experimental paradigm to the corresponding decoding model. Each type of brain-computer interface experimental paradigm is mapped to a set of decoding models, and each set of decoding models has at least one candidate decoding model.

[0033] The comprehensive score based on decoding performance metrics and model complexity is calculated as follows:

[0034] ,

[0035] in, Let be the classification accuracy of the i-th element in the sequence. Let i be the information transmission rate of the i-th element in the sequence. For model complexity, The maximum value in the classification accuracy sequence. This represents the maximum value in the information transmission rate sequence. It is the maximum value in the inverse sequence of model complexity. The weights for classification accuracy, As a weight for information transmission rate, These are the weights for model complexity;

[0036] The optimal decoding model, i.e. the initial decoding model, is obtained based on the highest comprehensive score.

[0037] As a preferred technical solution, the setting of initial personalized configuration parameters includes at least one of the following:

[0038] A feature importance vector used to weight input EEG features;

[0039] The set of hyperparameters of the initial decoding model;

[0040] The initial decoding model includes some network structure adjustment parameters.

[0041] As a preferred technical solution, the reinforcement learning adopts an actor-critic architecture, wherein:

[0042] The state is a concatenation of historical average features and current features;

[0043] The action is updated according to the type of personalized goal. Specifically, when the personalized goal is feature selection, the feature selection method is adopted and the action is a feature weight vector, where each element represents the retention weight of the corresponding feature. When the personalized goal is parameter optimization, the hyperparameter optimization method is adopted and the action is the hyperparameter adjustment amount.

[0044] Based on the updated actions, obtain the optimal network parameters, and then configure a personalized initial decoding model based on the optimal network parameters.

[0045] As a preferred technical solution, the online decoding and adaptive module includes:

[0046] A data buffer is used to cyclically store newly generated EEG data and decoding results from the user;

[0047] The performance detection unit is used to detect the decoding performance of the personalized initial decoding model in real time.

[0048] The incremental update unit is used to analyze the decoding performance. When the decoding performance is lower than a preset threshold or reaches a preset update cycle, the stored data is used to fine-tune the initial decoding model, wherein the learning rate during fine-tuning is lower than the training learning rate of the offline personalization module.

[0049] As a preferred technical solution, the fine-tuning performed by the incremental update unit is limited to updating the classification layer parameters of the personalized initial decoding model, or updating the network layer parameters that are associated with the high weights of the personalized configuration parameters.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] First, this invention fundamentally changes the traditional "one-size-fits-all" interaction mode of BCI systems by allowing users to autonomously choose personalized experimental paradigms for brain-computer interfaces based on subjective comfort, significantly improving user acceptance and willingness to use them. Second, this invention constructs a complete personalized technology chain by implementing a three-level personalized architecture of "paradigm level + model level + parameter level," from interaction personalization paradigm selection and decoding model matching to internal parameter optimization. Third, this invention greatly reduces the user burden by rapidly searching for the optimal personalized configuration with limited calibration data through reinforcement learning. Fourth, this invention solves the performance degradation problem caused by the non-stationarity of EEG signals by using incremental learning for adaptation, ensuring continuous model optimization while avoiding catastrophic forgetting. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram of the structure of the personalized brain-computer interface system according to an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the human-computer interaction interface of the model personalization paradigm selection module in an embodiment of the present invention;

[0055] Figure 3This is a schematic diagram of the workflow of the model matching and selection module in an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram illustrating the working principle of the personalized brain-computer interface system according to an embodiment of the present invention. Detailed Implementation

[0057] 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.

[0058] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0059] Please see Figure 1 This embodiment provides a personalized brain-computer interface system 10, including an EEG data engineering module 11, a personalized paradigm library module 12, a personalized paradigm selection module 13, a model matching and selection module 14, an offline personalization module 15, and an online decoding and adaptation module 16. The modules work together to achieve full-process personalization from paradigm to decoding model, effectively balancing the user experience, personalization depth, and long-term decoding performance of the BCI system.

[0060] Specifically, the system architecture is as follows:

[0061] The EEG data engineering module 11 is used to collect the user's EEG signals and preprocess the EEG signals.

[0062] In this embodiment, the EEG data engineering module uses a non-invasive EEG acquisition device to collect the user's EEG signals. The acquisition device includes an EEG cap and a multi-channel electrode array. The electrodes are arranged on the user's scalp surface according to the international 10–20 system. The number of channels is preferably 32 to 64, and the sampling frequency is preferably 250 Hz to 1000 Hz.

[0063] Let the raw EEG signals collected within time period 𝑡 be represented as:

[0064] ,

[0065] in, For the number of channels, Indicates the number of sampling time points.

[0066] The acquired raw EEG signals are first preprocessed. In this embodiment, the preprocessing includes the following steps:

[0067] Bandpass filtering is applied to remove low-frequency drift and high-frequency noise.

[0068] Notch filtering is used to suppress 50 Hz power frequency interference;

[0069] Artifact removal employs Independent Component Analysis (ICA) to identify and remove artifacts such as electrooculography (EOG) and electromyography (EMG).

[0070] If the reference electrode is loose, the filtered signal is subjected to common average reference processing.

[0071] Remove bad conductors and bad segments, eliminating electrodes and signal segments with poor signal quality;

[0072] Downsampling: To reduce computation and speed up the process, the collected EEG signals are downsampled to 250 Hz.

[0073] After the above processing, a preprocessed EEG signal matrix is ​​obtained, which serves as the input data for the subsequent decoding and learning modules.

[0074] The personalized paradigm library module 12 stores pre-validated brain-computer interface experimental paradigms. The paradigms use relational data storage for paradigm metadata, with each paradigm including a paradigm ID, paradigm type, paradigm description, pre-validation metrics, and a recommendation index. Specifically,

[0075] The paradigm ID is a unique identifier;

[0076] The paradigm types include motion imagination paradigm, visual motion imagination paradigm, spatial navigation imagination paradigm, SSVEP paradigm, P300 paradigm, voice imagination paradigm, music imagination paradigm, and image imagination paradigm.

[0077] The paradigm description includes textual and graphical explanations that illustrate how the paradigm works and the mental tasks that users need to perform.

[0078] The pre-validation metrics include decoding accuracy on small sample pre-experiments. Average information transmission rate ;

[0079] The recommendation index is a score calculated based on historical user selection feedback.

[0080] Each paradigm has undergone rigorous validation. During the pre-experiment phase, a small number of volunteers are recruited for each paradigm to conduct small-scale experiments and collect subjective comfort ratings. By reviewing literature and actually running various models to obtain different accuracies and average information transmission rates, the score of this paradigm is calculated according to the following formula. To determine whether to put the item into storage:

[0081] ,

[0082] The weighting coefficients are α+β+γ=1.

[0083] The personalized paradigm selection module 13 is used to display brain-computer interface experimental paradigms and select brain-computer interface experimental paradigms based on quantitative indicators and user feedback.

[0084] like Figure 2 As shown, this embodiment provides a user-friendly human-computer interaction interface to guide users in completing the paradigm selection. On the paradigm display page, paradigms are displayed in card format, including paradigm ID, icon, name, brief description, and recommendation index. The paradigm selection process is as follows:

[0085] First, the system presents users with multiple candidate paradigms, and users select one or more candidate paradigms as a preliminary set of paradigms based on their own understanding difficulty and subjective preferences.

[0086] Subsequently, the system collected short-term pre-experimental EEG data under each candidate paradigm and evaluated the fit of each candidate paradigm. The quantitative indicators include, but are not limited to:

[0087] Decoding performance metrics are as follows:

[0088] ,

[0089] in, Let be the classification accuracy of the i-th classification accuracy in the classification accuracy sequence. Let i be the i-th information transmission rate in the information transmission rate sequence. The maximum value in the classification accuracy sequence. This represents the maximum value in the information transmission rate sequence. The weights for classification accuracy, The weight is the information transmission rate.

[0090] User subjective experience metrics are as follows:

[0091] ,

[0092] in, Score the subjective difficulty level. Rate the degree of fatigue. For focus self-assessment, For subjective difficulty rating weights, As the weight of the fatigue level score, The system assigns weight to the self-assessment of focus levels. The system then performs a weighted comprehensive score based on the above indicators:

[0093] ,

[0094] in, For comprehensive evaluation and scoring, For decoding performance metrics, This refers to subjective user experience metrics. As the weight of decoding performance metrics, Weighting of user subjective experience metrics.

[0095] The models used for decoding in each candidate paradigm are the best-performing models in the pre-experiments. , , . .

[0096] Ultimately, the paradigm with the highest overall score was selected as the user's personalization paradigm.

[0097] like Figure 3 As shown, the model matching and selection module 14 is used to determine at least one candidate decoding model applicable to the target paradigm according to the preset paradigm-model correspondence, and to determine the initial decoding model from the candidate decoding models based on the user's calibrated EEG data.

[0098] In this embodiment, the model matching and selection module is a key bridge connecting paradigm selection and personalized learning, realizing intelligent mapping from paradigm to specific decoding model. The system pre-stores the following mapping table:

[0099] ,

[0100] Where M is the major paradigm category (e.g., the motor imagination category), P is the minor category based on M (e.g., the fist-clenching imagination category), and m1, m k Each paradigm type corresponds to a decoding model, and this mapping table maps each paradigm type to a set of the most suitable decoding models. This mapping is based on:

[0101] (1) Review the top journals and conferences of the past 5 years and compile statistics on the commonly used models for each paradigm;

[0102] (2) Test the data collected in the pre-experiment and evaluate the performance of different models.

[0103] When the user selects a paradigm Then, the system loads the corresponding k models {m1,…,m k The user conducts calibration experiments, obtaining N trials. The standard calibration data is then divided into a training set in a 7:3 ratio. and verification set Calculate the accuracy under each model. and The optimal model is selected using a multi-objective optimization method:

[0104] ,

[0105] in, Let be the classification accuracy of the i-th classification accuracy in the classification accuracy sequence. Let i be the i-th information transmission rate in the information transmission rate sequence. For model complexity, The maximum value in the classification accuracy sequence. This represents the maximum value in the information transmission rate sequence. The maximum value in the inverse sequence of model complexity is λ1+λ2+λ3=1. Therefore, we choose... As the initial decoding model.

[0106] The offline personalization module 15 is used to set the initial personalization configuration parameters, train the initial decoding model using reinforcement learning, and obtain the trained personalized initial decoding model.

[0107] The initial personalized configuration parameters include at least one of the following:

[0108] A feature importance vector used to weight input EEG features;

[0109] The set of hyperparameters of the initial decoding model;

[0110] The initial decoding model includes some network structure adjustment parameters.

[0111] When training the initial decoding model using reinforcement learning, a personalized configuration is quickly learned for the selected initial decoding model based on user calibration data. Reinforcement learning employs an actor-critic architecture, specifically:

[0112] state It is composed of historical average features and current features:

[0113] ,

[0114] in, It is the EEG feature vector at time t, and d is the feature dimension. This indicates the average operation.

[0115] action There are two design options depending on the individualized goals:

[0116] When the personalization goal is feature-level personalization, the object of action is the input feature space, and the output is the feature importance weights. That is, feature selection is used to update the action, as shown in the following formula:

[0117] ,

[0118] in, For feature dimension, For the sigmoid function, Represents the weight matrix. This indicates the bias term.

[0119] Each element in the output represents the retention weight of the corresponding feature.

[0120] When the personalization goal is parametric personalization, the object of the action is the model parameter space, and the output is the hyperparameter adjustment. The action is updated using hyperparameter optimization, as shown in the following formula:

[0121] ,

[0122] in, This represents the adjustment amount of the hyperparameters for the learning rate, regularization function, and dropout rate. This is the transpose symbol.

[0123] reward function The core design principle is a relative improvement in performance:

[0124]

[0125] in The loss of the model on the validation set when it is not personalized. To apply the action in step t The new losses that followed.

[0126] The actor network parameters are updated using a gradient strategy, as shown in the following equation:

[0127] ,

[0128] in, For policy gradient, Let the policy objective function be... Expectation operator For the logarithmic policy gradient, For the dominant function, For the first Step state.

[0129] It's worth explaining that feature-level personalization operates on the feature space, aiming to make the model focus on the feature channels or frequency bands most important to the current user. In this case, the action is a feature weight vector, used in scenarios requiring the selection of input features. Parameter-level personalization operates on the model space, aiming to optimize the hyperparameter configuration of the model training process itself. This action is a hyperparameter adjustment, suitable for scenarios requiring adjustments to the learning strategy. These two approaches are parallel and can be chosen based on specific application needs in practice.

[0130] The online decoding and adaptive module 16 is used to decode the user's real-time EEG signals using a trained personalized initial decoding model, and to continuously optimize the personalized initial decoding model using an incremental learning method based on the EEG data and decoding results generated during the user's online use.

[0131] In this embodiment, the online decoding and adaptive module is responsible for the real-time operation and long-term performance maintenance of the system. To cope with the non-stationarity of EEG signals, the system adopts triggered incremental learning.

[0132] The online decoding and adaptive module further includes:

[0133] Data buffer 161 is used to cyclically store newly generated EEG data and decoding results from the user;

[0134] Specifically, maintain a fixed-capacity B first-in-first-out buffer. Each time a control operation is completed, the data is transferred to... Stored in the buffer, where To infer the true label based on the control results (successfully) =Prediction correct, failure =Prediction failed).

[0135] Define sliding window performance metrics:

[0136] ,

[0137] Where W is the window size. For indicator functions, This indicates the true label inferred from the control results.

[0138] The performance detection unit 162 is used to detect the decoding performance of the personalized initial decoding model in real time.

[0139] The incremental update unit 163 analyzes decoding performance. When the decoding performance falls below a preset threshold or reaches a preset update cycle, it fine-tunes the initial decoding model using stored data. The learning rate during fine-tuning is lower than the training learning rate of the offline personalization module. Fine-tuning by the incremental update unit is limited to updating the classification layer parameters of the personalized initial decoding model, or updating the network layer parameters with high weights associated with the personalized configuration parameters. The main logic here is that reinforcement learning identifies which features or network layer parameters are important to the current user's decoding performance (assigning them high weights) during the learning process, and then makes more detailed adjustments to these important parts (high-weight parts) during subsequent training and online fine-tuning.

[0140] An update is triggered when the performance of a continuous window falls below a threshold τ. The update triggering conditions are as follows:

[0141] ,

[0142] When an update is triggered, from the buffer Take all B data samples, construct an incremental learning loss function, and perform gradient updates. Indicates the number of consecutive windows.

[0143] Through the collaborative work of the above modules, the problems of insufficient personalization, time-consuming calibration, and poor online adaptability of existing BCI systems are effectively solved.

[0144] Example 2: Application example of brain-controlled mobile device based on personalized brain-computer interface.

[0145] In a more specific scenario, this embodiment uses a brain-controlled mobile device as a specific application scenario to illustrate the actual application effect of the personalized brain-computer interface system described in this invention, but it should not be construed as a limitation on the scope of application of this invention.

[0146] I. Application Scenarios

[0147] In certain application scenarios, brain-computer interface technology has broad application prospects in the field of intelligent assisted control, such as brain-controlled cars and wheelchairs. In these applications, users typically send control commands directly to external devices via electroencephalogram (EEG) signals to perform basic operations such as moving forward, backward, and turning.

[0148] However, in practical applications, there are significant differences in the spatial distribution, spectral characteristics, and decodeability of EEG signals among different users. Furthermore, some users exhibit "unusable" or "low decodeability" brain-computer interfaces under specific experimental paradigms, a phenomenon known as brain-computer interface illiteracy (BCI illiteracy). Traditional brain-control systems employing fixed paradigms and unified decoding models struggle to simultaneously meet the diverse needs of different users, resulting in poor system compatibility, high training costs, and limited practical usability.

[0149] To address the aforementioned issues, this invention proposes a brain-computer interface system that supports user-initiated paradigm selection and personalized model optimization, thereby enhancing the system's adaptability to different users and improving the success rate and stability of brain-controlled devices in real-world application scenarios.

[0150] II. System Application Methods.

[0151] like Figure 4 As shown, in this embodiment, the personalized brain-computer interface system is connected to the mobile device control system. The mobile device control system can be a brain-controlled car, a brain-controlled wheelchair, or other mobile platform that can respond to control commands, and is used to realize motion control such as forward, backward, left turn, right turn, and stop.

[0152] The user first wears an EEG acquisition device to collect signals. In this embodiment, the EEG acquisition device is a 32-lead EEG cap, with the electrode arrangement conforming to the international 10–20 system standard, where A1 and A2 are used as reference electrodes, and the sampling frequency is set to 1000 Hz.

[0153] After wearing the device, users can browse a personalized paradigm library through the system interface. The paradigm library includes, but is not limited to, motion imagery paradigms, steady-state visual evoked potential (SSVEP) paradigms, and image imagery paradigms. The system provides task descriptions and difficulty reference information for each paradigm. Based on their own level of understanding and preferences, users can select the left and right hand clenched fist motion imagery paradigm and the image imagery paradigm as candidate paradigms.

[0154] The system then conducted short-term pre-experimental data acquisitions of approximately 5–10 minutes under the aforementioned candidate paradigms. The acquired data underwent bandpass filtering (0.5–40 Hz), notch filtering (50 Hz), artifact removal based on independent component analysis, and was downsampled to 250 Hz before being used as decoding input data.

[0155] The system evaluates the suitability of two candidate paradigms. Quantitative indicators include decoding accuracy, information transmission rate, and user subjective ratings. The subjective ratings are collected through a standardized questionnaire, covering dimensions such as perceived task difficulty, focus level, and fatigue level.

[0156] In one implementation scenario, the system evaluation results showed that the left and right hand clenched fist paradigm outperformed the image imagination paradigm in both decoding performance and subjective scoring. Therefore, the left and right hand clenched fist paradigm was identified as the user's personalized paradigm.

[0157] The system then guides the user to perform a calibration experiment under the defined paradigm and collects corresponding calibration EEG data. Based on the paradigm-model mapping relationship, the system identifies two candidate decoding models: the EEGNet model and the CNN-LSTM-EEG model. The system uses the calibration data to evaluate the performance of the two models and selects the one with better performance as the initial decoding model based on a comprehensive scoring function. In one implementation, the CNN-LSTM-EEG model is selected as the initial decoding model.

[0158] After determining the initial decoding model, the system enters the offline personalized training phase. The reinforcement learning agent uses calibration data features as state input and optimizes the policy by adjusting model feature weights and related parameters. Guided by the reward function, it gradually converges to obtain personalized configuration parameters adapted to the current user, and trains a personalized initial decoding model accordingly. In one implementation scenario, the decoding performance of the personalized optimized model is significantly improved compared to the initial model.

[0159] The user then enters the formal control phase. During this process, the system interface provides real-time feedback, including decoding category prompts and control status indicators, to help the user adjust their brain control strategy. The user generates different types of EEG patterns by performing left and right hand motor imagery tasks. The system maps the decoding results to corresponding control commands, such as left hand imagery corresponding to a left turn command, right hand imagery corresponding to a right turn command, and relaxed hands corresponding to a stop command. The system sends these control commands to the mobile device's control module via wireless communication, enabling real-time control of the mobile device's movement. During system operation, decoding performance is continuously monitored, and incremental learning updates are automatically triggered when a performance decline is detected, thus maintaining stable system operation.

[0160] In a continuous usage scenario, the system can maintain a high control success rate over a relatively long period of time, demonstrating good individual adaptability and stability.

[0161] III. The role of personalization mechanisms in application.

[0162] Because the system allows users to choose experimental paradigms based on their own understanding of the difficulty and subjective comfort, and obtain decoding models adapted to the current user through model matching and personalized training, even in cases where there are significant differences in users' EEG characteristics or the risk of being blind to brain-computer interfaces in the traditional sense, the system can still improve the success rate of effective control through adjustments at the paradigm and model levels.

[0163] In actual use, the system continuously monitors the control effect through online decoding and adaptive modules, and incrementally optimizes the model based on the EEG data generated by the user in real time, thereby maintaining the long-term stable operation of the brain-controlled mobile device.

[0164] Through the aforementioned personalized paradigm selection and model adaptive optimization mechanism, this invention can maintain stable decoding performance even when there are significant differences in EEG among users, reducing the risk of system failure due to individual differences, thereby improving the reliability and usability of brain-controlled mobile devices in actual use environments.

[0165] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0166] 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 personalized brain-computer interface system, characterized in that, include: The EEG data engineering module is used to collect the user's EEG signals and preprocess them. The Personalized Paradigm Library module is used to store pre-validated brain-computer interface experimental paradigms. The paradigms use relational data storage for paradigm metadata. Each paradigm includes a paradigm ID, paradigm type, paradigm description, pre-validation metric, and recommendation index. The personalized paradigm selection module is used to display brain-computer interface experimental paradigms and select brain-computer interface experimental paradigms based on quantitative indicators and user feedback; the quantitative indicators include decoding performance indicators and user subjective experience indicators. The system performs a weighted comprehensive score on the quantitative indicators, as shown in the following formula: ， ， in, For comprehensive evaluation and scoring, For decoding performance metrics, This refers to subjective user experience metrics. As the weight of decoding performance metrics, Weighting of user subjective experience metrics; The decoding performance metrics are as follows: ， in, Let be the classification accuracy of the i-th classification accuracy in the classification accuracy sequence. Let i be the i-th information transmission rate in the information transmission rate sequence. The maximum value in the classification accuracy sequence. This represents the maximum value in the information transmission rate sequence. The weights for classification accuracy, Weights for information transmission rate; The user subjective experience indicators are specifically as follows: ， in, Score the subjective difficulty level. Rate the degree of fatigue. For focus self-assessment, For subjective difficulty rating weights, As the weight of the fatigue level score, Weighting for self-assessment of focus; The model matching and selection module is used to determine at least one candidate decoding model suitable for the target paradigm according to the preset paradigm-model correspondence, and to determine an initial decoding model from the candidate decoding models based on the user's calibrated EEG data; the model matching and selection module maps the corresponding decoding model from the selected brain-computer interface experimental paradigm, each type of brain-computer interface experimental paradigm is mapped to a set of decoding models, and each set of decoding models has at least one candidate decoding model; The comprehensive score based on decoding performance metrics and model complexity is calculated as follows: ， in, Let be the classification accuracy of the i-th classification accuracy in the classification accuracy sequence. Let i be the i-th information transmission rate in the information transmission rate sequence. For model complexity, The maximum value in the classification accuracy sequence. This represents the maximum value in the information transmission rate sequence. It is the maximum value in the inverse sequence of model complexity. The weights for classification accuracy, As a weight for information transmission rate, These are the weights for model complexity; The optimal decoding model, i.e. the initial decoding model, is obtained based on the highest comprehensive score based on decoding performance metrics and model complexity. The offline personalization module is used to set the initial personalization configuration parameters and train the initial decoding model using reinforcement learning to obtain the personalized initial decoding model. The online decoding and adaptive module is used to decode the user's real-time EEG signals using a personalized initial decoding model, and to continuously optimize the personalized initial decoding model using an incremental learning approach based on the new EEG data and decoding results generated during the user's online use.

2. The personalized brain-computer interface system according to claim 1, characterized in that, The EEG data engineering module acquires multi-channel raw EEG signal sequences at a set time and performs preprocessing on the multi-channel raw EEG signal sequences, including bandpass filtering, notch filtering, artifact removal, baseline correction, rereference, and downsampling.

3. The personalized brain-computer interface system according to claim 1, characterized in that, The paradigm ID is a unique identifier; The paradigm types include motion imagination paradigm, visual motion imagination paradigm, spatial navigation imagination paradigm, SSVEP paradigm, P300 paradigm, voice imagination paradigm, music imagination paradigm, and image imagination paradigm. The paradigm description includes textual and graphical explanations that illustrate how the paradigm works and the mental tasks that users need to perform. The pre-verification metrics include decoding accuracy and average information transmission rate; The recommendation index is a score calculated based on historical user selection feedback.

4. The personalized brain-computer interface system according to claim 3, characterized in that, The personalized paradigm library module collects subjective comfort scores and calculates paradigm scores based on decoding accuracy and average information transmission rate, and stores the corresponding pre-validated brain-computer interface experimental paradigms according to the paradigm scores.

5. The personalized brain-computer interface system according to claim 1, characterized in that, The initial personalized configuration parameters include at least one of the following: A feature importance vector used to weight input EEG features; The set of hyperparameters of the initial decoding model; The initial decoding model includes some network structure adjustment parameters.

6. The personalized brain-computer interface system according to claim 1, characterized in that, The reinforcement learning method employs an actor-critic architecture, in which... The state is a concatenation of historical average features and current features; The action is updated according to the type of personalized goal. Specifically, when the personalized goal is feature selection, the feature selection method is adopted and the action is a feature weight vector, where each element represents the retention weight of the corresponding feature. When the personalized goal is parameter optimization, the hyperparameter optimization method is adopted and the action is the hyperparameter adjustment amount. Based on the updated actions, obtain the optimal network parameters, and then configure a personalized initial decoding model based on the optimal network parameters.

7. The personalized brain-computer interface system according to claim 1, characterized in that, The online decoding and adaptive module includes: A data buffer is used to cyclically store newly generated EEG data and decoding results from the user; The performance detection unit is used to detect the decoding performance of the personalized initial decoding model in real time. The incremental update unit is used to analyze the decoding performance. When the decoding performance is lower than a preset threshold or reaches a preset update cycle, the stored data is used to fine-tune the initial decoding model, wherein the learning rate during fine-tuning is lower than the training learning rate of the offline personalization module.

8. The personalized brain-computer interface system according to claim 7, characterized in that, The fine-tuning performed by the incremental update unit is limited to updating the classification layer parameters of the personalized initial decoding model, or updating the network layer parameters that are associated with the high weights of the personalized configuration parameters.

Images