Information processing method for motor imagery electroencephalogram classification under multi-session conditions

Abstract

The application discloses an information processing method for motor imagery electroencephalogram classification under multiple conversation conditions, which comprises the following steps: firstly, uniformly pre-processing and sliding window modeling of multiple conversation motor imagery electroencephalogram signals; then, extracting time domain and space features and frequency domain and structure features of the electroencephalogram samples through double-branch feature coding, and fusing to obtain a uniform embedded representation; and then, constructing a category prototype based on the current conversation sample, and performing multi-dimensional similarity joint evaluation, comprehensive migration score calculation and screening on the historical samples to realize effective utilization of the historical samples. By introducing the current conversation category prototype as a central reference, the stability and compactness of the current conversation category representation are enhanced, thereby improving the accuracy of the historical sample screening, reducing the negative migration risk, and enhancing the continuity, robustness and engineering application value of the multiple conversation motor imagery electroencephalogram classification process.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, specifically to an information processing method and apparatus for classifying motor imagery EEG under multi-conversation conditions. Background Technology

[0002] Brain-computer interface (BCI) technology is a type of intelligent control technology that enables information exchange between the human brain and external devices by collecting and analyzing electroencephalogram (EEG), magnetoencephalogram (MEG), or cortical electrical activity signals. Among these, EEG-based BCIs are widely used in rehabilitation training, intelligent prosthetic limb control, wheelchair navigation, human-computer interaction, neurofunctional assistance, and external device control due to their non-invasiveness, portability, relatively low cost, and good real-time performance. Among the various implementation paradigms of BCIs, motor imagery BCIs are an important technological approach. Motor imagery refers to the process by which a subject, without actually performing limb movements, actively imagines a specific limb movement, thereby inducing changes in related neural activity in the sensorimotor cortex. The system collects and analyzes these EEG signals to identify and classify the subject's motor intentions.

[0003] In existing technologies, EEG classification for motor imagery typically involves steps such as EEG acquisition, preprocessing, feature extraction, and pattern recognition. First, multi-lead EEG acquisition devices record the EEG signals generated by subjects under different task cues. Second, the raw EEG signals undergo bandpass filtering, notch filtering, artifact removal, segmentation, and standardization to reduce interference from electrooculography (EOG), electromyography (EMG), and environmental noise. Then, feature extraction methods are used to obtain discriminative time-domain, frequency-domain, time-frequency-domain, or spatial-domain features. Finally, a classification model is used to identify different motor imagery tasks. Commonly used feature extraction techniques in traditional methods include co-spatial pattern recognition, filter bank co-spatial pattern recognition, power spectral density analysis, wavelet transform, and combinations thereof. Classification models often employ machine learning methods such as linear discriminant analysis, support vector machines, K-nearest neighbors, and random forests. The above methods can achieve certain classification results under the conditions of single session, controlled environment and relatively stable sample distribution, but they are highly dependent on manual feature engineering and have limited adaptability to individual differences and session changes, making it difficult to meet the requirements of stability and robustness in actual continuous use scenarios.

[0004] In recent years, with the development of deep learning technology, EEG classification methods based on convolutional neural networks, recurrent neural networks, and attention mechanisms have gradually become a research focus. Compared with traditional methods, deep learning methods can automatically learn discriminative deep features directly from raw or weakly preprocessed EEG signals, reducing the need for manual feature design and demonstrating better recognition performance on multiple public datasets. Especially in motor imagery tasks, convolutional neural networks can effectively extract local spatiotemporal patterns of EEG signals and enhance the ability to express complex neural activity patterns through multi-layer nonlinear mapping, thus gaining widespread application in current technologies.

[0005] However, motor imagery EEG signals exhibit significant non-stationarity and intra-individual dynamic drift characteristics. Even for the same subject, EEG signals acquired on different dates, at different times, under different mental states, or with different electrode contact conditions can still show significant changes in amplitude distribution, frequency band energy, spatial patterns, and statistical structure. These changes typically manifest as inter-session differences; that is, a classification model trained in one acquisition session often shows decreased recognition performance in new acquisition sessions. The main reasons for this problem include: signal amplitude shifts due to changes in electrode impedance, spatial distribution changes caused by changes in scalp contact conditions, rhythmic feature drift due to subject fatigue, fluctuations in attention, or changes in task proficiency, and external disturbances caused by changes in the acquisition environment and equipment status. These factors work together to create a complex distribution shift between current and historical session data, thus limiting the model's generalization ability in multi-session scenarios.

[0006] To address the performance degradation in classification under multi-session conditions, existing technologies have proposed various solutions, including recalibration, session transfer, domain adaptation, and instance transfer. Recalibration methods typically rely on re-collecting a certain number of labeled samples in each new session to retrain or correct the classification model. However, this increases the system's overhead and reduces the convenience of brain-computer interfaces. Domain adaptation methods primarily reduce inter-session bias by aligning the feature distributions of the source and target domains. However, these methods often rely on complex distribution matching or adversarial learning mechanisms, which can easily lead to under-alignment or over-alignment when the sample size is small or the distribution bias is strong. Instance transfer methods attempt to select samples from historical sessions that are more relevant to the current session and introduce them into the classification process, thereby reducing recalibration costs and improving current session performance. Compared to directly using all historical data, instance transfer can mitigate the risk of negative transfer to some extent, thus showing good application potential.

[0007] Existing technologies employ convolutional neural networks to extract EEG features from multi-conversation motor imagery, and evaluate the correlation between historical and current conversation samples based on sample similarity. Then, highly correlated historical samples are selected for current conversation classification based on a preset threshold. This approach demonstrates that more historical data is not necessarily better; correlation-based filtering is necessary to effectively improve cross-conversation classification performance. This technical approach addresses, to some extent, the problem of negative transfer caused by directly introducing all historical samples, providing a valuable approach for multi-conversation motor imagery EEG classification.

[0008] However, existing technologies still have the following shortcomings. First, most schemes only use a single similarity index to judge the relationship between historical and current samples, which is difficult to comprehensively reflect the differences in EEG signals in terms of feature orientation, statistical structure, category center, and predictive stability. Second, the selection of relevant samples usually relies on fixed empirical thresholds, while the optimal selection criteria are often inconsistent for different subjects, different task categories, and different conversation conditions, resulting in insufficient adaptability of the methods. Third, existing schemes usually adopt a binary sample utilization method of "retaining" or "removing," which cannot reflect the actual contribution of different historical samples to the current classification task in a fine-grained manner, and is not conducive to the reasonable utilization of moderately relevant samples. In addition, existing technologies generally lack a mechanism for continuous maintenance of high-value historical information, and cannot dynamically update representative samples, category centers, and conversation drift information as subsequent conversations progress, thus making it difficult to adapt to long-term continuous use scenarios.

[0009] Based on this, existing multi-conversation motor imagery EEG classification technologies still suffer from problems such as insufficient accuracy in recognizing cross-conversation related samples, relatively crude utilization of historical samples, limited negative transfer suppression capabilities, and weak stability in continuous application. Therefore, it is necessary to propose a new information processing method for motor imagery EEG classification under multi-conversation conditions to improve the system's classification accuracy, stability, and continuous application capability in multi-conversation scenarios, providing more reliable technical support for the practical application of motor imagery brain-computer interfaces. Summary of the Invention

[0010] The main objective of this invention is to provide an information processing method for EEG classification of motor imagery under multi-conversation conditions, so as to overcome the shortcomings of related technologies.

[0011] To achieve the above objectives, according to a first aspect of the present invention, an information processing method for EEG classification of motor imagery under multi-session conditions is provided. The method includes: collecting raw EEG signals generated by the same subject performing different categories of motor imagery tasks under multiple independent session conditions; preprocessing the raw EEG signals of each session to obtain standardized EEG signals; constructing EEG window samples by using a sliding time window method on the standardized EEG signals to form a current session sample set and a historical session sample set; performing a unified embedding representation on the EEG window samples in the current session sample set and the historical session sample set based on a two-branch feature encoding method; determining the category prototype vector of each motor imagery category in the current session based on the embedding vector of the current session sample set; performing a multi-dimensional similarity joint evaluation on each historical sample in the historical session sample set, and determining a comprehensive transfer score based on the obtained joint evaluation results for use by subsequent classification models.

[0012] According to a second aspect of the present invention, an information processing apparatus for EEG classification of motor imagery under multiple conversation conditions is provided, comprising a sample determination unit for collecting raw EEG signals generated by the same subject when performing different categories of motor imagery tasks under multiple independent conversation conditions, performing preprocessing on the raw EEG signals of each conversation to obtain standardized EEG signals; constructing EEG window samples by using a sliding time window method on the obtained standardized EEG signals to form a current conversation sample set and a historical conversation sample set; and an embedding vector determination unit for uniformly embedding the EEG window samples in the current conversation sample set and the historical conversation sample set based on a two-branch feature encoding method to obtain the spatiotemporal feature vector and the structural feature vector of the current conversation sample set, respectively; and the historical... The system comprises: a spatiotemporal feature vector and a structural feature vector of a session sample set; fusion of the spatiotemporal feature vector and the structural feature vector to obtain the embedding vector of the current session sample set and the embedding vector of the historical session sample set, respectively; a sample filtering unit, used to determine the category prototype vector of each motion imagination category in the current session based on the embedding vector of the current session sample set; for each historical sample in the historical session sample set, performing a multidimensional similarity joint evaluation with the corresponding category prototype vector of the current session to obtain directional similarity, distribution structure similarity, drift compensation similarity, and prediction confidence correlation index information; determining the comprehensive transfer score of each historical sample based on the obtained joint evaluation results; and filtering historical samples based on the comprehensive transfer score for subsequent model classification.

[0013] The information processing method for EEG classification of motor imagery under multi-conversation conditions provided in this embodiment first performs unified preprocessing and sliding window sampling on multi-conversation motor imagery EEG signals; then, it extracts the temporal and spatial features, as well as the frequency and structural features of the EEG samples through dual-branch feature encoding, and fuses them to obtain a unified embedding representation; next, it constructs a category prototype based on the current conversation samples, and performs multi-dimensional similarity joint evaluation, comprehensive transfer score calculation, and screening on historical samples to achieve effective utilization of historical samples. By introducing the current conversation category prototype as a central reference, the stability and compactness of the current conversation category representation are enhanced, thereby improving the accuracy of historical sample screening, reducing the risk of negative transfer, and enhancing the continuity, robustness, and engineering application value of the multi-conversation motor imagery EEG classification process. Attached Figure Description

[0014] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0015] Figure 1 This is a flowchart of an information processing method for EEG classification of motor imagery under multi-conversation conditions according to an embodiment of the present invention; Figures 2 to 4 This is a schematic diagram illustrating the application of the information processing method for EEG classification of motor imagery under multi-conversation conditions according to an embodiment of the present invention; Figure 5 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] According to embodiments of the present invention, an information processing method for EEG classification of motor imagery under multi-conversation conditions is provided, such as... Figure 1 As shown, steps 101 to 103 are included below: Step 101: Collect raw EEG signals generated by the same subject when completing different types of motor imagery tasks under multiple independent conversation conditions. Perform preprocessing on the raw EEG signals of each conversation to obtain standardized EEG signals. Use a sliding time window method to construct EEG window samples from the obtained standardized EEG signals to form the current conversation sample set and the historical conversation sample set.

[0020] In this step, EEG signals are collected from the same subject as they complete different motor imagery tasks under multiple independent session conditions. These multiple sessions may correspond to data collection processes at different dates, time periods, experimental batches, or training phases. The motor imagery task categories include, but are not limited to, left-hand motor imagery, right-hand motor imagery, bipedal motor imagery, gait motor imagery, resting tasks, or other pre-defined limb movement intention categories. For the first... The raw EEG signals obtained from each session are denoted as ,in, The number of EEG acquisition channels is represented by T, and the number of sampling points is represented by T. Since raw EEG signals are susceptible to electrooculography (EOG), electromyography (EMG), power frequency noise, unstable electrode contact, transient artifacts, and changes in the external environment, directly using them for multi-conversation classification modeling would increase inter-conversation non-stationarity and reduce the accuracy of subsequent transfer judgments. Therefore, this invention first performs uniform preprocessing on the raw EEG signals to improve the comparability and standardization of different conversation samples at the input level.

[0021] The preprocessing workflow includes bandpass filtering, notch filtering, anomaly channel repair, artifact suppression, normalization, and task segment truncation. Bandpass filtering is used to preserve the main frequency components related to motion imagery, with a preferred frequency range of [missing information]. or To preserve Rhythm, While suppressing excessively high-frequency noise, the rhythm and related low-frequency information are also preserved; notch filtering is used to suppress power frequency interference, and the notch center frequency can be set according to the actual power supply environment. or For channels exhibiting drift, poor contact, or local anomalies, repair methods such as neighbor channel interpolation, spatial correlation reconstruction, or reference channel mapping can be employed. For eye movement artifacts, electromyography artifacts, and abnormal fluctuations in instantaneous amplitude, suppression methods such as independent component analysis, artifact subspace reconstruction, adaptive threshold removal, or combinations thereof can be used. Through these processes, the consistency of basic signal quality across different sessions is improved while preserving as much motor imagery EEG discrimination information as possible.

[0022] After filtering and artifact suppression, to reduce amplitude scale differences between different sessions, channels, and time periods, the EEG signals were further standardized. For the first... In the first session Each channel at time signal value Its standardization result can be expressed as: in, , Indicates the first In the first session The average value of the signals in each channel. Indicates the corresponding standard deviation. To prevent extremely small positive numbers from being introduced into the denominator (zero), standardization can reduce amplitude shifts caused by different acquisition batches and variations in electrode impedance, thereby improving scale consistency when EEG samples are incorporated into a unified feature coding model.

[0023] After standardization, task segments were extracted from the continuous EEG signals based on the start and end times of the task cues in the experimental paradigm. The second attempt at motion visualization was in the [number]th [number]. The effective time period in each session is Then the corresponding trial segment is extracted. To improve sample utilization and preserve local temporal features, a sliding time window method was further employed for sample construction. Let the sliding window length be... Step size is Then the fragment constructed from the m-th trial fragment is the first... A single EEG window sample can be represented as: in, This represents the r-th EEG window sample corresponding to the m-th trial in the s-th session. Furthermore, the window endpoint should not exceed the end position of the corresponding trial segment. The preferred sliding window length is... The preferred step size is By using sliding window sampling, the number of training samples can be increased without changing the original acquisition paradigm, while preserving the dynamic changes of EEG signals on local time scales.

[0024] After the above processing, a current session sample set and a historical session sample set are formed. Let the current session be denoted as c, then its sample set can be represented as: ,in, Indicates the first in the current session One EEG window sample, This indicates the corresponding motion imagery category label. This represents the total number of samples in the current session. Correspondingly, the sample set of all historical sessions can be represented as... ,in, Indicates the first The number of samples from each historical session. Through the above processing, the original multi-session EEG signals are transformed into a windowed sample set with a uniform format and clear labels, providing a foundation for subsequent cross-session modeling.

[0025] Step 102: Based on the dual-branch feature encoding method, perform unified embedding representation on the EEG window samples in the current session sample set and the historical session sample set to obtain the spatiotemporal feature vector and structural feature vector of the current session sample set; the spatiotemporal feature vector and structural feature vector of the historical session sample set; fuse the spatiotemporal feature vector and the structural feature vector to obtain the embedding vector of the current session sample set and the embedding vector of the historical session sample set.

[0026] In this step, refer to Figure 2 To enhance the representation of frequency band energy relationships and statistical structure changes, frequency domain and structural transformations were performed on the same EEG window samples to obtain input representations. ,in This refers to spectrum analysis, time-frequency transformation, power spectrum estimation, covariance statistics, or combinations thereof. The u may include... Band energy, Features such as band energy, power spectral density, interband power ratio, time-frequency representation, and channel covariance matrix are considered. This characterization is input to the frequency domain-structure branch, and its mapping function is set as follows. Then the structural feature vector is obtained. ,in, , This indicates the output feature dimension of this branch. This branch enhances the characterization of frequency band power redistribution, covariance structure variations, and cross-session statistical shifts.

[0027] In obtaining temporal-spatial features Frequency Domain-Structural Features Then, the two types of features are fused to form a unified embedded representation. Preferably, a gated fusion method is used to obtain the final embedded representation. Represented as In explicit gating implementations, it can be further represented as ,in, This represents the Sigmoid activation function. Represents the gating weight vector. and For learnable parameters, This represents element-wise multiplication. and This represents the two branches of features after dimensionality mapping. This fusion method allows for adaptive adjustment of the contribution ratio of the two types of features based on different sample features.

[0028] Furthermore, to ensure that current session samples and historical session samples can be directly compared within the same embedding space, this invention uses the same set of dual-branch feature encoding parameters for both. For both current and historical session samples, a unified feature mapping function is used. Obtain the embedded representation By sharing parameters, it can be ensured that different session samples are projected into the same feature space under a unified representation rule, which provides a foundation for subsequent category prototype construction, multidimensional similarity evaluation and historical sample transfer scoring.

[0029] Step 103: Based on the embedding vector of the current session sample set, determine the category prototype vector of each motion imagination category in the current session; for each historical sample in the historical session sample set, perform multi-dimensional similarity joint evaluation with the corresponding category prototype vector of the current session to obtain directional similarity, distribution structure similarity, drift compensation similarity and prediction confidence related index information; determine the comprehensive transfer score of each historical sample based on the joint evaluation results, and select historical samples based on the comprehensive transfer score for use by subsequent classification models.

[0030] In this step, after obtaining the current session sample embedding representation, a category prototype is further constructed based on the labeled samples of the current session to form a centralized category representation for the current session classification task. Let the set of current session embedding samples be: For the first All motion-based imagination tasks will be satisfied. The sample embeddings are summarized to form a subset of samples for that category. If the number of samples for that category in the current session is... Then the first The class prototype vector is defined as ,in, This represents the class prototype vector of the k-th class in the current session. Indicates the first in the current session The class of Each sample embedding vector. By constructing a category prototype, the overall distribution trend of similar samples in the current session can be compressed into a central vector, thereby reducing computational redundancy in subsequent relevance assessment and minimizing the impact of local noise.

[0031] In a preferred embodiment, to enhance the robustness of the category prototype to low-quality and anomalous samples, a weighted prototype construction method can also be used. Let the current session be... Class 1 The quality weight or confidence weight of each sample is: Then the prototype of this category can be represented as ,in, The similarity can be determined based on sample signal quality, current classification confidence, consistency before and after enhancement, or other stability indicators. Weighted prototype construction improves the stability and representativeness of category center representation and reduces interference from low-quality and anomalous samples. After completing the current conversation category prototype construction, a multi-dimensional similarity joint evaluation is performed based on the relationship between historical conversation samples and the current conversation category prototype to determine the transferability of historical samples relative to the current classification task. Since differences in multi-conversation motor imagery EEG signals manifest not only as changes in feature direction but also as changes in statistical structure, category center drift, and differences in sample prediction stability, using only a single cosine similarity is insufficient to fully reflect the transfer value of historical samples. Therefore, this invention calculates directional similarity, distribution structure similarity, drift compensation similarity, and prediction confidence correlation indicators for historical samples separately.

[0032] As an optional implementation of this embodiment, the directional similarity is used to measure the directional proximity between historical session samples and category prototype vectors in the embedding space; the distribution structure similarity is used to measure the consistency of the statistical structure between historical samples and the category prototype vectors corresponding to the current session; the drift compensation similarity is used to compensate for the drift of historical samples based on the drift of category centers between sessions, so as to reflect the actual matching degree between them and the category prototypes corresponding to the current session; the prediction confidence correlation index information is used to reflect the reliability of historical samples. Specifically, when the prediction results of historical samples are more concentrated and the preset classification model is more clear about their category judgment, the larger the index information, the higher the reliability of the historical sample; when the prediction distribution of historical samples is more dispersed and the preset classification model is difficult to make a clear classification of them, the smaller the index information, the higher the uncertainty of the historical sample.

[0033] In this optional implementation, refer to Figure 3 The diagram illustrates a multi-dimensional similarity assessment, assuming the historical sample embedding set is... ,in, Represents the historical sample embedding vector. This indicates the corresponding motion imagery category label. This represents the total number of historical samples. For any given historical sample... If it belongs to the first If a class is selected, its relationship with the current session's first iteration is calculated. Class prototype directional similarity This metric measures how close the historical samples are to the current class prototype in terms of orientation in the embedding space.

[0034] Furthermore, a distribution structure similarity is introduced to measure the consistency of statistical structure between historical and current sessions. Let the historical conversation be numbered. Class and current session The structure matrix of the class is and Then define the dimorphism distance. ,in, The structural distance metric function can be represented by Riemann distance, Frobenius distance, or other matrix distances. Furthermore, structural similarity is defined as... This formula converts structural distance into a similarity representation through an exponential mapping. The closer two sessions are in statistical structure, the higher their similarity. The smaller, The larger and closer to 1, the more significant the statistical structural differences between the two sessions. The larger, The smaller the value, the more intuitively it reflects the degree of matching between historical and current sessions at the statistical structure level, and provides a basis for subsequent assessment of the migration value of historical samples.

[0035] Furthermore, to compensate for category center drift between sessions, a drift compensation similarity is introduced. Let the historical prototype be... The category drift vector of the current session relative to the historical reference is defined as follows: Based on this, the similarity of historical samples after drift compensation is expressed as: This formula first compensates for the drift of historical samples and then calculates their directional similarity with the current category prototype, thus reflecting the actual matching degree of the historical sample after considering the overall session drift. Therefore, it can identify historical samples that appear dissimilar due to overall session drift but still have transfer value after compensation.

[0036] In addition, to reflect the reliability of the historical samples themselves, a prediction confidence correlation index is introduced. By inputting historical samples into the current classification model, the posterior probability vector of the class is obtained. ,in, This represents the posterior probability distribution of historical samples under the current classification model; The model indicates that the historical sample is classified as the first... The probability of a class This represents the total number of categories. To measure the central tendency of this probability distribution, an entropy function is used to measure its predictive uncertainty. ,in, The information entropy represents the prediction results of historical samples. A higher information entropy indicates a more dispersed prediction probability across multiple categories, and a less certain level of uncertainty in the model's classification. Conversely, a lower information entropy indicates a more concentrated prediction probability within a single category, and a clearer classification. In other words, the entropy value reflects the predictive stability and discriminative clarity of historical samples under the current classification model.

[0037] Furthermore, the prediction confidence correlation index is defined as follows: ,in, Used to normalize entropy values, The value range is more stable, making it easier to integrate with other similarity indicators. Therefore, the more concentrated the prediction results of historical samples and the clearer the model's category judgment, the better. The smaller, The larger the value and the closer it is to 1, the higher the reliability of the sample; when the predicted distribution of historical samples is more dispersed and the model has difficulty making a clear classification of them, The larger, The smaller the value, the higher the uncertainty of the sample.

[0038] After calculating the above four categories of indicators, this invention performs a weighted fusion to obtain a comprehensive migration score for historical samples. Its expression is ,in, These represent the weighting coefficients corresponding to the four indicators, satisfying... The aforementioned constraints stipulate that all weights are non-negative and sum to 1, thus enabling the comprehensive transfer score to characterize the combined contribution of different indicators to the transfer value of historical samples on a unified scale. This weighted fusion method maps the performance of historical samples across multiple levels, including feature orientation, statistical structure, session drift compensation, and sample reliability, into a single quantitative result. Therefore, it not only determines whether a historical sample is highly relevant to the current session but also further measures the overall transfer value of that historical sample, providing a unified basis for subsequent screening and tiered utilization. Thus, highly relevant historical samples can be directly screened based on scores, or further optimized screening can be employed.

[0039] As an optional implementation of this embodiment, filtering historical samples based on the comprehensive migration score includes: determining the filtering threshold for different categories based on the comprehensive migration score and a predefined adaptive filtering threshold mechanism for different categories; and further filtering historical samples based on the threshold.

[0040] In this optional implementation, adaptive filtering thresholds are established for different categories. Let the set of historical samples with comprehensive transfer scores corresponding to the k-th category in the current session be . ,in, Indicates that the historical sample belongs to the first The number of samples in each class. Calculate the mean for each class. with standard deviation , , Therefore, the adaptive filtering threshold for this category is defined as follows: ,in, Adaptive filtering threshold for historical samples; This represents the mean of the overall migration scores of historical samples in this category, used to reflect the average level of the overall migration value of historical samples in this category; The standard deviation of the overall migration score of historical samples in this category is used to reflect the dispersion of the migration value distribution of historical samples in this category. β is the threshold adjustment coefficient. When β is larger, the resulting threshold is higher, and the sample selection process is more stringent; when β is smaller, the resulting threshold is lower, and the sample selection process is relatively lenient. Using this method, selection criteria for corresponding categories can be adaptively generated based on the overall distribution characteristics of historical sample migration scores for different categories. This allows the historical sample selection process to no longer rely on a single fixed empirical threshold, but to dynamically adjust according to category distribution and the current session state.

[0041] As an optional implementation of this embodiment, filtering historical samples based on the comprehensive migration score includes: determining the filtering thresholds for different categories based on the comprehensive migration score and a predefined adaptive filtering threshold mechanism for different categories; determining the migration weights based on the comprehensive migration score, the filtering thresholds, and a preset weight allocation mechanism; and utilizing the historical samples in a tiered manner based on the migration weights.

[0042] Preferably, after determining the adaptive threshold, a continuous soft weight allocation mechanism can be further constructed so that different historical samples participate in the current model training with different intensities according to the relationship between their comprehensive transfer score and the category threshold, that is, the historical samples are used in a graded manner.

[0043] As an optional approach in this embodiment, for historical samples that meet preset conditions, the migration weights are... The allocation is performed using an exponential mapping and normalization method. ,in, This is the temperature coefficient.

[0044] In this optional implementation, for historical samples that meet the minimum participation condition, their migration weights... The allocation is performed using exponential mapping and normalization, expressed as follows: ,in, This is the temperature coefficient.

[0045] You can also set a minimum participation threshold, for example, only if historical samples meet the requirements. Only then does it enter the soft weight allocation process; otherwise, its weight is reset to zero.

[0046] By employing the aforementioned comprehensive transfer scoring, adaptive threshold, and continuous soft weight allocation mechanism, we can achieve differentiated utilization of high-value historical samples, moderately relevant samples, and low-confidence samples, reduce the risk of negative transfer, and provide more reliable historical information support for subsequent multi-session classification.

[0047] As an optional implementation of this embodiment, the method further includes updating specified information of historical samples based on the new session.

[0048] In this optional implementation, refer to Figure 4The core of the dynamic memory transfer mechanism is "building a memory bank + dynamic maintenance," which runs continuously as new sessions progress. It identifies high-value historical samples for the current session (screened by comprehensive transfer score), the current session category prototype, and inter-session drift information. The screened high-value historical samples (including their embedding vectors, comprehensive transfer scores, and soft weights), the category prototypes corresponding to each session, and the inter-session distribution drift vectors are stored in the memory bank.

[0049] The dynamic maintenance process selectively writes high-value historical samples selected from the current session, the category prototypes constructed in the current session, and the drift information between the current and historical sessions into the memory (retaining only high-value information to avoid redundancy). When a new session starts and completes its own category prototype construction, it calls upon the historical information in the memory, combines it with the feature distribution of the new session, re-evaluates the transfer value of historical samples in the memory, updates their comprehensive transfer scores and soft weights, and iteratively updates the category prototypes and session drift information (i.e., specified information) to adapt to the distribution characteristics of the new session. The memory content is periodically filtered to remove outdated samples with low transfer value, samples with significant differences from the latest session features, and invalid drift information, ensuring that the memory always contains information with continuous transfer value. Through this dynamic write-update-clear cycle, the memory adapts in real-time to the distribution drift of EEG signals from multiple sessions, providing accurate historical information support for model training in each subsequent new session, reducing the risk of negative transfer, and achieving long-term continuous classification adaptation.

[0050] This embodiment focuses on the EEG classification task of multi-conversation motor imagery, forming a complete technical implementation process consisting of EEG acquisition and unified preprocessing, task segment extraction and sliding window sampling, dual-branch feature encoding and fusion representation construction, current conversation category prototype generation, joint evaluation of multi-dimensional similarity of historical samples, comprehensive transfer score calculation, adaptive threshold screening, and continuous soft weight allocation. Through the above technical solution, accurate identification and hierarchical utilization of useful information from historical conversations are achieved, enhancing the ability to characterize inter-conversation distribution drift, and improving the targeting and effectiveness of historical sample screening and transfer utilization. This effectively improves the accuracy, stability, and continuous application capability of EEG classification for motor imagery in multi-conversation scenarios.

[0051] While retaining valuable information from historical conversations, this invention introduces a more adaptive similarity modeling and transfer utilization mechanism to improve the ability to characterize inter-conversational distribution drift and enhance the refinement of historical sample selection and utilization. This improves the system's classification accuracy, stability, and continuous application capability in multi-conversation scenarios, providing more reliable technical support for the practical application of motor imagery brain-computer interfaces. This embodiment no longer simply reduces the usability of historical conversation samples for the current classification task to a single binary judgment of similarity or dissimilarity. Instead, it jointly models the transfer value of historical samples from multiple levels, including feature direction relationships, statistical structure relationships, conversation drift compensation relationships, and sample prediction stability. Based on this, it constructs a comprehensive transfer scoring, adaptive threshold, and continuous soft weight allocation mechanism, enabling historical samples with different levels of relevance to participate in the current conversation classification in a differentiated manner. Simultaneously, this invention introduces the current conversation category prototype as a central reference to enhance the stability and compactness of the current conversation category representation. Furthermore, it continuously maintains high-value historical information through dynamic memory transfer, thereby improving the accuracy of historical sample selection, reducing the risk of negative transfer, and enhancing the continuity, robustness, and engineering application value of the multi-conversation motor imagery EEG classification process.

[0052] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0053] According to an embodiment of the present invention, an information processing device for EEG classification of motor imagery under multiple conversation conditions is also provided, including a sample determination unit for collecting raw EEG signals generated by the same subject when completing different categories of motor imagery tasks under multiple independent conversation conditions, performing preprocessing on the raw EEG signals of each conversation to obtain standardized EEG signals; constructing EEG window samples by using a sliding time window method on the obtained standardized EEG signals to form a current conversation sample set and a historical conversation sample set; and an embedding vector determination unit for uniformly embedding the EEG window samples in the current conversation sample set and the historical conversation sample set based on a two-branch feature encoding method to obtain the spatiotemporal feature vector and the structural feature vector of the current conversation sample set, respectively; and the historical... The system comprises: a spatiotemporal feature vector and a structural feature vector of a session sample set; fusion of the spatiotemporal feature vector and the structural feature vector to obtain the embedding vector of the current session sample set and the embedding vector of the historical session sample set, respectively; a sample filtering unit, used to determine the category prototype vector of each motion imagination category in the current session based on the embedding vector of the current session sample set; for each historical sample in the historical session sample set, performing a multidimensional similarity joint evaluation with the corresponding category prototype vector of the current session to obtain directional similarity, distribution structure similarity, drift compensation similarity, and prediction confidence correlation index information; determining the comprehensive transfer score of each historical sample based on the obtained joint evaluation results; and filtering historical samples based on the comprehensive transfer score to train a preset classification model.

[0054] According to embodiments of the present invention, the present invention also provides an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the methods described in any of the above embodiments.

[0055] According to embodiments of the present invention, the present invention also provides a readable storage medium storing computer instructions that enable a computer to perform the methods described in any of the above embodiments when executed.

[0056] According to embodiments of the present invention, the present invention also provides a computer program product that, when executed by a processor, can implement the methods described in any of the above embodiments.

[0057] Figure 5A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices.

[0058] like Figure 5 As shown, the electronic device 300 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of the electronic device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0059] Multiple components in electronic device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of displays, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows electronic device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0060] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as the object matching method. For example, in some embodiments, the object matching method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of the methods described above may be performed.

[0061] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0062] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0063] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

Claims

1. An information processing method for EEG classification of motor imagery under multi-conversation conditions, characterized in that, include: Raw EEG signals were collected from the same subject when they completed different types of motor imagery tasks under multiple independent conversation conditions. The raw EEG signals from each conversation were preprocessed to obtain standardized EEG signals. The obtained standardized EEG signals are sampled using a sliding time window method to construct EEG window samples, forming a current conversation sample set and a historical conversation sample set. Based on the dual-branch feature encoding method, the EEG window samples in the current session sample set and the historical session sample set are uniformly embedded to obtain the spatiotemporal feature vector and structural feature vector of the current session sample set; the spatiotemporal feature vector and structural feature vector of the historical session sample set; the spatiotemporal feature vector and structural feature vector are fused to obtain the embedding vector of the current session sample set and the embedding vector of the historical session sample set, respectively. Based on the embedding vectors of the current session sample set, the category prototype vectors of each motion imagination category in the current session are determined; for each historical sample in the historical session sample set, a multi-dimensional similarity joint evaluation is performed between it and the corresponding category prototype vector of the current session to obtain information on directional similarity, distribution structure similarity, drift compensation similarity, and prediction confidence correlation indicators; based on the results of the joint evaluation, a comprehensive transfer score for each historical sample is determined, and historical samples are selected based on the comprehensive transfer score for use by a preset classification model.

2. The information processing method for EEG classification of motor imagery under multi-conversation conditions according to claim 1, characterized in that, Historical samples selected based on the comprehensive migration score include: Based on the comprehensive migration score and a predefined adaptive screening threshold mechanism for different categories, screening thresholds for different categories are determined; historical samples are further screened based on the thresholds.

3. The information processing method for EEG classification of motor imagery under multi-conversation conditions according to claim 1, characterized in that, Historical samples selected based on the comprehensive migration score include: The screening thresholds for different categories are determined based on the comprehensive migration score and a predefined adaptive screening threshold mechanism for different categories. The migration weight is determined based on the comprehensive migration score, screening threshold, and preset weight allocation mechanism. Based on the migration weights, historical samples are used in a tiered manner.

4. The information processing method for EEG classification of motor imagery under multi-conversation conditions according to claim 1, characterized in that, Based on the embedding vectors of the current session sample set, the category prototype vectors for each motion imagination category in the current session are determined as follows: Let the current session embedding sample set be... Among them, for the first All motion-based imagination tasks will be satisfied. The sample embeddings are summarized to form a subset of samples for that category; if the number of samples for that category in the current session is... Then the first The class prototype vector is defined as ,in, This represents the class prototype vector of the k-th class in the current session. Indicates the first in the current session The class of Each sample embedding vector.

5. The information processing method for EEG classification of motor imagery under multi-conversation conditions according to claim 1, characterized in that, The directional similarity is used to measure the directional proximity between historical session samples and category prototype vectors in the embedding space; The distribution structure similarity is used to measure the consistency in statistical structure between historical samples and the corresponding category prototype vectors of the current session; The drift compensation similarity is used to compensate historical samples for drift based on the drift of class center between sessions, so as to reflect the actual degree of matching between them and the class prototype corresponding to the current session. The prediction confidence correlation index information is used to reflect the reliability of historical samples. When the prediction results of historical samples are more concentrated and the preset classification model is clearer in its classification judgment, the larger the index information indicates that the historical sample has higher reliability. When the prediction distribution of historical samples is more dispersed and the preset classification model has difficulty in making a clear classification, the smaller the index information indicates that the historical sample has higher uncertainty.

6. The information processing method for EEG classification of motor imagery under multi-session conditions according to claim 2 or 3, characterized in that, Based on the results of the joint evaluation, a comprehensive migration score is determined for each historical sample. Based on this comprehensive migration score and a predefined adaptive screening threshold mechanism for different categories, screening thresholds for different categories are determined, including: Comprehensive migration score ,in, These represent the weighting coefficients corresponding to the four indicators, satisfying... ; Let the set of historical samples with comprehensive transfer scores corresponding to the k-th class in the current session be: in, Indicates that the historical sample belongs to the first The sample size of each class is calculated, and their mean is calculated respectively. with standard deviation : Therefore, the adaptive filtering threshold for this category is defined as follows: in, Adaptive filtering threshold for historical samples; This represents the mean of the overall migration score for historical samples in this category; This represents the standard deviation of the overall transfer score of historical samples in this category; This is the threshold adjustment coefficient.

7. The information processing method for EEG classification of motor imagery under multi-conversation conditions according to claim 3, characterized in that, Based on comprehensive migration scores, screening thresholds, and a pre-defined weighting mechanism, migration weights are determined as follows: For historical samples that meet preset conditions, their migration weights The allocation is performed using an exponential mapping and normalization method. ,in, This is the temperature coefficient.

8. The information processing method for EEG classification of motor imagery under multi-conversation conditions according to claim 7, characterized in that, Only when historical samples meet When the weight is set to zero, a soft weight allocation mechanism is executed; otherwise, its weight is reset to zero.

9. The information processing method for EEG classification of motor imagery under multi-conversation conditions according to claim 1, characterized in that, The method also includes updating specified information of historical samples based on the new session.

10. An information processing device for EEG classification of motor imagery under multi-conversation conditions, characterized in that, include: The sample determination unit is used to collect the raw EEG signals generated by the same subject when completing different types of motor imagery tasks under multiple independent conversation conditions, and to perform preprocessing on the raw EEG signals of each conversation to obtain standardized EEG signals. The obtained standardized EEG signals are sampled using a sliding time window method to construct EEG window samples, forming a current conversation sample set and a historical conversation sample set. The embedding vector determination unit, based on a dual-branch feature encoding method, performs unified embedding representation on EEG window samples in the current session sample set and the historical session sample set, respectively obtaining the spatiotemporal feature vector and structural feature vector of the current session sample set; the spatiotemporal feature vector and structural feature vector of the historical session sample set; and fuses the spatiotemporal feature vector and structural feature vector to obtain the embedding vector of the current session sample set and the embedding vector of the historical session sample set, respectively. The sample filtering unit is used to determine the category prototype vector of each motion imagination category in the current session based on the embedding vector of the current session sample set; for each historical sample in the historical session sample set, it performs a multi-dimensional similarity joint evaluation with the corresponding category prototype vector of the current session to obtain directional similarity, distribution structure similarity, drift compensation similarity and prediction confidence correlation index information; based on the obtained joint evaluation results, it determines the comprehensive transfer score of each historical sample, and filters out historical samples based on the comprehensive transfer score for subsequent classification processing.

Images