Compression method and device for electroencephalogram characteristics

Abstract

The invention discloses an electroencephalogram characteristic compression method and device. The compression method comprises the steps that electroencephalogram signals of a subject are collected and preprocessed; performing feature extraction on the preprocessed electroencephalogram signals; carrying out secondary preprocessing and enhancement on the extracted features; performing primary screening on the enhanced feature set by using mutual information, and performing secondary screening on the feature set subjected to primary screening by using F statistics to obtain a final target feature set; according to the method, on the premise of ensuring that the judgment information is fully reserved, the data dimension and calculation burden are greatly reduced, the modeling efficiency, generalization ability and engineering deployment feasibility are remarkably improved, and the method is suitable for scenes such as edge computing platforms, wearable emotion sensing systems and high-frequency interactive brain-computer interface equipment.

Description

Technical Field

[0001] This invention relates to the field of signal processing, and in particular to a method and apparatus for compressing electroencephalogram (EEG) features. Background Technology

[0002] The inherent structural complexity of EEG signals presents significant challenges for practical applications. On one hand, these signals typically have dozens of channels and thousands of sampling points per second, with raw data dimensions often exceeding tens of thousands. On the other hand, EEG signals contain substantial physiological and environmental noise, and their stability is affected by individual differences, equipment status, experimental environment, and other factors. Furthermore, the discriminant features corresponding to different emotional states are often distributed across specific frequency bands, specific brain regions, or certain nonlinear features, exhibiting significant information sparsity and uneven distribution. These factors mean that directly modeling based on the entire raw EEG data not only consumes large computational resources and carries a high risk of overfitting, but also severely limits the practical deployment of the system in edge devices or mobile scenarios.

[0003] To alleviate these problems, researchers have proposed a series of preprocessing strategies based on feature extraction and dimensionality compression to improve signal representation efficiency and downstream modeling quality. Typical methods include: ① wavelet transform and energy spectrum decomposition, projecting EEG signals onto a multi-scale frequency domain basis to extract time-varying patterns of alternating high and low frequencies; ② principal component analysis (PCA) and independent component analysis (ICA), constructing low-dimensional representation subspaces while preserving data variance or independent structure; ③ band power aggregation (e.g., θ, α, β band energy calculation), used to capture cortical excitability and neural oscillation characteristics. These techniques, to some extent, improve model training efficiency, mitigate noise interference, and provide structurally cleaned input representations for subsequent neural network modeling.

[0004] While these methods have achieved practical results in feature compression, their performance retention, cross-scenario adaptability, and expression fidelity under higher compression ratios still need improvement. On the one hand, some methods decouple their compression strategies from the EEG task structure, meaning the compression process may not simultaneously preserve the feature subspace closely related to the specific cognitive or emotion decoding target. On the other hand, most existing strategies employ a uniform static dimensionality reduction process, lacking adaptation mechanisms to individual signal differences or dynamic task characteristics, making it difficult to achieve stable modeling under conditions of multi-source input, multiple emotion categories, and multiple participants. Furthermore, some compression processes are difficult to embed into end-to-end deep network frameworks, resulting in weak re-optimization of the compression results and limited gradient propagation capabilities, thus restricting their potential for integration and joint optimization in modern neural network systems. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for compressing EEG signals that integrates multi-domain feature engineering, structural screening, and deep neural networks. This method significantly reduces data dimensionality and computational burden while ensuring sufficient preservation of discriminative information, thereby substantially improving modeling efficiency, generalization ability, and engineering deployment feasibility. It is applicable to scenarios such as edge computing platforms, wearable emotion sensing systems, and high-frequency interactive brain-computer interface devices.

[0006] To achieve the above objectives, the present invention provides a method for compressing EEG features, comprising: acquiring EEG signals from a subject and preprocessing the EEG signals; extracting features from the preprocessed EEG signals, the extracted features including time-domain features, frequency-domain features, time-frequency-domain features, and inter-channel features; performing secondary preprocessing and enhancement on the extracted features; using mutual information to perform a first screening of the enhanced feature set, and then using the F-statistic to perform a second screening of the first-screened feature set to obtain the final target feature set.

[0007] Optionally, preprocessing the EEG signal includes: using a bandpass filter to limit the frequency of the EEG signal to a preset threshold, removing low-frequency drift and high-frequency noise components, and retaining effective components within the emotional frequency band; employing a combined strategy of independent component analysis and artifact detection algorithms to remove physiological interference signals; and introducing standardization processing to convert the signals of each channel into a standard normal distribution with zero mean and unit variance, thereby unifying the feature scale, taking into account the differences in data distribution across different channels.

[0008] Optionally, the time-domain features include one or more of the following: mean, standard deviation, variance, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, median, interquartile range, root mean square, third moment, fourth moment, 5th and 95th percentiles, mean / standard deviation / variance of the first difference, and zero crossover rate.

[0009] Optionally, the frequency domain features are extracted by: estimating the power spectral density of the signal using the Welch method; calculating the relative power and peak frequency of each frequency band around the main frequency band of the neural oscillation; and introducing structural features such as spectral centroid, spectral bandwidth, and spectral entropy to characterize the concentration and distribution complexity of the signal spectrum, reflecting the energy contribution structure of different frequency bands to the emotional state.

[0010] Optionally, the time-frequency domain features are extracted in the following way: Morlet wavelet is selected as the mother function using continuous wavelet transform, and multi-scale decomposition is performed within a preset scale range; mean, standard deviation, maximum value, variance and energy are extracted from the wavelet coefficients as statistics to capture the transient oscillation mode and local time-frequency coupling structure of the signal, effectively reflecting the short-term dynamic response induced by emotions.

[0011] Optionally, the extracted features are preprocessed and enhanced, including: using the Isolation Forest algorithm to identify and remove outlier samples in the feature space; using a robust scaler to standardize the cleaned data; multiplying the standardized data sample vector by a random scaling factor within a preset interval; and adding Gaussian noise with a standard deviation of 2% of the original sample to the scaled samples, as shown in the following formula:

[0012] ;

[0013] in, Represents the original feature vector. This represents the random scaling factor, and N represents Gaussian noise. It represents the standard deviation of the original sample.

[0014] Optionally, the expression for the mutual information is as follows:

[0015] ;

[0016] in, The entropy of the category label. This represents the conditional entropy given feature F.

[0017] Optionally, the expression for the F-statistic is as follows:

[0018] ;

[0019] Where α is the harmonic coefficient, This represents the mutual information between features and categories. This represents the corresponding F-value statistic.

[0020] On the other hand, the present invention also provides a compression device for EEG features, comprising: a first preprocessing unit for acquiring EEG signals of a subject and preprocessing the EEG signals; a feature extraction unit for extracting features from the preprocessed EEG signals, wherein the extracted features include time-domain features, frequency-domain features, time-frequency-domain features, and inter-channel features; a second preprocessing unit for performing secondary preprocessing and enhancement on the extracted features; and a filtering unit for performing a first filtering on the enhanced feature set using mutual information, and then performing a second filtering on the first-filtered feature set using the F statistic to obtain the final target feature set.

[0021] The advantages of this invention over the prior art are:

[0022] 1) This invention achieves efficient dimensionality reduction of the original high-dimensional EEG signal through multi-domain feature engineering and structural compression methods. While retaining key information, it increases the data compression rate to 99.47%, effectively reducing computational overhead and storage burden.

[0023] 2) This invention designs a joint feature selection mechanism based on mutual information and F statistics, and combines intelligent enhancement and robust standardization strategies to significantly improve the model's adaptability to noise interference, class imbalance and individual differences.

[0024] 3) The overall method proposed in this invention has good task discrimination performance and engineering adaptability. It is suitable for high-performance computing environments and can also be extended to resource-constrained real-time emotion recognition scenarios such as mobile terminals and edge devices. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for compressing electroencephalogram (EEG) features provided by the present invention;

[0026] Figure 2 This is a diagram illustrating the effect of the EEG feature compression method provided by this invention on storage reduction analysis;

[0027] Figure 3 This is an analysis diagram of the impact of the EEG feature compression method provided by this invention on storage in a compression pipeline;

[0028] Figure 4 This is an analysis diagram of the EEG feature compression method provided by the present invention under batch processing I / O and memory load;

[0029] Figure 5 This is a comparison of the overall performance improvement results of the EEG feature compression method provided by this invention;

[0030] Figure 6 This is an analysis diagram of the EEG feature compression method provided by the present invention in terms of non-specific feature recognition and cleaning;

[0031] Figure 7 This is a comparative analysis diagram of the robustness of the EEG feature compression method provided by this invention across different scenarios;

[0032] Figure 8 This is a structural diagram of a brainwave feature compression device provided by the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0034] Reference Figure 1 This embodiment provides a method for compressing EEG features, including the following steps:

[0035] S10: Collect the subject's electroencephalogram (EEG) signals and preprocess the EEG signals.

[0036] In this embodiment, multi-channel EEG signals were acquired from the subjects to construct a three-dimensional raw EEG data matrix with 3 channels, 63 time steps, and a spatial resolution of 32×32, fully covering the dynamic neural activity of emotion-related brain regions. To ensure signal quality, the acquired EEG data underwent a systematic preprocessing procedure.

[0037] Specifically, firstly, a bandpass filter is used to limit the signal frequency to 4–45 Hz, eliminating low-frequency drift and high-frequency noise components while retaining effective components within the main emotional frequency bands. Secondly, a combined strategy of independent component analysis (ICA) and artifact detection algorithms is employed to remove physiological interference signals such as eye movement (EOG) and electromyography (EMG), enhancing signal purity. Finally, to address the differences in data distribution across different channels, standardization is introduced to convert the signals from each channel into a standard normal distribution with zero mean and unit variance, unifying the feature scale and providing a highly consistent and comparable input basis for subsequent feature extraction modules.

[0038] After this stage of processing, the non-task-related noise in the EEG signal is significantly reduced, the amplitude scale is uniform, and it has strong temporal continuity and spatial consistency, meeting the technical requirements of downstream multi-domain feature engineering.

[0039] S20: Feature extraction is performed on the preprocessed EEG signal. The extracted features include time-domain features, frequency-domain features, time-frequency domain features, and inter-channel features.

[0040] Specifically, after signal preprocessing, a highly discriminative information representation is extracted from the original EEG signal based on a structured feature extraction process. In this embodiment, a 158-dimensional multi-domain composite feature representation is finally obtained by compressing the 24,192-dimensional original features. The overall feature design revolves around four aspects: time-domain statistical characteristics, spectral distribution structure, time-frequency dynamic response, and inter-channel cooperative modes.

[0041] In terms of constructing time-domain features, 20 basic statistics were calculated for each channel signal, including mean, standard deviation, variance, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, median, interquartile range (IQR), root mean square (RMS), third moment, fourth moment, 5th and 95th percentiles, mean / standard deviation / variance of the first difference, and zero crossover rate. These indicators characterize the overall properties of the signal in terms of amplitude, waveform shape, and instantaneous changes.

[0042] In terms of frequency domain feature construction, Welch's method is used to estimate the power spectral density (PSD) of the signal. The relative power and spectral peak frequency of each frequency band are calculated around the main frequency band of neural oscillation (δ, θ, α, β, γ1, γ2). Furthermore, 13 advanced spectral structure features such as spectral centroid, spectral bandwidth, and spectral entropy are introduced to characterize the concentration and distribution complexity of the signal spectrum and reflect the energy contribution structure of different frequency bands to emotional states.

[0043] In constructing time-frequency features, Continuous Wavelet Transform (CWT) is introduced, and Morlet wavelet is selected as the mother function. Multi-scale decomposition is performed within the scale range s=1~10. Five statistical measures, including mean, standard deviation, maximum value, variance, and energy, are extracted from the wavelet coefficients to capture the transient oscillation mode and local time-frequency coupling structure of the signal, effectively reflecting the short-term dynamic response induced by emotions.

[0044] In terms of channel collaboration modeling, a cross-channel feature set was further constructed. By analyzing the Pearson correlation coefficient matrix and power ratio matrix between different channels, eight statistical features, including maximum, minimum, mean, and standard deviation, were extracted to reflect the synchronicity and collaborative activation intensity between brain regions, thereby enhancing the expressive power of the features at the spatial structure level.

[0045] Ultimately, this multi-domain feature engineering module integrates information flows from multiple perspectives to construct a compact and highly discriminative intermediate feature representation from the original redundant EEG signals. This provides a high-quality and highly expressive input foundation for subsequent task-oriented compression and deep modeling, significantly improving modeling efficiency and emotion recognition performance.

[0046] S30: Perform secondary preprocessing and enhancement on the extracted features.

[0047] In this embodiment, the extracted 158-dimensional feature vectors are further cleaned, standardized, and enhanced to improve feature quality, robustness, and the model's generalization ability.

[0048] Specifically, this includes: outlier detection and data cleaning.

[0049] An Isolation Forest algorithm is employed to identify and remove outlier samples from the feature space. In this approach, the contamination rate parameter is set to 0.1, indicating that 10% of the dataset is expected to contain outliers. Isolation Forest is an unsupervised anomaly detection method based on ensemble learning. It uses a tree structure (Isolation Tree) constructed from random hyperplanes to segment samples, and outliers typically have shorter average path lengths. The calculated anomaly score effectively identifies samples in noisy or abnormal states, thereby removing potential interference factors from the data. This method is stable on high-dimensional data, has low computational cost, and does not rely on data distribution assumptions, making it suitable for outlier removal from complex EEG features.

[0050] Feature standardization:

[0051] A robust scaler is used to standardize the cleaned data. This method removes the median and uses interquartile range (IQR) scaling. The standardization formula is as follows:

[0052]

[0053] in, These represent the 25th, 50th (median), and 75th percentiles of the feature, respectively. Compared to traditional normalization methods, this scheme is more robust to residual outliers and can further stabilize the feature scale.

[0054] Data augmentation:

[0055] To alleviate the imbalance between emotion categories, an oversampling strategy combining Gaussian noise and random scaling is employed. Generated samples are added to the category with fewer samples, increasing its sample size to twice that of the majority class. The specific method is as follows:

[0056] 1. Random scaling: The original sample vector is multiplied by a random scaling factor that follows the interval [0.95, 1.05].

[0057] 2. Gaussian perturbation: Add Gaussian noise with a standard deviation of 2% of the original sample to the scaled sample, as shown in the following formula:

[0058]

[0059] in, Represents the original feature vector. This represents the random scaling factor, and N represents Gaussian noise. It represents the standard deviation of the original sample.

[0060] This enhancement strategy improves data diversity and expands the distribution range of minority class samples in the feature space while maintaining the semantic rationality of the samples, thereby helping to learn more robust classification boundaries.

[0061] S40: Use mutual information to perform a first screening of the enhanced feature set, and then use the F statistic to perform a second screening of the feature set after the first screening to obtain the final target feature set.

[0062] Specifically, the enhanced high-dimensional feature set is subject to task-oriented screening to ensure that the final retained feature subset has the highest discriminative power for the emotion classification task.

[0063] A two-stage SelectKBest screening process is adopted. First, the top 150 features are selected based on mutual information (mutual_info_classif), and then the F statistic (f_classif) is used to further filter out the final 128 core features. This method integrates nonlinear and linear relationship modeling, and can take into account both complex feature dependencies and significant mean differences.

[0064] Specifically, mutual information measures the statistical dependency between features and class labels, calculated based on information entropy, and is defined as follows:

[0065]

[0066] in, The entropy of the category label. This represents the conditional entropy given feature F.

[0067] The F-statistic, based on the concept of analysis of variance (ANOVA), measures the difference in the mean of a feature between different categories. Its calculation is expressed as follows:

[0068]

[0069] Where α is the harmonic coefficient. This represents the mutual information between features and categories. This represents the corresponding F-value statistic.

[0070] The two-stage screening strategy achieves effective aggregation of multi-angle feature information by first coarsely screening nonlinear dependencies and then finely screening linearly separable features.

[0071] It should be noted that the above screening process can be equivalent to a weighted scoring function Si, used to measure the overall discriminative power of the i-th feature:

[0072]

[0073] Based on the scores of Si, the first 128 features are retained to form the final compressed feature matrix used for model training.

[0074] As can be seen from the above description, the EEG feature compression method provided in this embodiment achieves compression from the original 24,192-dimensional feature vector to a 128-dimensional highly discriminative expression through the above process, with a compression ratio of 0.53% and a corresponding data compression rate of 99.47%. While ensuring the preservation of highly discriminative information, it significantly reduces redundant dimensions, lowers model complexity and training costs, and provides an efficient and interpretable data support foundation for the subsequent construction of a lightweight emotion classification model.

[0075] The beneficial effects of the EEG feature compression method provided in this embodiment will be further explained below with reference to specific experimental data:

[0076] Example 1: Storage Reduction Based on Data Compression

[0077] This embodiment verifies the comprehensive advantages of the structured EEG compression method proposed in this invention in terms of input dimensionality reduction, storage saving, and model training efficiency, especially in systematically solving the engineering bottleneck of high-dimensional EEG data in deep emotion recognition tasks.

[0078] The experiment was based on the DEAP dataset, with the original input consisting of EEG signals with 32 channels × 8064 sampling points per segment, resulting in a total dimension of 24,192, which presents significant feature redundancy and storage pressure. To address this, we first constructed a multi-level feature extraction process covering the time domain, frequency domain, and time-frequency domain, extracting a 1024-dimensional task-related intermediate representation. Based on this, we introduced an outlier detection mechanism to remove artifact samples and used robust standardization to unify the distribution of multiple participants, ensuring the stability and consistency of the compressed data.

[0079] Subsequently, the feature selection module progressively filters to the final 128-dimensional representation through a joint mechanism of mutual information (MI) and F-test, serving as the compressed input for the model. This process significantly compresses the original input space while maximizing the preservation of discriminative information. Calculated with float32 precision, the size of a single sample is reduced from the original 6.2MB to only 0.5KB, achieving an overall compression rate of 99.47%. The space requirement on the complete dataset is reduced from 189MB to approximately 1MB, significantly reducing storage and I / O load.

[0080] In terms of model construction, the 128-dimensional compressed input was uniformly fed into a task-oriented ResNet structure for training. An optimized batch size of 64, combined with 5-fold cross-validation and early stopping strategies, was employed to control overfitting risk while maintaining convergence speed. During training and inference, the loading speed of the compressed input on the GPU was significantly improved, while memory usage and computational latency decreased simultaneously, making it particularly suitable for real-time EEG processing needs in resource-constrained scenarios.

[0081] Reference Figure 2-4 Experimental results show that the invention, through task-oriented feature cleaning and compression techniques, reduces EEG data storage requirements by 99.5% while improving model accuracy by 5.7%. Despite the significant reduction in input dimensionality, this method still maintains strong competitiveness in classification performance. While maintaining a uniform input structure, the compressed model achieves an average accuracy improvement of 5.7% compared to the unfiltered traditional model. AUC and F1-score are also significantly optimized, and the performance variance is minimized under multi-fold validation, demonstrating excellent generalization stability.

[0082] The structural compression mechanism proposed in Example 1 not only significantly reduces the data size and storage burden of the original EEG signals, but also maintains or even improves the classification performance under a unified 128-dimensional input, providing a reliable technical foundation for the subsequent deployment of embedded, edge, and low-power EEG emotion recognition systems.

[0083] Example 2: Improved real-time performance and processing speed

[0084] The main purpose of Example 2 is to verify the processing speed advantage of the compressed pipeline based on multi-domain feature engineering and task-oriented screening proposed in this invention in real-time EEG emotion recognition, and to compare and analyze it with three traditional high-dimensional data processing methods, highlighting the technical advantages of this method in significantly improving system response speed and computational efficiency while maintaining high-precision classification performance.

[0085] This experiment, based on the standard EEG emotion recognition dataset, constructed a compression pipeline system and a three-class contrastive architecture. The compression pipeline compresses the original 24,192-dimensional EEG data to 128-dimensional discriminative features through multi-domain feature engineering (time-domain statistics, Welch frequency-domain analysis, Morlet wavelet time-frequency transform, and inter-channel collaborative modeling), and then performs task-oriented optimization by combining a two-stage screening mechanism of mutual information and F-statistic. The contrastive methods, on the other hand, perform real-time inference calculations directly on the original high-dimensional data or after only simple dimensionality reduction preprocessing.

[0086] The performance metrics of this embodiment are evaluated using the following four metrics:

[0087] Single-sample inference delay: end-to-end time from EEG input to sentiment output;

[0088] Batch throughput: The number of EEG samples that can be processed per unit of time;

[0089] Peak memory usage: The maximum memory consumption during the inference process;

[0090] Computational complexity: number of floating-point operations (FLOPs).

[0091] The experimental results are shown in Table 1:

[0092] Table 1 Comparison of experimental results between this embodiment and existing methods

[0093] method Single-sample delay (ms) Throughput (samples / s) Peak memory usage (MB) FLOPs (M) Compression pipeline (this invention) 2.3 434.8 0.5 0.013 Raw data + ResNet 45.7 21.9 24.6 2.41 PCA dimensionality reduction + CNN 38.2 23.1 18.7 1.89 ICA preprocessing + Transformer 52.1 19.2 12.3 3.15

[0094] The above-described Example 2 verifies that the compressed pipeline proposed in this invention achieves millisecond-level latency, a hundredfold increase in throughput, and extremely low memory usage in real-time EEG emotion recognition, significantly outperforming traditional high-dimensional modeling methods. While maintaining classification accuracy, this method effectively overcomes the real-time bottleneck, providing an efficient and feasible technical path for portable brain-computer interfaces, mobile emotion monitoring, and edge computing platforms.

[0095] Example 3: Task-Oriented Adaptability Enhancement

[0096] This embodiment focuses on the interference of non-specific dimensions on classification performance and generalization robustness during EEG feature compression and discriminative modeling. By introducing a task-oriented feature cleaning and filtering mechanism, the feature quality and model stability in complex environments are effectively improved.

[0097] First, a feature set containing 1024 dimensions was constructed based on the original multi-domain EEG data. To eliminate potential non-discriminatory interference dimensions, a two-stage feature cleaning process was implemented. The first stage used the IsolationForest algorithm to detect and remove outliers between feature dimensions, thus initially eliminating features with discrete distributions or abnormal structures. The second stage combined mutual information (MI) and the ANOVA statistic for joint screening, further retaining dimensions significantly relevant to the task label, ultimately generating a cleaned input representation compressed to 128 dimensions. The output of this process is the "task-oriented filtered" EEG feature set.

[0098] As a control group, we used the standard PCA method to compress the original features, also obtaining a 128-dimensional input, but without involving anomaly removal or correlation optimization. After the two sets of features were compressed, they were respectively input into a task-oriented ResNet model with an isomorphic structure for training, maintaining consistency in the training process and fairness in the evaluation.

[0099] Regarding the cleaning effect, we first compared the differences between the two groups in terms of the proportion of discriminative features. For example... Figure 5 As shown, the effective feature ratio of the task-oriented group reached 82.7%, while that of the PCA group was only 60.9%, with the former showing a relative improvement of 35.6%, which verifies the advantage of the cleaning process in feature purity.

[0100] Subsequently, based on feature cleaning, cross-scenario testing was conducted to comprehensively evaluate its robustness improvement effect. In the cross-subject generalization task, the training data did not include the target subjects, testing whether the model could adapt to new individual data; in multi-class classification, the original binary classification task was extended to four-class classification to examine the adaptability of the feature space to complex label structures; in noise resistance analysis, we injected 10dB Gaussian noise into the original signal to test the model's anti-interference ability; finally, in five-fold cross-validation, we calculated the overlap rate between the selected feature sets in each round to evaluate the consistency and stability of the feature selection mechanism.

[0101] Figure 6 The results show a comparison of the performance of the two groups across four metrics. The results indicate that the task-oriented group significantly outperformed the control group in all dimensions: cross-subject accuracy reached 84.4%, multi-class expansion attenuation was controlled at 3.8%, accuracy was maintained at 78.9% even under noisy conditions, and feature selection consistency reached 92.1%, fully demonstrating the stable adaptability of the method under high-dimensional noise conditions.

[0102] Finally, to present the overall performance improvement, we have visualized the four core metrics mentioned above and constructed a comparison chart as follows: Figure 7 As shown, the task-oriented group achieved a relative improvement of 22.5% in average indicators, with particularly significant improvements in noise robustness and feature consistency, highlighting the practical value of the strategy of this invention in complex EEG environments.

[0103] In this embodiment, under a unified 128-dimensional compressed EEG feature, the task-oriented ResNet model of the present invention significantly outperforms traditional shallow networks and traditional machine learning methods, with an average accuracy improvement of more than 5.7%. It also shows significant advantages in AUC and F1-score, while exhibiting the smallest performance variance. This verifies its discriminative power and stability under "high compression rate + low-dimensional representation", providing a solid basis for its subsequent application in edge deployment and wearable EEG devices.

[0104] Reference Figure 8 The present invention also provides a device for compressing electroencephalogram (EEG) features, comprising:

[0105] The first preprocessing unit 100 is used to collect the subject's electroencephalogram (EEG) signals and preprocess the EEG signals. It should be noted that since the specific preprocessing methods and processes have been described in detail in step S10 of the above-mentioned method for compressing EEG features, they will not be repeated here.

[0106] The feature extraction unit 200 is used to extract features from the preprocessed EEG signal. The extracted features include time-domain features, frequency-domain features, time-frequency-domain features, and inter-channel features. It should be noted that since the specific feature extraction method and process have been described in detail in step S20 of the above-mentioned EEG feature compression method, they will not be repeated here.

[0107] The second preprocessing unit 300 is used to perform secondary preprocessing and enhancement on the extracted features. It should be noted that since the specific preprocessing methods and processes have been described in detail in step S30 of the above-mentioned method for compressing EEG features, they will not be repeated here.

[0108] The screening unit 400 is used to perform a first screening of the enhanced feature set using mutual information, and then to perform a second screening of the feature set after the first screening using the F statistic to obtain the final target feature set. It should be noted that since the specific screening method and process have been described in detail in step S40 of the above-mentioned EEG feature compression method, they will not be repeated here.

[0109] In addition, embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium may store a program, which, when executed, includes some or all of the steps of the compression method for any EEG features described in the above method embodiments.

[0110] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0111] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0112] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0113] The above description, with reference to the accompanying drawings, illustrates an exemplary flowchart for implementing a method for compressing electroencephalogram (EEG) features according to an embodiment of the present invention. It should be noted that the numerous details included in the above description are merely illustrative of the invention and not intended to limit it. In other embodiments of the invention, the method may have more, fewer, or different steps, and the order, inclusion, function, and other relationships between the steps may differ from those described and illustrated.

Claims

1. A method for compressing electroencephalogram (EEG) features, characterized in that, include: Collect the subject's electroencephalogram (EEG) signals and preprocess the EEG signals; Feature extraction was performed on the preprocessed EEG signals. The extracted features included time-domain features, frequency-domain features, time-frequency-domain features, and inter-channel features. The extracted features are then preprocessed and enhanced. The enhanced feature set is filtered once using mutual information, and then the F-statistic is used to filter the feature set a second time to obtain the final target feature set.

2. The compression method according to claim 1, characterized in that, Preprocessing of the EEG signals includes: A bandpass filter is used to limit the frequency of the EEG signal to a preset threshold, eliminating low-frequency drift and high-frequency noise components, and retaining the effective components within the emotional frequency band. A combined strategy of independent component analysis and artifact detection algorithms is adopted to remove physiological interference signals; To address the differences in data distribution across different channels, a standardization process is introduced to convert the signals from each channel into a standard normal distribution with zero mean and unit variance, thereby unifying the characteristic scale.

3. The compression method according to claim 1, characterized in that: The time-domain features include one or more of the following: mean, standard deviation, variance, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, median, interquartile range, root mean square, third moment, fourth moment, 5th and 95th percentiles, mean / standard deviation / variance of the first difference, and zero crossover rate.

4. The compression method according to claim 1, characterized in that, The frequency domain features are extracted in the following way: The Welch method is used to estimate the power spectral density of the signal; Calculate the relative power and spectral peak frequency of each frequency band around the main frequency band of neural oscillation; The structural features of spectral centroid, spectral bandwidth, and spectral entropy are introduced to characterize the concentration and distribution complexity of the signal spectrum, reflecting the energy contribution structure of different frequency bands to emotional states.

5. The compression method according to claim 1, characterized in that, The time-frequency domain features are extracted in the following way: Using continuous wavelet transform, Morlet wavelet is selected as the mother function, and multi-scale decomposition is performed within a preset scale range. By extracting the mean, standard deviation, maximum value, variance, and energy from wavelet coefficients as statistical quantities, the transient oscillation mode and local time-frequency coupling structure of the signal are captured, effectively reflecting the short-term dynamic response induced by emotions.

6. The compression method according to claim 1, characterized in that, The extracted features undergo secondary preprocessing and enhancement, including: The isolated forest algorithm is used to identify and remove outlier samples in the feature space. A robust scaler is used to standardize the cleaned data. Multiply the standardized data sample vector by a random scaling factor within a preset range; Add Gaussian noise, which follows a standard deviation of 2% of the original sample, to the scaled sample, as shown in the following formula: ； in, Represents the original feature vector. This represents the random scaling factor, and N represents Gaussian noise. It represents the standard deviation of the original sample.

7. The compression method according to claim 1, characterized in that, The expression for the mutual information is as follows: ； in, The entropy of the category label. This represents the conditional entropy given feature F.

8. The compression method according to claim 1, characterized in that, The expression for the F-statistic is as follows: ； Where α is the harmonic coefficient, This represents the mutual information between features and categories. This represents the corresponding F-value statistic.

9. A device for compressing brainwave characteristics, characterized in that, include: The first preprocessing unit is used to collect the subject's electroencephalogram (EEG) signals and preprocess the EEG signals; The feature extraction unit is used to extract features from the preprocessed EEG signal. The extracted features include time-domain features, frequency-domain features, time-frequency-domain features, and inter-channel features. The second preprocessing unit is used to perform secondary preprocessing and enhancement on the extracted features; The filtering unit is used to perform a first filtering of the enhanced feature set using mutual information, and then to perform a second filtering of the feature set after the first filtering using the F statistic to obtain the final target feature set.

10. A computer-readable storage medium comprising a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of a method for compressing electroencephalogram (EEG) features as described in any one of claims 1 to 8.

Images