EEG signal abnormity identification method and system based on big data analysis

Abstract

The invention discloses an EEG signal abnormity identification method and system based on big data analysis. The method comprises the following steps: S1, collecting an EEG signal of a target testee, and carrying out filtering, denoising and baseline correction preprocessing; s2, a Sinc VAE anomaly recognition model with a learnable Sinc filter is built; s3, adopting a multi-target squirrel search algorithm to adaptively optimize the model; s4, constructing an EEG signal big data feature database and determining reconstruction error threshold ranges of normal and abnormal states; s5, performing semi-supervised training on the model by using the normal state data; s6, inputting EEG signal data to be detected, calculating a reconstruction error and identifying abnormity; and S7, carrying out classification processing on the abnormal signals and carrying out early warning. According to the method, the EEG anomaly recognition precision and real-time performance are improved, and the method is suitable for clinical electroencephalogram signal anomaly diagnosis and auxiliary decision making.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and biomedical signal technology, and in particular to a method and system for identifying EEG signal anomalies based on big data analysis. Background Technology

[0002] Electroencephalography (EEG) signals, due to their non-invasive and real-time ability to reflect brain neurophysiological activity, have been widely used in the clinical diagnosis of epilepsy, sleep disorders, mental illnesses, and brain dysfunction. Current EEG signal anomaly identification techniques mainly rely on traditional time-frequency analysis methods, such as Fourier transform, wavelet transform, and short-time Fourier transform, to analyze specific frequency bands or features manually or semi-automatically to identify abnormal signals. However, EEG signals possess nonlinear, non-stationary, and high-dimensional characteristics, and these traditional methods often suffer from low recognition accuracy, poor generalization ability, and insufficient real-time processing efficiency when processing complex EEG data.

[0003] In recent years, with the rapid development of artificial intelligence technology, deep learning algorithms such as convolutional neural networks, recurrent neural networks, and variational autoencoders have been gradually introduced into the field of EEG signal analysis, improving the accuracy and stability of abnormal signal recognition through automatic feature learning. In particular, anomaly detection methods based on variational autoencoders can effectively capture the probability distribution characteristics of data, thus finding some application in EEG anomaly recognition. However, existing deep learning methods typically still require manual pre-setting or empirical selection of model structures, hyperparameters, and feature extraction methods, making it difficult for the models to accurately adapt to diverse clinical scenarios and affecting the real-time performance and accuracy of anomaly detection.

[0004] Furthermore, the integration of big data analysis and optimization algorithms has become a new trend in the field of EEG signal anomaly recognition. For example, optimization methods such as genetic algorithms and particle swarm optimization have been gradually applied to the selection and optimization of hyperparameters in neural networks. However, existing optimization algorithms suffer from shortcomings such as low search efficiency and susceptibility to getting trapped in local optima, making it difficult to simultaneously meet the multi-objective real-time optimization requirements in EEG anomaly detection scenarios. Therefore, existing technologies have certain limitations in terms of model real-time performance, adaptive feature selection, and overall generalization ability, making it difficult to effectively support the needs of efficient and accurate clinical diagnosis.

[0005] Therefore, how to provide a method and system for identifying EEG signal anomalies based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for identifying EEG signal anomalies based on big data analysis. This invention has the technical advantages of automated frequency band selection, adaptive optimization of model parameters, high accuracy of anomaly identification, and strong real-time performance.

[0007] An EEG signal anomaly identification method based on big data analysis according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect the EEG signal of the target subject, and preprocess the EEG signal through filtering, noise reduction and baseline correction methods to obtain EEG signal data;

[0009] S2. Construct a SincVAE anomaly recognition model with a learnable Sinc filter;

[0010] S3. Adaptive optimization of the SincVAE anomaly recognition model is performed using a multi-target squirrel search algorithm to obtain an optimized SincVAE anomaly recognition model.

[0011] S4. Construct a big data feature database for EEG signals. By conducting big data analysis on historical EEG signal data in normal and abnormal states, determine the reconstruction error threshold range of EEG signal data in normal and abnormal states.

[0012] S5. Use normal EEG signal data to perform semi-supervised training on the optimized SincVAE anomaly recognition model to obtain the trained SincVAE anomaly recognition model.

[0013] S6. Input the EEG signal data to be detected into the trained SincVAE anomaly recognition model. Process the EEG signal data to be detected, calculate the reconstruction error of the EEG signal data to be detected, and determine whether the EEG signal data to be detected is abnormal based on the reconstruction error threshold range.

[0014] S7. Classify the identified abnormal EEG signal data according to the preset abnormal signal classification criteria, obtain the abnormal signal classification results, and output them to the clinical diagnostic system for early warning.

[0015] Optionally, the SincVAE anomaly detection model with a learnable Sinc filter includes a two-stage Sinc filter module, an adaptive multi-scale spatiotemporal fusion convolution module, and a two-stream interactive self-attention latent variable encoding module.

[0016] The dual-stage Sinc filter module is used to perform adaptive frequency band feature extraction on the input EEG signal data, including an initial Sinc filtering stage and a secondary adaptive adjustment stage.

[0017] The adaptive multi-scale spatiotemporal fusion convolution module is used to extract high-order spatiotemporal features from the output of the dual-stage Sinc filter module, including three parallel convolutional feature channels.

[0018] The dual-stream interactive self-attention latent variable encoding module is used to perform latent variable encoding on the output of the adaptive multi-scale spatiotemporal fusion convolution module, and includes a first attention encoding stream and a second attention encoding stream.

[0019] Optionally, the dual-stage Sinc filter module specifically includes an initial Sinc filtering stage and a secondary adaptive adjustment stage:

[0020] In the initial Sinc filtering stage, the input EEG signal data is divided into segments using a short-time sliding window. The short-time energy ratio of the EEG signal within each sliding window in each preset frequency band is calculated. The objective function is to maximize the energy ratio of the EEG signal within the short-time window. The low cutoff frequency and high cutoff frequency parameters of multiple parallel Sinc filters are iteratively adjusted to obtain preliminary frequency band characteristic data.

[0021] In the secondary adaptive adjustment stage, the goal of the secondary adaptive adjustment is to maximize the multi-dimensional spatiotemporal differences of the initial frequency band feature data. The interaction information contribution between the temporal features and spatial electrode position features of the initial frequency band feature data is analyzed. A nonlinear iterative optimization method is used to sort the interaction intensity of the frequency band contribution. Based on the sorting of the interaction intensity of the frequency band contribution, the low cutoff frequency parameter and high cutoff frequency parameter of each filter in the Sinc bandpass filter bank are adjusted to output optimized frequency band feature data with spatiotemporal differences.

[0022] The preliminary frequency band feature data and the optimized frequency band feature data are aligned with the index positions of the input EEG signal data and then output.

[0023] Optionally, the adaptive multi-scale spatiotemporal fusion convolution module specifically includes large-scale convolutional feature channels, medium-scale convolutional feature channels, and small-scale convolutional feature channels:

[0024] The large-scale convolutional feature channel uses a convolution kernel with a size larger than the length of a single time-domain period of the input optimized frequency band feature data to perform joint convolution operations on the input optimized frequency band feature data in the spatial domain electrode dimension and the time-domain signal dimension. Through a single convolution, it captures the spatiotemporal features of the overall trend of EEG signal data in the entire electrode space and multiple consecutive time-domain periods, and obtains global spatiotemporal feature data.

[0025] The mesoscale convolutional feature channel uses a convolution kernel with a size equal to the length of a single temporal period of the input optimized frequency band feature data. It performs convolution operations on the input optimized frequency band feature data in the spatial and temporal dimensions in a multi-layer stacked convolution manner to obtain mesoscale local features with distinct patterns in the local region of the spatial electrode location and in the temporal period, respectively. The local spatiotemporal feature data is obtained by fusing the multi-scale local feature weighting method.

[0026] The small-scale convolutional feature channel uses a convolution kernel with a size smaller than the length of a single time-domain period of the input optimized frequency band feature data. It performs convolution calculations on the spatial electrode position and time-domain signal of the input optimized frequency band feature data in a deep, step-by-step iterative convolution method. This extracts the interaction difference features between spatially adjacent electrodes and the subtle amplitude fluctuation pattern features within a shorter time period from the EEG signal data, thereby obtaining fine-grained spatiotemporal feature data.

[0027] After aligning the index positions of the global spatiotemporal feature data, local spatiotemporal feature data, and fine-grained spatiotemporal feature data, multi-scale spatiotemporal feature data is obtained by multi-scale convolution fusion.

[0028] Optionally, the dual-stream interactive self-attention latent variable encoding module specifically includes a first attention encoding stream and a second attention encoding stream:

[0029] The first attention encoding stream takes multi-scale spatiotemporal feature data as input and calculates the feature weights of EEG signals at different time positions in the time domain dimension through a self-attention mechanism to obtain temporal latent variable encoding with prominent features in the time dimension.

[0030] The second attention coding stream takes multi-scale spatiotemporal feature data as input and calculates the feature weights of different spatial electrode positions of the EEG signal in the spatial domain electrode position dimension through a self-attention mechanism to obtain spatial latent variable codes with prominent features in the spatial dimension.

[0031] The first attention encoding stream and the second attention encoding stream exchange feature weight parameters of temporal latent variable encoding and spatial latent variable encoding in real time through an interactive feedback mechanism, and use an iterative weighted update method to adjust the weight parameters in their respective self-attention mechanisms to generate optimized latent variable encoding feature data.

[0032] Optionally, S3 specifically includes:

[0033] S31. Establish an initial squirrel population. The position of each squirrel individual is encoded in the form of a real number vector, which corresponds to the number of filters and frequency range of the two-stage Sinc filter module, the kernel size of the adaptive multi-scale spatiotemporal fusion convolution module, the dimension of the latent variables of the two-stream interactive self-attention latent variable encoding module, and the number of model training iterations.

[0034] S32. Input the model parameter configuration corresponding to each individual squirrel into the SincVAE anomaly recognition model, and calculate the four performance indicators of the model output: classification accuracy, recall, model complexity, and real-time performance, to form a multi-objective evaluation objective.

[0035] S33. Based on the multi-objective evaluation objectives, perform non-dominated ranking of the squirrel population, identify and retain elite squirrel individuals that meet the multi-objective evaluation requirements;

[0036] S34. For non-elite squirrel individuals, based on the positional difference between them and elite squirrel individuals, update the positional parameters of non-elite squirrel individuals according to a non-linear adjustment method to form a new generation of squirrel population.

[0037] S35. Re-input the location parameters of each squirrel individual in the updated new generation squirrel population into the SincVAE anomaly detection model, and re-evaluate the classification accuracy, recall, model complexity, and real-time performance. Update the multi-objective evaluation objective and perform non-dominated sorting again.

[0038] S36. Repeatedly execute the steps of non-dominated sorting, elite individual selection and non-elite individual position parameter nonlinear adjustment until the preset maximum number of iterations is reached, and obtain the optimized SincVAE anomaly recognition model that satisfies the optimal multi-objective performance index.

[0039] Optionally, S4 specifically includes:

[0040] S41. Obtain historical EEG signal data and divide the historical EEG signal data into historical normal state EEG signal data and historical abnormal state EEG signal data according to the clinical diagnosis results.

[0041] S42. Calculate the reconstruction error values ​​of each historical normal state EEG signal data and historical abnormal state EEG signal data respectively, and obtain the reconstruction error set of historical normal state EEG signal data and the reconstruction error set of historical abnormal state EEG signal data.

[0042] S43. Perform statistical analysis on the set of reconstruction errors of EEG signal data under historical normal conditions, determine the probability distribution of reconstruction errors of EEG signal data under historical normal conditions, and obtain the probability density function of reconstruction errors of EEG signal data under historical normal conditions.

[0043] S44. Perform statistical analysis on the set of reconstruction errors of EEG signal data under historical abnormal conditions, determine the probability distribution of reconstruction errors of EEG signal data under historical abnormal conditions, and obtain the probability density function of reconstruction errors of EEG signal data under historical abnormal conditions.

[0044] S45. Based on the probability density function of the reconstruction error of historical normal EEG signal data and historical abnormal EEG signal data, determine the boundary point between the reconstruction error of normal EEG signal data and abnormal EEG signal data, and obtain the threshold range of reconstruction error for normal and abnormal EEG signal data.

[0045] S46. Based on the reconstruction error threshold range, construct a big data feature database of EEG signals and determine the reconstruction error threshold range of the EEG signal data to be detected.

[0046] Optionally, S6 specifically includes:

[0047] S61. Divide the EEG signal data to be detected into multiple time-domain subsequences, and input each time-domain subsequence into the trained SincVAE anomaly recognition model. Through the dual-stage Sinc filter module, adaptive multi-scale spatiotemporal fusion convolution module and dual-stream interactive self-attention latent variable encoding module inside the model, obtain the latent variable feature data corresponding to each subsequence.

[0048] S62. Decode the latent variable feature data of each subsequence to obtain the corresponding reconstructed subsequence data, and calculate the local reconstruction error value of each subsequence.

[0049]

[0050] in, This represents the local reconstruction error value of the j-th time-domain subsequence. This represents the actual EEG signal value at the m-th sampling point in the j-th time-domain subsequence. Let L represent the reconstructed EEG signal value at the m-th sampling point in the j-th time-domain subsequence, L represent the number of sampling points in a single time-domain subsequence, and δ be a minimal positive real constant, where 0 < δ ≤ 10. -6 ;

[0051] The formula calculates the difference between the actual EEG signal data and its corresponding reconstructed data for each subsequence, deriving a sensitive and stable local reconstruction error value. The formula first calculates the squared difference between the actual value and the reconstructed value for each sampling point, then introduces a relative error coefficient to highlight signal segments with significant abnormal fluctuations; a very small positive real constant δ is used to avoid zero denominators, ensuring computational stability.

[0052] S63. Calculate the overall reconstruction error ε of the EEG signal data to be detected based on the local reconstruction error values ​​of all time-domain subsequences. global :

[0053]

[0054] Where, ε global The value represents the overall reconstruction error of the EEG signal data to be detected, J represents the total number of time-domain subsequences, and μ represents the total number of time-domain subsequences. n σ represents the mean of the local reconstruction error of historical normal EEG signal data. nThis represents the standard deviation of the local reconstruction error of EEG signal data under historical normal conditions.

[0055] The formula calculates the overall reconstruction error of the EEG signal data under test using an exponential weighting method based on the local reconstruction error values ​​of each time-domain subsequence. The formula utilizes the difference between the mean reconstruction error of each local reconstruction error and the reconstruction error of historical normal EEG signal data, combined with the standard deviation of historical data, to amplify anomalous segments with significant local differences using an exponential function, thereby achieving accurate identification of the overall error.

[0056] S64. Based on the EEG signal big data feature database, determine the reconstruction error threshold range for normal and abnormal EEG signal data, and set the overall reconstruction error value ε. global By comparing the data with the threshold range, it can be determined whether the EEG signal data to be detected belongs to a normal or abnormal state.

[0057] Optionally, an EEG signal anomaly identification system based on big data analysis includes the following modules:

[0058] The data acquisition and preprocessing module is used to acquire the EEG signals of the target subject and preprocess the EEG signals through filtering, noise reduction and baseline correction methods to obtain EEG signal data;

[0059] The SincVAE anomaly detection model is used for adaptive frequency band feature extraction, multi-scale spatiotemporal feature fusion, and latent variable feature encoding of EEG signal data.

[0060] The model optimization module is used to adaptively optimize the SincVAE anomaly recognition model using a multi-target squirrel search algorithm.

[0061] The reconstruction error threshold database construction module is used to build a big data feature database of EEG signals, divide historical normal and abnormal EEG signal data, and determine the reconstruction error threshold range of normal and abnormal EEG signal data.

[0062] The anomaly detection module is used to divide the EEG signal data to be detected into multiple time-domain subsequences, calculate the local reconstruction error value of each time-domain subsequence, and calculate the overall reconstruction error value. By comparing it with the reconstruction error threshold range, it determines whether the EEG signal data to be detected is abnormal.

[0063] The beneficial effects of this invention are:

[0064] (1) This invention builds a SincVAE anomaly recognition model with a learnable Sinc filter and uses a multi-target squirrel search algorithm for adaptive optimization, thereby realizing the automated and intelligent selection of EEG signal frequency band features and the precise optimization of model parameters. This effectively improves the accuracy and generalization ability of EEG signal anomaly recognition and significantly enhances the adaptability and real-time processing capability in clinical EEG signal anomaly diagnosis scenarios.

[0065] (2) By constructing a big data feature database of EEG signals and performing statistical analysis on the reconstruction error of historical normal and abnormal state data, this invention can accurately determine the reconstruction error threshold range of the EEG signal to be detected, significantly improve the reliability and accuracy of abnormal signal detection, and show better robustness and application effect in complex EEG signal scenarios.

[0066] (3) In the detection of abnormal EEG signal data, this invention effectively solves the defects of insufficient sensitivity to minor abnormalities and unclear local abnormal signal location in traditional methods by using an innovative method for calculating the numerical value of local and overall reconstruction errors. It breaks through the bottleneck of existing technologies that make it difficult to accurately identify and locate minor abnormal signals, and realizes more accurate and faster identification and location of abnormal signals, thereby effectively improving the efficiency and reliability of clinical EEG signal abnormality diagnosis and auxiliary decision-making. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is a flowchart of an EEG signal anomaly identification method based on big data analysis proposed in this invention;

[0069] Figure 2 This is an architecture diagram of an EEG signal anomaly identification system based on big data analysis proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] refer to Figures 1-2 A method and system for identifying EEG signal anomalies based on big data analysis, comprising the following steps:

[0072] S1. Collect the EEG signal of the target subject, and preprocess the EEG signal through filtering, noise reduction and baseline correction methods to obtain EEG signal data;

[0073] S2. Construct a SincVAE anomaly recognition model with a learnable Sinc filter;

[0074] S3. Adaptive optimization of the SincVAE anomaly recognition model is performed using a multi-target squirrel search algorithm to obtain an optimized SincVAE anomaly recognition model.

[0075] S4. Construct a big data feature database for EEG signals. By conducting big data analysis on historical EEG signal data in normal and abnormal states, determine the reconstruction error threshold range of EEG signal data in normal and abnormal states.

[0076] S5. Use normal EEG signal data to perform semi-supervised training on the optimized SincVAE anomaly recognition model to obtain the trained SincVAE anomaly recognition model.

[0077] S6. Input the EEG signal data to be detected into the trained SincVAE anomaly recognition model. Process the EEG signal data to be detected, calculate the reconstruction error of the EEG signal data to be detected, and determine whether the EEG signal data to be detected is abnormal based on the reconstruction error threshold range.

[0078] S7. Classify the identified abnormal EEG signal data according to the preset abnormal signal classification criteria, obtain the abnormal signal classification results, and output them to the clinical diagnostic system for early warning.

[0079] By constructing a SincVAE anomaly recognition model with a learnable Sinc filter and employing a multi-objective squirrel search algorithm for adaptive model optimization, the extraction of EEG signal frequency band features and the selection of model parameters are automated, overcoming the shortcomings of traditional methods where frequency band features require manual selection and model parameters are difficult to optimize precisely. Simultaneously, by establishing a large-scale feature database of EEG signals and clarifying the reconstruction error threshold range between normal and abnormal states, this invention improves the accuracy and stability of anomaly recognition, demonstrating significant advantages in real-time performance, accuracy, and clinical decision support capabilities for EEG signal anomaly recognition.

[0080] In this embodiment, the SincVAE anomaly detection model with a learnable Sinc filter includes a two-stage Sinc filter module, an adaptive multi-scale spatiotemporal fusion convolution module, and a two-stream interactive self-attention latent variable encoding module.

[0081] The dual-stage Sinc filter module is used to perform adaptive frequency band feature extraction on the input EEG signal data, including an initial Sinc filtering stage and a secondary adaptive adjustment stage.

[0082] The adaptive multi-scale spatiotemporal fusion convolution module is used to extract high-order spatiotemporal features from the output of the dual-stage Sinc filter module, including three parallel convolutional feature channels.

[0083] The dual-stream interactive self-attention latent variable encoding module is used to perform latent variable encoding on the output of the adaptive multi-scale spatiotemporal fusion convolution module, and includes a first attention encoding stream and a second attention encoding stream.

[0084] By setting up a two-stage Sinc filter module, an adaptive multi-scale spatiotemporal fusion convolution module, and a dual-stream interactive self-attention latent variable encoding module, adaptive frequency band feature extraction, accurate fusion of multi-scale spatiotemporal features, and efficient encoding of latent variable features for EEG signal data are achieved. This overcomes the limitations of traditional models that rely on manually preset frequency band parameters and single-scale feature extraction, effectively improving the sensitivity and accuracy of EEG signal anomaly recognition, and enhancing the model's generalization ability and real-time processing performance for recognizing complex EEG signal abnormalities.

[0085] In this embodiment, the dual-stage Sinc filter module specifically includes an initial Sinc filtering stage and a secondary adaptive adjustment stage:

[0086] In the initial Sinc filtering stage, the input EEG signal data is divided into segments using a short-time sliding window. The short-time energy ratio of the EEG signal within each sliding window in each preset frequency band is calculated. The objective function is to maximize the energy ratio of the EEG signal within the short-time window. The low cutoff frequency and high cutoff frequency parameters of multiple parallel Sinc filters are iteratively adjusted to obtain preliminary frequency band characteristic data.

[0087] In the secondary adaptive adjustment stage, the goal of the secondary adaptive adjustment is to maximize the multi-dimensional spatiotemporal differences of the initial frequency band feature data. The interaction information contribution between the temporal features and spatial electrode position features of the initial frequency band feature data is analyzed. A nonlinear iterative optimization method is used to sort the interaction intensity of the frequency band contribution. Based on the sorting of the interaction intensity of the frequency band contribution, the low cutoff frequency parameter and high cutoff frequency parameter of each filter in the Sinc bandpass filter bank are adjusted to output optimized frequency band feature data with spatiotemporal differences.

[0088] The preliminary frequency band feature data and the optimized frequency band feature data are aligned with the index positions of the input EEG signal data and then output.

[0089] By using the initial Sinc filtering stage and the secondary adaptive adjustment stage of the dual-stage Sinc filter module, with the goals of maximizing the energy ratio of the short-time sliding window and maximizing the multi-dimensional spatiotemporal differences of frequency band features, the sensitive frequency band parameters of EEG signal data are adaptively determined and optimized. This effectively overcomes the limitations of traditional methods that rely on manual experience to preset frequency band parameters, achieving more accurate and efficient frequency band feature extraction and significantly improving the accuracy and real-time performance of abnormal EEG signal detection.

[0090] In this embodiment, the adaptive multi-scale spatiotemporal fusion convolution module specifically includes large-scale convolutional feature channels, medium-scale convolutional feature channels, and small-scale convolutional feature channels:

[0091] The large-scale convolutional feature channel uses a convolution kernel with a size larger than the length of a single time-domain period of the input optimized frequency band feature data to perform joint convolution operations on the input optimized frequency band feature data in the spatial domain electrode dimension and the time-domain signal dimension. Through a single convolution, it captures the spatiotemporal features of the overall trend of EEG signal data in the entire electrode space and multiple consecutive time-domain periods, and obtains global spatiotemporal feature data.

[0092] The mesoscale convolutional feature channel uses a convolution kernel with a size equal to the length of a single temporal period of the input optimized frequency band feature data. It performs convolution operations on the input optimized frequency band feature data in the spatial and temporal dimensions in a multi-layer stacked convolution manner to obtain mesoscale local features with distinct patterns in the local region of the spatial electrode location and in the temporal period, respectively. The local spatiotemporal feature data is obtained by fusing the multi-scale local feature weighting method.

[0093] The small-scale convolutional feature channel uses a convolution kernel with a size smaller than the length of a single time-domain period of the input optimized frequency band feature data. It performs convolution calculations on the spatial electrode position and time-domain signal of the input optimized frequency band feature data in a deep, step-by-step iterative convolution method. This extracts the interaction difference features between spatially adjacent electrodes and the subtle amplitude fluctuation pattern features within a shorter time period from the EEG signal data, thereby obtaining fine-grained spatiotemporal feature data.

[0094] After aligning the index positions of the global spatiotemporal feature data, local spatiotemporal feature data, and fine-grained spatiotemporal feature data, multi-scale spatiotemporal feature data is obtained by multi-scale convolution fusion.

[0095] By setting up an adaptive multi-scale spatiotemporal fusion convolution module, three parallel convolutional feature channels of large scale, medium scale, and small scale are used to extract overall trend features, local regional difference patterns, and fine-grained micro-amplitude fluctuation patterns from the optimized frequency band feature data, respectively. Finally, these are fused into multi-scale spatiotemporal feature data, which significantly improves the comprehensiveness and refinement of EEG signal feature extraction. This effectively solves the problem that traditional single-scale feature extraction cannot simultaneously take into account global trends and local micro-features, and improves the accuracy and generalization ability of abnormal signal detection.

[0096] In this embodiment, the dual-stream interactive self-attention latent variable encoding module specifically includes a first attention encoding stream and a second attention encoding stream:

[0097] The first attention encoding stream takes multi-scale spatiotemporal feature data as input and calculates the feature weights of EEG signals at different time positions in the time domain dimension through a self-attention mechanism to obtain temporal latent variable encoding with prominent features in the time dimension.

[0098] The second attention coding stream takes multi-scale spatiotemporal feature data as input and calculates the feature weights of different spatial electrode positions of the EEG signal in the spatial domain electrode position dimension through a self-attention mechanism to obtain spatial latent variable codes with prominent features in the spatial dimension.

[0099] The first attention encoding stream and the second attention encoding stream exchange feature weight parameters of temporal latent variable encoding and spatial latent variable encoding in real time through an interactive feedback mechanism, and use an iterative weighted update method to adjust the weight parameters in their respective self-attention mechanisms to generate optimized latent variable encoding feature data.

[0100] By designing a dual-stream interactive self-attention latent variable encoding module, the first attention encoding stream is used to encode the temporal salience of EEG signals, and the second attention encoding stream is used to encode the spatial electrode location salience. The latent variable encoding and weight parameters are exchanged in real time through an interactive feedback mechanism, which realizes efficient collaboration and deep fusion between spatiotemporal features. This effectively improves the sensitivity and accuracy of EEG signal anomaly detection and enhances the model's performance and adaptability in recognizing complex EEG signal abnormalities.

[0101] In this embodiment, S3 specifically includes:

[0102] S31. Establish an initial squirrel population. The position of each squirrel individual is encoded in the form of a real number vector, which corresponds to the number of filters and frequency range of the two-stage Sinc filter module, the kernel size of the adaptive multi-scale spatiotemporal fusion convolution module, the dimension of the latent variables of the two-stream interactive self-attention latent variable encoding module, and the number of model training iterations.

[0103] S32. Input the model parameter configuration corresponding to each individual squirrel into the SincVAE anomaly recognition model, and calculate the four performance indicators of the model output: classification accuracy, recall, model complexity, and real-time performance, to form a multi-objective evaluation objective.

[0104] S33. Based on the multi-objective evaluation objectives, perform non-dominated ranking of the squirrel population, identify and retain elite squirrel individuals that meet the multi-objective evaluation requirements;

[0105] S34. For non-elite squirrel individuals, based on the positional difference between them and elite squirrel individuals, update the positional parameters of non-elite squirrel individuals according to a non-linear adjustment method to form a new generation of squirrel population.

[0106] S35. Re-input the location parameters of each squirrel individual in the updated new generation squirrel population into the SincVAE anomaly detection model, and re-evaluate the classification accuracy, recall, model complexity, and real-time performance. Update the multi-objective evaluation objective and perform non-dominated sorting again.

[0107] S36. Repeatedly execute the steps of non-dominated sorting, elite individual selection and non-elite individual position parameter nonlinear adjustment until the preset maximum number of iterations is reached, and obtain the optimized SincVAE anomaly recognition model that satisfies the optimal multi-objective performance index.

[0108] By employing a multi-objective squirrel search algorithm to adaptively optimize the SincVAE anomaly recognition model, an initial squirrel population was constructed. Classification accuracy, recall, model complexity, and real-time performance were used as multi-objective evaluation indicators. The population was updated through non-dominated sorting and non-linear position adjustment, achieving efficient automatic optimization of model parameters. This overcame the limitation of traditional single-objective optimization being prone to getting trapped in local optima, and effectively improved the overall performance and generalization ability of the model.

[0109] In this embodiment, S4 specifically includes:

[0110] S41. Obtain historical EEG signal data and divide the historical EEG signal data into historical normal state EEG signal data and historical abnormal state EEG signal data according to the clinical diagnosis results.

[0111] S42. Calculate the reconstruction error values ​​of each historical normal state EEG signal data and historical abnormal state EEG signal data respectively, and obtain the reconstruction error set of historical normal state EEG signal data and the reconstruction error set of historical abnormal state EEG signal data.

[0112] S43. Perform statistical analysis on the set of reconstruction errors of EEG signal data under historical normal conditions, determine the probability distribution of reconstruction errors of EEG signal data under historical normal conditions, and obtain the probability density function of reconstruction errors of EEG signal data under historical normal conditions.

[0113] S44. Perform statistical analysis on the set of reconstruction errors of EEG signal data under historical abnormal conditions, determine the probability distribution of reconstruction errors of EEG signal data under historical abnormal conditions, and obtain the probability density function of reconstruction errors of EEG signal data under historical abnormal conditions.

[0114] S45. Based on the probability density function of the reconstruction error of historical normal EEG signal data and historical abnormal EEG signal data, determine the boundary point between the reconstruction error of normal EEG signal data and abnormal EEG signal data, and obtain the threshold range of reconstruction error for normal and abnormal EEG signal data.

[0115] S46. Based on the reconstruction error threshold range, construct a big data feature database of EEG signals and determine the reconstruction error threshold range of the EEG signal data to be detected.

[0116] By constructing a big data feature database of EEG signals, the reconstruction error values ​​of historical normal and abnormal EEG signal data are calculated respectively, and the reconstruction error probability density function of the two types of data is determined. This clarifies the threshold range of reconstruction error for normal and abnormal EEG signals, effectively improving the accuracy and stability of abnormal signal identification and overcoming the shortcomings of traditional methods, such as vague threshold settings and difficulty in accurately distinguishing abnormal signals.

[0117] In this embodiment, S6 specifically includes:

[0118] S61. Divide the EEG signal data to be detected into multiple time-domain subsequences, and input each time-domain subsequence into the trained SincVAE anomaly recognition model. Through the dual-stage Sinc filter module, adaptive multi-scale spatiotemporal fusion convolution module and dual-stream interactive self-attention latent variable encoding module inside the model, obtain the latent variable feature data corresponding to each subsequence.

[0119] S62. Decode the latent variable feature data of each subsequence to obtain the corresponding reconstructed subsequence data, and calculate the local reconstruction error value of each subsequence.

[0120]

[0121] in, This represents the local reconstruction error value of the j-th time-domain subsequence. This represents the actual EEG signal value at the m-th sampling point in the j-th time-domain subsequence. Let L represent the reconstructed EEG signal value at the m-th sampling point in the j-th time-domain subsequence, L represent the number of sampling points in a single time-domain subsequence, and δ be a minimal positive real constant, where 0 < δ ≤ 10. -6 ;

[0122] S63. Calculate the overall reconstruction error ε of the EEG signal data to be detected based on the local reconstruction error values ​​of all time-domain subsequences. global :

[0123]

[0124] Where, ε global The value represents the overall reconstruction error of the EEG signal data to be detected, J represents the total number of time-domain subsequences, and μ represents the total number of time-domain subsequences. n σ represents the mean of the local reconstruction error of historical normal EEG signal data. n This represents the standard deviation of the local reconstruction error of EEG signal data under historical normal conditions.

[0125] S64. Based on the EEG signal big data feature database, determine the reconstruction error threshold range for normal and abnormal EEG signal data, and set the overall reconstruction error value ε. global By comparing the data with the threshold range, it can be determined whether the EEG signal data to be detected belongs to a normal or abnormal state.

[0126] By innovating the local and global reconstruction error calculation methods, the EEG signal data to be detected is divided into time-domain subsequences. The reconstruction error of each local subsequence is calculated and then fused exponentially to obtain a more sensitive and refined global reconstruction error index. This effectively solves the problem that traditional global error calculation methods are insufficient in responding to local minor abnormal signals, and significantly improves the recognition accuracy of abnormal signals and the ability to locate local anomalies.

[0127] In this embodiment, an EEG signal anomaly identification system based on big data analysis includes the following modules:

[0128] The data acquisition and preprocessing module is used to acquire the EEG signals of the target subject and preprocess the EEG signals through filtering, noise reduction and baseline correction methods to obtain EEG signal data;

[0129] The SincVAE anomaly detection model is used for adaptive frequency band feature extraction, multi-scale spatiotemporal feature fusion, and latent variable feature encoding of EEG signal data.

[0130] The model optimization module is used to adaptively optimize the SincVAE anomaly recognition model using a multi-target squirrel search algorithm.

[0131] The reconstruction error threshold database construction module is used to build a big data feature database of EEG signals, divide historical normal and abnormal EEG signal data, and determine the reconstruction error threshold range of normal and abnormal EEG signal data.

[0132] The anomaly detection module is used to divide the EEG signal data to be detected into multiple time-domain subsequences, calculate the local reconstruction error value of each time-domain subsequence, and calculate the overall reconstruction error value. By comparing it with the reconstruction error threshold range, it determines whether the EEG signal data to be detected is abnormal.

[0133] High-quality EEG signals are acquired through a data acquisition and preprocessing module. Adaptive feature extraction and encoding are achieved using a SincVAE anomaly recognition model with a learnable Sinc filter. The model is optimized by combining a multi-target squirrel search algorithm and a reconstruction error threshold database is built to improve classification accuracy. Finally, the anomaly detection module achieves accurate identification of abnormal EEG signal states, effectively improving the real-time performance, accuracy, and robustness of abnormal signal recognition, and meeting the needs of clinical applications.

[0134] Example 1:

[0135] To verify the feasibility of this invention in practice, it was applied to a real-time EEG abnormality signal detection and diagnostic assistance system in a large tertiary hospital to improve the diagnostic accuracy and efficiency of diseases such as epilepsy, sleep disorders, and neurological lesions. In this clinical application scenario, traditional EEG signal detection and abnormality identification typically rely on clinicians' visual observation and automated software-assisted diagnostic tools based on traditional signal processing methods. However, the accuracy of manual interpretation of EEG abnormal signals is highly dependent on the doctor's clinical experience and is easily affected by doctor fatigue and subjective factors; the interpretation speed is also difficult to meet the real-time processing requirements of massive amounts of data. Traditional automated methods (such as Fourier transform and wavelet transform) require manual preset of frequency band parameters, which cannot adapt to the complex nonlinear changes in EEG signals, resulting in low abnormality detection accuracy and high misjudgment rate, making it difficult to effectively meet actual clinical needs.

[0136] To address the aforementioned issues, in practical applications, EEG signals from multiple patients under various clinical conditions were first acquired using a standard clinical EEG sensor, including normal state, epileptic seizure state, and abnormal sleep state. Each signal segment lasted approximately 10 minutes, with a sampling rate set to 256Hz. After signal acquisition, power frequency interference and baseline drift were removed using a digital bandpass filter (frequency range 0.5–50Hz), and random noise interference was eliminated using a wavelet threshold denoising algorithm to obtain clear EEG signal data for subsequent processing.

[0137] In the model construction process, this invention establishes a SincVAE anomaly recognition model with a learnable Sinc filter. The initial filtering stage of the dual-stage Sinc filter module uses a short-time sliding window (window length 2 seconds, step size 0.5 seconds) to segment the EEG signal data. It calculates the short-time energy ratio of the EEG signal within the sliding window in a preset frequency band (e.g., Delta, Theta, Alpha, Beta, Gamma bands). Based on maximizing the signal energy ratio within each short-time window, the filter frequency parameters are initially determined. The secondary adaptive adjustment stage further optimizes the filter frequency parameters again using a nonlinear iterative method, aiming to maximize the multi-dimensional spatiotemporal differences of the frequency band characteristics (50 iterations, dynamically adjusted step size). Finally, this dual-stage filtering method automatically generates the optimal combination of frequency band parameters, significantly improving the accuracy and stability of frequency band selection.

[0138] Subsequently, the adaptive multi-scale spatiotemporal fusion convolution module extracts spatiotemporal features of EEG signals at different scales through three parallel convolution channels of large scale, medium scale, and small scale, achieving comprehensive capture and fusion of overall trend features, local difference features, and fine-grained features; the dual-stream interactive self-attention latent variable encoding module further encodes the above features into high-dimensional latent variable representations, realizing deep expression and feature fusion of EEG signals.

[0139] In the model optimization stage, a multi-objective squirrel search algorithm was used for automatic parameter optimization. An initial squirrel population (30 squirrels) was established. The model classification accuracy, recall, model complexity, and real-time performance were used as multi-objective performance indicators. The optimal solution was selected by non-dominated ranking (Pareto front). After 40 rounds of nonlinear position parameter adjustment and optimization iterations, the algorithm automatically determined the optimal number of filters (16), latent variable dimension (32 dimensions), and number of training iterations (100 times), achieving efficient and accurate model optimization.

[0140] Next, we will construct a big data feature database of EEG signals. By statistically analyzing the reconstruction errors of 5,000 historical EEG normal state signal data and 2,000 historical EEG abnormal state signal data, we will calculate the reconstruction error probability density function of the two types of signals respectively, and determine the threshold range of reconstruction error for the two states (normal: 0 to 0.08, abnormal: above 0.09) to achieve more accurate classification of abnormal signals.

[0141] Finally, the 1000 EEG signal data to be detected were divided into multiple subsequences (each subsequence with 512 sampling points), and the local reconstruction error of each subsequence was calculated sequentially. The results were then fused and judged using an innovative overall reconstruction error calculation formula. Experimental results show that the method of this invention achieves an average classification accuracy of 97.8% for abnormal EEG signal identification, while the traditional method (based on a single Sinc filter and CNN model) only achieves an accuracy of 88.6%, representing a significant improvement of approximately 10%. To objectively demonstrate the technical advantages of this invention, some measured data comparing this invention with traditional techniques are listed below:

[0142] Table 1 Comparison of EEG signal anomaly identification accuracy results

[0143]

[0144]

[0145] As shown in Table 1, the method of the present invention achieves an accuracy of 98.2% in identifying normal state signals, which is significantly higher than the 89.5% of the traditional method; for identifying epileptic seizure signals, the accuracy of the present invention is 97.5%, an improvement of 11.5% compared to the traditional method; and in detecting sleep abnormal signals, the accuracy of the present invention is improved by 7.5%. Overall, the present invention significantly improves the accuracy and stability of EEG abnormality identification.

[0146] Furthermore, through clinical testing, the average processing time for a single EEG signal data point using this invention is approximately 0.15 seconds, significantly outperforming traditional methods (average processing time 0.7 seconds) in real-time processing capabilities, thus meeting the practical needs of efficient clinical diagnosis. Based on the detailed experimental data and practical application analysis described above, this invention demonstrates significant technical advantages in the field of EEG signal abnormality identification, making it suitable for large-scale clinical application and possessing considerable potential for widespread adoption.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying EEG signal anomalies based on big data analysis, characterized in that, Includes the following steps: S1. Collect the EEG signal of the target subject, and preprocess the EEG signal through filtering, noise reduction and baseline correction methods to obtain EEG signal data; S2. Construct a SincVAE anomaly recognition model with a learnable Sinc filter; S3. Adaptive optimization of the SincVAE anomaly recognition model is performed using a multi-target squirrel search algorithm to obtain an optimized SincVAE anomaly recognition model. S4. Construct a big data feature database for EEG signals. By conducting big data analysis on historical EEG signal data in normal and abnormal states, determine the reconstruction error threshold range of EEG signal data in normal and abnormal states. S5. Use normal EEG signal data to perform semi-supervised training on the optimized SincVAE anomaly recognition model to obtain the trained SincVAE anomaly recognition model. S6. Input the EEG signal data to be detected into the trained SincVAE anomaly recognition model. Process the EEG signal data to be detected, calculate the reconstruction error of the EEG signal data to be detected, and determine whether the EEG signal data to be detected is abnormal based on the reconstruction error threshold range. S7. Classify the identified abnormal EEG signal data according to the preset abnormal signal classification criteria, obtain the abnormal signal classification results, and output them to the clinical diagnostic system for early warning.

2. The EEG signal anomaly identification method based on big data analysis according to claim 1, characterized in that, The SincVAE anomaly detection model with a learnable Sinc filter includes a two-stage Sinc filter module, an adaptive multi-scale spatiotemporal fusion convolution module, and a two-stream interactive self-attention latent variable encoding module. The dual-stage Sinc filter module is used to perform adaptive frequency band feature extraction on the input EEG signal data, including an initial Sinc filtering stage and a secondary adaptive adjustment stage. The adaptive multi-scale spatiotemporal fusion convolution module is used to extract high-order spatiotemporal features from the output of the dual-stage Sinc filter module, including three parallel convolutional feature channels. The dual-stream interactive self-attention latent variable encoding module is used to perform latent variable encoding on the output of the adaptive multi-scale spatiotemporal fusion convolution module, and includes a first attention encoding stream and a second attention encoding stream.

3. The EEG signal anomaly identification method based on big data analysis according to claim 2, characterized in that, The dual-stage Sinc filter module specifically includes an initial Sinc filtering stage and a secondary adaptive adjustment stage: In the initial Sinc filtering stage, the input EEG signal data is divided into segments using a short-time sliding window. The short-time energy ratio of the EEG signal within each sliding window in each preset frequency band is calculated. The objective function is to maximize the energy ratio of the EEG signal within the short-time window. The low cutoff frequency and high cutoff frequency parameters of multiple parallel Sinc filters are iteratively adjusted to obtain preliminary frequency band characteristic data. In the secondary adaptive adjustment stage, the goal of the secondary adaptive adjustment is to maximize the multi-dimensional spatiotemporal differences of the initial frequency band feature data. The interaction information contribution between the temporal features and spatial electrode position features of the initial frequency band feature data is analyzed. A nonlinear iterative optimization method is used to sort the interaction intensity of the frequency band contribution. Based on the sorting of the interaction intensity of the frequency band contribution, the low cutoff frequency parameter and high cutoff frequency parameter of each filter in the Sinc bandpass filter bank are adjusted to output optimized frequency band feature data with spatiotemporal differences. The preliminary frequency band feature data and the optimized frequency band feature data are aligned with the index positions of the input EEG signal data and then output.

4. The EEG signal anomaly identification method based on big data analysis according to claim 2, characterized in that, The adaptive multi-scale spatiotemporal fusion convolutional module specifically includes large-scale convolutional feature channels, medium-scale convolutional feature channels, and small-scale convolutional feature channels: The large-scale convolutional feature channel uses a convolution kernel with a size larger than the length of a single time-domain period of the input optimized frequency band feature data to perform joint convolution operations on the input optimized frequency band feature data in the spatial domain electrode dimension and the time-domain signal dimension. Through a single convolution, it captures the spatiotemporal features of the overall trend of EEG signal data in the entire electrode space and multiple consecutive time-domain periods, and obtains global spatiotemporal feature data. The mesoscale convolutional feature channel uses a convolution kernel with a size equal to the length of a single temporal period of the input optimized frequency band feature data. It performs convolution operations on the input optimized frequency band feature data in the spatial and temporal dimensions in a multi-layer stacked convolution manner to obtain mesoscale local features with distinct patterns in the local region of the spatial electrode location and in the temporal period, respectively. The local spatiotemporal feature data is obtained by fusing the multi-scale local feature weighting method. The small-scale convolutional feature channel uses a convolution kernel with a size smaller than the length of a single time-domain period of the input optimized frequency band feature data. It performs convolution calculations on the spatial electrode position and time-domain signal of the input optimized frequency band feature data in a deep, step-by-step iterative convolution method. This extracts the interaction difference features between spatially adjacent electrodes and the subtle amplitude fluctuation pattern features within a shorter time period from the EEG signal data, thereby obtaining fine-grained spatiotemporal feature data. After aligning the index positions of the global spatiotemporal feature data, local spatiotemporal feature data, and fine-grained spatiotemporal feature data, multi-scale spatiotemporal feature data is obtained by multi-scale convolution fusion.

5. The EEG signal anomaly identification method based on big data analysis according to claim 2, characterized in that, The dual-stream interactive self-attention latent variable encoding module specifically includes a first attention encoding stream and a second attention encoding stream: The first attention encoding stream takes multi-scale spatiotemporal feature data as input and calculates the feature weights of EEG signals at different time positions in the time domain dimension through a self-attention mechanism to obtain temporal latent variable encoding with prominent features in the time dimension. The second attention coding stream takes multi-scale spatiotemporal feature data as input and calculates the feature weights of different spatial electrode positions of the EEG signal in the spatial domain electrode position dimension through a self-attention mechanism to obtain spatial latent variable codes with prominent features in the spatial dimension. The first attention encoding stream and the second attention encoding stream exchange feature weight parameters of temporal latent variable encoding and spatial latent variable encoding in real time through an interactive feedback mechanism, and use an iterative weighted update method to adjust the weight parameters in their respective self-attention mechanisms to generate optimized latent variable encoding feature data.

6. The EEG signal anomaly identification method based on big data analysis according to claim 1, characterized in that, S3 specifically includes: S31. Establish an initial squirrel population. The position of each squirrel individual is encoded in the form of a real number vector, which corresponds to the number of filters and frequency range of the two-stage Sinc filter module, the kernel size of the adaptive multi-scale spatiotemporal fusion convolution module, the dimension of the latent variables of the two-stream interactive self-attention latent variable encoding module, and the number of model training iterations. S32. Input the model parameter configuration corresponding to each individual squirrel into the SincVAE anomaly recognition model, and calculate the four performance indicators of the model output: classification accuracy, recall, model complexity, and real-time performance, to form a multi-objective evaluation objective. S33. Based on the multi-objective evaluation objectives, perform non-dominated ranking of the squirrel population, identify and retain elite squirrel individuals that meet the multi-objective evaluation requirements; S34. For non-elite squirrel individuals, based on the positional difference between them and elite squirrel individuals, update the positional parameters of non-elite squirrel individuals according to a non-linear adjustment method to form a new generation of squirrel population. S35. Re-input the location parameters of each squirrel individual in the updated new generation squirrel population into the SincVAE anomaly detection model, and re-evaluate the classification accuracy, recall, model complexity, and real-time performance. Update the multi-objective evaluation objective and perform non-dominated sorting again. S36. Repeatedly execute the steps of non-dominated sorting, elite individual selection and non-elite individual position parameter nonlinear adjustment until the preset maximum number of iterations is reached, and obtain the optimized SincVAE anomaly recognition model that satisfies the optimal multi-objective performance index.

7. The EEG signal anomaly identification method based on big data analysis according to claim 1, characterized in that, S4 specifically includes: S41. Obtain historical EEG signal data and divide the historical EEG signal data into historical normal state EEG signal data and historical abnormal state EEG signal data according to the clinical diagnosis results. S42. Calculate the reconstruction error values ​​of each historical normal state EEG signal data and historical abnormal state EEG signal data respectively, and obtain the reconstruction error set of historical normal state EEG signal data and the reconstruction error set of historical abnormal state EEG signal data. S43. Perform statistical analysis on the set of reconstruction errors of EEG signal data under historical normal conditions, determine the probability distribution of reconstruction errors of EEG signal data under historical normal conditions, and obtain the probability density function of reconstruction errors of EEG signal data under historical normal conditions. S44. Perform statistical analysis on the set of reconstruction errors of EEG signal data under historical abnormal conditions, determine the probability distribution of reconstruction errors of EEG signal data under historical abnormal conditions, and obtain the probability density function of reconstruction errors of EEG signal data under historical abnormal conditions. S45. Based on the probability density function of the reconstruction error of historical normal EEG signal data and historical abnormal EEG signal data, determine the boundary point between the reconstruction error of normal EEG signal data and abnormal EEG signal data, and obtain the threshold range of reconstruction error for normal and abnormal EEG signal data. S46. Based on the reconstruction error threshold range, construct a big data feature database of EEG signals and determine the reconstruction error threshold range of the EEG signal data to be detected.

8. The EEG signal anomaly identification method based on big data analysis according to claim 1, characterized in that, S6 specifically includes: S61. Divide the EEG signal data to be detected into multiple time-domain subsequences, and input each time-domain subsequence into the trained SincVAE anomaly recognition model. Through the dual-stage Sinc filter module, adaptive multi-scale spatiotemporal fusion convolution module and dual-stream interactive self-attention latent variable encoding module inside the model, obtain the latent variable feature data corresponding to each subsequence. S62. Decode the latent variable feature data of each subsequence to obtain the corresponding reconstructed subsequence data, and calculate the local reconstruction error value of each subsequence. S63. Calculate the overall reconstruction error ε of the EEG signal data to be detected based on the local reconstruction error values ​​of all time-domain subsequences. global ; S64. Based on the EEG signal big data feature database, determine the reconstruction error threshold range for normal and abnormal EEG signal data, and set the overall reconstruction error value ε. global By comparing the data with the threshold range, it can be determined whether the EEG signal data to be detected belongs to a normal or abnormal state.

9. A system for identifying EEG signal anomalies based on big data analysis, comprising executing the method for identifying EEG signal anomalies based on big data analysis as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to acquire the EEG signals of the target subject and preprocess the EEG signals through filtering, noise reduction and baseline correction methods to obtain EEG signal data; The SincVAE anomaly detection model is used for adaptive frequency band feature extraction, multi-scale spatiotemporal feature fusion, and latent variable feature encoding of EEG signal data. The model optimization module is used to adaptively optimize the SincVAE anomaly recognition model using a multi-target squirrel search algorithm. The reconstruction error threshold database construction module is used to build a big data feature database of EEG signals, divide historical normal and abnormal EEG signal data, and determine the reconstruction error threshold range of normal and abnormal EEG signal data. The anomaly detection module is used to divide the EEG signal data to be detected into multiple time-domain subsequences, calculate the local reconstruction error value of each time-domain subsequence, and calculate the overall reconstruction error value. By comparing it with the reconstruction error threshold range, it determines whether the EEG signal data to be detected is abnormal.

Images