Noise reduction method, system, equipment and medium

Automatically denoising EEG data noise is performed through the target noise reduction model, which solves the problem of manual intervention in the existing technology, and realizes efficient denoising of EEG data, improving the signal-to-noise ratio.

CN118916609APending Publication Date: 2024-11-08NAT UNIV OF DEFENSE TECH
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202410744097.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-11
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The prior art requires manual intervention in the process of denoising EEG data, resulting in low noise reduction efficiency and poor effect.

Method used

Using the target noise reduction model, including an encoder and decoder or generator and discriminator, automatically perform noise reduction processing by training noisy EEG data and noiseless data. The specific steps include extracting using the encoder or generator. Features and removes noise, uses a decoder or discriminator to reconstruct data, and achieves noise reduction without manual intervention.

Benefits of technology

It realizes efficient EEG data noise reduction without manual intervention, improves noise reduction efficiency and effect, can effectively remove EOG and EMG noise, and improves signal-to-noise ratio.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916609A_ABST
    Figure CN118916609A_ABST
Patent Text Reader

Abstract

The embodiment of the invention discloses a noise reduction method, system and device and a medium, and relates to the technical field of data processing, and the method comprises the following steps: obtaining electroencephalogram data to be subjected to noise reduction; the electroencephalogram data to be subjected to noise reduction are subjected to noise reduction through a target noise reduction model, target noise reduction data are obtained, the target noise reduction model comprises an encoder and a decoder, or the target noise reduction model comprises a generator and a discriminator, and the target noise reduction model is obtained through training according to the electroencephalogram data with noise and the electroencephalogram data without noise.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a noise reduction method, system, device and storage medium. Background Art

[0002] Data denoising is a core task in the data preprocessing process, which aims to reduce or eliminate the noise in the data to improve the accuracy and reliability of data analysis. Data denoising can be applied to many fields such as image processing, audio processing, communications, financial data analysis, medical image processing, meteorological data processing, drones, etc. For example, EEG data in the field of medical image processing is an important tool for neuroscience research, but various noises in reality pollute EEG data. In related technologies, methods such as empirical mode decomposition and independent component analysis are often used to achieve denoising of EEG data, but the denoising process still requires manual intervention, resulting in low denoising efficiency and poor denoising effect. Summary of the invention

[0003] In view of this, one of the purposes of the present application is to provide a noise reduction method, system, electronic device and computer-readable storage medium, wherein the entire process does not require human intervention and has high noise reduction efficiency and good effect.

[0004] To achieve the above purpose, the technical solution of this application is implemented as follows:

[0005] In a first aspect, an embodiment of the present application provides a noise reduction method, comprising:

[0006] Obtaining the EEG data to be denoised;

[0007] The target denoising model is used to denoise the EEG data to be denoised to obtain target denoised data, wherein the target denoising model includes an encoder and a decoder, or the target denoising model includes a generator and a discriminator, and the target denoising model is trained according to the noisy EEG data and the noise-free EEG data.

[0008] In a possible implementation, the target denoising model includes an encoder and a decoder, and the target denoising model is used to denoise the EEG data to be denoised to obtain target denoised data, including:

[0009] using the encoder to determine key features of the EEG data to be denoised, and the data to be denoised;

[0010] The decoder is used to reconstruct the data according to the key features of the EEG data to be denoised and the data from which the noise is removed to obtain the target denoised data.

[0011] In one possible implementation, the encoder includes at least one first convolution block, each of which includes a convolution layer with padding added, a pooling layer with padding added, and a first activation function, and the decoder includes at least one deconvolution block, each of which includes a deconvolution layer with padding added and a second activation function;

[0012] The target denoising model is used to denoise the denoised EEG data to obtain target denoised data, including:

[0013] Determine the key features of the EEG data to be denoised and the data to be denoised using the padded convolutional layer, the padded pooling layer and the first activation function;

[0014] By using the deconvolution layer with padding and the second activation function, data reconstruction is performed according to the key features of the EEG data to be denoised and the data to be denoised, and the target denoised data is obtained.

[0015] In a possible implementation, the first activation function and the second activation function are both ReLU functions, the number of convolution blocks and the number of deconvolution blocks are the same, and the number of convolution blocks and the number of deconvolution blocks include any one of 4, 5, and 6.

[0016] In a possible implementation, the target denoising model includes a generator and a discriminator, and the target denoising model is used to denoise the EEG data to be denoised to obtain target denoised data, including:

[0017] The generator is used to perform dimensionality reduction processing on the denoised EEG data to obtain intermediate denoised data;

[0018] The discriminator is used to determine the target denoised data from the intermediate denoised data.

[0019] In one possible implementation, the generator includes a bidirectional long short-term memory network layer and a unidirectional long short-term memory network layer, and the discriminator includes at least one second convolutional block and a fully connected layer;

[0020] The target denoising model is used to denoise the denoised EEG data to obtain target denoised data, including:

[0021] The EEG data to be denoised is sequentially input into a bidirectional long short-term memory network layer and a unidirectional long short-term memory network layer to obtain intermediate denoised data;

[0022] The intermediate denoised data is sequentially input into at least one second convolutional block, and the target denoised data is output through a fully connected layer.

[0023] In a possible implementation, each second convolution block includes a one-dimensional convolution layer, an activation function, and a one-dimensional pooling layer, and the number of second convolution blocks is 5.

[0024] In a second aspect, an embodiment of the present application provides a noise reduction system, the system comprising:

[0025] An acquisition module, used for acquiring EEG data to be denoised;

[0026] The denoising module is used to denoise the EEG data to be denoised using a target denoising model to obtain target denoised data, wherein the target denoising model includes an encoder and a decoder, or the target denoising model includes a generator and a discriminator, and the target denoising model is trained based on noisy EEG data and noise-free EEG data.

[0027] In a third aspect, an embodiment of the present application provides an electronic device, the electronic device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the noise reduction method provided in the first aspect is implemented.

[0028] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by one or more processors, the noise reduction method provided in the first aspect is implemented.

[0029] The embodiment of the present application provides a denoising method, which first obtains the EEG data to be denoised; then uses a target denoising model to denoise the EEG data to be denoised, to obtain target denoised data, wherein the target denoising model includes an encoder and a decoder, or the target denoising model includes a generator and a discriminator, and the target denoising model is trained based on noisy EEG data and noise-free EEG data. The embodiment of the present application achieves denoising of the EEG data to be denoised through the target denoising model, and the whole process does not require manual intervention, and the denoising efficiency is high and the effect is good. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. It should be understood that the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0031] Figure 1 A flowchart of a noise reduction method provided in an embodiment of the present application;

[0032] Figure 2 A schematic diagram of a noise reduction process included in a noise reduction method provided in an embodiment of the present application;

[0033] Figure 3 A model structure diagram of Novel AE involved in a noise reduction method provided in an embodiment of the present application;

[0034] Figure 4 A graph showing changes in average loss values ​​of a multi-model cross-validation method provided in an embodiment of the present application;

[0035] Figure 5 A rendering of the effect of removing EOG and EMG noise data involved in a noise reduction method provided in an embodiment of the present application;

[0036] Figure 6 A PSD diagram of EOG and EMG noise data removed according to a noise reduction method provided in an embodiment of the present application;

[0037] Figure 7 A diagram showing the effect of removing EOG noise data using a multi-model involved in a noise reduction method provided in an embodiment of the present application;

[0038] Figure 8 A diagram showing the effect of removing EMG noise data using a multi-model involved in a noise reduction method provided in an embodiment of the present application;

[0039] Fig. 9 A model structure diagram of Novel GAN ​​involved in a denoising method provided in an embodiment of the present application;

[0040] Fig.10 A graph showing changes in average loss values ​​of another multi-model cross-validation method provided in an embodiment of the present application;

[0041] Fig.11 Another effect diagram of removing EOG and EMG noise data involved in a noise reduction method provided in an embodiment of the present application;

[0042] Fig.12 Another PSD diagram of EOG and EMG noise data removed according to a noise reduction method provided in an embodiment of the present application;

[0043] Fig.13 This is another effect diagram of removing EOG noise data using a multi-model involved in a noise reduction method provided in an embodiment of the present application;

[0044] Fig.14 This is another effect diagram of removing EMG noise data using a multi-model involved in a noise reduction method provided in an embodiment of the present application;

[0045] Fig.15 A schematic diagram of functional modules of a noise reduction system provided in an embodiment of the present application;

[0046] Fig.16 A diagram of the internal structure of an electronic device provided in an embodiment of the present application.

[0047] Description of reference numerals:

[0048] Denoising system 1500 , acquisition module 1510 , denoising module 1520 . DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0050] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0051] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0052] In various embodiments of the present application, the expression "or" or "at least one of A or / and B" includes any combination or all combinations of the words listed at the same time. For example, the expression "A or B" or "at least one of A or / and B" may include A, may include B, or may include both A and B.

[0053] In the description of the present application, it should be noted that if the terms "upper", "lower", "inside", "outside", etc. appear to indicate an orientation or position relationship, it is based on the orientation or position relationship shown in the drawings, or is the orientation or position relationship in which the product of the invention is usually placed when used. It is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application.

[0054] In addition, the terms “first”, “second”, etc., if used, are merely used to distinguish between the descriptions and should not be understood as indicating or implying relative importance.

[0055] It should be noted that, in the absence of conflict, the features in the embodiments of the present application may be combined with each other.

[0056] In order to solve the technical problems in the background technology, the embodiments of the present application provide a noise reduction method, system, electronic device and computer readable storage medium. The noise reduction method provided by the embodiments of the present application is first introduced below.

[0057] See also Figure 1 , Figure 1 A flowchart of a noise reduction method provided in an embodiment of the present application, the noise reduction method can be applied to electronic devices, the electronic devices include personal computers, servers, mobile devices, cloud computing platforms and supercomputers, etc., the noise reduction method specifically includes the following steps:

[0058] Step 110, obtaining the EEG data to be denoised;

[0059] Step 120, denoising the EEG data to be denoised using a target denoising model to obtain target denoised data, wherein the target denoising model includes an encoder and a decoder, or the target denoising model includes a generator and a discriminator, and the target denoising model is trained based on the noisy EEG data and the noise-free EEG data.

[0060] The denoising method provided in the embodiment of the present application first obtains the EEG data to be denoised; then uses the target denoising model to denoise the EEG data to be denoised, to obtain the target denoised data, wherein the target denoising model includes an encoder and a decoder, or the target denoising model includes a generator and a discriminator, and the target denoising model is trained based on noisy EEG data and noise-free EEG data. The embodiment of the present application realizes denoising of the EEG data to be denoised through the target denoising model, and the whole process does not require manual intervention, and the denoising efficiency is high and the effect is good.

[0061] The following will be Figure 1 Each step of the method is described in detail.

[0062] In step 110, the EEG data to be denoised may be acquired by an electronic device.

[0063] Among them, electroencephalogram (EEG) data is an important tool for neuroscience research and clinical applications. However, in the process of recording EEG data, various physiological noises are often recorded at the same time, such as electrooculography, electromyography, electrocardiography, and non-physiological noises such as power frequency interference. Various noises pollute EEG data, cover up the time-frequency characteristics of EEG data, and bring great difficulties to subsequent research and application.

[0064] It should be noted that, in the process of collecting the original EEG data, i.e., the EEG data to be denoised, it usually contains electrooculogram (EOG) noise data and electromyogram (EMG) noise data.

[0065] Among them, if EEG data without noise data is called Clean EEG data, in order to obtain noisy EEG data containing EOG noise data or EMG noise data, it can be obtained by adding EOG noise data or EMG noise data to Clean EEG data, which can be recorded as Mixed EEG data.

[0066] In step 120, the target denoising model has two forms, one of which includes an encoder and a decoder, which can achieve denoising without losing the features of the EEG data to be denoised. The other includes a generator and a discriminator, which mainly finds the mapping relationship between the EEG data without noise and the EEG data with noise through the generator, and gradually removes the relevant noise in the iterative process of generating a signal close to the real signal.

[0067] Noisy EEG data may include EOG noise data and / or EMG noise data.

[0068] See also Figure 2 , Figure 2 A schematic diagram of a denoising process included in a denoising method provided in an embodiment of the present application, wherein the Block in the area where the encoder is located represents a convolution block, and the Block in the area where the decoder is located represents a deconvolution block. In a possible implementation, the target denoising model includes an encoder and a decoder, and the denoising of the denoised EEG data is performed using the target denoising model to obtain target denoised data, including:

[0069] using the encoder to determine key features of the EEG data to be denoised, and the data to be denoised;

[0070] The decoder is used to reconstruct the data according to the key features of the EEG data to be denoised and the data from which the noise is removed to obtain the target denoised data.

[0071] Specifically, the EEG data to be denoised can be input into the encoder, and the EEG data to be denoised is essentially the Mixed EEG data in the above embodiment. The encoder will extract the key features in the Mixed EEG data and remove the noise data at the same time. The decoder will reconstruct the data according to the key features of the EEG data to be denoised and the data to remove the noise, and obtain the EEG data with the noise removed, that is, the target noise reduction data, which can also be called Denoised EEG data.

[0072] In this embodiment, the following embodiments, and all drawings, the target noise reduction model including the encoder and the decoder can be represented by Novel AE.

[0073] The encoder and decoder of the target denoising model are composed of a similar symmetrical structure formed by stacking multiple convolutional structures (blocks or convolutional blocks) or deconvolutional structures. Each convolutional block also contains a pooling layer and an activation function. In the experiment, it was found that the model performance was not ideal when the decoder used a structure composed of upper pooling, deconvolutional layers and activation functions.

[0074] Considering that most encoder models composed of convolutional layers and deconvolutional layers process planar bitmap data with multiple dimensions and multiple spatial information, the features of the downsampled data are more concentrated. However, the EEG data to be denoised is a time series signal, and the EEG data to be denoised in a single channel does not have complex spatial features. After downsampling, the number of features of the single-channel EEG data to be denoised is drastically reduced. Even if the data is restored with upsampling, it is impossible to make up for the useful information that has been lost in the single-channel EEG data to be denoised.

[0075] In one possible implementation, the encoder includes at least one first convolution block, each of which includes a convolution layer with padding added, a pooling layer with padding added, and a first activation function, and the decoder includes at least one deconvolution block, each of which includes a deconvolution layer with padding added and a second activation function;

[0076] The target denoising model is used to denoise the denoised EEG data to obtain target denoised data, including:

[0077] Determine the key features of the EEG data to be denoised and the data to be denoised using the padded convolutional layer, the padded pooling layer and the first activation function;

[0078] By using the deconvolution layer with padding and the second activation function, data reconstruction is performed according to the key features of the EEG data to be denoised and the data to be denoised, and the target denoised data is obtained.

[0079] In order to ensure that the features of the single-channel EEG data to be denoised are not lost, in some embodiments, padding can be added to both the convolutional layer and the pooling layer in the encoder, and the upper pooling layer is removed. At the same time, considering the efficiency of the target denoising model in the training stage, in order to accelerate the training of the target denoising model and enhance the nonlinear ability of the target denoising model, a Relu activation function is added except for the last output layer.

[0080] During the entire model training, the sample size of the EEG data is kept unchanged while the maximum pooling layer in the encoder is used to separate sparse features, thereby increasing the proportion of useful features. The decoder uses multiple deconvolution layers to restore the signal. It obtains multiple outputs for the input single-channel signal and uses the corresponding hierarchical cascade structure to capture detailed information while also improving the problem of sparse feature map expansion information. When used in conjunction with the convolution layer, it can reconstruct the input EEG data well.

[0081] In order to prevent the target denoising model from having too many layers leading to overfitting and too few layers leading to underfitting. After multiple experiments and verifications, in one possible implementation, the first activation function and the second activation function are both ReLU functions, the number of convolution blocks and deconvolution blocks is the same, and the number of convolution blocks and deconvolution blocks includes any one of 4, 5, and 6.

[0082] Among them, when the number of convolution blocks or deconvolution blocks of the encoder and decoder is equal to 5, the denoising effect is the best. When the number of convolution blocks or deconvolution blocks is equal to 4, the training speed is the fastest, but underfitting will occur. When the number of convolution blocks or deconvolution blocks is equal to 6, the fitting effect is basically the same as when the number is equal to 5, but the training time cost of the target denoising model is several times that when the number is equal to 5.

[0083] See also Figure 3 , Figure 3 A model structure diagram of Novel AE involved in a denoising method provided in an embodiment of the present application. In some embodiments, the encoder includes 5 similar or identical convolution blocks, and the decoder includes 5 similar or identical deconvolution blocks. Each convolution block in the encoder includes a one-dimensional convolution layer, a Relu activation function, and a one-dimensional maximum pooling layer. The convolution kernel and pooling kernel of the convolution layer and pooling layer of each convolution block are (3, 5, 7, 9, 11), respectively, and padding is added. The step size of all pooling layers is 1. In order to better extract the features of the EEG data to be denoised, the feature channels of each convolution block are increased from 32 to 256 (1-32-32-64-128-256). The five deconvolution blocks of the decoder are composed of deconvolution layers and ReLU layers, with padding added. The last layer has no activation function. The deconvolution kernels of each layer are (11, 9, 7, 5, 3), and the feature channel symmetric encoder is reduced from 256 to 32 (256-128-64-32-32-1). Finally, the target denoised data after noise removal is output. The target denoised data can also be called Denoised EEG data.

[0084] In some embodiments, training a target noise reduction model based on noisy EEG data and noise-free EEG data includes:

[0085] The original EEG data, EOG data and EMG data can be downloaded from the publicly available database. The type, size, signal component and sampling frequency of the downloaded data are different, and they need to be preprocessed before they can be used. In some embodiments, the downloaded original EEG data, EOG data and EMG data need to be preprocessed first, and then cut into equal length segments after a series of filtering and denoising to ensure that the data is clean and usable. The data length obtained after cutting is set to 2 seconds, which is enough to restore the spectral characteristics of the EEG signal.

[0086] In addition, during the EEG experimental data collection process, it is difficult to obtain noise-free EEG segments of more than 2 seconds due to the subjects' random blinking or myoelectric noise.

[0087] In some specific embodiments, the downloaded EEG data may form an EEG data set including 52 participants, and 64 channels of EEG data are recorded simultaneously at a sampling frequency of 512 Hz. The raw data is bandpass filtered in the range of 1-80 Hz, and notch filtered to remove power frequency interference. The clean EEG data is used as the ground truth, and the EEG data has been processed using the ICA method for 64 channels of data, and finally cut into 2-second single-channel segments.

[0088] For the preparation of EOG noise data, additional EOG channel data from multiple public datasets can be used, the original signals of the datasets are band-pass filtered at 0.3-10 Hz, power frequency interference is removed, and then segmented into 2-second single-channel EOG segments.

[0089] The EMG data segment selected is the facial muscle dataset, which is one of the main sources of myogenic noise, with a sampling rate of 2048 Hz. The original electromyographic signal is band-pass filtered at 1-80 Hz and the power frequency interference is removed, and then it is cut into 2-second single-channel EMG data segments.

[0090] In order to facilitate unified training and processing, all raw data can be downsampled / upsampled to 256Hz, so that each segment is 2 seconds long and has 512 sample points, and then standardized. Then it is saved as a Python Numpy matrix file, and finally 4514 EEG segments, 3400 EOG segments, and 5598 EMG segments are obtained.

[0091] In the actual data collection process, the signal-to-noise ratio (SNR) of the EEG signal is related to the noise content. The smaller the noise, the better the signal quality and the larger the signal-to-noise ratio. In one embodiment, the Mixed EEG data containing noise data of different levels (-7dB to 2dB) can be synthesized by formula (1). In formula (1), y is the contaminated EEG data (Mixed EEG), x is the original clean EEG data (Clean EEG), n is the added EOG noise data or EMG noise data, and λ is a related parameter for controlling the signal-to-noise ratio of the contaminated EEG signal y.

[0092] The signal-to-noise ratio can be controlled by formula (2). The larger the λ, the smaller the signal-to-noise ratio, which means that the EEG is more seriously contaminated by noise. In formula (2), SNR is the signal-to-noise ratio of the signal, RMS (Root Mean Square) is the root mean square, which is defined as shown in formula (3), and N is the sample length.

[0093] y=x+λ·n Formula (1)

[0094]

[0095] For the generation of training and test sets for training models, since there are only 3400 EOG data segments, in order to facilitate calculation and training, 3000 segments can be randomly selected from the EEG data set formed by EEG data and the EMG data set formed by EMG data for training, and 400 segments can be randomly selected for testing. And the synthesis is performed according to formula (1). After randomly mixing EOG or EMG noise, 10 levels of noise data (-7dB, -6dB, -5dB...1dB, 2dB) are generated, and the data sets generated by five different random selections will be used for random cross-validation. Finally, 30,000 training segments of 10 different levels of noise data containing EOG or EMG noise and 4000 test segments are synthesized. The overall training set size is 60,000 single-channel samples, and the test set size is 8,000 single-channel samples.

[0096] After obtaining the training set and the test set through the above embodiment, the collected original EEG data is simulated with Mixed EEG data, and the Clean EEG data is used as the noise-free data that needs to be fitted. The Mixed EEG data is input into the deep learning model, and the model continuously optimizes the fitting so that the output denoised EEG data approaches the Clean EEG data.

[0097] To facilitate training, we can standardize the input Mixed EEG (y) signal and Clean EEG (x) signal by dividing them by the standard deviation according to formula (4). In formula (4), σ y is the standard deviation of the corresponding Mixed EEG segment y.

[0098] In addition, the choice of the loss function of the autoencoder model is very important. In the end, our goal is to make the output of the model completely equivalent to the Clean EEG data, and then we get the Novel AE in the above embodiment. Therefore, the correlation between DenoisedEEG data and Clean EEG data is the most important evaluation indicator of the performance of the model. In some embodiments, the mean square error (MSE) can be used as the loss function such as formula (5). The smaller the MSE, that is, the smaller the Loss value, the better the model effect. The learning method can use gradient descent to maximize the correlation between Mixed EEG data and Clean EEG data.

[0099] In formula (5), N represents the number of samples in each data set (training or testing), and is the i-th sample point of each sample after normalization, represents the output of the network, is the corresponding Clean EEG data.

[0100]

[0101] In addition, Simple CNN, Complex CNN, Novel CNN, and DeepSeparator models can be selected as benchmark models to compare the denoising effect of the target denoising model Novel AE obtained in the above embodiments of the present application, where:

[0102] Simple CNN mainly consists of four convolutional layers and one fully connected layer. Each convolutional layer has a batch normalization layer and a Relu activation function.

[0103] Complex CNN adds a cross-layer connected residual block between two sequentially connected blocks. Secondly, in order to extract multi-scale features, Compelx CNN repeatedly stacks residual blocks twice, uses parallel multi-scale convolution kernels twice, and arranges six groups of three residual block branches in parallel. The output results of the last three layers of parallel convolution are concatenated and input into the last convolution layer and the fully connected layer to restore the normal dimension of the data.

[0104] Novel CNN consists of 7 double-layer convolutional layers and a fully connected layer. The convolutional layers of the first 6 blocks all contain Relu activation functions, and finally an average pooling layer is connected at the end of the block. In the seventh block, the pooling layer is replaced by a flattening layer. The 7 blocks make the data size constantly change, and finally the data is restored to its dimension through the flattening layer and the fully connected layer.

[0105] The DeepSeparator model consists of multiple identical stacked initial blocks to form the encoder and decoder, as well as the noise separator, which are all composed of four initial blocks, and the output of the first block and the input of the fourth block are connected across layers. The initial block consists of four convolutional layers with different convolution kernels to capture features of different scales, and then the outputs are merged and concatenated, and then filtered by the activation function. In short, we reproduced the above four models and tried to set them to the same basic model training parameters as the Novel AE model, the same optimizer, learning rate and batch size (BatchSize).

[0106] The model trained in the embodiment of the present application can use the Pytorch deep learning framework. Each model can be trained on a single Nvidia 3060Ti GPU purchased in the laboratory. The batch size is set to 200, the learning rate is 0.0001, and the number of training rounds (Epoch) is 50. Specific network performance evaluation indicators can use the correlation coefficient CC, power spectral density PSD, and relative root mean square error RRMSE in the time-frequency domain T , RRMSE S .

[0107] See also Figure 4 , Figure 4 A graph showing changes in average loss values ​​of a multi-model cross-validation method provided in an embodiment of the present application. Figure 4 (a) is the Loss diagram of several models with poor fitting ability. Figure 4 (b) is the Loss diagram of the model with better fitting ability. Figure 4 (a) and Figure 4 In (b), the horizontal axis uses Epoch to represent the number of training rounds, and the vertical axis uses MSE to represent the mean square error.

[0108] Under the premise that all five models have certain feature extraction capabilities, the training loss functions of DeepSeparator and Novel AE show a downward trend, and the training loss reaches the minimum at about epoch = 40. Simple CNN and Complex CNN have serious overfitting, while the test loss value of Novel CNN remains high, performing the worst among the five models. This shows that on this data set, compared with the other four models, Novel AE has the best performance in both the training set and the test set, proving that the Novel AE model in the embodiment of the present application has better generalization ability.

[0109] See also Figure 5 , Figure 5 This is a diagram showing the effect of removing EOG and EMG noise data involved in a noise reduction method provided in an embodiment of the present application. Figure 5 (a) Figure 5 (b) Figure 5 (c) Figure 5 The horizontal axis of (d) represents time in seconds, and the vertical axis represents amplitude in microvolts.

[0110] It can be found that when the EEG data distortion is the largest (SNR=-7dB), the specific effect of removing the EOG noise data and the EMG noise data on a certain sample is achieved. Figure 5 The green curve is the original Clean EEG data, the blue curve is the Mixed EEG data after adding noise, and the orange curve is the Denoised EEG data after denoising. Regardless of whether it is EOG noise data or EMG noise data, the Denoised EEG signals of the two models can fit the Clean EEG data well. Taking the EOG noise data as an example, in Figure 5 In (b), the fitting effect of the yellow-green curve is better than Figure 5 (a) The fitting effect is better and the overlap is higher, that is, NovelAE is better than the DeepSeparator model in restoring the details of EEG data, which also means that the Novel AE model in the embodiment of the present application has better noise reduction performance and stronger robustness.

[0111] See also Figure 6 , Figure 6A PSD graph of EOG and EMG noise data removed according to a noise reduction method provided in an embodiment of the present application. The power spectral density (PSD) graphs of the DeepSeparator model and the Novel AE model at different frequencies on the corresponding samples are calculated when the signal-to-noise ratio is -7dB. The first four of the eight sub-graphs are the effects of removing EOG noise data, and the last four are the effects of removing EMG noise data. Each of the four sub-graphs is a PSD graph of CleanEEG data before and after adding different noises, a PSD graph of Mixed EEG data, and a spectral feature restoration graph of noise removed by the DeepSeparator model and the Novel AE model. Different color areas are used to represent the main signal components of different frequency bands, red for delta waves (0.1-4Hz), orange for theta waves (4-8Hz), yellow for alpha waves (8-13Hz), green and purple for beta waves (13-30Hz) and gamma waves (30-80Hz), respectively.

[0112] Compare the effects of the two models in removing EOG noise data and EMG noise data. First, take the removal of EOG noise as an example: Figure 6 (a) is the power spectrum of Clean EEG data. After adding EOG noise, Figure 6 (b) It can be seen that there is an obvious high-power delta wave component in the low-frequency band. Figure 6 (c) and Figure 6 In (d), we can see that the energy amplitude of the two sets of DenosiedEEG data is greatly reduced, that is, the low-frequency, high-power delta band components in the EOG signal are effectively removed. The Novel AE model is better at removing electrooculogram noise than the DeepSeparator model, and the restored power spectrum is closer to the original power spectrum. However, DeepSeparator still has a set of obvious EOG noise data near 4Hz that has not been removed, and its denoising ability is relatively weak, which is also the reason why its Denoised EEG data deviates more from the Clean EEG data in the time domain.

[0113] exist Figure 6 In (f), it can be found that the added EMG noise data has wide-band and high-frequency characteristics, widely distributed in the range of 15 Hz to 80 Hz. Figure 6 (g) and Figure 6In (h), we can see the signals after the two models have removed noise. The low-frequency noise around 1Hz and the EMG noise data above 15Hz are effectively removed. The high-frequency noise in the β and γ bands is well suppressed in the restored signal. The overall denoising effect of the two models is similar, but the denoised data of the Novel AE model is closer to the original clean data Clean EEG data than the denoised data of the DeepSeparator model.

[0114] Please attend Figure 7 and Figure 8 ,in, Figure 7 This is an effect diagram of a multi-model for removing EOG noise data involved in a noise reduction method provided in an embodiment of the present application. Figure 8 This is a diagram showing the effect of removing EMG noise data using a multi-model involved in a noise reduction method provided in an embodiment of the present application.

[0115] exist Figure 7 and Figure 8 In the figure, the specific performance of the Novel AE model and the four compared models in removing two types of noise at different signal-to-noise ratios on the test set after five random cross-validations was quantitatively analyzed. The sub-figures are the average time domain relative root mean square error, the average frequency domain relative root mean square error, and the average correlation coefficient of all samples at each signal-to-noise ratio between Denoised EEG data and Clean EEG data, respectively.

[0116] Figure 7 and Figure 8 The green curves of each sub-image in the figure are the denoising effects of the Novel AE model. Performance comparison shows that among the five networks, our model has significantly better denoising ability when the signal-to-noise ratio is low, with higher CC and RRMSE. T and RRMSE S When the signal-to-noise ratio is -7dB, the correlation coefficient of the Novel AE model for removing EOG noise is 0.815, which is at least 0.12 higher than other models. When removing EMG noise, the correlation coefficient is 0.64, which is about 0.08 higher than DeepSeparator. T and RRMSE S , the Novel AE model also has the smallest error at each SNR level, especially in removing EOG noise. EMG noise is relatively difficult to remove in EEG due to its higher frequency range and greater noise interference. In general, among the five models, the Novel AE model has the best overall performance, followed by DeepSeparator, while Novel CNN performs the worst in all three indicators.

[0117] In addition, the average denoising performance indicators of all samples in the test set were further calculated, and the results are shown in Tables 1 and 2. Table 1 is the average performance table of the five models for removing EOG noise data when SNR = [-7dB ~ 2dB], and Table 2 is the average performance table of the five models for removing EMG noise data when SNR = [-7dB ~ 2dB].

[0118] Table 1

[0119]

[0120] Table 2

[0121]

[0122] It can be seen that Novel AE has the best performance in removing EOG noise, with various indicators including CC=0.889, RRMSE T =0.467, RRMSE S =0.481. In terms of removing EMG noise, the indicators of Novel AE are CC=0.827, RRMSE T =0.554, RRMSE S =0.510, that is, the deviation between the data after our model removes noise and Clean EEG is the smallest, and the prediction accuracy is the highest. These results show that the autoencoder that combines convolutional layers and deconvolutional layers can better reconstruct EEG data and effectively restore the original details of the EEG signal.

[0123] In a possible implementation, the target denoising model includes a generator and a discriminator, and the target denoising model is used to denoise the EEG data to be denoised to obtain target denoised data, including:

[0124] The generator is used to perform dimensionality reduction processing on the denoised EEG data to obtain intermediate denoised data;

[0125] The discriminator is used to determine the target denoised data from the intermediate denoised data.

[0126] This embodiment uses a target denoising model including a generator and a discriminator to perform denoising on the EEG data to be denoised, and does not require labeled clean data for training, which enables the target denoising model to achieve data denoising even without a large amount of labeled data. In addition, the generator and the discriminator can continuously improve their performance through adversarial training and gradually improve the effect of denoising. This adversarial training can help the target denoising model adapt to various types of noise and data distribution.

[0127] In this embodiment, the following embodiments, and all drawings, the target denoising model including the generator and the discriminator can be represented by Novel GAN.

[0128] Considering that the EEG data to be denoised has obvious non-stationary changes, but still has continuity and periodicity in the time domain. In a possible implementation, the generator includes a bidirectional long short-term memory network layer and a unidirectional long short-term memory network layer, and the discriminator includes at least one second convolutional block and a fully connected layer;

[0129] The target denoising model is used to denoise the denoised EEG data to obtain target denoised data, including:

[0130] The EEG data to be denoised is sequentially input into a bidirectional long short-term memory network layer and a unidirectional long short-term memory network layer to obtain intermediate denoised data;

[0131] The intermediate denoised data is sequentially input into at least one second convolutional block, and the target denoised data is output through a fully connected layer.

[0132] Among them, the bidirectional long short-term memory network (BiLSTM) layer and the unidirectional long short-term memory network (LSTM) layer can both be used to process time series information.

[0133] The generator of this embodiment includes a bidirectional long short-term memory network layer and a unidirectional long short-term memory network layer, which can improve the global expression ability of the generator network and the discrimination ability of nonlinear systems. The discriminator in this embodiment can also include multiple second convolution blocks to further improve the confrontation ability of the discriminator.

[0134] See also Fig. 9 , Fig. 9 A model structure diagram of Novel GAN ​​involved in a denoising method provided in an embodiment of the present application, wherein the generator of Novel GAN ​​consists of a BiLSTM layer and a LSTM layer, wherein each network layer contains 40 hidden units, the hidden layer initialization parameter is 0, and is finally output after dimension reduction by the fully connected layer, and the length of the EEG data sequence to be denoised does not change after input into Novel GAN. In one possible implementation, each second convolution block includes a one-dimensional convolution layer, an activation function, and a one-dimensional pooling layer. In order to improve the adversarial ability of the discriminator, the number of second convolution blocks can be set to 5. Among them, the convolution kernels of the convolution layer are all 3. In the last layer of the discriminator, a fully connected layer and a Tanh activation function can be added to identify the EEG data.

[0135] Like most GAN networks, the least squares loss function (LS) can be used to generate the distance between the data and the decision boundary. The gradient update of G and D can be implemented by the Adam optimizer, which can alleviate the problem of gradient disappearance in the GAN network. In formula (6) and formula (7), V LSAAN (D) is the loss value for judging whether a label is true or false, V LSGAN (G) is the loss value between the denoised data and the true value. Parameters a and b represent Clean EEG (true) and generated Denoised EEG (false), respectively; c is the hyperparameter that G hopes D will recognize the generated samples as real samples.

[0136] In the training of the Nocel GAN ​​network, due to the competitive relationship between V(D) and V(G), their loss values ​​will fluctuate as the training iterations continue to deepen. If their loss values ​​are basically unchanged, it means that the performance of the network has reached a local saddle point and the training should be stopped. In addition, in order to ensure a sufficient number of training rounds so that the model can be trained to reach a suitable balance point, the number of training rounds can be the same as the number of Novel AE training rounds in the above embodiment.

[0137]

[0138] In some embodiments, the activation function in each second convolutional block includes a Relu activation function.

[0139] To demonstrate the denoising effect of the Novel GAN ​​model in the above example, see Fig.10 , Fig.10 Another graph of changes in average loss values ​​of multi-model cross-validation involved in a denoising method provided in an embodiment of the present application.

[0140] exist Fig.10(a) shows the training process of the Raw GAN model in the related art and our improved Novel GAN ​​model. Among them, Raw G Loss and Raw D Loss represent the training loss curve of the Raw GAN model, and Novel GLoss and Novel D Loss represent the training process of the Novel GAN ​​model. It can be clearly seen that compared with the Raw GAN model, the G Loss and D Loss of the Novel GAN ​​model fluctuate less, which can explain that the NovelGAN model in the embodiment of the present application rarely has the problem of mode collapse. However, it is undeniable that the GAN model itself is prone to fall into a local optimal solution during repeated training. When the number of training rounds (epoch) is not enough, the effect of the generator G is usually not very ideal. When the epoch is greater than 20, the training effect will be better. The general trend is that the model gradually deteriorates after achieving a good effect, and then achieves a better effect after several epochs. The above repeated process well reflects the process of the game between the generator and the discriminator.

[0141] During the training of the GAN model, when D Loss approaches 0.5, the best result is that it is impossible to distinguish true and false signals, but it also means that the model stops updating. Compared with the Novel GAN ​​model, the D Loss value of the Raw GAN model reaches 0.5 early, but the actual discrimination ability of the discriminator has not improved, resulting in a decrease in the model's fitting ability. For the GAN model, G and D have a long enough healthy competition iteration process, and maintain a delicate balance during the competition, so that the learning ability of the two will continue to improve.

[0142] Fig.10 In (b), only the Novel AE model and DeepSeparator model that performed better on Dataset 1 are given, and the training loss value curve on Dataset 2 is shown. Overall, whether it is training loss or test loss value, the Novel AE model has better fitting performance than the DeepSeparator model.

[0143] See also Fig.11 , Fig.11 Another effect diagram of removing EOG and EMG noise data involved in a denoising method provided in an embodiment of the present application. Among them, due to the poor actual performance of the Raw GAN model, only the denoising effects of the Novel GAN ​​model and the Novel AE model with better performance are shown in the figure. Fig.11 (a) and Fig.11 (b) The effect of removing EOG noise data of Novel AE and NovelGAN respectively. Fig.11 (c) and Fig.11(d) is the effect diagram of removing EMG noise data.

[0144] It can be found that whether it is removing EOG noise data or removing EMG noise data, the green curve Denoised EEG data of the Novel GAN ​​model is closer to the original yellow curve Clean EEG data than the Novel AE model, and the overlap between the two is higher. Especially in removing EOG noise data, compared with EOG noise data, EMG noise data has a larger amplitude and a wider frequency domain, and the denoising effect is worse than that of removing EOG noise data.

[0145] See also Fig.12 , Fig.12 This is another PSD diagram of removing EOG and EMG noise data involved in a noise reduction method provided in an embodiment of the present application. It can be found that compared with Figure 6 (b) and Figure 6 (f) Fig.12 (b) and Fig.12 The noise signal in (f) has greater energy. Fig.12 In (b), due to the addition of low-frequency EOG noise data, the energy index of the delta wave is too high, and the power spectrum characteristics of the EEG data are almost invisible. Fig.12 (a), (b), (c), and (d) show that among the Denoised EEG data generated by the two models, the signal generated by the Novel GAN ​​model is closer to the Clean EEG data, while the Novel AE model cannot restore the frequency domain characteristics of the Clean EEG data well. Fig.12 This can also be found in (h). This may be because the BiLSTM layer in the Novel GAN ​​network can better learn the periodic characteristics of the EEG timing signal and can effectively remove the noise signal, so it has a better restoration effect on the original Clean EEG signal.

[0146] This application also gives the Novel AE, Novel GAN, DeepSeparator and Raw GAN models, and the average performance of three evaluation indicators of multiple random cross-validation to remove EOG noise data and EMG noise data on different signal-to-noise ratios of the test set of Dataset2, such as Fig.13 and Fig.14 As shown. Fig.13 This is another effect diagram of removing EOG noise data using a multi-model involved in a noise reduction method provided in an embodiment of the present application. Fig.14 This is a diagram showing the effect of removing EMG noise data using another multi-model involved in a noise reduction method provided in an embodiment of the present application.

[0147] It can be found that when the noise level is higher, the overall performance of the Novel GAN ​​model shows the best level, followed by the Novel AE model. When SNR = -14dB, the average correlation coefficient of the Novel GAN ​​model after removing the EOG noise data is 0.705, and the average correlation coefficient after removing the EMG noise data is 0.495.

[0148] In Tables 3 and 4, several indicators of each model on the test set are further averaged. Table 3 is the average performance table of EOG noise data removed by SNR = [-14dB ~ 4dB], and Table 4 is the average performance table of EMG noise data removed by SNR = [-14dB ~ 4dB].

[0149] It can be found that when removing EOG noise data, the correlation coefficients of the Novel GAN ​​model and the Novel AE model are 0.846 and 0.809 respectively. When removing EMG noise data, the correlation coefficients of the two models are 0.745 and 0.671 respectively. In terms of the relative root mean square error in the time-frequency domain, the Novel GAN ​​model also has the smallest deviation, followed by the Novel AE model.

[0150] Although the Novel GAN ​​model performs well in all three indicators, especially when the noise component is large, such as when SNR = -14dB, it is significantly better than other networks in all three indicators. However, its performance is not satisfactory when the noise component is small. Fig.13 and Fig.14 In the figure, when SNR=4dB, the correlation coefficient of the Novel GAN ​​model is lower than that of the two autoencoder models. The reason is that compared with the CNN network, BiLSTM and LSTM can learn more time-frequency differences between EEG data and noise data due to their unique advantages in processing continuous time domain signals, so even when the noise component is large, the EEG data can be restored well. However, its overly strong denoising ability will remove part of the CleanEEG data as noise data when there are fewer noise components in the signal. The convolutional coding characteristics of the CNN network, as well as the characteristics of parameter sharing and sparse connection, enable it to better retain the characteristics of Clean EEG data at different noise levels, but when the noise is too large, its characteristics of noise data are also retained, resulting in a decrease in the denoising performance of the model. In general, the two models proposed in the embodiments of the present application: Novel AE and Novel GAN ​​both have good comprehensive noise reduction performance.

[0151] Table 3

[0152]

[0153] Table 4

[0154]

[0155] Corresponding to the above method embodiment, the present application embodiment also provides a noise reduction system, see Fig.15 , Fig.15 A functional module diagram of a noise reduction system provided in an embodiment of the present application, wherein the noise reduction system 1500 includes:

[0156] An acquisition module 1510 is used to acquire EEG data to be denoised;

[0157] The denoising module 1520 is used to denoise the EEG data to be denoised using a target denoising model to obtain target denoised data, wherein the target denoising model includes an encoder and a decoder, or the target denoising model includes a generator and a discriminator, and the target denoising model is trained based on noisy EEG data and noise-free EEG data.

[0158] The noise reduction system provided in the embodiment of the present application can achieve the following Figure 1 The various processes implemented in the method embodiments can achieve similar or identical technical effects, and will not be described again here to avoid repetition.

[0159] In a possible implementation, the target noise reduction model includes an encoder and a decoder, and the noise reduction module 1520 is further configured to:

[0160] using the encoder to determine key features of the EEG data to be denoised, and the data to be denoised;

[0161] The decoder is used to reconstruct the data according to the key features of the EEG data to be denoised and the data from which the noise is removed to obtain the target denoised data.

[0162] In one possible implementation, the encoder includes at least one first convolution block, each of which includes a convolution layer with padding added, a pooling layer with padding added, and a first activation function, and the decoder includes at least one deconvolution block, each of which includes a deconvolution layer with padding added and a second activation function;

[0163] The noise reduction module 1520 is further used for:

[0164] Determine the key features of the EEG data to be denoised and the data to be denoised using the padded convolutional layer, the padded pooling layer and the first activation function;

[0165] By using the deconvolution layer with padding and the second activation function, data reconstruction is performed according to the key features of the EEG data to be denoised and the data to be denoised, and the target denoised data is obtained.

[0166] In a possible implementation, the first activation function and the second activation function are both ReLU functions, the number of convolution blocks and the number of deconvolution blocks are the same, and the number of convolution blocks and the number of deconvolution blocks include any one of 4, 5, and 6.

[0167] In a possible implementation, the target denoising model includes a generator and a discriminator, and the denoising module 1520 is further used to:

[0168] The generator is used to perform dimensionality reduction processing on the denoised EEG data to obtain intermediate denoised data;

[0169] The discriminator is used to determine the target denoised data from the intermediate denoised data.

[0170] In one possible implementation, the generator includes a bidirectional long short-term memory network layer and a unidirectional long short-term memory network layer, and the discriminator includes at least one second convolutional block and a fully connected layer;

[0171] The noise reduction module 1520 is further used for:

[0172] The EEG data to be denoised is sequentially input into a bidirectional long short-term memory network layer and a unidirectional long short-term memory network layer to obtain intermediate denoised data;

[0173] The intermediate denoised data is sequentially input into at least one second convolutional block, and the target denoised data is output through a fully connected layer.

[0174] In a possible implementation, each second convolution block includes a one-dimensional convolution layer, an activation function, and a one-dimensional pooling layer, and the number of second convolution blocks is 5.

[0175] The present application also provides an electronic device. The present application also provides an electronic device. Fig.16 , Fig.16 An internal structure diagram of an electronic device provided for an embodiment of the present application. The electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the electronic device stores an operating system and may also store a computer program, which, when executed by the processor, enables the processor to implement the noise reduction method applied to the electronic device in the above embodiment. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to execute the noise reduction method. Those skilled in the art will appreciate that Fig.16 The structure shown in the figure is merely a block diagram of a partial structure related to the scheme of the present application, and does not constitute a limitation on the electronic device to which the scheme of the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0176] The embodiment of the present application further discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the xxx method in the method embodiment is implemented.

[0177] An embodiment of the present application provides a computer program product, which is stored in a storage medium. The program product is executed by at least one processor to implement the various processes of the embodiment of the noise reduction method as described above, and can achieve similar or identical technical effects. To avoid repetition, it will not be described here.

[0178] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRA M), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRD RAM), and memory bus dynamic RAM (RDRAM), etc.

[0179] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A noise reduction method, characterized in that: The method comprises: Obtaining the EEG data to be denoised; The target denoising model is used to denoise the EEG data to be denoised to obtain target denoised data, wherein the target denoising model includes an encoder and a decoder, or the target denoising model includes a generator and a discriminator, and the target denoising model is trained based on noisy EEG data and noise-free EEG data.

2. The noise reduction method according to claim 1, characterized in that: The target denoising model includes an encoder and a decoder, and the target denoising model is used to denoise the EEG data to be denoised to obtain target denoised data, including: Determining key features of the EEG data to be denoised and removing noise from the data using the encoder; The decoder is used to reconstruct data according to the key features of the EEG data to be denoised and the data from which the noise is removed, so as to obtain the target denoised data.

3. The noise reduction method according to claim 2, characterized in that: The encoder includes at least one first convolution block, each of which includes a convolution layer with padding added, a pooling layer with padding added, and a first activation function, and the decoder includes at least one deconvolution block, each of which includes a deconvolution layer with padding added and a second activation function; The method of using the target denoising model to denoise the EEG data to be denoised to obtain target denoised data includes: Determine the key features of the EEG data to be denoised and the data to be denoised by using the convolutional layer with added padding, the pooling layer with added padding and the first activation function; The deconvolution layer with added padding and the second activation function are used to reconstruct data according to key features of the electroencephalogram data to be denoised and the data from which noise is removed, so as to obtain the target denoised data.

4. The noise reduction method according to claim 3, characterized in that: The first activation function and the second activation function are both ReLU functions, the number of the convolution blocks and the number of the deconvolution blocks are the same, and the number of the convolution blocks and the number of the deconvolution blocks include any one of 4, 5, and 6.

5. The noise reduction method according to claim 1, characterized in that: The target denoising model includes a generator and a discriminator. The target denoising model is used to denoise the EEG data to be denoised to obtain target denoised data, including: Using the generator to perform dimensionality reduction processing on the EEG data to be denoised to obtain intermediate denoised data; The target denoised data is determined from the intermediate denoised data using the discriminator.

6. The noise reduction method according to claim 5, characterized in that: The generator includes a bidirectional long short-term memory network layer and a unidirectional long short-term memory network layer, and the discriminator includes at least one second convolutional block and a fully connected layer; The method of using the target denoising model to denoise the EEG data to be denoised to obtain target denoised data includes: Inputting the EEG data to be denoised into the first bidirectional long short-term memory network layer and the first unidirectional long short-term memory network layer in sequence to obtain the intermediate denoised data; The intermediate denoised data is sequentially input into at least one of the second convolution blocks, and the target denoised data is output through one of the fully connected layers.

7. The noise reduction method according to claim 6, characterized in that: Each of the second convolution blocks includes a one-dimensional convolution layer, an activation function and a one-dimensional pooling layer, and the number of the second convolution blocks is 5.

8. A noise reduction system, characterized in that: The system comprises: An acquisition module, used for acquiring EEG data to be denoised; A denoising module is used to denoise the EEG data to be denoised using a target denoising model to obtain target denoised data, wherein the target denoising model includes an encoder and a decoder, or the target denoising model includes a generator and a discriminator, and the target denoising model is trained based on noisy EEG data and noise-free EEG data.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein a computer program is stored in the memory, and the computer program implements the method according to any one of claims 1 to 7 when executed by the processor.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by one or more processors, the method according to any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Electroencephalogram signal feature extraction and classification method of combining DAE (denoising auto encoder) and CNN (convolutional neural network)

    CN107844755A

  • Visual image reconstruction system based on brain-computer interface

    CN111539331A

  • Stacked denoising auto-encoder model determination method

    CN112861625A

  • Electroencephalogram EEG signal denoising method and system based on deep neural network technology

    CN113208622A

  • Electroencephalogram data reconstruction method based on reinforcement learning auto-encoder

    CN113408697A