阅读说明：本技术 鲁棒性脑电信号的生成方法及装置 (Robust electroencephalogram signal generation method and device ) 是由 夏立坤 张达 于 2021-02-04 设计创作，主要内容包括：本发明提出了一种鲁棒性脑电信号的生成方法及装置。方法包括：步骤Y,获取脑电信号；步骤X,对所述脑电信号进行噪声分离处理,获取去噪后脑电信号和噪声信号；步骤U,将去噪后脑电信号和噪声信号输入生成对抗网络,生成模拟脑电信号并输出；结合定量评估方法和定性评估方法对获取的所述模拟脑电信号进行质量分析。本发明采用WGAN-GP作为基础框架,并使用LSTM代替2DCNN搭建的框架网络,使用脑电信号噪声代替高斯噪声作为生成对抗网络中生成器的输入数据,提高了模拟脑电信号的生成效率和质量。通过从定量评估结果中筛选的多段模拟脑电信号迭加得到的ERP数据与真实ERP的波形对比,从视觉角度对模拟脑电信号的质量进行评估。(The invention provides a method and a device for generating robust electroencephalogram signals. The method comprises the following steps: step Y, acquiring an electroencephalogram signal; step X, carrying out noise separation processing on the electroencephalogram signals to obtain denoised electroencephalogram signals and noise signals; step U, inputting the denoised electroencephalogram signal and the denoised noise signal into a generation countermeasure network, generating and outputting a simulated electroencephalogram signal; and performing quality analysis on the acquired simulated electroencephalogram signals by combining a quantitative evaluation method and a qualitative evaluation method. The method adopts WGAN-GP as a basic frame, uses LSTM to replace a frame network built by 2DCNN, uses EEG signal noise to replace Gaussian noise as input data of a generator in a generation countermeasure network, and improves the generation efficiency and quality of the simulated EEG signal. And comparing the ERP data obtained by superposing a plurality of sections of simulated electroencephalograms screened from the quantitative evaluation result with the waveform of the real ERP, and evaluating the quality of the simulated electroencephalograms from a visual angle.)

1. A method for generating robust electroencephalogram signals, comprising:

step Y, acquiring an electroencephalogram signal;

step X, carrying out noise separation processing on the electroencephalogram signals to obtain denoised electroencephalogram signals and noise signals;

and step U, inputting the denoised electroencephalogram signal and the noise signal into a generation countermeasure network, generating and outputting a simulated electroencephalogram signal.

2. The method of generating a robust brain electrical signal according to claim 1, further comprising:

and step A, performing quality analysis on the acquired simulated electroencephalogram signals by combining a quantitative evaluation method and a qualitative evaluation method.

3. The robust brain electrical signal generation method of claim 2, wherein said qualitative assessment method comprises:

step A1, preprocessing and superposing average operation are carried out on a plurality of sections of acquired electroencephalogram signals to obtain real ERP waves;

step A2, preprocessing and superposing average operation are carried out on a plurality of sections of simulated electroencephalogram signals screened from quantitative evaluation results to obtain simulated ERP waves;

and A3, comparing the real ERP wave with the simulated ERP wave, and performing quality analysis on the simulated electroencephalogram signal.

4. The robust electroencephalogram signal generation method according to claim 1, wherein in the step X, noise separation processing is performed on the electroencephalogram signal by using independent component analysis.

5. The robust electroencephalogram signal generation method according to any one of claims 1 to 4, wherein in the step U, the generation countermeasure network adopts WGAN-GP as a basic framework, and uses LSTM to replace a framework network built by 2 DCNN.

6. An apparatus for generating a robust electroencephalogram signal, comprising:

the acquisition module is used for acquiring an electroencephalogram signal;

the noise separation processing module is used for carrying out noise separation processing on the electroencephalogram signals to obtain denoised electroencephalogram signals and noise signals;

and the analog signal generation module is used for inputting the denoised electroencephalogram signal and the noise signal into a generation countermeasure network, generating and outputting an analog electroencephalogram signal.

7. The robust brain electrical signal generation apparatus of claim 6, further comprising:

and the quality analysis module is used for performing quality analysis on the acquired simulated electroencephalogram signal by combining a qualitative evaluation method and a quantitative evaluation method.

8. The robust brain electrical signal generation apparatus of claim 7, wherein the quality analysis module comprises: the device comprises a qualitative analysis module and a quantitative analysis module, wherein the qualitative analysis module comprises:

the first calculation module is used for carrying out preprocessing and superposition averaging operation on the acquired multiple sections of electroencephalogram signals to obtain a real ERP wave;

the second calculation module is used for preprocessing and superposing average operation on a plurality of sections of the simulated electroencephalogram signals screened from the quantitative evaluation result by the quantitative analysis module to obtain a simulated ERP wave;

and the comparison analysis module is used for performing quality analysis on the simulated electroencephalogram signals by comparing the real ERP waves with the simulated ERP waves.

9. The robust electroencephalogram signal generation apparatus of claim 6, wherein the noise separation processing module performs noise separation processing on the electroencephalogram signal by using independent component analysis.

10. The robust electroencephalogram signal generation apparatus according to any one of claims 6 to 9, wherein the analog signal generation module adopts WGAN-GP as a basic framework, and uses LSTM to replace 2DCNN to build the generation countermeasure network.

Technical Field

The invention relates to the technical field of machine learning, in particular to a method and a device for generating robust electroencephalogram signals.

Background

Electroencephalogram (EEG), which is a multi-channel physiological signal containing a large amount of human body information, has become important data in brain science research. In recent years, feature extraction of EEG using a deep learning model has become an important means of studying EEG. Such techniques typically require training large amounts of data to build robust models. However, due to the tedious EEG acquisition process and the noise effects in the acquisition environment, partial interference signals are present in the data, and it is extremely difficult to obtain a data set that satisfies the requirements of the training model. Therefore, generating robust EEG based on a small number of EEG samples and applying it to deep learning research work is of great significance to the development of brain science.

Generation of countermeasure Networks (GAN) has been applied to various fields (images, audio, etc.) as a mainstream neural network framework for small sample processing. Since the conventional GAN is mainly built with a Fully Connected (FC), the FC cannot capture complex associations between features and correlation attributes of time series signals, so that the model is prone to lose time correlation information when processing EEG.

In addition, the evaluation of GAN still mainly includes qualitative evaluation methods and some quantitative evaluation methods with poor reliability. The qualitative assessment method has the influence of human subjectivity, consumes a large amount of manpower, and cannot evaluate generated samples by using manual assessment in some samples without dominant features.

The existing quantitative assessment method can evaluate the GAN in the aspects of definition, similarity, accuracy and the like. From the viewpoint of clarity, there are currently available characterization methods: (ii) The Inclusion Score (IS) (S.T. Barratt and R.Sharma, A note on The acceptance Score, CoRR, vol.abs/1801.01973,2018.) and The Frechet acceptance Distance (FID) (S.Hochreiter, The variable mapping reducing receiving network and protocol solutions, in Proc. International Journal of Ucertaint, fuzzy and KnowledBased Systems,1998, vol.6, No.2, pp.107-116.); from a similarity perspective, Waterstein Distance (WD) (F.Otto and M.Westdickenberg, Eurrian calsulus for the linkage in the Wasserstein Distance, SIAM J.Math.analysis (PAMI),2005, vol.37, No.4, pp.1227-1255.), Euclidean Distance (ED) (L.Wang, Y.Zhang, and J.Feng, On the emultiude Distance of images, in Proc.IEEE trade. Pattern Analysis and Machine Analysis, 2005, MM27, vol.8, pp.1334-1339) and Kernel Maxim Distance spectrum (D) [26] et al compare the degree of similarity between samples from sample Distance distribution; from an accuracy perspective, 1-Nearest Neighbor accuracycary (m.govindarajan and r.m.chandarsekaran, Evaluation of k-Nearest Neighbor classifier performance for direct marking, in proc.expert system. appl.,2010, vol.37, No.1, pp.253-258.) determines the accuracy of generating a sample based on a classifier.

However, these methods can only measure the performance of GAN from the evaluation results of the generated samples, and cannot identify whether there are problems such as pattern collapse and overfitting in the generation process. Currently, there is no reasonable evaluation method that can effectively evaluate GAN performance. For EEG and other data which cannot be evaluated manually, the evaluation difficulty is increased.

In summary, constructing a stable countermeasure network and designing a reasonable evaluation method are key issues to be solved in the robust EEG signal generation method.

Disclosure of Invention

The invention provides a method and a device for generating a robust electroencephalogram signal, and aims to solve the technical problem of how to generate and evaluate a high-quality EEG signal.

The robust electroencephalogram signal generation method provided by the embodiment of the invention comprises the following steps:

step Y, acquiring an electroencephalogram signal;

step X, carrying out noise separation processing on the electroencephalogram signals to obtain denoised electroencephalogram signals and noise signals;

and step U, inputting the denoised electroencephalogram signal and the noise signal into a generation countermeasure network, generating and outputting a simulated electroencephalogram signal.

According to the robust electroencephalogram signal generation method, electroencephalogram signal noise is adopted to replace Gaussian noise to serve as input data for generating the countermeasure network, authenticity of a generated sample can be improved, meanwhile, fitting speed of the generator is increased, and generation efficiency and quality of the analog electroencephalogram signal are improved.

According to some embodiments of the invention, the method further comprises:

and step A, performing quality analysis on the acquired simulated electroencephalogram signals by combining a quantitative evaluation method and a qualitative evaluation method.

In some embodiments of the invention, the qualitative assessment method comprises:

step A1, preprocessing and superposing average operation are carried out on a plurality of sections of acquired electroencephalogram signals to obtain real ERP waves;

step A2, preprocessing and superposing average operation are carried out on a plurality of sections of simulated electroencephalogram signals screened from quantitative evaluation results to obtain simulated ERP waves;

and A3, comparing the real ERP wave with the simulated ERP wave, and performing quality analysis on the simulated electroencephalogram signal.

According to some embodiments of the invention, in the step X, noise separation processing is performed on the electroencephalogram signal by using independent component analysis.

In some embodiments of the present invention, in the step U, the generation of the countermeasure network uses WGAN-GP as a basic framework, and uses LSTM to replace a framework network built by 2 DCNN.

The robust electroencephalogram signal generation device provided by the embodiment of the invention comprises the following components:

the acquisition module is used for acquiring an electroencephalogram signal;

the noise separation processing module is used for carrying out noise separation processing on the electroencephalogram signals to obtain denoised electroencephalogram signals and noise signals;

and the analog signal generation module is used for inputting the denoised electroencephalogram signal and the noise signal into a generation countermeasure network, generating and outputting an analog electroencephalogram signal.

According to the robust electroencephalogram signal generation device, noise in the electroencephalogram signal is obtained through the noise separation processing module, the electroencephalogram signal noise is adopted to replace Gaussian noise to serve as input data for generation of the countermeasure network, authenticity of a generated sample can be improved, meanwhile, fitting speed of a generator is increased, and generation efficiency and quality of the simulated electroencephalogram signal are improved.

According to some embodiments of the invention, the generating means further comprises:

and the quality analysis module is used for carrying out quality analysis on the acquired simulated electroencephalogram signal by combining a quantitative evaluation method and a qualitative evaluation method.

In some embodiments of the invention, the mass analysis module comprises: the device comprises a qualitative analysis module and a quantitative analysis module, wherein the qualitative analysis module comprises:

the first calculation module is used for carrying out preprocessing and superposition averaging operation on the acquired multiple sections of electroencephalogram signals to obtain a real ERP wave;

the second calculation module is used for preprocessing and superposing average operation on a plurality of sections of the simulated electroencephalogram signals screened from the quantitative evaluation result through the quantitative analysis module to obtain a simulated ERP wave;

and the comparison analysis module is used for performing quality analysis on the simulated electroencephalogram signals by comparing the real ERP waves with the simulated ERP waves.

According to some embodiments of the invention, the noise separation processing module performs noise separation processing on the electroencephalogram signal by using independent component analysis.

In some embodiments of the invention, the analog signal generation module adopts WGAN-GP as a basic framework, and uses LSTM to build the generative countermeasure network instead of 2 DCNN.

Drawings

FIG. 1 is a flow chart of a method for generating robust electroencephalogram signals according to an embodiment of the present invention;

FIG. 2 is a flow chart of a method for generating robust electroencephalogram signals according to an embodiment of the present invention;

FIG. 3 is a block diagram of a generator and arbiter according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a device for generating robust electroencephalogram signals according to an embodiment of the present invention.

The generation means (100) is generated by,

the system comprises an acquisition module 10, a noise separation processing module 20, an analog signal generation module 30, a quality analysis module 40, a qualitative analysis module 410, a first calculation module 411, a second calculation module 412, a comparison analysis module 413 and a quantitative analysis module 420.

Detailed Description

To further explain the technical means and effects of the present invention adopted to achieve the intended purpose, the present invention will be described in detail with reference to the accompanying drawings and preferred embodiments.

It should be noted that, in the method steps of the present invention, the execution order of the steps is not necessarily related to the step numbers, for example, step Y may precede step X, and the execution order is not performed according to the alphabetical ordering of the steps.

The robust electroencephalogram signal generation method provided by the embodiment of the invention comprises the following steps:

step Y, acquiring an electroencephalogram signal;

it should be noted that, after acquiring the electroencephalogram signal, the electroencephalogram signal may be preprocessed, and in the preprocessing stage of the electroencephalogram signal, the electroencephalogram signal is normalized, and the value range of the real electroencephalogram signal is compressed to [ -1 μ v,1 μ v ], so as to solve the problem of the deviation of the value range of the generated electroencephalogram signal, and to ensure that the value range of the generated sample is consistent with the value range of the real sample.

Step X, carrying out noise separation processing on the electroencephalogram signals to obtain denoised electroencephalogram signals and noise signals;

and step U, inputting the denoised electroencephalogram signal and noise signal into a generation countermeasure network, generating and outputting a simulated electroencephalogram signal.

According to the robust electroencephalogram signal generation method, electroencephalogram signal noise is adopted to replace Gaussian noise to serve as input data for generating the countermeasure network, authenticity of a generated sample can be improved, meanwhile, fitting speed of the generator is increased, and generation efficiency and quality of the analog electroencephalogram signal are improved.

According to some embodiments of the invention, the method further comprises:

and step A, performing quality analysis on the acquired simulated electroencephalogram signals by combining a quantitative evaluation method and a qualitative evaluation method.

Further, the qualitative assessment method comprises:

step A1, preprocessing and superposing average operation are carried out on the obtained multiple segments of electroencephalogram signals to obtain real ERP waves;

step A2, preprocessing and superposing average operation are carried out on a plurality of sections of simulated electroencephalogram signals screened from quantitative evaluation results to obtain simulated ERP waves;

and step A3, performing quality analysis on the simulated brain electric signal by comparing the real ERP wave with the simulated ERP wave.

It should be noted that, because the qualitative assessment method needs to consume a large amount of human resources, the multiple segments of generated EEG are superimposed to obtain ERP data, and the quality of the generated EEG is assessed visually by comparing the waveform difference between the real ERP and the generated ERP.

According to some embodiments of the invention, in step X, noise separation processing is performed on the electroencephalogram signal by using independent component analysis. Independent Component Analysis (ICA) can reliably separate the brain electrical signals into de-noised brain electrical signals and noise signals.

In some embodiments of the invention, in step U, the generation of the countermeasure network adopts WGAN-GP as a basic framework, and uses Long Short-Term Memory network (LSTM) instead of the framework network built by 2 DCNN.

It should be noted that, compared to CNN, LSTM can better process timing signals and reduce the corruption of timing information. According to the invention, the LSTM is used for building the WGAN-GP framework for generating the EEG, so that the stability of the model training process is improved on one hand, and the quality of the generated sample is improved on the other hand.

The robust electroencephalogram signal generation apparatus 100 according to the embodiment of the present invention includes: an acquisition module 10, a noise separation processing module 20 and an analog signal generation module 30.

The obtaining module 10 is used for obtaining an electroencephalogram signal.

It should be noted that, after acquiring the electroencephalogram signal, the electroencephalogram signal may be preprocessed, and in the preprocessing stage of the electroencephalogram signal, the electroencephalogram signal is normalized, and the value range of the real electroencephalogram signal is compressed to [ -1 μ v,1 μ v ], so as to solve the problem of the deviation of the value range of the generated electroencephalogram signal, and to ensure that the value range of the generated sample is consistent with the value range of the real sample.

The noise separation processing module 20 is configured to perform noise separation processing on the electroencephalogram signal to obtain a denoised electroencephalogram signal and a noise signal;

the analog signal generating module 30 is configured to input the denoised electroencephalogram signal and the noise signal into a countermeasure network, generate an analog electroencephalogram signal, and output the analog electroencephalogram signal.

According to the robust electroencephalogram signal generation device 100 provided by the embodiment of the invention, the noise in the electroencephalogram signal is obtained through the noise separation processing module 20, and the electroencephalogram signal noise is adopted to replace Gaussian noise as the input data of the generator, so that the authenticity of a generated sample can be improved, meanwhile, the fitting speed of the generator is increased, and the generation efficiency and quality of the analog electroencephalogram signal are improved.

According to some embodiments of the invention, the generating means 100 further comprises: and the quality analysis module 40 is used for performing quality analysis on the acquired simulated electroencephalogram signals by combining a quantitative evaluation method and a qualitative evaluation method.

Further, the mass analysis module 40 includes: a qualitative analysis module 410 and a quantitative analysis module 420, the qualitative analysis module 410 comprising: a first calculation module 411, a second calculation module 412, and a comparative analysis module 413.

The first computing module 411 is configured to perform preprocessing and superposition averaging on the acquired multiple segments of electroencephalogram signals to obtain a real ERP wave;

the second calculation module 412 is configured to perform preprocessing and superposition averaging operations on multiple segments of simulated electroencephalogram signals screened from the quantitative evaluation result through the quantitative analysis module 420 to obtain a simulated ERP wave;

the comparison analysis module 413 is configured to perform quality analysis on the simulated brain electrical signal by comparing the real ERP wave and the simulated ERP wave.

It should be noted that, because the qualitative evaluation method consumes a lot of human resources, the present invention obtains ERP data by superimposing multiple segments of generated EEG, and evaluates the quality of generated EEG from a visual perspective by comparing the waveform difference between the real ERP and the generated ERP.

According to some embodiments of the present invention, the noise separation processing module 20 performs noise separation processing on the electroencephalogram signal using independent component analysis. Independent Component Analysis (ICA) can reliably separate the brain electrical signals into de-noised brain electrical signals and noise signals.

In some embodiments of the present invention, the analog signal generation module 30 uses WGAN-GP as a base framework and uses LSTM instead of 2DCNN building to generate the countermeasure network.

It should be noted that, compared to CNN, LSTM can better process timing signals and reduce the corruption of timing information. According to the invention, the LSTM is used for building the WGAN-GP framework for generating the EEG, so that the stability of the model training process is improved on one hand, and the quality of the generated sample is improved on the other hand.

The method and apparatus for robust brain electrical signals according to the present invention are described in detail in one embodiment with reference to the accompanying drawings. It is to be understood that the following description is only exemplary in nature and should not be taken as a specific limitation on the invention.

When electroencephalogram signals are processed, the existing network framework often has the following problems:

(1) problems of overfitting, mode collapse, gradient disappearance and the like easily occur in the model training process, so that the diversity of generated samples is insufficient; (2) the CNN destroys the time correlation in the channel information in the EEG processing process, so that the generated sample loses the time correlation information; (3) the existing model usually improves the stability of model training by increasing the complexity of the model, so that the fitting speed of the model is reduced. (4) The generated sample range is offset from the true sample range. (5) The quantitative evaluation method cannot comprehensively evaluate the quality of a generated sample, and the qualitative evaluation consumes a large amount of human resources.

Aiming at the problems, the invention provides a novel network model combining the LSTM and the WGAN-GP. Firstly, compared with other derivative models of GAN, the loss function of WGAN-GP can prevent problems such as overfitting and gradient disappearance; secondly, the LSTM can avoid the problems of gradient explosion and the like caused by overlong signal time step, and further improve the training stability of the model. On the basis of this model, the following improvements are proposed for generating robust EEG signals: (1) in the preprocessing stage of EEG, the EEG signal is normalized, the range of real EEG is compressed to [ -1 μ v,1 μ v ], and the problem of EEG range deviation is solved; (2) the filtered noise in the EEG is used as input data instead of the gaussian noise in the original GAN, improving the fitting speed of the model.

In terms of quantitative evaluation, the quality of the generated sample is evaluated in terms of diversity, similarity, and the like using various evaluation indexes; in the aspect of qualitative evaluation, the generated samples screened according to the quantitative evaluation result are superposed to obtain ERP data, and the authenticity of the generated samples is verified by comparing the ERP characteristic waveforms of the real samples and the generated samples.

As shown in fig. 1 and 2, the robust EEG generation method according to the present invention includes: a signal pre-processing phase, an EEG generation phase, and an EEG quality assessment phase.

Firstly, noise reduction filtering processing is carried out on EEG signals in an opening source data set, the EEG signals are separated into de-noised EEG signals and noise signals through Independent Component Analysis (ICA), and the EEG and the noise are divided according to time points so as to carry out qualitative evaluation; then, inputting the data into WGAN-GP built by LSTM to obtain the generated EEG; thereafter, the quality of the generated EEG was verified by qualitative and quantitative evaluation methods.

EEG noise:

to improve the realism of the generated samples and the fitting speed of the generator, filtered interfering noise in the real EEG signal is used instead of gaussian noise as input data to the generator. The unprocessed EEG signal not only contains brain cell potential activity information, but also contains tested eye movement, electrocardio, myoelectricity and other information, which can interfere the work of EEG analysis, and the information has a distribution similar to that of EEG, and can be used as noise to enable a generator to better fit the distribution of a real sample.

EEG generation:

the present invention uses WGAN-GP as the base framework for generating EEG. And (3) constructing a frame by using LSTM instead of 2DCNN, outputting a pseudo EEG signal after the input data of a generator is EEG noise and then inputting the real EEG and the pseudo EEG into a discriminator, and obtaining a true and false result of the sample through the discriminator. The network framework of generators and discriminators is shown in fig. 3.

The present invention uses the same noise signal as the EEG dimension as the input data to the generator, goes through one fully connected layer of 1024 cells and is activated using leak Relu, then goes through another fully connected layer and converts the output dimension to a three-dimensional tensor of 1 x 120 x 128. And the dimension of the output tensor of the previous layer is increased to 1 multiplied by 240 multiplied by 128 by adopting a bicubic interpolation method for up-sampling. To process the output using LSTM, the output dimensions are again converted to 240 × 28, and the resulting output is fed back into the two layers of LSTM. The input of the LSTM is integrated into the length of the real EEG by the slice layer. And finally, integrating the output of the upper layer by using a layer of LSTM to ensure that the dimension of the LSTM meets the input dimension of the discriminator.

The inputs to the discriminator are the true EEG signal with dimensions 231 x 1 and the pseudo signal output by the generator. In the invention, Gaussian white noise is added into the discriminator to corrode the input signal so as to help improve the training stability of the model. Then, the processed output is analyzed and flattened through three LSTMs with Leaky Relu activation functions, a tensor with the dimension of 29568 multiplied by 1 is provided for the full connection layer, and finally the identification result of the discriminator is output through the full connection layer.

The qualitative evaluation method comprises the following steps:

ERP is an evoked potential reflecting the activity of neuron cells when the brain activates the object, and is obtained by subjecting a plurality of segments of EEG signals to a series of operations such as preprocessing, superposition averaging, and the like. The method adopts a difference wave analysis method, and evaluates the quality of the generated sample from a visual angle by comparing the generated EEG with the ERP characteristic difference of the real EEG. Compared with EEG, the ERP features are obvious, and the EEG data are obtained by overlapping multiple sections of EEG data, so that the difficulty and cost of manual evaluation are reduced.

In summary, the EEG generation method proposed by the present invention includes three phases: a signal pre-processing phase, an EEG generation phase, and an EEG quality assessment phase. Firstly, extracting EEG noise and a real EEG signal in an EEG preprocessing stage, and carrying out standardization processing on the real EEG signal; then, establishing a WGAN-GP framework by using the LSTM, and respectively using EEG noise and the EEG signal after the standardization processing as input data of a generator and a discriminator to obtain the generated EEG; in the EEG evaluation phase, the similarity and diversity of the generated EEG are evaluated through SWD and MS, a plurality of generated EEG are superposed into ERP, and the authenticity of the generated EEG is judged through P300 characteristics.

While the invention has been described in connection with specific embodiments thereof, it is to be understood that it is intended by the appended drawings and description that the invention may be embodied in other specific forms without departing from the spirit or scope of the invention.