CN114469135A

CN114469135A - Artifact removing method and device for multi-lead electroencephalogram signal and brain-computer interface

Info

Publication number: CN114469135A
Application number: CN202210051158.6A
Authority: CN
Inventors: 向绍鑫; 郝慎才; 王晓岸
Original assignee: Beijing Brain Up Technology Co ltd
Current assignee: Beijing Brain Up Technology Co ltd
Priority date: 2022-01-17
Filing date: 2022-01-17
Publication date: 2022-05-13

Abstract

The application discloses an artifact removing method and device based on multi-lead electroencephalogram signals and a brain-computer interface, and relates to the technical field of new generation information, wherein the method comprises the following steps: preprocessing a multi-lead electroencephalogram signal to be processed, decomposing the multi-lead electroencephalogram signal into a plurality of independent components by using an Independent Component Analysis (ICA) algorithm, converting the plurality of independent components into an electroencephalogram characteristic diagram, identifying preset components to which the plurality of independent components belong according to a pre-constructed artifact identification model, removing the independent components belonging to artifacts, and then reconstructing to obtain a pure electroencephalogram signal without artifacts; according to the method and the device, the number of the maps input into the artifact identification model is increased, information provided for the artifact identification model is enriched, the artifact identification model can comprehensively consider the effect of the distribution characteristics of the artifacts in different maps, the purpose of improving the artifact identification precision is achieved by keeping the local characteristics of signals in different maps, and the purity of the reconstructed electroencephalogram signal is improved.

Description

Artifact removing method and device for multi-lead electroencephalogram signal and brain-computer interface

Technical Field

The application relates to the technical field of new generation information, in particular to the technical field of processing brain-computer interface data, and particularly relates to an artifact removing method and device for a multi-lead electroencephalogram signal and a brain-computer interface.

Background

Electroencephalogram (EEG) provides a bioelectric signal that reflects brain activity, and has a very important role in the field of research of human brain. The electroencephalogram signal is very weak and has time-varying sensitivity, and the acquisition process is easily influenced by irrelevant factors, so that the acquired electroencephalogram signal has artifacts such as electrooculogram, myoelectricity, electrocardio, noise and the like. The presence of these artifacts greatly affects the analysis and recognition of brain electrical signals. The related artifact removal technology has the problems that only one artifact can be identified at a time, and the identification precision of the artifact is poor due to the fact that the artifact is identified by adopting a single feature.

Disclosure of Invention

The application provides an artifact removing method and device for a multi-lead electroencephalogram signal, an electronic device, a computer-readable storage medium, a chip and a brain-computer interface, which can solve at least one technical problem. The technical scheme is as follows:

in a first aspect, an artifact removing method for multi-lead electroencephalogram signals is provided, and the method comprises the following steps:

acquiring a multi-lead electroencephalogram signal to be processed;

decomposing the multi-lead electroencephalogram signal into a plurality of independent components by utilizing a preset independent component analysis algorithm;

carrying out map conversion on the plurality of independent components to obtain an electroencephalogram characteristic diagram;

identifying the electroencephalogram characteristic diagram according to a pre-constructed artifact identification model to obtain component probabilities that a plurality of independent components respectively belong to various preset components and respectively correspond to the preset components;

determining the preset components to which the independent components belong based on component probabilities corresponding to the preset components to which the independent components belong;

if the preset component to which any independent component belongs is an artifact, removing the any independent component, and then reconstructing the remaining independent components to obtain a pure electroencephalogram signal with the artifact removed.

Further, the electroencephalogram feature map at least comprises:

brain topography and power spectral density PSD plots.

Further, the step of carrying out map conversion on the plurality of independent components to obtain an electroencephalogram characteristic map comprises the following steps:

fourier transformation is carried out on the independent components and square calculation is carried out on the independent components to obtain Power Spectral Density (PSD) graphs corresponding to the independent components;

and respectively carrying out power spectrum extraction on the plurality of independent components according to the same components of the channel and carrying out spatial interpolation to construct a brain topographic map.

Further, before the step of identifying the electroencephalogram characteristic diagram according to the pre-constructed artifact identification model to obtain component probabilities that the plurality of independent components respectively belong to various preset components and respectively correspond to the preset components, the method further comprises the following steps: determining a pattern for identifying the artifact; the method comprises the following steps of identifying an electroencephalogram characteristic diagram according to a pre-constructed artifact identification model to obtain component probabilities that a plurality of independent components respectively belong to various preset components and respectively correspond to the preset components, and comprises the following steps: and if the mode for identifying the artifacts is a mode for simultaneously identifying multiple artifacts, identifying the brain topographic map and/or the Power Spectral Density (PSD) map by using an artifact identification model to obtain component probabilities corresponding to various preset components of the independent components.

Further, the step of determining the component classification to which each of the plurality of independent components belongs based on the component probability that each of the plurality of independent components belongs to each of the predetermined components, respectively, includes:

determining the maximum value of the component probability of each independent component based on the component probability corresponding to each preset component belonging to each independent component;

and determining the preset components to which the independent components belong according to the maximum value of the component probability of each independent component.

Further, the step of identifying the electroencephalogram characteristic diagram according to the pre-constructed artifact identification model to obtain component probabilities that the plurality of independent components respectively belong to various preset components and respectively correspond to the preset components comprises the following steps:

if the mode for recognizing the artifacts is a mode for recognizing various artifacts one by one, recognizing the brain topographic map and the power spectral density PSD map by utilizing the pre-constructed artifact recognition models respectively corresponding to various artifacts to obtain component probabilities respectively corresponding to various preset components of a plurality of independent components.

Further, the step of determining the predetermined components to which the plurality of independent components belong based on component probabilities corresponding to the respective types of the predetermined components to which the plurality of independent components belong includes:

determining the maximum value of the component probability of each preset component based on the component probability corresponding to each preset component belonging to each independent component;

and determining the preset components to which the independent components belong according to the maximum value of the component probability of each preset component.

Further, before the step of identifying the electroencephalogram feature map according to the pre-constructed artifact identification model to obtain component probabilities that the independent components respectively belong to various predetermined components and respectively correspond to the predetermined components, the method further comprises the following steps:

determining a neural network model for training based on a predetermined model framework, the neural network model comprising in sequence:

the device comprises an input layer, a Stem layer, a first introduction layer, a first Reduction layer, a second introduction layer, a second Reduction layer, a third introduction layer, an average pooling layer, a drop layer and a Softmax layer.

Further, before the step of identifying the electroencephalogram characteristic diagram according to the pre-constructed artifact identification model to obtain component probabilities that the plurality of independent components respectively belong to various preset components and respectively correspond to the preset components, the method further comprises the following steps:

the method comprises the steps that a neural network model used for training is determined based on a preset model framework, and the neural network model sequentially comprises an input layer, a first pooling layer, a channel merging layer, a parallel superposition residual error neural network, a second pooling layer and an output layer, wherein the parallel superposition residual error neural network comprises a plurality of parallel convolution blocks provided with an inner side fast channel and an outer side fast channel, any parallel convolution block comprises a plurality of convolution layer branches, a branch channel merging layer and a scaling factor, the output of the plurality of convolution layer branches is the input of the branch channel merging layer, and the output of the branch channel merging layer is the input of the scaling factor.

Further, the method further comprises:

acquiring a sample set, wherein the sample set comprises electroencephalogram characteristic maps of a plurality of independent components respectively corresponding to a plurality of electroencephalograms, and component identifications of the electroencephalogram characteristic maps of the independent components respectively corresponding to the plurality of electroencephalograms;

and training the neural network model by using the sample set to obtain a trained artifact identification model.

Further, the step of training the neural network model by using the sample set to obtain a trained artifact recognition model includes:

configuring a neural network model based on a plurality of preset groups of parameters to obtain a plurality of candidate identification models;

respectively training a plurality of candidate recognition models by using a sample set; determining the respective accuracy of the candidate recognition models according to the training result;

and taking the candidate recognition model with the highest accuracy as a trained artifact recognition model.

Further, after the step of acquiring the multi-lead electroencephalogram signal to be processed, the method further comprises the following steps:

preprocessing a multi-lead electroencephalogram signal;

the preprocessing comprises at least one of channel position positioning, useless electrode deleting, re-reference, filtering, sampling rate reduction, baseline correction and data segmentation; or

The preprocessing comprises channel position locating, useless electrode deleting, re-reference, filtering, sampling rate reduction, baseline correction and data segmentation in sequence.

In a second aspect, an artifact removing apparatus for multi-lead electroencephalogram signals is provided, the apparatus comprising:

the electroencephalogram signal acquisition module is used for acquiring a multi-lead electroencephalogram signal to be processed;

the electroencephalogram signal decomposition module is used for decomposing the multi-lead electroencephalogram signal into a plurality of independent components by utilizing a preset independent component analysis algorithm;

the signal map conversion module is used for performing map conversion on the plurality of independent components to obtain an electroencephalogram characteristic map;

the component probability determination module is used for identifying the electroencephalogram characteristic diagram according to the pre-constructed artifact identification model to obtain component probabilities corresponding to the fact that the independent components respectively belong to various preset components;

the component classification determining module is used for determining the preset components to which the independent components belong based on the component probabilities corresponding to the preset components to which the independent components belong;

and the artifact removal processing module is used for removing any independent component and reconstructing the residual independent component to obtain the artifact-removed pure electroencephalogram signal if the preset component to which the independent component belongs is the artifact.

Further, the electroencephalogram feature map at least comprises:

brain topography and power spectral density PSD plots.

Further, the signal map conversion module comprises:

the first conversion submodule is used for performing Fourier transform and square calculation on the plurality of independent components respectively to obtain Power Spectral Density (PSD) graphs corresponding to the plurality of independent components respectively;

and the second conversion submodule is used for performing power spectrum extraction and spatial interpolation on the plurality of independent components according to the same components of the channel respectively to construct a brain topographic map.

Further, before the step of identifying the electroencephalogram characteristic diagram by the component probability determining module according to the pre-constructed artifact identification model to obtain the component probabilities corresponding to the preset components of the independent components, the device further comprises: the mode determining module is used for determining a mode for identifying the artifact; the component probability determination module comprises: and the first component probability determining submodule is used for identifying the brain topographic map and/or the Power Spectral Density (PSD) map by using an artifact identification model if the artifact identification mode is a mode for simultaneously identifying multiple artifacts, so as to obtain component probabilities corresponding to various preset components of the independent components.

Further, the component classification determination module includes:

the first maximum value determining submodule is used for determining the maximum value of the component probability of each of the plurality of independent components based on the component probability that each of the plurality of independent components belongs to each of the preset components and corresponds to each of the preset components;

and the first component category determining submodule is used for determining the preset components to which the independent components belong according to the maximum value of the component probability of each independent component.

Further, the component probability determination module includes:

and the second component probability determination submodule is used for identifying the brain topographic map and the power spectral density PSD map by utilizing the pre-constructed artifact identification models respectively corresponding to various artifacts if the artifact identification mode is a mode for identifying various artifacts one by one so as to obtain component probabilities respectively corresponding to various preset components belonging to the independent components.

Further, the component classification determination module includes:

the second maximum value determining submodule is used for determining the maximum value of the component probability of each preset component based on the component probability of each preset component belonging to each preset component;

and the second component category determination submodule is used for determining the preset components to which the independent components belong according to the maximum value of the component probability of each preset component.

Further, before the component probability determining module identifies the electroencephalogram feature map according to the pre-constructed artifact identification model to obtain component probabilities corresponding to each of a plurality of independent components belonging to each of various predetermined components, the component probability determining module further includes:

a network model setting submodule for determining a neural network model for training based on a predetermined model framework, the neural network model comprising in order: the device comprises an input layer, a Stem layer, a first introduction layer, a first Reduction layer, a second introduction layer, a second Reduction layer, a third introduction layer, an average pooling layer, a drop layer and a Softmax layer.

Further, before the step of identifying the electroencephalogram feature map according to the pre-constructed artifact identification model to obtain component probabilities corresponding to each of a plurality of independent components belonging to each of various predetermined components, the component probability determination module further includes:

Still further, the component probability determination module further comprises:

the sample acquisition sub-module is used for acquiring a sample set, wherein the sample set comprises electroencephalogram characteristic maps of a plurality of independent components respectively corresponding to a plurality of electroencephalograms and component identifications of the electroencephalogram characteristic maps of the independent components respectively corresponding to the electroencephalograms;

and the network model training submodule is used for training the neural network model by utilizing the sample set to obtain a trained artifact identification model.

Further, the network model training submodule includes:

the candidate model determining unit is used for configuring the neural network model based on a plurality of preset groups of parameters to obtain a plurality of candidate identification models;

the candidate model training unit is used for respectively training the candidate recognition models by utilizing the sample set; the candidate model accuracy rate determining unit is used for determining the respective accuracy rate of the plurality of candidate recognition models according to the training result;

and the artifact identification model determining unit is used for taking the candidate identification model with the highest accuracy as the trained artifact identification model.

Further, after the step of acquiring the multi-lead electroencephalogram signal to be processed, the electroencephalogram signal acquisition module further comprises:

the signal preprocessing submodule is used for preprocessing the multi-lead electroencephalogram signals;

In a third aspect, an electronic device is provided, which includes:

one or more processors;

a memory;

one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs configured to: and executing the artifact removing method of the multi-lead electroencephalogram signal.

In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, which when executed by a processor, implements the above-described artifact removal method for multi-lead electroencephalogram signals.

In a fifth aspect, a chip for performing the artifact removal method according to the multi-lead electroencephalogram signal is provided.

In a sixth aspect, a brain-computer interface is provided, comprising a wearable device applying a chip for executing the artifact removal method according to the multi-lead brain electrical signal.

The technical scheme provided by the embodiment of the application has the following beneficial effects: the method comprises the steps of obtaining a multi-lead electroencephalogram signal to be processed, decomposing the multi-lead electroencephalogram signal into a plurality of independent components by utilizing a preset independent component algorithm, carrying out map conversion on the plurality of independent components to obtain an electroencephalogram characteristic map, identifying the electroencephalogram characteristic map according to a pre-constructed artifact identification model to obtain component probabilities corresponding to the independent components belonging to various preset components respectively, determining the preset components belonging to the independent components respectively based on the component probabilities corresponding to the independent components belonging to the preset components respectively, reconstructing the residual independent components after removing any one independent component to obtain an artifact-removed readable electroencephalogram signal, and mapping the electroencephalogram signal which is not easy to observe into a high-readability map by adopting a mode of identifying the artifacts through a brain characteristic map, the method not only provides auxiliary data for visual analysis of the electroencephalogram signals, but also achieves the purpose of enriching information provided for the artifact identification model by increasing the number of the maps input to the artifact identification model, so that the artifact identification model can realize automatic identification of artifacts and finally remove artifact components, comprehensively consider the distribution characteristics of the artifacts in different maps, and retain a large amount of electroencephalogram information on the basis of automatic identification of various artifacts by retaining the local characteristics of the signals in different maps, thereby improving the artifact identification precision.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings used in the description of the embodiments of the present application will be briefly described below.

FIG. 1 is a schematic structural diagram of an artifact removal method for multi-lead electroencephalogram signals provided by an embodiment of the present application;

FIG. 2a is a schematic structural diagram of an embodiment of an artifact identification model in the artifact removal method for multi-lead electroencephalogram signals provided by the embodiment of the present application;

fig. 2b is a schematic structural diagram of an embodiment of a Stem layer in an embodiment of an artifact identification model in the method for removing artifacts from multi-lead electroencephalogram signals provided by the embodiment of the present application;

fig. 3 is a schematic structural diagram of each inclusion layer of an embodiment of an artifact identification model in the artifact removal method for multi-lead electroencephalogram signals provided by the embodiment of the present application;

FIG. 4 is a schematic structural diagram of another embodiment of an artifact identification model in the artifact removal method for multi-lead electroencephalogram signals provided by the embodiment of the present application;

FIG. 5 is a schematic diagram illustrating a preprocessing flow according to an embodiment of an artifact removal method for multi-lead electroencephalogram signals provided by the present application;

FIG. 6 is a schematic workflow diagram of an application system according to an embodiment of a method for artifact removal of multi-lead electroencephalogram signals provided by the embodiment of the present application;

fig. 7 is a schematic flowchart of a sample set obtaining process for training an artifact recognition model in an application system according to an embodiment of the method for removing artifacts from multi-lead electroencephalogram signals provided in the embodiment of the present application; and

fig. 8 is a schematic structural diagram of an artifact removing apparatus for multi-lead electroencephalogram signals provided in an embodiment of the present application.

Detailed Description

Reference will now be made in detail to the embodiments of the present application, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary only for the purpose of explaining the present application and are not to be construed as limiting the present application.

As used herein, the singular forms "a", "an", "the" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Further, "connected" or "coupled" as used herein may include wirelessly connected or wirelessly coupled. As used herein, the term "and/or" includes all or any element and all combinations of one or more of the associated listed items.

The following describes the technical solutions of the present application and how to solve the above technical problems with specific embodiments. The following several specific embodiments may be combined with each other, and details of the same or similar concepts or processes may not be repeated in some embodiments. Embodiments of the present application will be described below with reference to the accompanying drawings.

The embodiment of the application provides an artifact removing method for a multi-lead electroencephalogram signal, as shown in fig. 1, the method comprises the following steps: step S101 to step S106.

S101, acquiring a multi-lead electroencephalogram signal to be processed.

Specifically, the electronic device acquires a multi-lead electroencephalogram signal to be processed. The electronic equipment comprises a mobile phone, a PC (personal computer), a server, wearable equipment, brain-computer interfaces such as an electroencephalogram cap and the like.

In an embodiment of the present application, the multi-lead electroencephalogram signal may be an electroencephalogram signal acquired through a brain-computer interface having a plurality of electrodes. For example, a multi-lead brain electrical signal is the brain electrical signal of a user under test acquired from a 32-channel brain electrical cap.

In an embodiment of the present application, a multi-lead electroencephalogram signal is used to characterize an electroencephalogram signal containing a plurality of acquisition points. Specifically, the electronic device may obtain the multi-lead electroencephalogram signal to be processed in real time from a device, such as an electroencephalogram cap or other wearable device, which sets the electrode position according to the 10-20 international standard lead system, or may store the signal acquired by the device, which sets the electrode position according to the 10-20 international standard lead system, locally in advance, and read the signal locally when needed.

And S102, decomposing the multi-lead electroencephalogram signal into a plurality of independent components by utilizing a preset independent component analysis algorithm.

The multi-lead electroencephalogram signals are decomposed through the independent component ICA algorithm, and the purpose of separating the mixed signals is achieved. Specifically, the specific process of the independent component ICA algorithm includes: assuming that the multi-lead electroencephalogram signal is an observation signal, x (T) ([ x1(T), x2(T), …, xn (T) ] T, s (T) ([ s1T), s2(T), …, sm (T) ] T are M mutually statistically independent source signals for generating the observation signal x (T), and the observation signal x (T) is generated by linearly mixing the source signals s (T) through an unknown matrix a, that is, x (T) ═ as (T), an unmixed matrix W is obtained by an independent component ICA algorithm, and an output signal y (T) ([ wx (T) ═ was (T) is obtained to make the output signal approach the true source signal s (T) as much as possible, so as to obtain an independent component.

And S103, performing map conversion on the plurality of independent components to obtain an electroencephalogram characteristic map.

In the embodiment of the application, the electroencephalogram feature map is used for representing the result of the independent components presented on the image. Specifically, the electroencephalogram feature map may be a representation of the independent components in the time domain, the frequency domain, and the time-frequency domain.

And S104, recognizing the electroencephalogram characteristic diagram according to the pre-constructed artifact recognition model to obtain component probabilities that the independent components respectively belong to various preset components and respectively correspond to the preset components.

Specifically, the artifact identification model may be obtained by training a pre-established neural network model, such as a long-term and short-term memory network LSTM, a convolutional neural network CNN, a recurrent neural network RNN, or the like.

Before application, a large number of electroencephalogram signals with artifacts are decomposed by utilizing an Independent Component Analysis (ICA) algorithm, so that independent components of different electroencephalogram signals are obtained, and after the independent components of the different electroencephalogram signals are subjected to artifact labeling, a sample set for training is obtained. Before the sample set is used for training a preset neural network model, map conversion is carried out according to the step S103 to obtain electroencephalogram characteristic maps of respective independent components of different electroencephalograms, and therefore the characteristic maps are used as input to train the neural network model to obtain an artifact identification model.

And S105, determining the preset components to which the independent components belong based on component probabilities corresponding to the preset components to which the independent components belong.

In an embodiment of the application, the predetermined components are used to characterize the class of components to which the signal belongs. Specifically, the predetermined components may be set as normal electroencephalogram signals and artifacts, wherein the artifacts may include electrooculogram, myoelectricity, electrocardiograms, and the like.

Specifically, the predetermined components corresponding to the respective individual components may be determined according to a predetermined condition. For example, the predetermined component to which the independent component belongs, that is, the component classification of the independent component is determined according to the maximum value of the component probability.

And S106, if the preset component to which any independent component belongs is an artifact, removing any independent component, and then reconstructing the remaining independent component to obtain a pure electroencephalogram signal with the artifact removed.

Specifically, a mixture matrix for the remaining independent components may be calculated using an independent component analysis ICA reconstruction algorithm, thereby reconstructing the remaining independent components from which the artifacts are removed.

The method comprises the steps of obtaining a multi-lead electroencephalogram signal to be processed, decomposing the multi-lead electroencephalogram signal into a plurality of independent components by utilizing a preset independent component algorithm, carrying out map conversion on the plurality of independent components to obtain an electroencephalogram characteristic map, identifying the electroencephalogram characteristic map according to a pre-constructed artifact identification model to obtain component probabilities corresponding to the independent components respectively belonging to various preset components, determining the preset components to which the independent components respectively belong based on the component probabilities corresponding to the independent components respectively belonging to the preset components, reconstructing the residual independent components after removing any one of the independent components by taking the preset component to which any one of the independent components belongs as an artifact to obtain a pure electroencephalogram signal without the artifact, and converting the electroencephalogram signal which is difficult to observe into a high-readability map by adopting a mode of identifying the artifact through a brain characteristic map, the method not only provides auxiliary data for visual analysis of the electroencephalogram signals, but also achieves the purpose of enriching information provided for the artifact identification model by increasing the number of the maps input to the artifact identification model, so that the artifact identification model can realize automatic identification of artifacts and finally remove artifact components, comprehensively consider the distribution characteristics of the artifacts in different maps, and retain a large amount of electroencephalogram information on the basis of automatic identification of various artifacts by retaining the local characteristics of the signals in different maps, thereby improving the artifact identification precision.

In some implementations, the electroencephalogram feature map includes a brain topographic map, a power spectral density PSD map, and the like. The brain topographic map is used for representing the distribution condition of brain signals of different channels in a brain area; a power Spectral density (psd) (power Spectral density) graph is a physical quantity that characterizes the power energy of a signal as a function of frequency. In specific application, the power spectrum density PSD diagram can be obtained by calculating the power spectrum of the independent component by using the power spectrum estimation of an Autoregressive (AR) model and then converting.

Specifically, the power spectrum of the same component of each channel can be extracted according to the obtained independent component, then spatial interpolation is carried out on each channel point in the international 10-20 electrode distribution plane diagram according to the obtained power spectrum value, the distribution of a continuous power spectrum on the surface of the cerebral cortex is constructed, and finally, the distribution image which can quantitatively reflect the brain function change is obtained according to the mapping of the value to different color representations, so that the complicated and changeable brain function change is changed into a popular and easy-to-understand graph. When the method is applied, according to different physical parameters and different display modes of pixel values, other topographic maps with different meanings can be formed, such as an electroencephalogram equipotential map reflecting spatial electroencephalogram potential distribution at any moment, BEAM (brain electric Activity map) data and corresponding contrasted BEAM are subjected to statistical processing, and the processed data are converted into a significant probability topographic map which takes test statistics as pixel values to be displayed, a percentage topographic map reflecting proportions among different frequency band combinations, an Activity topographic map which is continuously and dynamically displayed according to time sequence, and the like.

In some implementations, step S103 further includes: and performing Fourier transform and square calculation on the plurality of independent components to obtain Power Spectral Density (PSD) graphs corresponding to the plurality of independent components.

In some implementations, step S103 further includes: and respectively carrying out power spectrum extraction on the plurality of independent components according to the same components of the channel and carrying out spatial interpolation to construct a brain topographic map.

When applied, the power spectrum can be estimated using an auto-regression based AR model whose output is a weighted sum of the current input and the past output, using the following difference equation:

wherein, a_kTable (

k

1,2, … …, p) is the parameters of the AR model, p is the AR model order, u (n) is zero mean and variance σ²The smoothed white noise sequence of (a), x (n), is treated as a white noise sequence u (n), is generated by an all-pole filter of a transfer function. Thus, power spectral density p of x (n)_x(e^jω) Expressed as:

when in use, the brain map construction mode is as follows: firstly, extracting the power spectrum of the same component of each channel according to the obtained independent component, and then carrying out spatial interpolation on each channel point in the international 10-20 electrode distribution plane diagram according to the obtained power spectrum value to construct the distribution of a continuous power spectrum on the surface of the cerebral cortex. And finally, mapping into different color representations according to the size of the value.

In some implementation manners, before the step S104 identifies the electroencephalogram feature map according to the pre-constructed artifact identification model to obtain component probabilities that the independent components respectively belong to various types of predetermined components, as shown in fig. 1, the method further includes: step S1041 (not shown): determining a pattern for identifying the artifact; step S104, recognizing the electroencephalogram characteristic diagram according to the pre-constructed artifact recognition model to obtain component probabilities that a plurality of independent components respectively belong to various preset components and respectively correspond to the preset components, and further comprising the following steps: step S1042 (not shown in the figure): if the mode for identifying the artifacts is a mode for simultaneously identifying a plurality of artifacts, the artifact identification model is used for identifying the brain topographic map and/or the Power Spectral Density (PSD) map to obtain component probabilities corresponding to various preset components of the independent components.

In an embodiment of the application, the pattern of the identification of the artefacts is used to characterize information which selects the model employed for identifying the artefacts. In particular, a numerical label may be utilized to represent a pattern that identifies artifacts. For example: "0" is used to indicate that the pattern of recognition of artifacts is: a mode for simultaneously identifying a plurality of artifacts; "1" is used to indicate that the pattern of recognition of artifacts is: and identifying each type of artifact one by one.

The application scenarios of the artifact identification model provided in the embodiment of the present application are as follows: meanwhile, various artifacts are recognized, namely, the purpose of recognizing various artifacts through one-time input can be achieved by the artifact recognition model provided by the embodiment of the application, and compared with the model for recognizing single artifacts, the number of the used models is reduced, the time for recognizing various artifacts is shortened, and the artifact recognition efficiency is improved.

In implementations utilizing the artifact recognition model provided in step S1042, step S105 further includes:

step S1051 (not shown in the figure), determining a maximum value of the component probabilities of the plurality of independent components based on the component probabilities of the plurality of independent components belonging to the predetermined components respectively;

step S1052 (not shown in the figure) determines a predetermined component to which each of the plurality of independent components belongs, based on a maximum value of component probability of each of the plurality of independent components.

In the embodiment of the present application, the result of the artifact identification model provided in step S1042 is that the same independent component belongs to component probabilities corresponding to different predetermined components, that is, the artifact identification model provided in step S1042 has higher identification accuracy for a single independent component, so that after the component probabilities identified by the same independent component are sorted, the component to which the maximum value of the component probabilities points is directly determined as the component category of the independent component.

For example, assume that the independent components obtained in step S102 include: a1, B1, and C1, the output results of the artifact recognition model for a1 include: the probability that A1 belongs to the electro-oculogram signal is 25%, the probability that A1 belongs to the electromyogram signal is 65%, the probability that A1 belongs to the normal electroencephalogram signal is 30%, the probability that A1 belongs to the electrocardio signal is 0.5%, and A1 can be determined as the electromyogram by comparing the probabilities that A1 belongs to various artifacts.

In some implementations, the step S104 of recognizing the electroencephalogram feature map according to the pre-constructed artifact recognition model to obtain component probabilities that the plurality of independent components respectively belong to various predetermined components, further includes:

step S1043 (not shown in the figure), if the artifact identification mode is a mode for identifying each type of artifacts one by one, identifying the brain topographic map and the power spectral density PSD map by using the pre-constructed artifact identification models corresponding to each type of artifacts, and obtaining component probabilities corresponding to each of the predetermined components belonging to each of the independent components.

The embodiment of the application provides a mode of respectively identifying the same independent component by using multiple models, namely the component probabilities that the same independent component belongs to different artifacts can be identified only by inputting the same independent component into different models one by one; meanwhile, the types of the artifacts needing to be identified in different service scenes are different, and the adaptability based on service requirements is improved by using a special model for identifying a single artifact, so that the method is more beneficial to timely adjusting the adopted model along with the service scenes.

In some implementations applying step S1043, step S105 further includes:

step S1053 (not shown in the figure), determining a maximum value of the component probabilities of each of the predetermined components based on the component probabilities of the independent components belonging to each of the predetermined components;

step S1054, not shown), determining the predetermined component to which each of the plurality of independent components belongs according to the maximum value of the component probability of each of the predetermined components.

Because the pre-constructed artifact identification models corresponding to various types of artifacts are used for identification, the output result of the artifact identification model corresponding to the same type of artifacts includes component probabilities corresponding to the same type of artifacts respectively by a plurality of independent components, that is, the model provided in this embodiment cannot determine which of the independent components is the same type of artifacts. Therefore, in the embodiment of the present application, the output result of the artifact identification model corresponding to the same type of artifact includes component probabilities corresponding to the plurality of independent components respectively for the same type of artifact, and the maximum value of the component probabilities belonging to the same type of artifact of the output of the artifact identification model corresponding to the same type of artifact is obtained, so that the independent component to which the maximum value of the component probabilities of the same type of artifact points is determined as the same type of artifact.

For example, assume that the independent components decomposed in step S103 include: a2, B2 and C2, the output result of the model pair a2 for identifying the eye charge includes: the probability of a2 belonging to the electrooculogram is 25%, the probability of B2 belonging to the electrooculogram is 5%, and the probability of C2 belonging to the electrooculogram is 95%, then C2 is determined as the electrooculogram.

In summary, the embodiments of the present application provide two models for identifying artifacts: one model can identify various artifacts simultaneously, namely the purpose of identifying the artifacts of a plurality of independent components can be achieved by providing the model; the other is to provide a plurality of specialized models to respectively identify different artifacts, namely, a plurality of models are provided to respectively identify corresponding artifacts, and the model for specifically identifying a single artifact has wider application range and higher identification precision; meanwhile, the types of the artifacts needing to be identified in different service scenes are different, and the adaptability based on service requirements is improved by using a special model for identifying a single artifact, so that the method is more beneficial to timely adjusting the adopted model along with the service scenes.

In some implementations, before step S104, the method further includes, as shown in fig. 1:

step S1044 (not shown in the drawings), determining a neural network model for training based on a predetermined model framework, the neural network model comprising in sequence: the device comprises an input layer, a Stem layer, a first introduction layer, a first Reduction layer, a second introduction layer, a second Reduction layer, a third introduction layer, an average pooling layer, a drop layer and a Softmax layer.

Specifically, the model framework can be set according to business needs. For example, a convolutional neural network CNN framework, a recurrent neural network RNN framework, or the like may be used as the predetermined model framework.

When applied, the neural network model may include as shown in fig. 2 a: an input layer, a Stem layer, an initiation-A layer, a Reduction-A layer, an initiation-B layer, a Reduction-B layer, an initiation-C layer, an Average Pooling layer Average Pooling, a dropout layer and a Softmax layer. The Stem layer is shown in FIG. 2 b. FIG. 3 is a schematic view of the structures of the inclusion-A, Reduction-A, B and C layers. As shown in fig. 2, the Stem layer includes a plurality of convolutional layers in parallel, and a deeper network structure is obtained by replacing the sequential connection of the original convolutional layers and pooling. The structure is that convolution is carried out for multiple times and pooling is carried out for 2 times, and the pooling adopts a structure of parallel convolution and pooling to prevent the bottleneck problem. Stem was followed by 3 total 14 inclusion layers. Each inclusion layer is provided with 4 branches for input, convolution layers or pooling layers with different sizes are respectively used, and finally the convolution layers or pooling layers are spliced together in a characteristic dimension. Inside which four convolution branches are connected in parallel as shown in fig. 3. The inclusion structure has two main characteristics: one is to use convolution of 1x1 to perform the lifting dimension; the second is to perform convolution re-aggregation on multiple sizes simultaneously. The Reduction layers (such as the Reduction-A layer and the Reduction-B layer) among the three types of inclusion layers play a role in pooling, and a parallel structure is also used for preventing the bottleneck problem. Three convolution branches are connected in parallel inside the Reduction layer. The Reduction structure is internally provided with 3 branches for input, convolution layers or pooling layers with different sizes are respectively used, and finally the convolution layers or pooling layers are spliced together in characteristic dimensions. Before application, a data set is divided into a training set, a verification set and a test set in a ratio of 8:1: 1. The data set sequentially passes through 1 Stem layer, 4 introduction-A layers, 1 Reduction-A layer, 7 introduction-B layers, 1 Reduction-B layer, 3 introduction-C layers, 1 Average Pooling layer Average Pooling, 1 drop layer and an output layer Softmax. The artifact identification model detects the effectiveness of the network model by using a cross entropy loss function, adjusts the parameters of the inclusion-v 4 neural network model by using an Adam optimizer and a back propagation mode to update the parameters of each layer, and stops training for multiple times until the accuracy of the network and the output value of the loss function are stable, so that the trained inclusion-v 4 neural network model is obtained.

In some implementations, before the step of identifying the electroencephalogram feature map according to the pre-constructed artifact identification model by the component probability determination module to obtain component probabilities corresponding to each of a plurality of independent components belonging to each of various predetermined components, as shown in fig. 1, the method further includes:

step S1045 (not shown in the figure), determining a neural network model for training based on a predetermined model frame, where the neural network model sequentially includes an input layer, a first pooling layer, a channel merging layer, a parallel superposition residual neural network, a second pooling layer, and an output layer, where the parallel superposition residual neural network includes a plurality of parallel convolution blocks provided with an inner fast channel and an outer fast channel, any parallel convolution block includes a plurality of convolution layer branches, and a branch channel merging layer and a scaling factor, an output of the plurality of convolution layer branches is an input of the branch channel merging layer, and an output of the branch channel merging layer is an input of the scaling factor.

When applied, the parallel superposition residual error neural network can comprise one parallel volume block, 2 parallel volume blocks or a plurality of parallel volume blocks. Specifically, the number of parallel volume blocks may be set according to traffic requirements.

In one embodiment of the present application, the parallel superposition residual neural network comprises 3 parallel convolution blocks as shown in fig. 4. The artifact identification model shown in FIG. 4 includes an input layer, a first pooling layer, a channel merge layer, a first parallel volume block, a second parallel volume block, a third parallel volume block, a second pooling layer, and an output layer. The 3 parallel-connected volume blocks in the artifact identification model shown in fig. 4 are all provided with a pooling layer, and dimensionality reduction is performed through the map arranged on the pooling layer, so that the purposes of reducing the calculated amount and increasing invariance and robustness of the robust map are achieved. In addition, the arrangement of the scaling factors in each parallel convolution block meets the subsequent input requirement on the image processing flow, avoids the problem that the output result of the previous processing flow cannot meet the input requirement of the next processing flow, and improves the image processing effect. In the artifact identification model shown in fig. 4, the convolution kernels included in the parallel convolution blocks are: 1X1, 3X3 and 5X5, wherein each parallel convolution block adopts a parallel structure and then converges to a branch channel merging layer, so that the width of the neural network is enlarged; because the internal residual block and the external residual block are added in each parallel convolution block, the output of the previous layer can be used as the input of the next layer and can be input in an interlayer mode, and therefore the neural network can better fit the equivalent function; and the scaling factor layer is added after the convolution channel merging layer, so that the training speed of the convolution neural network is improved and the convolution neural network can be converged better.

Three convolutional layer branches are connected in parallel inside the first parallel convolutional block, the first convolutional layer branch is connected with 1 convolutional layer of 1X1 and an average pooling layer in series, the second convolutional layer branch is connected with 5 convolutional layers of 3X3 and an average pooling layer in series, the third convolutional layer branch is connected with 5 convolutional layers of 5X5 and an average pooling layer in series, and each convolutional layer inside the first parallel convolutional block comprises 32 convolutions; meanwhile, the second convolutional layer branch and the third convolutional layer branch respectively comprise two internal residual error blocks, an internal fast channel exists between the input of the first convolutional layer and the output of the second convolutional layer, and between the input of the fourth convolutional layer and the output of the fifth convolutional layer, the three convolutional layer branches are connected to a convolutional layer with a convolutional core of 1X1 and containing 16 convolutions after passing through a channel merging layer, and finally are connected to a scaling factor layer. An outer residual block is also included at the outermost side of the first parallel volume block, so that an outer fast path exists between the input of the first parallel volume block and the output of the scale factor layer. The structure of the inner part of the second parallel volume block is similar to that of the first parallel volume block, and the outermost side of the second parallel volume block also comprises an outer quick channel. Different from the above, each convolutional layer inside the second parallel convolutional block is 64 convolutions, three convolutional layer branches pass through one channel merging layer and are connected to 32 convolutional layers with convolution kernels of 1X1, and other structures are the same, which is not described herein again. The structure of the inner part of the third parallel rolling block is similar to that of the first parallel rolling block and the second parallel rolling block, and the outermost side of the third parallel rolling block also comprises an outer quick channel. Different from the above, each convolutional layer inside the third parallel convolutional block is 128 convolutions, and three convolutional layer branches inside the third parallel convolutional block are connected to a convolutional layer with convolution kernel of 1X1 and containing 64 convolutions after passing through a channel merging layer, and other structures are the same and are not described herein. The inner fast channels of the three parallel convolution blocks are arranged in the parallel branch, and the structure is that two convolution layers are skipped from the output end of the upper layer and are directly connected to the input end of the lower layer; the outer fast channel is outside the parallel convolution block and has a structure that the whole parallel convolution block is skipped from the input end of the parallel convolution block and is directly connected to the output end of the parallel convolution block. The inner and outer fast channels realize interlayer input, so that the convolutional neural network can better fit the equivalent function.

The last pooling layer in the artifact identification model shown in FIG. 4 is the second pooling layer, whose pooling kernel is 3X 3. The step length of all pooling layers in the artifact identification model is 2, a Relu nonlinear activation function is used, and initial weights are randomly generated according to Gaussian distribution. Before application, a data set is divided into a training set, a verification set and a test set according to the ratio of 8:1:1, the data set sequentially passes through a convolutional layer, a first pooling layer, a channel merging layer, a first parallel convolutional block, a second parallel convolutional block and a third parallel convolutional block from an input layer, and is finally connected to the second pooling layer and the output layer. During training, the effectiveness of the network model is detected by using a cross entropy loss function, parameters of the parallel superposition residual convolution neural network model are adjusted in an Adam optimizer and back propagation mode to update parameters of each layer, the circulation is performed for multiple times until the accuracy of the network and the output value of the loss function are stable, and the training is stopped, so that the trained parallel superposition residual convolution neural network model is obtained.

In an implementation manner including step S1041, as shown in fig. 1, the method further includes:

step S1046 (not shown in the figure), obtaining a sample set, wherein the sample set comprises electroencephalogram feature maps of a plurality of independent components respectively corresponding to a plurality of electroencephalograms, and component identifications of the electroencephalogram feature maps of the plurality of independent components respectively corresponding to the plurality of electroencephalograms;

and step S1047 (not shown in the figure), training the neural network model by using the sample set to obtain a trained artifact identification model.

Specifically, the obtaining process of the sample set may include: firstly, acquiring a plurality of electroencephalogram signals; secondly, processing the plurality of electroencephalogram signals by adopting an independent component ICA algorithm with reference to the step S102 to obtain independent components corresponding to the plurality of electroencephalogram signals respectively; thirdly, marking independent components corresponding to the electroencephalogram signals respectively by using preset sample marking tools, such as object marker, exe-iteye, BBox-Label-Tool and the like; and fourthly, carrying out map conversion on the independent components respectively corresponding to the plurality of electroencephalogram signals to obtain electroencephalogram characteristic maps of the independent components respectively corresponding to the plurality of electroencephalogram signals.

In the case of constructing the second initial neural model according to the above step S1044, the processing may be performed with reference to step S1043 and step S1044, which is not described herein again.

In some implementations, step S1047 further includes:

respectively training a plurality of candidate recognition models by using a sample set;

determining the respective accuracy of the candidate recognition models according to the training result;

and taking the candidate identification model with the highest accuracy as an artifact identification model.

Specifically, a grid optimization SVM algorithm of the neural network can be used for providing a plurality of groups of parameters for the neural network model, and the neural network model is subjected to parameter configuration according to the plurality of groups of parameters to obtain candidate recognition models configured with different parameters.

Specifically, the candidate recognition models with the highest accuracy may be screened out from the plurality of candidate recognition models and used as the trained artifact recognition model.

In the embodiment of the application, the candidate recognition model with the highest accuracy is directly loaded and used in the model application stage. When the method is applied, the parameters adopted by the candidate recognition model with the highest accuracy rate can be directly output, and therefore the artifact recognition model is obtained through configuration according to the parameters.

In some implementations, after step S101, the method further includes, as shown in fig. 5:

step S1011 (not shown in the figure), preprocessing the multi-lead electroencephalogram signals;

Specifically, the preprocessing generally includes filtering, dessication, baseline correction, and the like. Filtering to enable the reserved electroencephalogram data to fall into a preset frequency band range; the baseline correction causes the brain electrical data to fluctuate around the horizontal axis from being skewed to one side of the horizontal axis.

In an embodiment of the present application, the location channel locations are used to characterize the process of identifying the brain region where each electrode is located. Specifically, a brain electrical file (including EEG data that is acquired after being tested and a positioning file for describing channel coordinate information) may be acquired by predetermined analysis software, so that a channel pointed by each electrode and a brain electrical signal of the channel are positioned through analysis of the brain electrical file. For example, EEG data is acquired according to a 10-20 international standard lead system, and then FP1, FP2, F3, F4, Cz, P3, P4 and Oz are determined through the processing of multiple positioning channel positions, and FP1, FP2, F3, F4, Cz, P3, P4 and Oz correspond to EEG signals respectively.

In the present embodiment, the elimination of the unwanted electrodes is used to characterize the process of removing the reference electrodes. Because the surface of the human body does not have real zero potential, most parts with less relative motion and not influenced by other bioelectric fields are selected as the positions of the reference electrodes, during data acquisition, REF is used as the reference electrode, the potential difference between all other electrodes and the reference electrode REF is used as the recorded value of the electrode, the information of the reference electrode REF and the grounding electrode GND cannot be used for subsequent analysis, and the reference electrode and the grounding electrode are required to be removed in subsequent processing.

In the embodiments of the present application, re-reference is used to characterize the process of reducing the potential difference between different electrodes. This is because the acquired data is affected by the reference of different electrode positions, for example, when a Cz electrode is used as a reference electrode, the potential difference recorded by the electrode at a position closer to Cz is naturally small, while the potential difference recorded by the electrode at a position farther away is larger, so that re-reference is needed to reduce the influence. For example, with the bilateral mastoid electrode points as the re-reference locations, the location is not susceptible to head motion, the corresponding electrodes are TP9 and TP10, all electrodes of the left brain will have the recorded values of TP9 subtracted, and all electrodes of the right brain will have the recorded values of TP10 subtracted.

In the embodiment of the present application, filtering is used to characterize a process of extracting a desired frequency band. The filtering of EEG data is typically done to reduce interference or noise signals while maintaining as much as possible the authenticity of the signal to be observed, so the EEG data can be acquired by filtering for analysis into the frequency band of interest. When the method is applied, if the data is collected, band-pass filtering is carried out between 0.1Hz and 50Hz, but 0.1Hz to 45Hz is generally used, and frequency bands corresponding to different rhythms are respectively extracted on the basis for analysis.

In the embodiment of the present application, the reduced sampling rate is used to characterize the process of controlling the number of data points recorded per unit time. For example, reducing the sampling rate to 100Hz, i.e. recording 100 data points within 1s, can effectively increase the speed of the calculation.

In the embodiment of the application, the baseline correction is used for characterizing the process controlling the difference of the starting values of different segments. Since the superposition averaging technique is often used when analyzing data, segmentation inevitably results in different starting values of different segments. Therefore, each piece of data can be at the same starting point by baseline correction.

In an embodiment of the present application, data segmentation is used to characterize the process of retrieving a data segment to be processed. For example, the data used for data acquisition is 8 minutes each in an open-closed eye state, and the data is segmented using a 5s time window for subsequent rapid and accurate analysis of the data.

In order to further explain the artifact removal method for multi-lead electroencephalogram signals provided by the embodiment of the present application, the following describes in detail the workflow of the system shown in fig. 6. The workflow of the system shown in FIG. 6 includes a pre-model application phase and a model application phase. The pre-application stage of the model includes steps S201 to S204.

Step S201, constructing a sample set, wherein the sample set comprises artifact forming and normal electroencephalogram forming diversity. The sample set may be obtained by referring to steps S401 to S406 shown in fig. 7. S401, acquiring an original electroencephalogram signal with artifacts; the original brain electrical signal can be acquired by a brain-computer interface (such as a brain electrical cap); s402, preprocessing an original electroencephalogram signal; the preprocessing may refer to step S1011 above, and perform preprocessing on the original electroencephalogram signal according to at least one of channel position locating, useless electrode deleting, re-referencing, filtering, sampling rate reduction, baseline correction, and data segmentation; or processing the original electroencephalogram signals according to the sequence of channel position positioning, useless electrode deleting, re-reference, filtering, sampling rate reduction, baseline correction and data segmentation; step S403, processing the electroencephalogram signal preprocessed according to the step S102, namely decomposing ICA into 19 different components; step S404, uploading the 19 different components in the step S403 to a labeling tool so that a user can label the 19 independent components with artifacts; s405, acquiring artifact components and normal electroencephalogram through a marking tool; and S406, circularly executing the steps S401 to S405 to obtain an artifact component set and a normal electroencephalogram component set.

And S202, constructing a neural network model. The step may be obtained by referring to the step S1044, and sequentially includes: the neural network model comprises an input layer, a Stem layer, a first introduction layer, a first Reduction layer, a second introduction layer, a second Reduction layer, a third introduction layer, an average pooling layer, a dropout layer and a Softmax layer; or obtaining a neural network model sequentially comprising an input layer, a first pooling layer, a channel merging layer, a parallel superposition residual neural network, a second pooling layer and an output layer by referring to the step S1045, wherein the parallel superposition residual neural network comprises a plurality of parallel convolution blocks provided with an inner fast channel and an outer fast channel, any parallel convolution block comprises a plurality of convolution layer branches, a branch channel merging layer and a scaling factor, the output of the plurality of convolution layer branches is the input of the branch channel merging layer, and the output of the branch channel merging layer is the input of the scaling factor.

And step S203, training a neural network model. When the method is applied, different sets of parameters can be respectively configured for the neural network model by referring to the step S1047 to obtain a plurality of candidate recognition models, then the sample set obtained in the step S201 is used for respectively training the plurality of candidate recognition models, the respective accuracy rates of the plurality of candidate recognition models are determined according to the training results, and the candidate recognition model with the highest accuracy rate is used as an artifact recognition model;

and step S204, outputting the optimal neural network model parameters, namely taking the parameters adopted by the candidate identification model with the highest accuracy in the step S203 as the optimal neural network model parameters.

The model application phase includes steps S301 to S307.

And S301, inputting original electroencephalogram data.

And S302, preprocessing the electroencephalogram signals. When the method is applied, operations of channel position locating, unnecessary electrode deleting, re-referencing, filtering, sampling rate reduction, baseline correction and data segmentation can be sequentially performed on the electroencephalogram signals according to the step S1011.

And step S303, constructing a brain map and a PSD curve graph of the independent components. When the method is applied, the electroencephalogram signal preprocessed according to the step S302 is decomposed into independent components by using an independent component ICA algorithm, for example, the electroencephalogram signal is decomposed into 19 independent components in the step S305 before the reference model is applied, and then the 19 independent components are subjected to map conversion, so that the aim of constructing an independent component brain topographic map and a PSD curve graph is fulfilled.

And S304, obtaining an artifact identification model according to the model parameters of the optimal neural network output in the previous stage of model application, and inputting the brain topographic map and the PSD curve graph obtained in the step S304 into the artifact identification model for classification. The classification process may be determined with reference to step S104 and its implementation, step S105 and its implementation.

And step S305, removing artifacts according to the classification result in the step S304.

And S306, reconstructing the data without the artifacts.

And S307, outputting the electroencephalogram signals with the artifacts removed.

Yet another embodiment of the present application provides an artifact removing apparatus for multi-lead brain electrical signals, as shown in fig. 8, the apparatus 80 includes: an electroencephalogram signal acquisition module 801, an electroencephalogram signal decomposition module 802, a signal map conversion module 803, a component probability determination module 804, a component category determination module 805, and an artifact removal processing module 806.

An electroencephalogram signal acquisition module 801, configured to acquire a multi-lead electroencephalogram signal to be processed;

the electroencephalogram signal decomposition module 802 is used for decomposing the multi-lead electroencephalogram signal into a plurality of independent components by using a set independent component analysis algorithm;

a signal map conversion module 803, configured to perform map conversion on the plurality of independent components to obtain an electroencephalogram characteristic map;

the component probability determination module 804 is used for identifying the electroencephalogram characteristic map according to the pre-constructed artifact identification model to obtain component probabilities corresponding to the independent components belonging to various preset components respectively;

a component classification determination module 805 configured to determine predetermined components to which the plurality of independent components belong based on component probabilities corresponding to the respective predetermined components to which the plurality of independent components belong;

and the artifact removal processing module 806 is configured to, if the predetermined component to which any one of the independent components belongs is an artifact, remove the any one of the independent components and then reconstruct the remaining independent components to obtain a pure electroencephalogram signal with the artifact removed.

Further, the electroencephalogram feature map at least comprises:

brain topography and power spectral density PSD plots.

Still further, the signal map conversion module comprises:

the first conversion submodule is used for performing Fourier transform and square calculation on the plurality of independent components respectively to obtain Power Spectral Density (PSD) graphs corresponding to the plurality of independent components respectively; and/or

And the second conversion submodule is used for performing power spectrum extraction and spatial interpolation on the plurality of independent components according to the same components of the channels respectively to construct a brain map.

Further, before the step of identifying the electroencephalogram feature map by the component probability determination module according to the pre-constructed artifact identification model to obtain component probabilities that the plurality of independent components respectively belong to various types of predetermined components and respectively correspond to the predetermined components, the device further comprises: the mode determining module is used for determining a mode for identifying the artifact; the component probability determination module comprises: and the first component probability determining submodule is used for identifying the brain topographic map and/or the Power Spectral Density (PSD) map by using the artifact identification model if the artifact identification mode is a mode for simultaneously identifying multiple artifacts, so as to obtain component probabilities corresponding to the fact that the independent components respectively belong to various preset components.

Further, the component classification determination module includes:

Further, the component probability determination module includes:

and the second component probability determining submodule is used for identifying the brain topographic map and the Power Spectral Density (PSD) map by utilizing the pre-constructed artifact identification models respectively corresponding to various artifacts if the artifact identification mode is a mode for identifying various artifacts one by one, so as to obtain component probabilities respectively corresponding to various preset components belonging to the independent components.

Further, the component classification determination module includes:

the second maximum value determining submodule is used for determining the maximum value of the component probability of each preset component based on the component probability that each independent component belongs to each preset component;

and the network model training submodule is used for training the neural network model by utilizing the sample set to obtain an artifact identification model with determined parameters.

Further, the network model training sub-module includes:

and the artifact identification model determining unit is used for taking the candidate identification model with the highest accuracy as the artifact identification model.

The artifact removing device for multi-lead electroencephalogram signals of the present embodiment can execute the artifact removing method for multi-lead electroencephalogram signals shown in the first embodiment of the present application, which is similar to the original implementation and is not described herein again.

Another embodiment of the present application provides a terminal, including: the processor executes the computer program to realize the artifact removal method of the multi-lead electroencephalogram signal.

In particular, the processor may be a CPU, general purpose processor, DSP, ASIC, FPGA or other programmable logic device, transistor logic device, hardware component, or any combination thereof. Which may implement or perform the various illustrative logical blocks, modules, and circuits described in connection with the disclosure. A processor may also be a combination of computing functions, e.g., comprising one or more microprocessors, a DSP and a microprocessor, or the like.

In particular, the processor is coupled to the memory via a bus, which may include a path for communicating information. The bus may be a PCI bus or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc.

The memory may be, but is not limited to, a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, an EEPROM, a CD-ROM or other optical disk storage, optical disk storage (including compact disk, laser disk, optical disk, digital versatile disk, blu-ray disk, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer.

Optionally, the memory is used for storing codes of computer programs for executing the scheme of the application, and the processor is used for controlling the execution. The processor is used for executing the application program codes stored in the memory so as to realize the action of the artifact removing device for the multi-lead electroencephalogram signals provided by the above embodiment.

Yet another embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, which when executed by a processor, implements the above-described artifact removal method for multi-lead brain electrical signals.

Yet another embodiment of the present application provides a chip for performing an artifact removal method according to the multi-lead electroencephalogram signal.

Yet another embodiment of the present application provides a brain-computer interface including a wearable device applying a chip for performing an artifact removal method according to the multi-lead brain electrical signal described above.

The above-described embodiments of the apparatus are merely illustrative, and the units illustrated as separate components may or may not be physically separate, may be located in one place, or may be distributed over a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

One of ordinary skill in the art will appreciate that all or some of the steps, systems, and methods disclosed above may be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). The term computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, as is well known to those of ordinary skill in the art. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can accessed by a computer. In addition, communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media as known to those skilled in the art.

While the present invention has been described with reference to the preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. An artifact removal method for a multi-lead electroencephalogram signal is characterized by comprising the following steps:

acquiring a multi-lead electroencephalogram signal to be processed;

identifying the electroencephalogram characteristic diagram according to a pre-constructed artifact identification model to obtain component probabilities of a plurality of independent components respectively belonging to various preset components;

2. The method of claim 1, wherein the step of performing map transformation on the plurality of independent components to obtain an electroencephalogram feature map comprises:

and/or

And respectively carrying out power spectrum extraction on the plurality of independent components according to the same components of the channel and carrying out spatial interpolation to construct the brain topographic map.

3. The method of claim 2,

before the step of identifying the electroencephalogram characteristic diagram according to the pre-constructed artifact identification model to obtain component probabilities that a plurality of independent components respectively belong to various preset components and respectively correspond to the preset components, the method further comprises the following steps:

determining a pattern for identifying the artifact;

the step of identifying the electroencephalogram characteristic diagram according to the pre-constructed artifact identification model to obtain component probabilities that a plurality of independent components respectively belong to various preset components and respectively correspond to the preset components comprises the following steps:

and if the mode for identifying the artifacts is a mode for simultaneously identifying a plurality of artifacts, identifying the brain topographic map and/or the power spectral density PSD map by using the artifact identification model to obtain component probabilities corresponding to various preset components of the independent components.

4. The method according to claim 3, wherein the step of determining the component classes to which the independent components belong based on component probabilities corresponding to the predefined components belonging to the classes, comprises:

determining the maximum value of the component probability of each of the plurality of independent components based on the component probability corresponding to each of the plurality of independent components belonging to each of the predetermined types of components;

5. The method according to claim 3, wherein the step of identifying the electroencephalogram feature map according to the pre-constructed artifact identification model to obtain component probabilities that a plurality of independent components respectively belong to various types of predetermined components and respectively correspond to the predetermined components comprises the steps of:

and if the mode for identifying the artifacts is a mode for identifying various artifacts one by one, identifying the brain topographic map and the power spectral density PSD map by utilizing artifact identification models corresponding to various pre-constructed artifacts respectively to obtain component probabilities corresponding to various preset components of the independent components.

6. The method according to claim 5, wherein the step of determining the predetermined components to which the independent components belong based on component probabilities corresponding to the respective types of the predetermined components to which the independent components belong comprises:

7. The method according to claim 1, wherein before the step of identifying the electroencephalogram feature map according to the pre-constructed artifact identification model to obtain component probabilities that the independent components respectively belong to various types of predetermined components, the method further comprises:

8. The method according to claim 1, wherein before the step of identifying the electroencephalogram feature map according to the pre-constructed artifact identification model to obtain component probabilities that the independent components respectively belong to various types of predetermined components, the method further comprises:

the method comprises the steps of determining a neural network model for training based on a preset model frame, wherein the neural network model sequentially comprises an input layer, a first pooling layer, a channel merging layer, a parallel superposition residual error neural network, a second pooling layer and an output layer, the parallel superposition residual error neural network comprises a plurality of parallel convolution blocks provided with an inner rapid channel and an outer rapid channel, any parallel convolution block comprises a plurality of convolution layer branches, a branch channel merging layer and a scaling factor, the output of the plurality of convolution layer branches is the input of the branch channel merging layer, and the output of the branch channel merging layer is the input of the scaling factor.

9. The method according to claim 7 or 8, characterized in that the method further comprises:

acquiring a sample set, wherein the sample set comprises electroencephalogram characteristic maps of a plurality of independent components corresponding to a plurality of electroencephalogram signals respectively, and component identifications of the electroencephalogram characteristic maps of the plurality of independent components corresponding to the plurality of electroencephalogram signals respectively;

and training the neural network model by using the sample set to obtain the trained artifact identification model.

10. The method of claim 9, wherein the step of training the neural network model with the sample set to obtain the trained artifact recognition model comprises:

configuring the neural network model based on a plurality of preset groups of parameters to obtain a plurality of candidate identification models;

respectively training a plurality of candidate recognition models by using the sample set; determining the respective accuracy of the candidate recognition models according to the training result;

and taking the candidate recognition model with the highest accuracy as the trained artifact recognition model.

11. The method of claim 1, wherein after the step of acquiring the multi-lead brain electrical signal to be processed, the method further comprises:

preprocessing the multi-lead electroencephalogram signals;

the preprocessing comprises at least one of channel position locating, useless electrode deleting, re-reference, filtering, sampling rate reduction, baseline correction and data segmentation; or

The preprocessing comprises channel position positioning, useless electrode deletion, re-reference, filtering, sampling rate reduction, baseline correction and data segmentation in sequence.

12. An artifact removing device for multi-lead electroencephalogram signals, comprising:

the component probability determination module is used for identifying the electroencephalogram characteristic diagram according to a pre-constructed artifact identification model to obtain component probabilities corresponding to a plurality of independent components which respectively belong to various preset components;

13. An electronic device, comprising:

one or more processors;

a memory;

one or more applications, wherein the one or more applications are stored in the memory and configured for execution by the one or more processors, the one or more programs configured to: performing the method according to any one of claims 1-11.

14. A computer-readable storage medium, on which a computer program is stored, which program, when being executed by a processor, is adapted to carry out the method of any one of claims 1-11.

15. A chip for performing the method according to any of claims 1-11.

16. A brain-computer interface comprising a wearable device applying a chip for performing the method according to any of claims 1-11.