KR102285546B1 - Deep learning-based eeg feature normalization system and method - Google Patents

Abstract

The present invention relates to a deep learning-based EEG feature normalization system and a method thereof. The present invention learns a measured EEG from a user and detects an EEG feature using an EEG feature detection unit, removes EEG in the resting state from an EEG feature in which the EEG in the resting state is detected, and includes an EEG feature normalization unit that normalizes the EEG feature from which the EEG in the resting state has been removed to improve analysis efficiency by providing a customized EEG analysis for an EEG signal of an individual user.

Description

DEEP LEARNING-BASED EEG FEATURE NORMALIZATION SYSTEM AND METHOD}

The present invention relates to a system and method for normalizing EEG features based on deep learning, and to a technology that enables EEG analysis for individual users.

The electrical activity of the brain reflected in EEG is determined by neurons, glia cells, and blood-brain barrier, and mainly occurs by neurons. Glia cells, which account for half of the brain's weight, control the flow of ions and molecules at the synapse, the area where neurons are connected, and play a role in maintaining, supporting, and repairing structures between neurons. EEG changes due to glial cells and blood-brain barrier occur little by little and slowly, whereas EEG changes due to neuronal activity are large, rapid, and diverse.

EEG shows a waveform that vibrates in a very complex pattern. In general, when observing EEG, a power spectrum classified according to frequency is used. Power spectrum analysis assumes that EEG is a linear combination of simple vibrations that vibrate at a specific frequency, and decomposes each frequency component in this signal to determine the size will be displayed

In general, they are called delta waves, theta waves, alpha waves, beta waves, and gamma waves depending on the shape of the frequency, and are also subdivided into low alpha waves, middle alpha waves, and high alpha waves.

Based on these brain waves, core human organs study the brain, from various brain-related diseases such as brain tumors, stroke, epilepsy, and dementia to various mental diseases such as schizophrenia, multiple personality and emotional disorders, anesthesia, pain, perception, cognition, He also conducts research related to brain development, education, emotions, meditation, and consciousness.

However, research on the brain is difficult to estimate theoretically on a specific state, so it is possible to give meaning and interpret it only through direct clinical trials and analysis.

In order to solve this problem, with the recent development of deep learning technology, a machine is making a development in brain engineering research by deriving theoretical estimations for a specific state instead. It is being used in various fields, such as deriving accurate results for various feedbacks by analyzing brain waves, judging people's intentions, and deriving types of mental disorders.

However, since each person's brain waves have different characteristics, it is somewhat difficult to derive the intended results by analyzing each person's brain waves using deep learning focused on a specific target.

Therefore, it is urgent to develop a personalized EEG analysis system to solve this problem.

delete

Republic of Korea Patent Publication No. 10-1724939 (2017.04.07.)

A deep learning-based EEG feature normalization system according to an aspect of the present invention includes: an EEG feature detecting unit for detecting EEG features by learning EEG measured from a user based on deep learning; a resting EEG removing unit for removing the EEG in the resting state from the detected EEG characteristic amount; and an EEG feature normalization unit for normalizing a feature amount of the EEG from which the EEG in the resting state is removed.

Preferably, it may further include an EEG classification unit for classifying the user's intention contained in the EEG feature amount input as the normalized EEG feature amount.

Preferably, the characteristic quantity of the EEG may be an EEG indicated by the user's intention.

Preferably, the resting EEG removing unit may use the additionally measured resting EEG of the user.

Preferably, the EEG feature normalization unit may generate the normalized EEG feature amount based on the EEG in a resting state.

Preferably, the EEG feature normalization unit may learn the EEG feature amount from which the EEG in the resting state is removed using a backpropagation algorithm.

Preferably, the EEG classification unit may classify one of a plurality of user intentions and output it as a probability value.

A deep learning-based EEG feature normalization method according to another aspect of the present invention includes: a normalized feature amount detection step of learning EEG measured from a user based on deep learning to detect EEG feature amounts; a resting EEG removing step in which the EEG in the resting state is removed from the detected EEG characteristic amount; an EEG feature amount normalization step in which the EEG feature amount from which the EEG in the resting state is removed is normalized; and an EEG classification step in which a user's intention contained in the input EEG feature amount is recognized and classified as the normalized EEG feature amount.

According to the present invention, the EEG deviation of individual users can be minimized by normalizing the EEG feature quantity, and the analysis efficiency can be improved by providing customized EEG analysis for the EEG signals of individual users who have not been trained, which was a limitation of the existing deep learning structure. can be improved

1 is a schematic diagram showing a deep learning environment for brain wave analysis according to an embodiment.
2 is a configuration diagram of a deep learning-based EEG feature normalization system apparatus according to an embodiment.
3 is a block diagram illustrating a learning process for normalizing EEG feature quantities according to an embodiment.
4 is a block diagram illustrating a process of normalizing EEG feature quantities according to an embodiment.
5 is a block diagram illustrating a process of analyzing an EEG for a user according to an exemplary embodiment.
6 is a flowchart illustrating a deep learning-based EEG feature quantity normalization method according to an embodiment.

Hereinafter, a deep learning-based EEG feature normalization system and method according to the present invention will be described in detail with reference to the accompanying drawings. In this process, the thickness of the lines or the size of the components shown in the drawings may be exaggerated for clarity and convenience of explanation. In addition, terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to the intention or custom of the operator. Therefore, definitions of these terms should be made based on the content throughout this specification.

Objects and effects of the present invention can be naturally understood or made clearer by the following description, and the objects and effects of the present invention are not limited only by the following description. In addition, in describing the present invention, when it is determined that a detailed description of a known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted.

1 is a schematic diagram showing a deep learning environment for brain wave analysis according to an embodiment.

As shown in Figure 1, the deep learning environment for EEG analysis according to an embodiment detects a feature by learning the user's EEG based on deep learning, and removes the resting EEG from the detected feature to generate a resting EEG. The feature quantity of the removed EEG can be normalized. In this case, EEG analysis may be performed for individual users with the normalized EEG feature amount.

In this case, the detected EEG feature amount may be detected through a long short-term memory (LSTM). Here, the LSTM is a modified form of a recurrent neural network (RNN) that receives information from the previous stage from the network of the next stage. can be derived.

In more detail, it will be described along with the configuration of the EEG feature normalization system device for deep learning-based EEG analysis.

2 is a block diagram of a deep learning-based EEG feature normalization system apparatus according to an embodiment.

As shown in FIG. 2 , the configuration of the deep learning-based EEG feature normalization system device according to an embodiment includes a normalized feature detector 100 , a resting EEG remover 300 , and an EEG feature normalization unit 500 . and may further include an EEG classification unit 700 .

The normalized feature amount detection unit 100 may detect the feature amount of the EEG by learning the EEG measured from the user based on deep learning. The characteristic quantity of the EEG is the EEG indicated by the user's intention. In this case, the measured characteristic quantity of the EEG may include the user's intention and the EEG in a resting state at the same time.

The resting EEG removing unit 300 may remove the EEG in the resting state from the characteristic amount of the detected EEG. Here, the brain wave in the resting state may be the brain wave in a state in which the user's thoughts and intentions are not embedded. That is, if the EEG in the resting state is removed from the EEG feature, only the EEG displayed by the user's movement or thoughts can be detected. Also, the resting EEG removing unit 300 may use the additionally measured resting EEG of the user.

The EEG feature normalization unit 500 may normalize the EEG feature amount from which the EEG in the resting state is removed. Since the EEG feature quantity from which the EEG in the resting state is removed can represent the EEG feature distribution for the same intention, the intention in a specific area can be more clearly derived by minimizing the overlapping intention in the EEG appearing according to the user's intention. there is. Also, it is possible to generate a characteristic quantity of the normalized EEG based on the EEG in the resting state.

The EEG classification unit 700 may classify the user's intention contained in the input EEG characteristic quantity as the normalized EEG characteristic quantity. Here, the input EEG characteristic quantity may be a resting state EEG characteristic quantity detected from a new user. That is, by detecting the EEG in the resting state from the input EEG feature amount, it is possible to derive the EEG feature amount according to the user's intention. It is possible to classify one of a plurality of user intentions and output this as a probability value.

Using this, it is possible to provide customized EEG analysis for individual users. In this case, the normalized EEG feature amount can be learned through a backpropagation algorithm.

,

), 두 개의 은닉층(

,

), 및 두 개의 출력층(

,

)으로 구성된 2-2-2 다층 퍼셉트론이 존재할 때, 입력층을 I라고 가정하면 I층의 i번째 노드와 I+1층의 j번째 노드를 연결하는 가중치 w를

으로 정의할 수 있다.Here, the backpropagation algorithm is one of the deep learning functions, and it is an algorithm that minimizes the error with two input layers (

,

), two hidden layers (

,

), and two output layers (

,

), assuming that the input layer is I, the weight w connecting the i-th node of the I-layer and the j-th node of the I+1 layer exists.

can be defined as

)는 다음 수학식으로 표현될 수 있다.At this time, the sigmoid function that is the applied active function (

) can be expressed by the following equation.

[Equation 1]

In Equation 1, z may be an input function used for the hidden layer.

The input function z between the input layer and the hidden layer can be expressed by the following equation.

[Equation 2]

)에 대한 시그모이드 활성 함수는

이고, 두번째 은닉층(

)에 대한 시그모이드 활성 함수는

일 수 있다.In Equation 2, the first hidden layer (

) is the sigmoid activation function for

, and the second hidden layer (

) is the sigmoid activation function for

can be

The input function z between the hidden layer and the output layer can be expressed by the following equation.

[Equation 3]

)에 대한 시그모이드 활성 함수는

이고, 두번째 출력층(

)에 대한 시그모이드 활성 함수는

일 수 있다.In Equation 3, the first output layer (

) is the sigmoid activation function for

, and the second output layer (

) is the sigmoid activation function for

can be

In this case, the error (J), which is a cost function of the output value, may be expressed by the following equation.

[Equation 4]

,

과 실제값 y₁, y₂는 고정된 값이므로 J₁과 J₂는 가중치

와

를 변수로 가지는 함수가 된다.In Equation 4, y ₁ and y ₂ are actual values corresponding to each node for training data. At this time, the input value

,

and the actual values y ₁ and y ₂ are fixed values, so J ₁ and J ₂ are weights

Wow

becomes a function with as a variable.

An error can be derived by calculating the output value of the output layer using the given weight value, and by excluding the value differentiated by each weight for the derived error from the existing weight, the number of learning times or when a value within the given tolerance range is reached Iteratively, the error can be minimized.

Here, the equation for changing the weight may be expressed as the following equation.

[Equation 5]

In equation 5, J is the sum of the error _totoal the node influence.

In other words, the input function (z) between the output increase and the hidden layer can be expressed as the following equation by the chain rule of differentiation.

[Equation 6]

에서 J_total 의 값은 J1이다. 따라서, J_total은 다음 수학식으로 계산될 수 있다.In Equation 6, the starting node of backpropagation is

In , the value of _{J total is J1.} Accordingly, J _total can be calculated by the following equation.

[Equation 7]

The differential value for the weight of the error used for weight change in backpropagation can be expressed only with the active function value of the starting node of the backpropagation, the active function value of the arrival node, and the actual value, which can be expressed by the following equation can

[Equation 8]

를

로 변형하면, 역전파에 의해 변경되는

는 다음 수학식으로 표현될 수 있고, 동일한 방법으로

를 구할 수 있다.In Equation 8,

cast

, which is changed by backpropagation

can be expressed by the following equation, and in the same way

can be obtained

[Equation 9]

Therefore, the error can be reduced to a minimum by the backpropagation algorithm.

The above-described backpropagation algorithm is described by setting arbitrary nodes to describe an embodiment, and is not necessarily limited to the above-described number of nodes, input layer, hidden layer, and output layer. In addition, the above-described backpropagation algorithm has been described as an embodiment, and should not be limited to the above-described algorithm.

3 and 4 are diagrams illustrating a learning process for normalizing EEG feature quantities according to an embodiment.

3 is a block diagram illustrating a learning process for normalizing EEG feature quantities according to an embodiment.

As shown in FIG. 3 , the learning process for normalizing EEG features according to an embodiment may include an EEG feature detecting module 110 and a user intention identification module 120 in the normalized feature detecting unit 100 , , it is possible to learn the result from which the user's intention is predicted by generating label data for the predicted result of the user's intention. A backpropagation algorithm may be used for learning.

4 is a block diagram illustrating a process of normalizing EEG feature quantities according to an embodiment.

As shown in FIG. 4 , in the EEG feature quantity normalization process according to an embodiment, the EEG feature quantity detection module 110 may detect the EEG feature quantity based on deep learning. In this case, the characteristic amount of the detected EEG may be an EEG in which the user's intention is implied.

The user intention identification module 120 may identify the user's intention contained in the EEG feature amount, and may normalize the EEG feature amount according to the identified EEG.

5 is a block diagram illustrating a process of analyzing an EEG for a user according to an exemplary embodiment.

As shown in FIG. 5 , the process of analyzing EEG for a user according to an embodiment may include a resting EEG feature amount detection module 710 and a user intention prediction module 720 in the EEG classification unit.

The resting EEG feature detection module 710 measures EEG from a plurality of users to detect the measured EEG, but removes the EEG in the resting state from the detected EEG characteristic amount to include the user's intention. can be detected. This is because the frequency, size, and pattern of brainwaves appear differently for each individual user, so by detecting the characteristic quantity of the user's brainwave in a stable state, it is possible to derive the characteristic quantity of the brainwave with the intention of the individual user.

The user intention prediction module 720 may derive the EEG characteristic quantity according to the user's intention by detecting the EEG in the resting state from the EEG characteristic quantity detected for an individual user.

In addition, it is possible to learn the EEG feature amount from which the normalized EEG feature amount is excluded through a backpropagation algorithm, and classify the EEG feature amount into the normalized EEG feature amount. This can facilitate the analysis of the EEG of a new user because the EEG characteristic of the resting period is removed for each user and classification is performed as a normalized EEG containing an intention-related EEG. That is, since the normalized EEG feature amount of a new user may appear similar to the normalized EEG feature amount used for learning, it is possible to provide a rapid EEG analysis for a new user by classifying the similar normalized EEG feature amount.

6 is a flowchart illustrating a deep learning-based EEG feature quantity normalization method according to an embodiment.

As shown in FIG. 6 , the deep learning-based EEG feature normalization method according to an embodiment includes a normalized feature amount detection step (S100), a resting EEG removal step (S300), an EEG feature amount normalization step (S500), and an EEG It may include a classification step (S700).

In the normalized feature amount detection step ( S100 ), the EEG measured from the user may be learned based on deep learning, and the EEG feature amount may be detected.

In the resting EEG removing step ( S300 ), the EEG in the resting state may be removed from the detected EEG feature amount.

In the EEG feature amount normalization step ( S500 ), the EEG feature amount from which the EEG in the resting state is removed may be normalized.

In the EEG classification step ( S700 ), the user's intention contained in the input EEG characteristic quantity as the normalized EEG characteristic quantity may be grasped.

Although the present invention has been described in detail through representative embodiments above, those of ordinary skill in the art to which the present invention pertains will understand that various modifications are possible within the limits without departing from the scope of the present invention with respect to the above-described embodiments. will be. Therefore, the scope of the present invention should not be limited to the described embodiments and should be defined by all changes or modifications derived from the claims and equivalent concepts as well as the claims to be described later.

100: normalized feature detection unit 300: resting EEG removal unit
500: EEG feature normalization unit 700: EEG classification unit
110: EEG feature detection module 120: User intention identification module
710: resting EEG feature amount detection module 720: user intention prediction module

Claims (8)

an EEG feature detecting unit that learns EEG measured from a user based on deep learning to detect EEG features of a user's intention in which the user's intention is embedded and EEG features in a resting state in which the user's intention is not embedded;
A resting EEG removing unit for deriving the EEG feature of the user intention representing only the user intention by removing the EEG feature of the resting state from the detected EEG feature of the user's intention (here, the EEG is the frequency and size of the EEG );
an EEG feature normalization unit for normalizing the size of an EEG generated by removing the EEG feature amount in a resting state from the derived EEG feature amount of the user's intention; and
Deep learning-based brain wave feature normalization system, characterized in that it comprises an EEG classifying unit for classifying user intention from the EEG measured as the EEG feature of the normalized user intention.
delete delete delete According to claim 1,
The EEG feature normalization unit is a deep learning-based EEG feature normalization system, characterized in that it minimizes the overlapping intention of the user according to the difference in the size of the EEG.
According to claim 1,
The EEG feature normalization unit is a deep learning-based EEG feature normalization system, characterized in that it learns the EEG feature amount from which the EEG feature amount in the resting state is removed with a backpropagation algorithm.
According to claim 1,
The EEG classification unit classifies the intentions of a plurality of users, and a deep learning-based EEG feature normalization system, characterized in that outputting the classified user intentions as a probability value.
In the deep learning-based EEG feature normalization method performed in the deep learning-based EEG feature normalization system of claim 1,
An EEG feature detecting step of learning EEG measured from the user based on deep learning to detect the EEG feature of the user's intention in which the user's intention is embedded and the EEG feature of the resting state in which the user's intention is not embedded;
a resting EEG removing step of deriving the EEG feature of the user's intention representing only the user's intention by removing the EEG in the resting state from the detected EEG feature of the user's intention (here, the EEG feature is the frequency and size of the EEG) ;
an EEG feature quantity normalization step of normalizing the size of EEG generated by the removal of the EEG feature quantity in the resting state from the EEG feature quantity of each EEG from which the resting state EEG has been removed; and
Classifying the EEG intention measured as the EEG feature of the normalized user intention, and a deep learning-based EEG feature normalization method comprising an EEG classification step of outputting the user intention as a probability value.

Images