KR102532757B1 - Apparatus for predicting dissolved gas concentration in aqueous solution based on Raman spectral signal and method therefor - Google Patents

Abstract

An apparatus for predicting dissolved gas concentration is provided. The apparatus includes a peak extraction unit for extracting peak data of a wavelength corresponding to a type of gas from the Raman spectrum data when Raman spectrum data for an aqueous solution is input, and a prediction model based on the peak data included in the aqueous solution. and a concentration predicting unit that calculates a predicted value of the gas concentration and derives and outputs the gas concentration from the calculated predicted value.

Description

Apparatus for predicting dissolved gas concentration in aqueous solution based on Raman spectral signal and method therefor}

The present invention relates to a technique for predicting dissolved gas concentration, and more particularly, to an apparatus and method for predicting dissolved gas concentration in an aqueous solution based on a Raman spectroscopic signal.

The Raman scattering method enables the detection of a small amount of gas in an aqueous solution that could not be previously measured due to the advantage of a highly sensitive detection technology. However, the relationship between the gas concentration and the Raman spectrum of the Raman scattering method is not linear, and the relationship cannot be easily derived. Therefore, in the past, the Raman spectrum data derived through experiments and the concentration data of dissolved gas were compared 1:1 to derive the relationship. However, these methods require numerous experimental databases.

Korean Patent Publication No. 2001-0022918 published on March 26, 2001 (Name: Real-time gas analysis method and device)

An object of the present invention is to provide an apparatus and method for predicting the concentration of dissolved gas in an aqueous solution based on a Raman spectroscopic signal.

An apparatus for predicting the dissolved gas concentration of an aqueous solution based on a Raman spectroscopy signal according to a preferred embodiment of the present invention for achieving the above object is input when the Raman spectrum data for the aqueous solution is input, the concentration of the gas from the Raman spectrum data A peak extractor for extracting peak data of a wavelength corresponding to the type, calculating a predicted value of the gas concentration contained in the aqueous solution through a predictive model based on the peak data, and deriving and outputting the gas concentration from the calculated predicted value It includes a concentration prediction unit.

The apparatus extracts learning peak data of a wavelength corresponding to the type of gas and gas concentration corresponding to the learning peak data from Raman spectrum data, which is experimental data, to prepare learning data, and sets some of the learning data as inspection data. and a model generator for learning a predictive model to calculate a predicted value of the gas concentration contained in the aqueous solution using the learning data, and examining the learning result using the test data.

The model generator controls the number of hidden layers and hidden nodes of the predictive model until both the accuracy of each of the learning data and the test data satisfies a preset condition, while learning the predictive model using the learning data And, it is characterized in that the inspection of the learning performance for the predictive model using the inspection data is repeated.

에 따라 상기 예측모델의 출력인 예측값과 경계 레이블의 차이인 경계손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 경계 최적화를 수행하고, 상기 Lboaderselected는 경계 손실함수이고, 상기 Oi는 상기 예측모델의 예측값이고, 상기 vi는 상기 출력값에 대응하는 경계 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스인 것을 특징으로 한다. The model generating unit converts the gas concentration corresponding to the learning peak data into a one-hot encoding vector based on a predetermined boundary value, sets it as a boundary label, inputs the learning peak data into a prediction model, and the prediction model converts the learning peak data into a prediction model. If a prediction value is calculated by performing a plurality of operations in which a plurality of inter-layer weights are applied to the peak data, the boundary loss function

According to , boundary optimization is performed to modify the weight of the prediction model so that the boundary loss, which is the difference between the predicted value, which is the output of the prediction model, and the boundary label, is minimized, the Lboaderselected is a boundary loss function, and the Oi is the prediction model A prediction value, vi is a boundary label corresponding to the output value, and i is an index corresponding to an output node of an output layer of the prediction model.

에 따라 상기 예측모델의 출력인 예측값과 경계 레이블의 차이인 경계 손실 및 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드의 출력값과 증강 레이블과의 차이를 나타내는 증강 손실을 포함하는 경계증강 손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 경계증강 최적화를 수행하고, 상기 Lboaderenhanced는 경계증강 손실함수이고, 상기 Oi는 상기 예측모델의 출력층의 출력인 예측값이고, 상기 vi는 상기 예측값에 대응하는 경계 레이블이고, 상기 hij는 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드의 출력값이고, 상기 rij는 상기 복수의 은닉노드의 출력값에 대응하는 증강 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스이고, 상기 j는 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드에 대응하는 인덱스인 것을 특징으로 한다. The model generator converts the gas concentration corresponding to the peak data for learning into a one-hot encoding vector based on the preset boundary value, sets it as the boundary label, and sets the reference vector corresponding to the hidden vector of the peak data for learning as an augmented label and inputs the peak data for learning to a prediction model, and the prediction model performs a plurality of operations in which a weight between a plurality of layers is applied to the peak data for learning, and output values of a plurality of hidden nodes of the last hidden layer and an output layer If the predicted value, which is the output of , is calculated, the boundary augmentation loss function

According to the boundary enhancement loss, including the boundary loss that is the difference between the predicted value and the boundary label, which is the output of the prediction model, and the enhancement loss representing the difference between the output value of a plurality of hidden nodes of the last hidden layer of the prediction model and the augmented label, Minimum Boundary enhancement optimization is performed to modify the weight of the predictive model so that Lboaderenhanced is a boundary enhancement loss function, Oi is a predicted value that is an output of the output layer of the predictive model, and vi is a boundary label corresponding to the predicted value. , wherein hij is an output value of a plurality of hidden nodes of the last hidden layer of the prediction model, rij is an augmentation label corresponding to an output value of the plurality of hidden nodes, and i is an output node of an output layer of the prediction model a corresponding index, and j is an index corresponding to a plurality of hidden nodes of the last hidden layer of the prediction model.

The model generator sets the gas concentration corresponding to the learning peak data as a numerical label, inputs the learning peak data to a prediction model, and the prediction model applies a plurality of inter-layer weights to the learning peak data. When a predictive value is calculated by performing the operation of the predictive model, numerical optimization is performed to modify the weight of the predictive model so that the numerical loss, which is the difference between the predicted value and the numerical label of the predictive model, is minimized.

에 따라 상기 예측모델의 출력값과 수치 레이블의 차이인 수치 손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 수치 최적화를 수행하며, 상기 Lvaule는 수치손실함수이고, 상기 Oi는 상기 예측모델의 출력인 예측값이고, 상기 Ci는 상기 예측값에 대응하는 수치 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스인 것을 특징으로 한다. The model generator is a numerical loss function

Numerical optimization is performed to modify the weight of the predictive model so that the numerical loss, which is the difference between the output value of the predictive model and the numerical label, is minimized according to A prediction value, Ci is a numerical label corresponding to the prediction value, and i is an index corresponding to an output node of an output layer of the prediction model.

및 2063

이고, 상기 가스가 아세트산염(Acetate)이면 상기 파장은 928

이고, 상기 가스가 낙산염(Butyrate)이면, 상기 파장은 877

인 것을 특징으로 한다. If the gas is CO, the wavelength is 1948

and 2063

, and if the gas is acetate, the wavelength is 928

, and if the gas is butyrate, the wavelength is 877

It is characterized by being

The apparatus further includes a spectrum measuring unit configured to emit light through a probe in the aqueous solution, measure scattering of the emitted light, acquire Raman spectrum data, and provide the acquired Raman spectrum data.

An apparatus for predicting the concentration of dissolved gas in an aqueous solution based on a Raman spectroscopic signal according to a preferred embodiment of the present invention for achieving the above object is for learning the wavelength corresponding to the type of gas from Raman spectral data, which is experimental data. A model generator for learning a predictive model, which is an artificial neural network algorithm, to prepare learning data by extracting peak data and a gas concentration corresponding to the peak data for learning, and using the learning data to calculate a predicted value of the gas concentration contained in the aqueous solution include

The model generator sets the gas concentration corresponding to the learning peak data as a numerical label, inputs the learning peak data to a prediction model, and the prediction model applies a plurality of inter-layer weights to the learning peak data. When a predictive value is calculated by performing the operation of the predictive model, numerical optimization is performed to modify the weight of the predictive model so that the numerical loss, which is the difference between the predicted value and the numerical label of the predictive model, is minimized.

에 따라 상기 예측모델의 출력값과 수치 레이블의 차이인 수치 손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 수치 최적화를 수행하며, 상기 Lvaule는 수치 손실함수이고, 상기 Oi는 상기 예측모델의 출력인 예측값이고, 상기 Ci는 상기 예측값에 대응하는 수치 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스인 것을 특징으로 한다. The model generator is a numerical loss function

Numerical optimization is performed to modify the weight of the predictive model so that the numerical loss, which is the difference between the output value of the predictive model and the numerical label, is minimized according to A prediction value, Ci is a numerical label corresponding to the prediction value, and i is an index corresponding to an output node of an output layer of the prediction model.

에 따라 상기 예측모델의 출력인 예측값과 경계 레이블의 차이인 경계손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 경계 최적화를 수행하고, 상기 Lboaderselected는 경계 손실함수이고, 상기 Oi는 상기 예측모델의 예측값이고, 상기 vi는 상기 출력값에 대응하는 경계 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스인 것을 특징으로 한다. The model generating unit converts the gas concentration corresponding to the learning peak data into a one-hot encoding vector based on a preset boundary value, sets it as a boundary label, inputs the learning peak data into a prediction model, and the prediction model converts the learning peak data into a prediction model. If a predicted value is calculated by performing a plurality of operations in which a plurality of inter-layer weights are applied to the data, the boundary loss function

According to , boundary optimization is performed to modify the weight of the prediction model so that the boundary loss, which is the difference between the predicted value, which is the output of the prediction model, and the boundary label, is minimized, the Lboaderselected is a boundary loss function, and the Oi is the prediction model A prediction value, vi is a boundary label corresponding to the output value, and i is an index corresponding to an output node of an output layer of the prediction model.

에 따라 상기 예측모델의 출력인 예측값과 경계 레이블의 차이인 경계 손실 및 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드의 출력값과 증강 레이블과의 차이를 나타내는 증강 손실을 포함하는 경계증강 손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 경계증강 최적화를 수행하고, 상기 Lboaderenhanced는 경계증강 손실함수이고, 상기 Oi는 상기 예측모델의 출력층의 출력인 예측값이고, 상기 vi는 상기 예측값에 대응하는 경계 레이블이고, 상기 hij는 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드의 출력값이고, 상기 rij는 상기 복수의 은닉노드의 출력값에 대응하는 증강 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스이고, 상기 j는 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드에 대응하는 인덱스인 것을 특징으로 한다. The model generator converts the gas concentration corresponding to the peak data for learning into a one-hot encoding vector based on the preset boundary value, sets it as the boundary label, and sets the reference vector corresponding to the hidden vector of the peak data for learning as an augmented label and inputs the peak data for learning to a prediction model, and the prediction model performs a plurality of operations in which a weight between a plurality of layers is applied to the peak data for learning, and output values of a plurality of hidden nodes of the last hidden layer and an output layer If the predicted value, which is the output of , is calculated, the boundary augmentation loss function

According to the boundary enhancement loss, including the boundary loss that is the difference between the predicted value and the boundary label, which is the output of the prediction model, and the enhancement loss representing the difference between the output value of a plurality of hidden nodes of the last hidden layer of the prediction model and the augmented label, Minimum Boundary enhancement optimization is performed to modify the weight of the predictive model so that Lboaderenhanced is a boundary enhancement loss function, Oi is a predicted value that is an output of the output layer of the predictive model, and vi is a boundary label corresponding to the predicted value. , wherein hij is an output value of a plurality of hidden nodes of the last hidden layer of the prediction model, rij is an augmentation label corresponding to an output value of the plurality of hidden nodes, and i is an output node of an output layer of the prediction model a corresponding index, and j is an index corresponding to a plurality of hidden nodes of the last hidden layer of the prediction model.

The model generator controls the number of hidden layers and hidden nodes of the predictive model until both the accuracy of each of the learning data and the test data satisfies a preset condition, while learning the predictive model using the learning data And, it is characterized in that the inspection of the learning performance for the predictive model using the inspection data is repeated.

The device emits light through a probe in an aqueous solution, measures the scattering of the emitted light, acquires Raman spectrum data, and provides a spectrum measurement unit for providing the acquired Raman spectrum data, and a gas measurement unit from the Raman spectrum data. A peak extractor for extracting peak data of a wavelength corresponding to the type; and based on the peak data, a predicted value of the gas concentration contained in the aqueous solution is calculated through the prediction model, and the gas concentration is derived and output from the calculated predicted value. It further includes a concentration prediction unit to.

A method for predicting the dissolved gas concentration of an aqueous solution based on a Raman spectroscopy signal according to a preferred embodiment of the present invention for achieving the above object is to emit light through a probe in the aqueous solution by the spectrum measuring unit, Acquiring Raman spectrum data by measuring the scattering of light, extracting peak data of a wavelength corresponding to the type of gas from the Raman spectrum data by a peak extraction unit, and predicting by a concentration prediction unit based on the peak data Calculating a predicted value of the gas concentration included in the aqueous solution through a model, and deriving and outputting a gas concentration from the predicted value calculated by the concentration predicting unit.

The method includes extracting learning data and inspection data from experimental data by the model generator before acquiring the Raman spectrum data, and setting the number of hidden layers and hidden nodes of the predictive model by the model generator. And, the model generator performs learning on a predictive model having a set number of hidden layers and hidden nodes using the training data, and checks the learning performance of the predictive model (PM) using the test data, Calculating, by a model generator, the accuracy of the training data and the accuracy of the test data, wherein the accuracy of the training data and the accuracy of the test data are both higher than the accuracy of the test data by the model generator and reach a threshold value Determining whether a condition equal to or above is satisfied, and if the condition is satisfied as a result of the determination, determining, by the model generating unit, the number of hidden layers and hidden nodes according to a current setting.

In addition, after the step of determining whether the condition is satisfied, if the condition is not satisfied as a result of the determination, the model generating unit resets the number of hidden layers and hidden nodes, and then calculating the accuracy and repeating the step of determining whether the condition is satisfied.

In the method, before the step of acquiring the Raman spectrum data, the model generator extracts learning peak data of a wavelength corresponding to the type of gas and gas concentration corresponding to the learning peak data from Raman spectrum data, which is experimental data, and extracts learning data. preparing a gas concentration corresponding to the learning peak data as a numerical label, the model generating unit inputting the learning peak data into a predictive model, and the predictive model Calculating a predicted value by performing a plurality of calculations to which weights between a plurality of layers are applied to the peak data for learning, and the model generating unit making the prediction such that the numerical loss, which is the difference between the predicted value and the numerical label of the predictive model, is minimized. and performing numerical optimization to modify the weights of the model.

에 따라 상기 예측모델의 출력값과 수치 레이블의 차이인 수치 손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 수치 최적화를 수행하며, 상기 Lvalue는 수치손실함수이고, 상기 Oi는 상기 예측모델의 출력인 예측값이고, 상기 Ci는 상기 예측값에 대응하는 수치 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스인 것을 특징으로 한다. The step of performing the numerical optimization is a numerical loss function by the model generator

Numerical optimization is performed to modify the weight of the predictive model so that the numerical loss, which is the difference between the output value of the predictive model and the numerical label, is minimized according to A prediction value, Ci is a numerical label corresponding to the prediction value, and i is an index corresponding to an output node of an output layer of the prediction model.

에 따라 상기 예측모델의 출력인 예측값과 경계 레이블의 차이인 경계손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 경계 최적화를 수행하는 단계를 더 포함한다. 여기서, 상기 Lboaderselected는 경계 손실함수이고, 상기 Oi는 상기 예측모델의 출력인 예측값이고, 상기 vi는 상기 예측값에 대응하는 경계 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스이다. Before acquiring the Raman spectrum data, the model generating unit converts the gas concentration corresponding to the learning peak data into a one-hot encoding vector based on a predetermined boundary value and sets it as a boundary label; The step of inputting peak data for learning into a predictive model, the step of calculating a predicted value by the predictive model performing a plurality of calculations to which weights between a plurality of layers are applied to the peak data for learning, and the boundary loss function performed by the model generator

The method further includes performing boundary optimization of modifying weights of the prediction model so that boundary loss, which is a difference between a predicted value output from the prediction model and a boundary label, is minimized according to the prediction model. Here, Lboaderselected is a boundary loss function, Oi is a prediction value that is an output of the prediction model, vi is a boundary label corresponding to the prediction value, and i is an index corresponding to an output node of an output layer of the prediction model. .

에 따라 상기 예측모델의 출력인 예측값과 경계 레이블의 차이인 경계 손실 및 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드의 출력값과 증강 레이블과의 차이를 나타내는 증강 손실을 포함하는 경계증강 손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 경계증강 최적화를 수행하는 단계를 더 포함한다. 여기서, 상기 Lboaderenhanced는 경계증강 손실함수이고, 상기 Oi는 상기 예측모델의 출력인 예측값이고, 상기 vi는 상기 예측값에 대응하는 경계 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스이고, 상기 hij는 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드의 출력값이고, 상기 rij는 상기 복수의 은닉노드의 출력값에 대응하는 증강 레이블이고, 상기 j는 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드에 대응하는 인덱스인 것을 특징으로 한다. In the method, after the step of performing the boundary optimization, the model generating unit converts the gas concentration corresponding to the peak data for learning into a one-hot encoding vector based on the preset boundary value, sets it as the boundary label, and sets the peak data for learning. Setting a reference vector corresponding to a hidden vector as an augmented label, inputting the peak data for learning by the model generator to a prediction model, and applying a weight between a plurality of layers to the peak data for learning by the prediction model Calculating the output values of the plurality of hidden nodes of the last hidden layer and the predicted value that is the output of the output layer by performing a plurality of operations, and the model generator performs a boundary enhancement loss function

According to the boundary enhancement loss, including the boundary loss that is the difference between the predicted value and the boundary label, which is the output of the prediction model, and the enhancement loss representing the difference between the output value of a plurality of hidden nodes of the last hidden layer of the prediction model and the augmented label, Minimum and performing boundary enhancement optimization of modifying the weights of the predictive model so that Here, Lboaderenhanced is a boundary enhancement loss function, Oi is a prediction value that is an output of the prediction model, vi is a boundary label corresponding to the prediction value, and i is an index corresponding to an output node of an output layer of the prediction model. , wherein hij is an output value of a plurality of hidden nodes of the last hidden layer of the prediction model, rij is an augmentation label corresponding to the output value of the plurality of hidden nodes, and j is a plurality of the last hidden layer of the prediction model It is characterized in that the index corresponding to the hidden node of.

및 2063

이고, 상기 가스가 아세트산염(Acetate)이면 상기 파장은 928

이고, 상기 가스가 낙산염(Butyrate)이면, 상기 파장은 877

인 것을 특징으로 한다. Here, if the gas is CO, the wavelength is 1948

and 2063

, and if the gas is acetate, the wavelength is 928

, and if the gas is butyrate, the wavelength is 877

It is characterized by being

A method for predicting the concentration of dissolved gas in an aqueous solution based on a Raman spectroscopy signal according to a preferred embodiment of the present invention for achieving the above object is the model generating unit corresponding to the type of gas from Raman spectral data, which is experimental data. preparing learning data by extracting peak data for learning of a wavelength of a wavelength and a gas concentration corresponding to the peak data for learning, and setting the gas concentration corresponding to the peak data for learning as a numerical label by the model generator; The step of inputting the peak data for learning into a predictive model by a model generator, the step of calculating a predictive value by performing a plurality of calculations in which a weight between a plurality of layers is applied to the peak data for the predictive model, and generating the model The method may further include performing numerical optimization of modifying a weight of the predictive model so that a numerical loss, which is a difference between the predicted value and the numerical label of the predictive model, is minimized.

에 따라 상기 예측모델의 출력값과 수치 레이블의 차이인 수치 손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 수치 최적화를 수행하며, 상기 Lvalue는 수치손실함수이고, 상기 Oi는 상기 예측모델의 출력인 예측값이고, 상기 Ci는 상기 예측값에 대응하는 수치 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스인 것을 특징으로 한다. The step of performing the numerical optimization is a numerical loss function by the model generator

Numerical optimization is performed to modify the weight of the predictive model so that the numerical loss, which is the difference between the output value of the predictive model and the numerical label, is minimized according to A prediction value, Ci is a numerical label corresponding to the prediction value, and i is an index corresponding to an output node of an output layer of the prediction model.

에 따라 상기 예측모델의 출력인 예측값과 경계 레이블의 차이인 경계손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 경계 최적화를 수행하는 단계를 더 포함한다. 여기서, 상기 Lboaderselected는 경계 손실함수이고, 상기 Oi는 상기 예측모델의 출력인 예측값이고, 상기 vi는 상기 예측값에 대응하는 경계 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스이다. In the method, after the step of preparing learning data and before the step of setting the gas concentration as a numerical label, the model generator converts the gas concentration corresponding to the peak data for learning into a one-hot encoding vector based on a preset boundary value, The step of setting a boundary label, the step of inputting the peak data for learning by the model generation unit to a prediction model, and the step of the prediction model performing a plurality of calculations in which weights between a plurality of layers are applied to the peak data for learning, thereby predicting a predicted value. Calculating, and the model generator boundary loss function

The method further includes performing boundary optimization of modifying weights of the prediction model so that boundary loss, which is a difference between a predicted value output from the prediction model and a boundary label, is minimized according to the prediction model. Here, Lboaderselected is a boundary loss function, Oi is a prediction value that is an output of the prediction model, vi is a boundary label corresponding to the prediction value, and i is an index corresponding to an output node of an output layer of the prediction model. .

에 따라 상기 예측모델의 출력인 예측값과 경계 레이블의 차이인 경계 손실 및 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드의 출력값과 증강 레이블과의 차이를 나타내는 증강 손실을 포함하는 경계증강 손실이 최소가 되도록 상기 예측모델의 가중치를 수정하는 경계증강 최적화를 수행하는 단계를 더 포함한다. 여기서, 상기 Lboaderenhanced는 경계증강 손실함수이고, 상기 Oi는 상기 예측모델의 출력인 예측값이고, 상기 vi는 상기 예측값에 대응하는 경계 레이블이고, 상기 i는 상기 예측모델의 출력층의 출력노드에 대응하는 인덱스이고, 상기 hij는 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드의 출력값이고, 상기 rij는 상기 복수의 은닉노드의 출력값에 대응하는 증강 레이블이고, 상기 j는 상기 예측모델의 마지막 은닉계층의 복수의 은닉노드에 대응하는 인덱스이다. In the method, after the step of performing the boundary optimization, the model generating unit converts the gas concentration corresponding to the peak data for learning into a one-hot encoding vector based on the preset boundary value, sets it as the boundary label, and sets the peak data for learning. Setting a reference vector corresponding to a hidden vector as an augmented label, inputting the peak data for learning by the model generator to a prediction model, and applying a weight between a plurality of layers to the peak data for learning by the prediction model Calculating the output values of the plurality of hidden nodes of the last hidden layer and the predicted value that is the output of the output layer by performing a plurality of operations, and the model generator performs a boundary enhancement loss function

According to the boundary enhancement loss, including the boundary loss that is the difference between the predicted value and the boundary label, which is the output of the prediction model, and the enhancement loss representing the difference between the output value of a plurality of hidden nodes of the last hidden layer of the prediction model and the augmented label, Minimum and performing boundary enhancement optimization of modifying the weights of the predictive model so that Here, Lboaderenhanced is a boundary enhancement loss function, Oi is a prediction value that is an output of the prediction model, vi is a boundary label corresponding to the prediction value, and i is an index corresponding to an output node of an output layer of the prediction model. , wherein hij is an output value of a plurality of hidden nodes of the last hidden layer of the prediction model, rij is an augmentation label corresponding to the output value of the plurality of hidden nodes, and j is a plurality of the last hidden layer of the prediction model It is an index corresponding to the hidden node of .

The method includes: after the step of performing the numerical optimization, the spectrum measuring unit emits light through a probe in the aqueous solution, and the emitted light is scattered to obtain Raman spectrum data, and the peak extraction unit acquires Raman spectrum data. Extracting peak data of a wavelength corresponding to the type of gas from data; calculating, by a concentration predicting unit, a predicted value of the gas concentration included in the aqueous solution through a prediction model based on the peak data; The method further includes deriving and outputting the gas concentration from the calculated predicted value.

According to the present invention, the dissolved gas concentration of the aqueous solution can be accurately predicted through machine learning based on the amplification amount of a specific wavelength of the Raman spectroscopy signal. Accordingly, the dissolved gas concentration can be obtained in real time.

1 is a diagram for explaining the configuration of an apparatus for predicting the dissolved gas concentration of an aqueous solution based on a Raman spectroscopy signal according to an embodiment of the present invention.
2 is a block diagram for explaining the configuration of a prediction unit for predicting the concentration of an aqueous solution from Raman spectrum data according to an embodiment of the present invention.
3 is a diagram for explaining the configuration of a prediction model for predicting gas concentration according to an embodiment of the present invention.
4 is a diagram for explaining nodes of a prediction model for predicting gas concentration according to an embodiment of the present invention.
5 is a flowchart for explaining a method of generating a prototype of a predictive model according to an embodiment of the present invention.
6 is a graph for explaining a method of generating a prototype of a predictive model according to an embodiment of the present invention.
7 is a flowchart for explaining a method for learning a predictive model according to an embodiment of the present invention.
8 is a graph showing Raman spectrum data for an aqueous solution containing acetate and butyrate in experimental data according to an embodiment of the present invention.
9 is a graph showing Raman spectrum data for an aqueous solution containing a mixture of acetate and butyrate and carbon monoxide in experimental data according to an embodiment of the present invention.
10 is a graph showing Raman spectrum data in which the concentration of gas contained in an aqueous solution is known in experimental data according to an embodiment of the present invention.
11 is a flowchart illustrating boundary learning for a predictive model according to an embodiment of the present invention.
12 is a diagram for explaining boundary learning for a predictive model according to an embodiment of the present invention.
13 is a flowchart for explaining numerical learning for a predictive model according to an embodiment of the present invention.
14 is a flowchart for explaining a method for predicting a dissolved gas concentration in an aqueous solution based on a Raman spectroscopy signal according to an embodiment of the present invention.
15 is a diagram illustrating a computing device according to an embodiment of the present invention.

Prior to the detailed description of the present invention, the terms or words used in this specification and claims described below should not be construed as being limited to a common or dictionary meaning, and the inventors should use their own invention in the best way. It should be interpreted as a meaning and concept corresponding to the technical idea of the present invention based on the principle that it can be properly defined as a concept of a term for explanation. Therefore, the embodiments described in this specification and the configurations shown in the drawings are only the most preferred embodiments of the present invention, and do not represent all of the technical ideas of the present invention, so various equivalents that can replace them at the time of the present application. It should be understood that there may be water and variations.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. At this time, it should be noted that the same components in the accompanying drawings are indicated by the same reference numerals as much as possible. In addition, detailed descriptions of well-known functions and configurations that may obscure the gist of the present invention will be omitted. For the same reason, some components in the accompanying drawings are exaggerated, omitted, or schematically illustrated, and the size of each component does not entirely reflect the actual size.

First, an apparatus for predicting the dissolved gas concentration of an aqueous solution based on a Raman spectroscopy signal according to an embodiment of the present invention will be described. 1 is a diagram for explaining the configuration of an apparatus for predicting the dissolved gas concentration of an aqueous solution based on a Raman spectroscopy signal according to an embodiment of the present invention. Referring to FIG. 1 , an apparatus for predicting the dissolved gas concentration of an aqueous solution based on a Raraman spectroscopic signal according to an embodiment of the present invention (PA: Prediction Apparatus, hereinafter abbreviated as 'prediction apparatus') is a probe 11 It includes a spectrum measuring unit 10 including a ) and a predicting unit 20.

As shown in FIG. 1 , the spectrum measurement unit 10 emits light through the probe 11 in the aqueous solution Q and measures the scattering of the emitted light to generate Raman spectrum data. These Raman spectrum data are provided to the prediction unit 20 .

The prediction unit 20 predicts the concentration of the aqueous solution from the Raman spectrum data through a prediction model (PM), which is an artificial neural network algorithm.

Then, the configuration of the prediction unit 20 will be described in more detail. 2 is a block diagram for explaining the configuration of a predictor for predicting the concentration of an aqueous solution from Raman spectrum data according to an embodiment of the present invention. Referring to FIG. 2 , the predictor 20 includes a model generator 100 , a peak extractor 200 and a concentration predictor 300 .

The model generator 100 prepares learning data from experimental data. That is, the model generating unit 100 may prepare learning data by extracting learning peak data of a wavelength corresponding to the type of gas and gas concentration corresponding to the learning peak data from Raman spectrum data, which is experimental data. The model generating unit 100 performs machine learning/deep learning using the prepared learning data to generate a predictive model (PM) to calculate a predicted value of the gas concentration contained in the aqueous solution. In addition, the model generating unit 100 may set some of the prepared learning data as test data and inspect a learning result using the test data. The learning method of the model generator 100 will be described in more detail below.

및 2063

이다. 또한, 가스가 아세트산염(Acetate)이면 파장은 928

이고, 가스가 낙산염(Butyrate)이면, 파장은 877

이다. When Raman spectrum data for an aqueous solution is input from the spectrum measuring unit 10, the peak extraction unit 200 extracts peak data of a wavelength (Raman shift) corresponding to the type of gas from the Raman spectrum data. Here, when the gas is CO, the wavelength is 1948

and 2063

am. Also, if the gas is acetate, the wavelength is 928

, and the gas is butyrate, the wavelength is 877

am.

The concentration prediction unit 300 calculates a predicted value of the gas concentration included in the aqueous solution through a predictive model based on the peak data extracted by the peak extractor 200, and derives and outputs the gas concentration from the calculated predicted value. As such, the present invention uses a predictive model (PM) from Raman spectrum data derived through a single measurement of the spectrum measurement unit 10 to measure the dissolved gas concentration of an aqueous solution derived by taking a sample through a conventional complex experiment. Predictable. Accordingly, the dissolved gas concentration can be obtained in real time.

Then, the configuration of the prediction model (PM) for predicting the gas concentration according to an embodiment of the present invention will be described in detail. 3 is a diagram for explaining the configuration of a prediction model (PM) for predicting gas concentration according to an embodiment of the present invention. 4 is a diagram for explaining a node of a prediction model (PM) for predicting gas concentration according to an embodiment of the present invention.

In an embodiment of the present invention, the predictive model (PM) is composed of a plurality of layers, and these multiple layers include an input layer (IL: Input Layer), at least one hidden layer (HL: Hidden Layer, HL1 to HLk) and an output It includes the layer (OL: Output Layer).

In addition, each of the plurality of layers IL, HL, and OL includes a plurality of nodes. For example, as shown, the input layer IL includes a number of input nodes i1 to ia, and the output layer OL includes one output node O. In addition, among the hidden layers (HL), the first hidden layer (HL1) includes b nodes (g1 to gb), and the last hidden layer, the k-th hidden layer (HLk) includes j nodes (h1 to hj) can do.

A plurality of nodes of a plurality of layers (IL, HL, OL) individually perform calculations. Nodes of different layers are connected by channels (shown as dotted lines) having a weight (W). In other words, the calculation result of one node is weighted and becomes the input of the next layer node. The configuration of any one node ND is shown in detail in FIG. 4 . An operation of the node ND will be described with reference to FIG. 4 . Each of the output values of a plurality of (n) nodes of the previous layer of the node ND is X=[X1, X2, . . . , Xn] with weights W=[W1, W2, . . . , Wn] is input, and after summing them all, the function F is taken on the sum value. Here, the function F is called an activation function or a transfer function. These activation functions include step function, sign function, sigmoid, hyperbolic tangent (tanh), exponential linear unit (ELU), rectified linear unit (ReLU), leaky ReLU, Maxout, Minout, and Softmax functions can be exemplified.

An output of the node ND is expressed in Equation 1 below.

의 값이 임계치 보다 작을 때 해당 노드가 활성화되지 않도록 하는 역할을 한다. The unexplained variable θ is a threshold or bias, and this threshold is

It serves to prevent the corresponding node from being activated when the value of is smaller than the threshold.

An operation performed by the node ND according to Equation 1 is performed as follows. As an example, it is assumed that the number of nodes in the layer prior to the node ND is three. Accordingly, the output of each of the three nodes of the previous layer is input with three weights W1, W2, and W3 applied to each of the three inputs (n=3) X1, X2, and X3 for the corresponding node ND. Accordingly, the node ND receives a value obtained by multiplying the three inputs X1, X2, and X3 by weights W1, W2, and W3, sums them all, and substitutes the summed value into a transfer function to calculate an output. . Specifically, suppose inputs [X1, X2, X3] = 0.5, -0.3, -0.1, and weights [W1, W2, W3] = 4, 5, 4. Also, for convenience of description, assuming that the activation function is a sign function, that is, 'sgn()', an output value is calculated as shown in Equation 2 below.

As described in FIG. 4, any one node of any one layer of the predictive model (PM) also receives a value obtained by applying a weight (W) to the outputs of a plurality of nodes of the previous layer. Then, the corresponding node sums the inputs and calculates the output value of the corresponding node by performing an operation by an activation function on the summed value. The output value may be an input of the next layer or a final output value of the predictive model (PM). Accordingly, when peak data is input, the prediction model (PM) performs a plurality of operations in which weights between a plurality of layers (IL, HL, and OL) are applied to the input peak data to predict the gas concentration of the aqueous solution. can be calculated.

More specifically, when the concentration prediction unit 300 receives peak data from the peak extraction unit 200, the peak data is converted into binary data, and the input node (IL) of the input layer (IL) of the prediction model (PM) After dividing according to the number of i1 to ia), input is distributed to a plurality of input nodes i1 to ia of the input layer IL.

Then, each of the plurality of first hidden nodes g1 to gb of the first hidden layer HL1 is distributed to a plurality of input nodes i1 to ia, and a weight is applied to each of the input binary data (ie, peak data) A value is received (indicated by a dotted line), all input values are summed, and an operation is performed according to an activation function on the summed values to calculate an output value of a plurality of first hidden nodes.

Subsequently, although not shown, each of the plurality of second hidden nodes of the second hidden layer receives a value to which a weight is applied to each of the plurality of output values of the plurality of first hidden nodes (g1 to gb), and all of the input values Output values of the plurality of second hidden nodes are calculated by summing and performing an operation according to an activation function on the summed values. In this way, in the hidden layer (HL), a previous node value is transmitted with a weight applied, and a current node value is calculated through an operation. By repeating this process, output values of a plurality of k-th hidden nodes of a plurality of k-th hidden nodes (h1 to hj) of the k-th hidden layer (HLk), which is the last hidden layer, may be calculated.

Accordingly, referring to FIG. 3, the output node O applies a weight w = [w1, w2, ... , wj] is applied (indicated by a dotted line), and after summing up all the input values, an operation is performed on the summed values according to an activation function to calculate an output value. The output node O of the output layer OL corresponds to the gas concentration of the aqueous solution. And the output value of the output node (O) is the predicted value of the gas concentration of the aqueous solution. For example, if the output value of the output node O is 0.089, the concentration of the corresponding gas is 9%, and if the output value of the output node O is 0.911, the concentration of the corresponding gas is 91%. In this way, when the prediction model (PM) calculates the predicted value (eg, 0.089, 0.911), the concentration predicting unit 300 converts the predicted value into a concentration and outputs it.

Then, a method of generating a prototype of the aforementioned predictive model (PM) will be described. 5 is a flowchart for explaining a method of generating a prototype of a predictive model according to an embodiment of the present invention. 6 is a graph for explaining a method of generating a prototype of a predictive model according to an embodiment of the present invention.

Referring to FIG. 5 , the model generating unit 100 extracts learning data and test data from experimental data in step S100. That is, the model generating unit 100 prepares learning data by extracting the learning peak data of the wavelength corresponding to the type of gas and the gas concentration corresponding to the learning peak data from the Raman spectrum data, which is experimental data, and prepares the experimental data, the Raman spectrum. Inspection data is prepared by extracting inspection peak data of a wavelength corresponding to the type of gas and gas concentration corresponding to the inspection peak data from the data.

Next, the model generator 100 sets the number of hidden layers and hidden nodes of the prediction model (PM) according to the initial values in step S110. As for the initial value, it is preferable that the number of hidden gates of the prediction model (PM) is one. In the example of FIG. 3 , this means that only the first hidden layer HL1 exists and the other hidden layers do not exist. However, the present invention is not limited thereto, and an initial value may be differently set based on trusted data.

The model generating unit 100 performs learning on a predictive model (PM) having a previously set number of hidden layers and hidden nodes using learning data extracted from experimental data in step S120. Then, the model generating unit 100 checks the learning performance of the predictive model (PM) using the test data in step S130. Subsequently, the model generating unit 100 calculates the accuracy of the training data and the inspection data in step S140.

The model generating unit 100 determines whether the accuracy of the training data and the test data satisfies the condition in step S150. Here, the condition means a case where the accuracy of the training data is higher than the accuracy of the test data and both the accuracy of the learning data and the accuracy of the test data are equal to or greater than the threshold value.

As a result of the determination in step S150, if the accuracy of the training data and the test data does not satisfy the set condition, the model generator 100 proceeds to step S160 to reset the number of hidden layers and hidden nodes. At this time, the model generator 100 increases or decreases the number of hidden layers and hidden nodes. Then, the model generator 100 repeats steps S120 to S150 described above.

On the other hand, if the determination result of step S150 and the accuracy of the learning data and the test data satisfy the set condition, the model generator 100 determines the number of hidden layers and the number of hidden nodes according to the current setting. In this way, the prototype of the predictive model PM is determined.

For example, as shown in the graph of FIG. 6 , when the number of hidden layers is 1 and 2, both the accuracy of the training data and the accuracy of the test data are less than the threshold value. And when the number of hidden layers is 4 or 5, the accuracy of the training data is very high, but the accuracy of the inspection data is less than the critical value. Accordingly, since the condition is satisfied only when the number of hidden layers is three, the prototype of the prediction model (PM) can be determined as a model with three hidden layers.

Then, the learning method for the predictive model (PM) according to an embodiment of the present invention will be described in more detail. 7 is a flowchart for explaining a learning method for a predictive model (PM) according to an embodiment of the present invention. 8 is a graph showing Raman spectrum data for an aqueous solution containing acetate and butyrate in experimental data according to an embodiment of the present invention. 9 is a graph showing Raman spectrum data for an aqueous solution containing a mixture of acetate and butyrate and carbon monoxide in experimental data according to an embodiment of the present invention. 10 is a graph showing Raman spectrum data in which the concentration of gas contained in an aqueous solution is known in experimental data according to an embodiment of the present invention.

의 피크 데이터를 추출한다. 또한, 도 8의 (B)는 낙산염(Butyrate)이 포함된 물과 배양액의 혼합액에 대한 라만 스펙트럼 데이터이다. 이와 같이, 수용액에 포함된 가스의 종류가 아세트산염(Acetate)이면 877

의 피크 데이터를 추출한다. 도 9의 (C)는 아세트산염(Acetate) 및 낙산염(Butyrate)의 혼합물 수용액에 대한 라만 스펙트럼 데이터이다. 수용액에 포함된 가스가 아세트산염(Acetate) 및 낙산염(Butyrate)이기 때문에 928

및 877

각각의 피크 데이터를 추출한다. 도 9의 (D)는 일산화탄소(CO)의 수용액에 대한 라만 스펙트럼 데이터이다. 이와 같이, 수용액에 포함된 가스의 종류가 일산화탄소(CO)이면 1948

및 2063

의 피크 데이터를 추출한다. 전술한 바와 같이, 실험 데이터는 해당 수용액의 가스 농도를 포함한다. 예를 들면, 도 10의 (E)는 2.5% 농도(concentration)의 일산화탄소를 포함하는 수용액의 라만 스펙트럼 데이터이고, (F)는 5% 농도의 일산화탄소를 포함하는 수용액의 라만 스펙트럼 데이터를 보인다. 이에 따라, 해당 실험 데이터로부터 수용액에 포함된 가스의 종류가 일산화탄소(CO)이기 때문에 1948

및 2063

의 피크 데이터를 추출하고, 그 피크 데이터에 대응하는 가스의 농도(2.5%, 5%)를 추출할 수 있다. Referring to FIG. 7 , the model generating unit 100 prepares learning data in step S210. As shown in FIGS. 8 to 10 , the experimental data includes Raman spectrum data of an aqueous solution containing gas and gas concentration of the aqueous solution. Accordingly, the model generating unit 100 prepares learning data by extracting learning peak data of a wavelength corresponding to the type of gas and gas concentration corresponding to the learning peak data from Raman spectrum data, which is experimental data. For example, (A) of FIG. 8 is Raman spectrum data of a mixture of water containing acetate and a culture medium. As such, if the type of gas contained in the aqueous solution is acetate, 928

Extract the peak data of In addition, (B) of FIG. 8 is Raman spectrum data for a mixture of water and culture medium containing butyrate. As such, if the type of gas contained in the aqueous solution is acetate, 877

Extract the peak data of 9(C) is Raman spectrum data for an aqueous mixture of acetate and butyrate. 928 because the gases contained in the aqueous solution are acetate and butyrate.

and 877

Extract each peak data. 9(D) is Raman spectrum data for an aqueous solution of carbon monoxide (CO). As such, if the type of gas contained in the aqueous solution is carbon monoxide (CO), 1948

and 2063

Extract the peak data of As mentioned above, the experimental data includes the gas concentration of the aqueous solution. For example, (E) of FIG. 10 shows Raman spectrum data of an aqueous solution containing carbon monoxide at a concentration of 2.5%, and (F) shows Raman spectrum data of an aqueous solution containing carbon monoxide at a concentration of 5%. Accordingly, from the experimental data, since the type of gas contained in the aqueous solution is carbon monoxide (CO), 1948

and 2063

Peak data of can be extracted, and the concentration (2.5%, 5%) of the gas corresponding to the peak data can be extracted.

After preparing the learning data, the model generating unit 100 may perform boundary learning using the learning data in step S220. This step S220 is optional and may be performed or omitted as needed. Boundary learning is for clarifying boundaries that are not clearly distinguished through learning. Referring to FIG. 10 , it is assumed that data of a carbon monoxide concentration of 2.5% and a carbon monoxide concentration of 5% are not clearly distinguished through learning. In this case, a value between 2.5% and 5% is set as a boundary value (e.g. 2.75%), and a label is set based on the boundary value for boundary learning to clearly distinguish values between concentrations of 2.5% and 5%. Do it. The specific contents of boundary learning will be explained in more detail below.

Next, the model generating unit 100 performs numerical learning using the learning data in step S230. Numerical learning is a procedure for learning a predictive model (PM) to actually predict the concentration of a gas in an aqueous solution by setting a target value, that is, a label, to a concentration (numerical value) obtained from experimental data. The specific contents of numerical learning will be explained in more detail below.

Next, boundary learning in the predictive model (PM) according to an embodiment of the present invention will be described in more detail. 11 is a flowchart illustrating boundary learning for a predictive model (PM) according to an embodiment of the present invention. 12 is a diagram for explaining boundary learning for a predictive model (PM) according to an embodiment of the present invention. Emphasizing again, FIG. 11 describes an embodiment of boundary learning in step S220 of FIG. 7 in more detail.

Referring to FIG. 11 , the model generating unit 100 prepares a plurality of learning data including peak data for learning wavelengths according to types of gases and corresponding gas concentrations in step S310. Then, in step S320, the model generating unit 100 converts the gas concentration corresponding to the peak data for learning into a one-hot-encoding vector based on a preset boundary value to set boundary labels. According to an embodiment, it is assumed that the boundary value is set to 2.75%, that is, 0.0275, in order to clearly distinguish between 2.5% and 5% carbon monoxide concentration. Then, gas concentrations above 2.75% are converted to vector 1, and gas concentrations below 2.75% are converted to vector 0. For example, for a carbon monoxide concentration of 2.5%, the label is set to vector 0, and for a carbon monoxide concentration of 5%, the label is set to vector 1.

When boundary labels are set, the model generator 100 performs learning to optimize boundary loss in step S330. At this time, when the model generating unit 100 inputs the peak data for learning to the prediction model (PM), the prediction model (PM) calculates a predicted value through a plurality of operations in which weights between a plurality of layers are applied to the peak data for learning. . Then, the model generating unit 100 performs boundary loss optimization for modifying the weight of the prediction model PM so that the boundary loss, which is the difference between the prediction value and the boundary label, is minimized. At this time, the boundary loss can be derived through a boundary loss function such as Equation 3 below.

In Equation 3, Lboaderselected represents a boundary loss function. In particular, Oi is a predicted value that is an output of the output node O of the output layer OL of the predictive model PM, and vi is a boundary label corresponding to the predicted value. Here, i is an index corresponding to the output node O of the output layer OL of the predictive model PM.

In summary, the model generator 100 derives the edge loss through the edge loss function of Equation 3, and then performs edge loss optimization of modifying the weight of the prediction model (PM) so that the edge loss is minimized. This boundary loss optimization is repeatedly performed using a plurality of different learning data, and such repetition may be performed until a desired accuracy is reached through an evaluation index.

When the learning according to the boundary loss optimization as described above is completed, the model generator 100 selects a reference vector from among a plurality of hidden vectors of each group classified based on the boundary value through the prediction model (PM) in step S140 do. A method for selecting a reference vector when classifying peak data for learning with a corresponding gas concentration equal to or greater than the threshold among a plurality of learning data into a first group and classifying peak data for learning with a corresponding gas concentration less than the threshold value as a second group. Is as follows. The learning peak data of the previously used training data (S130) is re-entered into the prediction model (PM) in which boundary loss optimization is performed, and the plurality of the last hidden layer of the prediction model (PM), that is, the k-th hidden layer (HLk) A hidden vector that is an output value of the hidden nodes (h1 to hj) is derived. That is, the hidden vector is H[h1, h2, h3, . . . , hj]. The model generator 100 embeds a plurality of hidden vectors having a plurality of dimensions into a predetermined vector space. 12(a) shows an example in which a plurality of hidden vectors are embedded in a predetermined vector space. Basically, by boundary loss optimization, the hidden vector (marked by a circle) for the first group having a concentration greater than or equal to the boundary value in the vector space and the hidden vector (marked by a square) for the second group having a concentration less than the boundary value are the boundary ( BORDER) is used for classification. However, as shown in (a) of FIG. 12, hidden vectors that cross the BORDER also exist. The model generator 100 selects a reference vector for each group from among a plurality of hidden vectors embedded in the vector space. As shown in (A) of 9, a median value among hidden vectors in the same group may be selected as the reference vector.

Next, the model generating unit 100 sets boundary labels and augmentation labels in step S350. The boundary label is set by converting the gas concentration corresponding to the peak data for learning into a one-hot encoding vector based on a preset boundary value, as in the previous step S320. The augmented label uses the reference vector selected for each group in step S340.

When boundary labels and augmentation labels are set, the model generator 100 performs learning to optimize augmentation loss in step S360. At this time, when the model generator 100 inputs the peak data for learning to the prediction model (PM), the prediction model (PM) sequentially performs the last concealment through a plurality of operations in which weights between a plurality of layers are applied to the peak data for learning. An output value of the layer (eg, the kth hidden layer) and a predicted value that is an output of the output layer are calculated. Then, the model generator 100 calculates the boundary loss representing the difference between the predicted value and the boundary label, and the output value H[h1, h2, h3, . . . . At this time, the boundary enhancement loss can be derived through a boundary enhancement loss function such as Equation 3 below.

In Equation 4, Lboaderenhanced represents the boundary enhancement loss function. In particular, Oi is a predicted value that is an output of the output node O of the output layer OL of the predictive model PM, and vi is a boundary label corresponding to the predicted value. Here, i is an index corresponding to the output node O of the output layer OL of the predictive model PM. In addition, hij is a hidden vector H [h1, h2, h3, . , hj], and rij is the hidden vector H [h1, h2, h3, . . . , hj]. This augmented label is a reference vector selected for each group. j is an index corresponding to a plurality of hidden nodes of the last hidden layer (eg, HLk) of the prediction model (PM).

In summary, the model generator 100 derives the boundary enhancement loss through the boundary enhancement loss function of Equation 4. The boundary enhancement loss is the boundary loss representing the difference between the predicted value and the boundary label, and the output value of the last hidden layer (eg, the kth hidden layer) H[h1, h2, h3, ... , hj] and the enhancement loss representing the difference from the enhancement label. In this way, when the boundary enhancement loss is derived, the model generating unit 100 performs boundary enhancement loss optimization of modifying the weight of the prediction model (PM) so that the boundary enhancement loss is minimized. Such boundary enhancement loss optimization is repeatedly performed using a plurality of different training data, and such repetition may be performed until a desired accuracy is reached through an evaluation index.

When learning by the boundary enhancement loss is completed, a plurality of hidden vectors on the vector space can be classified as shown in FIG. 12(B). That is, the plurality of hidden vectors are moved toward the reference vector on the vector space. In particular, it can be seen that the distinction between the first group and the second group becomes clearer as the hidden vector existing on the BORDER moves toward the reference vector. That is, as learning is performed through boundary enhancement loss optimization, the performance of discriminating numerical differences in the prediction model (PM) can be improved.

Next, numerical learning in the predictive model (PM) according to an embodiment of the present invention will be described in more detail. 13 is a flowchart for explaining numerical learning for a predictive model (PM) according to an embodiment of the present invention. Emphasizing again, FIG. 13 explains the numerical learning of step S230 of FIG. 7 in more detail.

Referring to FIG. 13 , the model generating unit 100 prepares a plurality of learning data including peak data for learning wavelengths according to types of gases and corresponding gas concentrations in step S410. Then, the model generation unit 100 sets the value of the gas concentration corresponding to the peak data for learning as a numerical label in step S420. For example, for a carbon monoxide concentration of 2.5%, the numerical label is set to 0.025, and for a carbon monoxide concentration of 5%, the label is set to 0.050.

When the numerical labels are set, the model generating unit 100 performs learning to optimize numerical loss in step S430. At this time, when the model generating unit 100 inputs the peak data for learning to the prediction model (PM), the prediction model (PM) calculates a predicted value through a plurality of operations in which weights between a plurality of layers are applied to the peak data for learning. . Then, the model generating unit 100 performs numerical loss optimization to modify the weight of the prediction model PM so that the numerical loss, which is the difference between the predicted value and the numerical label, is minimized. At this time, the numerical loss can be derived through a numerical loss function such as Equation 5 below.

In Equation 3, Lvalue represents a numerical loss function. Oi is a predicted value output from the output layer OL, and Ci is a numerical label corresponding to the predicted value output from the output layer OL. And i is an index corresponding to an output node of the output layer OL of the predictive model PM. The first term of the numerical loss function Lvalue is the L1-norm loss, and the second term represents SSIM (Structural Similarity Index). That is, optimization may be performed so that the sum of L1-norm and SSIM is minimized.

In summary, the model generator 100 derives the numerical loss through the numerical loss function of Equation 5, and then performs numerical loss optimization to modify the weight of the prediction model PM so that the numerical loss is minimized. This numerical loss optimization is repeatedly performed using a plurality of different training data, and such repetition may be performed until a desired accuracy is reached through an evaluation index.

When the predictive model (PM) is generated through the method described above, the predictor (PA) according to an embodiment of the present invention can predict the dissolved gas concentration of the aqueous solution in real time. These methods will be described. 14 is a flowchart for explaining a method for predicting a dissolved gas concentration in an aqueous solution based on a Raman spectroscopy signal according to an embodiment of the present invention.

Referring to FIG. 14, the spectrum measuring unit 10 emits light through the probe 11 in the aqueous solution as shown in FIG. 1 in step S510, and measures the scattering of the emitted light to obtain a Raman spectrum. Acquire data. These Raman spectrum data are provided to the prediction unit 20 .

및 2063

이다. 또한, 가스가 아세트산염(Acetate)이면 파장은 928

이고, 가스가 낙산염(Butyrate)이면, 파장은 877

이다. When Raman spectrum data for the aqueous solution is input from the spectrum measuring unit 10 in step S520, the peak extraction unit 200 of the prediction unit 20, the wavelength (Raman spectrum) corresponding to the type of gas included in the aqueous solution from the Raman spectrum data Shift) to extract the peak data. 8 to 9, when the gas is CO, the wavelength is 1948

and 2063

am. Also, if the gas is acetate, the wavelength is 928

, and the gas is butyrate, the wavelength is 877

am.

The concentration prediction unit 300 calculates a gas concentration predicted value from the peak data through the predictive model (PM) in step S530. At this time, the concentration prediction unit 300 inputs the peak data extracted by the peak extraction unit 200 to the prediction model (PM). Then, the prediction model (PM) calculates the predicted value of the gas concentration included in the aqueous solution through a plurality of calculations to which weights learned between the plurality of layers (IL, HL, and OL) are applied to the peak data.

Next, the concentration prediction unit 300 converts the predicted value into a concentration in step S540 and outputs the converted concentration as the dissolved gas concentration of the aqueous solution. For example, when the prediction model (PM) calculates predicted values such as 0.089 and 0.911 for carbon monoxide (CO), the concentration predicting unit 300 converts the predicted values into concentrations such as 9% and 91% to dissolve carbon monoxide (CO). Output the gas concentration as 9%, 91%, etc.

15 is a diagram illustrating a computing device according to an embodiment of the present invention. The computing device TN100 of FIG. 15 may be a device described in this specification (eg, a prediction device (PA), etc.).

In the embodiment of FIG. 15 , the computing device TN100 may include at least one processor TN110, a transceiver TN120, and a memory TN130. In addition, the computing device TN100 may further include a storage device TN140, an input interface device TN150, and an output interface device TN160. Elements included in the computing device TN100 may communicate with each other by being connected by a bus TN170.

The processor TN110 may execute program commands stored in at least one of the memory TN130 and the storage device TN140. The processor TN110 may mean a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor on which methods according to embodiments of the present invention are performed. Processor TN110 may be configured to implement procedures, functions, methods, and the like described in relation to embodiments of the present invention. The processor TN110 may control each component of the computing device TN100.

Each of the memory TN130 and the storage device TN140 may store various information related to the operation of the processor TN110. Each of the memory TN130 and the storage device TN140 may include at least one of a volatile storage medium and a non-volatile storage medium. For example, the memory TN130 may include at least one of read only memory (ROM) and random access memory (RAM).

The transmitting/receiving device TN120 may transmit or receive a wired signal or a wireless signal. The transmitting/receiving device TN120 may perform communication by being connected to a network.

On the other hand, the method according to the embodiment of the present invention described above can be implemented in the form of a program readable through various computer means and recorded on a computer-readable recording medium. Here, the recording medium may include program commands, data files, data structures, etc. alone or in combination. Program instructions recorded on the recording medium may be those specially designed and configured for the present invention, or those known and usable to those skilled in computer software. For example, recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks ( magneto-optical media), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of the program command may include a high-level language that can be executed by a computer using an interpreter, as well as a machine language generated by a compiler. These hardware devices may be configured to act as one or more software modules to perform the operations of the present invention, and vice versa.

The present invention has been described above using several preferred examples, but these examples are illustrative and not limiting. As such, those skilled in the art to which the present invention belongs will understand that various changes and modifications can be made according to the doctrine of equivalents without departing from the spirit of the present invention and the scope of rights set forth in the appended claims.

10: spectrum measuring unit
20: prediction unit
100: model generator
200: peak extraction unit
300: concentration prediction unit

Claims (20)

An apparatus for predicting the concentration of dissolved gas in an aqueous solution based on a Raman spectroscopy signal,
a peak extraction unit for extracting peak data of a wavelength corresponding to a type of gas from the Raman spectrum data when Raman spectrum data for an aqueous solution is input; and
a concentration prediction unit that calculates a predicted value of the gas concentration contained in the aqueous solution through a predictive model based on the peak data, and derives and outputs the gas concentration from the calculated predicted value;
Including,
Preparing learning data by extracting learning peak data of a wavelength corresponding to the type of gas and gas concentration corresponding to the learning peak data from Raman spectrum data, which is experimental data;
Set some of the learning data as test data;
Learning a predictive model to calculate a predicted value of the gas concentration contained in the aqueous solution using the learning data;
a model generating unit inspecting the learning result using the inspection data;
Including more,
The model generator
The gas concentration corresponding to the learning peak data is converted into a one-hot encoding vector based on a preset boundary value and set as a boundary label,
Input the peak data for learning into a predictive model,
When the predictive model calculates a predicted value by performing a plurality of calculations to which weights between a plurality of layers are applied to the peak data for learning,
boundary loss function

Performs boundary optimization for modifying weights of the prediction model so that boundary loss, which is a difference between a predicted value output from the prediction model and a boundary label, is minimized according to
The Lboaderselected is a boundary loss function,
The Oi is a predicted value of the predictive model,
vi is a boundary label corresponding to the output value,
Characterized in that i is an index corresponding to the output node of the output layer of the predictive model
A device for predicting dissolved gas concentrations. delete According to claim 1,
The model generator
Learning the predictive model using the learning data while adjusting the number of hidden layers and hidden nodes of the predictive model until both the accuracy of each of the learning data and the test data satisfies a preset condition, and the test Characterized in repeating the inspection of the learning performance for the predictive model using data
A device for predicting dissolved gas concentrations. delete According to claim 1,
The model generator
Based on a preset boundary value, the gas concentration corresponding to the peak data for learning is converted into a one-hot encoding vector and set as the boundary label,
Set a reference vector corresponding to the hidden vector of the learning peak data as an augmentation label,
Input the peak data for learning into a predictive model,
When the predictive model performs a plurality of operations in which weights between a plurality of layers are applied to the peak data for learning, output values of a plurality of hidden nodes of the last hidden layer and predicted values that are outputs of the output layer are calculated,
Boundary Augmentation Loss Function

Depending on the
Boundary enhancement loss, including boundary loss, which is the difference between the predicted value and the boundary label, which is the output of the prediction model, and enhancement loss representing the difference between the output value of a plurality of hidden nodes of the last hidden layer of the prediction model and the augmented label, is minimized performing boundary enhancement optimization to modify the weights of the prediction model;
The Lboaderenhanced is a boundary enhancement loss function,
The Oi is a prediction value that is an output of the output layer of the prediction model,
vi is a boundary label corresponding to the predicted value,
The hij is an output value of a plurality of hidden nodes of the last hidden layer of the prediction model,
rij is an augmented label corresponding to the output values of the plurality of hidden nodes;
The i is an index corresponding to the output node of the output layer of the predictive model,
The j is an index corresponding to a plurality of hidden nodes of the last hidden layer of the prediction model
characterized by
A device for predicting dissolved gas concentrations. According to claim 5,
The model generator
Set the gas concentration corresponding to the peak data for learning as a numerical label,
Input the peak data for learning into a predictive model,
When the predictive model calculates a predicted value by performing a plurality of calculations to which weights between a plurality of layers are applied to the peak data for learning,
Characterized in that performing numerical optimization of modifying the weight of the predictive model so that the numerical loss, which is the difference between the predicted value and the numerical label of the predictive model, is minimized
A device for predicting dissolved gas concentrations. According to claim 6,
The model generator
Numerical loss function

Performs numerical optimization to modify the weight of the prediction model so that the numerical loss, which is the difference between the output value of the prediction model and the numerical label, is minimized according to
The Lvaule is a numerical loss function,
The Oi is a prediction value that is an output of the prediction model,
Ci is a numerical label corresponding to the predicted value,
Characterized in that i is an index corresponding to the output node of the output layer of the predictive model
A device for predicting dissolved gas concentrations. According to claim 1,
If the gas is CO, the wavelength is 1948

and 2063

ego,
If the gas is acetate, the wavelength is 928

ego,
If the gas is butyrate, the wavelength is 877

characterized in that
A device for predicting dissolved gas concentrations. An apparatus for predicting the concentration of dissolved gas in an aqueous solution based on a Raman spectroscopy signal,
Preparing learning data by extracting learning peak data of a wavelength corresponding to the type of gas and gas concentration corresponding to the learning peak data from Raman spectrum data, which is experimental data;
a model generator for learning a predictive model, which is an artificial neural network algorithm, to calculate a predicted value of a gas concentration contained in an aqueous solution using the learning data;
Including,
The model generator
Set the gas concentration corresponding to the peak data for learning as a numerical label,
Input the peak data for learning into a predictive model,
When the predictive model calculates a predicted value by performing a plurality of calculations to which weights between a plurality of layers are applied to the peak data for learning,
Perform numerical optimization to modify the weight of the predictive model so that the numerical loss, which is the difference between the predicted value and the numerical label of the predictive model, is minimized,
Numerical loss function

Performs numerical optimization to modify the weight of the predictive model so that the numerical loss, which is the difference between the output value of the predictive model and the numerical label, is minimized according to
The Lvaule is a numerical loss function,
The Oi is a prediction value that is an output of the prediction model,
Ci is a numerical label corresponding to the predicted value,
Characterized in that i is an index corresponding to the output node of the output layer of the predictive model
A device for predicting dissolved gas concentrations. delete According to claim 9,
The model generator
The gas concentration corresponding to the peak data for learning is converted into a one-hot encoding vector based on a preset boundary value and set as a boundary label,
Input the peak data for learning into a predictive model,
When the predictive model calculates a predicted value by performing a plurality of calculations to which weights between a plurality of layers are applied to the peak data for learning,
boundary loss function

After performing boundary optimization for modifying the weight of the prediction model so that the boundary loss, which is the difference between the predicted value and the boundary label, is minimized according to
Based on a preset boundary value, the gas concentration corresponding to the peak data for learning is converted into a one-hot encoding vector and set as the boundary label,
Set a reference vector corresponding to the hidden vector of the learning peak data as an augmentation label,
Input the peak data for learning into a predictive model,
When the predictive model performs a plurality of operations in which weights between a plurality of layers are applied to the peak data for learning, output values of a plurality of hidden nodes of the last hidden layer and predicted values that are outputs of the output layer are calculated,
Boundary Augmentation Loss Function

Depending on the
Boundary enhancement loss, including boundary loss, which is the difference between the predicted value and the boundary label, which is the output of the prediction model, and enhancement loss representing the difference between the output value of a plurality of hidden nodes of the last hidden layer of the prediction model and the augmented label, is minimized performing boundary enhancement optimization to modify the weights of the prediction model;
The Lboaderselected is a boundary loss function,
The Lboaderenhanced is a boundary enhancement loss function,
The Oi is a prediction value that is an output of the output layer of the prediction model,
vi is a boundary label corresponding to the predicted value,
The hij is an output value of a plurality of hidden nodes of the last hidden layer of the prediction model,
rij is an augmented label corresponding to the output values of the plurality of hidden nodes;
The i is an index corresponding to the output node of the output layer of the predictive model,
The j is an index corresponding to a plurality of hidden nodes of the last hidden layer of the prediction model
characterized by
A device for predicting dissolved gas concentrations. According to claim 9,
a spectrum measuring unit that emits light through a probe in an aqueous solution, measures the scattering of the emitted light, acquires Raman spectrum data, and provides the obtained Raman spectrum data;
a peak extraction unit extracting peak data of a wavelength corresponding to the type of gas from the Raman spectrum data; and
a concentration prediction unit that calculates a predicted value of the gas concentration included in the aqueous solution through the predictive model based on the peak data, and derives and outputs the gas concentration from the calculated predicted value;
characterized in that it further comprises
A device for predicting dissolved gas concentrations. A method for predicting a dissolved gas concentration in an aqueous solution based on a Raman spectroscopy signal,
Acquiring Raman spectrum data by emitting light through a probe in an aqueous solution by a spectrum measuring unit and measuring scattering of the emitted light;
Extracting peak data of a wavelength corresponding to the type of gas from the Raman spectrum data by a peak extraction unit;
Calculating a predicted value of the gas concentration contained in the aqueous solution through a predictive model based on the peak data by a concentration predictor; and
deriving and outputting a gas concentration from the predicted value calculated by the concentration prediction unit;
Including,
Before acquiring the Raman spectrum data,
preparing learning data by extracting learning peak data of a wavelength corresponding to a type of gas and a gas concentration corresponding to the learning peak data from Raman spectrum data, which is experimental data, by the model generator;
setting a gas concentration corresponding to the peak data for learning as a numerical label by the model generating unit;
inputting the peak data for learning into a predictive model by the model generating unit;
calculating, by the predictive model, a predicted value by performing a plurality of calculations to which weights between a plurality of layers are applied to the peak data for learning;
performing numerical optimization by the model generating unit to modify weights of the predictive model so that a numerical loss, which is a difference between the predicted value and the numerical label of the predictive model, is minimized;
characterized in that it further comprises
A method for predicting dissolved gas concentrations. According to claim 13,
Before acquiring the Raman spectrum data,
extracting learning data and test data from the experimental data by the model generator;
setting the number of hidden layers and hidden nodes of the prediction model by the model generator;
The model generator performs learning on a predictive model having a set number of hidden layers and hidden nodes using the training data, inspects the learning performance of the predictive model (PM) using the test data, and generates the model additionally calculating the accuracy of the learning data and the accuracy of the inspection data;
determining, by the model generating unit, whether the accuracy of the training data is higher than that of the test data and a condition that both the accuracy of the learning data and the accuracy of the test data are equal to or greater than a threshold value is satisfied; and
determining, by the model generator, the number of hidden layers and hidden nodes according to a current setting, if the condition is satisfied as a result of the determination;
characterized in that it includes
A method for predicting dissolved gas concentrations. According to claim 14,
After determining whether the above condition is satisfied,
As a result of the determination, if the above condition is not satisfied,
repeating the step of calculating the accuracy and the step of determining whether the condition is satisfied after the model generator resets the number of hidden layers and hidden nodes;
characterized in that it further comprises
A method for predicting dissolved gas concentrations. delete According to claim 13,
The step of performing the numerical optimization is
The model generation unit
Numerical loss function

Performs numerical optimization to modify the weight of the prediction model so that the numerical loss, which is the difference between the output value of the prediction model and the numerical label, is minimized according to
The Lvalue is a numerical loss function,
The Oi is a prediction value that is an output of the prediction model,
Ci is a numerical label corresponding to the predicted value,
Characterized in that i is an index corresponding to the output node of the output layer of the predictive model
A method for predicting dissolved gas concentrations. According to claim 13,
Before acquiring the Raman spectrum data,
converting, by the model generator, the gas concentration corresponding to the peak data for training into a one-hot encoding vector based on a predetermined boundary value, and setting the boundary label;
inputting the peak data for learning into a predictive model by the model generating unit;
calculating, by the predictive model, a predicted value by performing a plurality of calculations to which weights between a plurality of layers are applied to the peak data for learning;
The model generation unit
boundary loss function

performing boundary optimization of modifying weights of the prediction model so that boundary loss, which is a difference between a predicted value output from the prediction model and a boundary label, is minimized according to;
Including more,
The Lboaderselected is a boundary loss function,
The Oi is a prediction value that is an output of the prediction model,
vi is a boundary label corresponding to the predicted value,
Characterized in that i is an index corresponding to the output node of the output layer of the predictive model
A method for predicting dissolved gas concentrations. According to claim 18,
After performing the boundary optimization,
The model generation unit converts the gas concentration corresponding to the peak data for learning into a one-hot encoding vector based on a preset boundary value, sets it as the boundary label, and sets the reference vector corresponding to the hidden vector of the learning peak data as an augmented label doing;
inputting the peak data for learning into a predictive model by the model generating unit;
calculating an output value of a plurality of hidden nodes of a last hidden layer and a predicted value that is an output of an output layer by performing a plurality of operations in which weights between a plurality of layers are applied to the peak data for learning by the prediction model;
The model generation unit boundary enhancement loss function

According to the boundary enhancement loss, including the boundary loss that is the difference between the predicted value and the boundary label, which is the output of the prediction model, and the enhancement loss representing the difference between the output value of a plurality of hidden nodes of the last hidden layer of the prediction model and the augmented label, Minimum performing boundary enhancement optimization to modify the weights of the prediction model so that
Including more,
The Lboaderenhanced is a boundary enhancement loss function,
The Oi is a prediction value that is an output of the prediction model,
vi is a boundary label corresponding to the predicted value,
The i is an index corresponding to the output node of the output layer of the predictive model,
The hij is an output value of a plurality of hidden nodes of the last hidden layer of the prediction model,
rij is an augmented label corresponding to the output values of the plurality of hidden nodes;
Characterized in that j is an index corresponding to a plurality of hidden nodes of the last hidden layer of the prediction model
A method for predicting dissolved gas concentrations. According to claim 13,
If the gas is CO, the wavelength is 1948

and 2063

ego,
If the gas is acetate, the wavelength is 928

ego,
If the gas is butyrate, the wavelength is 877

characterized in that
A method for predicting dissolved gas concentrations.

Images