Nanoparticle estimation model training method, nanoparticle estimation method using same, nanoparticle estimation model training device and nanoparticle estimation device using same

Abstract

The present invention relates to a nanoparticle estimation model training method, a nanoparticle estimation method using same, a nanoparticle estimation model training device and a nanoparticle estimation device using same, and, more specifically, to a nanoparticle estimation model training method, a nanoparticle estimation method using same, a nanoparticle estimation model training device and a nanoparticle estimation device using same, which select and use only training data whose distribution corresponds to a reference normal range before a nanoparticle estimation model is trained to increase the reliability of the training data, compare whether unknown nanoparticle data corresponds to a training data category, and input only data corresponding to the training data category into the trained nanoparticle estimation model so as to improve nanoparticle estimation accuracy.

Description

Nanoparticle estimation model training method, nanoparticle estimation method using the same, nanoparticle estimation model training device and nanoparticle estimation device using the same

[0001] The present invention relates to a method for training a nanoparticle estimation model, a method for estimating nanoparticles using the same, a nanoparticle estimation model training device, and a nanoparticle estimation device using the same. More specifically, the invention relates to a method for training a nanoparticle estimation model that improves nanoparticle estimation accuracy by selecting and using only training data whose distribution falls within a reference normal range before training the nanoparticle estimation model, thereby increasing the reliability of the training data, comparing whether unknown nanoparticle data falls within the training data category, and inputting only the data that falls within the training data category into the trained nanoparticle estimation model. The invention also relates to a method for estimating nanoparticles using the same, a nanoparticle estimation model training device, and a nanoparticle estimation device using the same.

[0002] Various organic and inorganic chemicals used in the manufacturing processes of products requiring high precision, such as displays and semiconductors, demand higher purity chemicals than currently available to prevent a decline in manufacturing yield; consequently, advanced analytical technologies are being developed and newly applied to verify the quality of these high-purity chemicals.

[0003] Among these, the importance of particle analysis is increasing, and since even small particles of 10 nanometers or less can affect the yield and high integration of semiconductor manufacturing processes, the development of stable analytical methods for quality control is required, and the scalability of the technology must be ensured so that it is possible to analyze the causes of defects that may occur during the process.

[0004] In conventional technologies regarding methods and devices for analyzing samples through artificial neural network learning, samples are analyzed by learning from artificial neural networks on emission images caused by light source irradiation; however, since the aforementioned conventional technology applies to methods for analyzing solid samples, there was a problem in that it could not be applied to the analysis of nanoparticles in solution.

[0005] In addition, when using artificial neural network models, there were cases where the variance or mean of the collected training data differed beyond the standard range.

[0006] This may be a problem caused by the collection of incorrect data due to reasons such as contamination during solution preparation, errors in the sample collection device, differences in the intensity of the incident laser resulting from abnormal operation or malfunction, or errors or malfunctions in the training data acquisition device.

[0007] Therefore, even if information regarding unknown nanoparticles is estimated using an artificial neural network model, a problem of reduced accuracy may occur.

[0008] One objective of the present invention is to provide a nanoparticle estimation model training method and a nanoparticle estimation model training device that can increase the reliability of training data by selecting and using only training data whose distribution corresponds to a reference normal range before training the nanoparticle estimation model, in order to solve the aforementioned problem.

[0009] In addition, the objective of the present invention is to provide a nanoparticle estimation method and a nanoparticle estimation device using the same, which ultimately increases the nanoparticle estimation probability by analyzing whether unknown nanoparticle data corresponds to a training data category during the process of estimating unknown nanoparticles based on a learned nanoparticle estimation model, and inputting only data corresponding to the training data category into the learned nanoparticle estimation model.

[0010] The objectives of the present invention are not limited to the problems mentioned above, and other objectives and advantages of the present invention not mentioned may be understood from the following description and will be more clearly understood from the embodiments of the present invention. Furthermore, it will be understood that the objectives and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims.

[0011] A method for training a nanoparticle estimation model according to an embodiment of the present invention may include the steps of: preparing first sample data corresponding to a solution containing a first standard nanoparticle; collecting second sample data corresponding to a solution containing a second standard nanoparticle with a different distribution from the first standard nanoparticle; calculating the similarity between the first sample data and the second sample data; distinguishing between training data and non-training data based on the result of the similarity calculation; and training a nanoparticle estimation model using the training data.

[0012] A nanoparticle estimation method according to an embodiment of the present invention may include the steps of: collecting unknown nanoparticle data by analyzing a solution containing unknown nanoparticles; comparing training data with the unknown nanoparticle data to determine whether the unknown nanoparticle data falls within the category of the training data; and, if the unknown nanoparticle data falls within the range of the training data, inputting the unknown nanoparticle data into a nanoparticle estimation model to estimate the size and type of the unknown nanoparticles.

[0013] A nanoparticle estimation model learning device according to an embodiment of the present invention may include a sample data collection unit that prepares first sample data corresponding to a first standard nanoparticle-containing solution and collects second sample data corresponding to a second standard nanoparticle-containing solution having a different distribution from the first standard nanoparticle; a training data selection unit that calculates the similarity between the first sample data and the second sample data and distinguishes between training data and non-training data based on the similarity calculation result; and a learning unit that trains a nanoparticle estimation model using the training data.

[0014] A nanoparticle estimation device according to an embodiment of the present invention may include a data preprocessing unit that collects unknown nanoparticle data by analyzing a solution containing unknown nanoparticles and compares training data with the unknown nanoparticle data to determine whether the unknown nanoparticle data falls within the category of the training data, and an estimation unit that, if the unknown nanoparticle data falls within the range of the training data, inputs the unknown nanoparticle data into a nanoparticle estimation model to estimate the size and type of the unknown nanoparticle.

[0015] According to an embodiment of the present invention, in the process of learning a nanoparticle estimation model, the reliability of the training data can be increased by comparing sample data with reference data and selecting data suitable as training data, thereby ensuring that the difference in the distribution of the training data falls within the reference normal range.

[0016] Furthermore, according to an embodiment of the present invention, the accuracy of nanoparticle estimation can be improved by comparing whether unknown nanoparticle data corresponds to a training data category and inputting only the data corresponding to the training data category into a trained nanoparticle estimation model.

[0017] The effects of the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art from the description below.

[0018] FIG. 1 is a block diagram illustrating the configuration of a nanoparticle estimation system according to an embodiment of the present invention.

[0019] Figure 2 is a graph illustrating an example of a data distribution difference occurring during the process of repeatedly collecting data N or more times.

[0020] FIG. 3 is a block diagram illustrating the configuration of a nanoparticle estimation model learning device according to an embodiment of the present invention.

[0021] FIG. 4 is a block diagram illustrating the configuration of a nanoparticle estimation device according to an embodiment of the present invention.

[0022] FIG. 5 illustrates an example of a learning factor formed by a sample data collection unit converting various types of raw data into one dimension according to an embodiment of the present invention.

[0023] FIG. 6 illustrates an example of comparing the overlap rate of the data distribution area of ​​individual factors of sample data according to an embodiment of the present invention.

[0024] FIG. 7 illustrates an example of comparing the overlap rate of two-dimensional data in which individual factors of sample data are converted into dimensionality reduction according to an embodiment of the present invention.

[0025] FIG. 8 is a flowchart illustrating a nanoparticle estimation model learning method according to one embodiment of the present invention.

[0026] FIG. 9 is a flowchart illustrating a nanoparticle estimation method according to one embodiment of the present invention.

[0027] Figure 10 illustrates an example of comparing training data and unknown nanoparticle data.

[0028] Figure 11 illustrates an example showing that the accuracy of a learning model is improved by selecting training data through the calculation of similarity based on the overlap rate of the data distribution regions of individual factors.

[0029] FIG. 12 is an example diagram illustrating the process of selecting training data by distinguishing the shape of vibration signals using machine learning and selecting only normal signals according to an embodiment of the present invention.

[0030] The present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the following embodiments, parts not directly related to the description are omitted to clearly explain the present invention; however, this does not mean that such omitted components are unnecessary when implementing a device or device to which the concept of the present invention is applied. Furthermore, the same reference numerals are used for identical or similar components throughout the specification.

[0031] In the following description, terms such as "first," "second," etc., may be used to describe various components, but said components should not be limited by said terms, and said terms are used solely for the purpose of distinguishing one component from another. Additionally, in the following description, singular expressions include plural expressions unless the context clearly indicates otherwise.

[0032] In the following description, terms such as “comprising” or “having” are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not excluding in advance the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0033] FIG. 1 is a block diagram illustrating the configuration of a nanoparticle estimation system according to an embodiment of the present invention.

[0034] Referring to FIG. 1, a nanoparticle estimation system (1) according to one embodiment of the present invention may include a nanoparticle estimation device (100) and a nanoparticle estimation model learning device (200).

[0035] Training data collected for the education of artificial intelligence models may have data distribution regions, specifically variance or mean, that differ beyond a standard range.

[0036] This may be a problem resulting from the collection of incorrect data due to reasons such as contamination occurring during the preparation of a solution containing nanoparticles, errors in the sample collection device, differences in the intensity of the incident laser caused by abnormal operation or malfunction, or errors or malfunctions in the training data acquisition device.

[0037] Referring to Figure 2, it can be seen that there are problems such as the median or mean value of the data distribution area of ​​the sample data collected as training data being different from the median or mean value of the data distribution area of ​​the reference data, having different dispersion, or having outliers added. Here, the reference data refers to the training data that was first used in the nanoparticle estimation model.

[0038] As such, if an artificial intelligence model is trained using sample data whose distribution differs from the reference data by more than a reference range, and information regarding nanoparticles is estimated based on the trained AI model, a problem of reduced estimation accuracy may occur.

[0039] To solve this problem, in the present invention, before training an artificial intelligence model using training data, the training data used in the artificial intelligence model is first set as reference data, and sample data corresponding to a normal standard range in terms of similarity comparison results with the reference data are selected as training data, so that only sample data corresponding to a normal range in terms of distribution difference are used as training data.

[0040] FIG. 3 is a block diagram illustrating the configuration of a nanoparticle estimation model learning device according to an embodiment of the present invention, and FIG. 8 is a flowchart for explaining a nanoparticle estimation model learning method according to an embodiment of the present invention.

[0041] Referring to FIG. 3, a nanoparticle estimation model learning device (100) according to an embodiment of the present invention may include a control unit (110), a sample data collection unit (120), a training data selection unit (130), and a learning unit (140). Not all blocks shown in FIG. 3 are essential components, and some blocks included in the nanoparticle estimation model learning device (100) in other embodiments may be added, changed, or deleted. For example, a verification unit (not shown) for verifying the trained results of the nanoparticle estimation model learning device (100) may be added.

[0042] Referring to FIGS. 3 and FIGS. 8, the control unit (110) can control the overall operation of the nanoparticle estimation model learning device (100).

[0043] The sample data collection unit (120) can prepare first sample data corresponding to a first standard nanoparticle-containing solution and collect second sample data corresponding to a second standard nanoparticle-containing solution.

[0044] The first sample data serves as reference data and refers to the training data previously used in the nanoparticle estimation model.

[0045] The second sample data can be described as data collected as candidates for training data before selecting the training data.

[0046] The sample data collection unit (120) can collect training data that has been used first through a nanoparticle estimation model as first sample data. The first sample data can be a set of training data whose distribution corresponds to a predetermined standard normal range. The distribution may refer to the mean or variance of the data distribution area.

[0047] The sample data collection unit (120) can acquire raw data composed of a one-dimensional or two-dimensional matrix from various measuring instruments, and can convert the features of the raw data into one-dimensional numerical values ​​through statistical processing such as machine learning or average calculation, and collect them as a set of learning factors. The measuring instruments may include at least one of a camera, a vibration sensor, a sound sensor, and an energy sensor.

[0048] FIG. 5 illustrates an example of an individual learning factor formed by a sample data collection unit using various types of raw data and data converted into one dimension, according to an embodiment of the present invention.

[0049] Referring to FIG. 5, when the raw data is an image, at least one learning factor (A, B, C, D) can be obtained by converting the two-dimensional image data into one dimension.

[0050] If the raw data is a vibration signal, the time series data of the vibration signal can be converted into one dimension to obtain the learning factor G.

[0051] When the raw data is a sound signal, the time series data of the sound signal can be converted into one dimension to obtain the learning factor H.

[0052] Accordingly, a set of individual factors including the aforementioned A~H can be obtained.

[0053] The method by which the training data selection unit (130) calculates similarity can be broadly divided into two methods: a method of analyzing the region distribution of the data through arithmetic processing, and a method of obtaining it through machine learning methods such as dimensionality reduction.

[0054] As a first method, similarity can be calculated through the overlap rate of the data distribution regions of individual learning factors.

[0055] To this end, the training data selection unit (130) can calculate the similarity between the first sample data and the individual factors from A to H described above, and distinguish between the training data and the non-training data based on the result of the similarity calculation.

[0056] Since the first sample data is a set of training data previously used through a nanoparticle estimation model, it can be considered training data having a distribution within a pre-set standard normal range.

[0057] Accordingly, the training data selection unit (130) can determine whether the difference between the data distribution area distribution of the second sample data and the data distribution area distribution of the first sample data falls within a standard normal range, and select the second sample data that falls within the standard normal range as training data. On the other hand, the second sample data that does not fall within the standard normal range due to the difference between the data distribution area distributions can be classified as non-training data.

[0058] The region defined by the aforementioned data distribution or the 2-dimensional data transformed by dimensionality reduction may be the region predicted by KDE (Kernel Density Estimation).

[0059] The overlap rate is the percentage value obtained by comparing data regions measured at different orders and dividing the overlapping region by the total data region of a given order. It can also be described as the average value of the overlap rates obtained from regions of different orders.

[0060] FIG. 6 illustrates an example of comparing the data distribution area overlap rate of individual factors of sample data according to an embodiment of the present invention.

[0061] The training data selection unit (130) can determine similarity based on the overlap rate predicted through KDE for each individual learning factor (A~G), which is an image, vibration signal, and sound signal.

[0062] For example, if the overlap rate of the region predicted through KDE is 100%, it can be seen that the similarity of the second sample data to the first sample data is 100%.

[0063] In particular, the training data selection unit (130) can improve the accuracy of the learning model by calculating similarity through the overlap rate of the data distribution area of ​​individual factors as shown in FIG. 11, and by selecting training data by selecting only learning factors that have a similarity of a certain level or higher.

[0064] Through Figure 11, it can be seen that the similarity of the training data, which is a vibration signal, is low compared to the image due to problems such as device error or sample contamination.

[0065] Accordingly, with reference to FIG. 11, the training data selection unit (130) sets a pre-set similarity standard and selects only individual factors for which the overlap rate of the data distribution area between the standard training data and the sample data is greater than or equal to the pre-set similarity standard, and selects them as training data.

[0066] In one example, the training data selection unit (130) sets the pre-set similarity criterion to 80% and can select learning factors (CCD: #1, #2, #3, #4, #5) with a similarity of 80% or more as training data.

[0067] Accordingly, it can be seen that the accuracy is more than 15% higher when a learning model is trained using learning factors with a similarity of 80% or higher as training data, compared to when a learning model is trained using all learning factors regardless of similarity as training data.

[0068] As a second method, the training data selection unit (130) converts each individual learning factor into at least one dimensionality reduction method among TSNE and PCA, LDA, UMAP, etc., calculates similarity between the dimensionality-reduced 2D data and reference data, and can distinguish between training data and non-training data based on the result of the similarity calculation.

[0069] FIG. 7 illustrates an example of comparing the overlap rate of two-dimensional data in which individual learning factors of sample data are converted into dimensionality reduction according to an embodiment of the present invention.

[0070] As illustrated in FIG. 7, the training data selection unit (130) can reduce the dimensionality of the learning individual factors A to I, which are the second sample data, to two dimensions, and then calculate the overlap rate with the first sample data of the same two dimensions. Subsequently, based on the overlap rate, the similarity between the second sample data and the first sample data is calculated as a result.

[0071] As illustrated in FIG. 12, the training data selection unit (130) may select training data by distinguishing between normal and abnormal shapes of vibration signals through machine learning and selecting only the normal signals.

[0072] Referring to Figure 12, before selecting the training data, it can be seen that the vibration signal in the overall DB similarity table is below the standard normal range of 80%.

[0073] Accordingly, the training data selection unit (130) can distinguish between normal and abnormal values ​​using an artificial intelligence model that has learned the shape of the vibration signal, remove the abnormal data, and include only the normal data in the training data.

[0074] Accordingly, as a result of recalculating the DB similarity, it can be confirmed that the training data similarity has increased compared to before due to the training data composed of selected vibration signals.

[0075] The learning unit (140) can train a nanoparticle estimation model using training data selected by the training data selection unit (130).

[0076] For example, the learning unit (140) can machine learn a nanoparticle estimation model. Machine learning can research and build a system and algorithms for training and estimating based on empirical data and improving its own performance.

[0077] The learning unit (140) learns a model composed of an artificial neural network using training data. The learning unit (140) can determine optimized model parameters of the artificial neural network by repeatedly training the artificial neural network using various learning techniques. In this specification, an artificial neural network whose parameters are determined by learning using training data may be referred to as a machine learning model or a trained model. At this time, the trained model can be used to infer result values ​​for new input data other than the training data. That is, the nanoparticle estimation model can be used to estimate the identity of nanoparticles for new unknown nanoparticle data.

[0078] The learning unit (140) may be configured to receive, classify, store, and output information to be used for data mining, data analysis, intelligent decision making, and machine learning algorithms and techniques.

[0079] The learning unit (140) may include one or more memory units configured to store data that is received, detected, sensed, generated, predefined, or output by another component, device, estimation device, or device communicating with the estimation device.

[0080] The learning unit (140) may include memory integrated into or implemented in the estimation device.

[0081] Specifically, the control unit (110) can induce the generation of laser-induced plasma by irradiating a pulsed laser beam onto a cell containing a solution containing standard nanoparticles with known size or type information in order to collect sample data, i.e., candidate training data, and can obtain feedback signals emitted as light, shock waves, and sound by the laser-induced plasma. At this time, the laser pulses can be varied differently, and a corresponding feedback signal can be obtained for each laser pulse.

[0082] To obtain a feedback signal emitted as light, the control unit (110) can obtain an image by capturing the light emitted from the laser-induced plasma through a camera, and obtain information on the position coordinates of the emitted light, the intensity of the light, and the magnitude of the light corresponding to the direction of light emission from the image.

[0083] In order to obtain a feedback signal emitted as a shock wave, the control unit (110) obtains the shock wave generated by the laser-induced plasma as a vibration signal through a vibration sensor, and obtains the frequency and intensity of the shock wave signal emitted by the laser-induced plasma from the vibration signal.

[0084] In order to obtain a feedback signal emitted as sound, the control unit (110) can obtain the sound generated by the laser-induced plasma as a sound signal and obtain the frequency and intensity of the sound signal emitted by the laser-induced plasma from the sound signal.

[0085] Additionally, the control unit (110) can measure the laser intensity before the pulsed laser beam is incident on the cell and measure the laser intensity after the pulsed laser beam is transmitted through the cell. At this time, the laser intensity can be measured using a dedicated measuring device or a known measuring method.

[0086] Additionally, the control unit (110) may obtain the laser pulse through setting information regarding the laser pulse of the laser beam emitting device or through user input information, and is not limited to these methods.

[0087] The control unit (110) can acquire feedback signals for each standard nanoparticle for each laser pulse, and acquire the intensity of each laser incident on the cell for each laser pulse and the intensity of each laser transmitted through the cell to build a database, thereby storing the database in memory (230).

[0088] Accordingly, the size or type of standard nanoparticles can be used as labeling data, and the intensity of each laser incident on a cell containing standard nanoparticles with known size or type information for each laser pulse, the intensity of each laser transmitted through the cell, and the feedback signal generated by the laser-induced plasma for each laser pulse can be obtained as sample data (candidates for training data).

[0089] The sample data obtained in this way can be separated into training data and non-training data by the aforementioned training data selection unit (130).

[0090] The learning unit (140) can machine learn a nanoparticle estimation model in solution based on training data selected by the training data selection unit (130).

[0091] The learning unit (140) can train the machine learning model to infer the size or type of nanoparticles, which are labeling data, when the intensity of the laser incident on the cell for each laser pulse, the intensity of the laser transmitted through the cell for each laser pulse, and the feedback signal generated by the laser-induced plasma for each laser pulse are input as training data to the machine learning model.

[0092] At this time, the learning model may infer result values ​​while mounted on the machine learning model learning device (100), or it may be transmitted to and mounted on another device such as a nanoparticle estimation device (200) via a communication device.

[0093] Additionally, when the learning model is updated, the updated learning model can be transmitted to and mounted on another device, such as an estimation device (100), via a communication device.

[0094] The nanoparticle estimation model can be trained using training data and used as a learning model.

[0095] In addition, the present invention may utilize supervised learning to train a machine learning model with labels for the training data. The labels serve as the correct answers that the nanoparticle estimation model must infer when the training data is input into the nanoparticle estimation model, and may include the size and type of the nanoparticles.

[0096] Training data represents multiple features, and the labeling of the training data implies that labels—specifically, information regarding the size and type of nanoparticles—are mapped to the features represented by the training data. In this case, the training data can represent the features of the input object in the form of a vector.

[0097] The nanoparticle estimation model can infer a function regarding the correlation between training data and labeled data using training data and labeled data. Furthermore, the parameters of the machine learning model can be determined (optimized) through the evaluation of the function inferred by the nanoparticle estimation model.

[0098] In one embodiment of the present invention, information of at least one of the previously known size and type of standard nanoparticles is a label, and an image of a solution containing standard nanoparticles, a vibration signal, a sound signal, laser intensity incident on a cell for each laser pulse, and laser intensity transmitted through a cell for each laser pulse may be training data.

[0099] Specifically, the training data may include at least one of an image of a solution containing standard nanoparticles, the frequency and intensity of a vibration signal, the frequency and intensity of a sound signal, the laser intensity incident on the cell for each laser pulse, and the laser intensity transmitted through the cell for each laser pulse.

[0100] Training data and labels corresponding to the training data constitute a single training set, and can be input into a nanoparticle estimation model in the form of training data.

[0101] In one embodiment, a machine learning-based learning model may include at least one machine learning model among the following structures: Deep Neural Network (DNN), Scalable Vector Extension (SVE), XGBoost algorithm, Light Gradient Boosting Model (GBM), Support Vector Machine (SVM), Gaussian Process Model (GPM), Random Forest (RF), Convolutional Neural Network (CNN), Region-based CNN (R-CNN), Convolutional Recursive Neural Network (C-RNN), Fast R-CNN, Faster R-CNN, Region-based Fully Convolutional Network (R-FCN), You Only Look Once (YOLO), and Single Shot Multibox Detector (SSD). However, any machine learning model other than those mentioned above may be applied to the present invention.

[0102] Meanwhile, the control unit (110) can identify the cause of the decrease in similarity based on the second sample data classified as non-training data because the difference between the distributions of the data distribution area does not fall within the standard normal range. Then, it can take measures to improve the identified cause of the decrease in similarity.

[0103] FIG. 4 is a block diagram illustrating the configuration of a nanoparticle estimation device according to an embodiment of the present invention, and FIG. 9 is a flowchart for explaining a nanoparticle estimation method according to an embodiment of the present invention.

[0104] Referring to FIG. 4, a nanoparticle estimation device (200) according to an embodiment of the present invention may include a control unit (210), a preprocessing unit (220), and an estimation unit (230).

[0105] Referring to FIGS. 4 and FIGS. 9, the control unit (210) can control the overall operation of the nanoparticle estimation device (200).

[0106] The preprocessing unit (220) can induce the generation of laser-induced plasma by controlling a laser emitting device to irradiate a pulsed laser beam onto a cell containing a solution containing unknown nanoparticles whose size or type information is to be estimated.

[0107] That is, a laser-induced plasma can be generated by directing a pulsed laser beam in a uniaxial direction into a cell containing a solution containing unknown nanoparticles so that the focus is formed inside the cell. At this time, the preprocessing unit (220) emits a laser beam for each laser pulse, and can obtain the intensity of the laser incident on the cell corresponding to each laser pulse, the intensity of the laser transmitted through the cell, and a feedback signal generated by the laser-induced plasma for each laser pulse.

[0108] To generate a laser-induced plasma, an Nd:YAG pulsed laser can be used, and the laser beam can be focused inside a fluid cell using a focusing lens. Since the laser beam contains numerous photons, it stimulates atoms, causing them to transition to an excited state and then emit light, shock waves, and sound as they return to the ground state.

[0109] Accordingly, the preprocessing unit (220) can obtain a feedback signal based on light, shock waves, and sound emitted by the laser-induced plasma.

[0110] That is, light emitted from a laser-induced plasma is captured through a camera to acquire an image, and from the image, information regarding the position coordinates of the emitted light corresponding to the direction of light emission, the intensity of the light, and the magnitude of the light can be obtained.

[0111] And, the preprocessing unit (220) can obtain the shock wave generated by the laser-induced plasma as a vibration signal through a vibration sensor, and obtain the frequency and intensity of the shock wave signal emitted by the laser-induced plasma from the vibration signal.

[0112] And, the preprocessing unit (220) can obtain the sound generated by the laser-induced plasma as a sound signal and obtain the frequency and intensity of the sound signal emitted by the laser-induced plasma from the sound signal.

[0113] In this way, the pretreatment unit (220) can collect unknown nanoparticle data by analyzing a solution containing unknown nanoparticles.

[0114] In particular, in the present invention, the preprocessing unit (220) can compare the reference training data with the unknown nanoparticle data to determine whether the unknown nanoparticle data falls within the category of the training data.

[0115] FIG. 10 illustrates an example of the result of the preprocessing unit (220) comparing training data and unknown nanoparticle data.

[0116] Referring to FIG. 10, the preprocessing unit (220) converts the training data and the unknown nanoparticle data into at least one dimensionality reduction method among TSNE and PCA, and then determines whether the location of the dimensionality-reduced unknown nanoparticle data is included in the KDE prediction area of ​​the dimensionality-reduced training data, thereby determining whether to select the unknown nanoparticle data as data to be input into the nanoparticle estimation model.

[0117] If the dimensionality-reduced unknown nanoparticle data falls within the KED prediction area (0% ≤ Density ≤ 100%) of the dimensionality-reduced training data, it is determined that the unknown nanoparticle data corresponds to the category of the training data, and the unknown nanoparticle data can be selected as data to be input into the nanoparticle estimation model.

[0118] On the other hand, if the dimensionality-reduced unknown nanoparticle data does not fall within the KED prediction area (0% ≤ Density ≤ 100%) of the dimensionality-reduced training data, it is determined that the unknown nanoparticle data does not fall within the category of the training data, and thus the unknown nanoparticle data may not be selected as data to be input into the nanoparticle estimation model.

[0119] In Figure 10, since the star-marked unknown nanoparticle data is included within the KDE prediction area of ​​the training data, the five star-marked unknown nanoparticle data can be selected as data to be input into the nanoparticle estimation model.

[0120] The estimation unit (230) can estimate the size and type of unknown nanoparticles by inputting the unknown nanoparticle data into a nanoparticle estimation model learned by the nanoparticle estimation model learning device (100) above when the unknown nanoparticle data falls within the category of training data. The nanoparticle estimation model corresponds to an artificial intelligence model learned through training data described with reference to FIGS. 3 and FIGS. 8.

[0121] Meanwhile, the control unit (210) can take measures to improve the learning model based on unknown nanoparticle data that is determined not to be in the category of training data.

[0122] The description of the embodiments of the present invention described above is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical spirit or essential features of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0123] The scope of the present invention is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present invention.

Claims

1. A step of preparing first sample data corresponding to a first standard nanoparticle-containing solution; A step of collecting second sample data corresponding to a second standard nanoparticle-containing solution; A step of calculating the similarity between the first sample data and the second sample data; A step of distinguishing between training data and non-training data based on the similarity calculation result; and A method for training a nanoparticle estimation model, comprising the step of training a nanoparticle estimation model using the above training data.

2. In Paragraph 1, The step of collecting the first sample data or the second sample data is A step of acquiring raw data composed of a one-dimensional or two-dimensional matrix from a measuring instrument; and A method for learning a nanoparticle estimation model, comprising the step of converting the features of the data measured through machine learning or statistical processing into one-dimensional numerical values ​​from the above raw data and collecting them as a set of learning factors.

3. In Paragraph 1, The step of preparing the first sample data above is, A nanoparticle estimation model learning method that collects training data previously used through the above nanoparticle estimation model as the above first sample data.

4. In Paragraph 1, The step of calculating the similarity between the first sample data and the second sample data is: A nanoparticle estimation model learning method that calculates similarity based on the overlap rate between a first data distribution area of ​​an individual factor constituting the first sample data and a second data distribution area of ​​an individual sample factor constituting the second sample data.

5. In Paragraph 1, The step of calculating the similarity between the first sample data and the second sample data is: A nanoparticle estimation model learning method that calculates similarity based on the overlap rate of a two-dimensional data region obtained by transforming the individual sample factors constituting the first sample data and the individual sample factors constituting the second sample data through dimensionality reduction.

6. A step of collecting unknown nanoparticle data by analyzing a solution containing unknown nanoparticles; A step of comparing the training data described in claim 1 with the unknown nanoparticle data to determine whether the unknown nanoparticle data falls within the category of the training data; and A nanoparticle estimation method comprising the step of estimating the size and type of the unknown nanoparticle by inputting the unknown nanoparticle data into a nanoparticle estimation model described in claim 1 when the unknown nanoparticle data corresponds to the range of the training data.

7. In Paragraph 6, The step of determining whether the above unknown nanoparticle data corresponds to the category of the above training data is, A nanoparticle estimation method that transforms the unknown nanoparticle data and the training data by dimensionality reduction, and then determines whether the location of the unknown nanoparticle data corresponds to the Kernel Density Estimation (KDE) prediction region of the training data.

8. In Paragraph 6, The step of collecting unknown nanoparticle data by analyzing a solution containing unknown nanoparticles is, A step of generating laser-induced plasma by irradiating a pulsed laser beam into a cell containing a solution; A step of obtaining a feedback signal based on light, shock waves, and sound generated by the laser-induced plasma; A step of acquiring the above feedback signal, the intensity of the laser incident on the cell, and the intensity of each laser transmitted through the cell for each laser pulse; A nanoparticle estimation method for collecting unknown nanoparticle data based on the feedback signal obtained for each laser pulse, the intensity of the laser, and the intensity of the laser transmitted through the cell.

9. A sample data collection unit that prepares first sample data corresponding to a first standard nanoparticle-containing solution and collects second sample data corresponding to a second standard nanoparticle-containing solution; A training data selection unit that calculates the similarity between the first sample data and the second sample data and distinguishes between training data and non-training data based on the similarity calculation result; and A nanoparticle estimation model learning device comprising a learning unit that trains a nanoparticle estimation model using the above training data.

10. In Paragraph 9, The above data collection unit is, A nanoparticle estimation model learning device that acquires raw data composed of a one-dimensional or two-dimensional matrix from a measuring instrument, and is a set of learning factors that convert the features of the measured data into one-dimensional numerical values ​​through machine learning or statistical processing of the raw data.

11. In Paragraph 9, The above data collection unit is, A nanoparticle estimation model learning device that collects training data previously used through the above nanoparticle estimation model as the above first sample data.

12. In Paragraph 10, The above training data selection unit is, A nanoparticle estimation model learning method that calculates similarity based on the overlap rate between a first data distribution area of ​​an individual factor constituting the first sample data and a second data distribution area of ​​an individual sample factor constituting the second sample data.

13. In Paragraph 10, The above training data selection unit is, A nanoparticle estimation model learning device that calculates similarity based on the overlap rate of a two-dimensional data region obtained by transforming the individual sample factors constituting the first sample data and the individual sample factors constituting the second sample data through dimensionality reduction.

14. A data preprocessing unit that analyzes a solution containing unknown nanoparticles to collect unknown nanoparticle data, compares the training data described in claim 1 with the unknown nanoparticle data, and determines whether the unknown nanoparticle data falls within the category of the training data; and A nanoparticle estimation device comprising an estimation unit that, when the unknown nanoparticle data falls within the category of the training data, inputs the unknown nanoparticle data into a nanoparticle estimation model described in claim 1 to estimate the size and type of the unknown nanoparticle.

15. In Paragraph 14, The above data preprocessing unit is, A nanoparticle estimation device that transforms the above unknown nanoparticle data and the above training data by dimensionality reduction, and then determines whether the location of the above unknown nanoparticle data corresponds to the Kernel Density Estimation (KDE) prediction region of the above training data.

16. In Paragraph 14, The above data preprocessing unit is, A nanoparticle estimation device that generates a laser-induced plasma by injecting a pulsed laser beam into a cell containing a solution, obtains a feedback signal based on light, shock waves, and sound generated by the laser-induced plasma, obtains the feedback signal, the intensity of the laser incident on the cell, and the intensity of each laser transmitted through the cell for each laser pulse, and collects the unknown nanoparticle data based on the feedback signal obtained for each laser pulse, the intensity of the laser, and the intensity of the laser transmitted through the cell.

Images