Fluid behavior analysis method

Abstract

A fluid behavior analysis method is provided. The fluid behavior analysis method comprises the steps of: generating an original image of a fluid by photographing the behavior of the fluid; generating a corrected image by correcting the original image; and generating a classification model by performing machine learning using the corrected image as image data.

Description

Fluid behavior analysis method

[0001] The present invention relates to a method for analyzing fluid behavior.

[0002] Traditionally, analysis of fluid thinning behavior has been based on rheological theory. Based on this theory, rheological properties such as extensional viscosity and relaxation time were derived, and the characteristic morphology of the fluid at the moment of thinning and cessation was used to classify the fluid or explain its morphological characteristics. However, this conventional analysis method can be influenced by the analyst's rheological knowledge during the process of deriving the results, resulting in discrepancies or reduced objectivity.

[0003] The present invention provides a fluid behavior analysis method having excellent performance.

[0004] Other objects of the present invention will become apparent from the following detailed description and the accompanying drawings.

[0005] A fluid behavior analysis method according to embodiments of the present invention includes a step of photographing the behavior of a fluid to create an original image of the fluid, a step of correcting the original image to create a corrected image, and a step of performing machine learning using the corrected image as image data to create a classification model.

[0006] The fluid behavior analysis method according to embodiments of the present invention exhibits excellent performance. For example, it can analyze the thinning behavior of a fluid simply, quickly, and objectively. Furthermore, machine learning reduces human error and maximizes information collection from data, facilitating the analysis of the rheological behavior of complex fluids. Furthermore, the fluid behavior analysis method can be applied to rapid inspection of coating quality for industrial equipment.

[0007] FIG. 1 is a drawing for explaining a step of generating an original image of a fluid in a fluid behavior analysis method according to one embodiment of the present invention.

[0008] Figure 2 shows a flowchart for finding an image (or frame) where the thinning behavior of a fluid begins among multiple images (multi-images or multi-frames) generated by taking continuous shots.

[0009] Figure 3 shows a flowchart for finding an image (or frame) where the thinning behavior of a fluid ends among multiple images (multi-images or multi-frames) generated by taking continuous shots.

[0010] FIG. 4 is a drawing for explaining a correction image generation step of a fluid behavior analysis method according to one embodiment of the present invention.

[0011] Figure 5 shows a flowchart of performing whitening during the correction image generation step of Figure 4.

[0012] Figure 6 shows a flowchart of performing stacking during the correction image generation step of Figure 4.

[0013] Figure 7 shows a flowchart of performing cropping and centering during the correction image generation step of Figure 4.

[0014] Figure 8 shows an example of a PCA process in the machine learning step of a fluid behavior analysis method according to one embodiment of the present invention.

[0015] Figure 9 shows an example of a PCA and K-NN-based classification model generated by a fluid behavior analysis method according to one embodiment of the present invention.

[0016] Figure 10 shows a flowchart for performing machine learning using PCA and K-NN.

[0017] Figure 11 shows a visualized form of a new feature obtained through PCA.

[0018] Figure 12 shows a scatter graph of data in a new coordinate system obtained through PCA.

[0019] Figure 13 shows the results of verifying the accuracy of the classification model created using PCA and K-NN.

[0020] Figure 14 shows an example of data amplification using principal components.

[0021] Hereinafter, the present invention will be described in detail through examples. The objectives, features, and advantages of the present invention will be readily understood through the following examples. The present invention is not limited to the examples described herein and may be embodied in other forms. The examples introduced herein are provided to ensure that the disclosure is thorough and complete and to sufficiently convey the spirit of the present invention to those skilled in the art. Therefore, the present invention should not be limited by the following examples.

[0022]

[0023] A fluid behavior analysis method according to embodiments of the present invention includes a step of photographing the behavior of a fluid to create an original image of the fluid, a step of correcting the original image to create a corrected image, and a step of performing machine learning using the corrected image as image data to create a classification model.

[0024] The behavior of the fluid may include a thinning behavior of the fluid. The starting point of the thinning behavior may be the point at which the angle formed between the needle discharging the fluid and the drop of the fluid becomes acute, and the ending point of the thinning behavior may be the point at which the fluid is completely cut off.

[0025] The above original image may include a plurality of images (multi-images) generated by continuously taking pictures, and the step of generating the corrected image may include a step of converting the multi-image into a black and white image, a step of merging the multi-images converted into black and white images into one image, and a step of aligning the merged images.

[0026] The above machine learning can be performed using principal component analysis (PCA) and the K-nearest neighbor algorithm (K-NN).

[0027] The above classification model generation step may include a step of extracting features that distinguish the image data, visualizing the features, clustering the image data based on similarity, and a step of classifying the clustered image data by labeling it based on geometric distance.

[0028] The above image data can be used as training data and test data in the machine learning.

[0029] The above fluid behavior analysis method may further include a step of verifying the accuracy of the classification model.

[0030]

[0031] [Creating original images of fluids (capturing fluid behavior)]

[0032]

[0033] FIG. 1 is a drawing for explaining a step of generating an original image of a fluid in a fluid behavior analysis method according to one embodiment of the present invention.

[0034] Referring to Figure 1, the thinning behavior of a fluid can be created and captured using a DoS-CaBER (Dripping onto Substrate - Capillary Break-up Extensional Rheometer) device and an ultra-high-speed camera. There are no limitations on the measurement equipment, and various measurement devices can be utilized. Since the brightness and shape of the measured image can affect the subsequent steps of generating a corrected image (image preprocessing) and the results of machine learning, when capturing the fluid behavior, the experimental environment and equipment must be set to identical conditions, excluding the fluid type.

[0035] Among the components of DoS-CaBER, the pump used was the PHD ULTRA Infusion Only model (Harvad Apparatus), the needle used was the N718 NDL 6 model (HAMILTON), and the substrate used was the slide glass of MARIENFELD's HSU-1000412 model. The high-speed camera used was the VEO710L 72G Mono Camera (Phantom) model equipped with a 5x magnification EDMUNDOPTICS lens. The pump discharge speed was set to 0.2 mm / hr, and the distance between the needle and the substrate was set to 3 times the inner diameter of the needle. The resolution of the high-speed camera was set to 960px*600px, and filming was performed at 10,000fps. Under the above conditions, the behavior of a total of seven types of solutions, including CMC solution, PEO solution, Carbopol solution, PEO+CMC mixed solution, PEO+Carbopol mixed solution, CMC+Carbopol mixed solution, and PEO+CMC+Carbopol mixed solution, were measured. The preparation methods of the above solutions are as follows. The CMC solution was prepared as a 1 wt% aqueous solution using Mw=250,000 g / mol CMC, and the PEO solution was prepared as a 1 wt% aqueous solution using Mv=2,000,000 g / mol PEO. The Carbopol solution was prepared as a 0.14 wt% aqueous solution using Carbomer941. The remaining mixed solutions were prepared by mixing CMC solution, PEO solution, and Carbopol solution in equal ratios according to the type of mixed solution. The mixing was performed at a speed of 300 rpm for 24 hours at room temperature. For the above 7 types of solutions, 30 data were measured for each type, and the measured images were saved as multi-frames.

[0036]

[0037] Figure 2 shows a flow chart for finding an image (or frame) where the thinning behavior of a fluid begins among a plurality of images (multi-images or multi-frames) generated by continuously shooting, and Figure 3 shows a flow chart for finding an image (or frame) where the thinning behavior of a fluid ends among a plurality of images (multi-images or multi-frames) generated by continuously shooting.

[0038] Referring to FIGS. 2 and 3, multi-frames are loaded chronologically into a list. The start point of the thinning behavior is set to be the list index of the first frame where the angle between the needle and the fluid drop becomes acute. The end point of the thinning behavior should be the point where the fluid is completely cut off, so it is set to be the list index of the first frame where there is a row composed entirely of white pixels without any black pixels corresponding to the fluid within the set ROI.

[0039]

[0040] [Creating a corrected image (image preprocessing)]

[0041]

[0042] Figure 4 is a diagram illustrating the correction image generation step of a fluid behavior analysis method according to one embodiment of the present invention. To improve machine learning performance, an image preprocessing process is performed to remove unnecessary data from the original image and add useful data. The image preprocessing process is performed on the original image to generate a correction image.

[0043] Referring to Figure 4, image preprocessing can be broadly divided into three stages. The first stage is Whitening, which converts the original image to black and white and removes lighting noise. The second stage is Stacking, which merges continuously captured multi-images into a single image. This stacking process can add useful information about the overall thinning behavior of the fluid. The third stage is Cropping & Centering, which removes nozzles and underlying materials that are not related to the fluid characteristics and centers the image. Cropping and centering can improve the accuracy of machine learning.

[0044]

[0045] Fig. 5 shows a flow chart for performing whitening during the correction image generation step of Fig. 4, Fig. 6 shows a flow chart for performing stacking during the correction image generation step of Fig. 4, and Fig. 7 shows a flow chart for performing cropping and centering during the correction image generation step of Fig. 4.

[0046] Referring to Figure 5, the original image is converted into an outline image with white pixels using a transformation algorithm (Canny edge detection algorithm). The first and last white pixels in each row of the outline image are found, and all pixels between them are converted to white. The same process is performed for all rows to obtain an image with the outline filled with white. This process is repeated for all measured frames.

[0047] Referring to Fig. 6, multi-frames for which whitening has been completed are loaded into a list. The start and end indices are found through an algorithm for finding the start and end points of the thinning behavior of the fluid described in Figs. 2 and 3, and only the frames with indices between those indices are extracted and loaded into a new list. The frames stored in an array format in the extracted new list are divided by a specific value (8000 in the embodiment of the present invention). The frames in the array format within the list are added to create a single image.

[0048] Referring to Fig. 7, the X coordinates of pixels with high pixel values ​​for each row in the stacked image are loaded into a list. The average of the X coordinates loaded into the list is obtained to obtain the X coordinate at which the center line is located. An image with the center aligned is obtained by cropping the same number of pixels on the left and right (200 pixels each in the embodiment of the present invention) based on the center X coordinate. Since the needle and the substrate are fixed in the same position when shooting, the images of the needle and the substrate are cropped by considering the corresponding position and entering a fixed value.

[0049]

[0050] [Creating a classification model (performing machine learning)]

[0051]

[0052] Machine learning is performed on the generated corrected image (the preprocessed image). This machine learning can be used to analyze the fluid's thinning behavior and obtain information. This machine learning utilizes two algorithms: principal component analysis (PCA) and the K-Nearest Neighbor (K-NN) algorithm. Hyperparameters for each algorithm are determined to achieve optimal performance.

[0053] The above principal component analysis (PCA) extracts features that distinguish image data, visualizes the features, and clusters the data based on similarity. Figure 8 shows an example of the PCA process that reduces the dimensionality of data with p variables to data with i variables. The graph is expressed as reducing the dimensionality of two-dimensional data to one-dimensional data for simplicity of expression. Through this dimensionality reduction process, the original data is projected into a coordinate system with an optimized dimension and clustered, and at this time, a new variable called a principal component (PC) can be obtained. The K-nearest neighbor (K-NN) algorithm labels and classifies the clustered image data based on geometric distance. When the PCA and K-NN algorithms are applied to the training data, a PCA-based classification model that can determine the class of the test data is created, as shown in Figure 9.

[0054]

[0055] Figure 10 shows a flowchart for performing machine learning using PCA and K-NN.

[0056] Referring to Figure 10, PCA is performed using the PCA module of the scikit learn library. When performing PCA, the PCA components must be determined. Initially, the PCA component values ​​are set to the minimum value of 1, and after performing PCA, the explained variance method is used to derive the explanatory power. This process is repeated by adding 1 to the PCA component values ​​until the set threshold explanatory power (in the embodiment of the present invention, the explanatory power is 0.99) or higher. The K-NN module of the scikit learn library is applied. The k value, a hyperparameter of K-NN, is initially set to the minimum value of 1. The classification accuracy test is performed using the score method of the K-NN module. This process is repeated by adding 1 to the k value until the set threshold accuracy (for example, an accuracy of 90%) or higher. Hyperparameters that have the set threshold explanatory power and classification accuracy are determined, and a classification model is constructed based on them.

[0057]

[0058] [Classification Model Verification and Utilization]

[0059]

[0060] Figure 11 shows a visualized form of a new feature obtained through PCA, Figure 12 shows a scatter graph of data in a new coordinate system obtained through PCA, Figure 13 shows the results of verifying the accuracy of a classification model created using PCA and K-NN, and Figure 14 shows an example of utilizing data amplification using principal components.

[0061] Referring to Figure 11, by performing machine learning and using the PCA method (the components_ method of the PCA module), principal components (PCs) can be accessed. By resizing each PC to the original image shape (360px*400px in the embodiment of the present invention), a visualized form of the PC can be obtained.

[0062] Referring to Figure 12, PCA can be used to obtain PCs, and the PCA method (the pca.transform method in the PCA module) can be used to project the original data into a new coordinate system. By projecting the training data into this new coordinate system using this method and labeling each data point, a graph can be obtained.

[0063] Referring to Fig. 13, the hyper-parameter of the classification model was set to PCA component = 6, and the hyper-parameter of K-NN was calculated for all accuracies from 1 to 20 for an overall comparison of the models according to the conditions. Accuracy evaluation was performed on the classification model with a total of four conditions applied depending on whether image preprocessing was performed or not and whether PCA was performed or not, and as a result, it was confirmed that the highest accuracy value of 93 to 100% was achieved when both the correction image creation (image preprocessing) according to the fluid behavior analysis method of the present invention and PCA were applied.

[0064] Referring to Fig. 14, by using the PCA score value, which is the value of data in a new coordinate system obtained through the PCA method (pca.transform method), and the PCs that can be obtained through the PCA method (components_ method), a PCA reconstruction image can be obtained according to the number of PCA components. By arbitrarily setting the PCA score value or making a slight change, various forms of image data that have differences from the existing image can be obtained. Image data amplification can be performed through this process. In addition, the amplified artificial image is in the form of a stacked image, and if this image is decomposed into each frame and made into the form of an image, image data amplification is also possible.

[0065]

[0066] We have discussed specific embodiments of the present invention. Those skilled in the art will appreciate that the present invention can be implemented in modified forms without departing from its essential characteristics. Therefore, the disclosed embodiments should be considered illustrative rather than restrictive. The scope of the present invention is set forth in the claims, not the foregoing description, and all differences within the scope equivalent thereto should be construed as being encompassed by the present invention.

[0067]

[0068] The fluid behavior analysis method according to embodiments of the present invention exhibits excellent performance. For example, it can analyze the thinning behavior of a fluid simply, quickly, and objectively. Furthermore, machine learning reduces human error and maximizes information collection from data, facilitating the analysis of the rheological behavior of complex fluids. Furthermore, the fluid behavior analysis method can be applied to rapid inspection of coating quality for industrial equipment.

Claims

1. A step of creating an original image of the fluid by photographing the behavior of the fluid; A step of generating a corrected image by correcting the original image; and A method for analyzing fluid behavior, comprising a step of performing machine learning using the above-mentioned correction image as image data to create a classification model.

2. In paragraph 1, A method for analyzing fluid behavior, characterized in that the behavior of the fluid includes thinning behavior of the fluid.

3. In paragraph 2, The starting point of the above thinning behavior is the point at which the angle formed by the needle discharging the fluid and the drop of the fluid becomes acute. A fluid behavior analysis method characterized in that the end point of the above thinning behavior is the point at which the fluid is completely cut off.

4. In paragraph 1, The above original image includes multiple images (multi-images) generated by continuously shooting, The above correction image generation step is, A step of converting the above multi-image into a black and white image, A step of merging the multi-images converted to the above black and white images into one image, and A method for analyzing fluid behavior, characterized in that it comprises a step of aligning the merged image.

5. In paragraph 1, A fluid behavior analysis method characterized in that the above machine learning is performed using principal component analysis (PCA) and the K-nearest neighbor algorithm (K-NN).

6. In paragraph 1, The above classification model creation step is: A step of extracting features that distinguish the image data, visualizing the features, and clustering the image data according to similarity, and A fluid behavior analysis method characterized by including a step of classifying the above clustered image data by labeling it based on geometric distance.

7. In paragraph 1, A fluid behavior analysis method characterized in that the above image data is used as training data and test data in the machine learning.

8. In paragraph 1, A fluid behavior analysis method further comprising a step of verifying the accuracy of the above classification model.

Images