Depth estimation method for surface flaw

A CNN-based method for flaw depth estimation in metal materials addresses inaccuracy by integrating image size and type information, enhancing precision and reducing error in flaw depth prediction.

JP2024164416APending Publication Date: 2024-11-27DAIDO STEEL CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
JP2023079871
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-15
Publication Date
2024-11-27

AI Technical Summary

Technical Problem

Conventional methods for estimating the depth of surface flaws in metal materials using magnetic particle flaw detection are inaccurate due to reliance on aspect ratio alone, neglecting small wavy patterns and local particle density variations, leading to ambiguous flaw type determination and depth estimation.

Method used

A method utilizing a trained convolutional neural network (CNN) to analyze magnetic particle pattern images, incorporating image size and flaw type information, and optionally resizing or adding zero-filled images to maintain aspect ratio, for precise flaw depth estimation.

Benefits of technology

Accurately estimates the depth of surface flaws on metal materials by improving prediction accuracy and reducing information loss, achieving low RMSE of 0.331-0.426, surpassing conventional rule-based processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024164416000001_ABST
    Figure 2024164416000001_ABST
Patent Text Reader

Abstract

To provide a depth estimation method for a surface flaw which can accurately estimate a flaw depth corresponding to a flaw type of a surface flaw of a metal material in magnetic particle flaw inspection.SOLUTION: A depth of a surface flaw of a flaw type corresponding to a flaw image in a magnetic particle pattern image is estimated by inputting the magnetic particle pattern image of a metal material obtained with magnetic particle flaw inspection to a convolutional neural network of an image recognition unit 21 constituted in a processing device 2. In this case, image size information of the flaw image and information about the flaw type are input to a fully-connected layer 23 that constitutes an output layer of the convolutional neural network.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a method for estimating the depth of a surface flaw, and more particularly to a method for estimating the depth of a surface flaw of a metal material from a magnetic powder pattern image of the metal material obtained by magnetic particle inspection. [Background technology]

[0002] In Patent Document 1, fluorescent magnetic powder is scattered on the surface of a magnetized metal material, the flaw location is identified from the magnetic powder pattern image drawn by the fluorescent magnetic powder, and the flaw type is determined from the aspect ratio of the flaw location. Then, a depth index value is obtained by multiplying the average width of the flaw location by the average brightness, and the flaw depth is calculated and estimated by multiplying the index value by a coefficient obtained in advance according to the flaw type. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 9-210969 Summary of the Invention [Problem to be solved by the invention]

[0004] In the above-mentioned conventional flaw depth estimation method, the type of flaw is determined from the aspect ratio of the flaw area. In reality, however, workers estimate the flaw depth sensorily, taking into account small wavy patterns and local changes in the shading of the magnetic powder, etc., and it is difficult to determine the type of flaw uniquely from the aspect ratio of the flaw area alone. As a result, there is a problem in that it is also difficult to accurately estimate the flaw depth.

[0005] Therefore, the present invention is intended to solve such problems, and aims to provide a method for estimating the depth of surface defects that can accurately estimate the depth of defects corresponding to the type of defect on the surface of a metal material in magnetic particle inspection. [Means for solving the problem]

[0006] In order to achieve the above object, in the first invention, a magnetic powder pattern image of a metal material obtained by magnetic particle inspection is input into a trained convolutional neural network to estimate the depth of a surface defect of a defect type corresponding to a defect image in the magnetic powder pattern image.

[0007] According to the first invention, by inputting a magnetic powder pattern image into a trained convolutional neural network, the magnetic powder pattern itself can be recognized, making it possible to accurately estimate the defect depth according to the defect type of the surface defect of a metal material.

[0008] In the second invention, the image size information of the defect image and the defect type information are input to a fully connected layer that constitutes an output layer of the convolutional neural network.

[0009] According to the second invention, by inputting image size information and defect type information into the fully connected layer of the convolutional neural network, it is possible to more accurately estimate the defect depth according to the defect type of the surface defect of the metal material without being affected by resizing (hereinafter referred to as "resizing") of the magnetic powder pattern image to a fixed size and fixed aspect ratio when inputting it into the convolutional neural network.

[0010] In the third invention, when the magnetic powder pattern image is input to the convolutional neural network, a zero-filling image is appropriately added to the edges of the magnetic powder pattern image to make the image sizes of all the magnetic powder pattern images the same.

[0011] In the third invention, it is possible to more accurately estimate the flaw depth according to the type of surface flaw on a metal material without being affected by resizing of the magnetic powder pattern image when it is input to a convolutional neural network. Effect of the Invention

[0012] As described above, according to the surface flaw depth estimation method of the present invention, it is possible to accurately estimate the flaw depth according to the type of surface flaw in a metal material in magnetic particle inspection. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram showing a schematic configuration of a magnetic flaw detection device. [Diagram 2] FIG. 1 is a diagram showing the configuration of a main part of a convolutional neural network. [Diagram 3] FIG. 13 is a diagram showing how a magnetic powder pattern image of a linear flaw is resized. [Figure 4] FIG. 13 is a diagram showing how a magnetic powder pattern image of a scab defect is resized. [Diagram 5] FIG. 2 is a diagram illustrating a configuration of a processing device. [Figure 6] FIG. 13 is a diagram showing an example of a magnetic powder pattern image for obtaining size data and determining the type of defect. [Figure 7] FIG. 13 is a diagram comparing predicted values ​​of flaw depth by an operator and a processing device. [Figure 8] FIG. 13 is a diagram showing a state in which a zero-filling image is added to a magnetic powder pattern image. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] It should be noted that the embodiment described below is merely an example, and various design improvements made by those skilled in the art without departing from the gist of the present invention are also included in the scope of the present invention.

[0015] (First embodiment) Figure 1 shows the schematic configuration of a magnetic flaw detector that uses the method of the present invention. A square steel piece M, which is the metal material to be inspected, is placed on a support S, and its inspection surface M1 is magnetized by a known magnetizer (not shown), and then fluorescent magnetic powder liquid is sprayed on it to create a magnetic powder pattern. A rail R is installed above the steel piece M in parallel with and along its length, and a flaw detector 1 is provided that can move along the rail R by guide rollers R1.

[0016] A black light 11 and a camera 12 are provided inside the flaw detector 1 facing the inspection surface M1 of the steel piece M, and the magnetic powder pattern of the magnetic powder that fluoresces when irradiated with ultraviolet light from the black light 11 is captured by the camera 12. Then, as the flaw detector 1 moves along the steel piece M, the magnetic powder pattern images of the inspection surface M1 of the steel piece M are sent successively to the processing device 2.

[0017] The processing device 2 has a built-in computer, and is composed of an image recognition unit 21 that is composed of the main parts of a trained CNN (convolutional neural network), and a defect size / defect type input unit 22. The configuration of the main CNN parts of the image recognition unit 21 is shown in Fig. 2, in which a required number of convolution layers 211 and a required number of pooling layers 212 are alternately provided, and finally a smoothing layer 213 is provided, and the two-dimensional image data (three-dimensional if it is a color image) from which features have been extracted is converted into a one-dimensional data string and output.

[0018] Incidentally, when a magnetic powder pattern image is read into the image recognition unit 21, the magnetic powder pattern image is usually converted to the same size, that is, resized. Examples of resizing are shown in Figs. 3 and 4. Fig. 3 shows an example of a linear flaw, and Fig. 4 shows an example of a scab flaw. As a result, size information of the magnetic powder pattern image is lost, and the accuracy of estimating the depth of the surface flaw is impaired. Therefore, in this embodiment, the above-mentioned flaw image size / flaw type input unit 22 is provided for inputting the flaw image size (area or diameter) in the flaw pattern image before resizing, and also inputting flaw type data.

[0019] The configuration of the processing device 2 is shown again in Fig. 5. In the flaw image size / flaw type input unit 22, the input flaw image size is converted into a continuous value or binary number, and the flaw type is one-hot encoded (an array in which the value is "1" if the flaw is of interest and the value is "0" otherwise). Then, the image data from the image recognition unit 21 converted into a one-dimensional data string, the flaw image size data converted into a continuous value or binary number, and the one-hot encoded flaw type data are input to a required number of fully connected layers 23 that constitute the output layer of the CNN, and flaw depth prediction data is output from the final fully connected layer 23.

[0020] The flaw image size data may be measured and calculated from the magnetic powder pattern image within the processing device 2, and the flaw type data may be determined and output from the magnetic powder pattern image by configuring a separate CNN within the processing device 2. As an example of a separate CNN, as shown in Fig. 6, by incorporating object detection technology, it is possible to simultaneously obtain the size data and determine the flaw type (A type, B type, etc.).

[0021] The CNN of the processing device 2 described above is given the necessary amount of training data to learn. An example of the training data is shown in Table 1, which is a set of a magnetic powder pattern image and the corresponding defect type and defect depth. The defect image size used as training data is obtained from the magnetic powder pattern image. Table 2 shows an example of the learning specifications.

[0022] [Table 1]

[0023] [Table 2]

[0024] The results of evaluating the predictive performance of the CNN trained and learned according to the above specifications are shown in Fig. 7. Fig. 7 plots both defect depths, with the horizontal axis representing the defect depth predicted by the operator and the vertical axis representing the defect depth predicted by the CNN. The RMSE (Root Mean Squared Error) was 0.375 as shown in Table 3, and the inclusion of a defect image size / defect type input unit ensured sufficiently high defect depth prediction accuracy. The processing time of the computer (GPU) in this case was 0.6 ms (Table 3), which was the same as that of each of the embodiments described below.

[0025] Second embodiment As described above, when the magnetic powder pattern image is read into the image recognition unit 21, resizing is performed, but there is a problem in that this resizing also significantly changes the aspect ratio of the defect, particularly in the case of a linear defect, as shown in Fig. 3. Therefore, as shown in Fig. 8, a zero-filled image fa with zero brightness is added to the edges of the magnetic powder pattern image (the upper and lower edges in this embodiment) to make all the magnetic powder pattern images input to the image recognition unit 21 the same size.

[0026] By doing so, the aspect ratio of the resized flaw is well maintained, and the RMSE of the flaw depth prediction data output from the processing device 2 is further improved to 0.331 as shown in Table 3.

[0027] Third embodiment Furthermore, even if a zero-filled image with zero brightness is added to the edges of the magnetic powder pattern image and all the magnetic powder pattern images input to the image recognition unit 21 are made the same size, without providing a flaw image size / flaw type input unit, the RMSE of the flaw depth prediction data output from the processing device 2 is 0.414, as shown in Table 3, and the flaw depth prediction accuracy is sufficiently high.

[0028] (Fourth embodiment) Even when the defect image size / defect type input section is not provided and a zero-filled image is not added, the RMSE of the defect depth prediction data output from the processing device 2 is 0.426, as shown in Table 3, and the defect depth prediction accuracy is sufficiently high.

[0029] [Table 3]

[0030] As described above, in the third and fourth embodiments, image recognition incorporating CNN has achieved prediction accuracy that is sufficiently higher than that of conventional rule-based processing, and further, in the first and second embodiments, even higher accuracy is achieved by processing that complements information loss in CNN. [Explanation of symbols]

[0031] 1... flaw detector, 11... black light, 12... camera, 2... processing device, 21... image recognition unit, 22... flaw size / flaw type input unit, M... steel piece, R... rail, S... support stand.

Claims

1. A method for estimating the depth of surface defects, comprising inputting a magnetic powder pattern image of a metal material obtained by magnetic particle inspection into a trained convolutional neural network to estimate the depth of a surface defect of a defect type corresponding to a defect image in the magnetic powder pattern image.

2. The method for estimating a depth of a surface defect according to claim 1, wherein image size information of the defect image and information on the defect type are input to a fully connected layer constituting an output layer of the convolutional neural network.

3. 3. The method for estimating the depth of a surface defect according to claim 1 or 2, wherein when the magnetic powder pattern image is input to the convolutional neural network, zero-filled images are appropriately added to edges of the magnetic powder pattern image so that the image sizes of all of the magnetic powder pattern images are made uniform.

Citation Information

Patent Citations

  • Automatic magnetic particle inspection device

    JP1997210969A