JP7318402B2 - Defect inspection method and defect inspection device - Google Patents

Description

The present invention relates to a defect inspection method and a defect inspection apparatus. More particularly, the present invention relates to an inspection method and inspection apparatus capable of detecting the presence or absence of defects in a specified depth region of an object to be inspected with high accuracy.

In recent years, industrial products such as electronic devices, composite materials, and resin molded products have become increasingly miniaturized and complicated. Demands on quality are becoming more stringent. Non-destructive inspection using X-rays, which inspects internal defects such as voids, cracks, and foreign matter inside an object that cannot be observed from the outside of the object, without destroying the object, is effective in solving these problems. The need for X-ray non-destructive inspection is increasing.

Japanese Patent Publication No. 2014-501818 JP-A-4-9606 Japanese Patent Application Laid-Open No. 2004-191112

However, since information overlaps in the depth direction in an X-ray inspection image obtained by general X-ray transmission imaging, there is a problem that it is difficult to extract the height position of the defect as information.

For example, in the mounting of electronic components on a double-sided board, if you want to inspect bonding failures by X-ray fluoroscopy and define bonding failures only on the front side (or back side) as defects and detect them, you can use general X-ray transmission imaging. In the obtained X-ray detection image, the image of the defect on the back side (or front side) overlaps, and the bonding failure on the back side (or front side) is also detected as a defect. That is, it is difficult to separately detect defects on the front side and defects on the back side.

In addition, carbon fiber reinforced plastic (CFRP), which is expected to be applied to aerospace applications, is often used as a laminated plate, and although it has high rigidity in the in-plane direction where the fiber is reinforced, The outward stiffness is weak and the material is highly anisotropic. There is a problem that delamination occurs due to stress in the out-of-plane direction, but it is useful for quality control, design, and improvement of products to know which layer delamination occurred in laminates. is. That is, for example, it is sometimes desired to define only delamination between the kth layer/k+1th layer as a defect and detect it by distinguishing it from other delamination. Since the images are detected overlapping, it is difficult to determine whether there is delamination between the kth layer/k+1th layer.

In addition, high-pressure hydrogen tanks mounted on fuel cell electric vehicles, which have been attracting attention in recent years, include, for example, a resin liner member that is lightweight and has excellent moldability, and a fiber-reinforced resin layer that covers the outside of the liner member. consists of A high-pressure gas tank manufactured using the molded article described in Patent Document 1 may undergo deformation or the like when repeatedly filling and releasing high-pressure gas (especially high-pressure hydrogen gas), which is a factor in reducing yield. was becoming The cause of such sudden anomaly is unknown, and there is no inspection method for it. Therefore, the present inventors have made intensive studies in order to solve the above-mentioned problems, and as a result, factors such as deformation of the tank exist, for example, in the jointed portion of the high-pressure tank member which is manufactured by dividing into two molded products. It was found that it was caused by impurities and voids. When a welding method generally used for joining resin molded products is used, a step of pressing the joints in order to achieve sufficient welding occurs. In this process, a swell of molten resin (hereinafter referred to as burr) is generated at the joint. Gaps and impurities in the burrs do not cause deformation of the high-pressure tank member. However, for the inspection of the inside of a resin molded product, as in the inspection method described in Patent Document 2, it is common to transmit X-rays and detect the presence or absence of internal impurities and voids from changes in the amount of transmission. is. This embodiment is shown in FIGS. FIG. 4 is a schematic diagram for explaining an example of a defect and a non-defect portion that cannot be discriminated in an inspection configuration based on general X-ray transmission imaging; In order to simplify the explanation, FIG. 3 shows only the joint cross section of the high-pressure tank member 2 on the side of the X-ray radiation means 1 . In this configuration, since the X-rays emitted from the X-ray radiation means 1 pass through both the joint portion to be inspected and the burr portion, the change in the transmission amount of the X-rays emitted from the X-ray radiation means 1 affects the joint. It was difficult to determine whether it was due to voids or impurities in the burr portion or due to voids or impurities in the burr portion. In addition, it is often used to specify the location of the defect by irradiating X-rays from multiple directions, but it is difficult to determine whether the defect is a burr or a joint from the location of occurrence.

X-ray computed tomography (abbreviation: CT) is known as means for obtaining information in the depth direction using X-rays. X-ray CT is a technique for constructing an internal image of an object by scanning the object using X-rays and processing it using a computer. Complicated internal shapes and positional information of defects can be obtained by the CT function. However, CT has problems such as the time it takes to measure, the size of an object to be inspected is limited, and the apparatus is expensive, making it difficult to use for in-line inspection of a production line.

On the other hand, Patent Document 3 discloses a method of picking up a plurality of inspection images under different optical conditions and judging the quality of an object to be inspected by neural network processing. A neural network is a mathematical model representing a neural network in the human brain, and has recently attracted attention in the field of image recognition. However, since the inspection method described in Patent Document 3 does not have an algorithm for calculating the height position of the defect candidate, it cannot be applied when the quality of the defect depends on the height position of the defect candidate.

The present invention has been made in view of such conventional problems, and in the X-ray nondestructive inspection of an object to be inspected such as an industrial product in which it is necessary to determine whether a defect candidate is a defect or not based on the height position of the defect candidate, An object of the present invention is to provide an inspection method and an inspection apparatus suitable for in-line inspection, capable of detecting the presence or absence of defects with high accuracy.

That is, a defect inspection apparatus related to one aspect of the present invention for solving the above problems comprises X-ray emitting means for emitting X-rays through a plurality of paths different from each other , and one or more X-rays for detecting X-rays transmitted through an object to be inspected. X-ray detection means and image processing means, wherein the image processing means identifies defect candidates using a trained neural network (a) for one or more images acquired by the X-ray detection means. defect candidate detection means for detecting; height position calculation means for calculating height positions of defect candidates using a neural network (b) trained from a plurality of images acquired by the X-ray detection means; and a sorting means for judging whether the defect candidate is good or bad from the position.

The defect inspection apparatus of the present invention uses one two-dimensional image data obtained by synthesizing the plurality of images as an explanatory variable, and performs regression using a two-dimensional convolutional neural network using the height position of the defect candidate as an objective variable. height position calculating means.

Further, the defect inspection apparatus of the present invention constructs one three-dimensional image data from the plurality of images, uses the three-dimensional image data as an explanatory variable, and detects the height position of the defect candidate by a three-dimensional convolutional neural network. It is preferable to have defect candidate height position calculation means for regressing using as an objective variable.

Also, in the defect inspection apparatus of the present invention, it is preferable that the X-ray radiation means comprises two or more X-ray radiation means.

Further, in the defect inspection apparatus of the present invention, the X-ray radiation means includes one or more X-ray radiation means and the at least X-ray radiation position moving means for moving one or more X-ray radiation means.

Further, in the defect inspection apparatus of the present invention, the X-ray radiation means comprises one or more X-ray radiation means and inspected object position moving means for moving the inspected object to two or more positions. is preferred.

Further, in the defect inspection apparatus of the present invention, the object to be inspected is preferably a member for a high-pressure tank.

Further, a defect inspection method related to one aspect of the present invention for solving the above problems is to radiate X-rays through a plurality of paths different from each other , detect the X-rays transmitted through the object to be inspected at one or more positions, and detect the X-rays. Defect candidates are detected using a trained neural network (a) for one or more X-ray images, and defect candidates are detected for a plurality of detected X-ray images using a trained neural network (b). is calculated, and the quality of defect candidates is selected from the height position.

Further, in the defect inspection method of the present invention, one two-dimensional image data obtained by synthesizing the plurality of images is used as an explanatory variable, and a two-dimensional convolutional neural network performs regression using the height position of the defect candidate as an objective variable. It is preferable to have defect candidate height position calculation means.

Further, in the defect inspection method of the present invention, one three-dimensional image data is constructed from the plurality of images, the three-dimensional image data is used as an explanatory variable, and height positions of the defect candidates are detected by a three-dimensional convolutional neural network. It is preferable to have defect candidate height position calculation means for regressing using as an objective variable.

Further, in the defect inspection method of the present invention, the method of radiating X-rays through a plurality of paths different from each other preferably radiates X-rays from two or more positions.

Further, in the defect inspection method of the present invention, the method of emitting X-rays through a plurality of paths different from each other is such that the X-rays are emitted from two or more different positions on the object to be inspected. It is preferable to move the position.

Further, in the defect inspection method of the present invention, the method of radiating X-rays through a plurality of mutually different paths radiates X-rays from one or more positions, and moves the object to be inspected to two or more positions. It is preferable to let

Further, in the defect inspection method of the present invention, the object to be inspected is preferably a member for a high-pressure tank.

According to the present invention, there is a possibility that a plurality of defect candidates exist in the depth direction. It is possible to provide a defect inspection device and a defect inspection method that are capable of accurate detection and suitable for in-line inspection.

FIG. 1 is a schematic diagram for explaining one embodiment of the present invention. FIG. 2 is a schematic diagram for explaining an examination configuration by general X-ray transmission imaging. FIG. 3 is a schematic diagram for explaining an example of a defect and a non-defect portion that cannot be distinguished by an inspection configuration based on general X-ray transmission imaging. FIG. 4 is a schematic diagram for explaining the difference in the detection positions of defect and non-defect portions detected by the X-ray detection means when X-rays are irradiated through a plurality of paths. FIG. 5 is an example of an image acquired by the X-ray detection means. FIG. 6 is a flowchart for explaining the processing flow of the image processing means. FIG. 7 is an example of subdividing an image acquired by the X-ray detection means into cells. FIG. 8 is a schematic diagram in which specific cells in a plurality of X-ray detection images obtained by irradiating X-rays through a plurality of paths are arranged adjacently. FIG. 9 is an example of a synthetic two-dimensional image input to a neural network that outputs height positions of defect candidates. FIG. 10 is a schematic diagram in which specific cells in a plurality of X-ray detection images obtained by irradiating X-rays through a plurality of paths are stacked and arranged. FIG. 11 is a schematic diagram for explaining a configuration provided with means for moving the X-ray radiation means in one embodiment of the present invention. FIG. 12 is a schematic diagram for explaining a configuration provided with means for moving an object to be inspected in one embodiment of the present invention. FIG. 13 is a schematic diagram for explaining another configuration example 1 in one embodiment of the present invention. FIG. 14 is a schematic diagram for explaining another configuration example 2 in one embodiment of the present invention. FIG. 15 is a schematic diagram showing an enlarged cross section of a joint portion of the high-pressure tank member. FIG. 16 is a schematic top view of the joint portion of the high-pressure tank member viewed from the X-ray radiation means 1 side. FIG. 17 is a schematic diagram of an example of an X-ray detection image of a joint portion of the high-pressure tank member. FIG. 18 is a schematic diagram for explaining a method of arranging cells in the example. FIG. 19 is a schematic diagram of adjacently arranging specific cells in a plurality of X-ray detection images obtained by irradiating X-rays through a plurality of paths in the example. FIG. 20 is a diagram plotting visually measured defect candidate height positions and predicted values of the defect candidate height positions according to the present invention in the embodiment. FIG. 21 is a schematic diagram in which specific cells are stacked and arranged in a plurality of X-ray detection images obtained by irradiating X-rays through a plurality of paths in the example. FIG. 22 is a diagram plotting visually measured defect candidate height positions and predicted values of the defect candidate height positions according to the present invention in the embodiment.

<Inspection device for high-pressure tank members>
DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment applied to a defect inspection apparatus of the present invention will be described below with reference to the drawings for a case in which an object to be inspected is a member for a high-pressure tank. In addition, the following embodiment illustrates one Embodiment of this invention, and this invention is not limited to the following description. The following examples can be modified without departing from the scope of the present invention.

First, an outline of the high-pressure tank member and the high-pressure tank to be inspected will be described. A high-pressure tank is a container for filling high-pressure gas such as compressed gas or liquefied gas. For example, when the high-pressure gas is hydrogen, it can be used as a container for fuel cell vehicles, a container for transporting high-pressure hydrogen, and a hydrogen station pressure accumulator. and so on. The structure of the high pressure tank is not particularly limited. For example, a high-pressure tank includes a liner member that is a high-pressure tank member, one or more reinforcing layers that cover the liner member, and a supply system (valve member, various piping systems) for supplying high-pressure gas to the fuel cell. etc.).

The shape of the high pressure tank is not particularly limited. In one example, the high pressure tank is generally cylindrical. The high-pressure tank is formed with an opening for filling high-pressure gas into the tank or taking out high-pressure gas from the tank. The opening is closed by a supply system. In the present invention, the high-pressure tank member is a member constituting a high-pressure tank, and includes a liner member, a member after forming a reinforcing layer on the liner member, and the like.

- Liner member The liner member is a member of the tank container that constitutes the housing of the high-pressure tank. The shape of the liner member is not particularly limited. In one example, the liner member is generally cylindrical and has a storage space formed therein. The accommodation space is filled with a high-pressure gas. The opening is formed in the liner member. Although the liner member may be composed of one member, it is generally composed of a plurality of divided members for ease of manufacture. In this case, the divided members can be integrated by joining or the like. Moreover, the method for producing the liner member includes blow molding, injection molding, and the like. On the other hand, the defect inspection method of the present invention is suitably used for inspecting joint surfaces where a liner member is joined to a plurality of divided members by injection molding.

The material of the liner member is not particularly limited. For example, the liner member is made of resin, metal such as aluminum or iron, or the like. Among them, the resin liner member is likely to be deformed or destroyed after being formed into a high-pressure tank when voids or impurities are formed in the joint portion. However, the defect inspection method of the present invention can appropriately detect voids and impurities as described later. Therefore, the defect inspection method of the present invention is particularly suitable when the liner member is made of resin. Resin has a higher X-ray absorption rate, and impurities in the liner member can be detected more accurately by an X-ray detector, which will be described later. It is preferable to include at least one of

Moreover, it is more preferable that the liner member contains a polyamide resin. Since the polyamide resin has a high X-ray absorption rate, voids and resin impurities in the polyamide resin are easily detected. In particular, when the high-pressure gas is hydrogen gas, hydrogen gas easily dissolves into the liner member because of its low molecular weight. As a result, even if there is a slight void or impurities in the joint portion of the liner member, the high-pressure tank for hydrogen gas is likely to deform or break at the joint portion. According to the defect inspection method of the present invention, which will be described later, such voids and resin impurities can be easily detected. Therefore, when the liner member is made of polyamide resin, the defect inspection method of the present invention can detect impurities and the like with high accuracy and can appropriately discriminate them.

- Reinforcing layer The outer surface of the liner member is preferably covered with one or more reinforcing layers in order to reinforce the liner member. The material of the reinforcing layer is not particularly limited. For example, the reinforcing layer is a fiber reinforced resin layer. Carbon fiber reinforced plastics (CFRP), glass fiber reinforced plastics, and the like are exemplified as fiber reinforced resins forming the fiber reinforced resin layer. These fiber reinforced resins may be used in combination. Moreover, the reinforcing layers made of the respective fiber-reinforced resins may cover the liner member doubly. When the fiber reinforced resin is, for example, carbon fiber reinforced plastic, the fiber reinforced resin layer consists of reinforcing fibers such as carbon fiber reinforced plastic wound around the outer surface of the liner member, and a thermosetting resin that binds the reinforcing fibers together. consists mainly of

Return to the description of the inspection method. The defect inspection method of the present invention is preferably carried out on the joint surface of the liner member before the reinforcing layer is provided in the high-pressure tank. A specific inspection method is to irradiate the liner member with X-rays from an X-ray radiation device and detect the X-rays transmitted through the liner member using an X-ray detector to determine whether the liner member is a non-defective product. Inspect for defective products.

FIG. 1 shows a schematic diagram for explaining the defect inspection apparatus of the present invention. The X-ray radiation means 1 is a device for radiating X-rays to the high-pressure tank member 2 . The shape and dimensions of the X-ray radiation device are not particularly limited. Also, the X-ray radiation means 1 may be accompanied by a power cable (not shown) for driving the X-ray radiation device. In this case, the power cable or the like preferably has a shape and dimensions that do not interfere with the high-pressure tank member 2 . Also, the emitted X-rays must irradiate the high-pressure tank member with X-rays through a plurality of paths. In the present invention, X-rays are emitted by four X-ray emitting means 1a, 1b, 1c and 1d. The emitted X-rays are detected by the X-ray detection means 3 after passing through the high-pressure tank member on the X-ray radiation means side and the high-pressure tank member on the X-ray detection means side, which will be described later. Although the arrangement of the X-ray radiation means 1 is not particularly specified, at least one of the plurality of X-ray radiation means has a transmission path on both the joint surface on the X-ray radiation means side and the joint surface on the X-ray detection means side. It is preferable to arrange so as not to Here, the X-ray radiation means 1a, 1b, 1c and 1d are arranged in parallel so as to sandwich the joint surface of the high-pressure tank member, and in each case the joint surface on the X-ray radiation means side and the joint surface on the X-ray detection means side are arranged. are arranged so as not to be an irradiation path.

The high-pressure tank member 2 is exemplified as a high-pressure tank member obtained by joining two divided resin moldings in a cylindrical shape.

The X-ray detection means 3 is a device for detecting X-rays that have passed through the high-pressure tank member 2 . The X-ray detection means may be composed of at least one or more X-ray detectors. When one X-ray detection means detects X-rays emitted from a plurality of X-ray emission means 1, a plurality of X-ray detectors may be used. X-ray radiation means may emit X-rays at different timings and detect the X-rays, or a plurality of X-ray detection means may be arranged in accordance with the number of the plurality of X-ray radiation means 1 to simultaneously X-rays may be detected, or the X-ray detection means may be moved to a position where X-rays emitted from a plurality of X-ray radiation means 1 can be detected with one X-ray detection means. In general, voids are detected more strongly than the surroundings because X-rays easily pass through them, and impurities are detected to be stronger or weaker depending on the specific gravity of the impurities and the specific gravity of the resin material that constitutes the high-pressure tank member. detected by In addition, since the burr portion is thicker than a normal high-pressure tank member, it is imaged weakly as a whole. Since the reference marker 8 is a reference for accurately determining the surface height position of the high-pressure tank member 2, it is preferable that the reference marker 8 can be clearly detected by the X-ray detection means, so it is often made of resin. It is preferable to use a metal material that is more difficult for X-rays to pass through for the high-pressure tank member. Further, since the reference marker 8 is used to obtain the surface height position of the high-pressure tank member 2 with high accuracy, a plurality of the reference markers 8 are arranged so that the X-ray detection means 3 can detect a plurality of them. Also good. Further, as a means for accurately determining the surface height of the high-pressure tank member by the reference marker 8, by preparing a mechanism (not shown) for mechanically positioning the fixing position of the high-pressure tank member with high accuracy, A predetermined height position reference value may be set without arranging anything similar to the reference marker 8 . Further, the height position reference value may be sequentially measured using a measuring means (not shown) for measuring the height position of the fixing position of the high-pressure tank member with high accuracy. In the description of the embodiment of the present invention, it is assumed that one reference marker 8 is arranged at a position within the detection range of the X-ray detection means 3 .

Incidentally, the X-ray detection means 3 may be a general-purpose X-ray detector. For example, the X-ray detector 3 may be a direct conversion type X-ray detector or an indirect conversion type X-ray detector. More specifically, the X-ray detection means 3 is an X-ray film, image intensifier, computed radiography (CR), flat panel detector (FPD) or the like.

The arrangement of the X-ray detection elements of the X-ray detection means 3 may be an area sensor type X-ray detector in which the detection elements are arranged two-dimensionally, or a line sensor type X-ray detector in which the detection elements are arranged one-dimensionally. The method of sequentially changing the inspection range may be optimized depending on which detection method is used. When using the area sensor method, a mechanism for sequentially switching the field of view according to the inspection field of the area sensor should be prepared, and when using the line sensor method, a mechanism for continuously moving the inspection field should be prepared. good.

The X-ray detection means 3 is preferably an indirect conversion type FPD in that the development process and the like are not required compared to the case where an X-ray film is used, and the time required for inspection can be shortened.

The indirect conversion type FPD is free from restrictions such as usable temperature compared to the direct conversion type detector. Therefore, the indirect conversion type X-ray detector is excellent in handleability. Further, the indirect conversion FPD preferably comprises a cellular scintillator. In indirect conversion FPDs, a scintillator panel is used to convert radiation into visible light. The scintillator panel contains an X-ray phosphor such as cesium iodide (CsI), and the X-ray phosphor emits visible light in response to emitted X-rays. X-ray information is converted into digital image information by converting it into an electrical signal with a (charge-coupled device). However, in the indirect conversion type FPD, when the X-ray phosphor emits light, visible light is scattered by the phosphor itself, and the sharpness of the image tends to be low. On the other hand, an FPD that employs a cellular scintillator has phosphors filled in cells partitioned by partition walls, and can suppress the influence of light scattering. As a result, the FPD equipped with the cellular scintillator has high sharpness and can detect impurities and voids in the high-pressure tank member 2 with high sensitivity.

The X-ray detection means 3 used in the defect inspection method of the present invention uses a photosensitive paste containing glass powder and is mainly made of glass because a large-area and high-sharp cellular scintillator can be easily formed. It is more preferable to use a cell-type scintillator produced by photolithographically processing a partition wall as a component. The pixel size of the sensor of the X-ray detection means 3 is not particularly limited. By way of example, the pixel size of the sensor is preferably 20-300 μm. If the pixel size is less than 20 μm, there is a tendency to detect minute impurities that do not contribute to deformation or destruction of the high-pressure tank member 2 and to erroneously determine good products as defective products. In addition, with such a pixel size, the image data tends to be enormous, and the time required for signal readout and image processing tends to be long. On the other hand, if the pixel size exceeds 300 μm, impurities and the like may not be sufficiently detected.

The image processing means 4 is connected to the X-ray detection means 3, and detects defect candidates using a trained neural network (a) for one or more images acquired by the X-ray detection means 3. defect candidate detection means; height position calculation means for calculating height positions of defect candidates using a neural network (b) that has been trained from a plurality of images; and screening for judging quality of defect candidates from the height positions means; Neural networks (a) and (b) differ in output objective variables, so in general, they are networks with different "weights", which are the strength of connections between neurons in the neural network, convolution filters, etc. . The hyperparameters such as the learning coefficient, the number of units in the intermediate layer, and the number of stages in the layer may also be optimized and designed separately. The flow of processing in the image processing means 4 will be described with reference to FIG. FIG. 6 is a flowchart for explaining the processing flow of the image processing means. The X-ray detection images 10a, 10b, 10c, and 10d are detection images obtained by detecting the X-rays emitted from the X-ray irradiation means 1a, 1b, 1c, and 1d by the X-ray detection means 3, respectively. In the detected image, the intensity of X-ray detection is output as a luminance value, and the luminance value is large (brighter) where X-rays are strongly detected, and the luminance value is smaller (darker) where X-rays are weakly detected.

S101 denotes defect candidate detection means for detecting a defect candidate area from an X-ray detection image. to detect defect candidates. Detection may be performed on at least one X-ray detection image among a plurality of X-ray detection images. It is performed on the image 10a. An example of the defect candidate detection method will be described in detail. As shown in FIG. 7, a wide-area X-ray detection image 10a is divided into a plurality of cells, and even if a defect candidate exists in a cell, the cell is subdivided so that the number of defect candidates becomes one. In FIG. 7, it is schematically shown divided into four cells 11a, 12a, 13a, and 14a. A class classification algorithm based on deep learning is used to determine the presence or absence of defect candidates for the cells 11a, 12a, 13a, and 14a of the X-ray detection image 10a. As a specific example, two-dimensional image data of each cell is input as an explanatory variable into a deep learning model configured with a two-dimensional convolutional neural network, and the classification class is output as an objective variable to determine the presence or absence of defect candidates. discriminate. Specifically, each cell is classified into two classes of "no defect candidate" and "with defect candidate". The classifying may further comprise an algorithm for classifying types of defect candidates. That is, each cell is classified into a class of "no defect candidate" and n+1 classes of "with defect candidate 1", "with defect candidate 2", "with defect candidate 3", ..., "with defect candidate n". do. Also, when the reference marker 8 exists, a class "reference marker present" indicating that the reference marker exists is added. Pre-learning is indispensable for implementing the class classification algorithm by this neural network. In the pre-learning, a plurality of teacher data are prepared in which two-dimensional images of cells and their class classifications are linked, and learning processing is performed so that the teacher data can be predicted correctly. That is, by optimizing the "weight", which is the strength of the connection between neurons in the neural network, and the convolution filter, etc., a model is constructed that enables class classification from the two-dimensional image of the cell. Note that class classification of cell images in training data is performed manually. The number of data required as training data depends on the number of classes and the complexity of cell images, so it is difficult to generalize. , the classification accuracy is higher when there are 100 data or more for each class.

S102 is height position calculation means, which measures the height of the same point detected in a plurality of X-ray detection images. In order to deepen the understanding of the principle of calculating the height position, a detailed description will be given with reference to FIGS. 4 and 5. FIG. FIG. 4 is a schematic diagram for explaining the difference in the detection positions of defect and non-defect portions detected by the X-ray detection means when X-rays are irradiated through a plurality of paths. In order to simplify the explanation, only the cross section of the joint on the X-ray radiation means 1 side of the high-pressure tank member 2 is shown. If there is a void defect 6 at the joint of the high-pressure tank member 2, a void 7 inside the burr at the burr portion, and a reference marker 8 on the surface of the high-pressure tank member 2, X-ray On the detector 3, the X-rays emitted from the X-ray radiation means 1a detect the gap defect 6 in the joint at the Xa1 coordinate position on the X-ray detector 3, and the gap 7 in the burr at the Xa2 coordinate position. Then, the reference marker 8 is imaged at the coordinate position Xa3 and detected as an image such as the X-ray detection image 10a shown in FIG. Depending on the X-rays emitted from the X-ray radiation means 1b, the gap defect 6 of the joint is at the coordinate position Xb1 on the X-ray detector 3, the gap 7 in the burr is at the coordinate position Xb2, and the reference marker 8 is at the coordinate position Xb2. It is imaged at the coordinate position of Xb3 and detected as an image such as the X-ray detection image 10b shown in FIG. At this time, the height position Hd0 of the void defect 6 of the joint is defined by f, the distance from the X-ray radiation means 1a and 1b to the X-ray detector 3, the X-ray radiation means 1a and the X-ray radiation means 1b is calculated as Equation 1, where w is the interval between .
(Formula 1) Hd0=f×w/(|Xa1−Xb1|)
Also, the height position Hf0 of the gap 7 in the burr is calculated by Equation (2).
(Formula 2) Hf0=f×w/(|Xa2−Xb2|)
Since each of the height positions Hd and Hf is a value indicating the distance from the X-ray radiation means, it is necessary to have height position information indicating how much height difference there is with respect to the height position of the reference marker 8. Therefore, the height position Hm0 of the reference marker 8 is calculated using Equation (3).
(Formula 3) Hm0=f×w/(|Xa3−Xb3|)
As a result, the height position Hd1 from the surface of the high-pressure tank member of the void defect 6 at the joint, which is a defect, can be obtained by Equation (4).
(Formula 4) Hd1 = Hd0 - Hm0
Similarly, the height position Hf1 from the surface of the high-pressure tank member of the gap 7 in the burr, which is not a defect, can be obtained by Equation (5).
(Formula 5) Hf1 = Hf0 - Hm0
In order to sequentially execute this height position calculation processing, the processing is executed using a trained neural network (b) in the embodiment of the present invention.

An example of the method for calculating the height position using the neural network (b) will be described in detail. Using the aforementioned defect candidate detection means, the height position calculation is executed in the cell of the detected X-ray image classified as "with defect candidate". A height position calculation method will be described using the cells 12a, 12b, 12c and 12d of the X-ray detection images 10a, 10b, 10c and 10d as examples. As shown in FIG. 8, four images of cells 12a, 12b, 12c, and 12d are arranged to synthesize one image (two-dimensional image 15 shown in FIG. 9). A single synthesized two-dimensional image data 15 is input as an explanatory variable into a deep learning model composed of a two-dimensional convolutional neural network, and the height position of the defect candidate is calculated by performing regression using the height position of the defect candidate as the objective variable. conduct. Regarding the height position of the reference marker 8, cells 13a, 13b, 13b, 13b, 13b, 13b, 13b, 13b, 13b, 13b, 13b, 13b, 13b, and 13d of the X-ray detection images 10a, 10b, 10c, and 10d classified as "with reference marker" are used in the same manner as in the defect candidate height position calculation method. The four images 13c and 13d are arranged to synthesize one image. One synthesized two-dimensional image data is input as an explanatory variable into a deep learning model configured with a two-dimensional convolutional neural network, and regression is performed using the height position of the reference marker as the objective variable.

Also, a three-dimensional convolutional neural network may be used for the height position calculation process. In this case, as an example, four two-dimensional images of cells 12a, 12b, 12c, and 12d are stacked as shown in FIG. Configure. For example, if one cell is two-dimensional image data of 64×64 pixels, by stacking four two-dimensional images, three-dimensional image data of 64×64×4 voxels can be obtained. The three-dimensional image data is input to a deep learning model composed of a three-dimensional convolutional neural network as an explanatory variable, and regression is performed using the height position of the defect candidate or the height position of the reference marker 8 as the objective variable.

Pre-learning is indispensable to implement the height position calculation algorithm by this neural network. In the pre-learning, one two-dimensional image obtained by arranging and synthesizing the four cell images having defect candidates (or one three-dimensional image obtained by stacking and synthesizing the four cell images) and its height position are linked. Prepare multiple sets of teacher data and perform learning processing so that the teacher data can be predicted correctly. In other words, it is possible to calculate the height position from the synthesized 2D image (or 3D image) by optimizing the "weight", which is the strength of the connection between neurons in the neural network, and the convolution filter. Building possible models. Note that the height position in the teacher data is calculated manually. The number of data required as teacher data depends on the complexity of the cell image, so it is difficult to generalize, but the larger the number of data, the better. Higher measurement accuracy.

With respect to the height information calculated by this method, the height of the part corresponding to the reference marker 8 is obtained as Hm0, and when Hm0 is subtracted from the height of each defect candidate, a position higher than the reference marker 8, that is, Height information having a positive value is output for a height portion of the high-pressure tank member 2 closer to the X-ray detection means 3 than the surface on which the reference marker 8 exists, and a position lower than the reference marker 8, that is, the high-pressure tank Height information having a negative value is output for a height portion on the X-ray radiation means 1 side of the surface of the member 2 on which the reference marker 8 exists.

S103 is a sorting means.

The selecting means 5 selects whether or not the defect candidate is a defect from the information of the height position calculated value of the defect candidate. Specifically, a height upper limit value and a height lower limit value are set as predetermined height thresholds, and a defect candidate having a height position calculated value within the range of these thresholds is determined as a defect occurring in the joint. , a defect candidate having a height position calculated value within the range of this threshold value is determined to be a void or impurity existing in burrs existing above and below the joint, or an erroneous detection. As the threshold value, a lower limit value and an upper limit value based on the thickness design value of the joint portion of the high-pressure tank member are preferably used. In other words, the lower limit of the threshold value is a depth value that is permissible in terms of design even if voids or impurities occur on the joint surface. Subtracted values are used. In the embodiment of the present invention, since the erroneously detected area 9 and the gap 7 in the burr are below the predetermined upper limit height, they are not determined as defects, and only the gap defect 6 at the joint is the upper limit height and the lower limit height. Since it is within the value range, it is sorted out as a defect.

An embodiment of the present invention has been described above with reference to the drawings, taking as an example the configuration provided with a plurality of X-ray radiation means 1 in order to obtain a plurality of X-ray radiation paths. The present invention can adopt the following modified embodiments, for example.
(1) A configuration in which the X-ray emitting means 1 is moved each time an X-ray is detected by the X-ray detecting means 3 in order to obtain a plurality of X-ray radiation paths. A schematic diagram is shown in FIG. FIG. 11 is a schematic diagram for explaining a configuration provided with means for moving the X-ray radiation means in one embodiment of the present invention. It is preferable that the means for moving the X-ray radiation means has a moving direction parallel to the X-ray detection means 3 and perpendicular to the joint surface.
(2) In order to obtain a plurality of X-ray radiation paths, the object to be inspected (the high-pressure tank member 2 in the schematic diagram) is moved each time detection is performed by the X-ray detection means 3 . A schematic diagram is shown in FIG. FIG. 12 is a schematic diagram for explaining a configuration provided with means for moving an object to be inspected in one embodiment of the present invention. It is preferable that the means for moving the object to be inspected has a moving direction parallel to the X-ray detecting means 3 and perpendicular to the arrival plane.

All embodiments of the invention can also be combined with further variants, for example any of the following.
(1) A configuration in which the X-ray radiation means is inserted inside the object to be inspected. A schematic diagram is shown in FIG. This arrangement can be realized if the X-ray emitting means are small relative to the opening of the specimen to be inspected. When the object to be inspected is a high-pressure tank member, in this configuration, the emitted X-rays pass through only one layer of the high-pressure tank member. Noise is less than in the case of the conventional method, and high-precision inspection can be realized.
(2) A configuration in which the X-ray detection means is inserted inside the object to be inspected. A schematic diagram is shown in FIG. This configuration can be realized if the X-ray detection means is small with respect to the opening of the object to be inspected. When the object to be inspected is a high-pressure tank member, the X-ray detection means 3 is installed outside the high-pressure tank member 2 because the emitted X-rays pass through only one layer of the high-pressure tank member 2 in this configuration as well. Noise is less than in the case of using the laser, and high-precision inspection can be realized.

A more specific embodiment will be described with respect to the modified example whose schematic diagram is shown in FIG. 12 . In FIG. 12, the distance from the X-ray radiation means 1 to the X-ray detector 3 is 323 mm, and the reference marker (FIG. 15 The inspection device and the high-pressure tank member to be inspected were arranged so that the distance to 17a or 18a) of was 51 mm. FIG. 15 is a schematic diagram showing an enlarged area indicated by reference numeral 16 in FIG. Areas 20 and 22 in the joint are burrs, and there is no quality problem even if there is a gap. The area 21 is a joint where no air gap should exist. A gap in the burr is not determined as a defect, and only a gap defect in the joint is selected as a defect. In this embodiment, the thickness of the region 21 is designed to be 2 mm, so if the detected defect candidate exists at a height of 51 mm or more and 53 mm or less from the X-ray radiation means 1, it is selected as a defect.

The fiducial marker described above is made of stainless steel and is formed near the junction in advance before X-ray imaging. The film thickness of the fiducial marker was thin enough to be ignored on the order of millimeters. The fiducial markers were positioned such that the junction was located midway between the fiducial markers 17a and 18a. Also, as shown in FIG. 16, a plurality of reference markers were arranged at intervals of 5 mm along the longitudinal direction of the joint. 16 is a top view of the joint viewed from the X-ray radiation means 1 side. The X-ray detection means 3 used an FPD equipped with a cellular scintillator. The pixel pitch of the FPD is 127 μm in the vertical and horizontal directions, and the number of pixels is 1920×1536.

The high-pressure tank member is moved from the position shown in 2a of FIG. A detection image was obtained. FIG. 17 shows a schematic diagram of an X-ray detection image obtained at the position of 2a. Next, as shown in FIG. 18, 80× is applied so as to pass through the respective midpoints of the reference markers 17a1 and 18a1, 17a2 and 18a2, 17a3 and 18a3, 17a4 and 18a4, 17a5 and 18a5, 17a6 and 18a6, 17a7 and 18a7. Twenty-four 80-pixel cells were arranged in a line (reference numeral 19a). The cells are arranged such that the horizontal coordinates of the center point of the cell in FIG. 18 coincide with the horizontal coordinates of the midpoint between the reference markers. In this manner, 24 cell images were obtained by X-ray imaging with the high-pressure tank member arranged on 2a. Twenty-four cell images were obtained from the X-ray detection images at the other three moving positions as well, by cells arranged in the middle of the reference markers in the same manner. In the four X-ray detection images, each cell having the same relative position from the reference marker was image synthesized in the arrangement shown in 24a to 24d in FIG. 19 to create a total of 24 synthesized images of 160×160 pixels. Another joint portion of the high-pressure tank member was also subjected to X-ray imaging in the same procedure as described above, and a composite image similar to that of FIG. 19 was created. A total of 2880 composite images were obtained from X-ray imaging of multiple joints. In this embodiment, the X-ray detection image is offset at the reference marker position, and the area where the defect candidate may exist is limited to the width of 80 pixels between the reference markers. It is not necessary to perform offset or area limitation as described in the form of .

A visual inspection of 2880 composite images revealed that 172 composite images had defect candidates. These 172 pieces are classified into a class of "with defect candidate". 172 images were selected at random from the remaining synthesized images and classified into the class "no defect candidate". FIG. 19 shows an example of a composite image classified as "with defect candidate". 172 shift amounts α (unit: pixel) in the image from the defect candidate position at the initial position shown in 24a in FIG. , was measured visually. If the defect candidate exists closer to the X-ray detection means 3 than the reference marker, the defect candidate moves within the cell as indicated by 24a to 24d in FIG. The shift direction in FIG. 19 of the defect candidate is defined as positive. If the defect candidate exists closer to the X-ray emitting means 1 than the reference marker, the shift amount of the defect candidate is observed as a negative value.

A method of converting the shift amount of α pixels into the height position H will be described. When the high-pressure tank member is moved by 9 mm from the position indicated by 2a in FIG. is calculated as Equation 6 below. The distance f from the X-ray radiation means 1 to the X-ray detector 3 is 323 mm, the distance w between the X-ray radiation means 1a and the X-ray radiation means 1b in FIG. When 9 mm and the height position Hm0 of the reference marker 8 of 51 mm are input, the moving distance of the reference marker in the X-ray detector is 57 mm.
(Formula 6) |Xa3-Xb3|=f×w/Hm0=323×9/51=57 mm
In this embodiment, the cell image is cut out after being offset at the reference marker position. is calculated as Equation 7.
(Formula 7) ΔX = |Xa3-Xb3|-0.127α (mm) = 57-0.127α (mm)
Therefore, the height position H (mm) of the defect candidate is calculated by Equation (8).
(Formula 8) H = f × w / ΔX = (323 × 9) / (57-0.127α)
If H is 51 mm or more and 53 mm or less, it is sorted out as a defect.

From a total of 344 synthesized images, 172 synthesized images classified as “no defect candidate” and 172 synthesized images classified as “with defect candidate”, 275 images corresponding to a ratio of 80% were obtained. A two-dimensional convolutional neural network (a) was trained by randomly selecting an image as teacher data for classifying "no defect candidate" or "with defect candidate". Neural network (a) uses TensorFlow, which is developed by Google and is open source, and uses Keras, an open source neural network library that runs on TensorFlow. -team/keras/blob/master/examples/cifar10_cnn.py) and implemented with code developed based on the sample code described. The 69 images corresponding to the remaining ratio of 20% are input to the trained two-dimensional convolutional neural network (a) as verification data, and classified into "no defect candidate" or "defect candidate present". , the correct answer rate for classification was 93%. If the number of training data is increased, it can be expected that the prediction accuracy will be further improved. In this embodiment, the input data is a synthesized image obtained by synthesizing four cell images, but it is also possible to construct a neural network that performs class classification using only the cell image at the position 2a, for example.

Next, 172 composite images classified as “with defect candidate” and height position data of the defect candidates are prepared, 137 images corresponding to 80% are randomly selected, and each A two-dimensional convolutional neural network (b) was trained to regress the height positions of defect candidates. Like neural network (a), neural network (b) was implemented with code developed based on the sample code described on the web page. Although the sample code is for a classification problem, it is a regression problem in which the height position of defect candidates is regressed. Therefore, in the output layer, the activation function was changed from a softmax function to a linear function. After learning, 35 images corresponding to the remaining 20% were used as verification data to predict the height positions of defect candidates. FIG. 20 plots the height positions of the defect candidates visually measured in advance and the predicted values of the height positions of the defect candidates by the neural network, respectively, plotted on the horizontal axis and the vertical axis. A coefficient of determination R ² representing the strength of the correlation between the actual measured value and the predicted value was calculated using Equation 9 and found to be R ² =0.59. Note that the number of samples n of the verification data in Expression 9 is 35.
(Formula 9)

Of the 35 defect candidates in the verification data, 10 data were selected as defects and 25 data were selected as non-defects based on height position measurement by visual inspection. When the prediction value of the neural network was used to sort out whether or not there was a defect, the classification correct answer rate for data that was a defect was 80%, and the classification correct answer rate for data that was not a defect was also 80%. In this embodiment, there are 137 pieces of teacher data for calculating the height position, but if the number of teacher data is increased, it can be expected that the prediction accuracy will be further improved.

Next, each of the 172 two-dimensional synthesized images classified as “with defect candidate” is divided again into four images before synthesis, and the four two-dimensional images are stacked to form four grids in the stacking direction. Construct a piece of three-dimensional image data having points. A more specific description will be given below with reference to the drawings. In FIG. 19, in the four X-ray detection images, the cell images 24a to 24d (80×80 pixels each) having the same relative position from the reference marker are arranged adjacent to each other in a plane, and one 160×160 pixel two A dimensional composite image was created, and the cell images 24a to 24d were stacked as shown in FIG. 21 to form one three-dimensional image data of 80×80×4 voxels having four grid points in the stacking direction. do. As described above, 172 three-dimensional synthesized image data classified as "with defect candidate" and height position data of the defect candidate are prepared, and 137 images corresponding to 80% are randomly selected. and trained a three-dimensional convolutional neural network (b) to regress the height position of each defect candidate. Like the two-dimensional convolutional neural network (b), the three-dimensional convolutional neural network (b) was implemented with code developed based on the sample code described on the web page. Since the sample code is for a two-dimensional convolutional neural network, the convolution filters, pooling layers, etc. have been changed to three dimensions. After learning, 35 images corresponding to the remaining 20% were used as verification data to predict the height positions of defect candidates. FIG. 22 plots the height positions of defect candidates visually measured in advance and the predicted values of the height positions of defect candidates by the neural network, respectively, plotted on the horizontal axis and the vertical axis. A coefficient of determination R ² representing the strength of the correlation between the actual measured value and the predicted value was calculated using Equation 9 and found to be R ² =0.50. Note that the number of samples n of the verification data in Expression 9 is 35.

Of the 35 defect candidates in the verification data, 10 data were selected as defects and 25 data were selected as non-defects based on height position measurement by visual inspection. When the predicted value of the neural network was used to determine whether or not the data had defects, the classification correct answer rate for data with defects was 80%, and the classification correct answer rate for data without defects was 72%. In this embodiment, there are 137 pieces of teacher data for calculating the height position, but if the number of teacher data is increased, it can be expected that the prediction accuracy will be further improved.

1 X-ray radiation means 1a X-ray radiation means 1b X-ray radiation means 1c X-ray radiation means 1d X-ray radiation means 2 High-pressure tank member 3 X-ray detection means 4 Image processing means 5 Sorting means 6 Void defect 7 at joint portion Burr Internal gap 8 Reference marker 9 False detection area 10a X-ray detection image a by X-ray radiation means a
10b X-ray detection image b by X-ray radiation means b
10c X-ray detection image c by X-ray radiation means c
10d X-ray detection image d by X-ray radiation means d
11a Cell 11 of X-ray detection image a
11b Cell 11 of X-ray detection image b
11c Cell 11 of X-ray detection image c
11d Cell 11 of X-ray detection image d
12a Cell 12 of X-ray detection image a
12b Cell 12 of X-ray detection image b
12c Cell 12 of X-ray detection image c
12d Cell 12 of X-ray detection image d
13a Cell 13 of X-ray detection image a
13b Cell 13 of X-ray detection image b
13c Cell 13 of X-ray detection image c
13d Cell 13 of X-ray detection image d
14a Cell 14 of X-ray detection image a
14b Cell 14 of X-ray detection image b
14c Cell 14 of X-ray detection image c
14d Cell 14 of X-ray detection image d
15 Composite two-dimensional image 16 Inspection location 17a Reference marker 17b Reference marker 18a Reference marker 18b Reference marker 17a1 Reference marker 17a2 Reference marker 17a3 Reference marker 17a4 Reference marker 17a5 Reference marker 17a6 Reference marker 17a7 Reference marker 18a1 Reference marker 18a2 Reference marker 18a3 Reference marker 18a4 Reference marker 18a5 Reference marker 18a6 Reference marker 18a7 Reference marker 19a Cell 20 Burr region 21 Joint region 22 Burr region 23 Joint 24a A cell image 24c extracted from the X-ray detection image at a position shifted by 3 mm A cell image 24d extracted from the X-ray detection image at a position shifted by 6 mm X at a position shifted by 9 mm Cell image extracted from line detection image

Claims (14)

X-ray emitting means for emitting X-rays through a plurality of paths different from each other, one or more X-ray detecting means for detecting X-rays transmitted through an object to be inspected, and image processing means, said image processing means is a defect candidate detection means for detecting defect candidates using a trained neural network (a) for one or more images acquired by the X-ray detection means; and a plurality of defects acquired by the X-ray detection means. A height position calculation means for calculating the height position of a defect candidate using a neural network (b) that has been learned from an image, and a selection means for determining the quality of the defect candidate from the height position .
A defect inspection apparatus, wherein the plurality of images acquired by the X-ray detection means are obtained by detecting X-rays radiated through a plurality of paths different from each other. Defect candidate height position calculation means for performing regression using one two-dimensional image data obtained by synthesizing the plurality of images as an explanatory variable and using the height position of the defect candidate as an objective variable by a two-dimensional convolutional neural network. , The defect inspection device according to claim 1. constructing one three-dimensional image data from the plurality of images, using the three-dimensional image data as an explanatory variable, and performing regression using a three-dimensional convolutional neural network using the height position of the defect candidate as an objective variable; 2. The defect inspection apparatus according to claim 1, further comprising height position calculation means. The defect inspection device according to any one of claims 1 to 3, wherein said X-ray radiation means comprises two or more X-ray radiation means. Said X-ray emitting means moves said at least one or more X-ray emitting means so as to emit X-rays from two or more different positions with respect to said object to be inspected. 4. The defect inspection apparatus according to claim 1, further comprising X-ray radiation position moving means. 4. Any one of claims 1 to 3, wherein said X-ray radiation means comprises one or more X-ray radiation means and inspected object position moving means for moving the inspected object to two or more positions. The defect inspection device according to the item. The defect inspection apparatus according to any one of claims 1 to 6, wherein the object to be inspected is a member for a high-pressure tank. A neural network that emits X-rays through a plurality of paths different from each other, detects X-rays that have passed through an object to be inspected at one or more positions, and has been trained for one or more detected X-ray images (a) to detect a defect candidate, and calculate the height position of the defect candidate using a trained neural network (b) for the detected plurality of X-ray images,
wherein the plurality of detected X-ray images are obtained by detecting X-rays radiated through a plurality of paths different from each other;
A defect inspection method for sorting out whether the defect candidate is good or bad from the height position. Defect candidate height position calculation means for performing regression using one two-dimensional image data obtained by synthesizing the plurality of X-ray images as an explanatory variable and using the height position of the defect candidate as an objective variable by a two-dimensional convolutional neural network. The defect inspection method according to claim 8, comprising: constructing one three-dimensional image data from the plurality of X-ray images, using the three-dimensional image data as an explanatory variable, and performing regression using a three-dimensional convolutional neural network using the height position of the defect candidate as an objective variable; 9. The defect inspection method according to claim 8, further comprising candidate height position calculation means. 11. The defect inspection method according to any one of claims 8 to 10, wherein the method of radiating X-rays through a plurality of paths different from each other radiates X-rays from two or more positions. 8 to 10, wherein the method of emitting X-rays through a plurality of paths different from each other moves the X-ray emission position so that the X-rays are emitted from two or more different positions with respect to the object to be inspected. The defect inspection method according to any one of . 11. Any one of claims 8 to 10, wherein the method of emitting X-rays through a plurality of paths different from each other emits X-rays from one or more positions and moves the object to be inspected to two or more positions. The defect inspection method described in the item. The defect inspection method according to any one of claims 8 to 13, wherein the object to be inspected is a high-pressure tank member.

Images