TWM604396U - Weld checking system based on radiography - Google Patents

Abstract

This model specification provides a radiographic-based weld bead inspection system, which includes a radiographic device, an image content discrimination device, and a neural network classification group. The radiographic device receives the penetrating rays that penetrate the pipeline to be inspected to generate a to-be-inspected image representing the intensity distribution of the penetrating rays; the image content discrimination device is electrically coupled to the radiographic device to obtain and analyze the Check the image to obtain the weld bead area image showing the weld bead in the image to be inspected; the neural network classification group is electrically coupled to the image content discrimination device to obtain the weld bead area image as input, and according to the input weld bead area The image judges the condition of defects in the weld bead.

Description

Radiographic-based weld bead inspection system

This new model relates to a weld bead inspection technology, in particular to a weld bead inspection system based on radiography.

Steel has the advantages of stable material, high strength and high toughness, so it has been widely used in various buildings. In order to stabilize the structure, the steel materials need to be joined by welding. However, in the process of heating and melting the welding material and rapid cooling, welding defects are likely to develop in the welding material and reduce the structural safety performance.

In the prior art, many methods have been provided to inspect the weld bead by non-destructive inspection to ensure the safety of the steel structure. The so-called non-destructive inspection refers to the inspection method that checks whether there are defects on the surface or inside without destroying the original shape, quality and function of the material. Common non-destructive inspection methods include: visual inspection, magnetic particle inspection, ultrasonic inspection and radiographic inspection. Among them, radiographic inspection is the radiograph (photographing the pipe weld bead) taken by professionals who have obtained the license. It is placed on the high-brightness judgement lamp to judge the defect position by human eyes, but because everyone's standards are different, it may cause the problem of unstable structural safety.

In view of this, the present specification proposes a radiographic-based weld bead inspection system. The weld bead inspection system can provide the consistency of the judgment result, and therefore can reduce the problem of unstable safety.

From one aspect, the present specification proposes a radiographic-based weld bead inspection system, which is suitable for detecting the state of the weld bead on the pipeline to be inspected. The weld bead inspection system includes a radiographic device, an image content discrimination device, and a neural network classification group. The radiographic device receives the penetrating radiation that penetrates the pipeline to be inspected to generate an image to be inspected that represents the intensity distribution state of the penetrating radiation. The image content judging device is electrically coupled to the radiography device to obtain an image to be inspected, and analyzes the image to be inspected to obtain an image of the weld bead area in the image to be inspected showing the weld bead. The neural network classification group is electrically coupled to the image content judging device to obtain an image of the weld bead area as an input, and the condition of the defect existing in the weld bead is determined according to the input of the weld bead area image.

In one embodiment, the device for determining image content includes an image position determining unit and an image pre-processing unit. The image position determining unit receives the image to be inspected and determines that the initial image of the pipeline to be inspected is present in the image to be inspected according to the hue, saturation and brightness of the image to be inspected. The image pre-processing unit is electrically coupled to the image position determining unit to obtain an initial image from the image position determining unit, and then performing a median filtering operation and an erosion operation on the content of the initial image to generate the aforementioned weld bead region image. Or, in another embodiment, the image content discrimination device includes a neural network. The neural network is electrically coupled to the radiographic device to obtain the image to be inspected as input. The neural network obtains the training result of image discrimination by training using multiple pipeline images with weld bead, and based on the image The training result is judged and the weld bead area image is obtained from the image to be inspected.

In one embodiment, the aforementioned neural network classification group includes a first deep learning neural network. The first deep learning neural network is electrically coupled to the image content discrimination device to obtain an image of the weld bead area as input, wherein the first deep learning neural network uses a plurality of first defect training images with weld bead defects The training is performed to obtain the first defect judgment training result, and the condition of the defect existing in the weld bead shown in the image of the weld bead area is judged according to the first defect judgment training result.

In one embodiment, the aforementioned neural network classification group further includes a second deep learning neural network. The second deep learning neural network is electrically coupled to the image content discrimination device to obtain an image of the weld bead area as input, wherein the second deep learning neural network uses a plurality of second defect training images with weld bead defects The training is performed to obtain the second defect judgment training result, and the condition of the defect existing in the weld bead shown in the image of the weld bead area is judged according to the second defect judgment training result. Wherein, the aforementioned first defect judgment training result is not completely the same as the second defect judgment training result.

In one embodiment, the aforementioned device for judging image content further includes an image capturer. The image picker takes a part of the image of the weld bead and outputs it to generate an image of the partial weld.

In one embodiment, the aforementioned third deep learning neural network is electrically coupled to the image content judging device to obtain the aforementioned partial weld bead region image as input, wherein the third deep learning neural network has A plurality of third defect training images of the weld bead defect are trained to obtain a third defect judgment training result, and the condition of the defect existing in the weld bead shown in the partial weld bead region image is judged based on the third defect judgment training result.

Based on the above, the weld bead inspection system proposed in the present specification uses a trained deep learning neural network to interpret radiographic films, which can not only achieve the effect of automation, but also the fixed training results can make the standard of film interpretation consistent. Reduce the problem of unstable safety caused by differences in interpretation standards.

Please refer to FIG. 1, which is a system block diagram of a weld bead inspection system according to an embodiment of the present invention. In this embodiment, the weld bead inspection system 10 includes a radiographic device 100, an image content discrimination device 110, and a neural network classification group 120. As shown in the figure, the radiographic apparatus 100 receives the penetrating rays IN that have penetrated a pipeline to be inspected (not shown), and generates corresponding images according to the intensity distribution of the received penetrating rays IN. (Hereinafter referred to as the image to be checked) P. The image content discrimination device 110 is electrically coupled to the radiographic device 100 to obtain the image P to be inspected generated by the radiographic device 100, and analyzes the image P to be inspected to obtain the image P showing the weld bead in the pipeline to be inspected Image of bead area P1. The neural network classification group 120 is electrically coupled to the image content judging device 110 to obtain the weld bead area image P1 as an input, and determine the condition of defects in the weld bead based on the input weld bead area image P1.

In detail, the radiation provided by the radiation source (not shown) will become the post-penetration ray IN after penetrating the pipeline to be inspected containing the weld bead, and the ray IN will then be irradiated onto the radiographic apparatus 100 after the penetration. The radiographic apparatus 100 can generate the corresponding image P to be inspected according to the content (generally referred to as intensity) of the received penetrating radiation IN according to the existing radiation detection technology. Please refer to FIG. 2, which is a schematic diagram of an image to be inspected generated by a weld bead inspection system according to an embodiment of the present invention. As shown in the figure, the image to be inspected 20 includes peripheral images 200 and 205 representing the peripheral area of the pipeline to be inspected, a pipeline image 210 representing the pipeline to be inspected, and a weld bead image 220 representing the location of the weld bead.

As shown in FIG. 1, after receiving the image P to be inspected generated by the radiographic apparatus 100, the image content discriminating device 110 analyzes the image P to be inspected, thereby obtaining the image P to be inspected which shows the pipeline to be inspected. The image of the weld bead area P1. Please also refer to FIG. 3A. The image content determining device 300A includes an image position determining unit 310 and an image processing unit 320, and the image processing unit 320 is electrically coupled to the image position determining unit 310 to receive data from the image position determining unit 310. Specifically, the image position determining unit 310 receives the image P to be inspected from the radiographic apparatus 100 and determines the pipeline area of the pipeline to be inspected in the image P to be inspected according to the hue, saturation, and brightness of the image P to be inspected. (For example, including the pipeline image 210 and the weld bead image 220 in FIG. 2), and the image data in this area is used as the initial image T; the image processing unit 320 first obtains the initial image T from the image position determining unit 310, and then compares the initial image The content of T is subjected to median filtering operations and erosion (Erosion) operations to eliminate noise or perform other image optimization operations to generate the weld bead area image P1.

In another embodiment, referring to FIG. 3B, the image content discrimination device 300B includes a neural network 340, wherein the neural network 340 is electrically coupled to the radiographic apparatus 100 to obtain the image P to be inspected as input. Before the image content discriminating device 300B is put into operation, the worker can first use multiple pipeline images containing weld bead to train the neural network 340. The trained image content discrimination device 300B has a set of specific logic for determining whether there is a weld bead in the pipeline image and the location of the weld bead (hereinafter referred to as the image discrimination training result), and the image content discrimination device 300B can The weld bead area image P1 is obtained from the image to be inspected according to the image discrimination training result. For example, it can be determined from the image to be inspected 20 in FIG. 2 that the image in the area 250 is the weld bead area image P1. It is worth mentioning that, because the neural network 340 is used to determine the image position, as the process of training the neural network 340 has different accuracy requirements, the accuracy of the final image judgment training result is also Will be different. The accuracy of different image discrimination training results will affect the difference between the area 250 and the weld bead image 220. The less accurate image discrimination training results generally require less training time, but a larger area 250 may be selected. In contrast, the more accurate image discrimination training results generally require more training time and more training materials , But the area 250 that is very close to the size of the weld bead image 220 may be selected. Users can perform different levels of image position judgment training according to their needs.

Next, as shown in FIG. 1, the weld bead area image P1 generated in the above embodiments will be sent to the neural network classification group 120, and the neural network classification group 120 will confirm whether there are defects in the weld bead And the distribution of defects. Please also refer to FIG. 4, which is a circuit block diagram of the neural network classification group of the weld bead inspection system according to an embodiment of the present invention. In this embodiment, the neural network classification group 400 mainly includes a first deep learning neural network 410 and a second deep learning neural network 420. The first deep learning neural network 410 is electrically coupled to the aforementioned image content judging device 110 to obtain the weld bead area image P1 as input. Before the first deep learning neural network 410 operates, the staff can use multiple The first deep learning neural network 410 is trained by the pipeline image with weld bead defects (hereinafter referred to as the first defect training image). The trained first deep learning neural network 410 has a set of specific logic for judging whether there are defects in the weld bead (hereinafter referred to as the first defect judgment training result), the first deep learning neural network 410 can be based on the first A defect judges the training result and judges whether there is a defect in the weld bead area image P1. Even if the first deep learning neural network 410 is trained to determine the type of defects, the first deep learning neural network 410 can also determine the types of defects in the weld bead region image P1.

The second deep learning neural network 420 is also electrically coupled to the aforementioned image content discrimination device 110 to obtain the weld bead area image P1 as input. Similar to the first deep learning neural network 410, before the second deep learning neural network 420 operates, the staff can use multiple pipeline images with weld bead defects (hereinafter referred to as the second defect training image) to compare the first The second deep learning neural network 420 is trained, and the trained second deep learning neural network 420 also has a specific set of logic for judging whether there are defects in the weld bead (hereinafter referred to as the second defect judgment training result), And the condition of the defect in the weld bead area image P1 can be judged subsequently according to the second defect judgment training result. It is worth mentioning that the aforementioned first defect judgment training result and the second defect judgment training result should not be exactly the same, so that different deep learning neural networks can be used to judge the weld bead according to the needs of different situations. defect. To achieve this goal, the pipeline image used when training the first deep learning neural network 410 may be different from the pipeline image used when training the second deep learning neural network 420, or it may be used when training the first deep learning neural network. Different parameter values (such as the safe size of flaws, etc.) are used when the road 410 and the second deep learning neural network 420 are trained.

Further, the neural network classification group 400 shown in FIG. 4 also includes a third deep learning neural network 430. The third deep learning neural network 430 is used together with the image capturer 330 shown in FIG. 3A. The image capturer 330 is electrically coupled to the image pre-processing unit 320 to obtain the weld bead area image P1 (or image The extractor 320 may also be electrically coupled to the neural network 340 to obtain the weld bead area image P1), and output a part of the weld bead area image P1 as a partial weld bead area image P2. Specifically, since a deep learning neural network usually undergoes convolution and sampling to reduce the data matrix, the entire weld bead area image P1 is input and subjected to convolution and sampling by the deep learning neural network. After processing, relatively subtle flaws (such as tiny bubble-like flaws) relative to the weld bead area image P1 may be ignored. Therefore, by inputting a part of the weld bead area image P2 that is smaller relative to the weld bead area image P1 into the deep learning neural network, these originally relatively subtle flaws will become relatively larger, which is more likely Discovered by deep learning neural networks.

According to the above description, the third deep learning neural network 430 can further improve the chance of finding defects in the weld bead. Therefore, the third deep learning neural network 430 can be used in the first deep learning neural network 410 or the second deep learning neural network. 420 is activated when no defects are found in the weld bead. Please also refer to FIG. 5, which is a logical definition diagram of the operation logic of the neural network classification group of the weld bead inspection system according to an embodiment of the present invention. As shown in the figure, in this embodiment, the situation in which the neural network judges that there is a defect and that there is actually a defect in the weld bead is called TP (True Positive), and the neural network judges that there is a defect, but the weld is actually not. The condition with defects is called False Positive (FP, False Positive), and the condition where the neural network judges that there is no defect but there are defects in the weld bead is called False Negative (FN, False Negative). The condition of judging that there is no defect and there is no defect in the weld bead is called FP (False Positive). Based on this, the following formulas (1) and (2) can be defined:

Accuracy = TP/(TP+FP)… (1)

Recall rate = TP/(TP+FN)… (2)

According to the above formula (1), under the premise that the deep learning neural network judges to be flawed, the smaller the number of flaws in the actual weld bead (that is, the smaller the FP), the higher the accuracy. Under the definition of this embodiment, the higher the accuracy rate, the lower the probability of misjudgment as flaws, that is, once the deep learning neural network judges that there are flaws in the weld bead, the greater the possibility of flaws in the weld bead in fact . Therefore, in order to increase the accuracy, the first deep learning neural network 410 or the second deep learning neural network 420 can be trained to reduce the defect judgment error (FP). In this way, it is possible to reduce the probability that the staff must perform re-inspection when the deep learning neural network determines that there is a defect in the weld bead.

Furthermore, according to the above formula (2), under the premise that there are defects in the actual weld bead, the smaller the number of defects judged by the neural network (that is, the smaller the FN), the higher the recall rate. Under the definition of this embodiment, the higher the recall rate, the lower the probability of misjudgment as flawless, that is, once the deep learning neural network judges that there is no flaw in the weld bead, the less likely there is a flaw in the weld bead actually . Therefore, in order to increase the recall rate, the third deep learning neural network 430 can be trained in the direction of reducing the faultless judgment error (FN). In this way, it can reduce the probability that the staff must perform re-inspection when the deep learning neural network determines that there are no defects in the weld bead.

Therefore, when the first deep learning neural network 410 determines that there is a defect in the weld bead, the control signal C1 sent by the first deep learning neural network 410 can be zero. When the third deep learning neural network 430 does not exist, the control signal C1 can be directly used as the output data of the neural network classification group 400 to indicate that there are defects in the weld bead; and when the third deep learning neural network 430 exists At that time, the control signal C1 is output to the third deep learning neural network 430, so that the third deep learning neural network 430 does not perform defect judgment and directly outputs the control signal C1 as data D, where data D is the network classification Output data of group 400.

In contrast, when the first deep learning neural network 410 determines that there are no defects in the weld bead, the control signal C1 sent by the first deep learning neural network 410 can be set to 1. When there is no third deep learning neural network 430, the control signal C1 can be directly used as the output data of the neural network classification group 400 to indicate that there are no defects in the weld bead; and when the third deep learning neural network 430 exists At the time, the control signal C1 is output to the third deep learning neural network 430 so that the third deep learning neural network 430 obtains the partial weld bead area image P2 and performs corresponding judgments to finally output data D (according to the judgment result , The data D may be 1 for flawless or 0 for flawed) as the output data of the network classification group 400.

The operation mode of the second deep learning neural network 420 and the design method of the control signal C2 are similar to the first deep learning neural network 410, and the description will not be repeated here.

Although the first deep learning neural network 410 or the second deep learning neural network 420 alone can already determine whether there are defects in the weld bead to a certain extent, if the third deep learning neural network 430 can be used together, There may be opportunities to further improve the accuracy of defect judgment.

In summary, the weld bead inspection system proposed in this new specification uses a trained deep learning neural network to interpret radiographic films. Not only can it achieve an automated effect, but the fixed training results can make the standard of film interpretation consistent. Reduce the problem of unstable safety caused by differences in interpretation standards.

10: Weld Bead Inspection System 20, P: image to be checked 100: Radiographic equipment 110, 300A, 300B: Image content judgment device 120, 400: neural network classification group 200, 205: Surrounding image 210: Pipeline image 220: Weld Bead Image 250: area 310: Image position judgment unit 320: image pre-processing unit 330: Image Extractor 340: Neural Network 410: The first deep learning neural network 420: The second deep learning neural network 430: The third deep learning neural network C1, C2: control signal IN: After penetration P1: Image of weld bead area P2: Part of the weld bead area image T: Initial image

Fig. 1 is a system block diagram of a weld bead inspection system according to an embodiment of the present invention. 2 is a schematic diagram of an image to be inspected generated by a weld bead inspection system according to an embodiment of the present invention. 3A is a circuit block diagram of an image content determination device of a weld bead inspection system according to an embodiment of the present invention. 3B is a circuit block diagram of an image content discrimination device of a weld bead inspection system according to another embodiment of the present invention. 4 is a circuit block diagram of a neural network classification group of a weld bead inspection system according to an embodiment of the present invention. 5 is a logical definition diagram of the operation logic of the neural network classification group of the weld bead inspection system according to an embodiment of the present invention.

10: Weld Bead Inspection System

100: Radiographic equipment

110: Image content discrimination device

120: Neural Network Classification Group

IN: After penetration

P: image to be checked

P1: Image of weld bead area

Claims (7)

A radiographic-based weld bead inspection system, suitable for detecting the state of a weld bead of a pipeline to be inspected, is characterized by including: A radiographic device that receives a penetrating ray that penetrates the pipeline to be inspected to generate an image to be inspected that represents the intensity distribution state of the penetrating ray; An image content discrimination device electrically coupled to the radiography device to obtain the image to be inspected, and analyze the image to be inspected to obtain an image of a weld bead region in the image to be inspected showing the weld bead; and A neural network classification group, electrically coupled to the image content judging device to obtain the weld bead area image as the input of the neural network classification group, and judge the weld bead according to the inputted image of the weld bead area The condition of existing defects. The weld bead inspection system according to claim 1, wherein the image content judging device includes: An image position determining unit that receives the image to be inspected and determines that the image to be inspected presents an initial image of the pipeline to be inspected according to the hue, saturation, and brightness of the image to be inspected; and An image pre-processing unit electrically coupled to the image position determining unit to obtain the initial image from the image position determining unit, and perform a median filtering operation and an erosion operation on the content of the initial image to generate the weld bead region image . The weld bead inspection system according to claim 1, wherein the neural network classification group includes: A first deep learning neural network, electrically coupled to the image content discrimination device to obtain the weld bead area image as input, the first deep learning neural network by using a plurality of first defects with weld bead defects The training image is trained to obtain a first defect judgment training result, and the condition of the defect in the weld bead presented in the weld bead area image is judged according to the first defect judgment training result. The weld bead inspection system according to claim 3, wherein the neural network classification group further includes: A second deep learning neural network is electrically coupled to the image content judging device to obtain the weld bead area image as input. The second deep learning neural network uses multiple second defects with weld bead defects The training images are trained to obtain a second defect judgment training result, and the condition of the defects in the weld bead presented in the weld bead area image is judged according to the second defect judgment training result, Wherein, the first defect judgment training result is not completely the same as the second defect judgment training result. The weld bead inspection system according to claim 3, wherein the image content discrimination device further includes: An image picker takes a part of the weld bead area image and outputs it as a part of the weld bead area image. The weld bead inspection system according to claim 5, wherein the neural network classification group further includes: A third deep learning neural network is electrically coupled to the image content judging device to obtain the partial weld bead area image as input. The third deep learning neural network uses a plurality of third deep learning neural networks with weld bead defects. The defect training image is trained to obtain a third defect judgment training result, and the condition of the defect in the weld bead shown in the partial weld bead region image is judged according to the third defect judgment training result. The weld bead inspection system according to claim 1, wherein the image content judging device includes: A neural network is electrically coupled to the radiographic device to obtain the image to be inspected as input. The neural network obtains an image discrimination training result by training with a plurality of pipeline images with weld bead, and according to The image determines the training result and obtains the image of the weld bead area from the image to be inspected.

Images