Intelligent detection method for x-ray flaw detection

Abstract

The application discloses an X-ray flaw detection intelligent detection method, which comprises the following steps: using an X-ray detection device to perform real-time dynamic detection on a steel pipe; a computer receives a dynamic image in real time, and sets a window position and a window width of the dynamic image; the detected dynamic image is saved in real time; an intelligent defect recognition model is established based on an artificial intelligence deep learning method; real-time dynamic image reasoning is performed through the dynamic recognition model, an alarm is given after a defect is detected, and a stop signal is sent to stop the steel pipe from moving; static image reasoning is performed on the static steel pipe through a static recognition model, a defect image is recognized and uploaded, and dynamic detection of the steel pipe is continued; after the detection of the whole steel pipe is completed, the detection work is stopped and the steel pipe is exited; and the beneficial effects are that a complete set of X-ray flaw detection solutions are provided by aiming at a weld X-ray flaw detection image, and the actual problems of users on site are well solved.

Description

Technical Field

[0001] This invention relates to X-ray nondestructive testing and intelligent image recognition, specifically to an intelligent X-ray flaw detection method. Background Technology

[0002] Due to varying environmental conditions and welding processes, various defects inevitably occur during welding (such as cracks, porosity, lack of fusion, incomplete penetration, and slag inclusions). Furthermore, due to the nature of these defects and the principles of radiographic imaging, some defects are relatively difficult to detect, making the entire flaw detection process increasingly complex and significantly increasing the difficulty of image interpretation. These defects directly jeopardize the quality of welded products and have a significant impact on the safety and performance of welded structures.

[0003] Currently, commonly used non-destructive testing methods include ultrasonic testing, magnetic particle testing, and radiographic testing. Among these, radiographic imaging can visually display the size and shape of volumetric defects inside welds, making it easy to determine the nature of the defects and thus having a wide range of applications. Compared to traditional film imaging, digital radiographic imaging is gradually being promoted due to its advantages such as high signal-to-noise ratio and the elimination of the need for film cleaning. However, the images obtained from flaw detection still mainly rely on manual identification, and the quality of identification depends heavily on the operator's skills, physical condition, and other factors.

[0004] Because X-rays have radiation properties, all equipment is placed in a separate X-ray room, including the object being inspected (steel pipe), the X-ray generator, and a digital flat panel detector. During the inspection, the X-ray generator emits rays that penetrate the object and strike the digital flat panel detector, which is connected to a computer via a gigabit network cable. Operators view the X-ray images to locate and determine defects, and then upload the images for evaluation. Currently, X-ray inspection of welds is primarily conducted manually by operators.

[0005] Existing manual evaluation and flaw detection techniques have the following drawbacks: 1. Due to the diversity of weld defect identification, specialized technicians observe and compare weld areas in X-ray images to identify defects. Manual evaluation is subjective and flexible, with significant limitations. It is highly dependent on the qualifications of the evaluators, resulting in poor objectivity. Other factors can significantly influence the identification results, such as the evaluator's sense of responsibility, skill level, experience, workload, and available time. Different evaluators may yield different evaluations of the same weld image, inevitably leading to variations and affecting the weld evaluation outcome.

[0006] 2. Radiographic testing is a highly specialized non-destructive testing (NDT) technique, demanding high intensity and pressure on operators. For the inspection of dynamic images of pipe bodies, the entire process relies on on-site operators visually monitoring the display screen. In dynamic situations, the images are blurry, fast-moving, and noisy, and the moving pipes easily cause visual fatigue. Due to the limitations of the human eye and the low quality of single-frame dynamic images, the movement speed of pipe inspection has been repeatedly reduced. Therefore, the weld quality of the pipes is significantly affected by subjective human factors. With the overall optimization of the production line, this step will become a bottleneck for accelerating the entire production line.

[0007] 3. Dynamic flaw detection data for the pipe body is not archived, making re-inspection and traceability impossible. Currently, no effective solution has been proposed to address these technical issues. Summary of the Invention

[0008] In response to the problems in related technologies, this invention proposes an intelligent X-ray flaw detection method to overcome the aforementioned technical problems existing in the existing related technologies.

[0009] Therefore, the specific technical solution adopted by the present invention is as follows:

[0010] An intelligent X-ray flaw detection method includes the following steps: real-time dynamic inspection of steel pipes using X-ray detection equipment.

[0011] The computer receives dynamic images in real time, sets the window level and width for the dynamic images, and saves the detected dynamic images in real time.

[0012] A defect intelligent identification model based on artificial intelligence deep learning methods;

[0013] Real-time dynamic image reasoning is performed using a dynamic recognition model. Upon detecting a defect, an alarm is triggered, and a stop signal is sent to halt the movement of the steel pipe.

[0014] Static image reasoning is performed on stationary steel pipes using a static recognition model to identify and upload defect images.

[0015] For example, continue the dynamic testing of the steel pipes;

[0016] The inspection of the entire steel pipe is completed; the inspection work is stopped and the steel pipe is removed.

[0017] Furthermore, the method for setting the window level and width of the dynamic image includes one-click automatic adjustment, mouse movement modification, and manual modification of values;

[0018] Furthermore, the one-click automatic adjustment of the window level and width settings for dynamic images includes the following steps:

[0019] Analyze the distribution of all grayscale values ​​in the current image; find the maximum and minimum grayscale values; calculate the window width by calculating the difference; calculate the window level by averaging the values.

[0020] The current image is normalized to a grayscale value of 0-255 based on the window level and width; the image with the automatically adjusted window level and width is then presented.

[0021] Furthermore, the defect intelligent identification model includes the dynamic identification model and the static identification model.

[0022] Furthermore, the method for establishing a defect intelligent identification model based on artificial intelligence deep learning includes the following steps:

[0023] Acquire image data;

[0024] Perform image data cleaning and complete defect annotation;

[0025] Based on the deep learning semantic segmentation network (UNet), dynamic recognition models and static recognition models were constructed respectively, and preprocessing and model training were performed on the two models respectively;

[0026] Post-processing of the model results;

[0027] The model is deployed, optimized, and tested to complete the construction of a model that meets the requirements of practical work. Furthermore, the preprocessing and training of the dynamic recognition model includes the following steps: converting the window level and window width of the original image;

[0028] The image size is transformed to the model input size using an image resizing algorithm: training, validation, and test sets are constructed based on the original dataset;

[0029] The training set was augmented with data through translation, rotation, and flipping; the dynamic recognition model was trained and optimized through manual parameter tuning.

[0030] Furthermore, the preprocessing and training of the static recognition model includes the following steps: using the steel pipe weld and heat-affected zone as the detection area, the original image is cropped to ensure that the cropped image covers all areas that need to be detected.

[0031] Perform window level and window width transformations on each of the cut sub-images;

[0032] The image size is transformed to the model input size using an image resizing algorithm; training, validation, and test sets are constructed based on the original dataset.

[0033] Data augmentation samples are added to the training set by translation, rotation, and flipping.

[0034] The dynamic recognition model is trained and optimized by manually adjusting parameters.

[0035] Furthermore, the post-processing of the model results includes the following steps:

[0036] The maximum probability is calculated based on the probability results output by the model, and its index is used as the category of that pixel; the contour coordinates of all defects in the image are found using a contour-finding algorithm;

[0037] The coordinates of each point are transformed to the coordinate system of the original image; the results of the dynamic recognition model and the static recognition model are then fused.

[0038] The final model results are filtered. For each defect, the size, position, contrast and confidence are calculated as feature vectors and corresponding thresholds are set. False defects that meet the conditions are filtered out, and finally accurate identification results are obtained.

[0039] Furthermore, the step of performing static image reasoning on a stationary steel pipe using a static recognition model, identifying and uploading defect images, and continuing dynamic detection of the steel pipe includes the following steps:

[0040] The system acquires static images of completely stationary steel pipes, and uses a static recognition model to infer the defects identified in the static images, displaying the defect types, locations, and sizes in real time on the display interface.

[0041] When the recognition result indicates a defect, the operator uploads the image of the intelligent recognition result to complete the scoring process and continues dynamic detection;

[0042] When the identification result indicates that there are no defects, the operator chooses to continue dynamic detection.

[0043] Furthermore, the inference of static images by calling the static recognition model includes the following steps:

[0044] A smoothed sample image is obtained by applying a smoothing filter to the static image; the smoothed sample image is then sharpened to obtain a sharpened sample image.

[0045] The sharpened sample image is subjected to defect detection, and defect information is calculated: the defect information of the static image is quantized and encoded into a static image feature vector;

[0046] The feature vector of the static image is compared with the static call model to obtain the judgment result.

[0047] The beneficial effects of this invention are as follows: It provides a complete X-ray flaw detection solution for weld radiographic images (including film-scanned digital images, DR images, and CR images), effectively solving practical problems on-site for users. Simultaneously, it integrates intelligent recognition functions for radiographic images, significantly reducing the workload of operators throughout the flaw detection and re-inspection process, improving the accuracy of defect judgment in the flaw detection system, reducing the rate of missed defects and false alarms, enhancing the objectivity of defect assessment, reducing personnel requirements, and achieving cost reduction and efficiency improvement, thereby increasing labor productivity. Furthermore, dynamic data can be saved and re-inspected, and even traceable for potential future problems if needed. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the intelligent X-ray flaw detection method according to an embodiment of the present invention:

[0050] Figure 2 This is a flowchart illustrating the software operation functions of the intelligent X-ray flaw detection method according to an embodiment of the present invention. Detailed Implementation

[0051] According to an embodiment of the present invention, an intelligent X-ray flaw detection method is provided.

[0052] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-2 As shown, the intelligent X-ray flaw detection method according to an embodiment of the present invention includes the following steps:

[0053] S1. Real-time dynamic inspection of steel pipes using X-ray detection equipment;

[0054] S2. The computer receives dynamic images in real time and sets the window position and width of the dynamic images; wherein, the method for setting the window position and width of the dynamic images includes one-click automatic adjustment, mouse movement modification, and manual modification of values.

[0055] The one-click automatic adjustment of the window level and width setting for dynamic images includes the following steps: S21, statistically analyzing the distribution of all grayscale values ​​in the current image; S22, finding the maximum and minimum grayscale values; S23, calculating the window width by the difference; S24, calculating the window level by the average value.

[0056] S25. Normalize the current image to a grayscale value of 0-255 according to the window level and window width; S26. Present the image after the window level and window width have been automatically adjusted.

[0057] The difference between the mouse-movement modification steps and the methods described above is that the calculation is no longer performed using the entire image, but rather using the region of interest selected by the mouse. This maximizes the display of the image within the mouse-selected area. Theoretically, when the mouse-selected area is the entire image, the calculated window level and width, as well as the displayed image, should match the conditions described above.

[0058] The steps to manually modify the window position and width are to normalize the manually entered window position and width before displaying it.

[0059] S3. Save the detected dynamic images in real time for easy re-inspection and traceability; S4. Build an intelligent defect recognition model based on artificial intelligence deep learning methods;

[0060] The defect intelligent identification model includes a dynamic identification model and a static identification model (the dynamic identification model is used to detect weld location, base metal defects, and other larger defects; the static identification model is used to detect defects on the weld and in the heat-affected zone at the weld edge); and establishing the defect intelligent identification model includes the following steps:

[0061] S41. Acquire image data (collect relevant image data; the user has actually archived over 450,000 still images on-site):

[0062] S42. Perform image data cleaning and complete defect annotation;

[0063] The process involves selecting valid images from the data (currently, over 20,000 static flaw detection images have been manually selected) and manually annotating the defects. By building a dedicated X-ray image annotation tool platform and combining the experience of on-site image evaluation experts and other image evaluation experts in the field of flaw detection, each image is manually annotated and verified (defect type, defect size, defect location), ultimately completing the defect annotation work.

[0064] S43. Construct dynamic and static recognition models based on the deep learning semantic segmentation network (UNet), and perform preprocessing and model training on the two models respectively;

[0065] The preprocessing and training of the dynamic recognition model includes the following steps:

[0066] S431. Convert the window level and window width of the original image;

[0067] S432. Convert the image size to the model input size using an image resizing algorithm;

[0068] S433. Construct training, validation, and test sets based on the original dataset;

[0069] S434. Data augmentation and sample expansion of the training set are performed using translation, rotation, and flipping methods.

[0070] S435. Train and optimize the dynamic recognition model through manual parameter tuning.

[0071] The preprocessing and training of the static recognition model includes the following steps:

[0072] S431': Using the steel pipe weld seam and heat-affected zone as the detection area, the original image is cropped to ensure that the cropped image covers all areas that need to be detected;

[0073] S432', Perform window level and window width conversion on each of the cut sub-images;

[0074] S433': The image size is converted to the model input size using an image resizing algorithm;

[0075] S434' Construct training, validation, and test sets based on the original dataset;

[0076] S435', Augment the training set by translation, rotation, and flipping; S436', Train and optimize the dynamic recognition model by manual parameter tuning; S44, Post-process the model results;

[0077] The post-processing of the model results includes the following steps:

[0078] S441. Calculate the maximum probability based on the probability results output by the model, and use its index as the category of the pixel;

[0079] S442. Find the contour coordinates of all defects in the image using a contour-finding algorithm;

[0080] S443. Perform coordinate transformation on each coordinate point, converting it to the coordinate system of the original image; S444. Fuse the results of the dynamic recognition model and the static recognition model;

[0081] S445. Filter the final model results. For each defect, calculate the size, position, contrast and confidence as feature vectors, and set the corresponding thresholds. Filter out false defects that meet the conditions to obtain accurate identification results.

[0082] S45. Optimize and test the model for deployment, and complete the construction of a model that meets the actual work requirements, that is, finally package it into a backend service or component that meets the actual inference performance in the field.

[0083] S5. Real-time dynamic image reasoning is performed through a dynamic recognition model. An alarm is triggered upon detecting a defect, and a stop signal is sent to stop the steel pipe from moving.

[0084] S6. Using a static recognition model, perform static image reasoning on stationary steel pipes to identify and upload defect images, and continue dynamic inspection of the steel pipes:

[0085] S6 includes the following steps:

[0086] S61. Obtain a static image of the completely stationary steel pipe, and perform inference on the static image by calling the static recognition model. The defect type, defect location, and defect size identified by the static image are displayed on the display interface in real time.

[0087] The inference of static images by calling the static recognition model includes the following steps:

[0088] S611. Apply a smoothing filter to the static image to obtain a smoothed sample image; S612. Sharpen the smoothed sample image to obtain a sharpened sample image;

[0089] S613. Perform defect detection on the sharpened sample image and calculate the defect information; S614. Quantize and encode the defect information of the static image into a static image feature vector;

[0090] S615. The feature vector of the static image is compared with the static call model to obtain a judgment result.

[0091] S62. When the recognition result indicates a defect, the operator uploads the image of the intelligent recognition result to complete the scoring process and continues dynamic detection;

[0092] S63. When the identification result indicates no defects, the operator selects to continue dynamic inspection. S7. The entire steel pipe inspection is completed; the inspection work is stopped and the steel pipe is removed.

[0093] In summary, by utilizing the above-mentioned technical solution of this invention, a complete X-ray flaw detection solution is provided for weld radiographic images (including film-scanned digital images, DR images, and CR images). This solution effectively addresses practical problems encountered by users on-site. Furthermore, by integrating intelligent recognition capabilities of radiographic images, it significantly reduces the workload of operators throughout the flaw detection and re-inspection process, improves the accuracy of defect determination in the flaw detection system, reduces the rate of missed defects and false alarms, enhances the objectivity of defect assessment, reduces personnel requirements, and achieves cost reduction and efficiency improvement, thereby increasing labor productivity. In addition, dynamic data can be saved and re-inspected, and even traceability is possible for potential future problems if needed.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent X-ray flaw detection method, characterized in that, The method includes the following steps: Real-time dynamic inspection of steel pipes is carried out using X-ray detection equipment; The computer receives dynamic images in real time and sets the window level and width of the dynamic images; The detected dynamic images are saved in real time; A defect intelligent identification model is established based on artificial intelligence deep learning methods; Real-time dynamic image reasoning is performed through a dynamic recognition model. An alarm is triggered after a defect is detected, and a stop signal is sent to stop the steel pipe from moving. Static image reasoning is performed on stationary steel pipes using a static recognition model to identify and upload defect images, and then dynamic detection of the steel pipes continues. The inspection of the entire steel pipe is completed; the inspection work is stopped and the steel pipe is removed. The defect intelligent identification model includes the dynamic identification model and the static identification model; The method for establishing a defect intelligent identification model based on artificial intelligence deep learning includes the following steps: Acquire image data; Perform image data cleaning and complete defect annotation; Based on a deep learning semantic segmentation network, dynamic recognition models and static recognition models were constructed respectively, and the two models were preprocessed and trained respectively. Post-processing of the results from the dynamic and static recognition models; Optimize and test the model for deployment, and build a model that meets the requirements of actual work. Post-processing of the results from dynamic and static recognition models includes the following steps: Calculate the maximum probability based on the probability results output by the model, and use its index as the category of the pixel; The algorithm for finding contours identifies the contour coordinates of all defects in the image. Perform coordinate transformation on each coordinate point to convert it to the coordinate system of the original image; The results of the dynamic recognition model and the static recognition model are fused together; The final model results are filtered. For each defect, the size, position, contrast and confidence are calculated as feature vectors and corresponding thresholds are set. False defects that meet the conditions are filtered out, and finally accurate identification results are obtained.

2. The intelligent X-ray flaw detection method according to claim 1, characterized in that, The methods for setting the window position and width of dynamic images include one-click automatic adjustment, mouse movement modification, and manual modification of values.

3. The intelligent X-ray flaw detection method according to claim 2, characterized in that, The one-click automatic adjustment of the window position and width settings for dynamic images includes the following steps: Analyze the distribution of all grayscale values ​​in the current image; Find the maximum and minimum values ​​of the grayscale values; The window width is obtained by calculating the difference. The window level is calculated using the average value. The grayscale values ​​of the current image are normalized based on the window level and window width. The image displays the automatically adjusted window position and width.

4. The intelligent X-ray flaw detection method according to claim 3, characterized in that, The preprocessing and training of the dynamic recognition model includes the following steps: The original image is converted in terms of window level and window width. The image size is converted to the model input size using an image resizing algorithm; Based on the original dataset, construct training, validation, and test sets; Data augmentation samples are added to the training set by translation, rotation, and flipping. The dynamic recognition model is trained and optimized by manually adjusting parameters.

5. The intelligent X-ray flaw detection method according to claim 4, characterized in that, Preprocessing and training the static recognition model includes the following steps: using the steel pipe weld and heat-affected zone as the detection area, the original image is cropped to ensure that the cropped image covers all areas that need to be detected; Perform window level and window width conversion on each of the cut sub-images; The image size is converted to the model input size using an image resizing algorithm; Based on the original dataset, construct training, validation, and test sets; Data augmentation samples are added to the training set by translation, rotation, and flipping. The static recognition model is trained and optimized by manually adjusting parameters.

6. The intelligent X-ray flaw detection method according to claim 5, characterized in that, Static image inference is performed on stationary steel pipes using a static recognition model to identify and upload defect images, followed by dynamic detection of the steel pipes. This includes the following steps: Obtain a static image of a completely stationary steel pipe, and use a static recognition model to infer the static image, displaying the defect type, defect location, and defect size identified in the static image on the display interface in real time. When the recognition result is determined to be defective, the operator uploads the image of the intelligent recognition result to complete the scoring work and continues to carry out dynamic detection. When the identification result indicates that there are no defects, the operator chooses to continue dynamic detection.

7. The intelligent X-ray flaw detection method according to claim 6, characterized in that, The inference of static images by calling the static recognition model includes the following steps: A smoothed sample image is obtained by applying a smoothing filter to a static image. The smoothed sample image is sharpened to obtain the sharpened sample image. Defect detection is performed on the sharpened sample image to calculate defect information; Defect information in static images is quantized and encoded into static image feature vectors; The static image feature vector is compared with the static recognition model to obtain a judgment result.

Images