Weld joint radiograph intelligent grading screening system and method based on deep learning

Abstract

The invention discloses an intelligent grading screening system and method for a welding seam radiograph based on deep learning, and the system and method achieve the automatic recognition, screening and analysis of defects in the welding seam radiograph through a three-stage screening process, a dynamic area selection mechanism and a man-machine interaction design. The system is combined with a transfer learning method, welding seam characteristics can be learned from a database, preliminary screening and analysis and amplification of suspected negative films are further carried out, and unqualified negative films are classified and sorted according to defect types. And the man-machine interaction design provides convenient confirmation and modification functions for film evaluation personnel, and finally, on the basis of efficient screening, the system is combined with manual confirmation, so that intelligence and automation of welding seam quality evaluation are realized, the detection efficiency and accuracy are remarkably improved, and the workload of the film evaluation personnel is effectively relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of non-destructive testing of special equipment, and in particular to the technical field of a system and method for intelligent grading and screening of weld radiographs based on deep learning. Background Art

[0002] Weld radiographic inspection technology uses X-rays, gamma rays, or digital radiographs to image the weld and analyze the resulting films for defects to determine the weld's internal quality. Traditional weld inspection defect assessment relies primarily on manual labor, requiring personnel to analyze numerous films under intense, long-term work. This can lead to fatigue and distraction, resulting in low inspection efficiency and the potential for misjudgments and missed detections, seriously impacting the accuracy of test results.

[0003] While some research has attempted to introduce computer vision and image recognition technologies into weld inspection, existing technologies still face numerous limitations. First, existing single-stage object detection models exhibit significant shortcomings in identifying small defects, with misjudgment rates exceeding 40% for cracks with critical dimensions (such as length and width) less than 0.5 mm. Second, existing models exhibit poor data adaptability. Significant differences in radiographic features across different equipment and process conditions limit the model's generalization capabilities and increase transfer training costs by over 300%. Furthermore, existing systems generally lack human-machine collaboration mechanisms, failing to establish a closed-loop workflow of "machine prediction + human correction," hindering efficient defect analysis and confirmation.

[0004] The existence of these problems has seriously restricted the intelligent development of weld detection technology, and there is an urgent need for an efficient weld evaluation system based on deep learning and transfer learning. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art and propose a deep learning-based intelligent grading and screening system and method for weld radiographs. By combining a dynamic selection mechanism and human-computer interaction design, the system further improves the automation level and detection accuracy of film evaluation through intelligent screening and classification functions, providing a solution for realizing intelligent and efficient weld detection.

[0006] To achieve the above objectives, the present invention proposes a deep learning-based intelligent grading and screening system for weld radiographs, comprising: The feature learning module uses deep learning algorithms and auto-encoding technology to automatically learn features in weld images from annotated film data and extract potential defect information; The structural parameter fusion module is used to receive structured data and perform multimodal data fusion with image features through a feature fusion network; The three-level screening module includes a primary screening unit, a fine screening unit, and a classification unit, which are used for preliminary screening, detailed analysis, and defect classification respectively; Dynamic selection interaction module, used to automatically identify defect areas and provide high-resolution defect images through dynamic adjustment of the magnification window and super-resolution reconstruction technology; Human-computer interaction design provides a user interface to support film reviewers in confirming or modifying system results.

[0007] Preferably, the structured data includes material properties, plate thickness, and welding process parameters.

[0008] Preferably, the feature learning module adopts a pre-trained model and a transfer learning strategy to adapt to the weld detection requirements under different equipment and scenarios, and automatically adjusts the model sensitivity threshold according to the material information input by the structural parameter fusion module.

[0009] Preferably, the three-level screening module includes: The initial screening unit, based on the EfficientNet-B4 deep learning model, performs preliminary screening of weld films, rejecting normal films and passing questionable films to the next stage; The fine screening unit uses a multi-channel residual autoencoder to generate a defect probability heat map and automatically mark the top three suspicious areas; The classification unit, combining ResNet50 pre-trained weights and graph neural networks, amplifies and classifies suspicious areas based on the fused multimodal features, achieving accurate classification of typical defects.

[0010] Preferably, the typical defects include pores, cracks and inclusions.

[0011] Preferably, the dynamic selection interaction module supports the user to select the region of interest through a frame selection or zoom tool, and triggers super-resolution reconstruction, recalculates in combination with local features, and optimizes the defect recognition results.

[0012] Preferably, the human-computer interaction design adopts multi-touch and gesture operation, supports heat map overlay display and dynamic adjustment, and allows film reviewers to manually confirm or modify the system results.

[0013] Another object of the present invention is to provide a method for an intelligent grading and screening system for weld radiographs based on deep learning as described in any one of the above items, comprising the following steps: Step S1, data uploading and preprocessing, uploading the annotated weld radiograph data and associated process parameters to the weld radiograph intelligent grading and screening system based on deep learning as described in any one of claims 1 to 7, and performing image normalization, denoising, and enhancement; Step S2: feature learning module training, using pre-training model and transfer learning strategy, combined with data enhancement technology, to perform feature extraction training on weld film data; Step S3: Dynamic area selection and super-resolution reconstruction, automatically identifying defect areas and improving image clarity through super-resolution reconstruction technology; Step S4: Human-computer interaction and confirmation: confirm or modify the system results through the user interface.

[0014] Preferably, the initial screening stage uses the EfficientNet-B4 model, the fine screening stage adopts a multi-channel residual autoencoder to generate a defect probability heat map and automatically mark the top 3 suspicious areas, the classification stage combines the ResNet50 pre-trained weights and graph neural network, and the dynamic selection and super-resolution reconstruction stage uses the ESRGAN algorithm to enlarge the image to 4K resolution.

[0015] Preferably, the process parameters in step S1 include material, thickness, and welding method.

[0016] Beneficial effects of the present invention's deep learning-based intelligent grading and screening system and method for weld radiographs: By constructing a deep learning model, a dynamic selection mechanism, and human-computer interaction design, the present invention achieves an intelligent upgrade for weld radiograph inspection. The system employs a three-level screening process, effectively reducing the workload of manual film evaluation and improving inspection efficiency and accuracy. Furthermore, the application of a meta-learning framework and data augmentation techniques significantly enhances the system's adaptability and generalization capabilities. By combining super-resolution reconstruction with local feature recalculation, the system can provide more refined defect analysis, offering a viable solution for the intelligent development of weld inspection.

[0017] The features and advantages of the present invention will be described in detail through embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the intelligent grading and screening system and method for weld radiographs based on deep learning of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. However, it should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the present invention.

[0020] In the description of the present invention, it should be noted that the terms "center", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or the orientations or positional relationships in which the inventive product is usually placed when in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", "third" and the like are only used to distinguish the description and cannot be understood as indicating or implying relative importance. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined. "Several" means one or more, unless otherwise clearly and specifically defined.

[0021] In the description of the present invention, it should also be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances. Example 1

[0022] The present invention provides a deep learning-based intelligent grading and screening system for weld radiographs, comprising: The feature learning module uses deep learning algorithms (such as convolutional neural networks (CNNs)) and autoencoding technology to automatically learn features from weld radiographs from large amounts of annotated data, efficiently and accurately extracting potential defect information. By leveraging pre-trained models (such as EfficientNet-B4 and ResNet50) and transfer learning strategies, the system can quickly adapt to the weld inspection needs of different equipment and scenarios.

[0023] The structural parameter fusion module allows users to input the structural and material parameters of the weld, including but not limited to: weld material (such as carbon steel, stainless steel, aluminum alloy, etc.), weld thickness, welding method, welding process parameters, etc. These parameters serve as auxiliary information throughout the screening process to optimize defect detection and classification results and improve the system's adaptability to welds of different materials and thicknesses.

[0024] The three-level screening module includes: Initial screening: This unit uses a trained deep learning model to perform preliminary screening, eliminating clear negatives and selecting questionable ones. The initial screening unit uses the EfficientNet-B4 deep learning model, combined with data augmentation and optimization strategies, to achieve efficient classification. Normal films (confidence level >95%) are directly filed, while questionable films proceed to the next stage.

[0025] Fine Screening Unit: This unit performs detailed analysis of suspected negatives and, through transfer learning, transfers the ability to identify specific defect types to new datasets to further improve defect recognition accuracy. The Fine Screening Unit uses a multi-channel residual autoencoder (MC-RAE) to generate a defect probability heat map and automatically annotate the top three suspicious areas.

[0026] Classification Unit: Based on suspected areas, a dynamic selection mechanism is used to select and magnify image regions, amplifying suspected defects to improve image assessment accuracy. Combining ResNet50 pre-trained weights with a graph neural network (GNN), a defect type discriminator is constructed to accurately classify typical defects such as pores, cracks, and inclusions.

[0027] The dynamic selection interactive module automatically identifies the defect area for each film and dynamically adjusts the magnification window to automatically select the area for magnification. This mechanism helps quickly focus on critical areas of weld quality issues and provides more detailed image analysis. The dynamic selection interactive module allows film reviewers to select a region of interest (ROI) using a selection box or magnification tool, triggering super-resolution reconstruction (ESRGAN algorithm, up to 4K resolution), recalculating the image based on local features to optimize defect identification results.

[0028] Defect classification and organization: For the final unqualified films, the system categorizes and organizes them according to defect type (such as pores, cracks, and inclusions) and generates a classification report for reviewers' reference. The classification unit combines ResNet50 pre-trained weights and a graph neural network (GNN) to construct a defect type discriminator, achieving accurate classification of typical defects.

[0029] Human-computer interaction design, the system is designed with a friendly human-computer interaction interface, allowing film reviewers to manually confirm or modify on the basis of automated processing. Users can quickly view the defect information of each film through the interface, and optimize the automated screening process by adjusting parameters. The interface design adopts multi-touch and gesture operations, supports heat map overlay display and dynamic adjustment, and enhances user experience. The human-computer interaction interface also integrates a parameter input module, which is convenient for film reviewers to view and modify the structural and material parameters of the weld at any time, and conduct a comprehensive evaluation of the system results based on these parameters. This embodiment realizes the intelligent upgrade of weld radiographic detection by constructing a deep learning model, a dynamic selection mechanism and a human-computer interaction design. The system adopts a three-level screening process, which effectively reduces the workload of manual film review and improves detection efficiency and accuracy. At the same time, the application of meta-learning framework and data enhancement technology significantly improves the adaptability and generalization ability of the system. By combining super-resolution reconstruction and local feature recalculation, the system can provide more refined defect analysis, providing a feasible solution for the intelligent development of weld detection. The process is as follows: Figure 1 shown. Example 2

[0030] This embodiment proposes a method for intelligent grading and screening of weld radiographs based on deep learning based on the system of embodiment 1.

[0031] Implementation Background A company (hereinafter referred to as "Company A") primarily installs pipelines and pressure vessels. The quality of its welds is directly related to the safe operation and service life of its equipment. Traditional weld radiographic inspection and evaluation relies on manual labor, which is subject to low efficiency, high subjectivity, and fatigue. To address these issues, Company A introduced a deep learning-based weld radiographic inspection and evaluation assistance system (hereinafter referred to as "the system") and applied it in actual production.

[0032] Implementation conditions Hardware requirements: The hardware required for system operation includes a GPU server (NVIDIA A100), a storage device (10TB storage space), an industrial-grade monitor (supporting 4K resolution), and a network environment (Gigabit network).

[0033] Software requirements: The system is developed based on deep learning frameworks such as TensorFlow 2.5 and PyTorch 1.9, and the operating environment is Linux Ubuntu 20.04.

[0034] Data Preparation Company A provided 5,000 annotated historical weld radiographs, containing various defect types such as normal welds, pores, cracks, and inclusions. In addition, there were 100,000 weld radiographs awaiting evaluation, which would be used for system training and validation.

[0035] Implementation process: Parameter input Before the system begins operation, the operator enters the weld's structural and material parameters through the parameter input module, including weld material (carbon steel), weld thickness (10 mm), welding method (manual arc welding), and welding process parameters (welding current, welding voltage, etc.). These parameters serve as auxiliary information to guide subsequent feature extraction, defect classification, and manual review.

[0036] Data uploading and preprocessing 5,000 annotated historical weld film images were uploaded to the system server and preprocessed, including image normalization, denoising, and enhancement. After preprocessing, the system randomly divided the images into training, validation, and test sets in a ratio of 7:2:1. Furthermore, 100,000 weld film images awaiting evaluation were initially screened and classified by the system.

[0037] Feature learning module training The system uses EfficientNet-B4 and ResNet50 as pre-trained models, combined with a transfer learning strategy, to perform feature extraction training on weld film data. During training, the system employs data augmentation techniques (such as rotation, flipping, and random cropping) to improve the model's generalization capabilities. After 500 epochs of training, the model achieved an accuracy of 97.8% and an R² value of 96.5% on the validation set.

[0038] Three-level screening process Initial screening: The system uses the EfficientNet-B4 model to perform a preliminary screening of all films. Experimental data shows that this initial screening process accurately identifies over 95% of normal films, significantly reducing the workload of manual film evaluation. For 100,000 weld films to be evaluated, the initial classification can be completed in approximately one hour.

[0039] Fine Screening: For films marked as "suspicious" during the initial screening, the system uses a multi-channel residual autoencoder (MC-RAE) to generate a defect probability heat map and automatically annotate the top three suspicious areas. The fine screening process achieved a defect detection rate of 98.2% and a false alarm rate of 4.7%.

[0040] Classification: Combining ResNet50 pre-trained weights with a graph neural network (GNN), the system zooms in on and classifies suspicious areas. Experimental results show that the system achieves 96.7% accuracy in classifying typical defects such as pores, cracks, and inclusions.

[0041] Dynamic selection and super-resolution reconstruction The system's dynamic selection module automatically identifies defect areas and uses the ESRGAN algorithm to upscale the image to 4K resolution. In practice, the upscaled image clarity is significantly improved, with defect details more visible, allowing film reviewers to identify defects more quickly and accurately.

[0042] Human-computer interaction and confirmation The system provides a friendly human-computer interaction interface, through which film reviewers can quickly view defect information and confirm or modify system results. Implementation Effect

[0043] Improved efficiency: The system can process 200 weld films per hour, four times more efficient than manual film evaluation. For 100,000 weld films to be evaluated, the system can complete preliminary classification and screening in approximately 50 hours.

[0044] Defect detection rate: The system's detection rate for defects such as pores, cracks, and inclusions reaches 98.2%, a significant improvement over the 90% detection rate of manual film evaluation.

[0045] Reduced false alarm rate: The system false alarm rate is controlled at 4.7%, while the manual false alarm rate is 12%, significantly improving detection accuracy.

[0046] Manual review time is reduced: Through the system's three-level screening and dynamic selection functions, manual review time is reduced by 50%, and the workload of film reviewers is significantly reduced.

[0047] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of the present invention. Therefore, based on the innovative concept of the present invention, changes and modifications to the embodiments described herein, or equivalent structural or equivalent process transformations made using the contents of the present invention's description and drawings, and direct or indirect application of the above technical solutions to other related technical fields, are all included in the scope of protection of the present invention's patent.

Claims

1. The intelligent grading and screening system for weld radiographs based on deep learning is characterized by: include: The feature learning module uses deep learning algorithms and auto-encoding technology to automatically learn features in weld images from annotated film data and extract potential defect information; The structural parameter fusion module is used to receive structured data and perform multimodal data fusion with image features through a feature fusion network; The three-level screening module includes a primary screening unit, a fine screening unit, and a classification unit, which are used for preliminary screening, detailed analysis, and defect classification respectively; Dynamic selection interaction module, used to automatically identify defect areas and provide high-resolution defect images through dynamic adjustment of the magnification window and super-resolution reconstruction technology; Human-computer interaction design provides a user interface to support film reviewers in confirming or modifying system results.

2. The deep learning-based intelligent grading and screening system for weld radiographs according to claim 1, characterized in that: The structured data includes material properties, plate thickness, and welding process parameters.

3. The deep learning-based intelligent grading and screening system for weld radiographs according to claim 1, characterized in that: The feature learning module adopts a pre-trained model and transfer learning strategy to adapt to the weld detection needs under different equipment and scenarios, and automatically adjusts the model sensitivity threshold according to the material information input by the structural parameter fusion module.

4. The deep learning-based intelligent grading and screening system for weld radiographs according to claim 1, characterized in that: The three-level screening module includes: The initial screening unit, based on the EfficientNet-B4 deep learning model, performs preliminary screening of weld films, rejecting normal films and passing questionable films to the next stage; The fine screening unit uses a multi-channel residual autoencoder to generate a defect probability heat map and automatically mark the top three suspicious areas; The classification unit, combining ResNet50 pre-trained weights and graph neural networks, amplifies and classifies suspicious areas based on the fused multimodal features, achieving accurate classification of typical defects.

5. The deep learning-based intelligent grading and screening system for weld radiographs according to claim 1, characterized in that: The typical defects include pores, cracks, and inclusions.

6. The deep learning-based intelligent grading and screening system for weld radiographs according to claim 1, characterized in that: The dynamic selection interaction module supports users to select an area of ​​interest through a frame selection or zoom tool, triggers super-resolution reconstruction, and recalculates based on local features to optimize defect recognition results.

7. The deep learning-based intelligent grading and screening system for weld radiographs according to claim 1, characterized in that: The human-computer interaction design adopts multi-touch and gesture operation, supports heat map overlay display and dynamic adjustment, and allows film reviewers to manually confirm or modify the system results.

8. A method for the deep learning-based intelligent grading and screening system for weld radiographs according to any one of claims 1 to 7, characterized in that: The following steps are involved: Step S1, data uploading and preprocessing, uploading the annotated weld radiograph data and associated process parameters to the weld radiograph intelligent grading and screening system based on deep learning as described in any one of claims 1 to 7, and performing image normalization, denoising, and enhancement; Step S2: feature learning module training, using pre-training model and transfer learning strategy, combined with data enhancement technology, to perform feature extraction training on weld film data; Step S3: Dynamic area selection and super-resolution reconstruction, automatically identifying defect areas and improving image clarity through super-resolution reconstruction technology; Step S4: Human-computer interaction and confirmation: confirm or modify the system results through the user interface.

9. The deep learning-based intelligent grading and screening method for weld radiographs according to claim 1, characterized in that: The initial screening stage uses the EfficientNet-B4 model, the fine screening stage adopts a multi-channel residual autoencoder to generate a defect probability heat map and automatically mark the top three suspicious areas. The classification stage combines the ResNet50 pre-trained weights and graph neural network. The dynamic selection and super-resolution reconstruction stage uses the ESRGAN algorithm to enlarge the image to 4K resolution.

10. The deep learning-based intelligent grading and screening method for weld radiographs according to claim 1, characterized in that: The process parameters in step S1 include material, thickness, and welding method.

Images