Ultrasonic phased array intelligent marking and defect diagnosis method for welding joint

Abstract

The invention relates to an ultrasonic phased array intelligent marking and defect diagnosis method for a welded joint, and belongs to the technical field of nondestructive testing and artificial intelligence crossing. The method comprises the following steps: constructing an annotation platform to complete intelligent pre-annotation and manual interactive annotation of an ultrasonic image, and training an initial annotation model; defect feature extraction and grade division are realized based on an improved semantic segmentation model; establishing a data closed-loop feedback mechanism, and distilling and optimizing annotation data through expert knowledge; a remote auxiliary diagnosis platform is constructed, one-key model deployment and version management are realized by adopting a containerized deployment technology, and interactive visual analysis, whole-process project management and remote collaborative diagnosis functions are integrated. The closed-loop optimization of labeling and diagnosis is realized, and the accuracy of defect identification and the diagnosis efficiency are improved.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of nondestructive testing and artificial intelligence, and specifically relates to an ultrasonic phased array intelligent annotation and defect diagnosis method for welded joints. Background Technology

[0002] In the industrial manufacturing sector, the quality of welded joints directly affects the safety and reliability of equipment operation. Ultrasonic phased array testing technology, with its flexible beam control capabilities and clear imaging effects, has become a core method for defect detection in welded joints. With the application of artificial intelligence technology in nondestructive testing, intelligent diagnostic solutions based on algorithms such as semantic segmentation are gradually emerging, aiming to improve the efficiency and accuracy of defect identification. However, current ultrasonic phased array intelligent annotation and defect diagnosis technologies for welded joints still have several limitations, making it difficult to meet the actual needs of industrial scenarios for high-precision and high-efficiency testing.

[0003] In the annotation stage, existing technologies face the dilemma of insufficient collaboration between intelligence and human expertise. Traditional annotation methods often rely on manual drawing of defect areas frame by frame, which is not only time-consuming and labor-intensive but also susceptible to the influence of the inspector's experience and subjective judgment, making it difficult to guarantee the consistency and accuracy of the annotated data. Some solutions that introduce intelligent pre-annotation often suffer from omissions and mislabeling due to the limited ability of the initial model to identify complex defects (such as microcracks and inclusions) in ultrasound images. Furthermore, the lack of convenient interactive correction tools makes it difficult for inspectors to quickly adjust the annotation results, ultimately leading to low efficiency in building high-quality training datasets and hindering the performance improvement of subsequent diagnostic models.

[0004] In the defect diagnosis stage, existing semantic segmentation models are not sufficiently adaptable to ultrasonic images of welded joints. These images often suffer from noise interference, blurred defect features, and low contrast between the background and the target. Traditional models struggle to accurately extract defect boundaries, easily leading to oversegmentation or undersegmentation. Furthermore, defect level classification often relies on fuzzy standards based on human experience, lacking unified judgment rules based on the physical quantities of defect features. This results in a lack of comparability of diagnostic results across different scenarios, and the accuracy of type identification and level determination for some complex defects (such as scenarios with multiple coexisting defects) is insufficient, failing to provide a reliable basis for subsequent quality assessment.

[0005] The lack of a data-driven model iteration mechanism means that existing technologies often present a one-way process where "diagnosis is the endpoint." Key data such as low-confidence results and complex defect cases generated during the diagnosis process are not effectively fed back to the annotation stage. Even when some solutions involve expert review, a systematic knowledge distillation mechanism has not been established, making it impossible to transform the expert-corrected annotation rules and judgment logic into guiding parameters for model optimization. This makes it difficult for diagnostic models to continuously iterate based on actual data, resulting in poor adaptability to joint defects under different materials and welding processes, and a tendency for diagnostic accuracy to decline after long-term use.

[0006] Weak remote collaboration and deployment management capabilities further restrict diagnostic efficiency. In industrial scenarios, testing sites are often widely distributed, while expert resources are mostly concentrated in core laboratories. When on-site testing personnel encounter difficult problems, it is difficult to quickly obtain expert support, leading to delays in problem handling. At the same time, the deployment process of existing diagnostic models is complex, requiring professional personnel to configure the environment and debug parameters, and lacks cross-terminal compatible access mechanisms, resulting in poor data synchronization and functional compatibility between desktop and mobile terminals. In addition, most solutions do not integrate full-process project management and interactive visualization analysis functions, making it difficult to achieve unified management and intuitive presentation of test data and diagnostic results, which is not conducive to traceability and quality control of the testing process.

[0007] In summary, the current technology's shortcomings in areas such as annotation collaboration, model accuracy, data closure, and remote collaboration have become bottlenecks restricting the practical application of ultrasonic phased array intelligent inspection technology for welded joints. To address these issues, a technical solution is urgently needed that can optimize the entire process from intelligent annotation to precise diagnosis, data feedback, and remote collaboration, thereby improving the accuracy and efficiency of defect detection and meeting the high standards of welding quality control required by industrial manufacturing. Summary of the Invention

[0008] To address the aforementioned problems in the existing technology, this invention provides an ultrasonic phased array intelligent annotation and defect diagnosis method for welded joints. The objective of this invention can be achieved through the following technical solutions: S1: Construct a labeling platform, acquire the original phased array ultrasonic images of the welded joint, establish an intelligent pre-labeling and manual interactive labeling mechanism to initially label the images, generate a labeled training dataset, and simultaneously train and test the initial labeling model for defect identification. S2: Based on the improved U-Net3 semantic segmentation model, the image is preprocessed and intelligently segmented to extract defect features. Based on predefined judgment rules, the defect level is divided and intelligent diagnosis results are output. At the same time, the intelligent diagnosis results are judged to identify the defect type. S3: Establish a data closed-loop feedback mechanism to feed the intelligent diagnostic results back to the annotation platform, screen and evaluate the diagnostic data based on diagnostic confidence and defect criticality, and use the expert knowledge distillation mechanism to annotate, correct and transform the diagnostic results to form optimized standard training data. S4: Establish a remote auxiliary diagnostic platform that supports multi-terminal access, adopt containerized deployment technology for one-click deployment and version management of semantic segmentation models and optimized annotation models, and build a comprehensive diagnostic environment that integrates interactive visual analysis, full-process project management and remote collaborative diagnosis.

[0009] As a preferred embodiment of the present invention, the method for initially labeling the image by combining intelligent pre-annotation and manual interactive annotation is as follows: The uploaded original phased array ultrasound image is automatically annotated by loading a pre-trained initial semantic segmentation model, generating initial annotation results containing defect regions. When the automatic annotation results do not meet the annotation requirements, a manual interactive annotation process is initiated, where the defect boundaries are manually drawn and corrected using the interactive tools provided by the annotation platform. During the annotation process, image panning operations are performed using key combinations to view different regions, and the annotation results can be undone and redone using function keys. Finally, through the coordinated efforts of intelligent pre-annotation and manual interactive annotation, the defect regions in the phased array ultrasound image are accurately marked.

[0010] Specifically, the method for training and testing the initial labeling model for defect identification is as follows: An initial labeling model based on an encoder-decoder structure is constructed. During the training phase, a multi-stage optimization strategy is adopted. Basic training is performed by setting a training period and batch sample size, and a dynamic learning rate adjustment strategy is used simultaneously to optimize the convergence effect. During the testing phase, a dual validation system is established. On the one hand, the segmentation effect is evaluated based on the region overlap, and on the other hand, the recognition accuracy is evaluated based on the confidence level. Finally, the stability of the model in practical applications is determined by cross-validation.

[0011] Specifically, the image preprocessing and intelligent segmentation method is as follows: A multi-level preprocessing process is adopted. Based on the target detection network, the B-type atlas region in the ultrasound image is located and adaptively cropped. Then, color space conversion and signal intensity analysis are used to achieve preliminary screening of electrofusion welding defects. In the intelligent segmentation stage, a semantic segmentation model with fused attention mechanism is adopted. Multi-scale feature extraction and feature pyramid fusion technology are used to enhance the recognition accuracy of defect boundaries. The segmentation results are also processed and optimized by combining connected component analysis.

[0012] Specifically, the predefined judgment rules for classifying defect levels are as follows: A multi-defect collaborative judgment mechanism is established to normalize the identified defect features and convert the physical quantities of defect features into standardized defect severity values. A hierarchical evaluation system is established based on defect types. For different types of defects, corresponding feature parameter extraction and threshold comparison methods are adopted. Based on the quantitative evaluation results of defect types, defect levels are classified in combination with predefined level thresholds.

[0013] The graded evaluation system based on defect type includes: for cold welding and over-welding defects, the grade is determined according to the degree of deviation between the characteristic line and the resistance wire spacing from the standard value; for hole defects, the grade is comprehensively evaluated based on the proportion of the missing part to the length of the fusion zone and the ratio of the hole height to the pipe wall thickness; for inclusion defects, different length threshold standards are used according to the degree of penetration with the inner cold welding zone; for socket misalignment defects, the grade is determined based on the length of the resistance wire at the unwelded pipe; and for resistance wire misalignment defects, the grade is determined based on the ratio of the misalignment amount to the resistance wire spacing.

[0014] Specifically, the method for determining the intelligent diagnostic results and identifying the defect type is as follows: An intelligent defect type discrimination mechanism is established, which performs multi-image defect level comparison analysis to obtain the highest defect level among all images of the same joint; discrimination is performed based on defect type priority rules, and when non-spacing defects exist, the non-spacing defects are selected as the primary defect type; feature value analysis is performed on the selected defect type, where the maximum feature value is selected for area defects, the minimum feature value is selected for spacing defects, and specific feature parameters are selected for comparison for deformation defects; and the final identification of defect types is completed simultaneously based on the mapping relationship between defect type and level, combined with predefined defect judgment thresholds.

[0015] Specifically, the data closed-loop feedback mechanism includes: A data interface is established between the diagnostic platform and the annotation platform. The ultrasound images that have completed intelligent diagnosis, along with their corresponding defect types, confidence levels, and characteristic parameters, are combined into a structured data package. Based on preset data filtering rules, the diagnostic results are automatically classified and prioritized to select the diagnostic data that meets the requirements. The filtered data package is synchronized to the processing queue of the annotation platform through a secure transmission protocol and automatically associated with the corresponding original image project to achieve closed-loop transmission of diagnostic data to the annotation platform.

[0016] Specifically, the method for screening and evaluating diagnostic data is as follows: A data screening mechanism based on multi-dimensional quality assessment was established. A dynamic threshold was set based on diagnostic confidence to screen key samples with confidence levels below a preset range. Simultaneously, a defect criticality assessment was combined to select complex cases containing high-level defects and multiple coexisting defects. Ultrasound image clarity was assessed to exclude image data with substandard signal quality. A sample representativeness assessment system was established, screening data defect categories based on the principle of balanced defect type distribution. The screening results were weighted and evaluated using a comprehensive data quality scoring model to form a dataset that meets the annotation requirements.

[0017] Specifically, the expert knowledge distillation mechanism includes: An expert review and knowledge transformation process is established, in which domain experts review and verify the intelligent diagnostic results through an annotation platform; based on ultrasound image features and detection standards, the domain experts correct and supplement the intelligent annotation results to form standardized annotation samples; by recording the expert annotation decision-making process and the basis for correction, an expert knowledge base is constructed; the difference between the expert annotation results and the original intelligent diagnostic results is analyzed to extract key discriminative features and correction rules; finally, the extracted expert knowledge is transformed into annotation specifications and model optimization guidance parameters.

[0018] Specifically, the remote auxiliary diagnostic platform supporting multi-terminal access includes: Establish a multi-terminal compatible system based on a web architecture, supporting cross-platform access from desktop and mobile terminals; ensure consistency in interface layout and interaction logic across different terminals by synchronizing data and adapting functions through a unified interface service layer; adopt responsive design technology to automatically adjust the display layout of the visual interface according to the screen size of the terminal device; establish a remote collaborative diagnostic mechanism to support multiple experts to access the same testing item in real time through different terminals, and control the functional access scope and data operation permissions of each terminal based on a permission management system.

[0019] Specifically, the containerized deployment technology includes: A dual-model deployment architecture based on container images is constructed, encapsulating the semantic segmentation model and the optimized annotation model as independent container images; a unified orchestration technology is adopted for one-click collaborative deployment of the two models, and the resource allocation and service coordination of the two models are automatically completed through preset orchestration templates; a dual-model version management system is established to support independent version control and dependency management of the semantic segmentation model and the optimized annotation model; the independent running environment of the two models is ensured through a container isolation mechanism, and a communication interface between the models is established to achieve collaborative diagnosis.

[0020] Specifically, the integrated diagnostic environment that combines interactive visualization analysis, full-process project management, and remote collaborative diagnosis is implemented using the following method: A 3D visualization analysis engine is built to support multi-plane reconstruction of ultrasound images and stereo rendering of defect areas, providing tools for image enhancement, measurement annotation, and comparative analysis; a full-process project management system is established to manage the entire lifecycle from project creation, data acquisition, diagnostic analysis to report generation, supporting project progress tracking and quality monitoring; a remote collaborative diagnosis mechanism is developed to provide real-time audio and video communication, annotation synchronization, and diagnostic opinion sharing functions, supporting online consultations by multiple experts and collaborative confirmation of diagnostic conclusions.

[0021] The beneficial effects of this invention are as follows: (1) By constructing a collaborative mechanism of "pre-annotation + interactive manual annotation", most of the ultrasound image annotation work is automatically completed with the help of pre-trained models, reducing the workload of manual frame-by-frame drawing; at the same time, the interactive tools of the annotation platform support manual correction of defect boundaries, ensuring the accuracy of the initial annotation. Subsequently, through data closed-loop feedback and expert knowledge distillation, the diagnostic data is screened, evaluated and the annotation is corrected, and the expert experience is transformed into standardized training data, which not only solves the problems of time-consuming and subjective differences in traditional annotation, but also provides training samples for the defect diagnosis model, helping to improve the model performance.

[0022] (2) An improved semantic segmentation model is adopted, combined with multi-scale feature extraction, attention mechanism and other technologies to enhance the defect boundary recognition capability, effectively deal with the problems of ultrasonic image noise interference and defect feature ambiguity, and reduce the situation of "over-segmentation" and "under-segmentation"; by establishing a normalized judgment rule based on the physical quantity of defect features, the standard for defect level classification is unified, avoiding the subjective error caused by the traditional reliance on human experience. At the same time, the data closed-loop feedback mechanism enables the diagnostic model to continuously receive optimized training data, continuously iterate and adapt to the defect features under different welding scenarios, further reduce the probability of misjudgment and missed judgment of complex defects (such as multiple defects coexisting), and improve the credibility of the diagnostic results.

[0023] (3) Containerized deployment technology is adopted to realize one-click deployment of diagnostic models and annotation models without the need for complex configuration by professional personnel, which reduces the technical threshold for model implementation; at the same time, a remote auxiliary diagnostic platform that supports access from multiple terminals is built, integrating visualization analysis, full-process project management and remote collaborative diagnostic functions. This not only makes it convenient for testing personnel to intuitively view defect information and track project progress through different terminals, but also supports real-time consultation of multiple experts across regions, solving the problem of delay in handling difficult problems caused by the dispersion of testing points and the concentration of expert resources in industrial scenarios, shortening the overall testing and diagnostic cycle and improving the efficiency of welded joint quality control. Attached Figure Description

[0024] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0025] Figure 1This is a flowchart illustrating an ultrasonic phased array intelligent annotation and defect diagnosis method for welded joints according to the present invention. Figure 2 This is an architecture diagram of an ultrasonic phased array intelligent annotation and defect diagnosis method for welded joints according to the present invention. Detailed Implementation

[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0027] Please see Figure 1-2 An intelligent ultrasonic phased array annotation and defect diagnosis method for welded joints: S1: Construct a labeling platform, acquire the original phased array ultrasonic images of the welded joint, establish an intelligent pre-labeling and manual interactive labeling mechanism to initially label the images, generate a labeled training dataset, and simultaneously train and test the initial labeling model for defect identification. S2: Based on the improved U-Net3 semantic segmentation model, the image is preprocessed and intelligently segmented to extract defect features. Based on predefined judgment rules, the defect level is divided and intelligent diagnosis results are output. At the same time, the intelligent diagnosis results are judged to identify the defect type. S3: Establish a data closed-loop feedback mechanism to feed the intelligent diagnostic results back to the annotation platform, screen and evaluate the diagnostic data based on diagnostic confidence and defect criticality, and use the expert knowledge distillation mechanism to annotate, correct and transform the diagnostic results to form optimized standard training data. S4: Establish a remote auxiliary diagnostic platform that supports multi-terminal access, adopt containerized deployment technology for one-click deployment and version management of semantic segmentation models and optimized annotation models, and build a comprehensive diagnostic environment that integrates interactive visual analysis, full-process project management and remote collaborative diagnosis.

[0028] Specifically, the method for initially labeling images by combining intelligent pre-annotation and manual interactive annotation is as follows: The uploaded original phased array ultrasound image is automatically annotated by loading a pre-trained initial semantic segmentation model, generating initial annotation results containing defect regions. When the automatic annotation results do not meet the annotation requirements, a manual interactive annotation process is initiated, where the defect boundaries are manually drawn and corrected using the interactive tools provided by the annotation platform. During the annotation process, image panning operations are performed using key combinations to view different regions, and the annotation results can be undone and redone using function keys. Finally, through the coordinated efforts of intelligent pre-annotation and manual interactive annotation, the defect regions in the phased array ultrasound image are accurately marked.

[0029] Specifically, the method for training and testing the initial labeling model for defect identification is as follows: An initial labeling model based on an encoder-decoder structure is constructed. During the training phase, a multi-stage optimization strategy is adopted. Basic training is performed by setting a training period and batch sample size, and a dynamic learning rate adjustment strategy is used simultaneously to optimize the convergence effect. During the testing phase, a dual validation system is established. On the one hand, the segmentation effect is evaluated based on the region overlap, and on the other hand, the recognition accuracy is evaluated based on the confidence level. Finally, the stability of the model in practical applications is determined by cross-validation.

[0030] Specifically, the image preprocessing and intelligent segmentation method is as follows: A multi-level preprocessing process is adopted. Based on the target detection network, the B-type atlas region in the ultrasound image is located and adaptively cropped. Then, color space conversion and signal intensity analysis are used to achieve preliminary screening of electrofusion welding defects. In the intelligent segmentation stage, a semantic segmentation model with fused attention mechanism is adopted. Multi-scale feature extraction and feature pyramid fusion technology are used to enhance the recognition accuracy of defect boundaries. The segmentation results are also processed and optimized by combining connected component analysis.

[0031] In this embodiment, during ultrasound image preprocessing and intelligent segmentation, the original ultrasound image of the electrofusion joint is first acquired using a phased array detection device. The device parameters include probe frequency, sampling frequency, scanning angle range, and gain value. The original image contains multiple functional areas such as the instrument interface, parameter display area, and B-mode ultrasound atlas, from which the effective ultrasound atlas region needs to be accurately extracted.

[0032] An improved object detection network is employed to automatically locate B-type atlas regions. This network uses depthwise separable convolutions to reduce the number of parameters during feature extraction and introduces a weighted feature pyramid network in the multi-scale feature fusion stage. Learnable weight coefficients are assigned to feature maps at different scales to optimize feature representation. The network outputs normalized coordinates and confidence scores for the B-type atlas regions. Based on these coordinates, the actual cropping region is calculated, and the cropped image is uniformly scaled to a standard size.

[0033] After image cropping, the image is converted from the RGB color space to the LAB color space using a conversion formula based on the CIE color space standard. Signal intensity analysis is then performed on the image in the LAB space. First, a color scale region is extracted, and a continuous mapping function from color values ​​to signal intensity is established using an interpolation algorithm. The average signal intensity I_avg of the resistance wire region is calculated. When I_avg is lower than a threshold proportion of the reference intensity, it is determined to be a case of poor electrofusion welding or poor coupling.

[0034] In the intelligent segmentation stage, an improved semantic segmentation network is employed. The encoder consists of multiple convolutional blocks, each containing a convolutional layer, a normalization layer, and an activation function. An attention mechanism module is introduced at specific layers to enhance the response of important features by reweighting feature channels. The decoder gradually restores spatial resolution through upsampling operations and establishes skip connections with the feature maps of corresponding layers in the encoder, effectively fusing low-level detailed features with high-level semantic features. Network training uses a combined loss function L_total = αL_seg + βL_aux, where L_seg is the primary segmentation loss and L_aux is the auxiliary loss. Multi-task learning is used to improve model performance.

[0035] Post-processing optimization of the segmentation results involves several steps. First, connected component analysis is performed to label all independent regions and calculate their area, centroid, and other features. Noisy regions are removed based on an area threshold, and adjacent regions are merged based on a spatial distance threshold. Next, morphological operations are performed: circular structuring elements are used to first dilate and fill holes, followed by erosion to smooth the boundaries. In the edge signal removal stage, the signal intensity of image boundary regions is detected. Regions with signal intensity lower than the overall average intensity are considered invalid and cropped. Finally, scale analysis is performed to identify scale marks in the image, calculate the conversion ratio between pixel size and physical size, and establish a physical coordinate system for the measurement results.

[0036] Through the complete processing flow described above, the system transforms raw ultrasound images into precise segmentation results. Key parameters at each processing stage are optimized and adjusted according to specific application scenarios to ensure the system's stability and accuracy under different detection conditions. The entire process forms a complete pipeline operation, providing a reliable data foundation for subsequent quantitative defect analysis.

[0037] Specifically, the predefined judgment rules for classifying defect levels are as follows: A multi-defect collaborative judgment mechanism is established to normalize the identified defect features and convert the physical quantities of defect features into standardized defect severity values. A hierarchical evaluation system is established based on defect types. For different types of defects, corresponding feature parameter extraction and threshold comparison methods are adopted. Based on the quantitative evaluation results of defect types, defect levels are classified in combination with predefined level thresholds.

[0038] The graded evaluation system based on defect type includes: for cold welding and over-welding defects, the grade is determined according to the degree of deviation between the characteristic line and the resistance wire spacing from the standard value; for hole defects, the grade is comprehensively evaluated based on the proportion of the missing part to the length of the fusion zone and the ratio of the hole height to the pipe wall thickness; for inclusion defects, different length threshold standards are used according to the degree of penetration with the inner cold welding zone; for socket misalignment defects, the grade is determined based on the length of the resistance wire at the unwelded pipe; and for resistance wire misalignment defects, the grade is determined based on the ratio of the misalignment amount to the resistance wire spacing.

[0039] In this embodiment, a gas company conducts quality inspection on a newly built PE pipeline project and uses an ultrasonic phased array device to scan a DN200 electrofusion joint.

[0040] First, the extracted physical quantities of defect features are normalized, converting them into uniform defect severity values ​​(ranging from 0 to 1). This process includes: Normalization of the distance between the characteristic line and the resistance wire: Let the standard distance be L0 (determined according to the pipe specifications, usually 2.5-3.5mm), and the measured distance be L, then the normalized distance parameter is... Normalization of bottom echo missing length: Let the nominal fusion zone length be L_f (usually 45mm) and the measured missing length be X, then the normalized missing parameter η = X / L_f.

[0041] Normalization of resistance wire misalignment: Let the normal resistance wire spacing be D (usually 2.0-3.0mm), and the measured misalignment be Δd, then the normalized misalignment parameter ε = Δd / D.

[0042] The following defect level determination formula system will be established simultaneously: (1) Determination of cold welding / over-welding defects Degree of cold welding: H = (1 - L′ / L) × 100%, Degree of over-soldering: H′=(L′ / L-1)×100%, Where L is the standard value of the feature line spacing in normal welding, and L′ is the measured value.

[0043] Preset grading standards: Level I (minor defects): H < 10%, the deviation in feature line spacing is not obvious, and the welding quality is acceptable; Level II (General Defects): 10% ≤ H < 30%, the feature line spacing deviates significantly, and welding quality needs to be monitored; Level III (Severe Defect): H≥30%, feature line spacing is seriously deviated, and welding quality is unqualified.

[0044] The grading criteria for over-welding defects are as follows: Level I: H′<20%; Level II: 20% ≤ H′ < 40%; Level III: H′≥40%; The defect level of holes is determined based on the proportion of missing parts, using a two-dimensional determination model: Single hole: Level LV=f(X / L,h / T), where X is the axial dimension of the defect, L is the nominal fusion zone length, h is the hole height, and T is the pipe wall thickness.

[0045] The specific grading standards are as follows: Level I: X / L < 5% and h / T < 5%, small hole size, minimal impact on structure; Level II: X / L < 10% and h / T < 10%, with medium-sized holes, structural safety needs to be assessed; Level III: Exceeds the Level II range; the hole size is too large, seriously affecting the structural integrity.

[0046] For combined holes, cumulative dimensions are used for evaluation: Level I: Cumulative dimension X / L < 10% and h / T < 5%; Level II: Cumulative dimension X / L < 15% and h / T < 10%; Level III: Exceeds the scope of Level II.

[0047] (4) Determination of defects in improper socket fitting Classification based on the length of the resistance wire at the unwelded section of the pipe: Grade I: Length is 0, socket position is correct; Grade II: Length ≤10mm, slight deviation in socket; Grade III: Length > 10mm, socket is seriously inadequate.

[0048] (5) Determination of resistance wire misalignment defects Misalignment degree: ε=Δd / D, where Δd is the misalignment amount and D is the normal resistance wire spacing.

[0049] The grading criteria are as follows: Grade I: No obvious misalignment (ε<5%), resistance wires are basically neatly arranged; Level II: ε<1 (misalignment is less than the distance between resistance wires), resistance wires are not aligned but are not in contact; Level III: ε≥1 or adjacent resistance wires are in contact with each other, and the resistance wires are arranged in a severely disordered manner.

[0050] Specifically, the method for determining the intelligent diagnostic results and identifying the defect type is as follows: An intelligent defect type discrimination mechanism is established, which performs multi-image defect level comparison analysis to obtain the highest defect level among all images of the same joint; discrimination is performed based on defect type priority rules, and when non-spacing defects exist, the non-spacing defects are selected as the primary defect type; feature value analysis is performed on the selected defect type, where the maximum feature value is selected for area defects, the minimum feature value is selected for spacing defects, and specific feature parameters are selected for comparison for deformation defects; and the final identification of defect types is completed simultaneously based on the mapping relationship between defect type and level, combined with predefined defect judgment thresholds.

[0051] In this embodiment, during a gas pipeline engineering inspection project, the system performs intelligent diagnosis on an electrofusion joint numbered WK-20231025-001. Six phased array ultrasonic images of this joint were acquired from different angles. The system then activates its intelligent defect type identification mechanism and begins the following judgment process: First, the system performs parallel analysis on six images, identifying defect features in each image. During image preprocessing, the system extracts key regions such as feature lines, resistance wires, and bottom echoes from each image using an improved semantic segmentation model, and calculates corresponding feature parameters. These parameters include quantitative indicators such as the distance between the feature lines and the resistance wires, the length of missing bottom echoes, the amount of resistance wire misalignment, and the size of the holes.

[0052] After feature extraction, the system performs a multi-image defect level comparison analysis. Taking images 1 and 2 as examples: Image 1 detects a cold welding defect, with the feature line spacing deviating from the standard value by 15%, and is rated as Level II; Image 2 finds a hole defect, with the missing part accounting for 8% of the fusion zone length, and is rated as Level II; Image 3 shows resistance wire misalignment, with the misalignment reaching 60% of the normal spacing, and is rated as Level II; the other images are all Level I or normal. The system automatically selects the highest level, Level II, as the defect level of the joint.

[0053] Next, the system makes a judgment based on the defect type priority rules. According to the preset rules, defect types are divided into three priorities: holes, inclusions, and improper socketing belong to the first priority (non-spacing defects); cold welding and over-welding belong to the second priority (spacing defects); and resistance wire misalignment belongs to the third priority (deformation defects). In this case, the system detected the presence of a hole defect (first priority), so it prioritizes the hole as the primary defect type.

[0054] After determining the main defect type, the system performs feature value analysis. For area-type defects such as holes, the system automatically selects the largest defect feature value from the six images as the evaluation criterion. The specific calculation process is as follows: The system identifies that in image 2, the axial dimension of the hole X = 4.5 mm, the fusion zone length L = 45 mm, and the missing proportion X / L = 10%; the hole height h = 1.2 mm, the pipe wall thickness T = 20 mm, and the height proportion h / T = 6%. According to the hole defect classification standard, the system classifies this defect as Level II.

[0055] In the final identification phase, the system completes the confirmation based on predefined defect judgment thresholds. For hole defects, the judgment rules invoked by the system include: Level I threshold for a single hole is X / L < 5% and h / T < 5%; Level II threshold is X / L < 10% and h / T < 10%; defects exceeding the Level II range are assessed as Level III. Simultaneously, the system also checks for the existence of combined holes to ensure the accuracy of the assessment results.

[0056] Throughout the entire assessment process, the system records the analysis results and decision-making basis for each step in real time, forming a complete assessment log. When boundary conditions occur (such as when the defect level approaches the threshold), the system automatically marks them and prompts for manual review. Finally, the system generates a structured diagnostic report, clearly indicating that the joint has a Class II hole defect, and listing the defect analysis data of each image in detail, providing a reliable basis for subsequent engineering processing.

[0057] Specifically, the data closed-loop feedback mechanism includes: A data interface is established between the diagnostic platform and the annotation platform. The ultrasound images that have completed intelligent diagnosis, along with their corresponding defect types, confidence levels, and characteristic parameters, are combined into a structured data package. Based on preset data filtering rules, the diagnostic results are automatically classified and prioritized to select the diagnostic data that meets the requirements. The filtered data package is synchronized to the processing queue of the annotation platform through a secure transmission protocol and automatically associated with the corresponding original image project to achieve closed-loop transmission of diagnostic data to the annotation platform.

[0058] Specifically, the method for screening and evaluating diagnostic data is as follows: A data screening mechanism based on multi-dimensional quality assessment was established. A dynamic threshold was set based on diagnostic confidence to screen key samples with confidence levels below a preset range. Simultaneously, a defect criticality assessment was combined to select complex cases containing high-level defects and multiple coexisting defects. Ultrasound image clarity was assessed to exclude image data with substandard signal quality. A sample representativeness assessment system was established, screening data defect categories based on the principle of balanced defect type distribution. The screening results were weighted and evaluated using a comprehensive data quality scoring model to form a dataset that meets the annotation requirements.

[0059] Specifically, the expert knowledge distillation mechanism includes: An expert review and knowledge transformation process is established, in which domain experts review and verify the intelligent diagnostic results through an annotation platform; based on ultrasound image features and detection standards, the domain experts correct and supplement the intelligent annotation results to form standardized annotation samples; by recording the expert annotation decision-making process and the basis for correction, an expert knowledge base is constructed; the difference between the expert annotation results and the original intelligent diagnostic results is analyzed to extract key discriminative features and correction rules; finally, the extracted expert knowledge is transformed into annotation specifications and model optimization guidance parameters.

[0060] Specifically, the remote auxiliary diagnostic platform supporting multi-terminal access includes: Establish a multi-terminal compatible system based on a web architecture, supporting cross-platform access from desktop and mobile terminals; ensure consistency in interface layout and interaction logic across different terminals by synchronizing data and adapting functions through a unified interface service layer; adopt responsive design technology to automatically adjust the display layout of the visual interface according to the screen size of the terminal device; establish a remote collaborative diagnostic mechanism to support multiple experts to access the same testing item in real time through different terminals, and control the functional access scope and data operation permissions of each terminal based on a permission management system.

[0061] Specifically, the containerized deployment technology includes: A dual-model deployment architecture based on container images is constructed, encapsulating the semantic segmentation model and the optimized annotation model as independent container images; a unified orchestration technology is adopted for one-click collaborative deployment of the two models, and the resource allocation and service coordination of the two models are automatically completed through preset orchestration templates; a dual-model version management system is established to support independent version control and dependency management of the semantic segmentation model and the optimized annotation model; the independent running environment of the two models is ensured through a container isolation mechanism, and a communication interface between the models is established to achieve collaborative diagnosis.

[0062] Specifically, the integrated diagnostic environment that combines interactive visualization analysis, full-process project management, and remote collaborative diagnosis is implemented using the following method: A 3D visualization analysis engine is built to support multi-plane reconstruction of ultrasound images and stereo rendering of defect areas, providing tools for image enhancement, measurement annotation, and comparative analysis; a full-process project management system is established to manage the entire lifecycle from project creation, data acquisition, diagnostic analysis to report generation, supporting project progress tracking and quality monitoring; a remote collaborative diagnosis mechanism is developed to provide real-time audio and video communication, annotation synchronization, and diagnostic opinion sharing functions, supporting online consultations by multiple experts and collaborative confirmation of diagnostic conclusions.

[0063] In this embodiment, the comprehensive diagnostic environment is constructed based on a modular architecture design. The 3D visualization analysis engine uses the WebGL 2.0 graphics interface and achieves stereoscopic visualization of ultrasound images through a multi-plane reconstruction algorithm. The system first performs spatial registration on the input 2D ultrasound image sequence to establish a 3D volumetric data model. In the rendering stage, physically based rendering technology is used to highlight defect areas by adjusting the transfer function. Specific operations include: adjusting image contrast using window width and window level, strengthening defect boundaries through edge enhancement algorithms, and using multi-plane reconstruction technology to achieve cross-sectional viewing from any angle. The measurement and annotation tool integrates an intelligent ruler function, which can automatically identify scale information in the image and achieve automatic conversion from pixel distance to physical size. The comparison analysis tool supports synchronous comparison of images at multiple time points, and accurately displays the evolution process of defects through image registration and difference detection algorithms.

[0064] The full-process project management system adopts a microservice architecture. The project creation module uses standardized templates to achieve structured input of project information, including basic project information, testing parameters, and quality requirements. The data acquisition module establishes real-time data connections with on-site testing equipment, supporting automatic parsing and metadata extraction of various image formats such as DICOM, PNG, and JPEG. The system drives the testing process through a workflow engine, setting quality checkpoints at each stage to achieve full-process tracking from task allocation, data acquisition, intelligent diagnosis to report generation. For quality monitoring, the system establishes a multi-dimensional quality indicator system, including image quality scoring, diagnostic consistency assessment, and report integrity checks, displaying project progress and quality status in real time through data dashboards. The report generation module, based on a configurable template engine, automatically extracts diagnostic results and key images to generate testing reports that conform to industry standards.

[0065] The remote collaborative diagnosis mechanism is built on a real-time communication platform based on WebRTC technology. The system establishes a distributed audio and video communication network, achieving low-latency audio and video transmission among multiple experts through an SFU architecture. During collaborative diagnosis, the system provides a shared whiteboard function, supporting real-time annotation by multiple parties. The annotation synchronization mechanism employs an operation conversion algorithm to ensure that annotation operations on the same image by multiple users are synchronized in real time and maintain consistency. The diagnostic opinion sharing module provides a structured opinion input interface, supporting the rapid input of standardized content such as defect marking, grade assessment, and treatment suggestions. The online consultation module establishes a case discussion room, supporting real-time voice discussions among experts and collaborative confirmation of diagnostic conclusions. The system also provides a permission management mechanism, assigning different operation permissions such as viewing, annotation, and review based on different roles to ensure the security and controllability of the diagnostic process.

[0066] In practical applications, inspectors first create an inspection project in the project management system, entering basic project information and configuring inspection parameters. During on-site inspection, the system automatically receives ultrasonic image data transmitted from the inspection equipment and triggers an intelligent diagnostic process. During the diagnosis, engineers can view defect morphology from multiple angles using 3D visualization tools and accurately quantify defect dimensions using measurement tools. When encountering difficult cases, a remote consultation function can be activated, inviting multiple experts to collaborate online. Experts exchange technical information through real-time audio and video, collaboratively annotate shared images, and ultimately reach a consistent diagnostic conclusion. The system automatically records the complete diagnostic process, including raw data, intelligent diagnostic results, expert discussion records, and the final diagnostic conclusion, providing a complete basis for subsequent quality traceability.

[0067] This implementation method, through the organic integration of three core modules, constructs a complete intelligent diagnostic working environment, realizes full-process digital management from data collection to diagnostic conclusions, and improves diagnostic efficiency and accuracy.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent ultrasonic phased array annotation and defect diagnosis of welded joints, characterized in that, include: S1: Construct a labeling platform, acquire the original phased array ultrasonic images of the welded joint, establish an intelligent pre-labeling and manual interactive labeling mechanism to initially label the images, generate a labeled training dataset, and simultaneously train and test the initial labeling model for defect identification. S2: Based on the improved U-Net3 semantic segmentation model, the image is preprocessed and intelligently segmented to extract defect features. Based on predefined judgment rules, the defect level is divided and intelligent diagnosis results are output. At the same time, the intelligent diagnosis results are judged to identify the defect type. S3: Establish a data closed-loop feedback mechanism to feed the intelligent diagnostic results back to the annotation platform, screen and evaluate the diagnostic data based on diagnostic confidence and defect criticality, and use the expert knowledge distillation mechanism to annotate, correct and transform the diagnostic results to form optimized standard training data. S4: Establish a remote auxiliary diagnostic platform that supports multi-terminal access, adopt containerized deployment technology for one-click deployment and version management of semantic segmentation models and optimized annotation models, and build a comprehensive diagnostic environment that integrates interactive visual analysis, full-process project management and remote collaborative diagnosis.

2. The method according to claim 1, characterized in that, The method for initial labeling of images by combining intelligent pre-labeling and manual interactive labeling is as follows: The uploaded original phased array ultrasound image is automatically annotated by loading a pre-trained initial semantic segmentation model to generate initial annotation results containing defect regions; when the automatic annotation results do not meet the annotation requirements, the manual interactive annotation process is initiated, and the defect boundaries are manually drawn and corrected using the interactive tools provided by the annotation platform. During the annotation process, image panning operations are performed using key combinations to view different areas, and annotation results can be undone or redone using function keys. Finally, through the coordinated efforts of intelligent pre-annotation and manual interactive annotation, the marking of defect areas in phased array ultrasound images is completed.

3. The method according to claim 1, characterized in that, The specific method for training and testing the initial labeling model for defect identification is as follows: An initial labeling model based on an encoder-decoder structure is constructed. A multi-stage optimization strategy is adopted during the training phase. Basic training is carried out by setting a preset training period and batch sample size, and a dynamic learning rate adjustment strategy is adopted simultaneously to optimize the convergence effect. During the testing phase, a dual validation system is established. On the one hand, the segmentation effect is evaluated based on the regional overlap, and on the other hand, the recognition accuracy is evaluated based on the confidence level. Finally, the stability of the model in practical applications is determined through cross-validation.

4. The method according to claim 1, characterized in that, The specific method for image preprocessing and intelligent segmentation is as follows: A multi-level preprocessing process is adopted. Based on the target detection network, the B-type atlas region in the ultrasound image is located and adaptively cropped. Then, color space conversion and signal intensity analysis are used to achieve preliminary screening of electrofusion welding defects. In the intelligent segmentation stage, a semantic segmentation model with fused attention mechanism is adopted. Multi-scale feature extraction and feature pyramid fusion technology are used to enhance the recognition accuracy of defect boundaries. The segmentation results are also processed and optimized by combining connected component analysis.

5. The method according to claim 1, characterized in that, The predefined judgment rules classify defect levels, and the specific method is as follows: A multi-defect collaborative judgment mechanism is established to normalize the identified defect features and convert the physical quantities of defect features into standardized defect severity values. A hierarchical evaluation system is established based on defect types. For different types of defects, corresponding feature parameter extraction and threshold comparison methods are adopted. Based on the quantitative evaluation results of defect types, defect levels are classified in combination with predefined level thresholds.

6. The method according to claim 1, characterized in that, The specific method for determining the intelligent diagnostic results and identifying the defect type is as follows: A defect type intelligent discrimination mechanism is established to perform multi-image defect level comparison analysis and obtain the highest defect level among all images of the same joint; discrimination is performed based on defect type priority rules, and when there are non-spacing defects, the non-spacing defects are selected as the main defect type; feature value analysis is performed on the selected defect type, and the final defect type is identified based on the mapping relationship between defect type and level and the predefined defect judgment threshold.

7. The method according to claim 1, characterized in that, The data closed-loop feedback mechanism includes: A data interface is established between the diagnostic platform and the annotation platform. The ultrasound images that have completed intelligent diagnosis, along with their corresponding defect types, confidence levels, and characteristic parameters, are combined into a structured data package. Based on preset data filtering rules, the diagnostic results are automatically classified and prioritized to select the diagnostic data that meets the requirements. The filtered data package is synchronized to the processing queue of the annotation platform through a secure transmission protocol and automatically associated with the corresponding original image project to achieve closed-loop transmission of diagnostic data to the annotation platform.

8. The method according to claim 1, characterized in that, The specific method for screening and evaluating diagnostic data is as follows: A data screening mechanism based on multi-dimensional quality assessment was established. A dynamic threshold was set based on diagnostic confidence to screen key samples with confidence levels below a preset range. Simultaneously, a defect criticality assessment was combined to select complex cases containing high-level defects and multiple coexisting defects. Ultrasound image clarity was assessed to exclude image data with substandard signal quality. A sample representativeness assessment system was established, screening data defect categories based on the principle of balanced defect type distribution. The screening results were weighted and evaluated using a comprehensive data quality scoring model to form a dataset that meets the annotation requirements.

9. The method according to claim 1, characterized in that, The expert knowledge distillation mechanism includes: An expert review and knowledge transformation process is established, in which domain experts review and verify the intelligent diagnostic results through an annotation platform; based on ultrasound image features and detection standards, the domain experts correct and supplement the intelligent annotation results to form standardized annotation samples; by recording the expert annotation decision-making process and the basis for correction, an expert knowledge base is constructed; the difference between the expert annotation results and the original intelligent diagnostic results is analyzed to extract key discriminative features and correction rules; finally, the extracted expert knowledge is transformed into annotation specifications and model optimization guidance parameters.

10. The method according to claim 1, characterized in that, The remote auxiliary diagnostic platform supporting multi-terminal access includes: Establish a multi-terminal compatible system based on a web architecture, supporting cross-platform access from desktop and mobile terminals; ensure consistency in interface layout and interaction logic across different terminals by synchronizing data and adapting functions through a unified interface service layer; adopt responsive design technology to automatically adjust the display layout of the visual interface according to the screen size of the terminal device; establish a remote collaborative diagnostic mechanism to support multiple experts to access the same testing item in real time through different terminals, and control the functional access scope and data operation permissions of each terminal based on a permission management system.

11. The method according to claim 1, characterized in that, The containerized deployment technology includes: A dual-model deployment architecture based on container images is constructed, encapsulating the semantic segmentation model and the optimized annotation model as independent container images; a unified orchestration technology is adopted for one-click collaborative deployment of the two models, and the resource allocation and service coordination of the two models are automatically completed through preset orchestration templates; a dual-model version management system is established to support independent version control and dependency management of the semantic segmentation model and the optimized annotation model; the independent running environment of the two models is ensured through a container isolation mechanism, and a communication interface between the models is established to achieve collaborative diagnosis.

12. The method according to claim 1, characterized in that, The integrated diagnostic environment that combines interactive visualization analysis, full-process project management, and remote collaborative diagnosis is described in the following specific method: A 3D visualization analysis engine is built to support multi-plane reconstruction of ultrasound images and stereo rendering of defect areas, providing image enhancement, measurement annotation, and comparative analysis tools; a full-process project management system is established to manage the entire lifecycle from project creation, data acquisition, diagnostic analysis to report generation, supporting project progress tracking and quality monitoring. Develop a remote collaborative diagnosis mechanism that provides real-time audio and video communication, annotation synchronization, and diagnostic opinion sharing functions, and supports online consultations by multiple experts and collaborative confirmation of diagnostic conclusions.

Images