CN-121981980-A - Ultrasonic phased array intelligent marking and defect diagnosis method for welding joint

CN121981980ACN 121981980 ACN121981980 ACN 121981980ACN-121981980-A

Abstract

The invention relates to an ultrasonic phased array intelligent marking and defect diagnosis method for a welding joint, and belongs to the technical field of nondestructive testing and artificial intelligent intersection. The method comprises the steps of constructing a labeling platform to complete intelligent pre-labeling and manual interactive labeling of an ultrasonic image, training an initial labeling model, realizing defect feature extraction and grading based on an improved semantic segmentation model, establishing a data closed-loop feedback mechanism, optimizing labeling data through expert knowledge distillation, constructing a remote auxiliary diagnosis platform, realizing model one-key deployment and version management by adopting a containerized deployment technology, and integrating interactive visual analysis, full-flow project management and remote collaborative diagnosis functions. Closed loop optimization of marking and diagnosis is realized, and accuracy and diagnosis efficiency of defect identification are improved.

Inventors

Zhou Bingbo
YAN BING
HE ZONGHUI
Zeng yuhuan
ZHANG JUN

Assignees

上海派普诺管道检测科技发展有限公司

Dates

Publication Date: 20260505
Application Date: 20260116

Claims (12)

1. An ultrasonic phased array intelligent marking and defect diagnosis method for a welding joint is characterized by comprising the following steps: S1, constructing a marking platform, acquiring an original phased array ultrasonic image of a welding joint, building an intelligent pre-marking and manual interaction marking mechanism to initially mark the image, generating a training data set with a label, synchronously training and testing an initial marking model for defect identification; S2, preprocessing and intelligently segmenting the image based on an improved U-Net 3-semantic segmentation model, extracting defect characteristics, classifying defect levels based on a predefined judgment rule, outputting an intelligent diagnosis result, judging the intelligent diagnosis result and identifying the defect type; S3, establishing a data closed-loop feedback mechanism, feeding the intelligent diagnosis result back to the labeling platform, screening and evaluating diagnosis data based on diagnosis confidence and defect criticality, and labeling, correcting and converting knowledge of the diagnosis result through an expert knowledge distillation mechanism to form optimized standard training data; And S4, establishing a remote auxiliary diagnosis platform supporting multi-terminal access, performing one-key deployment and version management of a semantic segmentation model and an optimization labeling model by adopting a containerized deployment technology, and constructing a comprehensive diagnosis environment integrating interactive visual analysis, full-flow project management and remote collaborative diagnosis.
2. The method according to claim 1, wherein the image is initially marked by combining intelligent pre-marking and manual interactive marking mechanisms, and the specific method is as follows: The method comprises the steps of loading a pre-trained initial semantic segmentation model to automatically label an uploaded original phased array ultrasonic image to generate an initial labeling result containing a defect area, starting a manual interaction labeling process when the automatic labeling result does not meet the labeling requirement, manually drawing and correcting the defect boundary through an interaction tool provided by a labeling platform, carrying out translation operation of the image through a combination key to check different areas in the labeling process, carrying out undo and redo operation of the labeling result through a function key, and finally completing labeling of the defect area in the phased array ultrasonic image through cooperative cooperation of the intelligent pre-labeling and the manual interaction labeling.
3. The method according to claim 1, wherein the training and testing of the initial marking model for defect identification is performed by: The method comprises the steps of constructing an initial marking model based on an encoder-decoder structure, adopting a multi-stage optimization strategy in a training stage, carrying out basic training through a preset training period and batch processing sample size, synchronously adopting a dynamic learning rate adjustment strategy to optimize convergence effects, and establishing a double verification system in a testing stage, wherein on one hand, the segmentation effect is evaluated based on region overlapping degree, on the other hand, the recognition accuracy is evaluated based on confidence degree, and finally, determining the stability of the model in practical application through cross verification.
4. The method according to claim 1, wherein the preprocessing and intelligent segmentation of the image is performed by: the method comprises the steps of adopting a multistage preprocessing flow, positioning a B-type map region in an ultrasonic image based on a target detection network, performing self-adaptive cutting, then realizing preliminary screening of electric fusion welding failure through color space conversion and signal intensity analysis, adopting a semantic segmentation model fusing a attention mechanism in an intelligent segmentation stage, enhancing the identification precision of a defect boundary through a multi-scale feature extraction and feature pyramid fusion technology, and performing processing optimization on a segmentation result by combining connected domain analysis.
5. The method according to claim 1, wherein the predefined decision rule classifies the defect level by: Establishing a multi-defect collaborative judging mechanism, carrying out normalization processing on the identified defect characteristics, converting the defect characteristic physical quantity into a standardized defect degree value, establishing a grading evaluation system based on defect types, adopting a corresponding characteristic parameter extraction and threshold comparison method aiming at different types of defects, and carrying out defect grading based on the quantitative evaluation result of the defect types and combining a predefined grade threshold.
6. The method according to claim 1, wherein the intelligent diagnosis result determination is performed to identify the defect type, and the specific method is as follows: The method comprises the steps of establishing an intelligent defect type judging mechanism, executing multi-image defect grade comparison analysis, obtaining the highest defect grade in all images of the same joint, judging based on a defect type priority rule, preferentially selecting non-interval type defects as main defect types when the non-interval type defects exist, executing characteristic value analysis on the selected defect types, synchronously carrying out final identification of the defect types based on the mapping relation between the defect types and the grades, and combining a predefined defect judging threshold value.
7. The method of claim 1, wherein the data closed loop feedback mechanism comprises: The method comprises the steps of establishing a data interface between a diagnosis platform and a labeling platform, forming an ultrasonic image with intelligent diagnosis and corresponding defect types, confidence and characteristic parameters into a structured data packet, automatically classifying and prioritizing diagnosis results based on preset data screening rules, screening out diagnosis data meeting requirements, synchronizing the screened data packet to a queue to be processed of the labeling platform through a safe transmission protocol, automatically associating the data packet to a corresponding original image item, and carrying out closed-loop transmission of the diagnosis data to the labeling platform.
8. The method according to claim 1, wherein the screening and evaluation of the diagnostic data is performed by: Establishing a data screening mechanism based on multi-dimensional quality assessment, setting a dynamic threshold based on diagnosis confidence, screening out key samples with the confidence lower than a preset range, simultaneously combining defect criticality assessment, selecting complex cases containing high-level defects and multi-defect coexistence, carrying out definition assessment on an ultrasonic image, eliminating image data with unqualified signal quality, establishing a sample representative assessment system, screening out data defect types based on defect type distribution balance principle, and carrying out weighted assessment on screening results through a data quality comprehensive scoring model to form a data set meeting labeling requirements.
9. The method of claim 1, wherein the expert knowledge distillation mechanism comprises: Establishing an expert review and knowledge conversion flow, carrying out review verification on an intelligent diagnosis result by a field expert through a labeling platform, correcting and supplementing the intelligent labeling result based on ultrasonic image characteristics and detection standards by the field expert to form a standardized labeling sample, constructing an expert knowledge base by recording expert labeling decision process and correction basis, carrying out difference analysis on the expert labeling result and an original intelligent diagnosis result, extracting key discrimination characteristics and correction rules, and finally converting the extracted expert knowledge into labeling standards and model optimization guiding parameters.
10. The method of claim 1, wherein the remote assisted diagnostic platform supporting multi-terminal access comprises: The method comprises the steps of establishing a multi-terminal compatible system based on a Web architecture, supporting cross-platform access of desktop terminals and mobile terminals, carrying out data synchronization and function adaptation of different terminals through a unified interface service layer, ensuring consistency of interface layout and interaction logic of the terminals, adopting a responsive design technology to automatically adjust display layout of a visual interface according to screen size of terminal equipment, establishing a remote collaborative diagnosis mechanism, supporting multiple experts to access the same detection item through different terminals in real time, and controlling function access range and data operation authority of the terminals based on an authority management system.
11. The method of claim 1, wherein the containerized deployment technique comprises: The method comprises the steps of constructing a dual-model deployment architecture based on a container mirror image, respectively packaging a semantic segmentation model and an optimization labeling model into independent container mirror images, adopting a unified arrangement technology to conduct one-key collaborative deployment of the two models, automatically completing resource allocation and service coordination of the two models through a preset arrangement template, establishing a dual-model version management system, supporting independent version control and dependency management of the semantic segmentation model and the optimization labeling model, guaranteeing independent operation environments of the two models through a container isolation mechanism, and simultaneously establishing a communication interface between the models to realize collaborative diagnosis.
12. The method according to claim 1, wherein the integrated interactive visual analysis, full-flow project management and remote collaborative diagnosis comprehensive diagnosis environment comprises the following specific steps: the method comprises the steps of constructing a three-dimensional visual analysis engine, supporting multi-plane reconstruction of ultrasonic images and three-dimensional rendering of defect areas, providing image enhancement, measurement labeling and comparison analysis tools, establishing a full-flow project management system, carrying out full-period management from project creation, data acquisition and diagnosis analysis to report generation, supporting project progress tracking and quality monitoring, developing a remote collaborative diagnosis mechanism, providing real-time audio-video communication, label synchronization and diagnosis opinion sharing functions, and supporting online consultation and collaborative diagnosis conclusion confirmation of multiple experts.

Description

Ultrasonic phased array intelligent marking and defect diagnosis method for welding joint Technical Field The invention belongs to the technical field of nondestructive testing and artificial intelligence intersection, and particularly relates to an ultrasonic phased array intelligent marking and defect diagnosis method for a welding joint. Background In the field of industrial manufacturing, the quality of a welding joint is directly related to the safety and reliability of equipment operation, and an ultrasonic phased array detection technology becomes a core means of welding joint defect detection by virtue of flexible beam control capability and clear imaging effect. With the application of artificial intelligence technology in the field of nondestructive testing, intelligent diagnosis schemes based on algorithms such as semantic segmentation and the like are gradually raised, and the purpose of improving the efficiency and accuracy of defect identification is achieved. However, the current ultrasonic phased array intelligent marking and defect diagnosis technology for welding joints still has various limitations, and is difficult to meet the actual requirements of industrial scenes on high-precision and high-efficiency detection. In the labeling link, the prior art faces the dilemma of 'insufficient intelligent and manual cooperation'. The traditional marking mode is mostly dependent on manual drawing of defective areas frame by frame, is time-consuming and labor-consuming, is easily affected by experience of detection personnel and subjective judgment difference, and causes difficulty in guaranteeing consistency and accuracy of marking data. The intelligent pre-labeling scheme is partially introduced, and because the initial model has limited identification capability on complex defects (such as micro cracks and inclusions) in an ultrasonic image, the phenomena of label missing and label misplacement often occur, and a convenient interactive correction tool is lacked, so that a detection personnel is difficult to quickly adjust a labeling result, and finally, the construction efficiency of a high-quality training data set is low, so that the performance improvement of a follow-up diagnosis model is restricted. In the defect diagnosis link, the existing semantic segmentation model has insufficient adaptability to the ultrasonic image of the welding joint. The image has the problems of noise interference, fuzzy defect characteristics, low contrast between the background and the target and the like, the defect boundary is difficult to accurately extract by the traditional model, the situation of over-segmentation or under-segmentation is easy to occur, meanwhile, the defect grade division is dependent on fuzzy standards formulated by manual experience, the unified judging rule based on the physical quantity of the defect characteristics is lacking, the comparability of the diagnosis results under different scenes is lacking, the type identification and grade judging accuracy of part of complex defects (such as multi-defect coexistence scenes) are difficult to reach the standard, and reliable basis can not be provided for subsequent quality evaluation. The model iteration mechanism driven by data is missing, and the prior art presents a unidirectional flow of diagnosis, namely end point. Even if part of schemes are introduced into expert review, a knowledge distillation mechanism of a system is not established, and marking rules and discrimination logic corrected by an expert cannot be converted into guide parameters for model optimization, so that the diagnosis model is difficult to continuously iterate through actual data, the adaptability to joint defects of different materials and under different welding processes is poor, and the diagnosis precision is easy to decay after long-term use. Remote collaboration and deployment management are weak, and diagnosis efficiency is further restricted. In an industrial scene, detection points are widely distributed, expert resources are concentrated in a core laboratory, when on-site detection personnel encounter difficult problems, expert support is difficult to obtain quickly, so that problem disposal is delayed, meanwhile, the deployment process of the existing diagnosis model is complex, the environment configuration and parameter debugging are required by the professional personnel, and an access mechanism compatible with a cross-terminal is lacked, so that data synchronization and functional suitability between a desktop and a mobile terminal are poor. In addition, most schemes do not integrate full-flow project management and interactive visual analysis functions, unified management and visual presentation of detection data and diagnosis results are difficult to realize, and tracing and quality control of a detection process are not facilitated. In summary, the defects of the prior art in labeling cooperativity, model accuracy, data closed-loop performance