Table of Content
- Executive Summary
- Introduction and Problem Context
- Operational Impact and Implementation Value
- Data Sources and Preprocessing
- Technical Approach
- Test and Validation Approach
- Practical Considerations and Future Work
- Conclusion
Executive Summary
Orbix NDE’s AI Data Integrity system solves two key challenges in radiographic weld inspection: detecting duplicated weld images and identifying inconsistencies in radiographic shots. The system operates in real-time at the point of data submission, demonstrating a high degree of accuracy in detecting both types of integrity issues. By catching potential fraud and integrity issues immediately, the solution prevents costly rework and infrastructure failures while maintaining compliance. The system integrates with existing workflows and supports human experts by automatically flagging suspicious cases for review.
Introduction and Problem Context
Radiographic Inspection is fundamental to ensuring structural integrity and safety across industrial applications. However, the industry faces persistent challenges with radiographic data integrity, particularly concerning the authenticity and proper representation of inspection results. These issues, when undetected, create significant downstream risks: compromised safety assessments, regulatory non-compliance, and costly rework of critical infrastructure.
Our AI Data Integrity System is designed to address these challenges through two distinct subsystems. The first is a duplicate weld detection system that compares incoming weld x-rays to a database of existing images, flagging submissions that exhibit strong similarities indicative of potential duplication. The second subsystem focuses on detecting discontinuities within the individual shots that comprise a single weld, identifying inconsistencies or overlaps that could compromise weld integrity. Together, these systems ensure the authenticity and completeness of inspection data, enabling compliance and operational efficiency in real-time.
Operational Impact and Implementation Value
The AI Data Integrity System detects fraudulent or compromised radiographic data at the point of submission, preventing issues that could lead to catastrophic infrastructure failures and costly remediation. By providing immediate analysis and intervention capabilities, the system enables organizations to address data integrity problems before they result in extensive rework, project delays, or regulatory violations. This preventive approach ensures inspection integrity from the start, rather than discovering issues after infrastructure is already in service.
Testing demonstrates that the system maintains consistent performance whether analyzing digitized film or direct Digital Radiography. This technical flexibility, combined with its real-time detection capabilities, allows organizations to enhance their inspection integrity regardless of their current imaging technology. The system efficiently flags potential issues for expert review while documenting all verification steps, providing inspection teams with an effective framework for preventing costly data integrity failures.
Data Sources and Preprocessing
Our solution builds on a comprehensive dataset of conventional radiographic Weld Inspections, comprising tens of thousands of digitized images from diverse field projects. The dataset spans girth welds from 2” to 24” in diameter and encompasses the natural variations seen in industrial radiographic techniques and imaging parameters.
Our preprocessing pipeline applies further standardization methods while ensuring the preservation of the critical radiographic features. This standardization allows for reliable analysis across different digitization sources while maintaining the essential image characteristics needed for integrity verification.
Technical Approach
Figure 1: Convolutional neural network (CNN) analysis of weld patterns within radiographic images.
Our solution employs computer vision and deep learning techniques to analyze the unique patterns present in weld images. The system is designed to identify characteristic features that distinguish individual welds while maintaining robustness to image variations such as rotation and offset positioning. This enables reliable detection of potential data integrity issues across diverse inspection scenarios and imaging conditions while ensuring the system can handle common tactics used to hide data integrity problems.
The system’s architecture was specifically developed to handle the unique challenges of radiographic weld inspection, focusing on both localized discontinuities within individual shots and pattern matching across complete weld profiles. Through extensive validation on real-world inspection data, we’ve optimized our approach to maintain consistent performance across diverse inspection images and acquisition hardware.
At its core, our AI system analyzes the distinctive patterns within the internal structure of each weld as captured in each radiograph. The system’s AI models focus exclusively on these physical features. By concentrating on these fundamental weld characteristics, rather than peripheral or superficial image elements, the system maintains consistent performance. This approach ensures that our AI integrity verification is grounded in the actual physical properties of the welds being inspected, enabling reliable detection of both subtle pattern discrepancies and the presence of unique physical features.
Test and Validation Approach
To accurately evaluate system performance, we implemented a standard machine learning testing methodology by separating our digitized conventional radiographic data into three distinct sets: training, validation, and testing. Critical to our approach was ensuring complete isolation between these sets - no welds from the same project appear in multiple sets.
The validation and test sets were specifically curated to include digitized radiographs from entirely different projects than those used in training. This approach ensures our performance metrics reflect the system’s true ability to analyze novel welds rather than recognizing patterns from familiar projects. The test set results presented in this paper represent genuine field performance, measured against previously unseen conventional radiographic data.
1. Performance Metrics (Sensitivity and Specificity)
Our dual-system approach demonstrates robust detection capabilities across different types of potential fraud in radiographic inspection workflows. The shot discontinuity detection system achieves a sensitivity of 95.49% (detecting 95 out of 100 problematic shots) and specificity of 93.18% (correctly clearing 93 out of 100 legitimate shots), effectively identifying cases where individual shots have overlap discontinuities. The weld duplication detection system, which reviews complete weld profiles to identify fraudulent reuse, shows similar strength with 96.84% sensitivity and 99.37% specificity - meaning it catches 97 out of 100 duplicated welds while only incorrectly flagging 1 out of 100 legitimate welds. These detection rates translate to low false positive rates of 6.82% and 0.63% respectively, minimizing unnecessary reviews while maintaining high true positive rates above 95%.
These performance metrics demonstrate the system’s ability to reliably flag suspicious cases while minimizing false alarms, making it an effective tool for auditor workflows. By automatically screening all submissions and elevating only the most concerning cases for expert review, the system enables auditors to focus their expertise where it matters most. This combination of automated screening and expert oversight creates a robust framework for maintaining inspection integrity, capable of detecting both shot-level inconsistencies and fraudulent weld reuse.
Practical Considerations and Future Work
Digital platforms enable real-time deployment of our AI Data Integrity Solution, allowing immediate analysis of radiographic submissions. By identifying potential issues at the point of submission, the system serves as a preventive measure against both unintentional and deliberate data integrity problems.
1 Operational Integration and Real-time Analysis
The system processes radiographic data upon batch submission, analyzing shots within their daily context rather than against a universal database. This focused approach enables efficient processing while maintaining reliable detection capabilities. Through digital inspection platforms like Orbix, organizations can deploy these capabilities across multiple locations while maintaining consistent analysis standards. When potential issues are detected, the system automatically routes cases to qualified auditors for review through the platform’s interface, enabling early intervention before problematic data enters the inspection workflow.
2 System Considerations and Quality Requirements
While our system validation and testing are based primarily on conventional digitized radiography, initial testing indicates strong performance across digital acquisition systems including DR and CR. For reliable analysis, input images must meet standard radiographic quality requirements for density, contrast, and definition. The system architecture allows for straightforward optimization when needed to accommodate specific imaging equipment or inspection requirements.
3 Future Development Directions
Development priorities include expanding our validation dataset for digital acquisition systems, automated assessment of radiographic quality, and enhanced workflow integration. Through our digital platform, we continuously track system performance metrics across deployments, enabling targeted refinements while maintaining consistency. These improvements aim to streamline the audit process while maintaining rigorous inspection standards.
Conclusion
The AI Data Integrity Solution provides a comprehensive approach to radiographic inspection verification through two complementary subsystems. The shot discontinuity detection system identifies inconsistencies between overlapping radiographs, while the weld duplication detection system prevents fraudulent reuse of inspection data. Both systems demonstrate robust performance, with sensitivity above 95% and specificity above 93%, enabling reliable detection while minimizing false alarms.
The system’s integration with digital inspection platforms enables immediate analysis upon submission, allowing organizations to prevent data integrity issues before they impact downstream operations. This preventive approach, combined with efficient human-in-the-loop workflows, provides a practical framework for maintaining inspection integrity across multiple locations and projects. While initially validated on conventional radiography, the system architecture supports adaptation to various imaging technologies, positioning it to address current and future industry needs.
The performance metrics and operational deployment demonstrate that automated analysis can effectively support human expertise in maintaining radiographic inspection integrity. By focusing on the fundamental patterns present in weld radiographs and providing immediate feedback, the system helps organizations maintain compliance, reduce costly rework, and ensure the integrity of critical infrastructure inspection data.
