Wed, 13 March, 2024
An Automated, Artificial Intelligence-Assisted, Defect Recognition Solution to Improve Reliability
In industry, the integrity of material joints is not just a priority but a necessity for safe operation. Traditional radiography of joints has played a pivotal role in upholding this integrity, although it's often affected by the limitations of human image interpretation. ‘Manual’ interpretation is not only time-consuming and operator-dependent but also comes with significant expense. Moreover, the primary factors affecting the quality of image interpretation are the inspector's individual expertise and psychological factors, such as fatigue or stress. Hence, the possibility of developing highly capable automated tools has emerged as a pivotal advancement with the potential to bolster the reliability of NDE.
TWI has collaborated with Industrial Member companies on a project to develop assisted defect recognition (ADR) utilising artificial intelligence (AI), with the aim of automating the defect detection process, to reduce inspection time, and provide accurate and consistent results.
Ideally, a large dataset would be used, from composite samples both with and without defects, and with various types of defects. One alternative is to simulate composite samples, however, following a feasibility study, it became apparent that simulating defect generation would be cost-prohibitive. This was due to the necessity to replicate the actual manufacturing process, requiring data for calibration and the implementation of a computationally intensive model. As an alternative, test samples were acquired, some with pre-agreed defects and others without, followed by computed tomography (CT) scanning. After scanning, the data for each sample was 40GB.
An AI ‘supervised classifier’ was developed, which achieved an accuracy of 98.9% in identifying pre-agreed defect types from the CT scans with no human intervention. However, as is common, some unforeseen defects emerged from the samples due to manufacturing variations. The author noticed that when defect types were not of the pre-defined types, the 'supervised classifier' model might struggle to provide predictions or even generate erroneous predictions. To address this challenge, the author recognised the need to train the model on a comprehensive set of potential defect classes it might encounter in the future. This task, however, proved to be challenging due to the inherent complexity of composite materials.
As an alternative, a ‘novel detector’ was developed, which, unlike the previously mentioned ‘supervised classifier,’ learns from defect-free samples, enabling it to detect a wide range of defects. This approach has two main advantages:
- Adaptability - It can detect unforeseen defects, a significant improvement over supervised learning-based approaches that struggle with unexpected defect types
- Future-ready - The ability to learn from defect-free samples makes it versatile and adaptable for future challenges, as listed below:
- Quality control during manufacturing: It can be used for real-time quality checks, ensuring high speed, reliable and efficient inspections during a manufacturing process
- A portable CT system equipped with the ‘novel detector’ could be used for onsite inspection of pipes, to assist in maintaining continuous and safe operation
This AI-assisted defect recognition tool has the potential to automate defect detection to reduce inspection times and avoid human factors that can degrade ability to provide accurate and consistent results.
Please contact us, at the email address below, to find out more…
- Kai Yang (Senior Project Leader)