THANK YOU FOR SUBSCRIBING
Manufacturing Technology Insights | Wednesday, October 26, 2022
Non-destructive testing (NDT) in civil engineering has been widely developed to evaluate the properties of the material or systems of infrastructures for identifying internal defects without causing any damage.
FREMONT, CA: In civil engineering, non-destructive testing (NDT) has become widely used to assess the qualities of a material, component, or infrastructure system to identify inherent flaws without inflicting any impact. NDT is a highly valuable technology since it can acquire information about the damaged area without cutting or breaking the material because it does not permanently modify the thing being inspected. The creation of an autonomous data acquisition framework for NDT in infrastructures has been announced by the Korea Institute of Civil Engineering and Building Technology. Numerous researchers have examined ways to make NDT systems more effective and dependable during the past few decades. The majority of these studies have concentrated on the creation of sophisticated signal processing or hardware systems. The development of an effective sample approach has not received much attention, even though the places being examined (sampling) constitute the essential component for successful damage localisation. Grid-based sampling, which still relies on human judgement, is nevertheless used as a standard sampling design in NDT procedures. The grid-based sample inspects sites from the full domain of the structure at regular intervals. Before doing NDT, their locations in this sampling should be guessed at. A research team in KICT, led by Dr Seung-Seop Jin, has developed a Gaussian process (GP)-assisted active learning for autonomous data acquisition framework in NDT to reduce any subjective judgement and prevent missing out on the unknown damages.
Stay ahead of the industry with exclusive feature stories on the top companies, expert insights and the latest news delivered straight to your inbox. Subscribe today.
This framework's foundation is active learning, which sequentially directs sampling toward the regions of interest that are the damaged sites. With a few samples, the framework starts the active learning process. A set of input-output pairs is then obtained as the initial training data after the initial samples are reviewed by NDT inspection. Gaussian Process regression is built as a learning algorithm for active learning based on the initial training data. To direct the sampling toward the damaged areas, active learning is gradually employed. Then, to enhance damage localisation, a new sample is added to the training data. In this context, a recently created framework might successively choose promising damage spots, and this autonomous framework can be used to collect data for any NDT procedures. With the help of training examples, the recently built framework can choose the best model for visualising damages. The most promising place for sampling can be deduced from the ideal model. Until there are no more resources, such as the maximum amount of samples, this process is consecutively iterated.
Modelling the GP regression from damage visualisation requires careful kernel selection. The prediction of GP regression can differ depending on the kernel under even identical training samples. To put it another way, choosing the right kernel will allow you to accurately forecast and see the damage on the training samples. The right kernel selection in this case is a crucial element of the recently established framework. This automatic model selection module can speed up the active learning's synergy for more accurate damage localization with fewer samples. Impact echo tests for concrete structures were used to evaluate the synergy established by Dr Jin's framework and to identify various interior damages, including deep and superficial delamination. IE tests provide thickness data on internal slab and pavement degradation. The findings show that by directing sampling toward damaged areas, the suggested framework has the potential to provide more informative samples. In this approach for autonomous data collecting in NDTs, it has been empirically demonstrated that automated model selection can create synergy. Dr Jin's approach consequently offers improved performance for damage identification with less training data. While the proposed approach successfully locates informative samples in all damage zones to identify all damage with greater damage resolution, the grid sampling fails to identify some damages.
More in News