[5B1] Development and comparison of RNN, LSTM and GRU neural network models for automated eddy current inspection of heat exchanger tubes using real-world industrial data
M Balaban and K Gündoğdu
AIS Field, Turkey
The field of non-destructive testing (NDT) is crucial for ensuring the safety and reliability of industrial equipment. One of the most widely used methods in NDT is eddy current testing, which utilises electromagnetic induction to detect and analyse flaws in conductive materials. Eddy currents are commonly used in heat exchanger tube inspection, to detect defects such as pitting, corrosion, longitudinal and circumferential cracking, fretting, wall thinning and erosion at tube support plates. Traditional approaches to analysing the vast amounts of data collected from eddy current inspections are prone to error, particularly in dynamic environments with changing conditions. To address this challenge, we developed neural network-based machine learning models for heat exchanger eddy current signal analysis and reporting. Our approach leverages ten years of inspection data collected from 89 different heat exchangers in real-world industrial sites, rather than simulated or laboratory data, and more than 100,000 potential defect areas are extracted using advanced signal processing methods. The artificial intelligence (AI) models are trained on data gathered from both the real and imaginary parts of eddy current signals, as well as features extracted exclusively from these parts. Potential areas are gathered using peaks in the signals and labelled as either ‘defect’ or ‘not defect’. These data, after certain preprocess steps, are then fed into various models as inputs. Three different neural network architectures (recurrent neural network (RNN), long short-term memory (LSTM) and gated recurrent unit (GRU)) are used for both ferrous and non-ferrous tubes, where a total of six different models are trained and compared. Our approach simplifies and automates the traditionally lengthy, human-driven and costly data analysis and reporting stages in eddy current inspections. Specifically, our AI-based digital assistant supports inspectors by streamlining the data analysis process, reducing the potential for human error and providing rapid and accurate results. Our models were tested using brass, stainless steel and carbon steel exchanger tube data collected from refineries and power plants and the most successful models achieved an accuracy of more than 99% in classification. Overall, our approach represents a significant advancement in NDT, offering a more efficient and reliable method for analysing and reporting eddy current inspection data. With further development and refinement, our approach has the potential to become a widely used tool in NDT and contribute to the safety and reliability of industrial equipment.
