[2B5] Towards high-cycle thermal fatigue monitoring via ultrasound with machine learning
L Clarkson and F Cegla
Imperial College London, UK
Pipe networks within nuclear power plants (NPP) are susceptible to high cycle thermal fatigue (HCTF) failures in the so-called mixing zones, where fluids of different thermal and hydraulic properties interact. Given the potentially deadly consequences following component failure in NPPs, a method for HCTF progression monitoring is expected to be of great use.
Previous work has shown that it is possible to predict the inaccessible (interior) pipe surface temperature to within ±2°C using ultrasonic time-of-flight with two physics-based inversion models: inverse thermal model (ITM) and assumed distribution model (AD). The current work investigates the use of the non-linear autoregressive with eXogenous inputs (NARX) machine learning model in place of physics-based models. The results from testing trained NARX models on an unseen simulated dataset showed that inaccessible surface temperature predictions were within ±2°C of the true value. Additionally, the NARX models had at least a 16 times lower root mean squared error compared with the predictions made with the ITM and AD models for the same dataset. Finally, a framework for computing an estimate of the damage accumulation due to HCTF from a temperature profile predicted by NARX was explored.
Previous work has shown that it is possible to predict the inaccessible (interior) pipe surface temperature to within ±2°C using ultrasonic time-of-flight with two physics-based inversion models: inverse thermal model (ITM) and assumed distribution model (AD). The current work investigates the use of the non-linear autoregressive with eXogenous inputs (NARX) machine learning model in place of physics-based models. The results from testing trained NARX models on an unseen simulated dataset showed that inaccessible surface temperature predictions were within ±2°C of the true value. Additionally, the NARX models had at least a 16 times lower root mean squared error compared with the predictions made with the ITM and AD models for the same dataset. Finally, a framework for computing an estimate of the damage accumulation due to HCTF from a temperature profile predicted by NARX was explored.
