A Machine Learning Method to Measure the Embedded Crack Length and Position in High-Density Polyethylene Using Ultrasound Time Signal

Abstract High-density polyethylene (HDPE) is a semicrystalline polymer used in several critical applications, ranging from cooling water pipelines in nuclear power plants and distribution pipelines for natural gas and hydrogen to biomedical implants. Embedded crack-like flaws form within HDPE during fabrication or operations, which may grow over time and can cause catastrophic failure if undetected. Large structures such as HDPE pipelines, where the location of a flaw is not known, require a fast, nondestructive evaluation (NDE) method where the sensor can move rapidly across the structure with a very short data collection window of microseconds at each location. This is only possible if the flaw is evaluated in HDPE and other polymeric structures using a microsecond time signal. Ultrasonic A-scan (time signal) allows for the rapid scan of large structures, whereas B-scan ultrasounds are limited, as they are slow and depend on postprocessing algorithms, where subtle information can be lost. We propose a methodology for training a convolutional neural network (CNN) using computer simulations of ultrasound on HDPE and applying the trained CNN to real-life experiments to decipher crack characteristics in HDPE or other polymer structures using ultrasound time (A-scan) signals. We show that a fully finite element simulation-trained CNN can accurately predict crack lengths (mean absolute percent errors (MAPE) 3.2%) and positions (MAPE 3.8%) in HDPE from experimentally measured ultrasound A-scan microsecond signals. The success of a 100% simulation-trained CNN without exposure to any prior experimental data in accurately predicting crack sizes from experimental time signal data underscores a promising path for next-generation NDE methodologies.