Title:A method to measure the embedded crack length and position in high-density polyethylene using microseconds ultrasound time signal

Abstract:High-density polyethylene (HDPE) is used in 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. Non-visible flaws can cause catastrophic failure if undetected. Large structures such as HDPE pipelines require a fast, non-destructive evaluation (NDE) method where the sensor can move rapidly across the structure. This is only possible if the flaw is evaluated using microseconds of time signal. We propose and show the accuracy of a machine learning-based Ultrasound NDE method that can rapidly and accurately predict embedded crack length and position simultaneously in HDPE with only tens of microseconds of time signal sensing. A method to quantify crack size in HDPE and other polymers using a very short Ultrasound time signal is lacking. We propose that an optimally trained machine learning model can decipher the crack characteristics using short measures of time signal, but a lack of large, well-distributed, and labeled datasets to train machine learning models continues to be a major limitation. To overcome this limitation, we have conducted computer simulations of ultrasound on HDPE to develop training data. We show that fully simulations trained convolutional neural network (CNN) can accurately predict crack lengths and positions in HDPE from experimentally measured ultrasound A-scan microsecond signals. Our method is based on the 1D time amplitude signal acquired over a very short time period and not based on 2D image analysis. The proposed methodology presents a pathway for training CNN using computationally generated data and applying the trained CNN in the field to quantify hidden cracks in large HDPE or other polymer structures.