GENERATION OF A SYNTHETIC DATASET OF DEFORMED PIPELINES WITH DEFECTS FOR STRESS PREDICTION BASED ON LASER SCANNING DATA

Abstract

The article presents a methodology for generating a synthetic dataset for modeling the stress-strain state of pipelines. The main goal is to create labeled data for developing and training machine learning models that will predict internal stresses in the pipeline walls based on measured coordinates of its outer surface. The dataset takes into account a wide range of realistic scenarios. It includes variations in geometric parameters such as the inner radius, the base wall thickness, and the total length of the cylinder. Special attention is paid to the integration of local thickness defects modeled using parameterized Gaussian functions, which allows controlling their characteristics. To increase the reliability and adapt the AI models to real materials, properties such as Young's modulus, Poisson's ratio, and thermal expansion coefficient are stochastically varied for each cylinder. The influence of complex external and internal loads is included: internal and external pressure, axial forces (tensile/compressive), bending moments (simulating non-uniform settlement) and temperature changes that cause thermal stresses. The key feature of the methodology is the simulation of real data collection conditions using laser scanners. Controlled random noise is added to the analytically calculated deformed coordinates of the outer surface of the pipeline. This reflects the typical inaccuracies of optical measurements, ensuring that the input data for the AI model are as close as possible to real field conditions. The proposed methodology allows for the efficient creation of large volumes of labeled data, indispensable for training and validation of deep learning algorithms that solve the inverse problem in materials mechanics. This opens up new opportunities for improving non-destructive testing, structural monitoring and prediction of the remaining resource of pipelines.