Deep-learning-based acoustic pipeline defect detection and characterization in dynamic environments involving varying...

Pipeline failures in industrial and municipal systems pose significant risks of environmental contamination, economic losses, and public safety hazards. However, it is often challenging to implement conventional pipeline inspection methods under dynamic environmental/operational conditions. Specifically, some of the most popular extended-range inspection methods for pipelines are based on acoustic guided-wave defect detection techniques, but they often exhibit degraded performance due to variable liquid flow rates and temperature fluctuations, resulting in critical gaps in continuous and reliable infrastructure monitoring. To address this challenge, this work presents a novel, deep-learning-based (DL-based) acoustic inspection method with two specialized deep learning architectures: a 1D convolutional autoencoder (CAE) and a 1D convolutional variational autoencoder (CVAE). The two architectures differ in how the latent representation is formed: the CAE learns a deterministic latent features, which favors stable and fine-grained multi-task regression, whereas the CVAE learns a probabilistic latent distribution, which explicitly models the uncertainty introduced by environmental variability and regularizes the latent space. These DL architectures are designed for noise-robust feature extraction and the disentangling of defect characteristics from environmental interference. To evaluate this new method, experiments were conducted on a test pipe in a flow loop with variable water flow rate and temperature. The tests used a 6-foot [183 cm] galvanized steel pipe (2-inch diameter) equipped with a sparse sensor network of three broadband transducers, generating rich experimental datasets. DL-based signal processing performed on these datasets demonstrated robust performance, with over 97 % physically simulated defect detection accuracy under flow rates from 30 to 122 mL/s and temperatures from 15 to 50 °C. The DL models achieved defect localization error under 10 cm and a relative defect quantification error under 10 %. Additionally, temperature and flow rate predictions by the DL methods yielded mean absolute errors below 0.83 °C and 1.56 mL/s, respectively. By integrating environment-invariant feature learning through convolutional layers and probabilistic latent space modeling, the proposed approach enabled simultaneous structural integrity assessment and environmental/operational parameter monitoring. These advancements have the potential to address a critical need for reliable inspection technologies in oil, gas, and energy infrastructure, particularly in dynamic environments where existing methods do not provide satisfactory solutions.