Modelos generativos aplicados à emissão acústica para detecção e localização de vazamentos em tubulações

Abstract

Leak detection in pipelines is critical to preventing financial losses and environmental damage; however, training artificial intelligence (AI) models for this task is hindered by the difficulty and high cost of acquiring real-world leak data. This study proposes a solution by generating synthetic, yet realistic, leak signals from a small set of real measurements. Acoustic Emission (AE) signals were collected from a 105-meter experimental pipeline, and a generative adversarial network (GAN) was trained to learn their physical characteristics and produce new samples for various leak positions, including locations where no measurements had been taken. To assess the quality of the generated data, a Vision Transformer (ViT) model was trained exclusively on the synthetic leak signals to detect and locate leaks from spectrogram images. The model, which had never been exposed to real leak data during training, was then evaluated on real samples. Results showed that it achieved 100% accuracy in leak classification and an average localization error on synthetic data of only 1.5 meters. These findings demonstrate that synthetic data generation is a viable and effective strategy for addressing data scarcity in anomaly detection, enabling the development of diagnostic systems without the need for extensive and costly field testing.