Unsupervised Learning-Based Plant Pipeline Leak Detection Using Frequency Spectrum Feature Extraction and Transfer Learning

The deterioration of power generation facilities built during the early stages of plant operation is becoming increasingly severe, raising concerns about potential socioeconomic harm from pipe leaks. Consequently, there is a pressing need for rapid leak detection and proactive responses. Prior research primarily relied on various signal processing techniques and supervised learning for leak detection. However, these approaches struggle with accurate detection amid environments with diverse background noises and weak leak signals, exacerbated by challenges in gathering sufficient real-world leakage data, which can lead to overfitting during model learning. Therefore, in this paper, an adaptable leak detection model suitable for various environments was proposed to ensure precise leak detection. Frequency spectrum feature extraction and transfer learning were utilized to achieve accurate leak detection, even with limited data. In addition, an unsupervised learning-based autoencoder model is employed to identify leaks accurately by learning general patterns, even when leakage data is limited. Experimental results demonstrate that the proposed model, integrating feature extraction techniques using the Uniform Manifold Approximation and Projection (UMAP) algorithm and employing transfer learning, achieved a higher accuracy performance with 6.35 percentage points (%p) compared to the model lacking these techniques. In addition, these findings confirm a slight decrease in accuracy performance even when using minimal learning data. Moreover, the leak detection performance was superior to the existing models considered in this study, achieving a high accuracy rate of 99.19%.