Pneumatic Piping Leakage Flow Rate Classification Based on Transfer Learning

Abstract

Leakage is the most common form of malfunction in pneumatic systems. When leakage occurs, the escaping gas creates a cooler area near the leakage point. Thermal imaging can visually capture these temperature variations, allowing assess the severity of the pneumatic system’s leakage. However, the accuracy of thermal imaging is often affected by varying background temperatures due to differing image feature distributions. To solve this problem, this article introduces a framework that combines key area extraction with discriminative feature transfer learning (KDFTL). First, to meticulously extract the core area of the leakage from the expanse of the image environment, the infrared images from both the source and target domains (TDs) are fed into the key area extraction (KAE) section. Subsequently, a distribution difference index is designed on class centers, and the inter class separability and intra class compactness are evaluated. Then, to make the features of samples with the same leakage rate at different temperatures similar for classification, the feature transfer matrix is calculated and the transferred features are obtained and classified. Finally, extensive experiments on the dataset of thermal images for pneumatic leakage show that KDFTL delivered an average classification accuracy of 72.59%, which is a significant improvement compared to 44.07% transfer joint matching (TJM).