Abstract 

Accurate and timely detection of boiler tube leakage in a thermal power plant is essential to maintain a stable power supply and prevent catastrophic failures. This paper proposes a novel, unsupervised learning-based leakage detection method, namely optimal temporal convolutional auto-encoder, which uses both acoustic emission signals and operating (i.e., temperature and pressure) signals. The proposed optimal temporal convolutional auto-encoder learns the characteristics of normal operating conditions by reconstructing input data and detects tube leakage by calculating its reconstruction error. Unlike conventional methods that mainly focus on modeling only inter-sensor relationships, the proposed method offers a deep learning structure that can effectively capture temporal as well as inter-sensor relationships. This paper also proposes a method to optimize the latent dimension of the auto-encoder structure by minimizing the entropy of the trained normal reconstruction errors. In a preprocessing step, the moving average filtering is used to reduce the effect of external noises in acoustic emission signals, thereby decreasing the number of false alarms. Kernel density estimation is adopted to automatically set the threshold. The effectiveness of the proposed method is verified through datasets acquired from real operating power plant. The results show that the proposed method can detect leakage more accurately than conventional methods.