Self-supervised combustion state diagnosis using a noise-augmented generative adversarial network and flame image sequences

Abstract

Reliable diagnosis of combustion states, particularly distinguishing between stable and unstable flame conditions, is crucial for maintaining power generation efficiency and stability. However, accurate detection of unseen unstable combustion states remains challenging due to the complex dynamics of flames and the limited availability of unstable flame data. To address this challenge, this study proposes a self-supervised combustion state diagnosing method based on a noise-augmented generative adversarial network (NAGAN) and flame image sequences. The proposed method employs a convolutional autoencoder (CAE) and principal component analysis (PCA) to extract abstract flame features from image sequences. A novel multi-generator NAGAN architecture, comprising a long short-term memory (LSTM)-based generator and two Gaussian noise-augmented generators, is designed to synthesize diverse unstable flame feature sequences with temporal dynamics and identify the combustion state. A Gaussian abnormal flame feature generator (GAFG) leveraging Gaussian noise and binary masking is introduced to simulate a wide range of anomalies, enabling the discriminator to learn diverse representations of unstable combustion states. Experimental results on methane-air flames show that the proposed NAGAN achieves an accuracy of 0.978 and an F1 score of 0.986 on the flame stability diagnosis, with a recall rate of 0.975 for unseen unstable flames, outperforming most existing unsupervised machine learning and deep-learning based diagnostic methods. These results demonstrate the potential of the proposed method to improve combustion state monitoring, enhancing the reliability and efficiency of power generation systems.