In a power plant, combustion condition monitoring is essential for maintaining stable operations and operational safety. Therefore, it is crucial to develop an intelligent combustion condition monitoring method. Existing methods not only need a large quantity of labeled data but also lack of generalization ability for monitoring the new condition. Aiming these problems, the present study presents a novel approach combining denoising autoencoder (DAE) and generative adversarial network (GAN) to monitor combustion condition. With the aid of the learning mechanism of the GAN, the learning ability is improved to learn representative features. These learned features are then fed into the Gaussian process classifier (GPC) for condition identification. Furthermore, new conditions can correctly be classified by simply retraining the established GPC using a small amount of labeled data under the new conditions, rather than training from scratch. Experiments were performed on a gaseous combustor and results indicate that the proposed approach can extract representative features accurately and provides higher accuracy for condition identification such as 99.4% for original conditions and 99.5% for the new conditions.
Keywords Combustion condition monitoring, Generative adversarial network, Gaussian process classifier