This article presents a methodology for predicting NOx emissions from a biomass combustion process through flame radical imaging and deep learning (DL). The dataset was established experimentally from flame radical images captured on a biomass-gas fired test rig. Morphological component analysis is undertaken to improve the quality of the dataset, and the region-of-interest extraction is introduced to extract the flame radical part and rescale the image size. The developed DL-based prediction model contains three successive stages for implementing the feature extraction, feature fusion, and emission prediction. The fine-tuning based on the prediction is introduced to adjust the process of the feature fusion. The effects of the feature fusion and fine-tuning are discussed in detail. A comparison between various image- and machine-learning-based prediction models show that the proposed DL prediction model outperforms other models in terms of root mean square error criteria. The predicted NOx emissions are in good agreement with the measurement results.

This article presents a methodology for the diagnosis of abnormal conditions in a combustion process through flame imaging and kernel principal component analysis (KPCA). A digital imaging system is used to capture real-time flame images and radiation signals, from which flame characteristics such as flame area, brightness, non-uniformity, and oscillation frequency are quantified. These characteristics are used as the variables to establish the KPCA model of the combustion process. With the use of Hotelling's T² and Q statistics, the monitoring of abnormal conditions of the combustion process is achieved. Unlike the traditional principal component analysis (PCA) method, the KPCA method is capable of dealing with nonlinear data via nonlinear mapping, which projects the original nonlinear input space into a high-dimensional linear feature space. The effectiveness of the methodology is demonstrated by applying the approach to processing the data obtained on a 9MW_th heavy oil fired combustion test facility. Experimental results obtained show that the KPCA method outperforms the traditional PCA in discriminating between the normal and abnormal combustion conditions, even in cases where the number of training samples is limited.

To reduce NO _x (nitrogen oxide) emissions from fossil fuel and biomass-fired power plants, online prediction of NO _x emissions is important in the plant operation. Data-driven models have been developed to predict NO _x emissions from various combustion processes with good accuracy. However, such models have initially been built based on known combustion conditions, which are historically “seen”. For new conditions, which are “unseen”, these models usually perform unwell. In this study, an online deep learning (ODL) model is proposed to predict NO _x emissions from an oxy-biomass combustion process for “seen” and “unseen” combustion conditions based on source deep learning and condition recognition models. The ODL model is mainly built based on “unseen” combustion conditions. A new objective function that consists of regression loss and distillation loss is introduced in the ODL model to improve the prediction accuracy. The ODL model is examined using boiler operation data, flame temperature maps, and NO _x data obtained under a range of oxy-biomass combustion conditions on an Oxy-Fuel Combustion Test Facility. Flame images acquired using a dedicated imaging system are used for computing the temperature distribution of the flame through two-color pyrometry. The results demonstrate that the proposed model is capable of predicting NO _x emissions under “seen” and “unseen” conditions with a mean absolute percentage error of less than 3%, for the first, second, and third updates.

This article presents a method for the multimode monitoring of combustion stability under different oxy-gas fired conditions based on flame imaging, principal component analysis (PCA), and kernel support vector machine (KSVM) techniques. The images of oxy-gas flames are segmented into premixed and diffused regions through the watershed transform method. The weighted color and texture features of the diffused and premixed regions are extracted and projected into two subspaces using the PCA to reduce the data dimensions and noises. The multi-class KSVM model is finally built based on the flame features in the principal component subspace to identify the operation condition. Two classic multivariate statistic indices, for example, Hotelling’s T² and squared prediction error, are used to assess the normal and abnormal states for the corresponding operation condition. The experimental results obtained on a lab-scale oxy-gas rig show that the weighted color and texture features of the defined diffused and premixed regions are effective for detecting the combustion state and that the proposed PCA-KSVM model is feasible and effective to monitor a combustion process under variable operation conditions.