Condition monitoring of an oxy-biomass combustion process through flame imaging and incremental deep learning

Condition monitoring of combustion processes in power plants is crucial for maintaining furnace stability, high efficiency, and low emissions, especially under flexible loads to meet fluctuating energy demands. Traditional machine learning approaches such as Random Weight Network (RWN) and Support Vector Machine (SVM) are trained on ‘seen’ combustion conditions and lack the ability to recognise newly ‘unseen’ combustion conditions. This paper proposes an Incremental Multi-mode Condition Monitoring (IMCM) model for recognising both ‘seen’ and ‘unseen’ conditions in an oxy-biomass combustion process to ensure the boiler operates under demanding conditions. A new recognition probability threshold strategy is also presented for the first time in this paper. Using flame temperature maps obtained from an Oxy-fuel Combustion Test Facility as input datasets, the IMCM model is built on a Source Multi-mode Condition Monitoring model by incrementally updating with newly ‘unseen’ datasets and a small portion of previously ‘unseen’ datasets to improve accuracy. Key hyperparameters, including loss balance weight, training epoch, and recognition probability threshold, were identified for optimal model performance. The IMCM model, with an established recognition probability threshold strategy, demonstrated high effectiveness with a Mean Recognition Success Rate of 92.40% after three updates. The IMCM model demonstrates considerable potential for practical multi-mode combustion condition monitoring in systems operating under variable conditions.