Identification of Combustion Conditions for C1–C4 Hydrocarbons Using CNN and CNN-LSTM Architectures

Accurate identification of combustion conditions from flame morphology remains a challenging task in combustion diagnostics due to the indistinct visual features. To tackle this issue, two deep learning architectures, a conventional CNN and a hybrid CNN-LSTM, were trained on flame images of C1–C4 hydrocarbons to perform fuel classification and equivalence ratio regression. The CNN demonstrated rapid convergence. In contrast, the CNN-LSTM required about twice the training duration but achieved superior identification performance. The CNN model achieved a classification accuracy of 89.92% for fuel type identification, whereas the CNN-LSTM model exhibited superior performance with 97.48%. The hybrid model showed particular strength in distinguishing visually similar flames. For the identification of equivalence ratio, the CNN yielded a mean absolute percentage error of 7.4%, whereas the CNN-LSTM reduced the error to 3.8%, providing more accurate identification across all fuel types. This work confirms the feasibility of using convolutional networks for non-contact flame diagnostics and provides a foundation for developing intelligent monitoring systems in combustion engineering.