Based on low-resolution infrared spectroscopy acquisitionand deep learning computational methods, an online detection methodfor flue gas temperature and CO concentration is proposed. The gas spectral radiation model was used to calculate the training data, thedistribution of flame flue gas temperature and CO concentration was inverted based on a multi-layer perceptron (MLP) neural network.Results show that the inversion errors of the MLP neural network model for temperature and CO and HO volume fractions are less than1%, and the prediction accuracies are all greater than 94.5%, which has good generalization and prediction capabilities. A set of on-linedetection device for flue gas temperature and CO concentration based on deep learning coupled with emission spectroscopy was established, and the ethylene diffusion flame and CH / NH partially premixed flame were investigated. The measurement results of flue gastemperature and CO volume fraction for the ethylene diffusion flame were consistent with the simulated flame results, which verified thefeasibility of the online detection method based on deep learning coupled with emission spectroscopy. Changing the ammonia doping ratio ofthe partially premixed flame and analyzing the temperature and CO concentration changes of the gas at different heights above the centralaxis of the flame, results show that the flue gas temperature at the same height increases with the increase of the doped ammonia, and theCO volume fraction shows a tendency to increase and then decrease sharply. The proposed method can detect the changes of temperatureand CO more sensitively, which can be used for combustion diagnostic studies of many kinds of flames, and also has some potential applications in the online detection of carbon emissions in power plants.