To address the data loss, distortion, and saturation of a single non-flame channel that may occur in a three-band infrared flame detector, a robust fusion algorithm for flame recognition based on a radial basis function (RBF) neural network entailing an improved Takagi-Sugeno (T-S) model is proposed in this paper. In this algorithm, the number of fuzzy rules required by the model is determined by a clustering algorithm. The membership degree of the feature component is added to the subsequent fuzzy polynomial to generate node output, and the weighted fuzzy node activation degree and feature characterization coefficient are defined to replace the Markov distance (fuzzy rule applicability) of the original model. Through the design of a three-band flame detector and routine and robustness experiments, it is shown that the proposed model significantly improves the number of nodes, convergence speed, accuracy, generalization ability, and robustness as compared with those of the traditional T-S model RBF neural network and genetic algorithm-back propagation models.