Aiming at the cognition problem of rotary kiln burning state, an intelligent cognition method based on deep transfer learning for rotary kiln burning state is explored. Firstly, based on the adaptive structure-based convolutional neural networks (ASCNN), the unstructured dynamic feature space is established with the determined mapping relationship of the flame image from global to local. Secondly, based on the feature discriminability measure and the variable precision rough set theory, under the uncertain condition of the finite field, the burning state cognitive intelligent decision information system with the discriminable and the unstructured dynamic feature representation is established from the perspective of information theory, to enhance the interpretability of the feature space of the burning states. Thirdly, a stochastic configuration networks (SCN) with universal approximation capability is built to establish a classified criterion for flame image burning states with stronger generalization ability. Finally, based on the generalized error and entropy theory, the entropy measure index of the uncertain cognition results of the flame image burning state is established, to evaluate the cognitive results of the burning state in real time. The dynamic transfer learning mechanism is constructed to realize the self-optimization adjustment and reconstruction of the multi-level differentiated feature space and its classified criteria for the burning state. The experimental results show that the intelligent cognitive model of flame image burning state based on deep transfer learning constructed in this paper is more feasible and superior to the existing method for the burning state recognition of cement rotary kiln.