<p>The accurate state identification of blast furnace (BF) top gas flow is important for optimizing control strategies and maintaining stable operational conditions. However, the complex and dynamic conditions inside the BF present significant challenges. Existing methods primarily focus on the central gas flow, often neglecting the edge gas flow, handcrafted features, and the challenges associated with image acquisition. To address these limitations, a novel method was proposed for recognizing BF top gas flow states based on deep learning of temporal images. First, a ResNet-based classification model was used to detect the abnormal gas flow and filter out normal images for subsequent recognition. Second, a YOLOv8-based instance segmentation model was employed to identify instances such as the chute and burden materials, thereby removing interfering data. Third, a data processing strategy was designed to extract effective temporal images and handcrafted features, thereby creating samples for recognition. Finally, a Res-BiLSTM-Longformer model was proposed to simultaneously recognize both the edge and central states of the BF top gas flow. The experimental results indicate that the model effectively identifies the BF top gas flow states, achieving an accuracy rate of 95.83%. Specifically, the recognition rate for the central gas flow state reaches 100%.</p>