<p>The furnace operation profile directly affects the economy, energy consumption, and condition of blast furnaces. However, it also presents a “black box” problem. In this research, a three-dimensional (3D) furnace operation profile recognition method, based on unsupervised learning and process experience, was proposed for blast furnaces. This method allows for the qualitative characterization of furnace operation profiles, helping to contribute to a higher degree of “transparency” in blast furnace operations. Thermocouple temperature data were used to characterize the furnace operation profile. Hierarchical clustering was applied for unsupervised clustering, with silhouette coefficient (SC), Davies–Bouldin (DB), and Calinski–Harabasz (CB) indices used as evaluation indicators. The best result was obtained with five categories, yielding scores of 0.61 for SC, 0.91 for DB, and 3940 for CH. The scores for categories A, B, C, D, and E were calculated by applying factor analysis on the basis of the blast furnace operation index and the slag layer stability index. Their scores were 0.35, −0.41, −0.83, −1.22, and −2.48, respectively, which were defined as the first through fifth levels of the furnace operation profile. On this basis, a similarity voting method is proposed to accurately localize the pattern of single-angle furnace operation profiles. Subsequently, a multi-angle scoring method combined with process experience was used to identify the 3D furnace operation profile pattern. When the total score was in the range of 33–40, 25–32, 17–24, 9–16, or 5–8, the furnace operation profile pattern was classified as optimal, well, general, poor, or abnormal, respectively. If the single angle was at the fifth level more often, the 3D furnace operation model was appropriately downgraded. This method closely aligned with field experience, and could help operators monitor blast furnaces online. Further, the economic performance and smelting status of blast furnaces would be improved.</p> Graphical Abstract <p></p>