A two-stage multisource heterogeneous information fusion framework for operating condition identification of industrial rotary kilns

The operating condition identification plays an irreplaceable role for the low-carbon and high-efficiency operation of industrial rotary kilns. However, existing single-stage multisource heterogeneous information fusion methods lack a unified framework to simultaneously fuse the complementary properties among visible images, infrared images, and process data, thus limiting the condition recognition accuracy. Moreover, smoke and dust interference make it challenging to extract critical image features such as flame brightness and blast pipe position, increasing the difficulty of condition recognition. To this end, this paper proposes a two-stage multisource heterogeneous information fusion (TSMHIF) framework for operating condition identification of industrial rotary kilns. First, in the initial fusion stage, a condition-aware visible and infrared image fusion network (CAVIF) is designed to generate fused images containing complementary properties of source images. In this network, a self-developed novel industrial system is utilized to collect aligned visible-infrared images of industrial rotary kilns. Next, an interpretable feature engineering is constructed by incorporating extracted shallow features based on mechanism knowledge and mined deep features with an autoencoder, and the blast pipe position in the shallow features is quantified by a keypoint detection algorithm based on a cascaded pyramid network (CPN). Then, in the comprehensive fusion stage, a multiplication operation is employed to fuse multisource heterogeneous deep features from fused images and process data to recognize the operating conditions. Finally, a joint training strategy is developed to balance the image fusion and condition classification networks. The classification loss, i.e., condition-aware loss, guides the training of the visible-infrared image fusion network to improve the visual quality of the fused images. The industrial experiments show that our proposed method exhibits superior performance in terms of identification accuracy, condition prediction deviation, and visual quality of fused images compared to other competitors.