CRPE-MViT: An enhanced transformer model with circular relative positional encoding for the condition identification of zinc oxide rotary kiln

Zinc oxide rotary kiln is an important equipment in the recovery process of zinc smelting. Accurate identifying its operating condition is crucial to ensure their efficiency and safety. Traditional methods rely on operator’s experience, which can be inconsistent and suboptimal due to the kiln’s dynamic environment and the varying skill levels of operators. To addres this issue, this paper introduces a novel condition identification approach that integrates Circular Relative Positional Encoding (CRPE) with a lightweight Mobile Vision Transformer (MobileViT) network. The specifically designed CRPE module takes into account the shape characteristics of the kiln and the rotational motion mechanism, and MobileViT provides a lightweight network structure. The resulting prediction model, CRPE-MobileViT (CRPE-MViT), significantly reduces the number of parameters while maintaining high accuracy. Experiments on a real large rotary kiln shown that the CRPE-MViT effectively identifies kiln conditions with an overall accuracy of 92.40%. Specifically, for the classification of normal-fired condition, the model achieved a precision of 92.21% and a recall of 97.25%, providing a reliable solution for real-time monitoring in industrial settings.