Abstract: To address the lag in f-CaO detection and the insufficient accuracy and generalization of existing data-driven models under nonlinear, small-sample conditions, we propose a small-sample f-CaO prediction method integrating an enhanced CNN-Transformer model with transfer learning. First, sliding average method and correlation analysis are employed to screen key process variables, followed by the construction of a hybrid CNN-Transformer model with enhanced feature extraction capabilities. Building upon this foundation, the model undergoes pre-training using large-scale data from the source production line. Subsequently, four transfer learning strategies are designed. By selectively freezing or fine-tuning parameters in the convolutional layers, cross-attention layers, Transformer encoder, and output layer, the model adapts to the small-sample data of the target production line. Experimental results demonstrate that the improved CNN-Transformer model achieves significantly enhanced prediction performance, with mean squared error (MSE) reduced by approximately 63% and coefficient of determination (R²) increased to 0.95. In small-sample scenarios, transfer learning effectively enhances the model's generalization capability on the target production line, with the strategy of freezing convolutional and cross-attention layers while fine-tuning the Transformer encoder and output layer yielding optimal results. The study demonstrates that the proposed method significantly alleviates the small-sample constraints in f-CaO prediction, exhibiting robust performance and application potential across complex production line conditions.