In the shipbuilding industry, the inefficiency of the successive approximation control method in CNC cold-bending machines has hindered productivity in steel bending manufacturing, particularly for rib profiles. This study proposes control methods for cold bending machines based on deep learning models to address this challenge, including CNN and Transformer-CNN (T-CNN), to predict the elastic spring-back rate of cold-processed metal profiles and generate precise control pulses for achieving target bending angles. Experimental validation using real-world datasets collected from a shipyard’s CNC cold bending machine demonstrates that the T-CNN model significantly reduces the number of steps required for each bending operation, achieving a 75% reduction in production time and substantially enhancing processing efficiency. By leveraging the strengths of CNNs and Transformer architectures, the T-CNN model excels at handling long sequence data and capturing global dataset characteristics. Results show that the T-CNN model outperforms traditional control methods and standard CNNs in prediction accuracy, stability, and efficiency, making it a superior choice for cold bending control.