Abstract: At present, iron and steel enterprises mainly use “after spot test ward” to control final product quality. However, it is impossible to realize on-line quality predetermining for all products by this traditional approach, hence claims and returns often occur, resulting in major economic losses of enterprises. In order to realize the on-line quality predetermining for steel products during manufacturing process, the prediction models of mechanical properties based on deep learning have been proposed in this work. First, the mechanical properties of deep drawing steels were predicted by using LSTM (long short team memory), GRU (gated recurrent unit) network, and GPR (Gaussian process regression) model, and prediction accuracy and learning efficiency for different models were also discussed. Then, on-line re-learning methods for transfer learning models and model parameters were proposed. The experimental results show that not only the prediction accuracy of optimized transfer learning models has been improved, but also predetermining time was shortened to meet real time requirements of on-line property predetermining. The industrial production data of interstitial-free (IF) steel was used to demonstrate that R² value of GRU model in training stage reaches more than 0.99, and R² value in testing stage is more than 0.96.