Prediction of mechanical properties of CSP using greedy unsupervised pre-trained BP neural network

A BP neural network based on greedy unsupervised pre-training is proposed to improve the prediction accuracy of mechanical properties in hot-rolled strips produced by thin slab casting and rolling (CSP) technique. With this method, unlabeled data are used for greedy pre-training. After the training process, trained weights and bias are used as the initial weights and bias of the BP neural network. The advantage of this method is that it not only can make full use of unlabeled data, but also improves the prediction accuracy of BP neural network. The measured ones showed that compared with BP neural network, prediction accuracy of mechanical properties with the method of greedy unsupervised pre-training BP neural network is improved about 2.2%, 4.4% and 4.2%. Compared with Support Vector Machine, prediction accuracy of mechanical properties with the method of greedy unsupervised pre-training BP neural network is improved about 0.4%, 2.6% and 0.4%. It is suitable for on-line mechanical property prediction of thin slab casting and rolling.