RUL Prediction of Continuous Casting Machine Based on Deep Transfer Learning

Abstract The operation and maintenance of continuous casting machines involve continuous casting rolls with varying degrees of newness, age, and maintenance history. This variability poses challenges in gathering adequate fault failure data for all continuous casting rolls, resulting in limited effectiveness of remaining useful life (RUL) prediction methods based on deep learning. This limitation arises from the insufficient availability of fault labeling data. To tackle this issue, we first construct a CNN-BiLSTM model for predicting the RUL of continuous casting rolls, utilizing available fault failure data. Subsequently, we introduce the concept of transfer learning, where the source domain consists of the old continuous casting roll dataset with ample fault failure data, while the target domain comprises the new continuous casting roll dataset with limited fault failure data. The network structure and parameters of the model are adjusted through the reciprocal exchange of data between the source domain and the target domain. This leads to the development of the TL-CNN-BiLSTM RUL prediction model, which combines transfer learning and deep learning. The efficacy of this model in predicting the RUL of continuous casting rolls with varying levels of fault failure data accumulation is subsequently demonstrated through illustrative examples.