Graph convolutional gated recurrent unit based instability prediction for continuous casting

Abstract

Due to the complexity of the internal environment for continuous casting, parameters of sensors are prone to instability, which leads to the decline of product quality, damage of enterprise profits, and even serious steel leakage accidents. Traditional instability prediction models are used to classify and warn states of sensors in advance by expert experience and traditional time series models. However, dynamic spatial correlations among temperature sensors in the instability evolution mechanism are often paid less attention, which leads to low accuracy and inability to provide more reliable decision-making suggestions for field workers. Inspired by the above problems, in this paper, a new graph convolutional gated recurrent unit based instability prediction model is proposed for continuous casting. Firstly, sliding window samples are constructed and the cosine similarity is used to calculate dynamic spatial correlations among multiple temperature sensors. Secondly, the graph convolutional gated recurrent unit is constructed to fully extract spatial–temporal fusion features of temperature time series, and the multi-layer perceptron based classifier is constructed to complete the instability prediction. Thirdly, two common instability states in the thin slab continuous casting technology of multi-mode continuous casting and rolling production line are used to test the accuracy of the proposed method by constructing binary and ternary classification datasets.