Abstract: In modern steel production, continuous casting serves as a crucial process that is evolving toward greater precision and integration. However, due to a large number of interrelated control variables in the continuous casting process, instability phenomena frequently occur, adversely affecting the product quality and production safety. Therefore, developing effective methods for predicting instability states in continuous casting process is of great significance. Existing methods typically treat multivariate data as equally important, neglecting the varying importance of different variables, and thereby affecting the prediction performance of continuous casting instability states. To address this issue, in this paper, an adaptive multivariate data fusion based instability state prediction framework is proposed for continuous casting process. Specifically, a channel attention mechanism is employed to adaptively assign fusion weights to different variables according to their relevance to process stability and various instability states. The weighted fusion is then performed based on those adaptive weights. Subsequently, the one-dimensional convolutional neural network is utilized to extract the features from fused data, which are then used for instability state prediction. Finally, two datasets are constructed using actual data from multi-mode continuous casting and rolling production line, which are used to conduct comparative experiments, and an interpretability analysis is also performed on the proposed method. Experimental results demonstrate that the proposed method achieved prediction accuracy of 99.62% and 96.72% in two cases, validating the effectiveness and superiority of the proposed framework.