In the cold rolling process of steel strip products, strip breakage is an undesired production failure which can lead to yield loss, reduced work speed and equipment damage. To perform a root cause analysis, conventional physics-based approaches which focus on mechanical and metallurgical principles have been applied in a retrospective manner. With the advancement of data acquisition technologies, numerous process monitoring data is collected by various sensors deployed along this process; however, conventional approaches cannot take advantage of these data. In this paper, a machine learning-based approach is proposed to characterise and model strip breakage in a predictive manner. First, to match the temporal characteristic of strip breakage which occurs instantaneously, historical multivariate time-series data of a cold rolling process were extracted in a run-to-failure manner, and a sliding window strategy was adopted for data annotation. Second, breakage-centric features were identified from three facets – physics-based approaches, empirical knowledge and data-driven features. Finally, these features were used as inputs for strip breakage modelling using recurrent neural networks (RNNs), which are specialised in discovering underlying patterns embedded in time-series data. An experimental study using real-world data collected from a cold-rolled electrical steel strip manufacturer revealed the effectiveness of the proposed approach.