A multi-source feature-level fusion approach for predicting strip breakage in cold rolling

As an undesired and instantaneous failure in the production of cold-rolled strip products, strip breakage results in yield loss, reduced work speed and further equipment damage. Typically, studies have investigated this failure in a retrospective way focused on root cause analyses, and these causes are proven to be multi-faceted. In order to model the onset of this failure in a predictive manner, an integrated multi-source feature-level approach is proposed in this work. Firstly, by harnessing heterogeneous data across the breakage-relevant processes, blocks of data from different sources are collected to improve the breadth of breakage-centric information and are pre-processed according to its granularity. Afterwards, feature extraction or selection is applied to each block of data separately according to the domain knowledge. Matrices of selected features are concatenated in either flattened or expanded manner for comparison. Finally, fused features are used as inputs for strip breakage prediction using recurrent neural networks (RNNs). An experimental study using real-world data instantaneous effectiveness of the proposed approach.