Using a Novel CNN Model for Predicting the Induction Heating Lines to Obtain a Desired Deformed Shape of Steel Plate

Abstract

Induction heating is a fast and energy-efficient method of heating steel plates in the shipyard industry due to the increasing demand for steel plate deformation. In this study, deep learning models are proposed with abstract representations of heatmap images allowing efficient recognition of complex and nonlinear patterns. The model for determining the heating type for the plastic zone is a fairly simple and effective approach to perform the heating type determination with the choice of weight optimization of network. A novel Convolutional Neural Network (CNN) is trained to predict heating regions from heatmap images with different vertical displacements of the required steel plate. We have proposed a Region based CNN (R-CNN) model using a sparse group to classify heating lines based on heatmap images of desired complex shape as input. The rotated region of interest pool layers and the rotated bounding box regression are also suggested to recognize inclined slender shapes of heating lines. The R-CNN is trained to detect heating regions bounded by labelled boxes within the confines of the steel plate. A frame-based event predictor for deformation areas is trained independently to analyse each individual box in the regions proposed. Then a post-processing step is based on the output of the R-CNN to determine the actual heating zones. By evaluating through the data sets, we have experimentally found that the proposed algorithm has performed well and achieved better results than other conventional methods.