Prediction of Mechanical Properties of Rolled Steel Based on Multi-Scale Expert System

The mechanical properties of carbon steel sheets directly depend on the chemical composition and process parameters of steel, how to establish the mapping relationship between key parameters and mechanical properties is the focus of predicting the mechanical properties of rolled steel. Traditional convolutional neural networks cannot effectively model the correlation between parameters, and there is a problem of feature loss during the feature transfer process. This paper proposed an expert system based on multi-scale for predicting carbon steel sheets' mechanical properties. A multi-scale module for the entire process was proposed to extract comprehensive features from the data, where a graph convolutional neural network captures the nonlinear causal relationships between production data, and the multi-scale convolution module ensures effective feature transfer. Secondly, the multi-activation module, composed of parallel channel and spatial attention, focuses on key features, effectively improving the network model's generalisation performance and computational efficiency.