Prediction of mechanical properties in steel heat treatment process using neural networks

Abstract

The present paper describes the application of a neural network technique for the prediction of tensile strength (TS) test results for steels heat treated using a batch heat treatment process. Industrial process data often contain outlying points, some of which can be spurious for a number of reasons. A data cleaning technique has been used to ensure that spurious data points are not present in the final neural model, which would otherwise hinder the model's representation of the true process. The effectiveness of this technique is demonstrated by comparison of the TS model trained on cleaned and uncleaned data. A model trained on cleaned data is generated for the prediction of TS in the form of an ensemble network, which was found to provide more reliable predictions and give a better representation of the degree of uncertainty in the network predictions. The performance of the model is evaluated from a metallurgical perspective. Application areas for the model are examined with particular attention being drawn to the need for caution when entering inputs into the neural model. An assessment of the model's ability to generalise to new treatment sites and new steel compositions is made, together with experimentation to determine the effect of measurement tolerances on the predicted values from the model.