Machine learning-enabled early prediction of dimensional accuracy for complex products of investment casting

Abstract

Abstract For the foundry industry, predicting the dimensional accuracy of investment precision castings is vital yet challenging. In order to reduce cost loss caused by out-of-tolerance phenomena, this work develops a data-driven framework for estimating and screening early products based on machine learning techniques. The hollow turbine blade is analyzed as a typical case for the proposed framework. Initially, a database was compiled from the same production line of wax patterns and corresponding castings. Feature engineering techniques were employed to choose the most important characteristics and simplify inputs, employing reject rate analysis and decision tree analysis. Random forest regressors (RFRs) were chosen as the fundamental models after a sensible selection of machine learning algorithms. To enhance the performance of RFRs, the results suggested that the thickness distribution of the source material is another crucial element in determining the dimensional correctness of castings. Ultimately, a hybrid metaheuristic strategy incorporating RFR and dung beetle optimization was developed. The results showed that the proposed prediction model could minimize the error by 22.4% compared to conventional models, making it an valuable tool for early dimensional quality evaluation and guiding wall thickness control for hollow turbine blades.