The present day’s requirement of the foundries is to produce good quality castings at the minimum cost and in the shortest possible time. Many castings especially the medium and large sized ones are produced in sand moulds, due to their cost effectiveness. The quality of the castings made in sand moulds is largely influenced by the mould properties and these properties, in turn, are dependent on the input process parameters. Hence, a good control on quality and cost of the castings can be achieved by choosing the appropriate binder and ingredients of the moulding sand systems. In the present study, three moulding sand systems, namely clay-bonded (green sand), sodium silicate-bonded carbon dioxide hardened and cement-bonded moulding sand systems have been considered. Among these moulding sand systems, clay-bonded one is the most popular and commonly used for most of the metals, whereas sodium silicate-bonded carbon dioxide hardened moulding sand system is used for ferrous metals to produce castings with good dimensional control. However, the cement-bonded moulding sand system is not well practised, due to the reason that it takes a long time to achieve the required strength. In the present work, an attempt is made to model these moulding sand systems using both conventional regression analysis as well as neural network-based modelling approaches. The first chapter starts with an introduction to the problem considered in the present study. A detailed literature survey is carried out on the above moulding sand systems. The gaps in the literature have been identified and the aims and objectives of the present work have been set. Chapter 2 describes the modelling tools and techniques utilized in the present study. For all these moulding sand systems, the input variables and mould properties that influence the casting process have been identified. Each moulding sand system is represented as an input-output model. Both linear as well as non-linear input-output relationships have been established using Design of Experiments (DOE) and Response Surface Methodology (RSM). These conventional statistical regression models have been tested for their statistical adequacy and their performances are compared with the help of 20 randomly-generated test cases. Among the above linear and non-linear models for each of the responses, the best one is chosen based on their average absolute % deviation values in prediction for the test cases. These models have been utilized in simulating a huge data and further the simulated data have been utilized to train the neural network-based models. The neural network-based models, namely Back-Propagation Neural Network (BPNN) and Genetic-Neural Network (GA-NN), have been developed for conducting both the forward as well as reverse mappings. The performances of these NN-based approaches were compared among themselves and with that of the conventional statistical regression models in case of forward mappings, whereas their performances have been compared among themselves, in case of reverse mappings. It is so, because it may not always be possible to carry out the reverse mapping using the conventional statistical regression analysis. This methodology has been adopted for modelling of the above three moulding sand systems.
Thesis submitted by Mahesh B. Parappagoudar, Guide: Prof. G.L. Datta and Dr. Dilip Kumar Pratihar, Year 2007