Performance predictive models in turning operations using neural networks

Turning as a practical machining operation has been extensively studied in the past. The nature of chip formation, tool life, tool wear and force predictive models were a major focus of research over the years for this indispensable manufacturing process. Empirical models and mechanics of cutting models to performance prediction have been used to develop reliable quantitative predictive models. While a reasonable quantitative reliability is achieved by these two methods the rigour involved with empirical methods and intricate modeling involved for mechanics of cutting models have forced investigators to look for alternative models. Both empirical and mechanics of cutting approaches have been found to be direct modeling techniques, where a set of process variables are fed into the respective models for a performance output. In recent years neural networks have been found to be applicable to variety of manufacturing applications and more particularly to decision making and as a tool for performance estimation. The quantitative reliability of the developed neural network models depends on the extent of training data. In this thesis a comprehensive review of the machining performance of models is carried out. The survey reviewed the mechanics of cutting models, empirical models and neural network architectures for performance prediction and their limitations. A comprehensive range of turning experiments covering a range of cutting variables, tool-geometrical variables, hardness and tool coating is carried out. This experimental investigation, while complementing the established trends in the literature, is used to train and test the developed neural network models. The qualitative and quantitative effects of major process variables on turning performance are established using statistical routines. A part of this investigation also covered the coated and un-coated cutting tool data for turning performance to check the claimed superiority of the coated tools. It has been found that developed neural network architecture can be used as both direct and inverse models for excellent quantitative reliability. This thesis is a step towards understanding the functioning of neural network architectures for performance estimation and eventual 'on line' use for performance prediction and control of practical machining operations.