Abstract The need for precision components and parts in manufacturing industries has bought an increase in the need for finishing operations that can satisfy this demand. In addition, there is a continuous demand for hard and tough materials that can withstand varying stress conditions to ensure prolonged service life of components and parts. The need to process these materials economically so as to meet stringent product quality requirements (generally expressed as composite of a family of properties, so-called multiple response characteristics) has become a real challenge for researchers and practitioners in manufacturing industries. Grinding has the potential to meet these critical needs for accurate and economic means of finishing parts, and generate the required surface topography. Despite this importance and popularity, grinding still remains one of the most difficult and least-understood processes due to lack of adequate inferential mechanistic and analytical multivariate models, for varied industrial situations. In this context, data-driven inferential linear or nonlinear multiple statistical regression, and artificial neural network modelling have become increasingly popular techniques for complex industrial grinding processes. Unfortunately, these techniques are either proposed and implemented in isolation or presented as a comparative evaluation grinding case study. A systematic solution methodology for inferential multivariate modelling, which addresses the different phases, starting from preliminary linear random x -case multivariate regression model, hypothesis testing of influence of addition of higher-order nonlinear terms to the adequate linear model (or presence of nonlinearity), and subsequent selection of a suitable nonlinear artificial neural network-based multivariate model, is lacking. In view of the above-mentioned conditional requirements, this paper attempts to provide a systematic methodology to develop a multivariate linear regression model, hypothesis testing for the influence of nonlinear terms to linear model, and accordingly selection of a suitable artificial neural network-based inferential model with improved prediction accuracy and control of grinding behaviour. The methodology suggests the use of various statistical techniques, such as Q–Q (quantile–quantile) plotting, data transformation, data standardization, outlier detection test, model adequacy test, model cross-validation and generalization. The suitability of the recommended methodology is illustrated with the help of an engine cylinder liner grinding (honing) case example, in a leading automotive manufacturing unit in India.
Read full abstract