LASER CUTTING OPTIMIZATION OF NON-METALLIC MATERIALS OVERALL QUALITY

Abstract

As the global stock of natural resources depletes the need of electricity efficient processes emerges. Laser cutting, an advance non-contact processing technique, outweighs the old methods such as hotwire and milling due to the requirement of retightening and replacement of cutting tools with time. Orthogonal array and Factorial design are selected as a design of experiment for modelling and optimization of Laser cutting process. The range adjustment of laser machine requires knowledge of experimental design, laser cutting process and material properties, otherwise missing values generate due to unsuccessful cutting. For this reason, many universities are unable to utilize these machines effectively. It is essential to formulate a technique which allows modelling the data with some missing values, consequently, it enhance the utilization of laser machines for research and other purposes. Initially, the qualities of output characteristic were modelled by Statistical and Neural network without missing values and then by supervised and novel Semi-supervised learning algorithms with missing values. The Statistical modelling results using one and two way analysis of variance with replication were better than other data mining techniques like linear and nonlinear regression, however, it is difficult to use these methods with missing values. Therefore, supervised neural network modelling is carried out and the effects of its parametric change are observed along the datasets size to model the orthogonal array. The neural network modelling results in edge quality and kerf width signal to noise ratio, it is acceptable, the edge quality indicates that modelling improves by pre-normalization, further improvement was made by increasing training data size to factorial design. It is observed that for the artificial neural network, supervised learning is not sufficient associated to orthogonal array, only due to edge quality mean modelling, average error were higher than the acceptable limit. The average error with factorial design was under 10%. The vast modelling experience of supervised learning engenders the development of novel Semi-Supervised learning algorithm. Consequently, the average error was reduced by utilizing the systematic randomize techniques to initialize the neural network weights and increase the number of initialization by using orthogonal array design of experiment, with up to 22% missing values. This algorithm reduces modelling time and cost thus reduces electricity consumption. The average error in Perspex sheet did not exceed 8.0% and 11.5% for edge quality and kerf width respectively. The overall quality was calculated by aggregation technique of data mining and a more generous and better aggregation is carried out by the novel combination of Fuzzy logic which provides overall quality for the customer while saving cost, time and Electricity.