Machine Learning Model for Micro Electro-Discharge Machining

Abstract

Abstract: Modern manufacturing techniques such as Micro–Electrical Discharge Machining (µEDM) demand not only great precision and quality but also require that the product be produced in the shortest amount of time possible. Thus, to produce the intended output, process parameters are controlled via machinist instincts and heuristic methods. This has resulted in a situation where knowledge of the optimal values for numerous input parameters is required to maximise or minimise a certain outcome. Understanding the intricate interactions between factors and responses during the machining of different workpiece materials in EDM necessitates a comprehensive model. This model should address the non-linear and stochastic nature of the manufacturing technique while also evaluating the inherent machining uncertainties. This research proposed artificial neural networks (ANNs) and model ensembling techniques to model outputs associated with µEDM operations on various workpieces, incorporating material properties such as thermal conductivity, melting point, and electrical resistivity to assist the user in predicting machining responses like machining time (MT), tool wear rate (TWR), overcut (OC), and taper angle (TA), together with the standard deviation (SD) of each response based on input factors like capacitance, voltage, feed rate, tool speed, tool diameter, workpiece thickness and workpiece type. The ANN model was trained with experimental data designed using the I-optimal design of experiment technique (DOE), and the network hyperparameters were optimised through the grid-search technique for optimum performance. The mean percentage accuracies of the models across the training, testing and validation sets for MT MEAN, SD MT, TWR MEAN, SD TWR, OC MEAN, and TA MEAN were determined to be 85.4%, 76.84%, 85%, 81.48%, 86.62%, and 84.79%, respectively. The outcomes of deployment tests reveal that the experimental outputs of MT, TWR, OC, and TA align closely with the predicted model outputs, where the uncertainty associated with these predicted values are quantified by four sigma (equivalent to a 99.99% confidence level). The proposed model will be able to enhance the precision and efficiency of µEDM operations of the different workpieces, aiding informed decision-making in manufacturing processes.