Hybrid Methodology To Predict Surface Texture Of The Honed Engine Cylinder Bore

The cylinder bore is a vital part of the engine as it acts as the combustor and a platform for transforming chemical energy into mechanical energy.It consists of a multi-scale surface texture with a crosshatch pattern,valleys,and plateaus to reduce friction in the piston ring assembly while preventing gas and oil leakage.To achieve the desired surface texture,a sequential honing process is adopted.Honing is an abrasive based finishing process.It consists of various input variables(pressure,stroke speed,rotational speed,honing time,grain size,and concentration)and multiple stages,with each stage having its specific target.Furthermore,the produced surface texture needs to have very tight tolerances.Considering all these variables makes it difficult to accurately predict the multi-scale surface texture of engine cylinder bore.In literature,two techniques were primarily employed to predict the multi-scale surface texture: analytical models(AM)and machine learning(ML)models.Both of these approaches offer certain advantages and limitations.AM provide the benefit of distinct relation among variables improving their interpretability.However,they are developed on simplified assumptions to trace the complex relationship among variables resulting in loss of accuracy under conditions when these assumptions are not satisfied.Oppose to AM;ML algorithms do not rely on predefined assumptions regarding the process mechanism.Recently these models have been extensively used to model abrasive-based finishing processes as they provide superior results compared to AMs’.However,these models are data-driven and rely heavily on data to develop a statistical model to explain the underlying process mechanism.Thus,they offer poor predictions for the feature space’s target areas not sufficiently explored during the training process.Unfortunately,with the growth of input variables that affect the output,the number of all possible configurations increases exponentially,resulting in a phenomenon known as the dimensionality curse.Therefore,gathering data for all possible configurations can be time-consuming and difficult for complex processes such as honing,which requires several input variables at each stage of the process.This research aims to develop hybrid models in which the ML algorithm can complement the AM to improve its prediction accuracy and efficiency in terms of computation time.Thereby,it preserves the benefit of employing the AM including the enhanced interpretability of the process mechanism.The developed hybrid models are based on hybrid boosting,an ensemble method that combines different models to form a strong learner.They utilize ML algorithms to reliably learn the function that characterizes the difference between actual and the simulated results instead of the original target function(surface texture).The proposed techniques result in a more robust model,i.e.,less sensitive than a model based on one of the two approaches applied alone and a significantly more efficient model than the AM.The main contributions made by this study included:1)Development of Analytical model for surface texture prediction: A tool topography model considering the abrasive shape,size,concentration,distribution,and wear was developed.The model was compared with the observed topography of the honing stone.The tool topography model was used alongside the kinematic model to simulate the sequential honing process.2)Hybrid model development to improve prediction accuracy: The prediction accuracy of the hybrid model was significantly improved by employing an artificial neural network model that could predict the error in AM as a correction factor.The proposed approach improved the prediction accuracy of AM from 79% to 96.6%.Furthermore,the model provided superior results than independent artificial neural network and a random forest model.3)Enhancing the efficiency of the hybrid model: The computation cost of the AM caused an increase in the computation time of the hybrid model.This limitation was rectified to a significant extent by replacing the AM in the hybrid model with an ANN that could predict the simulated surface texture.The approach resulted in a 76.47% more efficient model that could predict the surface texture with an accuracy of 92.6%.