A NEW MACHINE LEARNING MODEL FOR PREDICTING RELIABILITY OF CRANES

Abstract

In this study, new machine learning models were developed for the assessment of reliability and availability for machinery maintenance using pattern recognition artificial neural network. The choice of a classification model stemmed from the nature of available data. Hence, the input data variables for cranes were obtained from Hyster RS45-27 CH and Konecranes Liftace TFC 45 97-2002. The artificial neural network model was developed using MATLAB. Trial and error was initially used to arrive at the neural network architecture that gave the lowest mean square error. The architecture that was finally selected consists of input layer, three hidden layers and an output layer. The first and the last hidden layers had a total of 10 neurons each, while the second hidden layer had a total of 20 neurons. However, for the cranes, the highest prediction accuracy went to PRN-LMA with an accuracy of 87%, followed by PRN-CGF with an overall prediction accuracy of 86.3%. Next were PRN-OSS and PRN-BFG with prediction accuracies of 84.9% and 84.6% respectively, while PRN-BR was the least accurate, with an accuracy of 76.6%. Generally, the PRN-LMA models gave the highest prediction accuracy, while the Bayesian regularization models (PRN-BR) gave the least prediction accuracy. Particularly, PRN-CGF model, followed by PRN-LMA model predicted the highest number of failure days, while both models gave the highest prediction accuracy for failure days.