A comparative analysis between the multilayer perceptron "neural network" and multiple regression analysis for predicting construction plant maintenance costs

• Published in: Journal of quality in maintenance engineering Vol. 6; no. 1; pp. 45 - 61
• Main Authors: Edwards, David J, Holt, Gary D, Harris, Frank C
• Format: Journal Article
• Language: English
• Published: Bradford MCB UP Ltd 01.03.2000; Emerald Group Publishing Limited
• Subjects: Artificial intelligence; Beneficiaries; Coal mining; Comparative analysis; Construction; Construction industry; Contractors; Costs; Earthmoving equipment; Expenditures; Forecasting techniques; Hydraulics; Hypotheses; Independent variables; Maintenance; Maintenance costs; Maintenance management; Mining industry; Multiple regression analysis; Neural networks; Plant and machinery; Productivity; Regression analysis; Studies; Variables; United Kingdom > UK; Plant and machinery; Costs; Maintenance; Multiple regression analysis; Artificial intelligence; Construction industry
• ISSN: 1355-2511; 1758-7832
• DOI: 10.1108/13552510010371376

Notes that the real test of maintenance stratagem success (or failure in financial terms) can only be resolved when a comparison of machine maintenance costs can be made to some benchmark standard. Presents a comparative study between two models developed to predict the average hourly maintenance cost of tracked hydraulic excavators operating in the UK opencast mining industry. The models use the conventional statistical technique multiple regression, and artificial neural networks. Performance analysis using mean percentage error, mean absolute percentage error and percentage cost accuracy intervals was conducted. Results reveal that both models performed well, having low mean absolute percentage error values (less than 5 percent) indicating that predictor variables were reliable inputs for modelling average hourly maintenance cost. Overall, the neural network model performed slightly better as it was able to predict up to 95 percent of cost observations to within ≤q £5. Moreover, summary statistical analysis of residual values highlighted that predicted values using the neural network model are less subject to variance than the multiple regression model.