پیش بینی هزینه نگه داری تجهیزات ساخت و ساز: مقایسه بین شبکه عصبی رگرسیون عمومی و مدل های سری های زمانی باکس-جنکینز Predicting the maintenance cost of construction equipment: Comparison between general regression neural network and Box–Jenkins time series models

۱٫ Introduction Managing the maintenance cost of construction equipment is an important task for contractors in the construction industry, especially for those engaged in heavy construction work with extensive equipment use. Construction equipment provides the functions of earthmoving, lifting, and logistic supplies and is subject to various types of maintenance work, which include preventive maintenance, predictive maintenance, and running repairs, to stay in normal working conditions. Peurifoy etc. emphasized that “the cost of repairs is normally the largest single component of machine cost, the repair cost constitutes 37% of machine cost over its service life” [۱], and Vorster [2] pointed out that costs of repair part and labor make up between 15% and 20% percent of the total equipment budget, and is the most difficult to estimate, decisions regarding repair costs affect the hourly rate as well as the economic life of a machine. Maintenance costs can significantly change depending on equipment characteristics, the maintenance strategies of contractors, working conditions, and operator skills, which bring difficulty to estimating equipment ownership and operating cost for management decisions. One crucial yet challenging management activity is predicting the maintenance costs of equipment at various levels of the equipment-owning organization. An accurate prediction of equipment maintenance costs in the planning horizon facilitates budget planning for equipment operations, maintenance resource allocations, equipment repair, overhaul, and replacement decisions. The modeling of equipment maintenance costs can also reveal the dynamic behavior of equipment maintenance costs as well as their factors, on which management decisions can be made to interfere proactively with and predict maintenance cost variations. Traditionally, equipment owners in the construction industry (i.e., contractors, government organization, and equipment rental companies) predict the maintenance costs of various construction equipment based primarily on past experience, for example, the maintenance cost of a piece of equipment can be estimated from the historical data of similar equipment under similar conditions. Adjustment factors can be applied to the benchmark values to account for the impact from various factors related to equipment (age, heath conditions, maintenance history, etc.), environment (workloads, working conditions, etc.), and organization (equipment management policy, business nature, etc.). However, judgmental forecasting of future maintenance costs based on experience, intuition and personal knowledge is unreliable due to the inherent random nature of equipment failures. With no consensus on the methodology among industrial practitioners, the statistical modeling of the maintenance cost of construction equipment provides a better quantitative approach to predict maintenance costs in the planning horizon. Previous research in this area, which has commonly employed linear or nonlinear regression by ordinary least squares, has been conducted by Manatakis and Drakatos [3], Edwards et al. [4–۶], Edwards and Holt [7], and Gillespie and Hyde [8], among others. Apart from these conventional regression models, the use of the time series approach in this area or in related fields gives further insights into obtaining a good model of the maintenance costs of construction equipment. Moore [9] found that the maintenance cost time series has an inherent autocorrelation among observed cost series. Edwards et al. [4] utilized the centered moving average to analyze the time series of the maintenance cost of construction equipment and isolated its trend of changes. Zhao et al. [10] established an autoregressive moving average (ARMA) model, also known as the Box–Jenkins method [11], to model equipment failures based on transformed data. Durango-Cohen [12] adopted the ARMA with exogenous input model (ARMAX) to model the performance behavior of transportation facilities with the application of the Kalman filter. All these attempts have been made to describe and predict the behavior of equipment performance and maintenance cost by using time series forecasting models and results of various degrees of accuracy were obtained. Although time series analysis has been traditionally conducted using Box–Jenkins models, artificial neural networks (ANN) have also been used for time series modeling and analysis because of its capability to identify the complex underlying nonlinear relationships among time series data. The use of ANN in modeling and in predicting the maintenance cost of construction equipment has been presented in a number of related research work. Edwards et al. [5] used multilayer perceptron (MLP) to predict future values of the maintenance cost of construction plants and found that MLP neural networks have better performance than that of other modeling algorithms such as multiple regression. Hong and Pai [13] modeled and predicted engine reliability by using various forms of models, which include general regression neural networks (GRNNs), support vector machine, and ARMA, and compared their performance in predicting engine reliability metrics. Following Moore [9], who found that the time series of equipment maintenance cost has autocorrelations among observed data, this study aims to develop and compare time series models for a cost analysis of construction equipment maintenance by using both traditional Box–Jenkins models and GRNN, a machine learning-based forecasting model. The study first presents a univariate modeling of the time series of maintenance cost by using ARMA and GRNN to predict the maintenance cost of construction equipment based on its historical observations. The impact of fuel consumption on the maintenance cost modeling of both traditional vector autoregression (VAR) and GRNN is then investigated to evaluate the performance of forecasting models after the incorporation of this parallel explanatory variable. Finally, the performance of traditional time series models and that of GRNN models is compared, and their advantages and disadvantages are then discussed.