5 Companies must determine the replacement time of machine parts correctly since it affects 6 their production costs and efficiencies. For this, it is aimed to determine the most appropriate 7 replacement time to minimize cost per unit. In this study, it is proposed to develop a hybrid 8 Artificial Neural Network (ANN)-Genetic Algorithm (GA) model to predict replacement time 9 without using a cost model. At first, a replacement cost model is developed to calculate 10 replacement times to use in

more »

... training the neural network. Nevertheless, the cost model needs 11 complex mathematical calculations. GA is used instead of the cost model to determine 12 replacement time, and thus, to achieve fast learning for the neural network. The hybrid ANN-13 GA model was applied to predict replacement time of bladder in tire manufacturing. 14 Furthermore, ANN and GA models, which were developed to increase the prediction 15 accuracy of the hybrid model, were used. The hybrid ANN-GA model showed better solution 16 according to the performance statistics than the other ANN and GA models. The values 17 indicate that the hybrid model is in good agreement with the cost model. Thus, it is 18 recommended that the hybrid model is used instead of the cost model. 19 Keywords: Replacement time; replacement cost model; artificial neural network; genetic 20 algorithm; hybrid ANN-GA model. 21 22 23 2 1 . Introduction 24 As the costs associated with machine parts correspond to a large proportion of the total 25 cost of production, economical machine part replacement times are very important especially 26 for expensive parts [1]. The optimal replacement time should be obtained to ensure 27 minimization of the expected average cost per unit time. Replacement cost models are based 28 on economic comparison between planned (preventive) and unplanned (failure) replacement 29 actions. In planned replacement, machine parts may be changed at only scheduled times, and 30 in this case, these parts might not have completed their useful life. In unplanned replacement, 31 replacement is made on failed machine parts during operations. This case may damage the 32 product that is being processed on the machine and therefore causes additional scrap product 33 costs. 34 Machine part replacement time is regarded as a random variable that is usually 35 modeled by Weibull distribution in replacement cost models. Weibull distribution is the most 36 widely used probability distribution of reliability studies because it is highly flexible in 37 compliance with random data and has the ability to be adapted for data with different 38 distributions [2-4]. In a replacement cost model, Weibull distribution parameters, α-scale and 39 β-shape, need be updated with new replacement data to revise replacement strategies [5]. It is 40 very time-consuming and labor-intensive to calculate the new replacement time based on the 41 cost model involving complex mathematical operations every time. 42 This study proposes a hybrid Artificial Neural Network (ANN) -Genetic Algorithm 43 (GA) model to predict a new replacement time without the need for a mathematical model. 44 The ANN method is used to predict the machine part replacement time which minimizes the 45 cost per unit. Since the ANN method evaluates not only present data but also past data, it 46 predicts replacement time more accurately than cost models. A replacement time cost model 47 is developed to provide the necessary data for the training of the ANN. The GA method is 48 3 used instead of the developed cost model to find replacement times, and thus, to accelerate the 49 learning of the neural network. The replacement times obtained with the GA method 50 correspond to output data needed for the training of the neural network. 51 Replacement time cost model studies in the literature are mostly concerned with 52 machine tool replacement [6-8] and machine replacement [9,10], rather than machine part 53 replacement. A number of studies have been carried out with different criteria than those 54 known in replacement cost models. In one of them, Wang et al. [11] used profit instead of 55 cost as the criterion of economics in their replacement model. In another study, Sheikh et al. 56 [12] used the number of products processed in the machine instead of the lifetime in the cost 57 function to determine the optimal machine tool replacement interval. 58 Hybrid ANN-GA algorithms have been used in the past for the cost minimization 59 problem by researchers. Hashami et al. [13] proposed a hybrid model including ANN 60 optimized by GA for estimating power plant project costs. ANN was used to predict the costs, 61 whereas GA was used to set the ANN's parameters such as number of hidden layers, number 62 of nodes per each hidden layer and the corresponding weights and biases. Seo [14] developed 63 a hybrid GA-ANN model to predict product life cycle costs. GA was used to improve ANN 64 by eliminating irrelevant factors, determining the number of hidden nodes and processing 65 elements and optimizing the connection weights between layers. Firouzi and Rahai [15] 66 achieved a hybrid ANN-GA model optimizing risk-based repair and maintenance actions and 67 yielding the minimum life cycle cost for concrete bridge decks. 68 There is a limited number of studies in the literature including the ANN method or GA 69 method or hybridized ANN and GA methods in determining replacement times. Al-Chalabi et 70 al. [16] presented a model-based ANN method to determine the economic replacement time 71 (ERT) of production machines. Aldhubaib and Salama [17] illustrated an approach to link 72 maintenance and replacement decisions. They used GA to optimally schedule maintenance 73 4 activities. Liu et al. [18] conducted a study using ANN, GA and Weibull distribution together. 74 They structured a model to determine long-run fuzzy expected replacement cost per unit time 75 and the optimal preventive replacement interval. The ANN method was used for parameter 76 estimation, reliability prediction and evaluation of the expected maintenance cost. The GA 77 method was used to find the values for the membership function at any cut level. The 78 effectiveness of the proposed method was illustrated using a two-parameter Weibull 79 distribution. 80 In the literature, in cost-based hybrid ANN-GA models, GA was mostly used for 81 tuning the parameter values of ANN. In this study, unlike others, GA is used to obtain 82 replacement time, which is the neural network model's outputs, based on the cost model. 83 Application of the developed hybrid model in a real setting was illustrated on a 84 bladder used in a curing press in tire manufacturing. ANN and GA models were individually 85 created to increase the replacement time prediction performance of the developed hybrid 86 model. The application results of the developed hybrid ANN-GA model, GA and ANN 87 models were separately compared to the results obtained with the proposed replacement cost 88 model. According to the performance statistics such as coefficient of determination (R 2 ), 89 mean absolute percentage error (MAPE), and root mean square error (RMSE) the hybrid 90 ANN-GA model had more similar results to the proposed replacement cost model's results 91 than those of the ANN and GA models. Hence, the hybrid ANN-GA model is recommended 92 to predict machine part replacement time instead of the cost model because it is more 93 convenient and practical, as well. 94 The contribution of this study may be summarized as follows: 95 • A machine part replacement cost model was developed,