Artificial Neural Networks and Adaptive Neuro-fuzzy Models for Prediction of Remaining Useful Life

A RT I FI C I A L N E U R A L N E T W O R K S A N D A DA P T I V E N E U R O - F U Z Z Y M O D E L S F O R P R E D I C T I O N O F R E M A I N I N G U S E F U L L I F E Razieh T av akoli Department of Civil Engineering The Univ ersity of T exas at Arlington Box 19308, Arlington, TX 76019 razieh.tavakoli@uta.edu Mohammad Najafi Department of Civil Engineering The Univ ersity of T exas at Arlington Box 19308, Arlington, TX 76019 najafi@uta.edu Ali Sharifara Department of Computer Science and Engineering The Univ ersity of T exas at Arlington Box 19308, Arlington, TX 76019 ali.sharifara@uta.edu A B S T R AC T The U.S. water distrib ution system contains thousands of miles of pipes constructed from different materials, and of v arious sizes, and age. These pipes suf fer from physical, en vironmental, structural and operational stresses, causing deterioration which e ventually leads to their failure. Pipe dete- rioration results in increased break rates, reduced hydraulic capacity , and detrimental impacts on water quality . Therefore, it is crucial to use accurate models to forecast deterioration rates along with estimating the remaining useful life of the pipes to implement essential interference plans in order to prevent catastrophic failures. This paper discusses a computational model that forecasts the R UL of water pipes by applying Artificial Neural Networks (ANNs) as well as Adapti ve Neural Fuzzy Inference System (ANFIS). These models are trained and tested acquired field data to identify the significant parameters that impact the prediction of R UL. It is concluded that, on a verage, with approximately 10% of wall thickness loss in existing cast iron, ductile iron, asbestos-cement, and steel water pipes, the reduction of the remaining useful life is approximately 50% K eywords Remaining useful life prediction · Artificial Neural Networks (ANNs) · Adaptiv e Neural Fuzzy Inference System (ANFIS) 1 Introduction Pipe deterioration is an unav oidable process that occurs ov er time due to structural, operational and hydraulic capacity failure. Structural failure is caused by an y kind of defects on the pipe w all that reduces the structural inte grity of the pipe segment. Likewise, the soil surrounding the pipe has an essential role in failure time of pipes. In general, cracks, internal and external corrosion, pipe deflection, misaligned joints, and breaks are the most common type of defects associated with structural failure [ 6 ]. Se veral f actors impact the structural deterioration of w ater mains and their failures, including pipe material, pipe size, pipe age, soil type, climate, and c yclic pressures. Ho wev er , the physical processes that cause pipe breakage are complicated and as most water pipes are buried, only very fe w research is av ailable about how they deteriorate and f ail [ 11 ]. Operational failure is the most common failure in a w ater distribution system and generally occurs by a physical cause and can be resolv ed during a maintenance procedure and usually does not af fect the structural integrity of the pipes [6]. Statistical models have been dev eloped to quantify the structural deterioration of water distribution pipes based on analyzing v arious lev els of historical data (Shahata and Zayed, 2012). Rajani and Kleiner (2001) pro vided a critical revie w of the statistical, deterministic and probabilistic models that attempt to quantify and predict water pipe breakage or structural pipe failures. Their results illustrated statistical methods using historical data on water main breakage [ 12 ]. Moreov er , Martins et al. , 2013; Osman et al. , 2011, Tscheikner et al. , 2011, compare the strengths, weaknesses, and limitations of those statistical models [ 15 , 20 , 30 ]. Most of the models use different strategies to handle scarce data situations, so ev en for limited data av ailability deterioration models can give v aluable information. [ 32 , 24 , 25 ]. Other modeling categories are artificial intelligence models (e.g. genetic algorithms [ 19 ], neural netw orks [ 29 ] or Neuro-fuzzy systems [ 1 ]. These are purely data-driv en approaches that enable solving of complex problems without the necessity of detailed explicitly kno wn model assumptions [32]. Like wise, Kulandai vel (2004) de veloped a sanitary sewer pipe condition prediction model based on Neural Networks, which can prioritize se wer pipes for risk of degradation. Furthermore, he e valuated the performance of the proposed model with dif ferent sewer pipes data sources. The results showed that the prioritization model can assist municipal agencies in efficient utilization for tar geting critical areas with predicting deteriorated sewer pipes [13]. Fahmy and Moslehi (2009) proposed a model to forecast the remaining useful life of cast iron water mains. They considered sev eral factors related to pipe properties, pipe operating conditions and the external en vironmental factors influencing pipe failures. Three different data-dri ven techniques were considered in their model de velopment; these techniques are multiple regression and two types of artificial neural networks: multi-layer perception and general regression neural networks. The data was used in their model dev elopment acquired from 16 municipalities in Canada and the United States. They showed data-dri ven modeling methods are ef fectiv e in forecasting the remaining useful life of cast iron water mains and it ov ercame limitations associated with existing models [7]. Fares and Zayed (2010) developed a risk model for water main failure, which ev aluated the risk associated with each pipe in a water network. They considered four main factors: environmental, physical, operational factors, and consequences of f ailure and 16 sub-factors. In order to de velop the risk f ailure model, Hierarchical Fuzzy Expert System (HFES) w as used to process the input data and generated the risk of failure inde x of each water main [ 9 ]. They presented that pipe age was the most significant factor of w ater main failure risk, follo wed by pipe material and breakage rate. The model validation w as 74.8%, which was reasonable considering the uncertainty in volved in the collected data [9]. Rogers (2011), proposed a performance-based approach to e valuate the present and future conditions of pipes from routine pipe in ventory and break record data. The Multi-criteria Decision Analysis (MCDA) w as performed to provide a pipe replacement ranking. He showed that the pipe f ailure prediction and MCD A modules are a quantifiable process for prioritizing short-term and long-term pipe renew al decisions [23]. Christodoulou and Deligianni (2009) presented a neuro-fuzzy decision support system for the performance of multi- factored risk-of-failure analysis and pipe asset management, as applied to urban water distribution networks. They showed that the combination of Neural Netw orks and Fuzzy logic is extremely ef fectiv e in risk-of-failure analysis and prev entive maintenance of water distrib ution networks [1]. Clair and Sinha (2011) presented a state-of-the-technology re view on water pipe condition, deterioration, and f ailure rate prediction models to identify the gap between the existing models in pre vious research and those currently used by water agencies. They proved limitations of model capabilities and comple xity in analysis and v alidating these models [26]. Osman and Bainbridge (2011), presented a comparison and analysis of rate-of-failure models (R OF) and transition-state models using a single dataset for cast-iron and ductile-iron pipes in the City of Hamilton, Ontario, Canada. They compared the models’ ability to support breakage forecasting, long-term strategic planning, and short-term tactical planning [20]. In abo ve research, v arious v ariables are considered in the analysis, including pipeline diameter , material type, installation year , break history , etc., along with other risk factors such as operating pressure, soil type and soil acidity . Howe ver , there is no standard procedure for recording data on leaks, breaks, and condition indicators. Moreover , the above research does not consider wall thickness loss to predict the remaining useful life of water pipes. According to previous studies about failure in w ater pipes and remaining useful life prediction, there is a lack of a comprehensi ve model to ev aluate the risk of water pipes failure and determine the remaining useful life for different types of water pipe and there is no research using a combination of Artificial Neural Network (ANN) and Adapti ve Neural Fuzzy Inference System (ANFIS) models to predict the remaining useful life of water pipes. In order to ov ercome the limitations of existing approaches, this paper aims to implement a new approach for prediction of remaining useful life of water pipes. in this research, the Remaining Useful Life (R UL) is defined as an estimation before pipe experiences a structural failure. This type of failure in volv es conditions where the pipe section cannot function properly without replacement such as a pipe break. This research determines the most significant v ariables influencing RUL. Se veral v ariables are considered for the model dev elopment including pipe age, length, wall thickness loss, installation year , the number of 2 Figure 1: Distribution of inputs and tar get variables breaks, and pipe materials for cast iron, ductile iron, asbestos cement and steel pipes with limited diameter ranges 4 in. to 24 in (100 mm to 635 mm). 2 Data Analysis Preliminary data analysis is carried out; which consists of sorting and ordering of data obtained from sources, comparing data and finalizing input variables. The validation data are embedded into most appropriate regression models for comparing results with the actual results. Data for this research was collected from se veral municipalities. The first dataset was obtained from Southg ate W ater District (SWD) in Den ver , Colorado. The SWD water system contained approximately 40% A C (Asbestos Cement) pipe, 34% DI (Ductile iron) pipe, 25% PVC (Poly V inyl Chloride) pipe and about 1% of CI (Cast Iron) and steel pipe combined. These pipes had extreme wall thickness losses due to age and deterioration process. The second dataset was collected from a municipality in La val, Moncton and Quebec, Canada, with deteriorated distribution pipes ov er a large area [ 31 ]. Iron pipes that were installed in 1954 ha ve 95 breaks from 1987 to 2001 and 11 breaks in 1999 all the breaks occurred when the pipes are of ages between 20 and 50 years [ 31 ]. The third dataset which is been used in this research was collected from the City of Montreal and considered due to a lar ge database of cast iron water pipes [ 34 ]. The last dataset was from Den ver water’ s distribution system, which selected due to the mixture of old and ne w materials ranging from “pit” cast iron installed in the late 1800s to recently installed polyvinyl chloride (PVC). There was a total of 4,112 dif ferent pipes constituting the 296 miles (476.36 km) of water mains with a total of 5,610 break records [ 23 ]. All above datasets include 7 v ariables (pipe age, pipe installation year , pipe length, pipe material, pipe diameter , wall thickness loss, and the number of breaks) and one target v ariable which is the Remaining Useful Life (R UL) of the pipes. These variables were selected for inclusion in the model due to their accuracy and a v ailability for the entire dataset. An initial Exploratory Data Analysis (EDA) has been performed on the dataset to find out missing data, the correlation between variables, and their distribution. Moreov er , since there were a few missing data for some of the variables in the entire dataset, we ha ve cleaned (remov ed) the missing v alues. In addition, the cate gorical v ariables hav e been con verted into numeric data since dealing with numeric data is often easier than categorical data [ 5 ]. Figure 2 illustrates the distribution of inputs and tar get v ariables of the dataset. As can be seen from the above figure, most of the pipes were cast iron pipes constituting about 46% of the sample with asbestos cement pipes is the second-highest number follo wed by ductile iron and steel pipes. The age of pipes in this 3 T able 1: Statistics of W ater Pipe Database Used in this paper . Featur es Min Max Mean Std Mode Age (Y ears) 1 131 49.78 30.31 43 Diam.(in.) 4 24 10.66 5.13 6 Len. (ft) 20.5 36,161.4 2,870.51 5,008.58 5,280 Mat. 1.67 8.35 6.146 2.75 8.35 No. of Breaks 0 95 5.09 7.74 6 Ins. Y ear 1,887 2011 1,961.15 28.78 1,969 W all Thick L. 1 59 29.64 14.81 33 R UL (Y ears) 3 90 40.65 20.46 36 study ranges between 1 and 130 years. Most of the water pipes are in the 4 to 6 in. (100 mm to 150 mm) category and pipe diameter ranges are from 4 in. to 24 in. (100 to 635 mm). The pipe installed from 1887 to 2011. Most of the pipes installed from 1960 to 1969. Majority of pipes are between 303 feet (92.35 m) and 800 feet (243.84 m). that most of the pipes break from 0 to 2 and 8 to 10 during their life that most of the water pipes have w all thickness loss around 30-40%. The calculation of remaining useful requires data on the installation year , pipe material, and breakage history . The installation year of the pipe determines the age of the pipe. Pipe material regulates the manufacturer’ s recommended service life, gi ven as a range, which does not consider other factors such as pipe diameter [ 18 ]. According to De vera (2013), R UL is the dif ference between the pipe’ s age and Anticipated Service Life. De vera (2013) calculated anticipated service life (ASL) as the mean of the manuf acture’ s service life due to the uncertainty in the service life of a pipe. In this research, the remaining useful life is predicted based on the ef fects of each independent variables (age, diameter , installation year , material, number of breaks, length and wall thickness loss), Incorporating additional f actors attempts to minimize the uncertainty and v ariation in a pipe’ s service life based on additional operating conditions because the data required for this model is usually a vailable at municipalities. The histogram shows most of the remaining useful life varies from 35 to 45 years. Data analysis in volv ed all collected data as means of defining initial factors af fecting remaining useful life of water pipe. Furthermore, the analysis was used as means of re vealing data inconsistencies and errors. Minimum, maximum, mean, mode, standard deviation, and correlation v alues were developed for all factors (T able 1). The correlation of an input provides an indication of whether an input will correctly , or acceptably , train with a neural network. Multiple regressions and one-way analysis of v ariance were generated to check the correlation between each v ariable (input) with remaining useful life (target). In addition, analysis of variance and t-test are generated to determine the statistical significance of the other input variables [ 16 ]. When a model passes both ANO V A and t-test, it is statistically significant (P-value <0.05), which means that the dependent v ariable (response) and the independent variables (predictors) hav e a significant relationship. The ANO V A test results showed the significance of input v ariables (the p-value is p < .05). A regression model tries to find the best fit between the actual data and the predicted value of the model [ 8 ]. The performance of the model is assessed by calculating coefficient of determination ( R 2 ) or fit index. The input with a good correlation (v alue close to either 1 or - 1) will usually be more significant than an input with a poor correlation (v alue close to 0). Pipe materials are classified based on T able 2 [ 34 ]. Fares, 2010; dev eloped a hierarchical fuzzy expert system (HFES) to estimate the risk of water main failure. The author de veloped a pipe material factor performance and attrib uted the impact v alue based on the risk of failure of dif ferent pipe materials. The author applied that model to different case studies to v erify the model [9]. Based on T able 1, if we assumed the av erage age of pipes is 50 years, and standard deviation is approximately 30, approximately 68% of the pipes ha ve the age between 20 and 80 years because the amount of Z-f actor is between -1 and 1 and area of -1<= Z <=1 is 68%. The Z-factor is a measure of ho w many standard de viations belo w or above the population mean a raw score is. A Z-factor is also known as a standard score and it can be placed on a normal distribution curv e. Z-factor range from -3 standard deviations (which would fall to the far left of the normal distribution curve) up to +3 standard de viations (which would fall to the far right of the normal distribution curv e) [16]. Figure 2 illustrates the relationship between wall thickness loss and remaining useful life in the dataset. Remaining useful life has a lower v alue with the increase of wall thickness loss. This relationship is a polynomial regression. Howe ver the relationship between wall thickness loss and remaining useful life is linear here, b ut it does not indicate the relationship is always linear . The wall thickness loss depends on pipe material and may vary along the length of the pipe segment depending on the rate of corrosion along the pipe se gment [33]. Figure 3 illustrates the relationship between the water ages and the remaining useful life. As age of the pipes increases remaining useful life decreases. Pipes with an age of 100 years or more have lowest remaining useful life with a 4 T able 2: Classification of Pipe Material [34]. Pipe Material Chance of Deterioration EA Polyethylene Extremely low 0.42 Ductile iron V ery low 1.67 PVC V ery low 1.67 Steel V ery low 1.67 Concrete Medium 5.01 Asbestos High 6.68 Cast iron V ery high 8.35 Figure 2: Relationship between Remaining Useful Life and W all Thickness Loss quadratic regression model of 82%. As can be inferred from Figure 3, the relationship between those v ariables is not linear . Therefore, there is a need to implement the other computational models such as ANN and ANFIS to create an accurate model. 3 Methods 3.1 Artificial Neural Networks in Pipeline Pr ediction approach Over the past decades, Artificial Neural Networks (ANNs) hav e been recognized as an alternati ve to traditional statistical models. ANNs use an algorithm inspired by research into the human brain which can “learn” directly from the data. It can be defined as “highly simplified models of the human nervous system, exhibiting abilities such as learning, generalization, and abstraction” [ 17 ]. One of the advantages of a neural netw ork model is that a well-defined mathematical process is not required for algorithmically con verting the input into an output. Once trained, a neural network can perform classification, clustering and forecasting tasks. An ANN model was chosen for this research because of its ability to cover nonlinear and compound behavior of water networks. Furthermore, it covers many variables that increase the system’ s performance reliability [ 17 ]. In this research, eight ANN models dev eloped with one hidden layer that is dif ferent in two aspects: the number of neurons in the hidden layer and the random groups of data sets. The ANN1, ANN2 ... and ANN8 models ha ve 3, 4, 5, 6, 7, and 10 neurons in their hidden layers, respectiv ely (Figure 3.1). There is a close relationship between age, length, material, wall thickness loss, and remaining useful life. Thus, pipe material, wall thickness loss, length, diameter , and age are selected from the dataset as the input v ariables for ANN models. Due to the difficulty of accessing to wall thickness loss for the majority of the pipelines, this factor has been remov ed from fi ve ANN models. The number of hidden layers was determined through trial runs of the model. The dataset is split randomly into training (75%), validation (10%) and testing (15%). For each ANN model, trials were performed to reach the lo west error . The performance of the models was assessed based on R 2 , mean absolute error 5 Figure 3: Relationship between Remaining Useful Life and Age of Pipes T able 3: ANN1 results Phases MAE RRSE MAPE RAE T raining 0.17 0.012 1.076 0.007 V alidation 1.304 0.001 8.047 0 T esting 0.88 0.001 5.431 0.007 (MAE), relativ e absolute error (RAE), root-relativ e square error (RRSE), and mean absolute percentage error (MAPE) according to Zangenemadar and Moslehi, 2016 [34]. 3.2 ANFIS Appr oach Adaptiv e Neural Fuzzy Inference System (ANFIS) creates a fuzzy inference system (FIS) whose membership function parameters are (adjusted) using either a back-propagation algorithm or in combination with a least squares type of method. This allows fuzzy systems to learn from the data they are modeling [ 27 ]. ANFIS works when the input that comprises the actual v alue is transformed into fuzzy values using the fuzzification process through its membership function, where the fuzzy v alue has a range between 0 and 1 [ 27 ]. ANFIS techniques provide a method for the fuzzy modeling procedure to learn information about a dataset to calculate the membership function parameters that allow the related fuzzy inference system to track the giv en input/output data [10]. ANFIS is employed to model the relationship between the input v ariables. The same set of variables as applied with the ANN are considered in the fuzzy inference system. The complete ANFIS consists of fiv e layers, the fuzzy layer , production layer , normalization layer, de-fuzzy layer , and total output layer . Each layer includes sev eral nodes, which are defined by the node function. 4 ANN Results T able 3 presents the calculated the value of MAE, PRSE, MAPE and RAE for training, testing and v alidating phases for the model one. Figure 4 presents the error results for the total samples of ANN models. The X-axis presents the eight ANN models and the Y -axis depicts all error results (RRSE, MAPE, RAE, and MAE). The results demonstrates that the PRSE values are approximately equal in all ANN models in the three phases except in ANN6. The PRSE values are higher in the ANN6 model in training, validation and testing, which prov es the previous assumption that this model is the least accurate one. The MAPE values are higher in v alidation rather than testing and training; howe ver , the differences are more than 10%. Because MAE, RAE, RRSE, and MAPE are the least for the ANN1 and ANN7 model, it appears to be the most precise models. Moreov er , the MAPE v alue is less than 10, which categorizes this forecasting as a high-accuracy prediction. 6 Figure 4: Frame works of ANFIS (MA TLAB R2017) Figure 5: Error Results for all the Samples 7 Figure 6: Predicted results versus estimated results. The comparison of predicted results and estimated results are sho wn in Figure 4 for the best ANN model. Estimated R UL is calculated based on actual data and predicted R UL is based on ANN results for best model. Most of the results fall in the near area of y=0.9112x+3.7663. The coef ficient of determination is 89%, which indicates that the proposed models hav e predicted the remaining useful life of the pipes, and is reliable for further analysis of the network. The model’ s precision can be increased with adding more input variables that can af fect the water conditions. 5 ANFIS Results The training data is imported into Fuzzy Logic T oolbox, and a membership function is selected. The chosen membership functions is Gaussian function. After the model is trained using the hybrid-learning rule, the results output by different membership functions were tested against the verification data. In addition, the precision of each membership function is determined using the root-mean-square-error (RMSE). Figure 5 shows the correlation of the input variables with R UL. The slopes of wall thickness, age and installation year are higher than the other v ariables. Therefore, those variables hav e the most impact on the remaining useful life. 6 Discussion of Results The results of this paper sho w that neural network and ANFIS were adept in capturing the relationships for prediction of remaining useful life. According to neural network results, age, and wall thickness loss were most significant parameters in R UL, while based on ANFIS, age, wall thickness loss, installation year are significant. From statistical analysis, age and wall thickness loss were the most important v ariables. Therefore, it is concluded that age and wall thickness loss hav e the most significant relationships with remaining useful life [28]. Figure 6 illustrates the relationship between wall thickness loss and remaining useful life in dataset in dif ferent ages for cast iron pipes. The results show that indeed with increasing wall thickness loss remaining useful life decreases and pipes in old ages hav e a high renege of w all thickness loss compared to the pipes in young ages. The results show with increasing 8% of wall thickness loss for pipes greater than 60 years old, the remaining useful life decreases 70% approximately . Similarly , with increasing 20% of w all thickness loss for pipes between 50 to 60 years old, the remaining useful life decreases 25% approximately . Figure 6 illustrates the relationship between wall thickness loss and remaining useful life in dataset in different ages for ductile iron pipes. The results show that indeed pipes in old ages hav e a high renege of wall thickness loss compared to the pipes in young ages. The results show with increasing 12% of w all thickness loss for pipes between 31 to 40 years old, the remaining useful life decreases 10% approximately . Similarly , with increasing 14% of wall thickness loss for pipes between 21-30 years, the remaining useful life decreases 20% approximately . Figure 6 illustrates the relationship between wall thickness loss and remaining useful life in dataset in dif ferent ages for asbestos cement pipes. The results show that indeed pipes in old ages hav e a high renege of wall thickness loss 8 Figure 7: Contour Surface for Relationship between Input v ariables and Output Figure 8: Remaining Useful Life Prediction for Cast Iron Pipes. 9 Figure 9: Remaining Useful Life Prediction for Ductile Iron Pipes. Figure 10: Remaining Useful Life Prediction for Asbestos Cement Pipes. compared to the pipes in young ages. The results sho w with increasing 20% of wall thickness loss for pipes between 51 to 60 years old, the remaining useful life decreases 60% approximately . Similarly , with increasing 30% of wall thickness loss for pipes between 41-50 years old, the remaining useful life decreases 40% approximately . Figure 6 illustrates the relationship between wall thickness loss and remaining useful life in dataset in different ages for steel pipes. The results show that indeed pipes in old ages ha ve a high rene ge of wall thickness loss compared to the pipes in young ages. The results show with increasing 8% of w all thickness loss for pipes between 1 to 20 years old, the remaining useful life decreases 23% approximately . Similarly , with increasing 8% of wall thickness loss for pipes between 41-50 years old, the remaining useful life decreases 20% approximately . 7 V alidation and Contribution to the Body of Knowledge The deterioration models are determined with the most significant variables (age and wall thickness loss). V ariables are added into the non-linear multi-variable regression ( X 1 : age, X 2 : wall thickness loss and Y : remaining useful life). The regression models are selected based on high correlation with v ariables started from degree one and repeated the process with degree two and three to find the best correlation and high value of the coefficient of determination. T able 4 presents deterioration models for different types of w ater pipes. The major contributions of this paper are: • This paper predicts remaining useful life of water pipes using combination of Artificial Neural Networks (ANNs) and ANFIS. 10 Figure 11: Remaining Useful Life Prediction for Steel Pipes. T able 4: Deterioration Models for Different Pipes Material in the Dataset. Pipe Mat. Non-linear Regression Models R 2 CI Y = − 0 . 342 A 2 + 0 . 0548 W + 48 . 163 0.78 DI Y = 0 . 004 A 3 − 0 . 025 W 2 + 0 . 11 AW + 51 0.74 Ac Y = 0 . 0038 A 2 − 0 . 49 W + 195 . 92 0.80 Steel Y = 0 . 005 A 3 − 0 . 012 W 2 − 0 . 989 AW − 0 . 012 0.73 T able 5: * Where "A" is age of pipes, "W" is wall thickness loss , and "Y" is R UL. • Abov e models hav e not been used in previous literature for determination of remaining useful life of water pipes. 8 Conclusions and Limitations of this resear ch In this paper , we have analyzed the remaining useful life of water pipes. It is concluded that the applications of neural networks and Adapti ve Neural Fuzzy Inference System (ANFIS) to solve the problem of remaining useful life prediction of water pipes is feasible and the precision of the model depends on obtaining a larger and more comprehensiv e sample pipe dataset. Moreover , pipe age and wall thickness loss are the most significant parameters to predict the remaining useful life of the water pipes. Additionally , ductile iron and steel pipes hav e more remaining useful life compared to cast iron and asbestos-cement pipes. The a vailability of fe wer numbers of deterioration parameters and limited data av ailability posed the primary disadvantage to effecti ve neural network training and caused the main limitation to this paper . En vironmental parameters affecting the pipe, such as ov erburden pressure, soil type and properties, underground water -table location and other factors identified in the literature were omitted due to lack of monitoring of the data. Revie w of literature showed these parameters are suitable measures for prediction of remaining useful life. Since the dev eloped model does not include several parameters thought to be important to water deterioration, the model dev eloped in this paper is not complete. While it determines the utility of using Artificial Neural Networks and ANFIS models for predicting water condition, further w ork for data collection and model de velopment is required to confirm that the model is more precise and reliable for future applications. The av ailability of detailed soils parameters, water -table location, fluctuation, joint condition, leakage history , water pressure, installation depth, temperature, and water corrosi ve conditions w ould be resources to model the deterioration of water and precisely predict the remaining useful life. Moreov er , tho model does not apply to the new pipes with the age of zero and no wall thickness loss. The neural network and Adapti ve Neural Fuzzy Inference System (ANFIS) for water remaining useful life prediction models, can then be combined with an inclusive infrastructure asset management system to aid the municipal agencies in better planning and spending of their limited av ailable budget. 11 References [1] CS Feeney , S Thayer, M Bonomo, and K Martel. Condition assessment of waste water collection systems. En vironmental Pr otection Agency (EP A). W ashington, DC: US EP A , 2009. [2] Neil S Grigg, Darrell G F ontane, and Johan van Zyl. W ater distribution system risk tool for in vestment planning . W ater Research Foundation, 2013. [3] Y ehuda Kleiner and Balv ant Rajani. Comprehensive re view of structural deterioration of water mains: statistical models. Urban water , 3(3):131–150, 2001. [4] André Martins, João Paulo Leitão, and Conceição Amado. Comparati ve study of three stochastic models for prediction of pipe failures in water supply systems. J ournal of Infrastructur e Systems , 19(4):442–450, 2013. [5] Hesham Osman and Ke vin Bainbridge. Comparison of statistical deterioration models for water distribution networks. Journal of P erformance of Constructed F acilities , 25(3):259–266, 2010. [6] Franz Tscheikner-Gratl. Inte grated Appr oach for Multi-Utility Rehabilitation Planning of Urban W ater Infrastruc- tur e: F ocus on Small and Medium Sized Municipalities . innsbruck uni versity press, 2016. [7] Daniel W inkler , Markus Haltmeier , Manfred Kleidorfer, W olfgang Rauch, and Franz Tscheikner-Gratl. Pipe failure modelling for water distribution networks using boosted decision trees. Structur e and Infrastructur e Engineering , 14(10):1402–1411, 2018. [8] Andreas Scheidegger , Joao P Leitao, and Lisa Scholten. Statistical failure models for water distrib ution pipes–a revie w from a unified perspective. W ater resear ch , 83:237–247, 2015. [9] Lisa Scholten, Andreas Scheidegger , Peter Reichert, and Max Maurer . Combining expert kno wledge and local data for improv ed service life modeling of water supply networks. En vir onmental Modelling & Software , 42:1–16, 2013. [10] John Nicklo w , Patrick Reed, Drag an Savic, T ibebe Dessalegne, Laura Harrell, Amy Chan-Hilton, Mohammad Karamouz, Barbara Minsker , A vi Ostfeld, Abhishek Singh, et al. State of the art for genetic algorithms and beyond in w ater resources planning and management. Journal of W ater Resour ces Planning and Management , 136(4):412–432, 2009. [11] Dung H T ran, A WM Ng, and BJC Perera. Neural networks deterioration models for serviceability condition of buried stormw ater pipes. Engineering Applications of Artificial Intelligence , 20(8):1144–1151, 2007. [12] Symeon Christodoulou and Alexandra Deligianni. A neurofuzzy decision frame work for the management of w ater distribution netw orks. W ater resour ces management , 24(1):139–156, 2010. [13] Mohammad Najafi and Guru Kulandai vel. Pipeline condition prediction using neural network models. In Pipelines 2005: Optimizing Pipeline Design, Operations, and Maintenance in T oday’s Economy , pages 767–781. 2005. [14] Mohamed Fahmy and Osama Moselhi. Forecasting the remaining useful life of cast iron water mains. Journal of performance of constructed facilities , 23(4):269–275, 2009. [15] Hussam Fares and T arek Zayed. Hierarchical fuzzy expert system for risk of failure of water mains. Journal of Pipeline Systems Engineering and Practice , 1(1):53–62, 2010. [16] Peter D Rogers. Prioritizing water main renew als: case study of the denv er water system. Journal of Pipeline Systems Engineering and Practice , 2(3):73–81, 2011. [17] Alison M St. Clair and Sunil Sinha. State-of-the-technology re view on water pipe condition, deterioration and failure rate prediction models! Urban W ater Journal , 9(2):85–112, 2012. [18] Y ong W ang. Deterior ation and condition rating analysis of water mains . PhD thesis, Concordia Univ ersity , 2006. [19] Z Zangenehmadar and O Moselhi. Application of neural networks in predicting the remaining useful life of water pipelines. In Pipelines 2016 , pages 292–308. 2016. [20] Carmelo Del Coso, Die go Fustes, Carlos Dafonte, Francisco J Nóv oa, José M Rodríguez-Pedreira, and Bernardino Arcay . Mixing numerical and categorical data in a self-organizing map by means of frequenc y neurons. Applied Soft Computing , 36:246–254, 2015. [21] L yle John Nemeth. A comparison of risk assessment models for pipe replacement and rehabilitation in a water distribution system. 2016. [22] Douglas C Montgomery and Geor ge C Runger . Applied statistics and pr obability for engineers . John W iley & Sons, 2010. [23] Julian J Faraway . Extending the linear model with R: gener alized linear , mixed ef fects and nonparametric r e gression models . Chapman and Hall/CRC, 2016. 12 [24] Sirous F Y asseri, Roohollah B Mahani, et al. Remaining useful life (rul) of corroding pipelines. In The 26th International Ocean and P olar Engineering Confer ence . International Society of Of fshore and Polar Engineers, 2016. [25] Mohammad Najafi and Guru Kulandai vel. Pipeline condition prediction using neural network models. In Pipelines 2005: Optimizing Pipeline Design, Operations, and Maintenance in T oday’s Economy , pages 767–781. 2005. [26] W ayan Suparta and Kemal Maulana Alhasa. Modeling of tropospheric delays using ANFIS . Springer, 2016. [27] Mohammad M Ghiasi, Milad Arabloo, Amir H Mohammadi, and T ohid Barghi. Application of anfis soft computing technique in modeling the co2 capture with mea, dea, and tea aqueous solutions. International J ournal of Gr eenhouse Gas Contr ol , 49:47–54, 2016. [28] Razieh T avak oli et al. REMAINING USEFUL LIFE PREDICTION OF W ATER PIPES USING ARTIFICIAL NEURAL NETWORK AND AD APTIVE NEUR O FUZZY INFERENCE SYSTEM MODELS . PhD thesis, 2018. is enabled. References [1] Christodoulou, S., and Deligianni, A. (2010). A neurofuzzy decision framework for the management of water distribution netw orks. W ater resources management, 24(1), 139-156. [2] Clark, R. M., Stafford, C. L., and Goodrich, J. A. (1982). W ater distribution systems: A spatial and cost ev aluation. Journal of the W ater Resources Planning and Management Division, 108(3), 243-256. [3] Cortez, H. (2015). A Risk Analysis Model for the Maintenance and Rehabilitation of Pipes in a W ater Distribution System: A Statistical Approach. [4] Das, S., Nayak, B., Sarangi, S. K., and Biswal, D. K. (2016). Condition Monitoring of Rob ust Damage of Cantile ver Shaft Using Experimental and Adapti ve Neuro-fuzzy Inference System (ANFIS). Procedia Engineering, 144, 328-335. [5] Del Coso, C., Fustes, D., Dafonte, C., Nóvoa, F . J., Rodríguez-Pedreira, J. M., and Arcay , B. (2015). Mixing numerical and categorical data in a Self-Organizing Map by means of frequency neurons. Applied Soft Computing, 36, 246-254. [6] EP A, (2009), “Condition Assessment of W astew ater Collection Systems. ” March 2009, www .Ep.gov/nrmrl. [7] Fahmy , M., Moselhi, O. (2009). Forecasting the remaining useful life of cast iron water mains. Journal of performance of constructed facilities, 23(4), 269-275. [8] Fara way , J. J. (2016). Extending the linear model with R: generalized linear , mixed effects and non-parametric regression models. Chapman and Hall/CRC. [9] Fares, H., Zayed, T . (2010). Hierarchical fuzzy expert system for risk of failure of water mains. Journal of Pipeline Systems Engineering and Practice, 1(1), 53-62. [10] Ghiasi, M. M., Arabloo, M., Mohammadi, A. H., and Bar ghi, T . (2016). Application of ANFIS soft computing technique in modeling the CO2 capture with MEA, DEA, and TEA aqueous solutions. International Journal of Greenhouse Gas Control, 49, 47-54. [11] Grigg, N. S., Fontane, D. G., and v an Zyl, J. (2013). W ater distribution system risk tool for in vestment planning. W ater Research Foundation. [12] Kleiner , Y ., and Rajani, B. (2001). Comprehensiv e re view of structural deterioration of water mains: Statistical models. Urban W ater , 3(3), 131–150. [13] Kulandai vel, G. Sewer Pipeline Condition Prediction Using Neural Netw ork Models. Master’ Thesis, Michigan State University 2004, 10. [14] Lin, B. T ., and Huang, K. M. (2013). An adapti ve-network-based fuzzy inference system for predicting springback of U-bending. T ransactions of the Canadian Society for Mechanical Engineering, 37(3), 335-344. [15] Martins, A., Leitão, J.P ., and Amado, C. (2013). Comparati ve study of three stochastic models for prediction of pipe failures in water supply systems. Journal of Infrastructure Systems, 19(4), 442–450 [16] Montgomery , D. C., and Runger , G. C. (2010). Applied statistics and probability for engineers. John W iley and Sons. [17] Najafi, M., and Kulandai vel, G. (2005). Pipeline condition prediction using neural network models. In Pipelines 2005: Optimizing Pipeline Design, Operations, and Maintenance in T oday’ s Economy (pp. 767-781). 13 [18] Nemeth, L., (2016), “ A comparison of Risk Assessment Model for Pipe Replacement and Rehabilitation in water Distribution System. ” Master’ Thesis, California Polytechnic State Univ ersity , San Luis Obispo, 2016. [19] Nicklow , J., Reed, P ., Savic, D., Dessale gne, T ., Harrell, L., Chan-Hilton, A., Karamouz, M., Minsker , B., Ostfeld, A., Singh, A., Zechman, E. (2010). State of the art for genetic algorithms and beyond in w ater resources planning and management. Journal of W ater Resources Planning and Management, 136(4), 412–432. [20] Osman, H., and Bainbridge, K. (2011). Comparison of statistical deterioration models for water distribution networks. Journal of Performance of Constructed Facilities, 25(3), 259–266. [21] P aradkar , A. B. (2013). An Evaluation of F ailure Modes for Cast Iron and Ductile Iron W ater Pipes. [22] Prasojo, R. A., Diwyacitta, K., and Gumilang, H. (2017). T ransformer paper e xpected life estimation using ANFIS based on oil characteristics and dissolved gases (case study: Indonesian transformers). Energies, 10(8), 1135. [23] Rogers, P . D. (2011). Prioritizing water main rene wals: case study of the Den ver water system. Journal of Pipeline Systems Engineering and Practice, 2(3), 73-81. [24] Scheidegger , A., Leitão, J.P ., and Scholten, L. (2015). Statistical failure models for w ater distribution pipes – A revie w from a unified perspective. W ater Research, 83, 237–247. [25] Scholten, L., Scheidegger , A., Reichert, P ., and Maurer, M. (2013). Combining expert kno wledge and local data for improv ed service life modeling of water supply networks. En vironmental Modelling and Software, 42, 1–16. [26] St. Clair , A. M., and Sinha, S. (2012). State-of-the-technology revie w on water pipe condition, deterioration and failure rate prediction models!. Urban W ater Journal, 9(2), 85-112. [27] Suparta, W ., and Alhasa, K. M. (2016). Modeling of tropospheric delays using ANFIS. Springer International Publishing. [28] T av akoli, R. (2018). Remaining useful life prediction of water pipes using artificial neural network and adapti ve neuro fuzzy inference system models (Doctoral dissertation). [29] T ran, D.H., Ng, A.W .M., and Perera, B.J.C. (2007). Neural networks deterioration models for serviceability condition of buried storm w ater pipes. Engineering Applications of Artificial Intelligence, 20(8), 1144– 1151. [30] Tscheikner-Gratl, F . (2016). Integrated approach for multi-utility rehabilitation planning of urban water infrastruc- ture. PhD Thesis, Innsbruck Univ ersity Press. [31] W ang, Y . (2006). Deterioration and condition rating analysis of water mains (Doctoral dissertation, Concordia Univ ersity). [32] W inkler , D., Haltmeier, M., Kleidorfer , M., Rauch, W ., and Tscheikner-Gratl, F . (2018). Pipe failure modelling for water distribution networks using boosted decision trees. Structure and Infrastructure Engineering, 14(10), 1402-1411. [33] Y asseri, S. F ., and Mahani, R. B. (2016, June). Remaining useful life (R UL) of corroding pipelines. In The 26th International Ocean and Polar Engineering Conference. International Society of Offshore and Polar Engineers. [34] Zangenemadar , Z., and Moslehi, O. (2016, July). Application of Neural Networks in Predicting the Remaining Useful Life of W ater Pipelines. In Proceedings of the Pipelines 2016 Conference. 14