Water Research 92 (2016) 52e60

Contents lists available at ScienceDirect

Water Research
journal homepage: www.elsevier.com/locate/watres

Predicting concrete corrosion of sewers using artificial neural network

Guangming Jiang*, Jurg Keller, Philip L. Bond, Zhiguo Yuan
Advanced Water Management Centre, The University of Queensland, St. Lucia, Queensland 4072, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Corrosion is often a major failure mechanism for concrete sewers and under such circumstances the
Received 12 November 2015 sewer service life is largely determined by the progression of microbially induced concrete corrosion. The
Received in revised form modelling of sewer processes has become possible due to the improved understanding of in-sewer
4 January 2016
transformation. Recent systematic studies about the correlation between the corrosion processes and
Accepted 13 January 2016
sewer environment factors should be utilized to improve the prediction capability of service life by sewer
Available online 21 January 2016
models. This paper presents an artificial neural network (ANN)-based approach for modelling the con-
crete corrosion processes in sewers. The approach included predicting the time for the corrosion to
Keywords:
Sewer
initiate and then predicting the corrosion rate after the initiation period. The ANN model was trained and
Corrosion validated with long-term (4.5 years) corrosion data obtained in laboratory corrosion chambers, and
Concrete further verified with field measurements in real sewers across Australia. The trained model estimated the
Hydrogen sulfide corrosion initiation time and corrosion rates very close to those measured in Australian sewers. The ANN
Artificial neural network model performed better than a multiple regression model also developed on the same dataset. Addi-
Multiple regression model tionally, the ANN model can serve as a prediction framework for sewer service life, which can be pro-
gressively improved and expanded by including corrosion rates measured in different sewer conditions.
Furthermore, the proposed methodology holds promise to facilitate the construction of analytical models
associated with corrosion processes of concrete sewers.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction To alleviate and control corrosion problems in concrete sewers,

various technologies have been employed to remove or reduce
Sewer networks established through continuous public invest- hydrogen sulfide levels in the sewage or sewer air. This includes
ment are important infrastructure supporting modern urban living. removal from sewage by dosing with chemicals such as nitrates or
With estimated values of one trillion dollars in the USA and $35 iron salts to reduce the formation or emission of H2S (Gutierrez
billion in Australia these are valuable assets, however, sewers are et al., 2008; Jiang et al., 2011, 2013; Jiang and Yuan, 2013; Zhang
degrading at a very high rate, in many cases due to microbially et al., 2009) or removal of H2S from sewer air (Sivret and Stuetz,
induced concrete corrosion (Brongers et al., 2002). The annual total 2010). Other technologies include construction of new sewers
cost of corrosion within the water and wastewater infrastructure is with corrosion-resistant pipe materials or to repair corroded con-
reported to be US$ 36 billion in USA (Koch et al., 2002). Corrosion crete surfaces with corrosion resistant layers (De Muynck et al.,
causes the loss of concrete mass and structural capacity, eventually 2009; Haile et al., 2010; Hewayde et al., 2007).
leading to ultimate structural collapse. The rehabilitation and It is thus critical to accurately estimate the sewer service life for
replacement of damaged sewers is costly, in the USA alone this was the purpose of planning sewer maintenance and rehabilitation and
estimated to be around $14 billion per year (Brongers et al., 2002). for evaluating and optimizing the effectiveness of various control
This cost is expected to increase as the aging infrastructure con- technologies. The most direct indicator for the effectiveness of
tinues to fail (Alexander et al., 2013; Boulos and Walker, 2015; these various technologies should be the sewer service life (L, year),
Sydney et al., 1996; US EPA, 1991). which is determined by the time for corrosion to initiate (ti, month)
and the corrosion rate (r, i.e. concrete depth lost over time, mm/
year) (Eq. (1)).
* Corresponding author.
E-mail addresses: [email protected] (G. Jiang), [email protected]
(J. Keller), [email protected] (P.L. Bond), [email protected]
(Z. Yuan).

http://dx.doi.org/10.1016/j.watres.2016.01.029
0043-1354/© 2016 Elsevier Ltd. All rights reserved.
 G. Jiang et al. / Water Research 92 (2016) 52e60 53

modelling tool (Hsu et al., 1995). ANN models ‘learn’ the patterns of
Nomenclature the underlying process from past data and generalize the gained
‘knowledge’ (or mathematical relationships between input and
ANN Artificial Neural Network output data) to predict an output given a new set of input variables
GP Gas-phase from the problem domain. It has the potential to predict any
MICC Microbially induced concrete corrosion complex system with high precision provided its architecture and
MR Multiple regression parameters are properly set. ANN models have found many
PS Partially-submerged different applications including data mining, pattern recognition,
SOB Sulfide-oxidizing bacteria clustering analysis and predictive modelling. In the field of corro-
RH Relative humidity sion, ANN is successfully used for the prediction of carbon steel
corrosion when exposed to the atmosphere or to crude oil (Kenny
et al., 2009; Pintos et al., 2000; Rebak, 2005).
This study proposes a data-driven approach based on ANN for
the estimation of both corrosion initiation time (ti) and corrosion
ti D rate (r). As discussed, the two parameters vary from site to site due
L¼ þ (1) to the local environmental conditions. Our previous laboratory in-
12 r
vestigations cover the development and established stages of sewer
Where D is the concrete depth (mm) that can be sacrificed before corrosion, and provide extensive data evaluating the effect of three
the end of sewer service life. This bilinear corrosion model was controlling factors: H2S concentration, gas temperature and rela-
proposed based on sewer concrete losses measured in the field tive humidity (Jiang et al., 2014a, 2015). Those studies also inves-
(Wells and Melchers, 2014). However, the model was not param- tigated the effect of locations within the sewer on corrosion,
eterized for application to different sewer conditions due to it being exposing concrete to either the sewer atmosphere (simulating the
based on very limited field data. crown) or by partially-submerging it in sewage (simulating the
Existing sewer process models mainly focused on the modelling sewer tidal region at the sewage/air interface). Consequently, we
of hydrogen sulfide production and wastewater characteristics in utilized this extensive dataset to build ANN models to predict ti and
sewer networks (Hvitved-Jacobsen et al., 2013; Sharma et al., 2014, r, which can be used to estimate the service life for a specific sewer.
2013, 2008; Vollertsen et al., 2008; Vollertsen et al., 2011, 2015). The performance and application of the proposed ANN model was
These models lack a thorough approach for the estimation of further evaluated by comparison with a classical regression model
corrosion rate based on the simulated profile of hydrogen sulfide and with observations in real sewer sites across Australia.
and other environmental factors. The well-known Pomeroy model
was used to calculate the deterioration rate of concrete sewer pipes 2. Material and methods
(Pomeroy, 1990). This empirical model was widely used although it
fails to take into account recent findings of the corrosion process 2.1. Long-term corrosion tests in laboratory chambers
and associated impacting factors. It is recently shown that both the
corrosion initiation time and corrosion rate depend on various Thirty-six identical corrosion chambers were constructed to
sewer environmental factors that include the H2S concentration, achieve controlled environments simulating that of real sewers
relative humidity and temperature (Jiang et al., 2014a, 2015). (Jiang et al., 2014a). The controlled factors include combinations of
Additionally, it was recently discovered that the corrosion devel- three gas-phase temperatures (17  C, 25  C and 30  C), two levels of
opment can be facilitated by internal cracking which is caused by relative humidity (RH) (100% and 90%) and six H2S levels (0 ppm,
the formation of corrosion products that include iron oxides 5 ppm, 10 ppm, 15 ppm, 25 ppm and 50 ppm). Each chamber
precipitating in concrete (Jiang et al., 2014b; Monteny et al., 2000; contained 2.5 L of domestic sewage that was collected from a local
Parande et al., 2006). sewer pumping station and replaced every two weeks (Fig. 1).
Another hurdle to the implementation of Eq. (1) is the absence Concrete coupons were prepared from fresh and precorroded
of a model estimating the initiation time for corrosion (ti). The sewer pipes to investigate the corrosion initiation and the active
corrosion initiation involves a combination of physical, chemical corrosion respectively. The fresh concrete coupons were prepared
and biological processes. The transition of the highly alkaline fresh from a new sewer pipe (1.2 m diameter and 0.07 m thickness)
concrete surface of the pipe to a more microbe habitable environ- obtained from a sewer pipe manufacturer (HUMES, Sydney,
ment occurs by acidification through carbonation (by CO2) and H2S Australia). The precorroded coupons were prepared from corroded
dissociation during the initial stages of corrosion (Joseph et al., concrete sewer slabs obtained from Sydney Water Corporation,
2012). This is facilitated by the physicochemical conditions in Australia. One of the original surfaces of the coupons, i.e. the in-
gravity sewers through adequate water availability (due to high ternal surface of the pipe, was designated as the surface to be
relative humidity), high concentrations of carbon dioxide, high exposed to H2S. Part of the coupons were embedded as pairs (1
concentrations of H2S and lowered surface pH (Wei et al., 2013). As fresh þ 1 precorroded) in stainless steel frames using epoxy (FGI
the surface pH lowers towards neutrality various sulfide-oxidizing R180 epoxy & H180 hardener) with the steel frame providing a
bacteria (SOB) will populate the corrosion layer and produce sul- reference point for determining the change in thickness due to
furic acid (Cayford et al., 2012; Okabe et al., 2007; Santo Domingo corrosion (Jiang et al., 2014b).
et al., 2011). For each chamber, eight pairs of enclosed coupons were exposed
Recently a model was developed with H2S concentration as the to the gas phase. These gas-phase concrete coupons represent the
key parameter and this was validated with laboratory data for the sewer pipe crown, a location which is highly susceptible to sulfide
concrete corrosion of the crown area in sewer pipes (Jiang et al., induced corrosion (Mori et al., 1992; Vollertsen et al., 2008).
2014a). However, due to the difficulty in using deterministic Another eight pairs of bare coupons (8 fresh þ 8 precorroded) were
models to simulate the complex processes involved in corrosion, partially submerged (approx. 20e30 mm) in the wastewater
black-box type models can be used to obtain predictions with simulating the concrete sewer pipe near the water level, which is
reasonable accuracy. The artificial neural network (ANN) modelling also a region of high corrosion activity.
approach, inspired by biological neural systems, is an effective The corrosion chambers were operated for up to 54 months
54 G. Jiang et al. / Water Research 92 (2016) 52e60

Fig. 1. Side view of a corrosion chamber with H2S concentration, relative humidity and gas temperature controlled by PLC. The orientation of the gas-phase and partially-submerged
coupons is shown.

since December 2009. Periodically, at intervals between 6 and 10 significant contributing factors (P < 0.05) were excluded in a step
months, one set of coupons (one pair of gas-phase coupons and one by step manner to achieve a minimum adequate model. The
pair of partially-submerged coupons) were retrieved from each regression models were then used to predict corrosion initiation
corrosion chamber for detailed analysis. A standard step-by-step times and corrosion rates for the conditions in the laboratory and
procedure of methods was employed to measure surface pH, fol- field sewers, As well, some sewer examples from the literature
lowed by sampling for sulfur species and then photogrammetry were used for these predictions. The performance of the regression
analysis (thickness change). The detailed sampling and analysis models were compared to the ANN based model outputs.
procedures were described previously (Jiang et al., 2014a).
The corrosion initiation time, ti, was determined as the time to 2.4. Artificial neural network modelling
reach a detectable level of sulfate on the fresh concrete surface
assuming a linear increase of sulfate production with time (Jiang Basic structure of an artificial neural network model is
et al., 2015). The mass loss data of precorroded coupons were composed of three layers, i.e. input, hidden, and output layer. The
used to calculate the corrosion rate as the thickness change per neural network modelling process used in this study may be
year, i.e. mm/year. described in four steps: (i) pre-processing of the original data set
(determination of ti and r from experimental data and identification
2.2. Long-term corrosion tests in real sewers of outliers); (ii) partitioning of the pre-processed data set into
learning, validation and test sets; (iii) ANN model architecture
Concrete corrosion was studied in real sewer systems to quan- setting, learning and testing; and (iv) validation using field data.
tify corrosion occurring in working Australian sewers under a range In the first stage the pre-processed data set was sorted by
of environmental and operating conditions. Six sewer sites were experimental factors (the location of concrete coupons, H2S con-
chosen in three cities, i.e. Sydney, Melbourne and Perth in Australia. centration, gas temperature and relative humidity) and the
The choice of three cities and two sites in each city enabled the learning, validation, and test data were constructed using a
study of different temperatures and different H2S levels. Concrete randomization procedure. The percentage of observations per data
coupons were fixed into the sewers as described (Wells and set were assigned to be 70%, 15% and 15%, for the learning, valida-
Melchers, 2015) and then recovered approximately every 6 tion and test sets, respectively. The learning data set was used to
months in the early stages of the trial and yearly in the later stages. train the ANN. The validation data set was used in conjunction with
The retrieved coupons were analyzed similarly to the laboratory the learning data set to determine when to stop the learning pro-
coupons for measurement of surface pH, sulfur compounds in the cess such that the resulting model exhibited good generalization
corrosion layer and the mass loss due to corrosion. Details of the properties. The test data set allows the assessment of the prediction
study were reported by Wells and Melchers (2015) and the corro- capabilities of the ANN model. The ANN was evaluated using as
sion initiation time and corrosion rate determined in real sewers performance criteria mean square error and residual distributions
were used to validate the ANN model in this study. over the learning and test data sets.
Step 3 involves the ANN architecture setting and optimization.
2.3. Statistical data analysis The ANN model we designed with three layers: one input layer, one
hidden layer and one output layer. Generally speaking one hidden
Statistical Analysis of the data was performed using R (ver 3.03, layer should be suffice for most practical problems, thus only one
http://www.R-project.org/). Multiple linear regression analysis was hidden layer was used in this study. This was verified by comparing
performed on the corrosion initiation time and corrosion rate, with the performance against models with more than one hidden layer.
explanatory factors including H2S concentration, relative humidity The nodes for the ANN input and output layer were set by the
and temperature. The coefficients for each of the explanatory fac- number of input factors (one category and three numerical, shown
tors were determined together with the standard error. The non- in Fig. 2) and the number of variables to be predicted (ti or r),
 G. Jiang et al. / Water Research 92 (2016) 52e60 55

Fig. 2. The network architecture of the ANN model for the prediction of ti.

respectively. The input layer includes four nodes that represent the captured and explained by this linear model. The fairly low R2 value
four influencing factors of sewer corrosion environment: location suggests that the relationship between the predictors and ti is
of the sewer pipe, H2S concentration, gas temperature and relative unlikely linear.
humidity. The number of nodes in the hidden layer was established Following the regression analysis, an ANN model was used to
before the ANN model architecture was completed. An exhaustive analyze the same dataset. The final structure of the ANN model has
search was conducted to determine the optimal number of neurons 8 neurons in the hidden layer (Fig. 2). The activation functions for
in the hidden layer using Alyuda NeuroIntelligence ver 2.2. The best the hidden and output layers of the ANN model were hyperbolic
architecture was then constructed for the training and validation tangent and logistic function, respectively. Sum of squares was used
analysis using Matlab R2014a. as the error function for the output layer. The training process was
conducted using the standard backpropagation algorithm as the
optimization procedure, with weights updated each time the
3. Results and discussion complete training data set was considered.
The ANN model obtained a good relationship from the training
3.1. Prediction of the initiation of corrosion e ti data, with an R value of 0.8291 (Fig. 3A), and it performed relatively
well in the validation and testing, although there is a high level of
An equation was generated from the multiple regression (MR) scattering of the data and likely some outliers (Fig. 3B and C).
analysis to predict the concrete corrosion initiation time (Eq. (2)). Overall, the ANN performs satisfactorily for the whole laboratory
This assumes the corrosion initiation time is linearly dependent on data set (Fig. 3D). It must be realized that this ANN model did not
the explanatory variables, i.e. the location of concrete, H2S con- consider other environmental factors which may affect ti values.
centration, RH and temperature (T). These factors may include fluctuations of the three controlling
factors and variability in different concrete, such as aggregate
ti ¼ 96:34 þ 1:68*Location  0:18*H2 S  0:54*RH  0:84*T levels. Also, the dataset supporting the ANN model is limited to the
(2) conditions investigated in the laboratory corrosion chambers. Due
to its data-driven nature, the ANN model can be improved pro-
The uncertainty and significance of the regression coefficients
gressively by training it with more observed data.
were determined and all explanatory factors were significant as
After developing the ANN model to predict ti based upon the
evident by the P values for each coefficient (Table 1). The intercept
laboratory data, a further step was carried out to validate its per-
value of around 96 months, indicates the maximum theoretical
formance using field data. The corrosion initiation time ti measured
initiation time is around 8 years. Location contributes about 3.4
for all the field sites, including two Perth sewers and two Mel-
months of difference to the ti, with gas phase coupons more
bourne sewers, varied from site to site but were in the range of
resistant to the initiation of corrosion. Similarly, one unit increase of
9e24 months. Also, we compared the predictions of ti for the field
H2S concentration (ppm), RH (%) and temperature ( C) causes re-
sites between the ANN model and the multiple regression equation
ductions of about 0.18, 0.54 and 0.84 months to the initiation time
(Eq. (1)) (Fig. 4). In comparison between the predicted ti and the
respectively. The regression results align with previous observation
measured ti for the four field sites it is clear that the ANN model
and analysis of the corrosion development (Jiang et al., 2015). The
achieved reasonable accuracy for the prediction of ti, while the
R2 value obtained for the multiple regression was 0.54, which im-
multiple regression model failed to give reasonable estimates
plies that only 54% of the variability in the observed ti could be
(Fig. 4). Some conditions at the field sites were far beyond the
ranges for those in the laboratory corrosion chambers (Wells and
Table 1 Melchers, 2015). In particular, the Perth sewer site had very high
The coefficients for the multiple regression analysis of corrosion initiation data. H2S concentrations, up to 830 ppm, and high temperatures, and in
Coefficient Estimate Std. Error t value P(>jtj)a Significance b that situation the MR model predicted a negative ti.
8 Compared to the designed lifespan of the sewer pipes (50 years
Intercept 96.34 15.03 6.41 3.49  10 ***
Location 1.68 0.77 2.19 3.25  102 * or more), ti does not constitute a significant length of time and
H2S 0.18 0.05 3.83 3.34  104 *** consequently might be ignored when calculating the service life-
RH 0.54 0.15 3.51 9.13  104 *** span of a sewer pipe. However, ti is important when a prevention
Temperature 0.84 0.14 5.83 3.04  107 *** strategy, such as sewer gas ventilation and chemical dosing in
a
P(>jtj) is the probability value using the t-test. sewage, is in place to prevent the initiation of corrosion. For
b
Significance codes represent P values of: 0e0.001:***; 0.001e0.01:**; example, the CO2 levels can be greatly reduced by ventilation from
0.01e0.05:*.
56 G. Jiang et al. / Water Research 92 (2016) 52e60

Fig. 3. Outputs of the ANN model using training (A), validation (B), test (C) and the complete (D) data sets for the corrosion initiation time. The concrete corrosion initiation times
were determined from coupons exposed to conditions simulating environments in sewers using laboratory corrosion chambers.

would also be desirable to extend ti during the operation of new

sewer systems by controlling the sewer environmental factors. The
service time of the sewer can be extended by delaying the initiation
of corrosion.

3.2. Prediction of corrosion rate e r

The linear equation for corrosion rate generated by multiple

regression was determined (Eq. (3)). However, the coefficients for
RH and temperature are shown to be insignificant (Table 2). Further
analysis using back selection was carried out to obtain the mini-
mum adequate model (Eq. (4)). From the regression coefficients for
location (0.45), it can be inferred that the difference of corrosion
rates for a sewer pipe in the gas phase (crown) and the partially
submerged (water line) area is about 0.9 mm/year. Also, an increase
of H2S by 10 ppm will cause about 0.28 mm/year increase of
corrosion rate. However, it is evident this regression model over-
predicts corrosion rates when the H2S concentration is low, e.g.
0.6e1.5 mm/year of corrosion is estimated when H2S is at 0 ppm.
The R2 value obtained for the multiple regression was 0.59, which
implies that only 59% of the variability in the observed r was
Fig. 4. Validation of the ANN and multiple regression (MR) model using corrosion captured and explained by this linear model.
initiation time observed in real sewers.
r ¼ 0:63  0:45*Location þ 2:82  102 *H2 S þ 8:69

0.2 to 1.2% in sewers to the atmospheric level (0.04%) and thus the  103 *RH þ 1:57  102 *T (3)
carbonation of concrete. The prediction capacity of ti can be used to
evaluate and optimize those corrosion prevention strategies. It r ¼ 1:03  0:45*Location þ 2:82  102 *H2 S (4)
 G. Jiang et al. / Water Research 92 (2016) 52e60 57

Table 2
The coefficients for the multiple regression analysis of corrosion rate data.
a b
Coefficients Estimate Std. Error t value P(>jtj) Significance

Full model
Intercept 0.17 1.26 0.138 0.891
Location 0.45 0.064 7.046 1.25  109 ***
H2S 2.82  102 3.89  103 7.252 5.36  1010 ***
RH 8.69  103 0.013 0.676 0.501
Temperature 1.57  102 0.012 1.31 0.195
Minimum adequate model
(Intercept) 1.03 0.0938 10.982 <2  1016 ***
Location 0.45 0.0644 7.037 1.14  109 ***
H2S 2.82  102 0.0039 7.243 4.82  1010 ***
a
P(>jtj) is the probability value using the t-test.
b
Significance codes represent P values: 0e0.001:***; 0.001e0.01:**; 0.01e0.05:*.

To improve the prediction of corrosion rate an ANN model was model was further validated using corrosion rates measured in four
developed in a similar approach used for the prediction of corrosion Australian sewer sites and those in literature reviewed by Wells and
initiation (section 3.1). The best architecture determined by Melchers (2015). The ANN model demonstrated accurate pre-
exhaustive searching, based on R-square, has one input layer (4 dictions of corrosion rates for most sewer sites, while under-
inputs including location, H2S concentration, RH and temperature), predicting the corrosion rates of two sites (Fig. 6). The under-
one hidden layer (9 nodes), and one output layer for the output predicted sites had corrosion rates of up to 11 mm/year, which is
variable of corrosion rate (r). The model was trained using the 72 nearly 3 times the maximum rates measured in the laboratory
sets of corrosion data obtained in the laboratory corrosion chamber. corrosion chambers. Likely, those low predictions occurred because
The ANN model showed excellent performance in both training, the ANN was trained solely on the laboratory data. To improve the
validation and test, with an overall R ¼ 0.9119 (Fig. 5D). This in- ANN model, the field and literature dataset was included in the
dicates that the model established a clear relationship between the model training process. It is clear this significantly improved the
environmental factors and the corrosion rate. The trained ANN ANN's capability in predicting the higher corrosion rates (Fig. 6).

Fig. 5. The outputs of the ANN model using training (A), validation (B) and test (C), and the complete (D) data sets of corrosion rates. The concrete corrosion rates were determined
from coupons exposed to conditions simulating environments in sewers using laboratory corrosion chambers.
58 G. Jiang et al. / Water Research 92 (2016) 52e60

The MR model did show promising performance especially for

corrosion rates lower than 4 mm/year. Thus, in certain circum-
stances it might be very useful to use such a simple model (Eq. (4))
to make reasonable predictions of corrosion rates. However, the
ANN model is a more sophisticated predictive tool with higher
accuracy that can be applied to a wider range of conditions.

3.3. Sensitivity and uncertainty of the prediction

Sensitivity analysis conducted on the trained neural network

revealed the effects of sewer environmental factors on both ti and r
(Fig. 7AeF). The neural network models, established in section 3.1
and 3.2, separated the effects of the variables and simulated the
effects of each individual factor (Fig. 7AeF). In most cases nonlinear
responses were evident, except for the semi-linear change of ti to
RH and temperature (Fig. 7B and C). Also, it was seen that the
sensitivity of ti to environmental factors are in the order of
H2S > Temperature > RH, for the range of factors used in the lab
corrosion chambers. The corrosion rate, it seems was most sensitive
to the H2S concentration (Fig. 7D). These sensitivity analyses
confirm previous findings identifying controlling factors for the
initiation period and for the active corrosion stage (Jiang et al.,
2014a, 2015).
Fig. 6. Validation of the ANN and multiple regression (MR) models using corrosion
Additionally, the sewer corrosion at the two different locations
rates observed in real sewers. The ANN predictions were improved by including
corrosion rates observed in real sewers in the model training process.

Fig. 7. Sensitivity of corrosion initiation times (A, B and C) and corrosion rates (D, E and F) to H2S concentration, RH and temperature for the ANN models described in sections 3.1 and 3.2.
 G. Jiang et al. / Water Research 92 (2016) 52e60 59

(gas phase (GP) and partially submerged (PS)) showed different  The ANN model trained with laboratory data provided better
levels of sensitivity to all sewer environmental factors. The corro- estimations to field measurements in real sewers than the
sion initiation of GP sewer is more sensitive to the change of multiple regression model. Although, the multiple regression
environmental factors while PS sewer is barely affected by RH and model can be used as a simple and quick tool to estimate
temperature. This reflects that submerged coupons were not so corrosion rate.
much affected by humidity levels because of their proximity to  The ANN model can be used to understand the corrosion
water. Also, the corrosion rates of PS sewer showed much higher mechanisms and be further improved by including more
sensitivity to H2S than that of GP sewer. Possibly, this is resulting affecting factors or training with corrosion data under broader
from PS sewer having higher maximum corrosion rates that are conditions including H2S concentration, humidity and
influenced by the availability of nutrients and water or the high temperature.
wash-off of corrosion products due to the shear force of
wastewater. Acknowledgments
The results of the sensitivity analysis reveal details of the un-
derlying mechanisms of sewer corrosion. For sewer corrosion, The authors acknowledge the financial support provided by the
primarily caused by acid production through sulfide oxidation, the Australian Research Council and many members of the Australian
reaction kinetics are usually expressed as an exponential function water industry through LP0882016 the Sewer Corrosion and Odour
(y ¼ k$xn ), where the exponent value relates to the long-term Research (SCORe) Project (www.score.org.au). Dr Guangming Jiang
corrosion. When the corrosion is controlled by H2S diffusion is the recipient of a Queensland State Government's Early Career
through the corrosion product, the value of the exponent is 0.5. Accelerate Fellowship.
From the ANN predictions, the values of n are 0.20 ± 0.01 and
0.53 ± 0.04 for GP and PS sewers respectively (Fig. 6D). This sug- References
gests that the corrosion products on PS sewer act as a form of
diffusion barrier. The extent of corrosion on GP sewer is lower and Alexander, M.G., Bertron, A., De Belie, N., 2013. Performance of Cement-based
thus the diffusion limitation is less obvious. Materials in Aggressive Aqueous Environments.State-of-the-art Report.
Springer. RILEM TC 211 e PAE.
There are still considerable unexplained variances in the neural Boulos, P.F., Walker, A.T., 2015. Fixing the future of wastewater systems with smart
network predictions. Such unexplained variances are mainly network modeling. J. Am. Water Works Assoc. 107 (4), 72e80.
attributable to two reasons: ignorance of other important affecting Brongers, M.P.H., Virmani, P.Y., Payer, J.H., 2002. Drinking Water and Sewer Systems
in Corrosion Costs and Preventative Strategies in the United States. United
variables and inherent scattering in the data. There are some po-
States Department of Transportation Federal Highway Administration.
tential affecting variables that have not been considered in the ANN Cayford, B.I., Dennis, P.G., Keller, J., Tyson, G.W., Bond, P.L., 2012. High-throughput
analysis. For example, our recent study indicates that, in compari- amplicon sequencing reveals distinct communities within a corroding concrete
son to even concentrations of gaseous H2S, daily fluctuating levels sewer system. Appl. Environ. Microbiol. 78 (19), 7160e7162.
De Muynck, W., De Belie, N., Verstraete, W., 2009. Effectiveness of admixtures,
that occur in real sewers can affect the H2S oxidation pathway and surface treatments and antimicrobial compounds against biogenic sulfuric acid
the total amount of H2S uptake by the sewer concrete (Sun et al., corrosion of concrete. Cem. Concr. Compos. 31 (3), 163e170.
2015). Another possible factor not considered is the periodic inoc- Gutierrez, O., Mohanakrishnan, J., Sharma, K.R., Meyer, R.L., Keller, J., Yuan, Z., 2008.
Evaluation of oxygen injection as a means of controlling sulfide production in a
ulation of the concrete surface that occurs in real sewers by sewage sewer system. Water Res. 42 (17), 4549e4561.
due to changing water levels (e.g. flooding) and splashing. This Haile, T., Nakhla, G., Allouche, E., Vaidya, S., 2010. Evaluation of the bactericidal
could accelerate the initiation of corrosion and corrosion rate on characteristics of nano-copper oxide or functionalized zeolite coating for bio-
corrosion control in concrete sewer pipes. Corros. Sci. 52 (1), 45e53.
frequently wetted concrete. Additionally, different sewer pipes are Hewayde, E.H., Nakhla, G.F., Allouche, E.N., Mohan, P.K., 2007. Beneficial impact of
made of different cements and aggregates with different water to coatings on biological generation of sulfide in concrete sewer pipes. Struct.
cement ratios. Therefore, the concrete properties and other vari- Infrastruct. Eng. 3 (3), 267e277.
Hsu, K.-l., Gupta, H.V., Sorooshian, S., 1995. Artificial neural network modeling of
ables could be included for expansion of the ANN model.
the rainfall-runoff process. Water Resour. Res. 31 (10), 2517e2530.
To obtain reliable predictions, the use of the ANN model should Hvitved-Jacobsen, T., Vollertsen, J., Nielsen, A.H., 2013. Sewer Processes: Microbial
be limited to the range of the factors used to train the model, i.e. and Chemical Process Engineering of Sewer Networks, second ed. CRC Press.
Jiang, G., Gutierrez, O., Sharma, K.R., Keller, J., Yuan, Z., 2011. Optimization of
H2S gaseous concentration between 0 and 50 ppm, RH between
intermittent, simultaneous dosage of nitrite and hydrochloric acid to control
90% and 100% and temperature between 17  C and 30  C. Never- sulfide and methane production in sewers. Water Res. 45 (18), 6163e6172.
theless, the ANN model has a strong capability of scatter-tolerance Jiang, G., Keller, J., Bond, P.L., 2014a. Determining the long-term effects of H2S
and it is expected that the prediction accuracy of the model will concentration, relative humidity and air temperature on concrete sewer
corrosion. Water Res. 65, 157e169.
increase with the consideration of more affecting factors and the Jiang, G., Sharma, K.R., Yuan, Z., 2013. Effects of nitrate dosing on methanogenic
inclusion of more data in the model training process. activity in a sulfide-producing sewer biofilm reactor. Water Res. 47 (5),
1783e1792.
Jiang, G., Sun, X., Keller, J., Bond, P.L., 2015. Identification of controlling factors for
the initiation of corrosion of fresh concrete sewers. Water Res. 80, 30e40.
4. Conclusions Jiang, G., Wightman, E., Donose, B.C., Yuan, Z., Bond, P.L., Keller, J., 2014b. The role of
iron in sulfide induced corrosion of sewer concrete. Water Res. 49 (0), 166e174.
Jiang, G., Yuan, Z., 2013. Synergistic inactivation of anaerobic wastewater biofilm by
This study developed a predicting tool for the estimation of the free nitrous acid and hydrogen peroxide. J. Hazard. Mater. 250, 91e98.
service life of concrete sewers based on the modelling of sewer Joseph, A.P., Keller, J., Bustamante, H., Bond, P.L., 2012. Surface neutralization and
corrosion through an artificial neural network. The ANN model was H2S oxidation at early stages of sewer corrosion: influence of temperature,
relative humidity and H2S concentration. Water Res. 46 (13), 4235e4245.
trained and validated with long-term (4.5 years) corrosion data zaris, J., 2009.
Kenny, E.D., Paredes, R.S.C., de Lacerda, L.A., Sica, Y.C., de Souza, G.P., La
obtained in laboratory corrosion chambers, and further verified Artificial neural network corrosion modeling for metals in an equatorial
with field measurements in real sewers across Australia. This has climate. Corros. Sci. 51 (10), 2266e2278.
Koch, G.H., Brongers, M.P.H., Thompson, N.G., Virmani, Y.P., Payer, J.H., 2002.
led to the following key findings:
Corrosion Costs and Preventive Strategies in the United States. NACE
International.
 Artificial neural network performs well in predicting the time Monteny, J., Vincke, E., Beeldens, A., De Belie, N., Taerwe, L., Van Gemert, D.,
for corrosion to initiate on sewers and the corrosion rate based Verstraete, W., 2000. Chemical, microbiological, and in situ test methods for
biogenic sulfuric acid corrosion of concrete. Cem. Concr. Res. 30 (4), 623e634.
on four parameters related to the sewer environmental Mori, T., Nonaka, T., Tazaki, K., Koga, M., Hikosaka, Y., Noda, S., 1992. Interactions of
conditions. nutrients, moisture and pH on microbial corrosion of concrete sewer pipes.
60 G. Jiang et al. / Water Research 92 (2016) 52e60

Water Res. 26 (1), 29e37. Sun, X., Jiang, G., Bond, P.L., Keller, J., 2015. Impact of fluctuations in gaseous H2S
Okabe, S., Odagiri, M., Ito, T., Satoh, H., 2007. Succession of sulfur-oxidizing bacteria concentrations on sulfide uptake by sewer concrete: the effect of high H2S
in the microbial community on corroding concrete in sewer systems. Appl. loads. Water Res. 81 (0), 84e91.
Environ. Microbiol. 73 (3), 971e980. Sydney, R., Esfandi, E., Surapaneni, S., 1996. Control concrete sewer corrosion via the
Parande, A.K., Ramsamy, P.L., Ethirajan, S., Rao, C.R.K., Palanisamy, N., 2006. Dete- crown spray process. Water Environ. Res. 68 (3), 338e347.
rioration of reinforced concrete in sewer environments. Proc. Inst. Civ. Eng.- US EPA, 1991. Hydrogen Sulphide Corrosion in Wastewater Collection and Treat-
Munic. Eng. 159 (1), 11e20. ment System (Technical Report).
Pintos, S., Queipo, N.V., Troconis de Rinco n, O., Rinco
 n, A., Morcillo, M., 2000. Vollertsen, J., Nielsen, A.H., Jensen, H.S., Wium-Andersen, T., Hvitved-Jacobsen, T.,
Artificial neural network modeling of atmospheric corrosion in the MICAT 2008. Corrosion of concrete sewers-The kinetics of hydrogen sulfide oxidation.
project. Corros. Sci. 42, 35e52. Sci. Total Environ. 394 (1), 162e170.
Pomeroy, R.D., 1990. The Problem of Hydrogen Sulphide in Sewers. Clay Pipe Vollertsen, J., Nielsen, L., Blicher, T.D., Hvitved-Jacobsen, T., Nielsen, A.H., 2011.
Development Association Limited, London. A sewer process model as planning and management tool e hydrogen sulfide
Rebak, R.B., 2005. Use of Artificial Neural Networks for Predicting Crude Oil Effect simulation at catchment scale. Water Sci. Technol. 64 (2), 348e354.
on CO2 Corrosion of Carbon Steels, pp. 1e13. Vollertsen, J., Revilla, N., Hvitved-Jacobsen, T., Nielsen, A.H., 2015. Modeling sulfides,
Santo Domingo, J.W., Revetta, R.P., Iker, B., Gomez-Alvarez, V., Garcia, J., Sullivan, J., pH and hydrogen sulfide gas in the sewers of San Francisco. Water Environ. Res.
Weast, J., 2011. Molecular survey of concrete sewer biofilm microbial commu- 87 (11), 1980e1989.
nities. Biofouling 27 (9), 993e1001. Wei, S., Jiang, Z., Liu, H., Zhou, D., Sanchez-Silva, M., 2013. Microbiologically induced
Sharma, K., Derlon, N., Hu, S., Yuan, Z., 2014. Modeling the pH effect on sulfido- deterioration of concrete: a review. Braz. J. Microbiol. 44 (4), 1001e1007.
genesis in anaerobic sewer biofilm. Water Res. 49, 175e185. Wells, T., Melchers, R.E., 2014. An observation-based model for corrosion of con-
Sharma, K., Ganigue, R., Yuan, Z., 2013. pH dynamics in sewers and its modeling. crete sewers under aggressive conditions. Cem. Concr. Res. 61e62 (0), 1e10.
Water Res. 47 (16), 6086e6096. Wells, T., Melchers, R.E., 2015. Modelling concrete deterioration in sewers using
Sharma, K.R., Yuan, Z., de Haas, D., Hamilton, G., Corrie, S., Keller, J., 2008. Dynamics theory and field observations. Cem. Concr. Res. 77, 82e96.
and dynamic modelling of H2S production in sewer systems. Water Res. 42 Zhang, L., Keller, J., Yuan, Z., 2009. Inhibition of sulfate-reducing and methanogenic
(10e11), 2527e2538. activities of anaerobic sewer biofilms by ferric iron dosing. Water Res. 43 (17),
Sivret, E., Stuetz, R., 2010. Sewer odour abatement practices e an Australian survey. 4123e4132.
Water 37 (7), 77e81.