3 3 Jiang - Advanced Water Management CentreIntercept 896.34 15.03 6.41 3.49×10 *** Locaon 1.68...
Transcript of 3 3 Jiang - Advanced Water Management CentreIntercept 896.34 15.03 6.41 3.49×10 *** Locaon 1.68...
Predic'ng the corrosion ini'a'on 'me of fresh concrete sewers by ar'ficial neural network
Guangming Jiang 14 June 2015
Corrosion processes & reac-ons -‐ air & liquid phase
Sulfide oxidation:H2S + 2O2 → H2SO4
Acid attack on cement:CaO·SiO2·2H2O + H2SO4 → CaSO4 + Si(OH)4 + H2OH2SO4 + CaCO3 → CaSO4 + H2CO3
H2SO4 + Ca(OH)2 → CaSO4 ·2H2O (gypsum) 3CaSO4·2H2O + 4CaO·Al2O3·13H2O + 14H2O →(CaO)3·Al2O3·(CaSO4)3·32H2O (ettringite)
Corrosion processes & reac-ons -‐ solid phase
(Jiang et al. 2014)
Development of sewer corrosion – ini-a-on processes
• Surface neutraliza-on
o Weak acids from air and
wastewater (H2S, CO2, organic
acids)
o Biologically generate sulfuric acid
• Biological processes
o SOB (from producing S0 to
sulfuric acid)
• Controlling environmental factors
o H2S levels; Temperature; Rela-ve humidity (Joseph et al. 2012; Jiang et al. 2015)
Bilinear model of sewer corrosion
• Many models for the corrosion rate
• No model available for the predic-on of tini-a-on (Wells and Melchers, 2014)
12initiation
servicet Dt
r= +
tservice: service life (year) tini-a-on: ini-a-on -me (month) D: concrete depth (mm) r: corrosion rate (mm/year)
Laboratory corrosion chambers Factors
Temperature (Gas phase) 16-‐18oC, 25oC & 30oC
Rela-ve Humidity 100% & 85-‐95%
H2S level (ppm) 0, 5, 10, 15, 25 & 50
3 x 2 x 6 = 36 chambers
Coupons Loca-on Gas-‐phase & par-ally submerged
Type Fresh concrete
Exposure Every 6-‐9 months up to 4.5 years
2 x 8 = 16 coupons each chamber
Preliminary analysis Ini-a-on -me tin Sta-s-cal analysis
Controlling factors for tin
Model development
Training Valida-on
Field data
black-‐box model
Raw data
Data analysis & model development
Explanatory analysis of tin
• GPC: Temperature and RH are key factors, with H2S as a secondary.
• PSC: H2S and temperature are key factors. RH is not.
Gas-‐phase concrete Par-ally-‐submerged coupons
ANOVA of tin GPC PSC
H2S (ppm) Rela've humidity (%) Temperature (oC)
Corrosion ini'a'
on 'me (m
onth)
Mul-ple linear regression model
296.34 1.68 0.18 0.54 0.84it Location H S RH T= + ∗ − ∗ − ∗ − ∗
Coefficients Es'mate Std. Error t value P(>|t|) Significance
Intercept 96.34 15.03 6.41 3.49×10-‐8 ***
Loca'on 1.68 0.77 2.19 3.25×10-‐2 *
H2S -‐0.18 0.05 -‐3.83 3.34×10-‐4 ***
RH -‐0.54 0.15 -‐3.51 9.13×10-‐4 ***
Temperature -‐0.84 0.14 -‐5.83 3.04×10-‐7 *** • Reasonably reflect the rela-onship between ti and controlling factors
• Max ti = 96 months (8 years); Loca-on difference = 3.5 months
• R2=0.54 à nonlinear rela-onship
Ar-ficial Neural Network model
• ANN model design • Network architecture design (BP)
• Layers (1 input, 1 hidden, 1 output)
• Neurons (4 input (1C+3N), 8 hidden, 1 output)
• Model training with back-‐propaga-on algorithm • Data: Training(70%)/Valida-on(15%)/Tes-ng(15%)
ANN model inspired by biological neuron systems generate the mathemaCcal relaConships between input and output data by idenCfying the paFerns in data. -‐ Data driven; black-‐box; suitable for any complex system.
ANN model performance
ANN platforms and tools: • Alyuda NeuroIntelligence
• Matlab
ANN model valida-on
• Validated ANN model
• Reasonable accuracy
• Higher confidence with more data
Sensi-vity analysis & uncertainty
• Sensitivity analysis o Nonlinear responses to environment
o Sensitivity order: H2S > Temperature > RH
o Different sensitivity for different locations
• Uncertainty (unexplained variances in predictions) o Constant vs. fluctuating H2S
o Processes contributing to corrosion
o Other factors (concrete properties)
Conclusions
• Artificial neural network (ANN) performs well in predicting tin based on sewer environmental conditions.
• The ANN model can be used to improve the understanding of corrosion mechanisms.
• The ANN model is a good base framework for further expansion by including more corrosion data.
Acknowledgements
• Queensland Government Accelerate fellowship
• ARC Linkage Project LP0882016, the Sewer Corrosion and Odour Research (SCORe) Project