Materials Performance

AUG 2018

Materials Performance is the world's most widely circulated magazine dedicated to corrosion prevention and control. MP provides information about the latest corrosion control technologies and practical applications for every industry and environment.

Issue link: https://mp.epubxp.com/i/1007306

Contents of this Issue

Navigation

Page 63 of 92

61 MATERIALS PERFORMANCE: VOL. 57, NO. 8 AUGUST 2018 TABLE 1. SELECTED VARIABLES FOR THE MLR AND PCA MODELS MLR PCA Atmospheric Zone Variables Entered R 2 Rules KMO Component Proportions of Variance (%) Industrial DEW, EXP, SUN, RF 0.81 Stepwise Regression (Entered if ≤0.05, Removed if ≥0 .10) 0.51 4 88.5 Coastal EXP, Cl – , TEMP 0.53 0.52 4 82.3 Rural RF, WS, WD 0.83 0.56 3 91.8 TABLE 2. CONSTRUCTION OF ANN MODEL Model Construction MAPE Values Atmospheric Zone Model Layer Trained RMSE Tested RMSE Trained R 2 Tested R 2 Corrosion Rate MAPE (%) Forecasting Interpretation Apply to ANN Industrial ANN One 0.0149 0.0432 0.99 0.94 8.1 Highly accurate Yes Two 0.0204 0.0415 0.98 0.95 8.1 Highly accurate MLR-ANN One 0.0656 0.0651 0.85 0.86 22.7 Reasonable No Two 0.0679 0.0671 0.84 0.85 21.5 Reasonable PCA-ANN One 0.0347 0.0481 0.96 0.92 9.8 Highly accurate Yes Two 0.0516 0.0535 0.91 0.90 13.3 Good Coastal ANN One 0.0404 0.0488 0.85 0.85 48.2 Not accurate No Two 0.0445 0.0568 0.82 0.79 49.7 Not accurate MLR-ANN One 0.0528 0.0824 0.75 0.56 55.4 Not accurate No Two 0.0346 0.0603 0.89 0.77 45.3 Not accurate PCA-ANN One 0.0428 0.0455 0.83 0.87 39.0 Reasonable No Two 0.0433 0.045 0.83 0.87 39.0 Reasonable Rural ANN One 0.0683 0.0716 0.81 0.85 24.2 Reasonable No Two 0.0636 0.0656 0.83 0.88 21.6 Reasonable MLR-ANN One 0.0496 0.0499 0.90 0.93 17.4 Good Yes Two 0.0466 0.0466 0.91 0.94 15.5 Good PCA-ANN One 0.0213 0.0494 0.99 0.81 11.7 Highly accurate Yes Two 0.0249 0.0449 0.98 0.85 10.9 Highly accurate locations in Taiwan, as shown in Table 2. For the industrial atmosphere, the ANN model with two hidden layers worked the best (trained RMSE = 0.0204, tested RMSE = 0.0415, trained R 2 = 0.98, and tested R 2 = 0.95). For the coastal atmosphere, the PCA- ANN model with one hidden layer worked the best (trained RMSE = 0.0428, tested RMSE = 0.0455, trained R 2 = 0.83, and tested R 2 = 0.87). For the rural atmosphere, the MLR-ANN model with two hidden layers worked the best (trained RMSE = 0.0466, tested RMSE = 0.0466, trained R 2 = 0.91, and tested R 2 = 0.94). The ratings for RMSE and MAPE values are shown in Table 3. Forecasting Power of Artificial Neural Construction Th e three atmosph eric zon es were trained and tested to obtain optimal neural structure results, and then a comparison of forecasting power was conducted . The PCA-ANN model (MAPE = 39.0%) for the coastal zon e had th e b est forecasting power. For the rural areas, the PCA-ANN model (MAPE = 10.9%) had the best fore- castin g p ower ; fol lowed by MLR -ANN (MAPE = 15.5%). In summary, it was clear that the models for the industrial atmo- spheric zone had highly accurate forecast- ing powers. For the coastal atmospheric zone, the MAPE values for the models were all larger than 20% and forecasting power was either reasonable or not accurate. Conclusion This study shows that in the industrial atmospheric zone, the corrosion rate val- ues were higher than the values in the Atmospheric Corrosion Prediction of Carbon Steel

Articles in this issue

Archives of this issue

view archives of Materials Performance - AUG 2018