Neural network fatigue life prediction in steel i-beams using mathematically modeled acoustic emission data

The purpose of this research is to predict fatigue cracking in metal beams using mathematically modeled acoustic emission (AE) data. The AE data was collected from nine samples of steel Ibeam that were subjected to three-point bending caused by cyclic loading. The data gathered during these tests were filtered in order to remove long duration hits, multiple hit data, and obvious outliers. Based on the duration, energy, amplitude, and average frequency of the AE hits, the filtered data were classified into the various failure mechanisms of metals using NeuralWorksRTM Professional II/Plus software based self-organizing map (SOM) neural network. The parameters from mathematically modeled AE failure mechanism data were used to predict plastic deformation data. Amplitude data from classified plastic deformation data is mathematically modeled herein using bounded Johnson distributions and Weibull distribution. A backpropagation neural network (BPNN) is generated using MATLABRTM. This BPNN is able to predict the number of cycles that ultimately cause the steel I-beams to fail via five different models of plastic deformation data. These five models are data without any mathematical modeling and four which are mathematically modeled using three methods of bounded Johnson distribution (Slifker and Shapiro, Mage and Linearization) and Weibull distribution. Currently, the best method is the Linearization method that has prediction error not more than 17%. Multiple linear regression (MLR) analysis is also performed on the four sets of mathematically modeled plastic deformation data as named above using the bounded Johnson and Weibull shape parameters. The MLR gives the best prediction for the Linearized method which has a prediction error not more than 2%. The final conclusion made is that both BPNN and MLR are excellent tools for accurate fatigue life cycle prediction.