Back propagation neural network prediction of failure strength of composite tensile specimens

Abstract

In this research work, a feed forward back propagation neural network model was generated from acoustic emission cumulative counts data taken during loading of unidirectional carbon/epoxy tensile specimens. Only data collected up to 50% of ultimate load and its corresponding failure strength of 12 specimens were utilised for training the network out of 18 specimens tested, remaining six specimens data were utilised for testing the software. The network structured as 66-15-1 was able to give the prediction results with the worst case error of 4.98%. Fifteen neuron middle layer was mapping the pattern for failure of specimens well before from the AE data. This work indicates that it is possible to proof test the composites more sophisticated at lower loads (may be 50% of ultimate load) then are currently being tested (70% to 80% of ultimate load). Unintentional degradation of composite materials while proof testing could be thus minimised.