Cross-grain fracture characterization in softwood using artificial neural network analysis of acoustic emissions

In an effort to better understand crack growth in the cross-grain direction, an acoustic emission (AE) approach was implemented to monitor the propagation of damage in eastern spruce under tensile loading. Supervised and unsupervised neural networks were applied to classify AE signals, which could then be associated with different damage mechanisms. Using a mel-frequency cepstral (MFC) representation of the AE signals, a self-organizing map was implemented to identify the number of different signal clusters, leading to a pattern recognition network. The experiments consisted of compact tension specimens with a 5-mm precrack loaded at three different end-grain orientations relative to initial crack direction: 0\(^\circ\), 45\(^\circ\), and 90\(^\circ\). The recorded AE events showed eight distinct clusters, which were used to train a two-layered feed-forward supervised network. The trained network was used to identify damage mechanisms as a function of crack propagation orientation. Once identified, the AE energy release associated with the damage mechanism could be measured. Results showed that earlywood cell wall tearing, dominant at 0\(^\circ\), produced higher energy release than cell wall separation, which dominates 90\(^\circ\) crack propagation. Fiber bridging, which was observed in later stages of crack growth, produced a large number of AE events, but minimal energy release.