Proceedings of the
9th International Conference of Asian Society for Precision Engineering and Nanotechnology (ASPEN2022)
15 – 18 November 2022, Singapore

A Machine Learning Approach to Defect Detection and Failure Analysis in Composites

Soumya Bhowmik1, Tan Long Bin2,a and Tan Beng Chye Vincent1

1Department of Mechanical Engineering, National University of Singapore, College of Design and Engineering, 9 Engineering Drive 1, #07-08 Block EA, Singapore 117575

2Engineering Mechanics Department, Institute of High Performance Computing, A*STAR, 1 Fusionopolis Way, #16-16 Connexis, Singapore 138632


This work showcases the application of Machine Learning (ML) for forensic and defect analyses. The first part uses ML to predict the initial conditions, such as the speed, angle, and location, of impact to a hemispherical shell. The conditions leading to an accidental impact event are usually unknown, but certain forensic signatures such as plastic deformation or dents are easily measurable. The developed Artificial Neural Network (ANN) model can be used to serve legal and insurance purposes, where knowledge of the conditions leading to an impact event is crucial. Our study consisted of collisions of a 3 m diameter and 5 mm thick steel shell. 192 finite element (FE) simulations with varying impact conditions were conducted and the nodal displacements for the entire shell, post-collision, were extracted. 154 case data were used for training and the remaining data for testing of the ANN. For the test cases, the ANN predicted locations are all within 10% of the true values. The mean error for longitude, latitude and speed are 2.6%, 2.2% and 9% respectively. The developed ANN was successfully able to predict the initial impact location and impact conditions. The methodology may be further expanded to predict loading conditions for structural damage or for car/plane crashes to better understand the root cause of the accident. The inverse modeling scheme may also be applied to determine manufacturing conditions for Thermoforming, Punching, etc based on post-manufactured features such as flash, smear, delamination or other defects.

The second part of the project attempts to predict the location % size of delamination for a two-ply (2.5 mm each) plain-woven Carbon Fiber Reinforced Polymer (CFRP) composite. The laminate material is used in place of steel from the previous study. The FE model for delamination location study has 10 delamination zones across a single meridian which are 7.5 degrees apart from each other ranging from 15 to 90 degrees. The delamination size for each zone is 25 cm, and separate analysis is conducted for each delamination location. Another four FE models are used for the delamination size analysis, with delamination sizes ranging from 25 cm to 70 cm in diameter. A total of 26 cases were analyzed numerically. For our study, only one delamination at any time is considered. The two shells are bonded by tie-constraints, except for the designated delamination region, which is unbonded. A tapping pulse is imparted to the shell and the normal displacement-time outputs extracted to form the ML training data set. For the delamination location prediction, the RMS Error is 3.66, which is lower than the 7.5-degree resolution and thereby able to accurately distinguish between the zones. For the delamination size prediction, the mean and max errors are 4.46% and 9.93% respectively. The ANN was successfully able to predict both the delamination location and delamination size reasonably well. The methodology may be implemented to use actual instrumentation data, for example vibration/acoustic register from aircraft or ship panels to help identify possible damage or delamination locations in composite materials.

Keywords: Machine Learning, Artificial Neural Network, Forensic, Defect Analysis, Composites, Delamination

PDF Download