A Machine Learning Driven Damage Quantification Algorithm in moisture-contaminated composites.

R.D. Guha
North Carolina State University,
United States

Keywords: damage detection, moisture interaction, free-bound water, relative permittivity, thresholding, machine learning


Non-destructive evaluation techniques is becoming extremely crucial in determining the location and extent of damage for in-service polymer composites. In this work, an attempt has been made to use the relative permittivity of moisture as an imaging agent to detect damage in a polymer composite sample. Localized damage is induced in the laminate specimens via low velocity impact damage of 9 Joules. A split post dielectric resonator coupled with a vector network analyzer has been used to determine the spatial variation in relative permittivity across the composite laminate. Generally, results show significant increase in relative permittivity towards the damage location compared to surrounding undamaged areas. This increase is indicative of internal damage as a result of micro-crack formation around the point of impact. The new free volume in the damaged area is primarily occupied by “free” water molecules which are characterized by a higher relative permittivity; driving a local increase in the permittivity due to the higher ratio of free to bound water. The tendency of polymer composites to absorb moisture in almost all environments coupled with the high sensitivity of our technique makes this relatively simple non-destructive examination novel, especially for the early detection of damage. For this presentation, I have tried to use machine learning to perform classification and clustering analysis on the permittivity data collected to detect the damage spot and delineate a damage boundary. The performance of different learning algorithms has been compared and the problem with manual thresholding and labeling of data points has also been discussed. Finally, a potential solution to eliminate the misclassification of data points has also been investigated which can be further developed to use this algorithm similar to an image classification problem.