Doctoral Defense - Subal Sharma

About the Event

The Department of Electrical and Computer Engineering 
Michigan State University 
Ph.D. Dissertation Defense 
Wednesday, July 30, 2025, at 12:00 pm 
Electrical and Computer Engineering Conference Room EB 2219 

Join Zoom:  

https://msu.zoom.us/j/94980617034

Passcode: Upon request. Contact Department or Advisor

ABSTRACT

FEATURE ENGINEERING OF ULTRASONIC CODA WAVES FOR NONDESTRUCTIVE EVALUATION AIDED BY MACHINE LEARNING

BY: SUBAL SHARMA

ADVISOR: DR. SUNIL KISHORE CHAKRAPANI

Coda waves are diffuse field ultrasonic signals generated by multiple scattering and reflections from boundaries and heterogeneities within materials. Their long propagation times and decaying amplitudes, resulting from interactions with micro- and macro-scale features, create complex waveforms that appear noisy due to phase in-coherency and superposition of multiple scattered modes. Extracting meaningful features from these intricate signals is challenging. This thesis focuses on three problems where the changes/degradation at microstructural level: (i) microstructural variation in Grade 91 steel tubes, (ii) barely visible impact damage (BVID) in carbon fiber reinforced polymer (CFRP) composites, and (iii) the characterization of microcrack arrays in rails. 

Grade 91 steel, used extensively in power plants, local microstructural changes such as grain size variation reducing component serviceability. Initial feature extraction using loss of correlation (LOCOR) differentiated microstructure states into good (G), mid-level (M), and bad (B) groups with ≥80% accuracy but could not fully resolve all eight targeted conditions. Statistical T-tests and clustering verified group distinguishability but indicated limitations in resolving subtle microstructural variation, motivating exploration of more sensitive approaches. 

To address these limitations, topological data analysis (TDA) methods were introduced for feature extraction from coda waves. TDA-derived features, including Carlsson’s Coordinates, the Tent Function, and Interpolating Polynomials, provided quantitative descriptors from persistence diagrams of time-domain coda wave data. Machine learning classifiers trained on TDA features achieved ≥90% accuracy (averaged across multiple cases), outperforming LOCOR, which did not exceed 57%. This demonstrated the capacity of TDA to resolve subtle microstructural features from coda waves. 

To distinguish eight microstructure conditions, a data fusion approach was used, integrating features from seven different ultrasonic experiments. Non-fused data models showed limited accuracy for eight-class classification, while both feature-level and decision-level fusion substantially improved performance. This confirmed that combining diverse features from multiple sources was essential for robust multi-class microstructure discrimination. 

Feature extraction from coda waves also proved critical for detecting BVID in composites. Carbon fiber reinforced polymer (CFRP) laminates of two different ply sequences (0° & 90°) were subjected to low-energy impacts ranging from 2J-4.5J. Coda wave features distinguished between damages in different laminate types, while coda wave interferometry (CWI) provided the highest sensitivity to damage severity. Using machine learning, damage in different laminate sequences was classified with high accuracy, with most models achieving perfect classification. Within each laminate, impact energy levels above 2J (for 0° laminates) and above 3.5J (for 90° laminates) were identified with at least 80% classification accuracy. 

Finally feature engineering was applied to rolling contact fatigue (RCF) damage, which is typically found in rails and originated at the surface at micrometer scale. Previous work showed that using transmission coefficient to characterize the number and depth of cracks was impossible. Instead, this work proposed extracting features from the scatter field of Rayleigh waves. Specifically, it used LOCOR features along with ML regression models to predict multiple outputs for characterizing the (i) depth and (ii) number of cracks associated with RCF damage. LOCOR features with ML regression enabled simultaneous prediction of RCF crack depth and number of cracks with ≥90% accuracy, while the transmission coefficient feature could achieve only ~50% accuracy. 

In summary, this thesis demonstrated that advanced feature engineering, encompassing statistical, topological, and data fusion strategies, substantially improved coda wave-based diagnostics, enabling accurate, actionable assessments across critical engineering materials. 

Persons with disabilities have the right to request and receive reasonable accommodation. Please call the Department of Electrical and Computer Engineering at 355-5066 at least one day prior to the seminar; requests received after this date will be met when possible.