Unsupervised feature-learning for out-of-distribution detection
Embedded Files
Unsupervised feature-learning for one-class anomaly detection: Feature extractor + Relationship learner.
Feature Extractors: Principal Component Analysis, Independent Component Analysis & Convolutional Autoencoders
Relationship Learner: one-class Support Vector Machines.
This research work is published in Composite Structures journal. Check it out here.
SPIE Smart Structures + NDE conference (March 2020) video presentation
Filtering out environment effects for guided wave using deep latent variable models
Filtering out environment effects for guided wave using deep latent variable models
The effect of temperature on guided waves interferes with the existing damage identification strategies.
A temperature compensation methodology using autoencoders.
Time-frequency analysis based on higher-order representations can be used.
Autoencoder transforms temperature-affected signals at the reference temperature.
Check it out here.
SPIE Smart Structures + NDE conference (March 2020) video presentation
Supervised damage detection & localization with domain knowledge priors for deep learning
Supervised damage detection & localization with domain knowledge priors for deep learning
This research work is published in Ultrasonics journal. Check it out here.
Domain knowledge-assisted ML = domain knowledge + deep learning.
CNN for damage detection (binary classification) and localization (multi-class classification).
Domain knowledge helps networks perform better than a direct DL
Advantages: Accuracy, time, sensor optimization, in-situ monitoring, and robustness towards the noise.
Supervised damage detection & localization with inverse deep surrogate models
Supervised damage detection & localization with inverse deep surrogate models
Forward model: A parallel implementation of a reduced-order spectral FEM
Datasets: Time-series dataset and time-frequency dataset.
Epistemic uncertainty: The datasets are corrupted with several levels of white Gaussian noise.
Inverse model: Deep learning networks like CNNs & RNNs.
Damage detection: Classification model & Damage localization + Severity: Regression model.
DL based on automatic features performs better than ML based on manual features.
NDT in Aerospace conference (Paris, Nov 2019) video
Damage detection & severity assessment of adhesive bonded joints
Damage detection & severity assessment of adhesive bonded joints
This research work is published in International Journal of Adhesion & Adhesives. Check it out here.
EMI conference at MIT (June 2018) video
Adhesive bonded structures have advantages over conventional joining methods.
Non-destructive testing and health monitoring of adhesive bonded structures is challenging
Piezoelectric transducers used in SHM are adhesively bonded and get disbonds from the host structure.
Transducer disbonds may interfere with the SHM of structural disbonds.
Electromechanical admittance model and experiment are conducted for classification of both types of disbonds.
Page updated
Google Sites
Report abuse