Showing papers by "Erdal Oruklu published in 2019"

Multi-Class Classification of Defect Types in Ultrasonic NDT Signals with Convolutional Neural Networks

Abstract: The key objective of this work has been to detect and classify different types of flaws encountered in ultrasonic Non-Destructive Testing (NDT) applications. The flaws that are examined for classification are Side Drilled Hole, Flat Bottom Hole, and Rectangular voids with different aspect ratios. Manual inspection of ultrasonic images is often inadequate for classification purposes since it is difficult to visually discriminate the flaws due to their similarities. In particular, voids and holes closer to the edge boundaries (i.e. front or bottom of the structure) make this task especially challenging. We propose to use deep-learning methods to automate the ultrasonic flaw classification. In the proposed setup, the specimen under test is a steel block and immersion based ultrasonic testing has been considered. OnScale multiphysics simulation software has been used for data synthesis and a Convolutional Neural Network based deep learning method has been developed to classify the flaws with one hot encoding method. Preliminary results have been promising with the average classification accuracy of 90%, 93%, 95% and 91% for flat bottom hole, rectangular void with greater width, side drilled hole and rectangular void with greater length, respectively.

Unsupervised Machine Learning for Ultrasonic Flaw Detection using Gaussian Mixture Modeling, K-Means Clustering and Mean Shift Clustering

Abstract: Supervised Machine Learning (ML) algorithms such as Neural Networks, Support Vector Machines and Logistic Regression have been successfully utilized in Ultrasonic Non-Destructive Evaluation (NDE) applications. In supervised learning algorithms, data outputs are labeled and classified for training. In contrast, Unsupervised Machine Learning (UML) algorithms identify and exploit the commonalities in the data and no "ground truth" is necessary. In this work, we use three different UML algorithms based on K-means clustering, Gaussian Mixture Modeling and Mean Shift Clustering in order to detect and locate flaw echoes in ultrasonic A-Scan data. All three algorithms have been shown to perform flaw classification successfully. In particular, Gaussian Mixture Modeling achieves highest detection accuracy at 93%.

Investigation of Feature Inputs for Binary Classification of Ultrasonic NDT Signals using SVM and Neural Networks

Abstract: Classification of Ultrasonic Non-Destructive Testing (NDT) signals can be done by Machine Learning models including Support Vector Machine (SVM) and Neural Networks (NN). The main objective of this study is to classify the ultrasonic A-scan data either as flaw echoes or clutter echoes (no flaw). The signal pre-processing has been done using Discrete Wavelet Transform (DWT). The refined low pass output has been used as feature input to the machine learning algorithms either directly or as a power signal. In case of SVM, direct low pass output in windowed format was tested with linear kernel and Radial Basis Function (RBF) kernel and the power signal of the low pass DWT in the windowed format was also tested with linear kernel and RBF kernel. SVM simulation results show that the direct low pass signal with linear kernel fails to converge while power of the low pass DWT achieves an accuracy of around 95%. RBF kernel accuracy was around 98% irrespective of the format of the signal. In case of Neural Network, both the direct low pass output and the power of the low pass output were tested and it was found that the direct low pass output with NN yielded 94% accuracy while the power of the low pass output with NN yielded an accuracy of 98%.

A New Classifier Network for Ultrasonic NDE Applications based on Ensemble Deep Learning

Abstract: This work presents a classifier architecture for Non-Destructive Evaluation (NDE) applications which can robustly detect the presence and location of flaws using an ensemble of deep learning networks. The ensemble draws upon the effective sequential time analysis of Long Short-Term Memory - Neural Networks (LSTM–NN), the function estimation and prediction properties of Wavelet Neural Networks (WNN), and the feature extraction capabilities of Convolution Neural Networks (CNN). Simulation results confirm that the proposed architecture offers highly reliable flaw detection and localization with significant Flaw to Clutter Ratio (FCR) enhancements.

Side-Channel Attack Resilient Design of a 10T SRAM Cell in 7nm FinFET Technology

Abstract: The non-invasive Side-Channel Attacks (SCA) for integrated circuits have been a concern for many years and Leakage Power Analysis (LPA) is among the leading threats for the IC security. For SRAM blocks, LPA would exploit the correlation between data in memory cell and its corresponding leakage power, and possibly decrypt the secret key inside the memory of crypto-systems. This paper proposes a new SRAM design countering against LPA, based on a recent low-power single-ended 9T cell design. Leakage balance issue for the 9T cell is discussed and a new cell design is presented. Simulation results confirm that the proposed SRAM cell preserves the advantages from the 9T design and have a well-balanced leakage behavior preventing against SCA.

From Design to Implementation - Algorithm Conversion and Transition to a Mobile Platform for an Artificial Pancreas

Abstract: An artificial pancreas system is implemented as a mobile application which connects a glucose sensor, an insulin pump and wearable physical activity sensors. It automatically delivers the optimal insulin amounts, based on a multivariable control algorithm. The algorithm, previously tested on a laptop, is to be hosted on a smartphone. This requires a fast, reliable method to translate the inherent functions. Validation tests of the new mobile application confirm there are no deviations from the original code.