Browsing by Author "Balci, Zekeriya"

Now showing 1 - 2 of 2

Autoencoder-Based Eggshell Crack Detection Using Acoustic Signal
(Wiley, 2024) Yabanova, Ismail; Balci, Zekeriya; Yumurtaci, Mehmet; Unler, Tarik
Breaks or cracks in eggshells offer substantial food safety issues. Bacteria and viruses, in particular, are more likely to enter the egg through breaks and cracks, increasing the risk of food poisoning. Furthermore, deformations in the shell may compromise the integrity of the protective shell, exposing the egg to more external variables and causing it to lose freshness and decay faster. To reduce such hazards, this research created an innovative crack detection system based on an autoencoder (AE) that uses acoustic signals from eggshells. A system that creates an acoustic effect by hitting the eggshell without damaging it was designed, and these effects were recorded through a microphone. Acoustic signal data of size 1 x 1000 was fed into k nearest neighbor (kNN), decision tree (DT), and support vector machine (SVM) classifiers. AE was employed to reduce data size in order to accommodate the raw data's unique features. This AE model, which reduces data size, was used with many classifiers and was able to accurately distinguish between intact and cracked eggs. The built AE-based classifier model completed the classification procedure with 100% accuracy, including microcracks that are invisible to the naked eye.
Photoacoustic Signal to Image Based Convolutional Neural Network for Defect Detection
(AIP Publishing, 2025) Balci, Zekeriya; Mert, Ahmet
In this paper, we propose a novel photoacoustic (PA) signal to image conversion based convolutional neural network (CNN) model for defect detection in materials. A low-cost computer aided PA triggering and acquisition device has been developed, and then, PA signals are stored for four types of defected and intact materials. Variational mode decomposition is applied to the dataset to extract intrinsic mode functions to convert PA signals to images as the first step of the feature extraction, and then, a lightweight CNN architecture is trained and tested using converted grayscale PA images to detect as defected or intact material. The proposed model is performed on the PA signals of aluminum, iron, wood, and plastic depending on the within-class and all-class evaluation strategies. The mean accuracy levels of 0.977 (up to 1.0) for within-class (material dependent) and 0.942 (up to 0.955) for all-class (material independent) are yielded.