ÇEÇağatay Ersin

Çağatay Ersin

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Çağatay Ersin has a total of 4 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Context-Aware AI-Based PPE Compliance and Occupational Risk Assessment in Construction Environments Using Lightweight YOLO Models

    by Çağatay Ersin

    Automated personal protective equipment (PPE) monitoring in construction environments requires not only accurate object detection but also contextual interpretation of workplace safety. This study proposes a context-aware framework integrating lightweight YOLO detectors, worker–PPE association, workplace-context validation, worker–equipment proximity analysis, and interpretable heuristic risk assessment. A publicly available Construction Site Safety dataset comprising 2,801 images was used, including 2,605 training, 114 validation, and 82 independent test images. YOLOv8n and YOLO11n were trained under identical conditions for 100 epochs using 640 × 640-pixel inputs. On the independent test set containing 760 annotated instances, YOLOv8n achieved a precision of 0.848, recall of 0.682, mAP@0.50 of 0.743, and mAP@0.50:0.95 of 0.455, outperforming YOLO11n in recall and both mAP metrics. Consequently, YOLOv8n was selected for contextual risk assessment. Of 155 detected persons, 126 were retained as context-applicable workers, while 29 were excluded due to insufficient workplace-context evidence. Worker-specific analysis identified 101 helmet-positive, 5 helmet-violation, 59 safety-vest-positive, and 48 safety-vest-violation cases. The heuristic risk assessment classified applicable workers as SAFE (57; 45.2%), MODERATE (52; 41.3%), HIGH (9; 7.1%), or CRITICAL (8; 6.3%). The findings indicate that combining lightweight object detection with contextual validation and transparent rule-based reasoning provides more meaningful occupational safety information than object detection alone. The proposed framework supports interpretable, camera-based safety monitoring without presenting heuristic risk categories as validated accident probabilities.

  • Open Access

    Lightweight One-Class Industrial Anomaly Detection Using a Normal-Calibrated Random-Subspace Ensemble

    by Çağatay Ersin

    Industrial anomaly detection is commonly performed under one-class conditions because defective samples are scarce or unavailable during training. This study presents a lightweight one-class anomaly detection framework that reduces the dependence of compact deep representations on a single arbitrarily selected feature subset. Intermediate features extracted from a frozen MobileNetV2 backbone are divided into five independently sampled 100-channel subspaces. A position-specific diagonal Gaussian model is fitted separately to each subspace using only normal training images. The resulting anomaly scores are robustly normalized and combined through a normal-calibrated reliability-weighting strategy, while image-level decisions are obtained from the highest-scoring local patches. To prevent information leakage, operating thresholds are determined through five-fold cross-fitted calibration using only normal training scores. The proposed compact random-subspace ensemble was evaluated on all 15 categories of MVTec AD and externally validated on the 12-category VisA benchmark. On MVTec AD, the ensemble increased the macro ROC-AUC from 0.9350 to 0.9437 and the macro F1 score from 0.8366 to 0.8504 compared with a single random 100-channel subspace. External validation on VisA showed the same improvement trend, with higher macro ROC-AUC, AP, and F1 values than the single-subspace configuration. The results demonstrate that combining multiple compact normal-data subspaces improves discrimination and reduces reliance on the evaluated single random representation while retaining a lightweight, anomaly-label-free training procedure.

  • Open Access

    Lightweight One-Class Anomaly Detection for Industrial Image Inspection Using Enhanced Unified Feature Representation

    by Çağatay Ersin

    This study presents an enhanced unified feature representation for classical one-class anomaly detection in industrial image inspection using the MVTec AD dataset. Experiments were conducted on all fifteen categories of the MVTec AD dataset, including object and texture categories. Unlike raw pixel-based representations, the proposed feature representation integrates raw intensity information, histogram of oriented gradients (HOG), local binary pattern (LBP), gray-level co-occurrence matrix (GLCM)-based texture descriptors, and basic statistical intensity features. These complementary descriptors were evaluated with classical one-class anomaly detection methods, including PCA-based reconstruction, kNN-density, Local Outlier Factor, One-Class SVM with RBF kernel, Isolation Forest, and Mahalanobis-PCA. Considering the class imbalance inherent in anomaly detection tasks, Average Precision (AP) was used as the primary evaluation metric, while ROC-AUC and F1@95%-training-quantile were reported as supporting metrics. The results showed that PCA-based reconstruction achieved the strongest overall performance with a mean ROC-AUC of 0.8289, mean AP of 0.9254, and mean F1@95%-training-quantile of 0.8343. Category-based analyses showed particularly strong discrimination in the bottle, toothbrush, leather, transistor, and zipper categories, whereas carpet, screw, and grid remained more challenging. The findings indicate that carefully designed classical feature representations can provide an explainable and CPU-friendly baseline for industrial anomaly detection, particularly at the model-inference stage, while feature extraction remains the dominant computational cost.

  • Open Access

    Blockchain Destekli Nesnelerin İnterneti (IoT) Varlıklarında Fiyat Tahmini İçin Model Karşılaştırması

    by Çağatay Ersin, Onur Köktürk

    Teknolojik gelişmelerin hız kazanmasıyla birlikte Nesnelerin İnterneti (IoT - Internet of Things), küresel ölçekte üretim süreçlerinden akıllı şehir uygulamalarına, kişisel sağlık izleme cihazlarından akıllı spor kıyafetlerine kadar geniş bir yelpazede dönüşüme katkı sağlamaktadır. Nesnelerin İnterneti, cihazların birbirleriyle ve veri merkezleriyle etkileşime geçerek büyük miktarda veri ürettiği bir ekosistem yaratmaktadır. Bu bağlamda, IoT teknolojilerine dayalı blok zinciri tabanlı kripto varlıklar da gelişim göstermiştir. Bu çalışmanın amacı, IoT temalı kripto varlıklardan Vechain (VET), IOTA ve Helium (HNT) için fiyat tahmini yapmaktır. Çalışmada 02.01.2020 - 10.12.2024 tarihleri arasındaki Vechain ve IOTA, 09.06.2020 - 10.12.2024 tarihleri arasındaki Helium coinin günlük kapanış fiyatları kullanılmıştır. Fiyat tahmini için Karar Ağaçları (Decision Trees), Rassal Orman (Random Forest) ve Destek Vektör Regresyonu (SVR) modelleri karşılaştırılmıştır. Karar ağaçları modeli, geçmiş fiyat verilerini baz alarak gelecekteki fiyat hareketlerini tahmin etmeyi amaçlamaktadır. Rassal Orman algoritması, birden fazla karar ağacının bir araya getirilmesiyle oluşan topluluk öğrenme yöntemini kullanarak daha sağlam ve genellenebilir tahminler sunmaktadır. Destek Vektör Regresyonu (SVR) ise veri noktaları arasındaki karmaşık ilişkileri daha iyi yakalamak için hiper düzlemleri kullanarak doğrusal olmayan regresyon modelleri oluşturur. Modellerin performansı Mean Squared Error (MSE) ve R² (R-kare) skoru gibi metriklerle değerlendirilmiştir. Sonuçlar, Rassal Orman modelinin Helium ve IOTA coinlerin tahminlerinde yüksek doğruluk sağladığını, VeChain coin için ise SVR’nin daha iyi modellediğini ortaya koymuştur. Karar Ağaçları modeli, bağımsız tahminler üretme konusunda başarılı olmakla birlikte diğer modellere göre nispeten daha düşük performans göstermiştir.

  • Open Access

    Optimization of Wood-Based Birch Plywood CO2 Laser Engraving Process Parameters with Taguchi Method

    by Mehmet Güneş, Çağatay Ersin

    In this study, the optimization of the parameters used in the engraving process of wooden birch plywood material with a CO2 laser machine was investigated using the Taguchi method. An industrial laser machine with a 150W glass tube was used during the experimental examination process. There are basic factors affecting the engraving process of wooden surfaces such as laser power (P), engraving speed (S) and laser head parts (F). The engraving depth (D) and engraving width formed on the surface during the engraving process are the main factors that determine the aesthetics of the product. Taguchi L25 orthogonal array was used in the experiments to determine and optimize the highly important parameters. The optimum combination of parameters in the laser engraving process was then evaluated. The research results showed that the effect of P and S factors played a leading role, and the F parameter had a small effect on the depth of wood scraping. Optimization results found that F:5, S:100 and P:30 gave the best engraving depth optimization, while F:6, S:300 and P:10 gave the lowest engraving depth optimization result.


  • Open Access

    Implementation and Comparison of Wearable Exoskeleton Arm Design with Fuzzy Logic and Machine Learning Control

    by Çağatay Ersin, Mustafa Yaz

    In this study, a wearable exoskeleton arm was designed and controlled with different control methods to help people with muscle disorders in their arms and support treatment. The developed robot arm was transferred to Simulink software with the Simmechanics application. Two electromyography (EMG) muscle sensors and the ADXL335 position and acceleration sensors attach to the human arm’s biceps and triceps muscle areas. As the human moved the arm, data were obtained from the EMG muscle sensors and the ADXL335 position and acceleration sensor. The received data were first trained with the fuzzy logic algorithm. The same data were then trained with machine learning algorithms in Simulink software. It has been determined that the best result is the quadratic support vector machine (SVM) algorithm. The fuzzy logic algorithm trained with the PID controller block and the received sensor data have been added to the degrees of freedom regions that will enable rotation in the block diagram of the previously exported system. Later, the fuzzy logic block was removed and the machine learning algorithm, the quadratic SVM algorithm, was added. The designed system was operated with two different control systems, and the control algorithm closest to the human arm movement was determined. In addition, each part of the system, whose design was prepared, was removed and assembled separately with a 3D printer. ESP32 microcontroller development board was used to control the system, and it was run in real-time with EMG muscle sensors and position sensors.


  • Open Access

    Comparison of KNN and Random Forest Algorithms in Classifying EMG Signals

    by Çağatay Ersi̇n, Mustafa Yaz

    Depending on the growing average age and busy work life, muscle disorders are also increasing. Disturbing use life hurts the upper limb due to casing. Electromyography (EMG) muscle sensors are used to detect muscle diseases. To obtain more accurate results, the perception of the data received with the EMG sensors is required. This evaluation was compared with electromyography (EMG) muscle sensors used as a muscle measurement tool and those taken from the upper limb and KNN explanations and Random Forest examinations, which are the predictions of machine learning in this context and give more accurate results than other effects. Three EMG muscle sensors are attached to the upper limb of the user and taken from 0o, 45o and 90o angles with the microcontroller development board. It has been read and tested with the resulting machine-learning readings. The percentages of the accuracy of the highest accuracy KNN and Random Forest locations were chosen for their assumptions and use in use.