Sürdürülebilir Kalkınma Amaçları
Yayınlar
- ArticleAçık Erişim2026Multi-Modal Attention-Based Automatic Diagnosis of Bearing Vibration Faults for Robust Mechanical Condition Monitoring
Tastimur, Canan; Kaya, Volkan
Bearing components are critical elements of rotating machinery, and their failures can lead to sudden shutdowns, reduced efficiency, and increased maintenance costs. To address the limitations of conventional fault diagnosis methods, this study proposes a multimodal deep learning framework that integrates raw time-domain vibration signals with stationary wavelet transform (SWT)-based time-frequency representations using a novel bidirectional cross-attention mechanism. This architecture enables effective and mutual interaction between heterogeneous features, significantly enhancing discriminative representation learning and fault classification performance in complex operating environments. The proposed model was evaluated on the public SUBVF1.0 dataset under three fault conditions, achieving training and validation accuracies of 99.17% and 95.63%, respectively. The model also demonstrated strong performance in terms of precision, recall, F1-score, and ROC analysis. An ablation study further confirms the contribution of each model component. These results indicate that the proposed approach is accurate, robust, and suitable for practical industrial fault diagnosis applications.
RDL Kalıcı Kaydıİncele - ArticleYalnız künye2026Automated fabric defect inspection with fuzzy c-means, conditional random fields and refined U-Net
Tastimur, Canan
In this study, multiple segmentation methods were applied sequentially to enhance the classification for fabric data. First, fuzzy c-means (FCM) segmentation was employed on fabric images, effectively separating in homogeneous regions. Subsequently, these results were refined using conditional random fields (CRF), which provided sharper and more accurate delineation of pixel boundaries. In the final stage, I applied the refined U-Net model to the segmentation results enhanced by CRF, ensuring the preservation of small details in complex and irregularly structured fabric images. In industrial applications such as fabric flaw identification, the distinct contribution of categorizing based on the segmentation results significantly enhanced classification capabilities, particularly in low-quality images. This innovative hybrid strategy serves as an effective technique for improving outcomes, especially in images with uneven textures. Compared to the original dataset, the results are markedly better and pave the way for future advancements in industrial fabric analysis procedures.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2026A new approach for lung cancer diagnosis: parallel feature learning
Tastimur, Canan
Lung cancer, one of the most common types of cancer worldwide, can be fatal. Early diagnosis saves lives. Computed tomography (CT) is used in the diagnosis of the disease. Since the radiology specialist evaluates this X-ray result, the specialist's interpretation can vary. Furthermore, the analysis by the radiologist is both time-consuming and costly. However, a cancer diagnosis approach based on deep learning models supports the radiologist's decision. In this study, a parallel feature learning architecture developed for lung CT images was designed. This architecture focuses on learning different features from each parallel path by using deformable and dilated convolution layers together. Dilated convolution captures semantic features in images with different dilated rates ratios by expanding receptive field, while deformable convolution better captures structural changes. This mechanism allows for more flexible and distinctive feature extraction without significantly increasing computational complexity. The proposed architecture was tested on three different lung cancer datasets: the Public Lung Cancer Dataset, IQ-OTH/NCCD, and LIDC-IDRI. Experimental findings demonstrate robust and consistent classification performance, achieving accuracy rates of 99.44%, 98.75%, and 98.62%, respectively. This shows that the proposed architecture offers a reliable solution for lung cancer diagnosis.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2026Explainable AI for accurate diagnosis of papillary thyroid carcinoma via fine-needle aspiration cytopathology
Kilicarslan, Ahmet; Tastimur, Canan; Gundogdu, Gokhan; Akin, Erhan
Papillary thyroid carcinoma (PTC) represents most thyroid cancer cases worldwide, making early and accurate diagnosis crucial for successful treatment. This study introduces an explainable AI framework for the automated analysis of fine-needle aspiration (FNA) cytopathology images. By integrating handcrafted features with deep learning-derived abstract features, the proposed model achieves high classification performance (~ 92% accuracy and ~ 96% AUC). The proposed architecture, incorporating multi-scale dilated and deformable convolutions, effectively captures irregularities in cell morphology. Visualizations demonstrate alignment with pathologists’ diagnostic criteria, enhancing the model’s clinical applicability and reliability. The code, dataset, and comprehensive documentation are publicly available via Zenodo (https://zenodo.org/records/18428647), ensuring full reproducibility of the experiments and results presented.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2026LAViT: Class-Aware Vision Transformer with Learnable Attention for Pest Recognition in Agricultural Fields
Tastimur, Canan; Kaya, Volkan
Insect pests pose a significant threat to agricultural productivity, making early and accurate identification essential for effective pest management. This study proposes a novel deep learning-based classification framework for multi-class pest recognition from field images. The proposed approach enhances discriminative region representation by integrating a patch embedding module, a Vision Transformer backbone, and a learnable spatial attention mask. This hybrid design enables the model to focus on critical visual cues without requiring segmentation-based preprocessing. The attention mask, learned via convolutional layers, is pooled and directly applied to Transformer-encoded patch tokens to refine spatial feature emphasis. Positional embeddings are further employed to preserve spatial context within the tokenized image representation. Experimental evaluations conducted on publicly available pest datasets with 5, 9, and 12 classes demonstrate the effectiveness and robustness of the proposed framework. The model achieves accuracies of 99.67%, 99.52%, and 97.00% on the Pest5, Pest9, and Pest12 datasets, respectively, indicating strong generalization across varying classification complexities. To enhance model transparency and reliability, visual interpretability is provided through Grad-CAM and attention heatmap visualizations that reveal the model’s focus regions. Additionally, t-SNE-based feature visualization illustrates clear separability in the learned embedding space. The proposed framework shows strong potential for practical deployment in smart agriculture and precision pest monitoring systems.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2026Noise-Robust Multi-Domain Feature Integration for Bearing Fault Diagnosis
Tastimur, Canan
In this study, a hybrid deep learning architecture is proposed for robust vibration-based fault diagnosis in industrial machinery by jointly modeling time-domain, frequency-domain, and temporal dynamics. In the time domain, convolutional layers combined with an Enhanced Gated Attention (EGA) mechanism emphasize informative signal components while suppressing noise. Temporal evolution is modeled using Neural Ordinary Differential Equations (Neural ODEs), enabling smooth and stable continuous-time feature representations. In parallel, a Fourier Neural Operator (FNO) extracts frequency-domain characteristics, augmented with gated attention to focus on fault-related spectral patterns. Long Short-Term Memory (LSTM) layers capture long-range dependencies, while Squeeze-and-Excitation (SE) blocks adaptively recalibrate channel-wise feature responses. A multi-scale attention-based fusion module integrates domain-specific representations and auxiliary features to enhance discrimination under varying operating conditions. The proposed model is evaluated on the SUBFv1.0 dataset through extensive ablation studies and experiments under multiple noise levels, achieving 98.41% accuracy in noise-free conditions and maintaining performance above 91% even at 5 dB SNR. Unlike existing multi-path approaches that combine heterogeneous features in a loosely coupled or discrete manner, the proposed architecture uniquely integrates multi-domain feature learning with bidirectional attention mechanisms and continuous-time temporal dynamics, enabling coherent cross-domain interaction and robust fault characterization. KEYWORDS: Bearing fault diagnosis; Feature fusion; Hybrid deep learning; Neural ODE; Self-attention.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2025Efficient and Real-Time Railway Track Fault Classification Using CNN Integrated with Convolutional Block Attention Module
Tastimur, Canan
The occurrence of various types of faults on railway rail surfaces can lead to accidents such as train derailments. In this study, a new approach that produces very fast and robust results is developed by combining the convolutional neural network and the convolutional block attention module for the classification of rail faults. The dataset used in this study contains four different fault types, and experimental studies were conducted on the public rail dataset. The impact of the convolutional block attention module on the performance of the proposed approach and its contribution to the model's generalization ability are examined, and the performance of the proposed approach increases by approximately 5% compared to the performance of the proposed approach without this module. It has been demonstrated that the proposed approach can be used effectively in railway track fault diagnosis by producing fast and effective results.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2025Fabric Faults Robust Classification Based on Logarithmic Residual Shrinkage Network in a Four-Point System
Tastimur, Canan; Akin, Erhan; Agrikli, Mehmet
Accurate and robust detection of fabric defects under noisy conditions is a major challenge in textile quality control systems. To address this issue, we introduce a new model called the Logarithmic Deep Residual Shrinkage Network (Log-DRSN), which integrates a deep attention module. Unlike standard residual shrinkage networks, the proposed Log-DRSN applies logarithmic transformation to improve resistance to noise, particularly in cases with subtle defect features. The model is trained and tested on both clean and artificially noised images to mimic real-world manufacturing conditions. The experimental results reveal that Log-DRSN achieves superior accuracy and robustness compared to the classical DRSN, with performance scores of 0.9917 on noiseless data and 0.9640 on noisy data, whereas the classical DRSN achieves 0.9686 and 0.9548, respectively. Despite its improved performance, the Log-DRSN introduces only a slight increase in computation time. These findings highlight the model’s potential for practical deployment in automated fabric defect inspection.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2025Mean shift based prototypical network for steel surface anomaly recognition
Tastimur, Canan
Steel production plays a crucial role in the global economy, and surface defects in hot-rolled steel sheets significantly affect product quality and consumer acceptance. This paper introduces Mean Shift based Prototypical network (MSPro-Net), a novel and effective Prototypical network based on mean drift for classifying surface defects early in the manufacturing process. Key contributions of this study include the introduction of MSPro-Net’s adaptive prototype computation, which uses the mean-shift method to improve defect detection. Unlike the classical Prototypical network (CL-ProNet), MSPro-Net produces a prototype that better represents all examples within a class, leading to significantly higher accuracy. Additionally, MSPro-Net is highly effective in few-shot learning scenarios, achieving excellent results with limited training data. Extensive experimental evaluations demonstrate that MSPro-Net consistently outperforms CL-ProNet in multiple N-way K-shot scenarios. For example, on the 6-way 25-shot NEU dataset, MSPro-Net achieves 98.67% accuracy com-pared to 50.67% for the classical network. Similar improvements are observed on the 7-way 25-shot XSDD and 10-way 25-shot GC10-Det datasets, with MSPro-Net achieving 96.00% and 90.00% accuracy, respectively, far exceeding the classical network’s performance. These results highlight the superior performance of MSPro-Net across both small and large datasets.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2024Diagnosis of Permanent Magnet Assisted Synchronous Reluctance Motor Winding Fault by Convolutional Neural Network
Akın, Erhan; Bayrak, Ayşe; Tastimur, Canan
In recent years, the use of machine learning models for fault detection has become commonplace. Its goal is to identify and fix problems with permanent magnet synchronous reluctance motors. This research’s primary goal is to identify and categorize errors in their early stages. We classified winding faults using machine learning approaches, such as Independent Component Analysis and Deep Learning models. We could distinguish between vibration and current signals from the engine signals by using Independent Component Analysis (ICA). We experimented on multiple architectures using the convolutional neural network (CNN) architecture we designed from scratch and the Transfer Learning technique, testing two distinct datasets we generated using the signals we got. According to experimental findings, the suggested scratch CNN model performed exceptionally well in classification, achieving 98.6% with current signals and 99.4% with vibration signals.
RDL Kalıcı Kaydıİncele - Conference ObjectAçık Erişim2024Analysis of Barrel Electroplating Line with Process Mining and Petri-Net Model
Ozates, Ayse; Akin, Erhan; Bilgic, Aykut; Akca, Emre; Tastimur, Canan
Nowadays, many companies carry out efficiency studies to increase their profitability. The most important step in productivity studies is to objectively analyze the processes and identify the points that can be improved. Process mining shows the actual functioning within the process. It also identifies points of improvement such as bottlenecks, waiting times and deviations in the process. In this article, it is aimed to analyze the production process of the barrel electrolytic coating line with the process mining method. In order to perform the analysis, firstly, the data related to the production process was obtained from the ERP system in. csvformat. Within the scope of process mining, data obtained from the event log were used to analyze and visualize the processes affecting the coating line and to identify areas where improvements can be made. Process mining techniques were applied to the event log of the production identified after the electrolytic coating process. ProM 6.12, a widely used open source tool for process mining, was used to perform our analysis. After examining the in-line processes of the product produced in the electrolytic coating line, the data from the event logs were processed and analyzed with process discovery algorithms. In the literature, there has been no research to date on using a process mining perspective to detect anomalous flows in electrolytic coating production.
RDL Kalıcı Kaydıİncele - Conference ObjectYalnız künye2024Skin Cancer Detection Based on YOLOv8 Through A Mobile Application
Deniz, Numan; Tastimur, Canan
Skin cancer is a type of cancer that can cause death and has a high mortality rate. Detecting this disease in its early stages is extremely important to prevent the worst effects. However, detection by a dermatologist is time-consuming and costly. Computer vision and image processing techniques can help dermatologists understand medical images. Deep learning strategies have been repeatedly used in computer vision and image processing techniques. YOLO((You Only Look Once) is one of the well-known deep learning models used to solve detection cases for small, medium, and large objects. This study evaluates the performance of YOLOv8 for skin cancer detection using the HAM10000 dataset. The proposed method achieved an accuracy rate of 92.80 %. These results suggest the use of YOLO, a large object detection model, to detect skin cancer at an early stage, which can be considered a reference for future research.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2023Bağımsız Bileşen Analizi Kullanılarak İplik Gerginlik Sinyalinde Anormallik Tespiti
Akın, Erhan; Tastimur, Canan; Agrikli, Mehmet
Anormallik tespiti, verilerde beklenen davranışı göstermeyen kalıpların elde edilmesine karşılık gelmektedir. Bu uygun olmayan kalıplar, farklı uygulama etki alanlarında anormallikler, istisnalar, aykırı değerler veya kirleticiler olarak adlandırılır. Anormallik tespitinin önemi, çeşitli uygulama alanlarındaki verilerdeki anormalliklerin önemli bilgilere dönüşmesidir. İplik gerginlik sinyalleri incelendiğinde herhangi bir nedenle arıza yaptığı durumlarda bu sinyallerde anormallik durumları gözlemlenmektedir. Bu farklılıktan yola çıkarak iplik bükme makinesinde, bir arıza olup olmadığı önceden tespit edilebilir. Böylece daha büyük bir sorunun önüne geçilmiş olmaktadır. Anormallikleri tespit etmek için literatürde birçok yöntem kullanılmaktadır: Küme Tabanlı Algoritmalar, İstatistiksel Yöntem Algoritmaları vb. Bu çalışmada, bağımsız bileşen analizi kullanılarak, iplik makinelerinde bulunan iplik gerilim sinyalleri incelenmiş ve hatalı sinyal bölgeleri incelenmiştir. belirlenen. Önerilen yöntem diğer yöntemlerle karşılaştırılmış ve önerilen yöntemle en yüksek başarı oranı elde edilmiştir.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2023A Hybrid Classification Approach for Fasteners Based on Transfer Learning with Fine-Tuning and Deep Features
Akın, Erhan; Tastimur, Canan
Deep learning, which has seen frequent use in recent studies, has helped solve the problem of classifying objects of many different types and properties. Most studies both create and train a convolutional neural network (CNN) from scratch. The time spent training the network is thus wasted. Transfer learning (TL) is used both to prevent the loss of time due to training the dataset and to more effectively classify small datasets. This study performs classification using a dataset containing eighteen types of fastener. Our study contains three different TL scenarios. Two of them use TL with fine-tuning (FT), while the third does so with feature extraction (FE). The study compares the classification performance of eighteen different pre-trained network models (i.e., one or more versions of EfficientNet, DenseNet, InceptionResNetV2, InceptionV3, MobileNet, ResNet50, Xception, and VGGNet) in detail. When compared to other research in the literature, our first and second scenarios provide excellent implementations of TL-FT, while our third scenario, TL-FE, is hybrid and produces better results than the other two. Furthermore, our findings are superior to those of most previous studies. The models with the best results are DenseNet169 with an accuracy of 0.97 in the TL-FT1 scenario, EfficientNetB0 with 0.96 in TL-FT2, and DenseNet169 with 0.995 in TL-FE.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2022GMRNet: A Novel Geometric Mean Relation Network for Few-Shot Very Similar Object Classification
Tastimur, Canan; Akin, Erhan
With the widespread use of Deep Learning (DL), the use of DL has increased to provide a solution to the problem of object recognition and classification. In addition to classifying many different types of objects, the Deep Metrics Learning(DML) technique is effective in classifying objects that are visually very similar to each other. In this study, a novel Relation Network (RN) based DML has been designed to classify objects in two different datasets we created. We distinguished groups of objects that had a high degree of similarity to each other. These objects have been categorized using few-shot learning(FSL) since they are quite similar to one another. The impact of changing the number of classes and samples in the database on the network's performance has been studied. It is shown how the network's accuracy varies depending on the N-way (number of classes) and K-shots (number of samples) combinations used in its design. Additionally, the performance of the network has improved by an average of 15% thanks to the contribution of the recently introduced geometric mean module to the RN in our study. The accuracy rate of our recommended RN in screw and spare parts datasets is 96.1% and 92.3%, respectively. The first dataset consists of 1800 screw images with 18 classes, while the second dataset consists of 4100 spare parts images with 20 classes. The effectiveness of our method is expressed by the two datasets that we have extensively experimentally studied.
RDL Kalıcı Kaydıİncele - ArticleAçık Erişim2022Fastener Classification Using One-Shot Learning with Siamese Convolution Networks
Tastimur, Canan; Akin, Erhan
Deep Learning has been widely used in image-based applications such as object classification, object detection, and object recognition in recent years. Classifying highly similar objects is a very difficult problem. It is difficult to classify datasets in this situation where object similarity between classes and differences between classes are high. In this study, Siamese Convolution Neural Network, which is a similarity measurement-based network, has been practiced to classify 6 types of screws, 5 types of nuts, and 7 types of bolts that are very similar to each other. In addition, this neural network formed with the One-Shot Learning technique is trained. Thanks to the OSL technique, there is no need to use large data sets. Also, there is no need to use large amounts of data from each class. Adding a new class to be classified is also made easier by the use of the OSL technique. The performance results of the proposed method are manifested in detail in the article.
RDL Kalıcı Kaydıİncele - Conference ObjectYalnız künye2021A Novel Method Based on Deep Learning and Image Processing Techniques for Wearing Inspection on the Pantograph Surface
Tastimur, Canan; Karaduman, Gulsah; Akin, Erhan
Wear is an important problem in the pantograph-catenary system of electric rail transportation vehicles, which are increasingly used in the world, which receive their energy from the overhead line. This problem causes an increase in unplanned repair costs and disruption of transportation. The pantograph current collector strips move the locomotive with the electrical energy they receive from the catenary. Wear failure on the surface of the current collector strips creates problems in the pantograph-catenary interaction, causing transportation disruption and the safety of transportation is significantly affected. In this study, wear on the current collector strip surface has been detected by the noncontact monitoring method. A Convolution Neural Network (CNN) architecture has been developed to diagnose wear. Before the current collector strip surface images are trained on the CNN architecture, it is aimed to clarify the wear images and to make the diagnosis correctly with some pre-processing techniques on the images. After the pre-processing steps applied to the images, the performance results regarding the training of the dataset and its training with the data with the raw data have been compared, and more successful results have been obtained from the model trained using the pre-processed dataset.
RDL Kalıcı Kaydıİncele - ArticleYalnız künye2021Knitting machinery spare classification using deep learning with differential privacy
Tastimur, Canan; Kasap, Songul; Akin, Erhan
Given their widespread use, knitting machines must be maintained regularly. When the spare parts that make up these machines break down or become unusable, they must be replaced with new ones. However, the code/name information of the spare parts is not available to the end user, and can only be accessed with high-cost catalog procurement. Manufacturing companies keep the code/name information of such machine parts confidential. When the literature is examined, there are no studies in which spare parts are classified with machine learning–based algorithms. In line with this, this study focuses on the classification of spare parts using machine learning–based algorithms. The deep learning–based Convolutional Neural Network (CNN) architecture developed in this study can classify highly similar spare parts. In addition, since the code/name information received from the manufacturer and the spare part sample images require confidentiality, the CNN architecture has been developed in combination with the Differential Privacy (DP) method to present the DP-CNN method. As a result of the application of the Differential Privacy method, there has been no great loss of accuracy. This is an important development for our study. In the article, many optimizer algorithms are tested on the proposed method and comparative results are given. A 99.41% accuracy ratio has been obtained with the DP-RMSProp optimization method, which produces the best results. Experimental results of our study are presented in detail.
RDL Kalıcı Kaydıİncele - Conference ObjectAçık Erişim2021A New Bearing Fault Diagnosis Method using Envelope based Feature Extraction
Tastimur, Canan; Karakose, Mehmet; Aydin, Ilhan; Akin, Erhan
The performance of conventional intelligent diagnostic methods depends on the feature extraction of faulty signals. Feature extraction from defective signals requires signal processing techniques, expert knowledge, and human exertion. In this study, a time series based classification technique was developed to detect malfunctions occurring in induction motors. The proposed algorithm only takes the vibration signals as input. The phase space of the vibration signals taken as input is generated according to the time delay and embedding dimension for each motor status. Each motor condition is divided into two classes: faulty and intact. After the envelope transform is enforced into each signal, phase space is obtained. Then the Gaussian mixture model is formed. After with K-means clustering algorithm, the classification has been made up of four classes which are normal, outer race, inner race, and ball. The proposed approach has been tested on normal and defective vibration signals and results with high accuracy have been obtained.
RDL Kalıcı Kaydıİncele - Conference ObjectYalnız künye2019Vibration Signal Processing Based Bearing Defect Diagnosis with Transfer Learning
Tastimur, Canan; Karakose, Mehmet; Akin, Erhan; Varol, A.; Yazici, A.; Yazici, M.A.; Varol, C.; Aygunes, G.S.; Cotur, A.
It is very important to diagnose the fault condition of the bearing machines in order to make the machine run healthier. There are many successful studies in the detection of failures in the bearing machine made by conventional machine learning methods. However, these studies produce successful results in cases where the machines operate under the same condition and feature space is the same. Therefore, a deep learning-based diagnostic has been proposed for changing machine operating conditions. In the scope of our study, Keras and Tensorflow libraries, CNN network from scratch, VGG16 model and VGG19 deep learning models have been used for the classification of vibration images. Freeze the weights for transfer learning and remove the last fully connected layer to update the network in a problem-specific manner. The number of iterations and batch size has been determined by experimental studies. In this study, four faulty conditions have been successfully classified. While the accuracy rate of CNN network from scratch is 25%, the accuracy rate obtained by the VGG16 transfer learning method is 93% and the loss rate is 0.17% and the accuracy rate obtained by the VGG19 transfer learning method is 95% and the loss rate is 0.13%.
RDL Kalıcı Kaydıİncele