Tez Koleksiyonu

Bu koleksiyon için kalıcı URI

https://hdl.handle.net/20.500.12451/529

Listeleniyor 1 - 5 / 5

Damage detection in aircraft engine borescope inspection using deep learning
(Aksaray Üniversitesi Fen Bilimleri Enstitüsü, 2023) Uzun, İsmail; Tolun, Mehmet Reşit
Aircraft engine inspection is a key pillar of aviation safety by maintaining adequate performance standards to ensure the airworthiness of an engine. In addition, it is vital for asset value retention. Borescope inspection is currently the most widely used aircraft engine visual inspection method. However, borescope inspection is a time consuming, subjective, and complex process which heavily depends on the experience of the inspector as well as on their concentration level during inspection. On the other hand, cost saving of airlines and maintenance, repair, and overhaul (MRO) centers expose pressure and workload on inspectors. These make an automated system to support damage detection during borescope inspection necessary to avoid potential risks. Deep learning has found very wide application and has proven to be very successful during the last 10-15 years in the image recognition domain. In this research, we suggest a deep learning based automated damage detection framework from aircraft engine borescope inspection images. Faster R-CNN based deep learning model with Inception v2 feature extractor is utilized for the architecture. Due to the limited number of images, data augmentation and other overfitting methods are employed. The framework supports crack, burn, nick and dent damage types across all modules of turbofan engines. It is trained and validated with moderate to complex borescope images obtained from the field. The framework achieves 92.05% accuracy for nick or dent, 92.64% for crack and 81.14% for burn damage classes. Moreover, it achieves 88.61% average accuracy.
An intelligent system for detecting Mediterranean fruit fly [Medfly; Ceratitis capitata (Wiedemann)]
(Aksaray Üniversitesi Fen Bilimleri Enstitüsü, 2023) Uzun, Yusuf; Tolun, Mehmet Reşit
Nowadays, the most critical agriculture-related problem is the harm caused to fruit, vegetable, nut, and flower crops by harmful pests, particularly the Mediterranean fruit fly, Ceratitis capitata, named Medfly. Medfly's existence in agricultural fields must be monitored systematically for effective combat against it. Special traps are utilised in the field to catch Medflies which will reveal their presence and applying pesticides at the right time will help reduce their population. A technologically supported automated remote monitoring system should eliminate frequent site visits as a more economical solution. This paper develops a deep learning system that can detect Medfly images on a picture and count their numbers. A particular trap equipped with an integrated camera that can take photos of the sticky band where Medflies are caught daily is utilised. Obtained pictures are then transmitted by an electronic circuit containing a SIM card to the central server where the object detection algorithim runs. This study employs a faster region-based convolutional neural network (Faster RCNN) model in identifying trapped Medflies. When Medflies or other insects stick on the trap's sticky band, they spend extraordinary effort trying to release themselves in a panic until they die. Therefore, their shape is badly distorted as their bodies, wings, and legs are buckled. The challenge is that the deep learning system should detect Medflies of distorted shape with high accuracy. Therefore, it is crucial to utilise pictures that contain trapped Medfly images with distorted shapes for training and validation. In this academical study, the success rate in identifying Medflies when other insects are also present is 94.05%, achieved by the deep learning system training process, owing to the considerable amount of purpose-specific photographic data. This rate may be seen as quite favourable when compared to the success rates provided in the literature.
Adopting several models to alleviate sparsity and cold-start problems using data mining techniques for recommendation systems
(Aksaray Üniversitesi Fen Bilimleri Enstitüsü, 2021) Al-Sabaawi, Ali Mohsin Ahmed; Yenice, Yusuf Erkan; Karacan, Hacer
The development of Web 2.0 and the rapid growth of available data have led to the evolution of multiple systems among them the Recommendation Systems (RSs) which can handle the information overload. Due to the fact that, RSs performance is substantially limited by sparsity and cold-start problems; this research study takes the route for a major objective to attenuate these problems. To realize this objective, four data mining techniques are proposed namely: multi-steps resource allocation- singular value decomposition (MSRA-SVD), MSRA-SVD++, clustering community detection, and overlapping community detection. The first two models are dedicated to tackle the data sparsity problem whereas the rest models are adopted to alleviate cold-start users' problem. The core strategy of the first two models is to use the (MSRA) method to identify hidden relations in social network. the MSRA method is applied to determine the probability of their relation. If the probability exceeds a threshold, a new relationship will be established. For the second model (MSRA-SVD++), an implicit feedback source is exploited as an additional source of information, which can be extracted via rating information. Additionally, clustering and overlapping models are adopted to overcome cold-start user problem. The main idea of these models is to apply a clustering technique to group users into several communities. In order to attain that, explicit and implicit social relations with the confidence values are integrated to compute distance values. Additionally, the partitioning around medoids (PAM) clustering algorithm is adopted. Later, for the last model, the average distances of all clusters are computed. For all users, the distance between users and the center node of a particular cluster is computed. If the distance is less than the average of this cluster, a new user for this cluster will be added. Moreover, the SVD++ method is employed for each cluster to compute the prediction value. The proposed models are evaluated through the usage of three real-world datasets. Ultimately, findings exhibited a great deal of insights on how the proposed models outperformed a number of the state-of-the-art studies in terms of prediction accuracy.
Detection of photovoltaic panel faults with thermal camera and UV led
(Aksaray Üniversitesi Fen Bilimleri Enstitüsü, 2020) Coşgun, Atıl Emre; Tolun, Mehmet Reşit
In this thesis, hot spot failures, panel surface fractures, and yellowing problems on 40 Watt polycrystalline photovoltaic (PV) panel were analysed through indoor laboratory experiments and outdoor fieldwork. The outdoor field tests were performed with a 40 Watt PV panel on a clear and sunny day in the garden of Ortaköy Vocational School of Aksaray University. It is observed that the output power value of the panel decreases in the open circuit fault which was created by disconnecting the series-connected cells within the panel. In addition to the observed change in the power value, temperature increase value due to broken glass surfaces and disconnected cells on the panel was observed with a thermal camera. In the indoor laboratory experiments, artificial shading and yellowing parts detection studies were performed with UV LED and photodiode which is designed as a rectangular prism and placed at 45-degree angles. To perform these works on the photovoltaic (PV) panel, the user interface was developed and the graphical values of the manually scanned panel were plotted on the screen. Besides, colour changes on the PV panel were tried to be determined at different spectrum values. Blue, green, yellow and red LED have been used to do this. It is understood from the figure that the blue led spectrum is most effective method to obtain PV diagnosis. According to the results obtained, colour changes on the PV panel could not be obtained in colours other than blue. Additionally, common faults in PV systems which are shading situation, line to line faults, panel soiling problem, increasing resistance value problem between panels, problems caused by bird droppings or tree leaves and branches, and the effect of temperatures on the photovoltaic (PV) panels were simulated, analysed, and P-V and I-V graphics of the panels were drawn.
Path planning for 3 degree of freedom (RRR) robot and its application
(Aksaray Üniversitesi Fen Bilimleri Enstitüsü, 2020) Demir, Hasan; Tolun, Mehmet Reşit
Bu çalışmada, 3 serbestlik derecesine sahip RRR robotun yörünge planlaması yapılmış ve 3d yazıcı ile üretilen model robot kol üzerinde uygulanmıştır. Yörünge planlaması yapılırken genetik algoritmalar kullanılarak zaman optimizasyonu yapılmıştır. Model olarak ilk üç eklemin Kartezyen uzayda konumu etkilemesi nedeniyle üç eklemi de döner eklem olan RRR yapıdaki robot kol kullanılmıştır. Robot kolun öncelikle ileri kinematik analizi yapılmıştır. İleri kinematik analiz yöntemi olarak en çok kullanılan analiz yöntemi olan Denativ-Hartenberg yöntemi seçilmiştir. İleri kinematik analiz sonucunda elde edilen bilgiler ile ters kinematik analiz yapılmıştır. Ters kinematik analiz sonucunda model robotun Kartezyen uzayda bir noktaya ulaşması için gerekli olan eklem değişkenlerini elde edilmiştir. Kinematik analizler robotun eklem uzayında yörünge planlamasını yapmak üzere gerçekleştirilmiştir. Robot kolun yörünge planlaması yapılırken hareketlerini en kısa sürede tamamlaması için genetik algoritmalar kullanılarak zaman optimizasyonu yapılmıştır. Eklem uzayında yörünge planlama sonucunda her bir eklem için elde edilen hız ve ivme denklemleri optimizasyonda amaç fonksiyonları olarak kullanılmıştır. Amaç fonksiyonunun sınırlarını robot kolun eklemlerinde kullanılan servo motorların hız ve ivme değerleri oluşturmuştur. Optimizasyon sonucunda her bir eklemin hareketini en kısa zamanda tamamlayabileceği süreler bulunmuştur. Model robot kol 3d yazıcıda üretilmiş ve deney seti oluşturulmuştur. Optimizasyon sonucunda elde edilen veriler deneysel çalışmalardan elde edilen veriler ile karşılaştırılmıştır. Bu çalışmanın sonuncunda genetik algoritmalar kullanılarak zaman optimizasyonu yapılmış ve her hangi bir robota uygulanabilecek hareketi en kısa sürede tamamlayan bir model oluşturulmuştur.

Güncel Gönderiler