Yapay zekâ ve sinyal işleme yöntemleri ile rulmanlarda taşlama yanığı hatasının tespiti
Detection of grinding burn defect in bearings with artificial intelligence and signal processing methods
- Tez No: 825943
- Danışmanlar: DOÇ. DR. SEZGİN KAÇAR
- Tez Türü: Doktora
- Konular: Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Makine Mühendisliği, Mekatronik Mühendisliği, Computer Engineering and Computer Science and Control, Mechanical Engineering, Mechatronics Engineering
- Anahtar Kelimeler: Rulmanlarda Taşlama Yanığı Tespiti, Derin Öğrenme, Makine Öğrenmesi, Sinyal İşleme, Grinding Burn Detection in Bearings, Deep Learning, Machine Learning, Signal Processing
- Yıl: 2023
- Dil: İngilizce
- Üniversite: Sakarya Uygulamalı Bilimler Üniversitesi
- Enstitü: Lisansüstü Eğitim Enstitüsü
- Ana Bilim Dalı: Mekatronik Mühendisliği Ana Bilim Dalı
- Bilim Dalı: Mekatronik Mühendisliği Bilim Dalı
- Sayfa Sayısı: 115
Özet
Metal endüstrisinin ana işlemlerinden ve rulman üretiminin temel aşamalarından biri taşlamadır. Taşlama ile rulman parçalarının yüzey pürüzsüzlüğü ve boyutlandırma hassasiyeti artırılır. Ancak taşlama sırasında büyük miktarda ısı açığa çıkar. Bu ısının çoğu iş parçasına aktarılır. Bu durumda, iş parçasının yüzeyinde veya iç yapısında yanıklar oluşabilir. Bu yanıklar yüzey kalitesini bozmakta ve parçanın mekanik özellikleri zarar görmektedir. Yüzeyinde veya iç yapısında taşlama yanığı olan rulman parçaları, faydalı kullanım ömrün dolmadan kırılmakta veya çatlamaktadır. Rulmanlar dönen makinaların en temel yataklama elemanları olduğundan, rulmanda meydana gelen bir arıza makinenin bozulmasına sebep olacaktır. Bu nedenle rulman üretim süreçlerindeki kontrol aşamalarında taşlama yanığının tespiti önemlidir. Taşlama yanığını tespit etmek için rulman üreticileri, bir çoğu tahribatlı yöntemler olan metalurjik yöntemler, gözle muayene veya sinyal işleme gibi dolaylı yöntemler kullanmaktadır. Metalurjik yöntemler gibi doğrudan yöntemler kesine en yakın sonuç vermesine karşın pahalı, zaman alıcı ve işleme sürecini durdurmayı gerektiren yöntemlerdir. Sinyal işleme gibi dolaylı yöntemler ise mutlak sonuç vermemekle birlikte ucuzdur, gerçek zamanlı sonuçlar üretebilirler ve tahribatsız yöntemlerdir. Bu tezde, rulman parçalarında taşlama yanığının tespiti için sinyal işleme, derin öğrenme ve makine öğrenmesi temelli bir yöntem önerilmektedir. Öncelikle gerçek üretim koşullarında, bir taşlama makinasında, rulman parçaları taşlanırken, akustik emisyon sensörleriyle sinyaller toplanmıştır. Daha sonra bu sinyaller ya sinyal olarak kaydedilmiştir veya zaman-frekans gösterim yöntemleriyle görüntülere dönüştürülmüştür. Son adımda bu sinyal ve görüntüler derin öğrenme ve makine öğrenmesi yöntemleri ile hatalı-normal diye 2 sınıfa ayrılarak sınıflandırılmıştır. Sinyal olarak sınıflandırmak için uzun dönemli kısa süreli bellek, destek vektör makineleri ve en yakın K komşuluğu yöntemleri kullanılmıştır. Bu yöntemlerden elde edilen sonuçlara göre %93 doğruluk oranına ulaşılmıştır. Zaman-frekans görüntüleri ise farklı transfer learning ağları (Googlenet, Resnet-50, Squeezenet ve Alexnet) ve tasarlanan yeni bir evrişimli sinir ağı kullanılarak yüksek doğrulukla sınıflandırılmıştır. Görüntü sınıflandırma uygulamalarında ise %100 doğruluk oranlarına ulaşılmıştır. Bu hem literatürdeki karşılaştırılabilir çalışmalara kıyasla hem de üretim süreçlerindeki testlere göre iyi bir sonuçtur. Gerçek üretim koşullarından elde edilen verilerle, yapay zekâ ve sinyal işleme yöntemleri kullanılarak rulman parçalarında taşlama yanığının tespit edilmesi tezin literatüre temel katkısı olarak gösterilebilir.
Özet (Çeviri)
Grinding is one of the main processes of the metal industry and one of the key stages of bearing production. Grinding is a high and specific energy machining process widely used in the manufacture of components that require fine tolerances and smooth surfaces. Grinding improves the surface smoothness and dimensional accuracy of bearing parts. Almost all of the energy expended in the grinding process is converted into heat and transferred to the workpiece. This raises the temperature of the work surface. This can cause burns on the surface or internal structure of the workpiece and adversely affect the mechanical properties of the workpiece, such as tensile stress. These burns deteriorate the surface quality and the mechanical properties of the part suffer. Grinding burns are one of the main constraints when trying to improve the surface finish of the components being ground, because any change that improves the surface finish also increases the risk of grinding burns. Grinding burn is the oxidation of the machining surface caused by the high temperature generated on the workpiece surface during grinding. The causes affecting the formation of grinding burn can be listed as follows. • Abrasive: The grain size and material of the abrasive used affect the formation of grinding burn. Fine grain abrasives increase the formation of grinding burn. • Material to be machined: As the hardness of the material increases, the risk of grinding burn also increases. • Application parameters: Parameters such as cutting speed, feed rate, contact pressure affect the formation of grinding burns. Increasing the cutting speed increases the risk of grinding burns, while decreasing the feed speed increases the risk of grinding burns. The effect of contact pressure on grinding burn is limited. Since bearings are the most basic bearing elements of rotating machines, a bearing failure will affect the failure of the machine. Bearing parts with grinding burns on the surface or internal structure break or crack before the end of their useful life. Therefore, it is important to detect grinding burns during the control stages of bearing manufacturing processes. Grinding burn detection is critical in bearing manufacturing to ensure that the final product is of high quality and can perform its intended function efficiently. If grinding burns are not detected during the manufacturing control stages, they can cause significant damage to the bearings, resulting in their failure and, eventually, the failure of the entire machine. As a result, detecting grinding burns early in the manufacturing process can aid in the prevention of such failures. Metallurgical methods and visual inspection are commonly used methods for detecting grinding burns in bearing manufacturing. These methods are called direct detection methods. These are costly, time-consuming, and incapable of detecting threats in real time. For example, metallurgical methods necessitate the use of specialized equipment and trained personnel to examine bearing material samples, which is a time-consuming process that can cause delays in the manufacturing process. Similarly, visual inspection is dependent on the operator's visual acuity and experience, both of which can be subjective and error-prone. In summary, direct detection methods provide the closest accuracy, but are expensive, time-consuming and disadvantageous for real-time implementation. In contrast, there is another detection method, called indirect detection methods, which is based on the measurement of the effects caused by the defect. In this method, the change in phenomena such as vibration, acoustic emission, force, electric current and voltage is measured and the defect is detected using signal processing, image processing and classification tools. Although this method is based on some estimation algorithms, it is an inexpensive, real-time and non-destructive detection method. According to this method, signals are received with sensors related to the measured phenomenon. The received signals are digitized and transferred to the computer. These data about the measured phenomenon are evaluated with signal processing methods, image processing methods or artificial intelligence based algorithms to detect faulty signals. This is often a classification problem that requires the separation of erroneous signals from error-free signals. For this reason, signal or image classification methods are used after the signals are acquired. One of the most commonly used measurement parameters in indirect methods is acoustic emission. Acoustic Emission are transient elastic waves produced by the rapid release of energy from a local source inside the material. Small frequency ranges can be successfully detected by acoustic emission measurement. It is a wise choice to use acoustic emission measurement for the detection of a fault type that affects particularly small frequency ranges. In this thesis, a method based on signal processing, deep learning and machine learning is proposed for the detection of grinding burn in bearing parts. The proposed method begins with the use of acoustic emission sensors to collect signals during the grinding process of bearing parts under real-world production conditions. The acoustic emission data used in this study came from experiments carried out in the Ortadou Rulman Sanayi Research and Development laboratories. Experiments on outer ring grinding and roller grinding were carried out. By varying 10 different parameters of the grinding machine, 101 different measurements were obtained. During the experiments, measurements were taken from the grinding of 41 roller and 60 outer ring specimens. Of these measurements, 18 were of defective specimens and 83 were of normal specimens. The machine parameters were recorded for each measurement. During the measurements, the machining characteristics of the machine are changed according to the production scenario. Measurements were taken by changing 10 different features of the machine such as cutting speed, cutting pressure, sharpening frequency.Signals were obtained using a 4MHz sampling frequency. The bearing parts to which the collected signals belong to are recorded. These bearing parts were then analyzed by some metallurgical methods and separated into those with grinding burns and those without grinding burns. Thus, it was determined which of the acoustic emission signals belonged to the sample with grinding burns and which belonged to the sample without grinding burns. Acoustic emission is a non-invasive monitoring technique that detects high-frequency stress waves produced by machining materials. These signals contain information about the material being machined and can be used to detect changes in its properties. The next step in the proposed method is signal preprocessing. In this step, the signals are first divided into smaller sub-signals, where the faulty signals are divided into smaller sub-signals, and then the faulty and non-faulty signals are combined into a single data file and labeled with class names. These signals are then classified using machine learning methods specialized for signal classification and the long-term short-term memory method developed in this thesis. Then, after some preprocessing steps, these signals were converted into images either as signals or with time-frequency representation methods and classified into 2 classes as faulty-normal. The images are then classified using deep learning techniques. For signal classification, the thesis employs a variety of algorithms, including Long Short-Term Memory (LSTM), Support Vector Machines (SVM), and K-Nearest Neighbor (KNN). According to the results obtained from these methods, 93% accuracy rate was achieved. For image classification, the thesis employs a variety of algorithms, including transfer learning networks such as Googlenet, Resnet-50, Squeezenet, and Alexnet. Additionally, a new CNN network is suggested in this thesis to categorize time-frequency pictures of signals obtained from bearing samples. In image classification applications, 100% accuracy was achieved. This is a good result both compared to comparable studies in the literature and to tests in production processes. In some applications, there is either insufficient training data or additional training instances are not feasible. In these cases, pre-trained networks can be used. In order to achieve successful results in deep learning studies, a good data set is needed. For this reason, institutions that pioneer deep learning studies are conducting some studies to create data sets that can be used in these studies. In deep learning studies, one of the measures of the reliability of the method studied is the reliability of the data set. In this respect, training and testing one's own network using a data set widely used in the literature is a necessary step for the success of the related method. Transfer learning methods are preferred by researchers in terms of meeting this need. Transfer learning is the use of previously trained neural networks. Googlenet, Resnet-50, Squeezenet, and Alexnet deep CNNs were pretrained for image recognition and then used to classify grinding burn signals using a time-frequency representation in this study. These networks perform a new task by utilizing transfer learning. The most common method for implementing transfer learning is to use pre-trained networks on the ImageNet data set. In addition, in this thesis, a new CNN network is proposed to classify time-frequency images of signals obtained from bearing samples. The number of layers (depth of the network) and the width of each layer of the CNN network are considerably less than the transfer learning networks used in the thesis, which reduces the computational load of the network. Despite this, it has been shown that the accuracy values achieved in transfer learning networks are also achieved in the designed CNN network. Achieving similar results with lower processing capacity is the prominent aspect of this network. All these algorithms are trained using a labeled dataset of images with and without grinding burns, allowing them to learn the features that distinguish between the two categories. Time-frequency images obtained from the signals were classified with high accuracy using different transfer learning networks (Googlenet, Resnet-50, Squeezenet and Alexnet) and a newly designed convolutional neural network. In most of the methods, 100% classification success was achieved. Overall, the thesis's proposed method provides a non-destructive, low-cost, and real-time approach to detecting grinding burns in bearing parts. The method can provide highly accurate results by combining signal processing, deep learning, and machine learning techniques, reducing manufacturing time and cost while ensuring high-quality bearings that are reliable and long-lasting. The main contribution of this thesis to the literature is the detection of grinding burns in bearing parts using artificial intelligence and signal processing methods with data obtained from real production conditions.
Benzer Tezler
- Improved fuzzy logic based edge detection method on clinical images
Klinik görüntülerde bulanık mantık temelli iyileştirilmiş kenar tespit yöntemi
MURAT MERT ÇELEN
Yüksek Lisans
İngilizce
2022
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik ÜniversitesiKontrol ve Otomasyon Mühendisliği Ana Bilim Dalı
DR. ÖĞR. ÜYESİ İLKER ÜSTOĞLU
- Sayısal haritalama teknikleri kullanılarak DNA dizilimleri üzerinden lösemi hastalığının temel türlerinin yapay zeka tabanlı algoritmalar ile sınıflandırılması
Classification of main types of leukemia disease with artificial intelligence-based algorithms on the DNA sequences using digital mapping techniques
FATMA AKALIN
Doktora
Türkçe
2023
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolSakarya ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
PROF. DR. NEJAT YUMUŞAK
- Aşırı öğrenme makineleri ile beta talasemi ve demir eksikliği anemisinin ayırt edilmesi
Discrimination of β-thalassemia and iron deficiency anemia through extreme learning machine and regularized extreme learning machine based decision support system
BETÜL SAYIN
Yüksek Lisans
Türkçe
2021
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolFırat ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
DOÇ. DR. TANER TUNCER
- Real-time crash risk analysis using deep learning
Derin öğrenmeyle gerçek zamanlı kaza risk analizi
SAEID MORADI
Yüksek Lisans
İngilizce
2022
Mühendislik Bilimleriİstanbul Teknik Üniversitesiİnşaat Mühendisliği Ana Bilim Dalı
PROF. DR. ALİ OSMAN ATAHAN
- Akciğer seslerinin makine öğrenmesi yöntemleri ile sınıflandırılması
Classification of lung sounds using machine learning techniques
MAHMUD ESAD ARAR
Doktora
Türkçe
2024
BiyomühendislikYıldız Teknik ÜniversitesiElektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
PROF. DR. HERMAN SEDEF