Geri Dön

CFRP plakalarda delaminasyon hasarının makina öğrenmesi ile tahmin edilmesi

Predicting delamination failure in CFRP composite plates with machine learning algorithms

PDF İndir

  1. Tez No: 849656
  2. Yazar: AMMAR TARIK DİNÇER
  3. Danışmanlar: PROF. DR. MUSTAFA BAKKAL
  4. Tez Türü: Yüksek Lisans
  5. Konular: Makine Mühendisliği, Mechanical Engineering
  6. Anahtar Kelimeler: Belirtilmemiş.
  7. Yıl: 2024
  8. Dil: Türkçe
  9. Üniversite: İstanbul Teknik Üniversitesi
  10. Enstitü: Lisansüstü Eğitim Enstitüsü
  11. Ana Bilim Dalı: Makine Mühendisliği Ana Bilim Dalı
  12. Bilim Dalı: Malzeme ve İmalat Bilim Dalı
  13. Sayfa Sayısı: 97

Özet

Karbon elyaf takviyeli polimer (CFRP) kompozit malzemeler, üstün mekanik özellikleri dolayısıyla havacılık ve uzay sanayisi başta olmak üzere pek çok alanda sıklıkla tercih edilmektedir. Mekanik özelliklerin yanı sıra geleneksel imalat yöntemleri ile çok parçalı olarak üretilebilen parçaları entegre olarak üretmeye imkan tanıyan kompozit malzemeler montaj içerisindeki toplam parça sayısında da düşüş sağlamaktadır. Bu düşüşe karşına yine de diğer malzemeler gibi kompozit parçaların da neredeyse tüm montaj arayüzlerinde vidalı bağlantı elemanları kullanılmaktadır. Bu nedenle CFRP parçaların, bilhassa uçak kabuğu gibi yüksek bağlantı arayüzü olan bölgelerdeki parçalar, üretiminde en çok kullanılan imalat yöntemlerinden birisi delik delmedir. Kompozit malzemeler, tıpkı metalik malzemeler gibi kendilerine has bazı hasar modlarına sahiptirler. CFRP malzemelerde delik delme söz konusu olduğunda en sık karşılaşılan hasar modlarından biri delaminasyon hasarıdır. Kompozit katmanlarının birbirinden ayrılması ile meydana gelen bu hasar, delme esnasındaki itme kuvveti ve sıcaklık gibi çeşitli fiziksel niceliklere bağlıdır. Bu fiziksel büyüklükler ise ilerleme hızı, takım çapı ve kesme hızı gibi parametrelere bağlıdır. Bu çalışmada, izah edilen 2 kademeli nedensellik ilişkisini analiz edebilmek; kesme parametreleri, fiziksel büyüklükler veritabanı ve delaminasyon faktörleri arasında bir öğrenme modeli kurmak amaçlanmıştır. Bu amaçla, 5 ilerleme hızı, 5 kesme hızı ve 3 takım çapından oluşan toplam 75 deliği kapsayan bir deney matrisi oluşturulmuş ve bu matris hem direkt delme hem gagalama delme prosesleri için tekrarlanmıştır. Toplam 150 delik için delme esnasında kuvvet, moment, ivme ve sıcaklık verileri sensörler yardımıyla toplanmıştır. Yeterli bir veritabanının oluşması için gerekli son unsur olarak, tüm delikler için takım giriş ve takım çıkış yüzeylerinden mikroskop yardımıyla delaminasyon çapları ölçülmüştür. Yüksek ölçüm frekansında uzun zaman serileri olarak elde edilen sensör sinyalleri, ham haliyle alınmış, uygun şekilde filtrelenmiş, downsampling işlemi uygulanmış ve istenen düzene göre istif edilmiştir. İşlemeye hazırlanan 4 fiziksel nicelik veriseti üzerinden toplam 14 öznitelik türetilmiştir. Delaminasyon çapları da giriş ve çıkış yüzeyleri için 1D delaminasyon faktörlerine dönüştürülmüştür. Bu şekilde veritabanı hazır hale getirilmiştir. İki kademeli ilişkiyi açıklamak için artarda entegre olacak şekilde iki farklı regresyon modeli oluşturulması düşünülmüş ve bu amaçla; karar ağacı ve polinom regresyonu zayıf öğrenicileri, destek vektör makinaları, güçlendirme algoritmaları, toplu öğrenme modelleri ve LSTM derin öğrenme algoritması kullanılmıştır. Her bir modelin hiperparametre optimizasyonu yapılmıştır. Optimizasyon sonrası koşturulan modellerin tahmin performansları, ilgili metrikler üzerinden açıklanmış ve sayısallaştırılmıştır. Makina öğrenmesi modellerinden ayrı olarak, eldeki direkt delme ve gagalama delme verileri üzerinden CFRP plakalarda gagalama delme işleminin, direkt delmeye kıyasla ne gibi değişikliklere neden olduğu, açığa çıkan fiziksel niceliklerde ve delaminasyon hasarının şiddetinde ne mertebede değişiklikler olduğu tespit edilmiştir. Son olarak ilgili öğrenme modellerinin hesaplama maliyeti hesaplanmıştır. Bunun için tüm modellerin, her iki regresyon safhası için de koşturma süreleri hesaplanmış ve kıyaslanmıştır. Sonuç olarak; hazırlanan veritabanı üzerinden %3'e kadar düşen hata oranları ile delaminasyon faktörünü tahmin edebilen bir öğrenme modeli geliştirilmiş, delik delmede aşırı hasar oluşumu ve sınır kesme değerlerine ilişkin bulgular paylaşılmış, CFRP malzemelerde direkt delme ve gagalama delme proseslerinin çıktılarında ne gibi farklara neden oldukları açıkça ifade edilmiştir.

Özet (Çeviri)

Carbon fiber-reinforced polymer (CFRP) composite is one of the most commonly used composite materials. Due to its superior mechanical properties, it is frequently used as a structural material in different fields, especially in aerospace industries. Another characteristic advantage of CFRP is reduction of total amount of parts in an assembly. Wide variety and nature of composite manufacturing techniques provide chance to produce complex geometries as integral parts and decrease the number of used fasteners which is not possible with traditional manufacturing methods. Decrease of total amount of parts result with advantages like weight reduction and assemblability. Despite the reduction in the number of total parts, big constructions like aircrafts have many assembly interfaces and many fasteners to be used as there are tens of thousands of fasteners used in a modern airliner with high ratio of composite parts. Therefore, drilling of CFRP composite emerges as a critical process. Depending on the fibrous structure of composite materials, they have some unique failure modes. One of the most common failure modes of composites is delamination which is an interlaminar phenomenon occurs by separation of laminates. In drilling process, delamination frequently emerges with different levels of severity. Occurrence and severity of delamination depends on different parameters such as thrust force and drilling temperature. These physical quantities are directly related to the cutting parameters like feed rate, cutting speed, tool diameter and tool material. A comprehensive study on this field is required to be examining the delamination failure, physical quantities and cutting parameters. Delamination problem is defined as the two-staged causal relationship between these three variable sets. To solve the aforementioned problem, various machine learning algorithms are utilized. Through this a statistical solution method is presented in this study. Unlike the analytical or numerical solution methods, here a data-driven solution is utilized. Therefore, there are different operations achieved in all phases of the study. Another topic studied here is pecking of CFRP plates. Pecking/ peck drilling is a type of hole drilling generally used in deep drilling. However this method is preferred for facilitating chip removal and preventing chip packing, it also has advantageous effects on thrust force and accuracy. For this reason, peck drilling is involved in the scope of this study and the effect of peck drilling on both the physical quantities and delamination failure is investigated. In the experimental phase of the study, a proper experiment matrix is created which is consisting of cutting parameters cutting speed, feed rate and tool diameter. In this matrix there are 5 different feed rates, 5 cutting speeds and 3 tool diameters. This matrix is repeated for direct drilling and peck drilling which makes 150 holes in total. For all these holes, CFRP plates and PCD (polycrystalline diamond) tools are used. For quantizing and evaluating the severity of delamination failure occurred, it is decided to utilize 1D delamination factor. Accordingly, delamination diameters are read by microscope from all drilled holes and both of the enter and exit side surfaces. Thus, it is aimed to examine the difference between push-down and peel-up delaminations in this specific material and to evaluate both delamination factors together to prevent the possible manipulation caused by stochastic nature of process itself. While drilling, multi-sensor measurements are made. Thrust force, moment, acceleration and drilling surface temperature are measured and recorded. Acquiring data from sensors and bringing them together with cutting parameters and delamination factors, a raw database is created. For effective use of sensor time-series, a pre-processing phase is required. In this study, sensor data is filtered through various filters according to the Fourier transform analysis. Through this, noise in the sensor data is cleaned and outliers in series are removed. After filtering, downsampling is applied to shorten the series, which is too long initially. In final step of pre-processing, all the database is formatted and organized to make it convenient for direct use in learning algorithms. Machine learning applications are increasingly used for solution of engineering problems and increasing efficiency of industrial operations. Especially in problems requiring predictive perspective such as tool condition monitoring or failure detection, machine learning algorithms are proven to be useful. In this study, learning algorithms are utilized for predicting delamination factors by input cutting parameters. In ML modelling, two-staged model framework is designed. In this framework, Stage-1 is a regression model between cutting parameters (as input) and physical quantity feature set (as output). Stage-2 is also a regression model but between physical quantity feature set (as input) and delamination factors (as output). After building and optimizing each independent regression models, they are connected by a function or a pipeline in Python. Thus, the integrated model is created. Instead of that it was possible to do it in a single regression model with only cutting parameters and delamination factors, it would not be mechanically reasonable. Because relation between these two variable sets is indirect. The fundamental assumption or approach here can be briefly explained as; cutting parameters directly cause physical quantities and these quantities directly cause the delamination failure. To include this mechanically reasonable perspective in the model, this framework is designed. Firstly, Decision Tree Regression (DTR), Linear Regression (LR) and Polynomial Regression (PR) algorithms are tested with prepared database. It has been seen that DTR gives better results with lower error rate. Therefore, it is decided to move on with DTR as a weak learner in next ensemble learning models. In the second part, DTR, Support Vector Mahine (SVR) and Random Forest Regression (RFR) models are created. Hyperparameter optimization is applied for all these models separately. As RFR being an ensemble learning algorithm derived from DTR, it was expected to generate better results and it did so. In the last phase of ML modelling, boosting algorithms are utilized. Gradient Boosting Regression (GBR) and AdaBoost Regression (ABR) models are created, optimized and run. Results has increased accuracy significantly compared to any other model tested in this study. This improvement is valid for both stages of integrated model. Beside the explained ML applications, Deep Learning (DL) techniques are considered very useful for large databases and solving complex relations. In current predictive manufacturing applications, many variations of neural networks are used and developed. Neural networks are the most fundamental representation of DL algorithms and they are proven to be effective in such applications. Depending on the topic, neural network type and details of framework vary. As a last learning model test, it is aimed to include and utilize deep learning techniques in the model. To this end, it is planned to extract latent features by deep learning model and use it in the integrated model instead of the former extracted statistical features in machine learning models. Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models are tested in this purpose. Firstly, it has been seen that CNN is not efficiently working in this way. Since the main capability of CNN is applications like image processing and object recognition, it was a predictable result. Beside this, LSTM, as a great algorithm for detecting features in sequences like time-series or language modelling, has worked well enough to obtain reasonable predictions but not to reach the accuracy level of applied boosting algorithms. DL model results can be considered as disappointing when compared to ML models. But it is also known that DL algorithms work efficiently with very big databases unlike the one in this study. Beside this, using DL algorithms only to extract features is quiet rare method in literature. Here, LSTM model is built to extract latent features and embedded into the two-staged integrated ML model. Beside the evaluation of models, it is clearly shown that there is a significant decrease in thrust force and moment values when holes are drilled with peck drilling method. The rate of decrease is far above the expected values. Delamination failure is also decreased in peck drilling. On the enter side of CFRP plate, delamination factor is decreased %12.3 in average and on the exit side it decreased %16. As a characteristic of delamination failures in this study, it has been seen that severity of delamination failure is greater on the exit surface. This situation has 2 reasons; first one, manufacturing method of specimen plates cause different properties on both surfaces. Second reason is the mechanical difference between peel-up and push-down delamination types. Beside final metrics, importance factors are generated throughout the tested models. Due to the unique and stochastic nature of each learning model, importance factors are varying in between. But from Stage-1 it is turned out that, feed rate is the 'most important' feature determining the measured physical quantities. In the Stage-2, feature importances from different models are averaged and the most important five features have been determined as average thrust force, maximum temperature, maximum thrust force, average moment and average temperature above the threshold value. Considering all tests and results, it has been seen that GBR is the best model with lower error rates. In first phase, RFR was the best model with %7.6 mean absolute percentage error (MAPE) on enter side and %10 MAPE on the exit side. But GBR has approximately %3 on enter surface and %4 on exit surface. To evaluate models, different metrics and residual values are checked and compared step by step.

Benzer Tezler

  1. Tez No

    765690

    Karbon fiber takviyeli kompozit malzemenin farklı kalınlıkta destek plakalar kullanılarak delinmesi

    Drilling of carbon fiber reinforced composite material using support plates of different thickness

    ÇAĞLAR BETGÜL

    Yüksek Lisans

    Türkçe

    Türkçe

    2022

    Makine MühendisliğiKaramanoğlu Mehmetbey Üniversitesi

    Makine Mühendisliği Ana Bilim Dalı

    PROF. DR. UĞUR KÖKLÜ

  2. Tez No

    595951

    Experimental and numerical investigation of damage induced by transverse impact in composite beams and plates

    Kompozit kiriş ve plakalarda düzlem dışı darbeye bağlı oluşan hasarın deneysel ve sayısal olarak incelenmesi

    MİRAÇ ONUR BOZKURT

    Yüksek Lisans

    İngilizce

    İngilizce

    2019

    Havacılık MühendisliğiOrta Doğu Teknik Üniversitesi

    Havacılık ve Uzay Mühendisliği Ana Bilim Dalı

    DOÇ. DR. DEMİRKAN ÇÖKER

    PROF. DR. KEMAL LEVEND PARNAS

  3. Tez No

    830780

    Karbon fiber takviyeli havacılık kompozitlerinin delme proseslerinde nanografen katkılı mmy soğutma metodunun delik kalitesine etkilerinin incelenmesi

    Examination of the effects of nanographene-enhanced mql cooling method on hole quality in drilling processes of carbon fiber reinforced aviation composites

    MUSTAFA BURAK SAĞENER

    Yüksek Lisans

    Türkçe

    Türkçe

    2023

    Makine MühendisliğiEskişehir Teknik Üniversitesi

    Makine Mühendisliği Ana Bilim Dalı

    PROF. DR. OĞUZ ÇOLAK

  4. Tez No

    518958

    Aşındırıcılı su jeti ile karbon fiber takviyeli plastik (CFRP) kompozit malzemenin delinmesi ve kesilmesi

    Carbon fiber reinforced plastic (CFRP) composite materials drilling and cutting with abrasive water jet

    MELTEM ALTIN KARATAŞ

    Doktora

    Türkçe

    Türkçe

    2018

    Makine MühendisliğiKarabük Üniversitesi

    Konstrüksiyon ve İmalat Ana Bilim Dalı

    PROF. DR. HASAN GÖKKAYA

  5. Tez No

    882902

    Betonarme kirişlerin karbon plaka ve kumaşlarla güçlendirilmesi

    Strengthening of reinforced concrete beams with carbon plates and fabrics

    AHMET KAĞAN UZEL

    Yüksek Lisans

    Türkçe

    Türkçe

    2024

    İnşaat MühendisliğiGazi Üniversitesi

    İnşaat Mühendisliği Ana Bilim Dalı

    PROF. DR. SABAHATTİN AYKAÇ

Geri Dön