Geri Dön

Kauçuk hamuru üretiminde makine öğrenmesi ve hibrit yapay zeka algoritmalarının uygulanması ile süreç optimizasyonu

Process optimization in rubber compound production through machine learning and hybrid artificial intelligence algorithms

PDF İndir

  1. Tez No: 1016021
  2. Yazar: ZEYNEP URUK
  3. Danışmanlar: DOÇ. DR. ALPER KİRAZ
  4. Tez Türü: Doktora
  5. Konular: Endüstri ve Endüstri Mühendisliği, Industrial and Industrial Engineering
  6. Anahtar Kelimeler: Belirtilmemiş.
  7. Yıl: 2025
  8. Dil: Türkçe
  9. Üniversite: Sakarya Üniversitesi
  10. Enstitü: Fen Bilimleri Enstitüsü
  11. Ana Bilim Dalı: Endüstri Mühendisliği Ana Bilim Dalı
  12. Bilim Dalı: Endüstri Mühendisliği Bilim Dalı
  13. Sayfa Sayısı: Belirtilmemiş.

Özet

Bu tez çalışması, kauçuk endüstrisinde makine öğrenimi ve optimizasyon tekniklerinin sağladığı avantajları inceleyen üç ayrı araştırma sunmaktadır. İlk çalışma, kauçuk hamurunun reometrik özelliklerini tahmin etmek amacıyla Yapay Sinir Ağları (YSA) ve iki hibrit yaklaşım olan Yapay sinir ağları-Parçacık Sürü Optimizasyonu ve Yapay sinir ağları-Genetik Algoritma tekniklerini kullanmıştır. Üç katmanlı bir Çok Katmanlı Algılayıcı ile gerçekleştirilen bu çalışmada, YSA ağları Levenberg-Marquardt geri yayılım algoritması ile eğitilmiş, hibrit algoritmalarda ise nöron ağırlıkları ve önyargıları Parçacık Sürü Optimizasyonu ve Genetik Algoritma ile optimize edilmiştir. Proses koşulları ve hamur formülasyonu gibi değişkenler kullanılarak yapılan analizlerde, hibrit yaklaşımların standart YSA'dan daha yüksek performans gösterdiği ve aşırı sapma noktalarını daha iyi bir şekilde ortadan kaldırdığı gözlemlenmiştir. İkinci çalışma, benzer reometrik özelliklerin tahmini için YSA'nın yanı sıra Gauss Süreç Regresyonu (GPR) ve Destek Vektör Regresyonu (SVR) algoritmalarını kullanmıştır. Bu çalışmada daha geniş bir veri seti ve ek proses değişkenleri ile tahminler yapılmıştır. Sonuçlar, GPR ve SVR algoritmalarının tahmin performanslarının birbirine yakın olduğunu, ancak YSA'nın minimum tork (ML) ve %60 pişme süresi (t60) özellikleri için, SVR ve GPR'nin ise maksimum tork (MH) ve %30 pişme süresi (t30) tahminleri için daha iyi sonuçlar verdiğini ortaya koymuştur. Son olarak, üçüncü çalışma kauçuk hamuru bileşenlerini optimize ederek maliyetleri düşürmeye odaklanmıştır. Plackett-Burman ve Box-Behnken tasarımları kullanılarak hamur bileşenleri ile teknik özellikler arasındaki ilişkiler belirlenmiş ve SVR ile entegre Genetik Algoritma kullanılarak maliyet optimizasyonu yapılmıştır. Yapılan analizlerde, hamur maliyeti 2.009 €/kg'dan 1.989 €/kg'a düşürülmüş ve toplam maliyet %1,4 oranında azaltılmıştır. Elde edilen sonuçlar, önerilen kauçuk hamurunun istenen teknik özellikleri sağlarken maliyetleri de düşürdüğünü göstermektedir. Bu üç çalışma, kauçuk endüstrisinde makine öğrenimi ve optimizasyon tekniklerinin ürün kalitesini artırma, üretim süreçlerini iyileştirme ve maliyetleri düşürme konusundaki potansiyelini ortaya koymakta ve veri odaklı karar verme süreçlerinin önemini vurgulamaktadır. Hibrit yaklaşımlar ve çeşitli regresyon tekniklerinin entegrasyonu, tahmin doğruluğunu artırırken, optimizasyon algoritmaları maliyetleri düşürmede etkili sonuçlar sağlamıştır. Bu araştırmalar, gelecekteki çalışmalara yönelik yeni araştırma yolları sunmaktadır.

Özet (Çeviri)

In the rapidly evolving rubber industry, enhancing product quality and optimizing production processes are crucial for maintaining competitiveness and achieving cost efficiency. Traditional methods of predicting rheometric properties and managing production costs have relied heavily on empirical knowledge and trial-and-error approaches, often leading to inefficiencies and suboptimal outcomes. The advent of machine learning and optimization techniques offers a promising alternative, providing data-driven insights and precise predictions that can significantly improve both product quality and manufacturing efficiency. This thesis explores the application of machine learning and optimization techniques in the rubber industry through three comprehensive studies, illustrating their significant advantages in improving rheometric property predictions and reducing production costs. The research leverages various machine learning methods, including Artificial Neural Networks (ANN), hybrid approaches integrating ANN with Particle Swarm Optimization (PSO) and Genetic Algorithms (GA), as well as other advanced regression techniques. Additionally, experimental design methods such as Plackett-Burman and Box-Behnken designs are employed to systematically investigate the relationships between component amounts and technical specifications of the rubber compound. By integrating these sophisticated techniques, this research aims to demonstrate their effectiveness in refining rheometric property predictions and reducing production costs, ultimately contributing to more efficient and sustainable practices in the rubber industry. First study investigates the prediction of rheometric properties of a rubber compound using ANN, along with two hybrid approaches ANN-PSO and ANN-GA. A three-layer Multi-Layer Perceptron (MLP) is employed across these methods. For the ANN model, the network is trained using the Levenberg-Marquardt backpropagation algorithm. In contrast, the hybrid approaches optimize the weights and biases of each neuron via PSO and GA. The input variables include process conditions such as time spent, energy absorbed, and maximum temperature reached during each stage of the process, as well as compound formulation parameters like the amount of masterbatch and chemicals (curing agents and accelerators) in kilograms. The output variables to be predicted are minimum and maximum torque (ML and MH), scorch time (ts2), and 90% cure time (t90). The dataset comprises 220 batches of the rubber compound, which are randomly divided into training (85%) and testing (15%) datasets. The performance of each method is evaluated based on the mean absolute percentage error (MAPE). Results demonstrate that while all methods yield reasonable predictions, the hybrid approaches consistently exhibit lower MAPE values compared to the standard ANN, indicating superior performance. Additionally, the hybrid methods are effective in mitigating extreme deviation points, providing more reliable and stable predictions. A comparison with existing literature reveals that the prediction errors in this study are substantially lower, highlighting the enhanced accuracy of the hybrid methods. Second study expands the analysis by incorporating Gaussian Process Regression (GPR) and Support Vector Regression (SVR) alongside ANN to predict rheometric properties. A larger dataset of 1128 batches is utilized, which is divided into training (70%), validation (15%), and test (15%) sets. The input variables include additional process conditions such as ram pressure, chamber water temperature, rotor water temperature, and drop door water temperature, in addition to the previously considered factors. The output variables for prediction are extended to include 30% and 60% cure times (t30 and t60), along with ML and MH. A detailed sensitivity analysis is conducted to identify the optimal hyperparameters for each technique. The results indicate that the best-performing algorithms achieve MAPE values of 2.12% for ML, 1.69% for MH, 2.61% for t30, and 2.71% for t60. These results demonstrate a significant improvement in prediction quality compared to existing literature. The performance of SVR and GPR is found to be closely comparable, suggesting their potential for interchangeable application in predicting rheometric properties. While ANN predictions are preferable for ML and t60, SVR and GPR exhibit superior performance for MH and t30 predictions, underscoring the necessity of these algorithms in rubber industry literature for accurate rheometric property estimation. Third study focuses on optimizing the composition of rubber compounds to reduce costs. Plackett-Burman and Box-Behnken experimental designs are utilized to identify the correlations between component amounts and technical specifications of the rubber compound. A Support Vector Regression integrated Genetic Algorithm is proposed to optimize the formulation and minimize costs. Key factors such as natural rubber, carbon black, white filler, stearic acid, zinc oxide, antiozonant, antioxidant, process oil, curing retarder, curing agent, and accelerator are initially screened using the Plackett-Burman design to determine the most significant factors. Subsequently, the Box-Behnken design is employed to examine four key factors (carbon black, process oil, curing agent, and accelerator) to reduce the number of tests while establishing accurate relationships between formulation and specifications. The optimization process, implemented in MATLAB R2022a, includes a sensitivity analysis to determine the best hyperparameters for the integrated Support Vector Regression and Genetic Algorithm approach. The base compound cost is reduced from 2.009 €/kg to 1.989 €/kg, resulting in a 1.4% decrease in overall production costs. The optimized rubber compound meets the required technical specifications while achieving a significant reduction in curing costs, demonstrating the effectiveness of the proposed method for cost optimization in rubber compounding. Overall, these studies highlight the substantial potential of machine learning and optimization techniques in the rubber industry. They demonstrate how advanced predictive and optimization models can significantly enhance product quality, streamline manufacturing processes, and reduce costs. The integration of hybrid and regression techniques offers promising results for accurate predictions and cost-efficient solutions, emphasizing the importance of data-driven decision-making and paving the way for future research in this field. Future research could further explore these methodologies by expanding and diversifying datasets, such as including different compound types, process conditions, and component combinations to improve model generalization. Additionally, incorporating seasonal variations, raw material fluctuations, and production process differences could enhance prediction reliability. Developing real-time prediction systems for dynamic process monitoring and optimization could offer significant benefits. The continuous updating of machine learning models with real-time data may facilitate more precise process control and rapid adaptation. Future studies could also investigate more complex interactions between factors and employ multi-level modeling techniques to provide more comprehensive and accurate optimization results. Including environmental and economic factors in the analysis could contribute to optimizing both costs and environmental impacts. Encouraging the development of user-friendly software and tools for machine learning and optimization results can enhance decision support systems and visualization methods for better data analysis by industry practitioners. In conclusion, the integration of machine learning and optimization techniques in the rubber industry has demonstrated considerable potential for improving product quality, optimizing production processes, and reducing costs, thereby contributing to the development of a more efficient, sustainable, and competitive rubber manufacturing sector.

Benzer Tezler

  1. Tez No

    389259

    qRT-PCR analysıs of gene expressıon level dıfferences between freeze-tolerant laboratory and ındustrıal straıns of Saccharomyces cerevisiae

    Donma-eri̇me stresi̇ne di̇rençli̇ Saccharomyces cerevisiae laboratuvar ve endüstri̇yel suşlarinin gen anlatim düzeyleri̇ndeki̇ farkliliklarinin qRT-PCR i̇le anali̇zi̇

    ELİF ENGİN

    Yüksek Lisans

    İngilizce

    İngilizce

    2015

    Biyoteknolojiİstanbul Teknik Üniversitesi

    Moleküler Biyoloji-Genetik ve Biyoteknoloji Ana Bilim Dalı (disiplinlerarası)

    PROF. DR. ZEYNEP PETEK ÇAKAR

  2. Tez No

    930824

    Paphlagonıa Hadrianopolisi Güney Nekropol tarihöncesi dönem çanak çömlekleri

    Prehistoric pottery assamblages from Southern Necropolis of Paphlagonia Hadrianopolis

    FERDA TATLI

    Yüksek Lisans

    Türkçe

    Türkçe

    2025

    ArkeolojiHitit Üniversitesi

    Arkeoloji Ana Bilim Dalı

    DR. ÖĞR. ÜYESİ EMRULLAH KALKAN

  3. Tez No

    945928

    Data-driven process mining for production line optimization using IIOT and big data technologies

    IIOT ve büyük veri teknolojileri kullanılarak veri odaklı süreç madenciliği ile üretim hattı uygulaması

    BEYZA YAPAKÇI

    Yüksek Lisans

    İngilizce

    İngilizce

    2025

    Endüstri ve Endüstri Mühendisliğiİstanbul Teknik Üniversitesi

    Veri Mühendisliği ve İş Analitiği Ana Bilim Dalı (disiplinlerarası)

    PROF. DR. ALP ÜSTÜNDAĞ

  4. Tez No

    650889

    Sıvı kauçukların kord bezi kaplama hamurunun ısıl, mekanik ve yapışma özelliklerine etkileri

    Effect of liquid rubbers on the thermal, mechanical and adhesion properties of the tire skim compound

    FURKAN ÇELTİK

    Yüksek Lisans

    Türkçe

    Türkçe

    2020

    Kimya MühendisliğiKocaeli Üniversitesi

    Polimer Bilim ve Teknolojisi Ana Bilim Dalı

    DOÇ. MUSTAFA ÖZGÜR BORA

  5. Tez No

    808616

    Elastomerlerin metal yüzeylere kimyasal yapışması için karışım geliştirilmesi

    Development of compositions for the chemical adhesion of elastomers to metal surfaces

    MURAT AYDIN

    Yüksek Lisans

    Türkçe

    Türkçe

    2023

    KimyaSakarya Üniversitesi

    Kimya Ana Bilim Dalı

    PROF. DR. SALİH ZEKİ YILDIZ

Geri Dön