Kauçuk, keşfedildiği 15. yüzyıldan bu yana endüstride ve günlük yaşantıda farklı alanlarda sıklıkla kullanılan bir malzeme olmuştur. Elastik yapısı, aşınmaya karşı direnci, su geçirmez özelliği sayesinde kauçuk, modern yaşamda farklı sektörler için hayati bir öneme sahiptir Lastik üretiminde, kayış ve conta üretiminde, konveyör bant üretiminde yoğun olarak kullanılan kauçuk, farklı dolgu malzemeleri, kimyasallar ve proses kolaylaştırıcılar ile bileşik haline getirilip, işlenerek hizmete sunulur. Kullanılacağı alanın gereksinimlerini karşılaması önemli olduğu için, kauçuk bileşik üretiminin başlangıcından sonuna kadar bir dizi gereksinim testleri yapılır. Yapılan bu testler ile kauçuk bileşiklerinin reolojik ve mekanik özellikleri belirlenir. Kauçuk bileşiklerinin mekanik özellikleri, zaman alıcı ve pahalı laboratuvar çalışmaları ile belirlenir. Öte yandan günümüzde oldukça popüler olan yapay zekâ yaklaşımları, herhangi bir numuneye veya laboratuvar deneyine ihtiyaç duymadan, kauçukların mekanik özelliklerini saniyeler içerisinde tahmin etmek için kullanılabilir. Bu çalışmada, kauçuğun mekanik özelliklerinden olan sertliği, çekme mukavemetini, kopma uzamayı ve kauçuk bileşiklerinin yırtılma mukavemeti test değerlerini tahmin etmek için yapay sinir ağları, regresyon ağaçları ve topluluk öğrenimi teknikleri kullanılmıştır. Kauçuk bileşiklerinin, proses parametreleri ve hammadde bileşimleri, bu çalışmada seçilen yapay zekâ tekniklerinde girdi olarak kullanılmıştır. Veri seti 76 farklı kauçuk bileşiği formülasyonundan oluşmaktadır. 76 farklı kauçuk bileşiği formülasyonu ve test değerleri üzerinde normallik testi yapılmıştır. Daha sonrasında veri setinde girdi olarak yer alan 29 hammadde kg ağırlıkları ve dk cinsinden proses süreleri ele alınarak test değerleri üzerindeki korelasyon ilişkisine bakılmıştır. Sertlik, çekme mukavemeti ve yırtılma mukavemeti test değerleri üzerinde anlamlı değişikliğe yol açan hammaddeler belirlenirken, kopma uzama test değerleri üzerinde hiçbir hammadde ve proses süresinin anlamlı bir değişikliğe yol açmadığı görülmüştür. Mekanik özelliklerin test değerleri üzerinde anlamlı olarak belirlenen hammaddeler her bir mekanik özellik için seçilerek veri seti her mekanik özellik için ayrı olarak oluşturulmuştur. Oluşturulan veri setleri, yapay sinir ağları, regresyon ağaçları ve topluluk öğrenimi yöntemlerinden olan ağaç topluluklarında girdi ve çıktı verileri olarak kullanılmıştır. Her bir yapay zeka tekniğinde oluşturulan modellerin performansını karşılaştırmak için ortalama hata karesi (MSE), ortalama mutlak yüzde hatası (MAPE) ve çoklu belirlilik katsayısı (R2) değerleri hesaplanmıştır. Hesaplanan bu değerler sonucunda ağaç toplulukları modellerinin sertllik, çekme mukavemeti ve yırtılma mukavemeti test değerlerini tüm veri seti için R2 = 0,99 olarak tahmin ettiği görülmüştür ve değerin 1'e yakın olması tahminin güçlü olduğunu göstermiştir. Regresyon ağaçları ise sertlik için R2 = 0.96, çekme mukavemeti için R2 = 0.98 ve yırtılma mukavemeti için R2 = 0.97 olarak hesaplanmıştır. Yapay sinir ağları ise belirlenen girdi ve çıktı değerleri ile başarılı olarak kabul edilebilecek bir R2 değeri bizlere sunmamaktadır. Çalışmanın sonucunda kauçuk bileşiklerin mekanik test özelliklerinin tahmininde ağaç toplulukları ve regresyon ağaçlarının kullanılabileceği görülmüştür.
Since its discovery in the 15th century, rubber has been a material that is frequently used in different fields in industry and daily life. Thanks to its elastic structure, resistance to abrasion and waterproof feature, rubber has a vital importance for different sectors in modern life. processed and put into service. Since it is important that it meets the requirements of the area in which it will be used, a series of requirements tests are performed from the beginning to the end of rubber compound production. With these tests, the rheological and mechanical properties of rubber compounds are determined. The mechanical properties of rubber compounds are determined by time-consuming and expensive laboratory work. On the other hand, artificial intelligence approaches, which are very popular today, can be used to predict the mechanical properties of rubbers in seconds, without the need for any samples or laboratory experiments. In this study, the test values of the mechanical properties of rubber and the parameters used to control the quality and process standards of rubber products produced in the industry, the hardness, tensile strength, elongation and tear strength of rubber compounds, which are measured with test devices suitable for use in the laboratory and expected to take place in the predetermined test intervals, are determined. Artificial intelligence techniques such as artificial neural networks, regression trees and ensemble learning, which are very popular today, have been used to predict predictions. The process time parameters in minutes, including the mixing times of the rubber compounds of the rubber products that form a part of the tire industry such as motorcycle tires, bicycle tires, agricultural agricultural vehicles tires in the industry, called Banbury, and the weights in kilograms of the raw materials used in the production of these rubber compounds, the artificial Artificial neural networks, regression trees and ensemble learning techniques, which are among the intelligence techniques, are used as algorithm input dataset. The dataset consists of 76 different rubber compound formulations used in tire production. After the dataset was prepared as input, a normality test was performed on the compounds and test values in 76 different rubber compound formulations in order to create correct analyzes in correlation analysis. Then, by considering the 29 raw material kg weights and the process times in minutes as input in the data set, the correlation relationship of the mechanical properties of each of the raw material and process parameters on the test values was examined in order to determine the raw materials and process times to be selected for use in artificial intelligence techniques. As a result of the correlation analysis, while the raw materials that cause significant changes on the hardness, tensile strength and tear strength test values were determined, it was seen that the process parameters did not cause a significant change on these test values. It was observed that no raw material and process time did not cause a significant change on the Tensile elongation test values, which is another selected mechanical property. The raw materials that were determined to cause a significant change on the test values of the mechanical properties were handled and used to create the data set for hardness, tensile strength and tear strength. In this way, the data set, which is provided as input to the artificial intelligence techniques to be predicted, has been prepared with data that creates a meaningful change on the test values. For one of the mechanical properties, elongation, this mechanical property was not considered in the estimation, as it was seen that it did not cause a significant change on any raw material and process parameter. The datasets created for stiffness, tensile strength and tear strength were first used for prediction in artificial neural networks. The algorithm parameters to be used in each mechanical feature of the artificial neural networks to be used in the study were determined. First, the learning algorithm and transfer function of the artificial neural networks to be used in the prediction of mechanical properties are determined. Levenberg-Marquart, Bayesian Networks and Scaled Conjugate Gradient learning algorithms and tan-sigmoid, log-sigmoid and radial basis functions, which are transfer functions, were tested in artificial neural networks created with empirically determined hidden layer and 5 neurons. According to the estimation results, learning algorithm and transfer function were determined separately for hardness, tensile strength and tear strength. While making these determinations, each artificial neural network was run 50 times and the mean squared error (MSE) and mean absolute percentage error (MAPE) values of the data used as test data in the estimation were examined. By comparing these values, the learning algorithm and transfer function were determined for each mechanical property to be estimated. After the learning algorithm and transfer functions are determined, in order to determine the number of hidden layers that will provide the best prediction performance in the artificial neural networks and the number of neurons in each layer, hypothetically 1-layer artificial neural networks with 5, 10, 20 neurons, respectively, and 2-layer and artificial neural networks have been created in a way that each layer will have 5, 10 and 20 neurons, respectively. Each artificial neural network created was run 50 times. MSE and MAPE values are considered to compare the prediction performances of the operated networks. The best performing hidden layer and neuron numbers from these values were determined for each mechanical property. With the learning algorithm, transfer function, number of hidden layers and number of neurons determined for each mechanical property, artificial neural networks were created for mechanical properties and predictions were made by running 50 times. For the regression trees technique, data sets prepared for each mechanical property were used as inputs. In order to test the number of leaves that will achieve the best prediction performance for each mechanical property, regression tree models with 4 leaves, 12 leaves and 36 leaves were created, respectively, and MSE, MAPE and multiple coefficient of determination (R2) values were obtained by running 50 times. The tree ensemble method, which is one of the ensemble learning methods, is very similar to the regression trees. For each mechanical feature, models with 4, 12, and 36 leaves were created, as in the number of leaves regression trees, respectively, which will provide the best performance for each mechanical feature. Estimation was made by running the models created with the number of leaves 50 times and MSE, MAPE and R2 values were checked to measure the results of the prediction performance. MSE, MAPE and R2 values are discussed to compare the performance of the models created in each artificial intelligence technique. As a result of these calculated values, it was seen that the tree assemblage models estimated the hardness, tensile strength and tear strength test values as R2 = 0.99 for the whole data set, and the value close to 1 showed that the estimation was strong. Regression trees were calculated as R2 = 0.96 for hardness, R2 = 0.98 for tensile strength and R2 = 0.97 for tear strength. Artificial neural networks, on the other hand, do not offer us an R2 value that can be considered successful with the determined input and output values. As a result of the study, it was seen that tree assemblages and regression trees could be used to predict the mechanical test properties of rubber compounds.
Except where otherwise noted, this item's license is described as http://creativecommons.org/licenses/by/4.0/
