Rekabetçi ve esnek üretim ortamında, ürünü müşteriye hızlı, istenilen kalitede ve oluşabilecek en düşük maliyette üreterek sevk edebilmek, dün de önemli olmakla birlikte günümüzde üreticiler için olmazsa olmaz bir hale gelmiştir. Bazı üretim süreçlerinde, üretimin çıktısının istenilen kalitede olup olmadığını önceden tahmin edebilmek ise, bu hedefe ulaşmada en etkin yollardan biridir. Günümüzde bu konuda genel yaklaşım uzman kişilerin bilgi ve tecrübelerinden yararlanmaya çalışmak olsa da, bu tarz yaklaşımlarla elde edilen başarılar, oldukça sınırlı kalmaktadır. Alüminyum, Dünya'da yükselen bir metaldir. Bu metalden elden edilen yassı rulo ürünlerin de yükselişi, alüminyumun cazibesinin artmasına paralel olarak devam etmektedir. Ortada olan yüksek pazar, yüksek rekabeti de beraberinde getirmektedir. Sürekli üretimin yapıldığı bu ürünlerin üretim prosesindeki en küçük iyileştirmeler bile, yüksek karlılık ve rekabet avantajı olarak firmalara geri dönmektedir. Bu çalışmada, Alüminyum Yassı Mamul üretiminin ilk aşaması olan Döküm Sıcak Haddeleme sürecine ilişkin, süreçteki girdi değişkenleri parametrelerine bağlı olarak nihai ürünün kalitesini (1. Kalite, 2. Kalite ya da 3. Kalite), nihai ürün sok kontrol aşamasına gelmeden, daha döküm aşamasında iken tespit edecek, böylelikle istenmeyen üretimlerin ve bunlara bağlı oluşan maliyetlerin önüne de geçerek üretim planlama esnekliğini artıracak, öğrenebilen akıllı bir yapı oluşturulmaya çalışılmıştır. Araştırma esnasında, Eğiticisiz Öğrenme metodu olan Kümeleme Yöntemleri ve Eğiticili Öğrenme, Sınıflandırma yöntemleri olan Makine Öğrenmesi ve Derin Öğrenme yöntemleri denenmiştir. Yöntemlerin uygulanmasından önce verinin sistemden alınması sonrasında veri üzerinde veri ön işleme süreçleri uygulanmıştır. Araştırma sonuçlarına göre, Kalite Tipi'nin eğiticisiz öğrenme teknikleri ile belirli bir seviyede ama düşük bir seviyede olmakla birlikte, eğiticili öğrenme teknikleri ile yüksek seviyede tahmin edilebileceği görülmüştür. Araştırma, mevcut durumda Eğiticili Öğrenme tekniklerinin yarı mamul bazında uygulanmasının en iyi sonuçları ortaya koyduğunu, bu durum içerisinde de farklı yarı mamul tiplerinde farklı algoritmaların en başarılı algoritmalar olabileceğini göstermiştir. Elde edilen sonuçlar ışığında yüksek derecede kazanımlar elde edilmiştir.Sonuçlar, fiili olarak kullanıma geçmiş olmakla birlikte, kullanılmakta olan yazılımlara entegrasyon süreci devam etmektedir.
In a competitive and flexible production environment, it is essential for today's producers to ship products to customers quickly, at the desired quality, and at the lowest possible cost. One of the most effective ways to achieve this goal in some production processes is to predict if the output of the production is of the desired quality. In the casting process, the raw flat aluminum coil produced undergoes different processes according to the variety of the final product and transforms into the end product. At this point, looking at the production process, it is generally observed that these raw coils are subjected to long-term casting, and the quality of these components can only be understood after the production process. If the resulting final product is not of the desired quality, it is observed that this semi-finished component, if not scrap, can be served as 2nd or 3rd quality. During the process, the ability to understand the quality of the semi-finished component that will emerge is extremely important. For example, if a coil expected to meet the final product with a 1st quality expectation is not of the desired quality during production, the entire planning will need to be changed. As a matter of fact, this situation will also cause the production of other semi-finished coils waiting for other customers to be delayed. This is a situation that creates bottlenecks in terms of planning. On the other hand, predicting the quality of the semi-finished coil that will emerge after casting will have many advantages. The most important of these is that if the coil that will emerge is scrap, urgent measures can be taken beforehand, and if a 2nd or 3rd quality semi-finished product will emerge, it can be quickly analyzed whether it can be substituted for other orders waiting in the plan, thus allowing these orders to be removed from the plan. In reality, expert personnel also try to make these predictions and form some forecasts. However, the complexity of the problem makes it very difficult for these predictions to be accurate. Today, using Artificial Intelligence and Data Mining techniques, high prediction success has been achieved with prediction algorithms in many different fields through data obtained from systems. Of course, for this, the available data must first be transformed into meaningful data through preprocessing steps using data mining methods, and then the most appropriate algorithms must be applied to the obtained data. In this study, considering many input variables and the quality of the output semi-finished rolls that are thought to affect the result tracked instantly during the casting process, it was tried to predict the quality level of the output roll based on the input variables. With the correct prediction, flexibility provided in the production planning of the semi-finished product increased efficiency in the production process, and some delays in the delivery times of the final product to the customer were prevented. In the first sections of the thesis, a Literature Review is presented. In the literature, it is possible to find many studies on predictions using Data Mining and Machine Learning methods. Within the scope of the literature review, the studies conducted were first examined in general on classification and prediction studies, then studies conducted specifically in the metal sector, and finally, studies on data mining and prediction algorithms on aluminum casting, which is the subject of our study, were examined. In the continuation of the thesis, the materials and methods used in the thesis are introduced. In the research, data from the casting process of a factory producing Flat Aluminum Coil was used. Within the data obtained, there are 63 different input variables. One year of data for these variables was taken from the system. The data obtained from the system were analyzed with personal computers and rented super powerful computers. The data obtained from the system were taken as Excel files and uploaded to the Jupyter Notebook library. Then, Clustering methods, which are among the Unsupervised Learning methods, were tried on the data. Clustering analysis is a group of multivariate techniques whose primary aim is to group objects (units) based on their characteristics. Clustering analysis groups objects in a very similar manner within the cluster and differently between clusters. If the clustering process is successful, when a geometric drawing is made, the objects will be very close to each other within the cluster and far from each other between clusters. At this point, Clarans, Birch, Spectral Clustering, and Hierarchical Clustering Methods were used. After testing the Unsupervised Learning algorithms, machine learning and deep learning algorithms, which are Supervised Learning methods, were tested. Here, twelve different algorithms were tested. In both cases where Unsupervised Learning algorithms were tested and where Supervised Learning algorithms were tested, the algorithms were first applied to the data set, considering the materials as a single data frame, then clustering and classification algorithms were applied to the data, and then the materials were taken one by one from the data set as individual data frames, and the clustering and classification algorithms were applied to the data belonging to the materials. To evaluate the performance of the methods, evaluation metrics from the literature were used. The silhouette score was used to evaluate the clustering algorithms. The Silhouette Score was developed to find the appropriateness of the clustered data within their respective clusters. Additionally, to evaluate the performance of both Supervised and Unsupervised Learning algorithms, the AUC Score and ROC Curve were used. This method is a good way to visualize the performance of a classifier to select an appropriate operating point or decision threshold. Another evaluation method used was the confusion matrix. The F1 Score, Recall, and Precision values calculated using the confusion matrix were used to compare the performance of the classification algorithms. In the world of machine learning, the runtime of the algorithm to be used should always be considered when selecting the most accurate algorithm for the data among the available algorithms. Therefore, when comparing algorithm performance, the runtime of the algorithms was also considered as a comparison criterion. The results of the study were implemented in the facility. Accordingly, the successful algorithms from the study were used during the planning phase based on input parameters through Jupyter Notebook. The success of the algorithms was tested by making necessary planning changes based on the findings obtained. At this point, very successful results were also achieved in the real-life application of these algorithms. The gains achieved include improvements in the planning preparation phase, a reduction in material changes, a decrease in the inventory of final products held, and a reduction in the amount of rework. At the end of the section, these gains and their contributions are presented to the reader. In conclusion, looking at the significant findings obtained from the thesis, the first noteworthy point is that this study has highly accelerated a process that was previously predicted based on individuals' experiences and took a considerable amount of time even then, making it possible to predict without human errors. In this context, the thesis is one of the few examples conducted specifically for the sector and problem. The thesis demonstrated that clustering algorithms applied to the entire data set performed poorly. However, the classification tested with machine learning and deep learning in the continuation of the thesis gave very positive results. In the initial evaluation conducted without separating semi-finished products as a single data set, the AUC value, accuracy value, and F1 score, which are the first evaluation criteria, ranged between 0.80-0.91 for the best algorithms. This is a very good success. Among the algorithms, the CATBOOST and GBM algorithms achieved these values, making them the most successful algorithms. However, in terms of values, the CATBOOST algorithm is slightly better than GBM. Therefore, it is stated that the CATBOOST algorithm is the most successful algorithm. Finally, when classification was applied with machine learning and deep learning algorithms on a semi-finished product basis, the obtained AUC, accuracy, and F1 score values ranged between 0.90-1.00 for most types of semi-finished products. At this point, it has been shown that the most successful solution for the problem in the study is the classification with machine learning and deep learning on a semi-finished product basis. The most successful algorithm at this point has been identified as the RNN algorithm. In light of these results, a software structure has been created where, in reality, a program that makes predictions by running the most successful algorithm from our study for the semi-finished product to be included in the production plan is used. This structure is currently in use and has yielded very positive results in real life. These results have reflected in reduced costs and increased profitability.
Except where otherwise noted, this item's license is described as http://creativecommons.org/licenses/by/4.0/
