Prediction of defective product with machine learning algorithms

In recent times, data collecting and storing has become easy with the current developments in technology. Many data are recorded and stored voluntarily or involuntarily on a day-to-day basis in our lives both in an individual sense and in an institutional sense. Although the collection of each data seems to be an advantage in recording, it is becoming increasingly difficult to extract meaningful and useful information from such large data. This situation has made the big data debatable. The data that can be analyzed and understood for humans have become meaningful and information. For this reason, special methods such as artificial intelligence are needed to process and analyze the data. The concept of artificial intelligence, which is inspired by the intelligence of human beings and aims to make the same system and fiction possible in robots and machines; has led to the concept of machine learning for purposes such as prediction, classification, clustering. The desire of machines and computers to learn like human beings and to be able to make the best decision is the main aim of machine learning concept. It is necessary to educate and teach computers like human beings to a certain level only by means of machine learning. With the help of machine learning, more accurate and valid results can be obtained by forward prediction and planning based on past data. In this study, the use of machine learning algorithms in predicting defective products is examined, and the performance of Random Forest, Naive Bayes, Support Vector Machines and k-Nearest Neighbor algorithms is evaluated through an application. In the study, firstly the concept of machine learning, machine learning methods and processes are mentioned. Then, machine learning algorithms are examined in general and working methods, performances and advantages of these algorithms are mentioned. There are different methods to use in machine learning. With the development of information and communication, these methods are developing and increasing day by day. Four of them are prominent in the machine learning method. These are supervised, unsupervised, semi supervised, and reinforcement learning.Supervised learning is a method which obtain the interaction and influence of inputs on outputs under the surveillance of supervisor . The main element in supervised learning is the existence of a training set that was previously made up of observations, and this training set is taught and introduced to the system by a supervisor. Unsupervised learning is a learning method in which there is no supervisor, only observations with input variables. It is a learning method that collects and groups the samples having close features and variables among the observations under the same set, which enables the system to find the relation between the input variables by self learning without the supervisor. The learning method applied to a data stack consisting of observations with only input variables in a very large quantity and observations with input and output variables in a small amount is called semi-supervised learning. Semi-supervised learning is used when it is difficult and costly to obtain observations with a target outcome. The method that learns how to reach the target through trial and error system in accordance with the target given to the system is learned to reinforce. The next step in the system with the award-penalty given at each step is a learning system that tries to reach the end by making use of the inferences from the previous step. Some processes should be followed in the face of problems to be solved by using machine learning. These are respectively as follows; definition of the problem, data analysis, preparation of the data, establishment of the model, evaluation of the model and use of the model. The first step in the problems that has solved by machine learning is defining the problem in the best way and clarifying what the purpose of the solution is . In this step, the success criterion for the purpose of the problem and what is in the present situation needs to be well defined. In the data analysis part, the core reason is to obtain the appropriate data. In addition, it should be checked whether the main feature that will be used in the model is met by that dataset needed. After the necessary data is obtained , the data can be used by machine learning algorithms. This process have called data preparing. The most appropriate model in order to solve the link between the variables and determine the most suitable algortihms for the model in the data set is the phase of establishing the model. At this stage, more than one suitable model is determined and applied to the data set in order to get the best result. Different models and algortihms is compared during the evaluation process of the model. In a model evaluation stage , it is determined that the results of the model will be beneficial or not in the future. In the process of using the model , the core reason is to benefit from definig model in an a suitable way in similar situations and researches with the problem . The implementation phase of the thesis, washing machine powder coating unit input and output measurements is provided by one of the leading companies in the durable good sector in Turkey Firstly, it is determined whether or not the laundry machine parts are in the range of the desired paint thickness value after leaving the powder coating unit, and the target variable is set in this direction for use in the algorithm. Later on, the number of observations was enriched and included in the new variable data set, which is thought to be more effective. And 327 line data set is used to evaluate in the algorithms. Machine learning algorithms have been evaluated in accordance with input variables, business objectives and data. In this direction, Random Forest, Naive Bayes, Support Vector Machines and k-Nearest Neighbor algorithms have begun to be applied in the data set. 70% of the data set was used as the training set and 30% as the test set. In the data set, RapidMiner named software was used to run these algorithms. The algorithms determined by RapidMiner are run on the data for classification result for each algorithm. Classification successes were measured and compared with each other using 4 different algorithms in machine learning. it is desirable to measure the classification performance of each algorithm in the wrong product estimation by using the accuracy values of the algorithms in comparison. Then feature engineering studies were performed on the data with data normalization and data scaling. Same machine learning algorithms are applies and their performances were compared on these data. Finally, the general validity of the results of the algorithm that has the highest performance is tested and interpreted. Random Forest algorithm among the Random Forest, Naive Bayes, Support Vector Machines and k-Nearest Neighbor algorithms in the washing machine's defective product prediction has shown the best performance. Another successful defective product classifier algorithm is the Naive Bayes algorithm. In addition, the most important input variables that will help the company to focus on the next improvements and work are also revealed in this analysis. The Random Forest algorithm has been identified as a high performance classification algorithm for defective products, which will help reduce costs for companies. Suggestions and recommendations have been made for the findings obtained and the results of the subsequent studies.