İş parçası kodlama ve sınıflandırma sistemlerinde yapay sinir ağı yaklaşımı

Abstract

Grup teknolojisi üretim sistemlerinde, kodlamave sınıflandırma çok özel bir yere sahiptir. Parça ailelerinin oluşturulması ve benzer parçalara, benzer kodların atanması pek çok faydalar sağlamaktadır. Ancak, bu tür işler tecrübeli elemanlar tarafından yapılsa dahi fazla zaman olmakta ve tek elden yürütülmediği takdirde hatalara sebep verebilmektedir. Bu çalışmada amaç; yapay sinir ağlan sistemi kullanarak daha kısa sürede ve hatasız olarak bu işlemleri gerçekleştirebilecek bir yapıyı ortaya koyabilmek veya en azından ileri çalışmalara temel hazırlayabilmektedir. Çalışmada, yapay sinir ağlan üzerinde kullanılan Backprogation metodundan faydalanılarak, OPTTZ kodlama sistemine çözüm aranmıştır.

Classification and coding system arc widely used in Group Technology Manufacturing System. There are many methods for classification and coding (CC). Classic methods for CC require lots of time, what is more, these workings are boring, tiresome and costly. Since techniques have developed, these kinds of problems should be developed. To achieve a code fora component in a factory, requires 3 minutes at least if a qualified code-maker would do it. These kind of time consuming may manage to firms less competitive. Neural Network approaches are widely used in computer and electrical researches, and the approaches are used in Industrial Engineering with the therm of 'Artificial Intelli eğence'. In an Artificial Neural Network approach, there is a matter of 'Learning". Assume that the network has learned, you could easily take a decision for solving your problem. In the Artificial Neural Network approach, first you should have input values, second; random weight values which are the variables Head to learn, third; target values. The input and the target values are constant. There are many neurons in the Artificial Neural Network Approaches. The neurons operate mathematical computations. VI An activation function is used in the computations which is the multiplication of the input vectors by their weights. Activation function is ex|x>nanu'al. It's range is limited between 0 and 1. The neurons are used for linking each neuron in the system is connected the other, and, informations can easily be processed through the Artificial Neural Network. There are three layers in a Multilayer Neural Network; input layer, output layer and hidden layer between the input and output layers. Informations flow through the Artificial Neural Network from the* input layer to output layer.. Neurons in the input layer are e-pnnected other neurons in the hidden layer. Input values becorhe output values of the hidden layer by the computation efforts and the outputs are the input values of the output layer which propagates output values of the output layer. The important thing is adjusting the weights for learning. The adjusting can be made step by step. There are many methods for adjusting the weights. Backpropagation method (BP) Is widety used in Multilayer Artificial Neural Network Approaches. The method requires the values which are just mentioned above, including target values. The activation function is used in each neuron in BP to produce output values. The output values can be computed easily but the results are generally not satisfactory for learning so, weights should\ be adjusted. The adjusting procedure requires the target values to produce error values by substracting the output value» of the output layer ftom the target values. Propagating the error value gives the name of BP because, the error values are served backthrough the Artificial Neural Network. The error \ values change the weights and they are polarize. The weights are adjusted step by step by using the polarize values. As a srtarting rule, first; random weights should be very small. After "\ many iterations network can learn but, sometimes network nfay not team. Because, problems for Multilayer Neural Network are non-linear and, network may find a local minima. MWroteation of the error values is objective function. VII In this study, Multilayer Neural Network requires the input, weight, output, and, target values with formulations below; NETQ=Pi*Wij P = the neurons in the input layer, input values. Q = the neurons in the hidden layer. OUTQ=1 / ( 1 + e-MET ) OUTQ = output values for the neurons in the hidden layer. NETQ = NET values for the neurons in the hidden layer. So, in the hidden layer, each neuron has NET and OUT values. In the second part of computation the formulations are changed by modifying indices between the hidden layer and the output layer. Modified formulas are below: NETR = OUTj*Wjk OUTR = 1/(1+ e-NET) R = the number of the neurons in the output layer. The computation can be made by setting the input values. In this study, it is assumed that the input values are like a map of a component. For example; VIII f 0000000000 1111111111 100000000 1 1111111111 0000000000 Fig. 1. A bitmap of the component. Figure 1 shows a tube shapped component which is a pipe, ft has a. length of 10 units and a diameter of 3 units. How is the map turned into a vector form? Consider that every element of the matrix is a vector, will have 10*5=50 vectors for the input layer. At this point, the Artificial Neural Network has 50 input vectors in the input layer. As the number of the input neurons are defined, the number of neurons in the hidden and output layers should be determined. The number of the neurons in the output layer is 10, because, there are 10 characters in each digit in OPITZ. AeMcsd number of the neurons in the hidden layer is the avarage number of the neurons in the input layer and output layer. However changing the number of the neurons in the hidden layer may cause a better learning. As it is mentioned above, the number of the neurons in the output layer is 10. The Artificial Neural Network will work for only one column in OPITZ. So, what is the meaning of 10 neurons in the output layer for making a decision for one column in OPITZ. Consider another coding which is not directly connected with OPITZ. 1 000000000 = 0 01 00000000 = 1 001 0000000 -2 ' IX 0000000000 = 9 There are 10 output values and 10 target values in training. If the target value for the first digit in OPITZ is 2, the output value should be 2, so, the Artificial Neural Network will try to compute 2. In order to define the value of 2 for the computer, 0010000000 vector form is selected. By means of this, the binary system is used. This provides the computational efficiency and, understanding of the computer logic. Finally, for one set of the input and target values, the programme is runned and results are determined, so, the weights are adjusted for the first set. For each set the programme requires (relatively) 1000-15000 iterations with a learning coefficient 0.01-0.1. The Artificial Neural Network will learn all exam pier input vectors seperately, in this study the learning is made and test should be made separately too. In this study, processing time with computer is approximately 2 minutes for learning. After learning, the testing phase will take only seconds. Recognizing time is short enough and the computer programme is efficient. However, the programme make RAM too busy. For example, only one file for learning has 4Kb. So the programme is not efficient for large cases because of the most computers have 4 Mb RAM.