İmalat sistemlerinin tasarlanması ve öncelik kurallarının belirlenmesinde yapay sinir ağlarının kullanılması

Tez (Doktora) -- İstanbul Teknik Üniversitesi, Fen Bilimleri Enstitüsü, 1997Thesis (Ph.D.) -- İstanbul Technical University, Institute of Science and Technology, 1997Günümüzde yapay zeka kapsam itibarıyla disiplinlerarası bir hal almış ve her alanda başarıyla kullanılan yeni teknikler ortaya koymuştur. Bununla beraber bilginin organizasyonu sunumu ve elde edilmesi kullanılan tüm bilimsel teknikler için ayrı bir önem arz etmeye başlamıştır. Tüm bu gelişmelere bağlı olarak yapay zeka ve özellikle makina öğrenmesi tekniklerini bir bütün olarak kullanan zeki imalat sistemleri ortaya çıkmıştır. Zeki imalat sistemleri genellikle birkaç tekniği birarada kullanabilen oldukça başarılı karma (hybrid) sistemlerdir. Yapay sinir ağlan biyolojiden esinlenerek düşünülmüştür. Yapay sinir ağlarının fonksiyonu insan idrakini anımsatır. Burada sinapslar sinyalleri alma noktasında diğer nöronlara kadar ulaşır. Burada toplanan bazı girdiler hücreyi etkileme eğilimi gösterirler. Yapay nöron biyolojik nöronun ilk emir karekteristiğini taklit etmek için tasarlanmıştır. Bu sebeple ağ tasanmcılan mevcut biyolojik bilgiden daha ileri gitmeye yararlı fonksiyonlar oluşturmaya yardımcı olacak kavramlar aramaya mecbur kalmışlardır. Yapay sinir ağlan, doğrusal olmayan hatta doğrusal olmama özelliği oldukça yüksek dolayısıyla klasik metodlarla modellenmesi çok zor ve modellenmesindeki basan beklentiside bu oranda az olan sistemlerde uygulanmakta ve çok mükemmel modellemeler ortaya çıkabilmektedir. Bunlar arasında en yaygın olarak kullanılan metod geriye yayılım algoritmasıdır. Çalışma prensibi olarak hataların geriye yayılmasını ve öğrenme işlemi bittiğinde elde edilen ağırlıkların problemi çözmede kullanılmasını esas alır. Yapay sinir ağlarının en önemli özelliği öğrenebilme yetenekleri ve adapte olabilme özellikleridir. Yapay sinir ağlarının eğitimi yani bir modeli uygun şekilde kodlama ve YSA'yı çalıştırma ve çalıştırma sonucu YSA'nın öğrenmesini sağlama önemli bir ayrıntıdır. Aynca öğrenip öğrenmediğini de test etmek ve sonuca varabilmek mihenk noktalarından birisidir. îmalat sistemleri tasarlanırken, gerçek sistemler üzerinde denemeler yapmak çok zaman alıcı ve maliyetli olduğundan simülasyon metodu tercih edilir çünkü modeli kurulan sistem üzerinde değişiklik yapmak, sistemin parametrelerini değiştirmek çok kolay, maliyeti daha az ve zaman açısmda da oldukça avantajlıdır. Fakat buna rağmen herhangi bir sistem arzu edilen performans ölçülerine göre tasarlanırken, istenilen değerlere ulaşmak için çok sayıda deneme yapılması gereklidir. Bu ise yine zaman kaybma yol açmaktadır. Burada daha mükemmel bir sistem tasarlamak amacıyla simülasyon ve yapay sinir ağı birlikte çalıştırılmıştır. Bu çalışmada zaman kaybı olmaksızın istenilen performans ölçülerine sahip sistemi ve sistemde kullanılacak olan öncelik kuralım belirlemek için tasarım metodunu öğrenmiş yapay sinir ağlan kullanılmıştır. SPT, EDD, CR ve FCFS öncelik kurallarına göre ayn ayn tasarımlar öneren yapay sinir ağlarının önerileri değerlendirilmeye alınmakta, en uygun tasarım ve öncelik kuralı tercih edilmektedir.There has been an increase in the use of simulation and artificial intelligence techniques to design manufacturing system. Because, increased competition in many industries has resulted in a greater emphasis on automation to improve productivity and quality and also to reduce costs. Since automated systems are more complex, they can typically be analyzed in only by simulation. Computing costs have been reduced by microcomputers and engineering work stations. Perhaps the greatest overall benefit of using simulation in a manufacturing environment is that it allows a manager or engineer to obtain a system-wide view of effect of local changes to manufacturing system. If a change is made at particular work station, its impact on the performance of this station may be predictable. On the other hand, it may be difficult, if not impossible, to determine ahead of time the impact of this change on the performance of overall system. The simulation process is one of the most widely used modeling techniques in tiie system and modeling. However model might be run several times to desired performance level by trial and error method. Use of neural network is a promising tool for the design of manufacturing system. Neural network reduce the number of trial run required in simulation modeling. This study is composed of eleven chapters. Chapter 1 presents an overview of artificial intelligence and its growing. Chapter 2 presents an overview of knowledge concept. Chapter 3 discusses machine learning concept and examines the elements and various classifications of learning. Chapter 4 discusses why we use artificial neural networks. Chapter 5 presents fundamentals of artificial neural networks and its biological prototype. Chapter 6 presents knowledge representation in neural networks. Chapter 7 presents perception learning. Chapter 8 presents backpropagation algorithm and its technical structure. Chapter 9 discusses simulation of manufacturing system and design. Chapter 10 presents intelligent manufacturing system. Chapter 11 provides detailed analysis and methodology to design of manufacturing system using artificial neural networks. In this study there are two major aims. First is to determine the working mecanism as well as the principles of Backpropagation algorithm and neural networks. Second is to design of manufacturing system and dispatching rules in a manufacturing system by using artificial neural network. xv CHAPTER 1 AI is a big business. After a lean period in the academic wilderness, it has suddenly become respectable. Head-hunters in expensively tailored suits prowl the corridors of prestigious universities with tempting offers designed to entice its leading practitioners into industry. AI has replaced biotechnology as the venture capitalists' flavour of the moment. This aspect of AI is simply an attempt to make computer software a little less dumb. The key notation is intelligent problem solving and the key to intelligent problem solving, as opposed to brute-force approach, is to apply the same kind of techniques that human use. The next major step forward was the idea of heuristic search. AI workers abandoned the attempt to build artificial brains from the ground up. Instead they looked on human thinking as a complex coordination of essentially simple symbol- manipulating tasks. Here they were on firmer ground since computers can do thinks like searching, comparing sysmbols and so on, which they identified as the foundation of intelligent problem solving. CHAPTER 2 Knowledge analysis consists of the determination of the elements of a knowledge and knowledge characterization. Knowledge decomposition, is breaking a knowledge into its primitive parts. It requires lexical and syntactic analysis abilities to detect primitive elements; and semantic analysis abilities to detect meaning attached to it. Knowledge characterization is pragmatic analysis of knowledge and consist of identification, of the characteristics of knowledge and its constituents. Selection of knowledge can be done in absence or the presence of conflict. Decision is selection in the absence of conflict. In multicriteria decision making, several criteria are used simultaneously to make the selection. There are two categories of knowledge evaluation: intrinsic and relative. Intrinsic evaluation of an entity is its assessment with respect to some criteria. Depending of nature of criteria used, there are two categories of intrinsic evaluation, descriptive and normative. Descriptive evaluation is done with respect to some descriptive criteria and includes syntactic. Normative evaluation is donş with respect to normative criteria and can be pragmatic or ethical evaluation. Relative evaluation of an entity is its assessment with respect to some other entity. Organization is essential to reduce and manage complexity. There are two categories of knowledge organization issues depending on whether knowledge exists in one or several files. In the first case, knowledge is either put in a predefined sequence or is arranged in groups. CHAPTER 3 Learning can be defined in general terms as the processes in which intelligent systems increase their knowledge and improve their skills. Learning processes include the acquisition of new commonsense, descriptive, definitive, technical and theoretical knowledge, the development of skills, organization of new knowledge into general, xvi effective representations, and the discovery of new facts and theories through observation and experimentation. There is also a connection between learning and understanding, which includes the whole natural language problem. Most of what we learn we get from other people, communicated to us in natural language. The study and computer modeling of learning processes constitutes the subject matter of machine learning. It has long been a goal of artificial intelligence to develop computer systems that could be taught rather than programmed. Computational modeling of learning processes can help in the understanding of human learning. Gaining insights into the principles underlying human learning abilities is also likely to lead to more effective educational techniques. One basic scientific objective of machine learning is the exploration of alternative learning mechanisms, the scope and limitations of certain methods, the information which must be available to the learner, the issue of coping with imperfect training data, and the discovery of general techniques applicable in many task domains. By construction and testing learning systems, one can determine the effectiveness and limitations of particular approaches to learning. Another aim of machine learning is to build systems which will function as intelligent research assistants or intelligent personal assistants. According to the understanding in the first stage, learning systems start with little or no initial structure or knowledge. The major thrust based on tabula rasa approach involved building neural nets or self organizing machines, with random or partially random initial structure. Learning in these systems consisted of incremental changes in the probabilities that threshold logic units would transmit a signal. The hope was that if a system were given a set of stimuli, a source of feedback and enough degrees of freedom to modify its own organization, it would adapt itself toward an optimum organization. Machine learning can be classified in several different ways regarding the underlying strategies, representation of knowledge, application domain, and according to the levels at which the outcomes and the methods of learning are describable. Machine learning had been classified and reviewed by various researchers from different perspectives. CHAPTER 4 Artificial neural networks are biologically inspired; that is, they are composed of elements that perform in a manner that is analogous to the most elementary functions of the biological neuron. These elements are then organized in a way that may or may not be related to the anatomy of the brain. Despite this superficial resemblance, artificial neural networks exhibit a surprising number of the brain's characteristics. For example, they learn from experience, generalize from previous examples to new ones, and abstract essential characteristics from inputs containing irrelevant data. Despite these functional similarities, not even the most optimistic advocate will suggest that artificial neural networks will soon duplicate the functions of the human brain. The actual intelligence exhibited by the most sophisticated artificial neural networks is below the level of a tapeworm; enthusiasm must be tempered by current reality. It is, however, equally incorrect to ignore the surprisingly brain like performance of certain artificial neural networks. These abilities, however xvn limited they are today, hint that a deep understanding of human intelligence may He close at hand, and along with it a host of revolutionary. The strategy has been to develop simplified mathematical models of brain-like systems and then to study these models to understand how various computational problem can be solved by such devices. The work has attracted scientists from a number of disciplines, neuroscientists who are interested in making models of the neural circuitry found in specific areas. CHAPTER 5 Artificial neural networks are biologically inspired; that is, researchers are usually thinking about the organization of the brain when considering network configurations and algorithms. At this point the correspondence may end. Knowledge about the brain overall operation is so limited that there is little to guide those who would emulate it. Hence network designers must go beyond current biological knowledge, seeking structure that perform useful functions. In many cases, this necessary shift discards biological plausibility; the brain becomes a metaphor; networks are produced that are organically infeasible or require a highly improbable set of assumptions about brain anatomy and functioning. Despite this tenuous, often nonexistent relationship with biology, artificial neural networks continue to evoke comparisons with the brain. Despite the preceding caveats, it remains profitable to understand something of the mammalian nervous system. It is an entity that successfully performs the tasks to which our artificial systems only aspire. Each neuron shares many characteristics with the other cells in the body, but has unique capabilities to receive, process and transmit electrochemical signals over the neural pathways that comprise the brain's communication system. The artificial neuron was designed to mimic the first-order characteristics of biological neuron. In essence, a set of inputs are applied, each representing the output of another neuron. Each input is multiplied by a corresponding weight, analogous to a synaptic strength, and all of the weighted inputs are then summed to determine the activation level of the neuron. CHAPTER 6 The process of determininthe features used by a trained neural network and the knowledge embedded within it has elements of inference, analysis and guesswork. It is not nearly as simple as printing out the list of rules and facts known to an expert system. Although there is no doubt that knowledge is incorporated in the trained network, extracting it for comparison and validation can be tricky and time consuming. Just as neurobiologists must infer the operation of the brain, neural network researcher who care about how a network solves a problem must also be prepared to act as knowledge. Consider how information is stored in an ordinary computer program. There are two kinds of knowledge maintained by a program: the list of instructions that comprise the program itself and the values of variables used by the program. Finding out exactly what the computer knows at any given time is accomplished by printing xvui out the current instruction list and the values of all variables. The provides a complate and accurate accounting of all the knowledge contained in the program, and it is an intelligible format as well. CHAPTER 7 The proof of perceptron learning theorem demonstrated that a perceptron could learn anything it could represent. It is important to distinguish between representation and learning. Representation refers to the ability of a perceptron to simulate a specified function. Learning requires the existence of a systematic procedure for adjusting the network weights to produce that function. Perceptron is the earliest artificial neural network paradigm. The architecture of perceptron is simple. It uses the feedforward scheme, usually two layers, with supervised learning. Processing unit computes the net weighted sum of input signals and then the threshold value. If the computed value is greater then the threshold value the output is 1; otherwise is -1. The major characteristic of perceptron is the linear separation due to its threshold function. Single layer perceptron is seriously limited in its representational ability; there are many simple machines that the perceptron cannot represent no matter how weights are adjusted. A perceptron is trained by presenting a set of patterns to its input, one at a time, and adjusting the weights until the desired output occurs for each of them. CHAPTER 8 The invention of the backpropagation algorithm has played a large part in the resurgence of interest in the artificial neural networks. Backpropagation is a systematic method for training multilayer artificial neural networks. It has a mathematical foundation that is strong if not highly practical. Despite its limitations, backpropagation has expanded the range of problems to which artificial neural networks can be applied, and it has generated many successful demonstrations of its power. The form of the reinforcement signal mentioned there, it appears as though the rule functions by making corrections for errors, the corrections being determined by the teacher input. In fact, the rule is typically applied to the case in which pairs of patterns, consisting of an input pattern and a target output pattern, are to be associated. Imagine a situation in which the set of input/output pairs are repeatedly presented. Then the change in weight Wji following pattern pis given by the product of the ith input element and the jth target element. Gradient descent is sure to find the single minimum error set of weights. With hidden units, derivative computation is not obvious and there is the danger of getting struck at a local minimum on the complicated error surface. Rumelhart shows that there does indeed exist a way of finding the elusive derivatives and that the problem of local minima is irrelevant in a wide variety of learning tasks. We shall have more to say about the necessity of relying on a XIX methodology that will fail in the worst case in the section dealing with the intractability of the network learning problem. For the purposes of studying simple learning by backpropagation, consider a layered feedforward network with a semilinear activation function. A layered feedforward network is specified by the foil owing characteristics. The bottom and top layers are for input and output, respectively. Every unit receives inputs from layers lower than its own and must send output to layers higher than its own. Given and input vector, the output vector is computed activity levels in the earlier layers CHAPTER 9 Engineers or managers use simulation to improve the performance of existing manufacturing organization, as well as to plan and design new systems. For example it is very unlikely that anyone todat would spend a substantial amount of money on building new plant without having thoroughly checked its future performance by some sort of simulation study. On the other hand researchers have been using simulation techniques to seek a better understanding of the charecteristics of various manufacturing systems in an attempt to develop more rational organization structures, more effective manufacturing strategies and more efficient production control policies. Some theoritical techniques, such as those of probability and statistical analysis and some results from queuing theory with particular attention paid to their applications in relation to computer simulation. The implementation of Advanced Manufacturing System creates major problems for the system designers and engineers who are responsible for the succesful design implementation and operation of the systems concerned. Computer simulation has found a wide range of applications in the design of advanced manufacturing system, due to the fact that computer capabilities, combined with the versatility of simulation techniques, from a very powerful and flexible tool to assist the tasks involved. A computer simulation model can cope with the complexity of the systems and, in provides an effective communication tool to explain their operation, clearly to prospective investors. CHAPTER 10 Manufacturing strategies are based on flexibility as the competitive thrust. Responsiveness to market, new organization forms, introduction of new products, international thrusts, upgrade in manufacturing system. Technical planning systems, decentralization, intelligent manufacturing system, computer integrated manufacturing planning and simplification are the results of these new manufacturing strategies. The purpose in modern intelligent system design is to specify, design and implement systems that have a high degree of machine intelligence so that these systems can learn autonomously and adapt in uncertain or partially known enviroments. Modern industrial or manufacturing system requires many different cooperating knowledge agent and forms of expertise to be fused together to accomplish a common purpose or goal. Charecteristics of intelligent manufacturing xx systems are explained in reference to the outline for a theory of intelligence proposed by Albus. CHAPTER 11 The artificial neural networks (ANN) are applied to systems which are diificult to model with nonlineer modelling methods, even if nonlineer characteristcs of teh systems are highly considerable and therefore modelling with classical modelling techniques is very hard thus the expectation from modelling success is very low. Most widely used technique among these techniques is Backpropagation algorithm. The working mechanism of this technique is described as that the technique is base on errors back propagation and when is finished the learning process it uses the obtained weight to solve problems. The most important characteristic of ANN is the ability of learning and adaptibility. To educate