Neuro-fuzzy variable structure control of robotic manipulators
Neuro-fuzzy variable structure control of robotic manipulators
Dosyalar
Tarih
2000
Yazarlar
Palaz, Hasan
Süreli Yayın başlığı
Süreli Yayın ISSN
Cilt Başlığı
Yayınevi
Fen Bilimleri Enstitüsü
Institute of Science and Technology
Institute of Science and Technology
Özet
Yapay sinir ağlarının (YSA) ana özellikleri, örnekleri tanıyabilmeleri, girişleri sınıflandırabilmeleri ve öğrenerek dinamik çevre karşısında kendilerini adapte edebilmeleridir. Fakat akıllı kontrolördeki YSA'larının eşleme özellikleri siyah kutu yaklaşımıdır. Siyah kutunun ne yaptığını anlamak mümkündür, fakat nasıl yaptığını anlamak mümkün değildir. Bulanık lojik (BL) insan bilgi ve tecrübesini kullanmakta ve çıkarım yapabilmektedir. Fakat BL teorisinde öğrenme mekanizması bulunmamaktadır. Bu dezavantajları aşmak için Bulanık YSA (BYSA) geliştirilmiştir. Bu yapıda YSA parçası öğrenme için, BL bilgilerin temsili ve sisteme aktanmında kullanılmaktadır. Adaptif BYSA sistemi gerçeklenen sistem performanslarında artışa sebep olduğundan, sistem tanımlama ve kontrolünde geniş uygulama alanı bulmuştur. Değişken yapılı kontrolün (DYK), YSA ve BL'in kontrol uygulamalannda birbirlerini tamamlayıcı özellikleri vardır. Bu yüzden bu tekniklerin tümleştirilerek lineer olmayan sistemlerin kontrolünde kullanılmasında kazanımlar olacağı açıktır. Böylece ortaya çıkan kontrolör lineer olmayan bir kontrolör olacaktır ve klasik DYK'de karşılaşılan problemlerin çözümünü sağlayabilecektir. BYSA kullanılan kontrol sistemlerinin kullanım alanları basit sistemlerden (sadece birkaç değişkenin olduğu ve deneme yanılmaya izin verilen çamaşır makinesi, klima ve fırın gibi benzeri cihazlar) karmaşık sistemlere genişledikçe (çok fazla değişkenin bulunduğu ve başarısız deneme yanılmaya müsade edilemiyen kimyasal süreçler, güç santralleri, robotik sistemler ve uçaklar gibi benzeri sistemler), bu kontrol sistemlerin sistematik tasarımlarına acil ihtiyaç vardır. Ayrıca bu kontrol sistemlerinin aşağıdaki özelliklere sahip olması istenmektedir, 1. İyi bir izleme performansı sağlarken çatırtıyı azaltabiimelidir ve hatta yok etmelidir, 2. Kurallar tablosunu kendisi çalışma esnasında oluşturabilmelidir, böylece sistem boş kural tablosu ile çalışmaya başlayabilmelidir, XVII 3. Kontrol ve öğrenmenin aynı anda birlikte çalışma esnasında yapıldığı on-line kontrolör olmalıdır, 4. Kontrol sistemi gürbüz ve hata giderebilir yapıda olmalıdır, 5. Kontrol sistemi kararlı olmalıdır, 6. Kontrol sistemi yapısal bulanık kuramını kullanmalı ve böylece sistemin anlaşılması ve yorumu kolay olmalıdır, 7. Kontrol sistemin içerdiği üyelik fonksiyonları ve kural tablolarının çalışma esnasında değişimleri kolayca gözlenebilmelidir, 8. Kullanıcıya kontrol stratejilerini kontrolöre eklemeye izin vermelidir, 9. Çalışma esnasında kontrol sistemi bütün parametrelerini kendisi ayarlamalı ve optimum performansı vermelidir. Bu doktora tezindeki BYSA'lı DYK sistemleri yukarıdaki problemleri çözmeyi amaçlamaktadır. Bu yüzden ilk beş ihtiyacı karşılamak ve klasik DYK'ün dezavantajlarını gidermek için Gauss fonksiyonu tabanlı BYSA'lı DYK kontrol sistemi geliştirilmiştir. Bu kontrol sistemi BL ile YSA arasında yumuşak geçiş olarak tanımlanan radyal tabanlı YSA içermektedir. Adaptif BYSA'lı kontrol sistemi ilk sekiz ihtiyacı gidermek için önerilmiştir. Bu yapıda eşdeğer kontrol kestirimi kayma yüzeyi değişkenini içeren kurallar tablosu kullanılarak yapılmaktadır. Dokuzuncu ve son ihtiyaç için kendinden ayarlamalı BYSA'lı DYK sistemi üzerinde çalışılmıştır.
The main characteristics of Neural Network (NN) are to recognize patterns and to classify input, and to adapt themselves to dynamic environments by learning, but the mapping structure of NN in the intelligent controllers is a black box. It is possible to understand what the box does, but not how it is done conceptually. In the other words, the resulting NN behavior in the control scheme is difficult to understand. In addition, Fuzzy Logic (FL) can cope with human knowledge and can perform inference, but FL does not fundamentally include learning mechanism. Neuro-fuzzy computing has developed for overcoming their disadvantages. In general, the NN part is used for learning, while fuzzy logic part is used for representing knowledge. Adaptive Neuro-Fuzzy Inference Systems (ANFIS) has greatly improved the realization performance of fuzzy system and has extensively been used for identification and control purposes. It is seen that the Variable Structure Control (VSC), NN and FL methodologies are complementary for control objectives and there can be much to be gained in using them in a combined manner for the control of nonlinear systems. The resulting controller is a nonlinear one and may be suitable to overcome the difficulties involved in using conventional VSC for nonlinear systems. As the application area of neuro-fuzzy control expands from simple systems (in the sense that there are only a few key variables, and that large number of trial-and-error experiments are permitted, like in washing machine, air conditioners, rice cookers, etc.) to more complex systems (in the sense that there are many key variables, and that unsuccessful trial-and-error experiments are not permitted, like in chemical processes, power plants, robotic systems, aircrafts, etc.), there is a urgent need for systematic design method of the neuro-fuzzy VSC systems which have the following properties, 1. They alleviate the chattering phenomena while maintaining sliding behavior with accurate tracking performance, 2. They have a self organizing capability that can start from an empty rule base, xv 3. They produce an on-line controller that means the control and learning takes place simultaneously 4. They have a high degree robustness and fault tolerance, 5. They are stable neuro-fuzzy control system, 6. They introduce structured fuzzy concept that is easy to understand, 7. They visualize the evaluation of membership functions and the rule bases during operation, 8. They enable the designer to incorporate the experiences of control strategy into controller, 9. They have a self-tuning capability that can tune its parameters in on-line. The neuro-fuzzy approaches to VSC system in this doctoral dissertation try to solve of this kind of problems. Roughly speaking, the following ideas are proposed. For the first five requirements, A Gaussian Neuro Fuzzy Variable Structure Control (GNFVSC) scheme is presented to tackle drawbacks of conventional VSC. This controller consists of a Gaussian Radial Basis Function NN (RBFNN), which is considered as a smooth transition between FL and NN. For the first eight requirements, the adaptive neuro-fuzzy control system is introduced to estimate an equivalent control inside the boundary layer by incorporating linguistic control rules based on knowledge of the sliding surface variable into the controller. Finally, for the ninth requirement, a self tuning neuro- fuzzy VSC system is studied to eliminate fine tuning problem of the controller.
The main characteristics of Neural Network (NN) are to recognize patterns and to classify input, and to adapt themselves to dynamic environments by learning, but the mapping structure of NN in the intelligent controllers is a black box. It is possible to understand what the box does, but not how it is done conceptually. In the other words, the resulting NN behavior in the control scheme is difficult to understand. In addition, Fuzzy Logic (FL) can cope with human knowledge and can perform inference, but FL does not fundamentally include learning mechanism. Neuro-fuzzy computing has developed for overcoming their disadvantages. In general, the NN part is used for learning, while fuzzy logic part is used for representing knowledge. Adaptive Neuro-Fuzzy Inference Systems (ANFIS) has greatly improved the realization performance of fuzzy system and has extensively been used for identification and control purposes. It is seen that the Variable Structure Control (VSC), NN and FL methodologies are complementary for control objectives and there can be much to be gained in using them in a combined manner for the control of nonlinear systems. The resulting controller is a nonlinear one and may be suitable to overcome the difficulties involved in using conventional VSC for nonlinear systems. As the application area of neuro-fuzzy control expands from simple systems (in the sense that there are only a few key variables, and that large number of trial-and-error experiments are permitted, like in washing machine, air conditioners, rice cookers, etc.) to more complex systems (in the sense that there are many key variables, and that unsuccessful trial-and-error experiments are not permitted, like in chemical processes, power plants, robotic systems, aircrafts, etc.), there is a urgent need for systematic design method of the neuro-fuzzy VSC systems which have the following properties, 1. They alleviate the chattering phenomena while maintaining sliding behavior with accurate tracking performance, 2. They have a self organizing capability that can start from an empty rule base, xv 3. They produce an on-line controller that means the control and learning takes place simultaneously 4. They have a high degree robustness and fault tolerance, 5. They are stable neuro-fuzzy control system, 6. They introduce structured fuzzy concept that is easy to understand, 7. They visualize the evaluation of membership functions and the rule bases during operation, 8. They enable the designer to incorporate the experiences of control strategy into controller, 9. They have a self-tuning capability that can tune its parameters in on-line. The neuro-fuzzy approaches to VSC system in this doctoral dissertation try to solve of this kind of problems. Roughly speaking, the following ideas are proposed. For the first five requirements, A Gaussian Neuro Fuzzy Variable Structure Control (GNFVSC) scheme is presented to tackle drawbacks of conventional VSC. This controller consists of a Gaussian Radial Basis Function NN (RBFNN), which is considered as a smooth transition between FL and NN. For the first eight requirements, the adaptive neuro-fuzzy control system is introduced to estimate an equivalent control inside the boundary layer by incorporating linguistic control rules based on knowledge of the sliding surface variable into the controller. Finally, for the ninth requirement, a self tuning neuro- fuzzy VSC system is studied to eliminate fine tuning problem of the controller.
Açıklama
Tez (Doktora) -- İstanbul Teknik Üniversitesi, Fen Bilimleri Enstitüsü, 2000
Thesis (Ph.D.) -- İstanbul Technical University, Institute of Science and Technology, 2000
Thesis (Ph.D.) -- İstanbul Technical University, Institute of Science and Technology, 2000
Anahtar kelimeler
Bulanık mantık,
Robot denetim,
Robot kolu,
Yapay sinir ağları,
Fuzzy logic,
Robot control,
Robot arm,
Artificial neural networks