Search academic texts

Information × Registration Number 0216U001497, 0113U000361 , R & D reports Title Neuro-fuzzy systems for on-line clusterization and classification of data sequences in the conditions of their distortion by absent and abnormal observations popup.stage_title Head Bodyanskiy Yevgeniy Volodymyrovych, Доктор технічних наук Registration Date 17-02-2016 Organization Kharkiv National University of Radioelectronics popup.description2 An object of this research is hybrid systems of computational intelligence based on soft and liquid computing and used for current clustering and classification of distorted data. The project's aim is further development of theories, new methods and means of computational intelligence which are designated for solving a wide class of Data Mining problems, especially clustering and classification with the help of the latest advances in this area (liquid computing, spike-neural networks, fuzzy-type-2 systems) in a real-time mode under conditions of a lack of initial information and its curvature. Hybrid adaptive online clustering and classification methods which can process data in the form of "object-property" tables or multivariate time series have been developed. New fast clustering/classification methods which make it possible (unlike existing well-known methods) to process information in a sequential real-time mode, while a number of gaps as well as their location in data arrays are assumed to be unknown. Such modes as output arrays' recovery as well as their direct processing are provided, while a number of gaps is commensurate with a number of undistorted data. If a very short data sample is given, then a new method of spatial extrapolation with a set of different types of metrics that has to work under conditions of distorted information. The proposed neuro-fuzzy methods are capable of processing significantly distorted information in an online mode. Online data clustering and classification methods have been developed and researched. They process data in the form of "object-property" tables and short time series (taking into consideration multivariable vector and matrix ones) in the presence of abnormal outliers of unknown nature. These methods are described in the form of recurrent procedures or fuzzy inference rules to be later used as learning/self-learning algorithms for hybrid neuro-fuzzy systems. A distinctive feature of the proposed approach is the fact that special similarity functions which suppress abnormal outliers are used instead of traditional learning robust criteria or clustering objective functions. These functions are based on the so-called "partial metrics" that makes it possible to use them for processing data with gaps. Thus, the proposed methods are twice robust both in terms of outliers and gaps that contain lost observations. There's one more feature of the introduced approach. It can work under conditions when classes overlap. This means that one observation can simultaneously belong to two or more classes with different membership levels (in terms of the fuzzy sets' theory). The methods' recurrent presentation allows learning fuzzy self-organizing maps with their help. At the same time, it is important that similarity functions, membership functions and neighborhood functions of the self-organizing maps are in essence kernel (bell-shaped) structures which may have the same form of Cauchy functions. This simplifies greatly the methods' numerical implementation and increases a processing speed of distorted data. Architectures and on-line learning rules for neuro-fuzzy systems for clustering, classification and diagnosis have been developed. They are based on distorted data to be sequentially processed. The data contain both gaps and abnormal outliers at the same time. Moreover, while learning/self-learning a neuro-fuzzy system, it's assumed that not only synaptic weights, but actually both activation/membership functions and an architecture will be tuned. This fact allowed processing large data sets based on the concepts of Big Data and Data Stream Mining in a real time mode unlike other well-known approaches. The introduced immune approach to data classification and time series prediction under conditions of abnormal outliers in input data improves an efficiency of intellectual information processing for these conditions. The obtained results correspond to international standards and achievements and have no analogues as all known analogues are focused on processing data in a batch mode while processing operations and abnormal outliers is considered separately based on statistical methods that are tightly connected to specific types of distributions. Online versions of compatible nonlinear signal processing are currently unknown. Product Description popup.authors Єлісеєв А.О. Єлаков С.Г. Єськов Р.Г. Іващенко Г.С. Ілюнін О.О. Агафонов В.В. Аксак Н.Г. Алексенко В.С. Барковская О.Ю. Безсонов О.О. Богучарський С.І. Бойко О.О. Винокурова О.А. Ганжа Д.Д. Гришко А.О. Дейнеко А. А. Денищук П.М. Долотов А.І. Дрюк О.Д. Езе Ф.М. Житник І.О. Заїка О.А. Кіріченко Л.О. Кобилін І.О. Кобицька Ю.А. Колчигін Б.В. Коляда М.О. Копаліані Д.С. Корабльов М.М. Костин Д.Ю. Краснояружська К.Г. Куценко Я.В. Кушнарьов М.В. Лєщенко О.В. Ламонова Н.С. Легедіна О.В. Малишева Д.М. Машошин Д.А. Машталір С.В. Нечитайло А.Ю. Плісс І.П. Попов С.В. Самітова В.О. Сапожников Ю.А. Скуратов М.В. Соковікова Н.С. Сорокіна І.В. Сотніков О.М. Стольнікова М.З. Тімофєєв В.О. Татарінова Ю.Є. Тесленко Н.О. Тихонов А.О. Тищенко О.К. Удовенко С.Г. Ушков К.І. Фомічов О.О. Харченко О.О. Цибулько В.І. Чепенко Т.Є. Шафроненко А.Ю. Шкловець А.В. Шкуро К.О. popup.nrat_date 2020-04-02 Close