About Complex Objects Defining Via Integration of Data from Various Sources Anatoly D. Khomonenko Alena I. Zimovets Sergey P. Plyaskin Emperor Alexander I St. Military Space Academy Military Space Academy Petersburg State Transport name’s A.F.Mozhaisky, name’s A.F.Mozhaisky, University, 190031, zim_alenka@ramble plyaskin.work@yan Military Space Academy r.ru dex.ru name’s A.F.Mozhaisky, khomon@mail.ru about digital publications. In other words, information published on the web pages of the authors. At the same time there are arrays of official scientific documentation including reports about work done, projects results, personal stuff Abstract information about their achievements and etc. The The analysis of up-to-day problems in problems of store and processing information information technologies is carried out. appear in our nowadays live. There are a lot of These problems are connected with various information about different objects. Of increasing volumes of data in the whole course that fact is positive but at the same time it world. There is considered the problem of gives us problem of processing and analysis. complex object defining via integration of Analysts of IBM Company estimate Worldwide data from various sources. The first volume of data this way [Bar09]: approach to solve this problem is to use 2003 – 5 exabytes (1 Eb = 1 billion Gb); neural-fuzzy nets. The other approach is to 2008 – 0,18 zettabytes (1 Zb = 1024 Eb); use invariants of the moments. 2015 – more than 6,5 zettabytes; 2020 – 40-44 zettabytes (expected); 2025 – volumes will increase in 10 times. 1 Introduction It is important to know how to deal with this We live in the age of information. Revolution data and know what conclusions we can make. progress in electronics and computer technologies Today we have an opportunity to get leads to state when the one of tendencies of information about one complex object from various contemporary science development is increasing sources. There is a question about using this data volumes of experimental data.1 the most effectively. In this paper, we talk about The tendency of experimental plants of new how to recognize complex objects with help of generation design appears in worldwide science information gotten from various sources. last days. Registration of metrological data, tasks of In itself, digital format consists of concealed biological mathematics, astrological observations, data that is why it is necessary to use special aviation, power engineering and instrument analysis methods. Human recognizing and analysis making are not the only areas to discover and these of information can take a lot of time. If we have branches give us colossal volumes of information. giant volumes of data, the operativeness is The number of various articles, including scientific impossible. It is clear that using of automation tools ones, is growing like an avalanche. Firstly, we talk increase the effectiveness of this process. There are 2 contemporary and effective ways of complex Copyright c by the paper's authors. Use permitted under Creative objects identification. The one is identification via Commons License Attribution 4.0 International (CC BY 4.0). integration of data with help of neural-fuzzy nets In: A. Khomonenko, B. Sokolov, K. Ivanova (eds.): Selected Papers of the Models and Methods of Information Systems Research [Kho18]. The other one is identification of complex Workshop, St. Petersburg, Russia, 4-5 Dec. 2019, published at objects via invariants [Kho16]. http://ceur-ws.org. 46 Perspective calculating complexes intended to solve tasks of multichannel processing of 2 Identification of complex objects via information in real time for functional and data integration with help of neural- automatic control tools. Technical calculation complexes ensure solving fuzzy net following tasks [Laz16]: From system analysis point of view perspective - automatic processing of information from system of data processing must provide complex receiving tools; objects classification, identification them by known - automated control of hardware complex signs and forecasting of complex objects functionality; development. - representation and documentation of We can increase quality of classification and processing results; identification of complex objects with help of - control, diagnostic, optimization of the work program means and integration of data gotten from and ensuring stability of instruments functions. different sources. The other method is using The main stages of recognizing are collecting algorithms of fuzzy output. and processing of complex objects parameters We are able to get information recurred for gotten from various sources, conducting of complex classification from one source in insufficient objects catalogue, measures identification with volume. In conditions of incompleteness and catalyzed information, detection of new complex inaccuracy of information, the construction of a objects and definition their parameters, mathematical model is problematic. The solution is recognizing complex objects. to use few sources. It gives an opportunity to Both following methods have mathematical recover information. In the result, quantity of data apparatus of fuzzy sets theory, nets and measuring increases. and theory of planning. We can improve The information integration system frees users effectiveness and quality of recognizing and from the need to know the data from which sources analysis via using this methods and automatic they use, what are the properties of these sources systems. We also can create complex of analytical and how to access them. Sources of data can have information and knowledge with help of intelligent different properties. We can choose method of information technologies. This complex will give us integration by source properties. With an increase an opportunity to get conclusions about complex in the volume of information, its duplication may objects in developing environment. take place. In connection with this, with an increase To achieve these goals, the following tasks must in quantitative indicators, qualitative indicators be solved: may deteriorate. Approximation is able to solve this 1) the creation of a single functional space problem. The universal and commonly used and indicators characterizing the state of complex method is least squares approximation. objects based on a central database of information Fuzzy output algorithms can used to recognize and knowledge with accumulation, storage, access complex objects, actually Neural-fuzzy net. Neural- and management; fuzzy net is a neural-fuzzy with fuzzy signals 2) integration existing local data bases in weights and activation function but at the same central information store; time with unification of xi and wi, p1 and p2 with 3) collecting, accumulation and using of using t-norma, t-conorma or other continuous experts’ knowledge in distributed bases for making operations [Kho18]. Inputs, outputs and weights of conclusions and recommendations; net are real numbers within limits of [0,1] segment. 4) continuous observation (complex analysis) In other words, neural-fuzzy net is a neural net of current situation; designed on the basis of layered architecture with 5) increasing effectiveness and quality of using of “&”, “OR” neurons. analytical instruments; Process of work with data based on neural net 6) automation of analytical reports; consists of three steps. There are preparation of 7) visualization of information via pictorial data, extraction of the rules and estimation of them. graphics; Preparation process must define and process 8) expert instrumental and informational gotten data to make it good for specific methods of support of analytical activities. intellectual analysis. The main methods of rules 47 extraction are LTE (Limited Relative Error) how to do conclusions on incomplete or inaccurate method, black box method, method of fuzzy rules information. extraction from fuzzy nets, algorithm of particular Learning of neural net as known is rules extraction (Partial-RE), algorithm of full rules accomplished on the basis of exactly known extraction (full-RE). Rules of estimation correspond parameters and proper output characteristics following tasks. There are finding optimal [Kho18]. If availing data for learning is not enough, sequences of rules extraction, checking of precision information fills up via data integration from the rules extraction, definition of knowledge quantity other sources or with help of approximation. In the in neural net. result, we get common functional dependence It is necessary to teach neural net before using. projected output characteristics of complex objects In deep learning we use experts’ knowledge. from their parameters. Fecundity of identification and classification of complex objects increases when the machine learn P P P 1 1 1 P P P 2 2 2 S P P P 1 3 3 3 P P P 1 1 1 О S P P P 2 2 2 2 P P P 3 3 3 S P P P 3 1 1 1 P P P 2 2 2 P P P 3 3 3 Figure 1: Net Structure (о – complex object; s – source on information; p – parameters) By this approach, neural net turns out functional equivalent of some model of dependence between 48 variable like those that built on traditional modeling. recognizing of current states of the objects. Solving The dignity of this approach is commonality and these tasks in technical diagnostics depends on universality of admissible input parameters. models. Those models connect complex objects states Formal task of neural net learning can be described and its representation in set of diagnostic signs. in the following way [Kho18]. Parameters compared with calculated values Teaching base: X = – set of parameters; M estimated for every kind of complex objects to define = {mi | i = 1…m} – set of classes. current state in the set of parameters. Set of diagnostic Variety of recognition parameters: signs can be considered as array X([n, p]) consisted N = {nj | j = 1 … n}; from n measuring by p telemetric parameters. To find Set of states in which complex objects are able to 𝐾𝑖 = arg max[𝐹̂ (𝐾𝑖 )], stay presented with help of two-dimensional 𝐾𝑖 ∈ 𝐾 invariants of the moments. In this case, the model of recognizing can be 𝒮={𝑆𝑖│𝑖=(1,𝑚)} – set of states; presented like graph (Figure 1). The graph shows the Π={𝜋𝑗│𝑗=(1,𝑛)} – set of diagnostic parameters, all relationship between the recognition parameters of states 𝑆𝑖 ∈ 𝒮 pairwise distinct. complex objects. Vertices of the graph must contain This way 𝑆𝑖 corresponded to its defined vector 𝜋𝑗 weights defining probability of complex objects characterized by values of diagnostic signs. recognizing with present parameters gotten with help In any moment of the time t in T vector 𝜋𝑗 can be of statistical method. A back propagation algorithm presented as two-dimensional function f(x, y) where x used to select the training parameters of the presented defines time t, y is diagnostic sign, amplitude f is value multilayer neural network. of current diagnostic sign at the moment t. Fuzzy output system advantage is transparency of This approach can be used to recognize fuzzy nets that possible by its linguistic interpretation information presented as images [Kho16]. Relevance in fuzzy production rules. Disadvantage is a priori of this problem consists in the fact that the main part components definition of these models [Dol14]. of information processed by human in process of Advantage of neural nets is opportunity to reveal of discovering design and control activity is graphic data sequences. In other words, there is extraction of information like photographs, diagrams, sketches and gotten knowledge. Disadvantage is complexity of net drafts. Apparatus of imagine representation and size and structure definition. Disadvantages description based on invariants calculation is needed compensated by combined using with their to solve identification tasks. advantages. Linguistic structure of rules base Central moment of digital image f(x,y) described promotes to comprehend and analyze the system. This way neural-fuzzy net is a processor with  pq =  ( x − x ) p ( y − y )q f ( x, y), massive parallelization of operations. This processor x y has essential capability to save experimental were x = m10, / m00 , y = m01 / m00 . knowledge and make it accessible for following usage. Nets are similar to human brain in two ways. Net gets In the process of analysis seven invariants are used. knowledge in learning process and use quantity of The parts of considered image are invariant to transfer intensity inter-neuron connects. turning axial symmetry pressing and stretching. 1 = 20 + 02 ; 3 Complex objects identification via invariants of the moments 3 = (30 − 312 ) 2 + (321 − 03 ) 2 ; Fecundity of identification and classification can be 4 = (30 + 12 ) 2 + (21 + 03 ) 2 ; increased with using program means and integration of data gotten from various sources and with help of invariants of moments. 5 = (30 − 312 )(30 + 12 )((30 +12 )2 − 3(21 +03 )2 ) + It is known that the set of complex objects states can be presented not only in the view of diagnostic +(321 −03 )(21 + 03 )(3(30 + 12 )2 − (21 + 03 )2 ); features but with set of invariants values [Kho16]. In our case, the identification and classification tasks must considered like common theory of image 49 6 = (20 −02 )((30 + 12 )2 − (21 + 03 )2 ) + +411 (30 + 12 )(21 + 03 ); Input Variables: • T is the longitude of the ascending angle;  7 = (321 −03 ) (30 + 12 )((30 +12 )2 − 3 ( 21+ 03)2 ) − • i is the inclination; • Ha – semi major axis;; −( 30 − 3 12)(21 + 30 )(3 ( 30+ 12 )2 − ( 21+  03)2 ). • Hp – pericenter. The output variable corresponds to various elements of space debris: In itself image recognizing consists of two relatively 1– Idle satellites. independent tasks. The first is classification of groups 2 – Overclocking blocks. based on the adjusted requests. The second is taking 3 – Parts of space rockets. object to the one of them. 4 – Worked out steps. When solving the first task it needed to prepare 5 – Technological elements. data and form set of states for diagnostic signs. It is 6 – Wreckage. necessary that forms of representation diagnostic data 7 – Small element (up to 10 cm). and 𝑆𝑖 and 𝜋𝑗 vectors to be correspondent. It can be The ANFIS Editor with training set shown on the fig. discrete or analog forms. 2. When solving the second task it is required comparison between reference values and calculated invariants. If these values do not diverge from reference ones more than allowed quantity, then the state is defined. Thus we can conclude that complex objects defining via invariants of the moments is a compound task. We can also say that this method is very effective in analysis of information gotten from sources in image way. 4 An example of a neural-fuzzy network for classifying objects Figure 2: ANFIS Editor with training set Let's consider an example of constructing a neural- The graph of the dependence of learning errors on fuzzy network ANFIS [Jyh93] in MATLAB to classify the number of training cycles presented on the fi. 3. space debris objects. In table 1 a fragment of the training set for the neural-fuzzy network ANFIS presented. Table 1: A fragment of the training Set T i Ha Hp Exit 92.5935 51.65 422 403 1 91.22653 51.56 324 333 1 170.9687 53.32 5185 4367 2 98.4683 43.56 301 344 2 92.5944 51.66 416 456 2 189.367 63.34 1991 1286 2 104.046 98.15 712 69S 3 96.1496 94.11 38654 35120 3 Figure 3: The dependence of learning errors on the 98.9342 65.02 981 678 3 number of training cycles 93.7549 74.75 354 454 3 The given example shows the possibility of using 104.975 74.03 1016 243 3 the ANFIS fuzzy neural network to solve the problem 718.157 69.68 11234 6918 3 of space debris classification when integrating data 1437.62 13.77 594 578 4 from various sources. 50 algorithms and cognitive graphical 5 Conclusion representation. – SPb.: MSA name’s А.F.Mozhaisky, p. 11-17. It is considered two ways of solving problems of complex objects recognizing by integration of data [Kop15] Е.V. Kopkin, V.А. Chikurov, V.V. Aleynik, from different sources. The one of them is using О.G. Lazutin (2015) Algorithm for neural-fuzzy nets. That method gives us an constructing a flexible program for technical opportunity to analyses data consisted of incomplete object diagnosing on the criterion of information. Another method is considered as using received information value. SPIIRAS invariants of the moments. This method is the most Proceedings. Vol. 4(41). – Pp. 106-130. effective recognizing complex objects in image form. [Kop15] Kopkin Е.V., Kravcov А.N., Lazutin О.G. (2015) Informational technologies influence on various Selection of discrete diagnostic features scientific fields. Nowadays instrumental tools develop taking into account their value for the in very active way. This tools give an opportunity to recognition of the technical condition of the get information about different complex objects like object. Information And Space. № 2. Pp. 111- human genome or data about distant star systems. We 117. can get detailed definition about all possible objects and concepts by big data analysis. [Bub15] V.P. Bubnov, М.L. Gluharev, А.А. Kornirnko, Proposed directions of development give us S.A. Krasnov, V.V. Rogalchuk, A.V. Tyrva, V.V. substantial building up possibilities of information Fedyanin, A.D. Khomonenko (2015) Models recognizing systems. Firstly it touches on efficiency of of information systems. Textbook. Moscow, complex objects identification and classification. 2015. [Isa13] E.A. Isaev, V.V. Kornilov (2013) The Problem of Acknowledgments Processing and Storage of Large Amounts of Scientific Data and Approaches to Its The work was partially supported by the grant of Solution. Mathematical Biology and the MES RK: project No. AP05133699 "Research Bioinformatics. Vol. 8, №1, Pp.49-65 and development of innovative information and telecommunication technologies using modern cyber [Man11] V.G. Manzhula, D.S. Fedjashov (2011) technical means for the city's intelligent transport Kohonen Neural Networks And Fuzzy system". Neural Networks in Data Mining. Fundamental Research. № 4. p. 108-115 References [Yan16] M.V. Yanaeva, E.V. Sinchenko (2016). [Kho18] A D. Khomonenko, E. L. Yakovlev (2018) The Indistinct neural nets in data mining. rationale for the architecture of the Scientific works of KubSTU. № 15. Pp.80-93 convolutional neural network for object [Kog10] M.R. Kogalovsky (2010). Methods of Data recognition on images on-board computer Integration in Information Systems. system. H&ES Research. №Vol. 10. No. 6, Pp. Electronic Publications of the RAS p.9 86-93. [Dol14] O.N. Dolinina, A.Y.Shvarts (2014). Several [Kho16] A D. Khomonenko, E. L. Yakovlev (2016) State aspects of increasing of the effectiveness of estimation of complex object via invariants. the making decisions by intelligent systems 3RD Russian conference "Mathematical using fuzzy inference. Fundamental modeling and information technologies", research. №12. Pp. 691-695. MMIT, Ceur Workshop Proceeding. Pp.44- 49. [Bar09] A.A. Bargesjan (2009). Data and process [Dmi87] A.K. Dmitryev, R.M. Yusupov (1987) analysis. Textbook. BHV-Petersburg. 512 p. Identification And Technical Diagnostics. [Jyh93] Jyh-Shing, Roger Jang. ANFIS: adaptive- Textbook for universities. Leningrad.: EM network-based fuzzy inference system. IEEE USSR.521 p. Transactions on Systems, Man and Cybernetics : journal. 1993. Vol. 23, No. 3. [Laz16] Lazutin O.G. (2016) The technique of bringing DOI:10.1109/21.256541. information about the technical condition of space vehicles using data compression 51