Introduction
Research aims and objectives
Relevance of the topic and motivation
The review of research and development
       The review of international sources
       The review of national sources
       The review of local sources
Review of existing diagnosis methods
Parametric identification
Neural network based techniques
Results, conclusions and prospects
References

Introduction

Modern design and development facilities are closely linked with the concepts of efficiency and reliability. In order to determine these parameters to date, there are various methods of diagnosis and verification, based on different approaches and using the latest achievements of science. These issues devoted many articles and publications written in the last ten or fifteen years abroad.

These days diagnosis includes a variety of industries and fields of knowledge such as medicine, engineering, economics and others. In what follows we will focus on technical diagnostics.

Research aims and objectives

Technical Diagnostics is part of the maintenance. The main goal of technical diagnostics is to reduce the maintenance costs of facilities and reduction of losses due to downtime resulting from failures [1]. Therefore it is important to constantly monitor and respond to various changes in the parameters of technical tools and devices.

The purpose of the master's final work is to study the methods and approaches in order to solve the problem of efficient parametric identification of linear analog devices, as well as seek opportunities in the introduction and use of this issue of neural network technologies.

To achieve these goals it is necessary to solve several problems, such as:

• review and analysis of parametric identification method as a method of detection and fault isolation of analog devices;
• development a method for parameter identification of linear analog circuits based on neural networks;
• development of an experimental parameter identification subsystem based on neural network;
• analysis of results and experimental studies of the model.

Relevance of the topic and motivation

Currently, most of the manufactured equipment belongs to the class of digital devices, but with an overall percentage reduction of the share of manufactured analog devices on the market, in absolute terms, their number increases [2]. With the development of technology, along with integration of observed phenomena such as reducing the size of devices down to nanotechnology, as well as their increasing complexity.

There are some problems associated with developing, designing and setting up specific analog devices, such as measurement error, the complexity of the process and modeling, control and accountability of states and modes of operation of all elements and components of the scheme.

Relevance of the topic is to propose a new approach to the problem of parametric identification using neural network technology, thereby reducing the number of performed actions designed to assess the serviceability and suitability of analog integrated circuits.

The review of research and development

Issues related to the diagnosis on the basis of analog neural networks, are being studied around the world, including scientists and engineers of the United States, China, Japan, India and Europe. However, in Russia and Ukraine, the issue is not paid enough attention, which indicates a lack of Russian-language information on both electronic and in printed form.

The review of international sources

Prominent scholars in the field of research methods of analog diagnosis using neural networks are V.K. Agrawal [3], A. Fort [4], P. Kabisatpathi [5], M.L. Bushnell [6], P. Sinha [7] and others. Their work, in particular, Fault detection and diagnosis in analog integrated circuits using artificial neural network in a pseudorandom testing scheme, Analog Automatic Test-Pattern Generation, SBT Soft Fault Diagnosis in Analog Electronic Circuits: A Sensitivity-Based Approach by Randomized Algorithms, reveal some of the issues in the field of analog diagnostki using neural network technology, as well as reflect the experimental findings.

The most complete neural network issues in the book Neural networks: a complete course by S. Haykin [8].

It is necessary to mention the Russian magazine Neurocomputers: development and application [9], aimed at studying some aspects of the theory of neural networks.

The review of national sources

At the national level on the analog diagnostics and neural networks devoted to a number of publications which are for the most part, familiarization.

The review of local sources

In the Donetsk National Technical University research on neural networks and diagnosis are conducted at the Department of Computer Engineering under the direction of Yury Zinchenko. Technical diagnostics and testing of analog devices studied by Masters Medgaus [10], Shigimagin [11] and Masyakin [12].

Review of existing diagnosis methods

Traditional methods applied in the 20th century, could not cope with the tasks of faster, better and cheaper diagnosis, which led to an increase in terms of design and testing costs. Therefore, further development of the technology required to improve the technical and, in particular, the diagnostic test [13]. The greatest use is made of the following techniques:

• DFT (Design For Test) or the technology of designing testability schemes - a technology that simplifies the development and implementation of production test, and also provides diagnosed microelectronic equipment;
• JTAG (Joint Test Automation Group) - boundary scan, which is used to verify that the in-circuit boards;
• ATPG or technology of automatic sample generation used for electrical testing of semiconductors, where the test set automatically generated by the program;
• BIST or built-in self test - the technology of designing additional hardware and software contained in the integrated circuit and allow verification of their work with their own schemes.

These technologies are partially implemented in software tools to create testability of electronic means, such as onTAP Boundary Scan Software, Pronto TEST-FIXTURE software, Galaxy Design Platform and Design For Test.

The above methods and software products are targeted for use in the diagnosis of digital electronic devices, printed circuit boards and integrated circuits. For the diagnosis of analog circuits using the following methods and approaches:

• method of reference;
• parametric identification;
• power supply control methods;
• leakage current monitoring methods;
• approximate methods and other.

For these techniques is characterized by common flaws, such as a large amount of computation, the need for access to all nodes in the scheme, the sensitivity to errors in calculations and, consequently, the difficulty of implementation.

Parametric identification

While constructing the models of testing schemes arises the problem of finding the numerical values of the unmeasurable constants from the available experimental data, that is, the values of measured variables (responses). This problem is called the problem of parametric identification. Its purpose is to find numerical values are unknown constants model, which would correspond to the solution of the problem, in a sense, the experimental data and thus these values do not contradict the physical sense and theoretical reasons, [14].

In the case of diagnosis of analog integrated circuits parametric identification is reduced to the analysis of the obtained values, as well as the selection of parameter values tested schemes that would lead to results typical of the standard model (operational).

In the process of analog devices are subjected to various influences caused by the different operating conditions. So, you can select settings and in-circuit circuit elements, such as current, voltage, current gain of transistors, as well as environmental parameters, which include room temperature, pressure, different vibration. Usually, the problems of adaptation for different devices, operating conditions are solved by another level of design and allowable ranges of values ??specified in the specification. Much more important is the determination of in-circuit parameters in order to analyze the health of individual sections of circuits and components, and determine whether the deviations of certain values.

For better understanding we must assume that there is a system consisting of several segments (blocks). Such a system can be represented as shown in Figure 1.

Figure 1 — An example of a simplified testing system

If we assume that both may deny only a block or section, the number of inoperable be Sn = N = 9. Discarding the unlikely failure (Blocks 6,7,8,9) can be obtained that the most probable number of faulty states of the Sn is equal to only 5. These states are:

• S1 - 1st block fault
• S2 - 2nd block fault
• S3 - 3rd block fault
• S4 - 4th block fault
• S5 - 5th block fault

In this example, attributes can be listed states such deviation parameters, such as the change in the amperage, voltage, resistance coils, contacts, and others.

In general, between the states of Si and their attributes Xj can meet the relationship, as shown below in Figure 2.

Figure 2 — Types of interaction of states and traits

To determine the type of relationship (or lack of relationship) between the selected states is commonly used logical analysis or a full-scale experiment. For this system can give the following example of cause and effect:

Figure 3 — Diagram showing the relationship of states and traits
(animation: 5 shots; 3 cycles repeat, 84 Kb)

The scheme allows you to select the minimum number of characters needed to monitor and detect all the five states do not work as well as eight pre-selected features is redundant. To facilitate the selection of a scheme of cause and effect relationships are not depicted in the graphic and in tabular (matrix) form, as shown below:

Table 1 - Matrix presentation of cause-effect relationships
Sign/State	S1	S2	S3	S4	S5
X1	0	0	0	1	0
X2	1	0	0	0	0
X3	1	1	1	1	1
X4	0	1	0	0	0
X5	0	1	0	0	0
X6	0	1	0	1	1
X7	0	1	0	1	1
X8	1	1	0	1	1

Table rows are formed by signs of states, and columns - states. The elements of the matrix are the zeros and ones, with a zero is put at the intersection of row and column if the corresponding attribute does not respond to the appropriate state. The unit is affixed to the opposite case.

Analysis of the table allows you to completely eliminate overlapping symptoms, ie those that repeat the combination of zeros and ones. Of the two interacting characters are usually removed one that is more difficult (expensive, long-term) control. In this example, are removed from further consideration by the signs of X4 (X4 sign of duplicate sign X5) and X6 (he, in turn, duplicated X7). In diagnosis think vzaimodubliruyuschie signs contain the same information.

If using all the controlled parameters during diagnosis (without screening them), the obtained diagnostic system will be overloaded with sensors, redundant circuits, and testing program will be very cumbersome. In this context, it becomes necessary to sample a minimum number of features, necessary and sufficient condition for recognition of each object. Usually this choice is based on the elements of information theory, some methods which can also help in solving the problem of determining the location of system failure (localization error) [15].

Neural network based techniques

Diagnosis is a special case of classification of events, with the greatest value is the classification of the events that are not in the training set of neural network. Neural networks are nonlinear systems to better classify objects and data than do linear methods, thus generalizing previous experience and applying it to new cases [16].

Currently, neural networks are used extensively in medicine and problems of medical diagnostics. However, there is a need to introduce and apply neural network technology in the field of technical diagnostics in order to ensure reliability and availability of devices, electronic circuits and equipment, which, though indirectly, but are also responsible for human life.

Modern developed devices and systems are characterized by high complexity of the processes. As a rule, the complex processes known only to information about the system response to input signals (the so-called black box), which leads to the description of processes based on input-output tables of data. Depending on the task of constructing the output signals from the input - the identification of models - consists of two phases: structural and parametric identification. Neural networks are an effective tool for modeling and can partially solve the problem of structural identification, because they have formally given the structure of connections between neurons and defined an algorithm for obtaining the output value. The main task of constructing neural network models is the training - parametric identification based on the training set. However, the same model in real applications based on the use of neural networks that contain hundreds of parameters to be optimized. In this regard, there is a problem of optimization and efficiency of the algorithms.

When building a neural network model aim to find the best algorithm for its work as an integrated system, the so-called global optimum, rather than individual sites (local optimum). The applied learning algorithms do not guarantee the globality of the solution found. Warranty to find the global optimum is the use of interval analysis techniques, however, interval optimization techniques have a low rate of convergence and are resource intensive, which leads to time-consuming, even with a small number of variables [17].

Another method for global optimization - a method of branch and bound. This method begins with determining the lower and upper bounds for the original problem. If the upper and lower bounds coincide, the result is an optimal value, and the method exits. Otherwise, the set of variables is divided into several proper subsets whose union coincides with the original set. These subtasks are descendants of the original. The algorithm is applied recursively to each of the subtasks, creating a tree of subproblems. If the optimal solution is found for some sub-tasks, it is achievable for the original problem (not necessarily optimal), but because it is achievable, it can be used to cut branches from the source tree. The search process continues for as long as each of the sub will not be resolved or thrown out or up until it reaches a preset threshold between the best solutions found and the lower bound of f(x) for all outstanding problems.

The method of distribution limitations or restrictions of the algorithm is computationally plan can be presented as an iterative process that successively narrows the area, guaranteed to contain all solutions. At each step of the iterative process is applied a narrowing of the operators, constricting the range of variables associated with a certain limitation. As a rule, the restriction operator specifies the domain of only one variable. Office of the iteration process is carried out by the data that is to be executed on the next step is chosen only the operator of restriction, the arguments which have been modified during the computation in the previous step.

Restriction operators can be represented in different ways. For example, the splitting of the initial system of equations for the primitive (unary and binary) relations through the introduction of additional variables. In this case, each primitive relation enables us to express each of the variables occurring in it by others. The whole set thus obtained the primitive equations and a set of operators will ask for the restriction of the original system. The advantage of this approach is the ability to solve problems with non-differentiable functions.

In 1992 he was described by the method of Hansen, whose essence lies in the sequential removal of the initial region of subdomains, which do not contain the global minimum.

The method of modeling the firing, which is representative of a family of Monte Carlo methods, based on the physical process of freezing liquids or metals during recrystallization annealing. As an estimate of the solution can be chosen point not only reduces the value of the objective function, but increases it. This behavior helps to find optimal solutions to avoid falling into local minimum point [18].

There are also a variety of algorithms, which include modifications and advantages of some of the above methods.

Results, conclusions and prospects

Technical diagnostics and testability design is an important step in the development of equipment. New technology and knowledge in this area, such as artificial intelligence and neural network methods for parameter identification and optimization, will continue to create highly intelligent devices and systems and put them into practice in various industries.

As part of this essay on the topic of the Master, a review of existing methods and diagnostic tools, analyzed modern software products, allowing the device to simulate the developed pre-production stage, and also lists the methods of optimization and parameter identification based on neural network models. In what is supposed to justify the use of the chosen method of parameter identification of analog devices and partially implement these methods in the form of ready-made solution or device.

As so as the master's work isn't complete yet, the full text of work and materials of topic can be obtained from the author or his scientific adviser after December 2012.

References

1. Machine fault diagnosis. Wikipedia. — Access mode: http://ru.wikipedia.org/wiki/...
2. Диагностика аналоговых схем с учетом тепловых режимов радиоэлементов. [Электронный ресурс]. — Режим доступа: http://www.quality-journal.ru/data/article/574/files/VolovikovaUvaysov@QJ0309.pdf
3. Jamuna.S, V.K. Agrawal. Implementation of BIST Structure using VHDL for VLSI Circuits. [Электронный ресурс]. — Режим доступа: http://www.ijest.info/docs/IJEST11-03-06-016.pdf
4. Cesare Alippi, Marcantonio Catelani, Ada Fort, Marco Mugnaini. SBT Soft Fault Diagnosis in Analog Electronic Circuits: A Sensitivity-Based Approach by Randomized Algorithms. [Электронный ресурс]. — Режим доступа: http://home.dei.polimi.it/alippi/articoli/01174051.pdf
5. Prithviraj Kabisatpathy, Alok Barua, and Satyabroto Sinha. Fault detection and diagnosis in analog integrated circuits using artificial neural network in a pseudorandom testing scheme. [Электронный ресурс]. — Режим доступа: http://www.buet.ac.bd/eee/publications2004/P013.pdf
6. Michael L. Bushnell, Vishvani D. Agraval. Analog Automatic Test-Pattern Generation. [Электронный ресурс]. — Режим доступа: http://masters.donntu.ru/2010/fknt/masyakin/library/article3_or.htm
7. Priyanka Sinha. Neural Network Automatic Test Pattern Generator. [Электронный ресурс]. — Режим доступа: http://www.eng.auburn.edu/~agrawvd/COURSE/E7250_05/REPORTS_TERM/Sinha_Neural.pdf
8. Хайкин С. Нейронные сети: полный курс, 2-е издание. : Пер. с англ. — М. :Издательский дом "Вильямс", 2006. — 1104 с. : ил. — Парал. тит. англ.
9. Журнал Нейрокомпьютеры: разработка, применение. [Электронный ресурс]. — Режим доступа: http://www.radiotec.ru/catalog.php?cat=jr7
10. Anton Medgaus. Development and research method of modeling faults of analog devices based on the language PSpice that describes analog-digital devices. Abstract. — Access mode: http://masters.donntu.ru/2011/fknt/medgaus/diss/indexe.htm
11. Alexander Shigimagin. Development and research of methods and structures of hardware analysis of FPGA based analog test reactions. Abstract. — Access mode: http://masters.donntu.ru/2010/fknt/shigimagin/diss/indexe.htm
12. Eugene Masyakin. Development and research of FPGA based hardware-oriented methods and structures of analog tests generating. Abstract. — Access mode: http://masters.donntu.ru/2010/fknt/masyakin/diss/indexe.htm
13. Зинченко Ю.Е. Методы и средства встроенного тестового диагностирования устройств специализированных сетей передачи данных реального времени. [Электронный ресурс]. — Режим доступа: http://hardclub.donntu.ru/zinchenko/cand.htm
14. Постановка задачи параметрической идентификации. [Электронный ресурс]. — Режим доступа: http://www.ccas.ru/konkom/ident.htm
15. Сафарбаков А.М., Лукьянов А.В., Пахомов С.В. Основы технической диагностики: учебное пособие. – Иркутск: ИрГУПС, 2006. – 216 с.
16. Нейронные сети в медицине и диагностике. Открытые системы. СУБД №04, 1997. [Электронный ресурс]. — Режим доступа: http://www.osp.ru/print_text.php?type=article&id=179201&isPdf
17. Сараев П.В. Программное обеспечение для исследования методов гарантированной параметрической идентификации. [Электронный ресурс]. — Режим доступа: http://www.econf.info/files/88/1187/0573.pdf
18. Долгов Ю.Г. Метод глобальной оптимизации на основе метода ветвей и границ. [Электронный ресурс]. — Режим доступа: http://www.sbras.ru/interval/Conferences/IMRO_04/Dolgov.pdf