Українська   Русский
DonNTU   Masters' portal

Abstract

Содержание

Introduction

The structure of the neural network came to the programming world straight from biology. Thanks to this structure, the machine acquires the ability to analyze and even memorize various information. Neural networks are also able not only to analyze incoming information, but also to reproduce it from their memory. In other words, a neural network is a machine interpretation of the human brain, which contains millions of neurons that transmit information in the form of electrical impulses.

A neural network is a sequence of neurons connected by synapses. The classification of neural networks by the nature of learning divides them into: -neural networks using supervised learning; -neural networks using unsupervised learning; -neural networks using supervised learning

1. Relevance of the topic

Nowadays, software development takes a lot of effort. Too much time is spent on what it would be logical to assign to computers - choosing a methodology, designing, writing a technical specification, testing - all this is done by a person, and does it very slowly. Doctor of Technical Sciences Maxim Shcherbakov in his lecture "Neural networks: maths & magic" LINK replied: "Yes, it is possible." It is good to have an automated assistant who will solve the problem simply by having certain criteria and templates, while he does not need to pay and he will do it in the shortest possible time, but what specific tasks can neural networks solve in our time? The development of machine learning nowadays is by leaps and bounds, not today, so tomorrow machines will be able to replace humans in routine tasks. Drawing up a technical assignment is not the most pleasant task, especially because of the need to paint everything in great detail and in form. Each company chooses a specific software development methodology and rarely changes it.

2. Purpose and objectives of the study, planned results

The aim of this work is to analyze and form the characteristics of artificial neural networks for solving the problems of “smart home”.

The main objectives of the study:

  1. Analysis of the characteristics of artificial neural networks.
  2. Evaluation of ways to reduce hardware costs by coding the states of Moore machines.
  3. Combining functionally different directions of optimization of neural networks in terms of hardware costs into a unified approach to synthesis and the formation of recommendations for its use.
  4. Development of alternative options for the main stages of a unified neural network synthesis process.

Research object: creation of neural networks.

Research subject: adaptive artificial neural network

As part of the master's work, it is planned to obtain relevant scientific results in the following areas:

  1. Adaptive Artificial Neural Network
  2. Neural network for solving "smart home" tasks
  3. Modification of well-known neural networks

3. Description of an artificial neural network for the "Smart Home"

3.1 Data scaling

Input variables must be scaled (scaled), that is, brought to a single range of variation. There are several reasons for the need to scale. After coding information, the inputs and outputs of the neural network can be heterogeneous quantities that vary in different ranges. It is desirable to bring all input variables to a single range and normalize (the maximum absolute value of input variables should not exceed one). Otherwise, errors due to variables varying over a wide range will have a stronger effect on network learning than errors due to variables varying over a narrow range. By ensuring that each input variable changes within the same range, we will ensure that each variable has an equal effect on the change in weights during training.

3.2 Creating tests and training datasets

In machine learning, there is a phenomenon called "overfitting". This happens when the models, while learning, become too confusing - they are well enough trained, but when new data is passed to them that they have never "seen", then the result they produce becomes poor. In other words, the models are not generated very well. To make sure we don't create overly complex models, it is common to split the dataset into training cases and test cases. The training dataset on which the model will learn, and the test dataset is the data on which the model will be tested upon completion of training. The amount of training data should always be greater than the test data. They usually occupy 60-80% of the dataset ..

3.3 Creating ANN

A biological neuron is imitated in the ANN through an activation function. In classification tasks (for example, the definition of spam messages), the activation function must have the characteristic of "switch". In other words, if the input is greater than a certain value, then the output must change the state, for example, from 0 to 1 or -1 to 1. This imitates a biological neuron "switching on". The activation function is usually used sigmoidal function:

4 Technologies of the "Smart Home" system

1-Wire Technology There are many power management technologies available today. FROM The most simple, popular and inexpensive to implement is 1-wire from Maxim / Dallas. 1-wire technology is far from new and has been known for about 10 years. 1-wire is widely used both in everyday life and in industrial systems. 1-wire advantages: - Simple and straightforward network architecture; - Low requirements for cables; - Long line length; - Low cost and simplicity of components; - Open protocol and available programming software; - Ability to do without food in certain situations. - Disadvantages: - Low data transfer rate; - Mandatory presence of a master leading the network. Communication devices are passive circuit elements. 1-wire components cannot send data to the network without a special request. 1-wire devices cannot communicate with each other without an “active” master. Network elements are always slaves. There is only one master in the 1-wire network. It initiates, monitors, and controls the operation of the network and devices connected to the network. Depending on the software used, the 1-wire line master can work with his wards in different ways, but only he can ultimately interrogate the state of a specific element, and also send a request to receive any information from this element. For example, if it is necessary to display the temperature value from multiple sensors in real time, then the network master in a cycle will poll these sensors in turn. Only one device can reply to the master over the network. An MC (microcontroller) can act as a master, which, with the appropriate software, can be directly connected to the 1-wire network to one of its input-output ports. A PC (personal computer) can also act as a master, using special interface elements that allow you to connect 1-wire to a COM port or USB port of a computer, such as DS9097 or DS9490. Ethernet technology is a very common standard. It is very well documented.

There are a huge number of controllers, microprocessors, individual chips, ready-made devices with embedded Ethernet. The Ethernet port integrates into home appliances, TVs, players, receivers, not to mention computers.

The minimum data transfer rate is 10/100 Mbps, which is more than enough for any home automation task. Ethernet technology has good noise immunity and reliability with a considerable segment length for a home. The technology contains all the necessary mechanisms for resolving collisions and monitoring the integrity of data transmission. Devices for "smart" networking (switches, routers) are sold in any store for almost a penny. In fact, Ethernet is now, if not the main one, then one of the main standards for organizing high-speed information exchange between various devices, both at home and in the office. In addition, in Ethernet technology, I was interested not only and not so much in the physical layer of the data model as in the network and transport.

It's no secret that Ethernet is most often used in conjunction with TCP / IP protocols. Smart Home is a set of simple devices that support TCP / IP, SNMP, SMTP and a Web interface based on the HTTP protocol, available for management, programming and monitoring from any standard client (computer, tablet PC, netbook, communicator, phone) as local network, and via the Internet, 3G / GPRS, Wi-Fi, etc. At the same time, devices can be controlled both directly and through a special controller (server), which could make such control more comfortable, including advanced features of the Web interface (Ajax, Flash). This is a very important point, since Ethernet does not have a mandatory network master and all devices can communicate with each other directly and remain available for management in case of any problems with the main controller. Many of the world's largest home automation companies are gradually migrating their product lines to Ethernet and TCP / IP technologies because it is convenient, fast, relatively cheap and easy. In addition, sockets, connectors, UTP / STP / FTP cables, stands, accessories, boxes, tools - all this is available anywhere in the world, which allows you to quickly and inexpensively deploy systems of any complexity. The topology of modern Ethernet-based networks is star. This means that it is necessary to lead its own cable to each device. This has a certain drawback, since in a house with ready-made finishing, such automation looks difficult.

4.1 Classification of neural networks

A neural network is a sequence of neurons connected by synapses REFERENCE. The structure of the neural network came to the programming world straight from biology. Thanks to this structure, the machine acquires the ability to analyze and even memorize various information. Neural networks are also able not only to analyze incoming information, but also to reproduce it from their memory. In other words, a neural network is a machine interpretation of the human brain, which contains millions of neurons that transmit information in the form of electrical impulses. Classification of neural networks by the nature of training divides them into:

Supervised learning assumes that for each input vector, there is a target vector that represents the desired output. Together they are called a learning pair. Usually the network is trained on a certain number of such training pairs. The output vector is presented, the network output is calculated and compared to the corresponding target vector. Then the weights are changed in accordance with an algorithm that seeks to minimize the error. The vectors of the training set are presented sequentially, the errors are calculated, and the weights are adjusted for each vector until the error throughout the training array reaches an acceptable level. Neural networks using unsupervised learning. Unsupervised learning is a much more believable learning model in terms of the biological roots of artificial neural networks. Developed by Kohonen and many others, it does not need a target vector for outputs, and therefore does not require comparison with predefined ideal responses. The training set consists only of input vectors. The training algorithm adjusts the weights of the network so that consistent output vectors are obtained, that is, so that presentation of sufficiently close input vectors gives the same outputs. The learning process therefore extracts the statistical properties of the training set and groups similar vectors into classes. .

Output

The creation of neural networks for various purposes and applications is being widely developed at the Donetsk National Technical University (Department of Computer Engineering).

Taking into account the above, that in the applied sense, a neural network allows you to get an acceptable solution in situations when it is difficult to find a suitable function, an algorithm from input parameters, or such a function will require too complex implementation and time. As applied to the "Smart Home" it is much easier and faster to teach the network by examples than to compose voluminous and complex programs. In other words, the process of learning (or self-learning) the network replaces the process of traditional programming. This sometimes saves a lot of developer time. Application of neural networks in Smart House systems:

You can think of many different uses for neural networks in home automation systems. In this section, I will publish my examples of using this approach.

List of sources

  1. Jesse, Russell Искусственная нейронная сеть / Jesse Russell. - М.: VSD, 2012. - 0 c.
  2. Jesse, Russell Нейрон / Jesse Russell. - М.: VSD, 2012. - 0 c.
  3. Барский, А. Б. Логические нейронные сети / А.Б. Барский. - М.: Интернет-университет информационных технологий, Бином. Лаборатория знаний, 2007. - 352 c.
  4. Барский, А.Б. Логические нейронные сети / А.Б. Барский. - М.: Интернет-Университет Информационных Технологий (ИНТУИТ), 2013. - 0 c.
  5. Бунаков, В. Е. Нейронная физика. Учебное пособие: моногр. / В.Е. Бунаков, Л.В. Краснов. - М.: Издательство Санкт-Петербургского университета, 2015. - 200 c.
  6. Головинский, П. А. Математические модели. Теоретическая физика и анализ сложных систем. Книга 2. От нелинейных колебаний до искусственных нейронов и сложных систем / П.А. Головинский. - М.: Либроком, 2012. - 234 c.
  7. Денис, Хусаинов Механизмы ритмической активности нейронов виноградной улитки / Хусаинов Денис , Иван Коренюк und Татьяна Гамма. - М.: LAP Lambert Academic Publishing, 2012. - 108 c.
  8. Как устроено тело человека. Выпуск 25. Нейроны. - М.: DeAgostini, 2007. - 30 c.
  9. Катехоламинергические нейроны. - М.: Наука, 1979. - 296 c.
  10. Круглов, В.В. Искусственные нейронные сети. Теория и практика: моногр. / В.В. Круглов, В.В. Борисов. - М.: Горячая линия - Телеком; Издание 2-е, стер., 2002. - 382 c.
  11. Мандельштам, Ю. Е. Нейрон и мышца насекомого: моногр. / Ю.Е. Мандельштам. - М.: Наука, 1983. - 168 c.
  12. Нейронные сети. Statistica Neural Networks. Методология и технологии современного анализа данных. - М.: Горячая линия - Телеком, 2008. - 392 c.
  13. Парвин, Манучер Из серого. Концерт для нейронов и синапсов / Манучер Парвин. - М.: Страта, 2015. - 408 c.
  14. Позин, Н. В. Моделирование нейронных структур / Н.В. Позин. - М.: Наука, 1970. - 264 c.
  15. Рассел, Джесси Вербализация нейронных сетей / Джесси Рассел. - М.: VSD, 2013. - 0 c.
  16. Рассел, Джесси Искусственный нейрон / Джесси Рассел. - М.: VSD, 2013. - 0 c.
  17. Татузов, А. Л. Нейронные сети в задачах радиолокации / А.Л. Татузов. - М.: Радиотехника, 2009. - 432 c.
  18. Толкачев, С. Нейронное программирование диалоговых систем / С. Толкачев. - Москва: РГГУ, 2016. - 192 c.
  19. Шибзухов, З. М. Конструктивные методы обучения сигма-пи нейронных сетей / З.М. Шибзухов. - М.: Наука, 2006. - 160 c.
  20. Юревич, Артур Нейронные сети в экономике / Артур Юревич. - М.: LAP Lambert Academic Publishing, 2014. - 80 c.tex-5, Spartan-3, and Newer CPLD Devices [Электронный ресурс]. – Режим доступа: http://www.xilinx.com/support....