Engineering science /12.àâòîìàòèçèðîâàííûå of a control system on manufacture
Zhikeyev
À.À.
The automation of technological operation usually consists of parts:
acceptance of the decision about necessity any of action and performance of
this action. The first part is most complex (difficult) in realization, as it
is traditional a task of acceptance of the decision the man incurs.
To facilitate process of acceptance of the decision the systems of
support of acceptance of the decision allow [1].
The process of acceptance of the decision with the help of system of
support of acceptance of the decisions is iterative process of interaction of
the expert and computer. The process consists of a phase of the analysis and
statement of a task for the computer which is carried out by the person,
accepting the decision, both phase of search of the decision and performance of
his (its) characteristics sold by the computer.
For the decision of tasks of support of acceptance of the decisions
methods of artificial intelligence, such as expert systems and artificial
neurons of a network often are used. The expert system, using knowledge
received from the experts in the given subject domain, decides (solves) the
same problems, the experts in which are these experts. Artificial neurons of a
network are trained on a set of the training data, then can decide (solve)
tasks for the data, which were not in training set. But there are reasons
limiting application of methods of artificial intelligence at the enterprises:
-
Complexity of preparation of the data to use.
- Complexity
of preparation of system to operation.
-
Complexity of problems, which are decided (solved) with the help of the given
means.
The assumption is done (made), that the
application of expert systems and neurons of networks in environment of uniform
information space allows lowering influence of the given reasons [2]:
- At use STEP as the standard on presentation of the
data, all data are represented uniformly. Each element of the data is one type. For each type given it is possible to define (determine)
rules of preparation them in expert systems and neurons networks.
- As all data are accumulated in uniform information
spaces, available there are large files for training in expert systems and
neurons networks.
- It is supposed to apply expert systems and íåéðîííûå of a network to
the decision enough small on dimension of tasks, that will allow effectively to
supervise their behavior.
- As at creation of uniform information space the
information model of the enterprise is under construction, the given knowledge
also can effectively be used.
For the purposes of automation of management of complex (difficult)
technological processes is used production expert system (PES). The systems
productions are one of the most widespread ways of representation of knowledge
in expert systems. The transformations, determining process, of the entrance
data are stored (kept) as rules (production), in such systems of knowledge.
[3].
The job production expert systems represent a sequence of iterations, in
each of which the rule gets out of base of knowledge which is applied to the
current contents of a context. The cycle of job production expert systems
(logic conclusion) comes to an end, when is deduced (removed) or the target
statement is denied. The logic conclusion can function on various algorithms
most widespread from which are the direct order of a conclusion and return
order of a conclusion.
Quality
of the decisions which are given out by expert system the higher, than more
than volume of knowledge of system. Accordingly, the problem of filling of
system by knowledge is one of prime for reception of efficient expert system
[4].
From various architecture neurons the networks would
like to allocate multilayer’s neurons of a network of direct distribution,
which have a number (line) of advantages at practical use:
- Uses as algorithm of training algorithm of return
distribution - first effective algorithm of training multilayer’s neurons of
networks;
- Plenty of applications for the decision of practical
tasks;
- Breadth of area of application - can be applied to
recognition of images, classification, forecasting, synthesis of speech,
control, adaptive management, construction of expert systems [5].
Multilayer’s neurons of a network of direct distribution represent
multilayer’s neurons of a network with consecutive communications
(connections). Neurons of the first (entrance) layer carry out only
distributive functions. They receive entrance signals and pass them neurons
second. Neurons of the second layer will transform a signal and pass it (him)
neurons of the third layer etc. Up to a target layer, this processes the
information from the previous layers and gives out a target signal. As function of activation neurons usually
use hyperbolic tangents or sigmoid.
The ability to training is determining property neurons of a network.
For the decision of any task by a traditional method it is necessary to know
rules, on which of the entrance data it turn out the decision of a task. With
the help neurons of a network the decision is possible to find, not knowing
rules, and having a set of examples, on which it is possible to train a
network.
For training multilayer’s neurons of networks use “training with the
teacher”. In training sample there is a correct answer to each entrance
example. The weights are insisted so that the network gave out meanings
(importance) as it is possible closer to known correct meanings (importance)
[6].
The purpose of training neurons of a network consists in achievement of
an opportunity of generalization. It means that the network allocates features
of the entrance data and begins to carry similar samples to one class. It also
raises stability to handicaps.
At training networks one of two following criteria of end of training,
as a rule, is used:
- End of training at achievement of the training, given in parameters,
of meaning (importance) of function of a mistake;
- End of training in case for all examples of training sample the
network gives out meanings (importance) appropriate reference.
Before training the initialization neurons of a network that is giving
to parameters of a network of some initial meanings (importance) is carried
out. As a rule, these initial meanings (importance) - some small random number.
For check of skills acquired by a network during
training, the imitation of functioning (testing) of a network is used.
It is possible to allocate the following parameters of
a problem, which show, that the put problem can effectively be decided (solved)
with the help of the device neurons of networks:
- The expert can not explain, how he decides (solves)
a problem.
- On a problem is saved (or the plenty of copies of
the decisions of the given problem for concrete entrance parameters can be
received).
- The problem can not be easily decided (easily
solved) with use of more traditional computing methods.
- The problem concerns to one of the following areas:
recognition of images, classification, forecasting, control, adaptive
management, experts' report.
- The decision of a task should be received in the
fixed time.
The choice
of architecture multilayer’s neurons of networks of direct distribution is
caused by wide variety of types of tasks solved by the given architecture, and
also large distribution of the given architecture in commercial
neurosimulations.
The literature:
1
Coltchin A.F.,
Ovsyannikov M.V., Strekalov A.F. etc. Management of life cycle of production. – Ì.:
Anaharsys, 2002.
2
Milayev V., Fatkin À., Ruleva Ò. Automation of management // PC Week/RE. – 2001. – ¹ 10.
3
Wotermen D. Management (manual) on expert systems. – Ì.: The world, 1989.
4
Shemelin V.Ê., Utrosyn V.V. Use of expert systems in environment
of information support of life cycle of a product // The incorporated scientific journal. – Ì., 2005. – ¹ 13.
5
Kruglov V.V., Borisov V.V. Artificial neurons of a
network. The theory and practice. - M.: a hot line - Telecom, 2002.
6
Golovko V.A. Neurons of a network: training,
organization and application: studies. The grant (manual) for high schools /
under general edition A.I. Galushkin – Ì.: 2001. – The book 4.