AUTHOR QUERIES

AUTHOR PLEASE ANSWER ALL QUERIES

PLEASE NOTE: Please note that we cannot accept new source files as corrections for your article. If possible, please annotate
the PDF proof we have sent you with your corrections and upload it via the Author Gateway. Alternatively, you may send us
your corrections in list format. You may also upload revised graphics via the Author Gateway.

Carefully check the page proofs (and coordinate with all authors); additional changes or updates WILL NOT be accepted
after the article is published online/print in its final form. Please check author names and affiliations, funding, as well as the
overall article for any errors prior to sending in your author proof corrections. Your article has been peer reviewed, accepted as
final, and sent in to IEEE. No text changes have been made to the main part of the article as dictated by the editorial level of

of
service for your publication.

AQ:1 = Please provide department name for the South Ural State University.
AQ:2 = Please confirm or add details for any funding or financial support for the
research of this article.

ro
AQ:3 = If you haven’t done so already, please make sure you have submitted a
graphical abstract for your paper. The GA should be a current figure or
image from your accepted article. The GA will be displayed on your
articles abstract page on IEEE Xplore. Please choose a current figure from
the paper and supply a caption at your earliest convenience for the
graphical abstract. Note that captions cannot exceed 1800 characters
EP
(including spaces). If you submitted a video as your graphical abstract,
please make sure there is an overlay image and caption. Overlay images are
usually a screenshot of your video that best represents the video. This is for
readers who may not have access to video-viewing software. Please see an
example in the link below: http://ieeeaccess.ieee.org/submitting-an-article/
AQ:4 = Note that equations (9)–(10) are missing in this article, so we have
renumbered the rest of the equations to maintain the sequential order.
Please check and confirm.
IEE

AQ:5 = Please note that Refs. [1] and [26] were identical in your originally
submitted manuscript. Hence, we have deleted Ref. [26] and renumbered
the subsequent references. This will also be reflected in the citations
present in the body text. Also, issue no. or month for the same.
AQ:6 = Please provide the publisher name and publisher location for Refs. [2], [3],
[9], [11], and [23].
AQ:7 = Please provide the journal title and issue no. or month for Ref. [26].
 Received September 19, 2020, accepted October 8, 2020. Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2020.3032326

New Methods for Control System Signal Sampling

in Neural Networks of Power Facilities
KONSTANTIN OSINTSEV, SERGEI ALIUKOV , AND YURI PRIKHODKO
AQ:1 South Ural State University, 454080 Chelyabinsk, Russia
Corresponding author: Sergei Aliukov ([email protected])

of
AQ:2 This work was supported by the Government of the Russian Federation (Resolution No. 211 of 16 March 2013) under Contract
02.A03.21.0011.

AQ:3 1 ABSTRACT The authors of the paper emphasize that when the Nyquist-Shannon sampling theorem is used
2 in practice, there arise several problems, which can be explained only through the use of new methodologies
3 and mathematical models. The review of the researchers’ works, as well as the authors’ own practical
4

10
ro
research in the course of processing the statistical sample, which is described by a wave-like sine-cosine
function, leads to the conclusion that it is necessary to take into account optimization criteria for high-
tech processes and innovative indicators of building functions for the statistical sample, for example, when
signals are transmitted and sampled using neural networks at production facilities. In the practice of economic
calculations, for example, when making a graphic presentation of trend lines based on the functions built
subject to the sampling conditions by the Nyquist theorem, the authors propose to use new methods for
approximating piece linear functions, which allow for achieving a smaller error as compared to standard
EP
11 calculation methods. The work resulted in the creation of a neural network regulation algorithm, which will
12 be trained based on the collected data and adapted to a specific type of a boiler unit. Besides, it was established
13 that the task of neural network algorithms in the program is to find the optimal value of the weight coefficient
14 for each argument of the resulting function to obtain the maximum number of predictions of the flare level
15 and the particle burn-up time, which are consistent with reality. The use of these methods for the first time
16 made it possible to significantly reduce the error, which is confirmed not only by calculations, but also by
17 experimental data.

18 INDEX TERMS Approximation, neural networks, piece linear functions, signal sampling.
IEE

19 I. INTRODUCTION to the sampling conditions by the Nyquist theorem, the 33

20 The relevance of improving the methods for modeling data author [1] proposes to use new methods for approximating 34

21 transfer processes is determined by the need to increase their piece linear functions, which allow us to achieve a smaller 35

22 analytical capabilities and obtain more accurate solutions error as compared to standard calculation methods. The 36

23 when using the achievements of several fields of science. This theoretical existence of such a possibility is confirmed by 37

24 is a response to the challenges of the formulated integration- the methods of the geometric algebra of multidimensional 38

25 convergent organizational and technological paradigm of space, which studies computational approaches to compres- 39

26 the complex system development. In their early works [1], sion and stretching as space transformation types. Later, these 40

27 the researchers emphasize that when the Nyquist-Shannon approaches formed the mathematical foundations of quantum 41

28 theorem is used, there arise several problems, which can be mechanics. 42

29 explained only through the use of new methodologies and The additive contribution of the authors of the article to 43

30 mathematical models. Having faced the use of neural network the scientific component of the world knowledge base about 44

31 algorithms in the calculation practice, for example, the con- the subject of research, namely, the use of new mathematical 45

32 struction of trend lines based on the functions built subject methods in the approximation of functions in the process 46

of signal transmission from control objects to the neural 47

network, consists in the development of an optimization 48

The associate editor coordinating the review of this manuscript and procedure to minimize errors in data processing by neural 49

approving it for publication was Junhua Li . networks. The use of these methods for the first time made it 50

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 8, 2020 1
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

51 possible to significantly reduce the error, which is confirmed In the section ‘‘Comparison of the calculation results’’ the 107

52 not only by calculations, but also by experimental data. authors compare the results of experiments with a neural 108

53 The structure of the article is as follows: network, results of computer modeling using new methods 109

54 The Introduction presents the main provisions of the of approximation and without these methods in the current 110

55 Nyquist-Shannon theorem, indicates the problems that in state of science and technology. 111

56 science and in practice are solved with its help. In addition, The most famous recent study is presented by Kok and 112

57 additive contributions to science at the time of publication of Tam [2], [3]. They summarize the works dealing with signal 113

58 this article are indicated here. transmission and processing systems, as well as mathematical 114

59 In the Review of the literature, the authors presented the relations allowing one to implement such systems in practice. 115

60 genesis of this scientific direction and the most striking publi- The works of Basak et al. [4], Gama et al. [5] present the 116

61 cations of the authors, who timely solved scientific problems basics of signal transmission. Ortega et al. [6] clearly shows 117

62 arising from the smooth development of branches of technol- the advantages of new signal sampling methods, the same 118

of
63 ogy associated with the transmission of signals from control author (Ortega et al. [7]) shows the disadvantages of using 119

64 objects. the Fourier series in signal processing and transmission. 120

65 The following is the formulation of the scientific problem The prerequisites for [1] are the fundamental work often 121

66 that the authors solve in this article. In the same chapter, referred to by researchers, namely Mayers et al. [8] dealing 122

67 the relevance of the solution to this problem and the prospects with mathematics in signal processing. 123

68 that open up when a ready-made solution is applied in prac- Besides, we should note that, until recently, the specialist 124

69 tice is given. Sozanski [9] has been engaged in analog signal sampling. 125

70

71

72

73

74

75

76
ro
In the section ‘‘The existing mathematical methods for
describing the signal sampling dynamics’’, the authors in
more detail than in the introduction and review of the lit-
erature reveal the scientific and practical problems that are
solved using the Nyquist-Shannon theorem. It also describes
the branches of economics and technology, where these solu-
tions are most in demand. However, here the authors empha-
size that at the moment in the development of science and
He also studied the data transmission stability [10].
Oshana’s [11] and Meyer-Base’s [12] studies deal with
the signal conversion for neural network algorithms; they are
close to the topic of our research.
The studies of O’Shaughnessy [13], Khobotov et al. [14]
deal with improving the efficiency of devices when process-
ing signals in neural networks.
The works of Malsagov M.Y. et al. study the weight
126

127

128

129

130

131

132
EP
77 133

78 technology, it was impossible to advance in terms of reducing coefficients [15] and training of neural networks [16]. The 134

79 the error in transferring data from the control object to the researchers Kryzhanovsky and Kryzhanovsky [17] studied 135

80 neural network without using a new mathematical apparatus. directly the signal sampling. Reference [18] indicates the 136

81 In the section ‘‘The use of piece linear functions for sig- advantages and disadvantages of various sampling meth- 137

82 nal sampling and new approximation methods of Doctor ods. During sampling, there arise problems addressed by 138

83 S.V. Alyukov exemplified by the functions from the Nyquist Kryzhanovsky B. et al. 139

84 theorem’’, the authors talk about their early achievements in The authors would also like to mention the books under- 140

85 the development of a new mathematical apparatus as applied lying the design of neural networks [19], their practical 141

86 to various branches of technology. In particular, it is described application [20], and the manifestation of oscillations during 142

87 in more detail about applications in the field of energy facil- solving [21]. 143
IEE

88 ities. Then, a proof is provided regarding signal sampling For example, in [22], methods for increasing the efficiency 144

89 methods based on the new approximation methods of Doctor of data transmission in an analog way were presented. In [23] 145

90 S.V. Alyukov exemplified by the functions from the Nyquist the same thing, but for in digital data processing. However, 146

91 theorem. in these and other similar works, the issue of reducing the 147

92 In the section ‘‘The application of the development to error in data transmission has not been fully resolved. 148

93 neural network algorithms and model adaptation based on the

94 new data sampling systems’’ the adaptation of the mathemat- II. CONTROL SYSTEM SIGNAL SAMPLING 149

95 ical model is carried out due to the maximum approximation A. PROBLEM STATEMENT 150

96 of the measurement frequency to the prediction step. It is Based on the review of the researchers’ works, as well as our 151

97 also emphasized here that training a neural network for an own practical research, for example, when processing a sta- 152

98 energy object is multi-criteria, and therefore should have the tistical sample described by a wave-like sine-cosine function, 153

99 smallest errors when processing each individual criterion in we can conclude that it is necessary to take into account the 154

100 the network. criteria of efficiency and optimization of high-tech processes 155

101 The section ‘‘An object for using the new neural net- and innovative indicators for building functions for the statis- 156

102 work signal sampling system and mathematical modeling’’ tical sample. Moreover, it is necessary to develop a mathemat- 157

103 contains a specific control object, as well as the results of ical apparatus, which could take into account the drawbacks 158

104 its mathematical modeling. It is emphasized here that new of using calculations by the Nyquist theorem. The paradigm 159

105 approximation methods can also be used to reduce the error defines the trends to ensure the sustainability of the results of 160

106 in mathematical modeling in standard software packages. the effective control system development through the use of 161

2 VOLUME 8, 2020
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

162 processes of improving individual elements and systems. This analysis depth and accuracy. However, we often observe the 214

163 theoretical and methodological approach allows us to deeper phenomena of poor matrix conditioning of equation systems, 215

164 analyze processes in production equipment using neural net- poor convergence of the infinite series of exact analytical 216

165 works and building trend lines. solutions, which affects the validity of sampling parameters. 217

The dynamics of systems can be described through the 218

166 B. THE RELEVANCE OF THE RESEARCH TOPIC use of linear or nonlinear differential or finite-difference 219

167 Currently, neural network algorithms are based on the pro- equations. Deterministic continuous functions or lattice func- 220

168 cessing of the data coming from the control object with a tions, which model phenomena and processes in technolog- 221

169 certain time interval. Signals from the control object are sam- ical systems, are often the solutions of these equations, but 222

170 pled. The reliability degree of the sampled signals does not sometimes they can be predetermined based on the theoretical 223

171 always satisfy the technical assignment to the development provisions or simply on the historically established modeling 224

172 of neural network algorithms for production equipment and practice. In this case, the use of continuous time is conve- 225

of
173 systems. nient for analytical modeling; it allows us to use differential 226

174 In connection with the drawbacks when the existing signal calculus tools. At the same time, discrete models facilitate 227

175 sampling methods are used, the study is aimed at developing the construction of algorithms for their computer implemen- 228

176 a methodology for an effective combination of numerical and tation. For spatial series of dynamics, there are regressions 229

177 analytical methods of mathematical modeling based on new with stochastic factor variables. Their identification is more 230

178 mathematics, which take into account the unsteady nature difficult and has its own specific features. 231

179 of the studied processes. That is, the authors identified the Mathematical economics of production facilities considers 232

180

181

182

183

184

185
need to take into account the contingence conditions and the
variety of boundary conditions.

C. NEW SCIENTIFIC RESEARCH METHODS

For the first time, we use the mathematical appara-
tus for approximating functions proposed by Professor
S.V. Alyukov, which allows us to expand the signal sam-
pling capabilities for more accurate regulation of jump-like
ro continuous models and their solutions to a greater extent,
while econometrics, which manipulates statistical data, uses
the results of sampling these models. Nevertheless, in the
well-known econometric literature, the sampling procedure
is viewed rather superficially, without analyzing the errors,
which may occur, and the conditions to be observed.
Most researchers confine themselves to a reference to the
Nyquist-Shannon discrete sampling theorem, according to
233

234

235

236

237

238

239
EP
186 240

187 processes. To eliminate the Gibbs effect and the sampling which the continuous time function ϕ(t) will be completely 241

188 error, we propose to approximate the initial step function by determined by the sequence of its values, which are apart 242

189 a sequence of recursive periodic functions. from each other by no more than by a polling step 243

190 In the work presented by us, mathematical dependencies

191 have been derived that help to minimize errors in data trans- 1 = fmax /2, (1) 244

192 mission to the neural network due to the approximation meth-

where fmax is the maximum frequency in the spectrum of the 245
193 ods of Doctor S.V. Alyukov. The authors confirm that earlier
sampled function. 246
194 such mathematical methods as applied to control objects and
The value inverse to the sampling period is called the linear 247
195 neural networks were not applied in practice.
sampling frequency 248
IEE

196 D. PRACTICAL IMPLICATIONS fs = 1/1. (2) 249

197 Signal sampling is needed to control process facilities through
198 the use of neural networks. When such technologies are used, We should pay attention to the fact that the sampling 250

199 for example, at power facilities, it significantly reduces the frequency in the time dynamics series is usually known, for 251

200 data transmission and processing error. This, in turn, affects example, an hour, a day, a week, ten days, a month, a quarter, 252

201 the qualitative and quantitative characteristics of the control a year, and can vary only within certain limits. 253

202 object. The developed mathematical models also allow us to For each model of a dynamic trajectory, we can calculate 254

203 optimize the control process and reduce the energy costs of its spectrum and frequency fmax and choose a relevant polling 255

204 the control object. For example, fuel consumption is reduced period 1 in time. It should be recognized that this can refer 256

205 after optimizing the parameters of a flare in the furnace of a both to temporal and spatial dynamics. At a high dynamics 257

206 boiler unit. described by the polynomials of the second and third order, 258

the step should be small. 259

207 E. THE EXISTING MATHEMATICAL METHODS FOR Let us illustrate this condition by the example of a 260

208 DESCRIBING THE SIGNAL SAMPLING DYNAMICS harmonic: it can be a component model of a dynamics 261

209 The combination of exact analytical and approximate numer- series or an integral part of a more complex model, Fig. 1. 262

210 ical approaches provides universality and simplicity with the According to the Nyquist-Shannon theorem, a harmonic 263

211 accuracy sufficient for practice. Computer facilities allowed signal with a period T can be adequately represented by 264

212 one to solve nonlinear differential equations, high-degree its observations, if its frequency does not exceed half the 265

213 algebraic equations and their systems with a sufficient sampling frequency; this frequency fN is called the Nyquist 266

VOLUME 8, 2020 3
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

FIGURE 1. ‘‘Correct’’ restoration of the sine-cosine function by a

frequency lower than the Nyquist frequency.

of
FIGURE 3. An example of pressure oscillations during the flame front
propagation.

example, automatic control theory, information and signal 298

transmission theory, equations of mathematical physics, dif-

267

268
FIGURE 2. Oscillations in an ‘‘incorrect’’ data sampling.

frequency
fN = fd /2 = 1/21.
Thus, there are negative consequences of an incorrect
ro (3)
ferential equations etc. Similar developments were proposed
by the authors in the field of power engineering [22], [23].
Besides, the authors initially developed a mathematical appa-
ratus for approximating piece linear functions [24], [25].
Only initial steps are being taken currently to apply systems
with step characteristics and functions in heat power engi-
neering. Such systems belong to highly nonlinear structures
299

300

301

302

303

304

305

306
EP
269

270 choice of a sampling frequency due to a significant differ- characterized by difficult obtaining of exact solutions. 307

271 ence between the real sampling frequency used to record the
272 indicator and the Nyquist frequency. Even more so, the con- III. NEW APPROXIMATION METHODS OF PROFESSOR 308

273 sequences are catastrophic if the optimal sampling frequency S. V. ALYUKOV EXEMPLIFIED BY THE FUNCTIONS 309

274 is chosen incorrectly when choosing a trend line in the stock FROM NYQUIST THEOREM 310

275 market, Fig. 2. The existence of the Gibb’s effect when functions are approx- 311

276 It is also reflected during construction, for example, when imated by trigonometric expressions also makes us take 312

277 processing the data for gas-dynamic discontinuous flows in a critical look at the proof of some important theorems, 313

278 the energy sector. An example of such currents is a propagat- which reduces the accuracy of calculations. In particular, 314

279 ing flame front. the Kotelnikov’s theorem, known in the English literature as 315

the Nyquist theorem, is widely used in the theory of signal

IEE

316
280 Discontinuous flows, shock waves, and compression
281 shocks occur in subsonic and supersonic gas flows. In this transmission. To prove the Nyquist theorem, the so-called 317

282 case, we introduce the concept of contact surfaces of a gap integral sine defined by the following expression is used to 318

283 between two gases, which remain on both sides of the surface approximate the functions 319

284 and do not mix. Oscillations appear when considering discon- Zx

sin t
285 tinuous gas flows jointly with the use of central-difference Si(x) = dt. (4) 320

286 methods, Fig. 3. t

0

287 F. THE USE OF PIECE LINEAR FUNCTIONS Based on the integral sine, Nyquist builds the following 321

288 FOR SIGNAL SAMPLING function to prove the theorem 322

289 The presence of a discontinuity of several processes, implicit Si (T (ω + ω1 )) − Si (T (ω − ω1 )) , (5) 323

290 oscillations inside the model due to the data transfer time
291 response – all this preconditioned the need to display step where ω is the argument, T , ω1 are some parameters. At the 324

292 functions. same time, he affirms that with an increase in T this function 325

293 Analytical methods are needed to provide a more profound tends to the limits shown in Fig. 4a, and it is equal to: zero at 326

294 analysis of the process, accuracy, and convergence of solu- ω > ω1 ; π at ω < ω1 . 327

295 tions. It is proposed to use step functions in the considered S.V. Alyukov proved that the graph of the limiting func- 328

296 types of nonstationary processes. Their traditional areas of tion will have the form shown in Fig. 3b. That is, for any, 329

297 application are technical and mathematical disciplines, for even arbitrarily large but finite values of the parameter T , 330

4 VOLUME 8, 2020
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

Proof: At points x = 0 and x = π/2, we have 366

fn (x) − f0 (x) = 0, ∀n ∈ N . (11) 367

Consequently, at these points 368

fn (x) −−−→ f0 (x) . (12) 369

n→∞

FIGURE 4. Graphs of the limiting function in the Nyquist theorem (a) and since 370

clarifications of S.V. Alyukov (b).

∀ε > 0 ∃n∗ ∈ N ∀n : n > n∗ 371

331 there will always be such ω < ω1 , for which the values ⇒ |fn (x) − f0 (x)| = 0 < ε. (13) 372

of the function built by Nyquist will differ from π, and

of
332

333 there will always be such ω > ω1, for which its values We can take, for example, n∗ = 1. 373

334 will be non-zero. Moreover, it is important to note that the Since 374

indicated difference with an increase in T does not tend

sin x > (2/π) · x, ∀x ∈ (0, π/2) ,
335
(14) 375
336 to zero but to a certain number other than zero, approxi-
337 mately equal to 0.281 [22]. This indicated number is known then, for any x ∈ (0, π/2) the condition is met 376

338 in the global scientific community as Alyukov’s constant,

339 CA = 0.281. fn (x) = sin ((π/2) · fn−1 (x)) > fn−1 (x) > . . . > f1 (x) > 0. 377

340

341

342

343

344

345

346
The mathematical apparatus proposed by S.V. Alyukov for
approximating functions allows us to expand the technical
and economic capabilities of more precise regulation of jump-
like processes.
ro
The expansion of function (3) in the Fourier series has all
the aforesaid disadvantages, namely the Gibbs effect and a
significant approximation error. To eliminate them, we pro-
posed to approximate the initial step function by a sequence
Then, the sequence fn (x) , ∀x ∈ (0, π/2) is positive,

which we denote

lim f n (x) = A ∈ R.
n→∞
(15)

increasing and bounded, and, consequently, has a finite limit,

(16)
378

379

380

381

382
EP
347
We obtain that 383
348 of recursive periodic functions
fn (x) |fn (x)
  A = lim sin((π/2) · fn−1 (x)) 384
 n→∞
349 (π/2) fn−1 (x) , ⊂ C ∞ [−π, π] . (6) = sin((π/2) · lim fn−1 (x)) = sin((π/2) · A), (17) 385
= sin n→∞
f1 (x) = sin x, n − 1 ∈ N

whence we find that A = 0 or A = 1. Since the sequence is 386
We assume that
of positive terms and increasing, then A = 1= f0 (x). Then,
350
387

351 {fn (x)} ⊂ L2 [0, π/2] . (7) in the considered interval, taking into account the previous 388

conclusion that the sequence converges at the points x = 0 389

AQ:4 352 Since, due to the limitation of functions fn (x) and the ratio and x = π/2, we conclude that 390
IEE

π/2
sup Var fn = 1 < ∞, (8) fn (x) −−−→ f0 (x) , ∀x ∈ [0, π/2]. (18) 391
353 n→∞
n∈N 0
This convergence is only pointwise but not uniform since 392
354 due to the monotoneness of the functions fn (x) in the section the function f0 (x) is not continuous in the section [0, π/2]. 393
355 [0, π/2], then, we obtain that Theorem 2: In Banach L1 [0, π/2] and Hilbert L2 [0, π/2] 394

356 sup sup |fn (x)| = 1 < ∞. (9) spaces of measurable functions a sequence of approximating 395

n∈N x∈[0, π/2] functions fn (x) converges in norm to the initial function f0 (x). 396

Proof: Let us introduce a sequence of functions minorant 397

Then, based on Helly’s theorem, in the sequence {fn (x)} we
to the sequence of fn (x) functions
357
398
358 can extract a subsequence converging at each point [0, π/2]
ηn (x) |ηn (x)
 
359 to some function f , wherein
⊂ C ∞ [0, π/2] . (19)
= (2/π) · arctg (n · π) ; n ∈ N
399
π/2 ___ π/2
360 Var f ≤ lim Var fn . (10)
0 n→∞ 0 We can show that 400

361 Let us show that such function f can be the initial function fn (x) ≥ ηn (x) , ∀n ∈ N , ∀x ∈ [0, π/2] . (20) 401
362 f0 (x) = sign(sin x).
363 Theorem 1: A sequence of functions fn (x) converges to the Notably, the measure of the discontinuing set of the func- 402

364 initial function f0 (x), wherein the convergence is pointwise tion f0 (x) is equal to zero. Then, taking into account the 403

365 but not uniform. sign of non-negativity and limitation of the functions fn (x) 404

VOLUME 8, 2020 5
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

of
FIGURE 5. Screening coal dust on a vibrating screen.

405 and ηn (x) in the considered section, in space L1 [0, π/2] we

406 obtain
Zπ/2
407 ||f0 (x) − fn (x)|| = (1 − f n (x))dx
0

408

409

410 Since
π πn
≤
Zπ/2

+
0
1
πn

π
(1 − ηn (x))dx = − arctg
2

· ln 1 + (π n)2 /4 .
πn
2

ro (21)
FIGURE 6. The appearance of the database for the neural network.

and flue gases, the average specific heat of fuel combustion,

emission of volatiles, fractional composition, excess air ratio,
carbon monoxide content in flue gases. A part of database
containing the relevant information is presented in Fig. 6.
Currently, the authors are working at the possibility to take
437

438

439

440

441
EP
 
1 
411 lim −arctg + · ln 1 + (πn)2 /4 = 0, (22) into account the slag formation of the furnace walls based 442
n→∞ 2 2 πn on the readings of the temperature difference on the heating 443

412 in this case surface, the temperature of the slag formation of the fuel, and 444

the chemical composition of slag-removing water. 445

413 kf0 (x) − fn (x)k −−−→ 0. (23)
n→∞ To create a software product, we chose the MicrosoftVisu- 446

414 Similarly, we can prove that the sequence fn (x) converges alBasic 6.0 programming support environment. The resulting 447

415 in norm to the function f0 (x) in space. Thus, the sequence product accepts for input the argument values through text 448

416 of the approximating functions fn (x) in spaces L1 [−π, π] boxes. The argument values include the values of the method- 449

417 L2 [−π, π] is fundamental. In space C [−π, π], the sequence ology appearing as non-computable and characterizing the 450

418 fn (x) is not fundamental. composition and properties of the fuel, the properties of the 451

flare, flue gas parameters, and the geometry of the boiler unit.
IEE

452

419 IV. APPLICATIONS One changeable argument is chosen from the arguments; the 453

420 A. THE APPLICATION OF THE DEVELOPMENT TO NEURAL limits of its change are set as a percentage of the specified 454

421 NETWORK ALGORITHMS value. 455

422 Notably, for the successful adaptation of the model, the mea- After pressing the ‘‘Calculate’’ button, all the functions 456

423 surement frequency should be brought as close as possible to are calculated for the arguments specified in the text fields. 457

424 the prediction step. The prediction step, in turn, is determined To build a graph of each function, we created a separate 458

425 by the dynamics of the process. Combustion processes are button. After pressing this button the function is linked with 459

426 fast, however, taking into account the decrease in the fuel the selected argument. To build a graph, we calculate the 460

427 heterogeneity at a high grinding degree of coal dust and high- average value of the function for the limits of change and 461

428 quality mixing, as well as the low rate of daily fluctuations the two extreme values, after which the graph is centered in 462

429 in the temperature of the outdoor air supplied for oxidation, the graphics window and scaled. Then, we calculate an array 463

430 we can assume that measurements taken once a minute will be of evenly distributed N points, which are plotted on the graph. 464

431 fairly close to each other in order to record the direct influence
432 of regulation on the combustion process. The regulation of the B. MODEL ADAPTATION BASED ON THE NEW DATA 465

433 coal dust grinding fineness is controlled by a neural network SAMPLING SYSTEMS 466

434 on a vibrating screen, Fig. 5. The speed of the model response to any changes in the com- 467

435 Prediction is based on a variety of readings, such as the bustion process dynamics is characterized by the adaptation 468

436 feed rate and temperature of the fuel, air, air-fuel mixture parameter. The model training process consists of choosing 469

6 VOLUME 8, 2020
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

470 the best adaptation parameter based on the samples on a

471 retrospective material.
472 Mathematical models are employed to describe time
473 series [23]. Imagine that the time series xτ generated by some
474 model can be represented as two components
475 xτ = ξτ + ετ , (24)

476 where ετ is a time series component generated by a random

477 non-autocorrelated process with a zero mathematical expec-
478 tation and finite dispersion, which affects only the value of
479 the synchronous series term; ξτ is a time series component
FIGURE 7. Prediction of the flare length with different weight
generated by a deterministic function or a random process,

of
480 coefficients at the temperature argument in the active combustion zone.
481 which determines the value of several or all subsequent series
482 terms.
483 When building the predictive model, a hypothesis is put
484 forward on the dynamics of the function ξτ depending on
485 the time argument. The considered models are endowed with
486 adaptive properties, the ability to adjust the initial hypothe-
sis or to replace it with another one, which is more adequate
487

488

489

490

491

492

493
in terms of prediction accuracy.
The simplest adaptive model is based on calculating the
exponential average. The basis of exponential smoothing is
formed by the calculation of exponential averages; it is car-
ried out according to the recurrence formula:
Sτ = γ · xτ + β · Sτ −1 ,
ro (25)
FIGURE 8. A grid laid on the solid model of the burner device.

random deviations, it is necessary to reduce γ . Thus, the task

of neural network algorithms in the program is to find the
optimal value of the weight coefficient at each argument
of the resulting function to obtain the maximum number of
522

523

524

525
EP
494 Sτ is the value of the exponential average at the moment τ ; γ predictions of the flare length and the particle burn-up time, 526

495 is the smoothing parameter, wherein γ = const and 0 < γ < which are consistent with reality. 527

496 1; β = 1 − γ .
497 Transforming the formula (23), we obtain: C. AN OBJECT FOR USING THE NEW NEURAL NETWORK 528

SIGNAL SAMPLING SYSTEM

Sτ = Sτ −1 + γ · (xτ − Sτ −1 ) .
529
498 (26)
During the mathematical modeling of the combustion process 530

499 The exponential average at the moment τ is expressed here in ANSYS, we built a solid three-dimensional model of the 531

500 as the exponential average of the previous moment plus the channel of the burner device and laid a grid on it, Fig. 8. 532

501 fraction γ of the difference between the current observation The flow of the fuel and air mixture was d in two cases, 533

502 and the exponential average of the past moment. namely with 12 percent and 16 percent oxygen. This con- 534
IEE

503 The value Sτ is the weighted sum of all the series terms. dition was necessary for complete modeling of the mixing 535

504 Moreover, the weights fall exponentially depending on the process and chemical reactions at different distances from the 536

505 antiquity of the observation. The value of Sτ has the same cut of the burner since the burners are located in two tiers 537

506 mathematical expectation as the series õ but a smaller above each other on a real boiler unit. Thus, on the lower 538

507 dispersion. tier, the condition is fulfilled, under which the fuel burns out 539

508 The smaller is γ , the more the dispersion of the exponential immediately and completely with 16 percent oxygen in the 540

509 average is reduced and the more the oscillations of the initial air. At the same time, on the second tier, at a certain dis- 541

510 series are suppressed. tance from the cut of the lower burners, oxygen is gradually 542

511 Fig. 7 presents a series of graphs showing the deviations of consumed, and its content approaches 12 percent. In the real 543

512 the predicted values of the flare length from its actual value boiler unit, the fuel burn-up degree can be estimated not only 544

513 at various values of the smoothing parameter, located next by the oxygen content but also by the presence of CO in 545

514 to the temperature argument in the active combustion zone, the combustion products. In fact, the higher the CO content 546

515 captured after changing the fuel type (13:00 08.04.2019). is, the lower the degree of the coal dust burn-up gets, and 547

516 In the case of a short-term prediction of combustion pro- the worse the process of the organic fuel combustion in the 548

517 cesses, it is necessary to reflect changes as quickly as possible furnace of a boiler unit is organized. These data are collected 549

518 and simultaneously clear the data from random fluctuations. by the neural network; the signals are sampled according to 550

519 Thus, on the one hand, it is necessary to increase the infor- the technology developed by the authors. 551

520 mation weight of more recent observations, which can be As a result of the mathematical modeling of the coal dust 552

521 achieved by increasing γ , but on the other hand, to smooth combustion process, we obtained approximately equivalent 553

VOLUME 8, 2020 7
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

of
FIGURE 9. Building a mathematical model of the velocity distribution
based on the data collected by the neural network.

FIGURE 12. The scheme of combustion automatic control: Br is fuel

consumption, nspp is speed of the solid fuel conveyor belt, Kfj , Ktj are
correction factors affecting the determination of the optimal flame
temperature Tfi and the temperature of the combustion products at the

ro outlet of the furnace Tt ’’i , which are compared with standard values of
these temperatures Tfjn , Tt ’’n
i
.

obtained from similar devices using the developed sampling

technology were used as initial conditions.
Regulation becomes possible in the presence of a propa-
gating combustion wave depending on the temperature and
concentration of the reacting substances µ : f (T , µ) =
561

562

563

564

565
EP
C −1 q(T , µ) − ξ (T − T1 ). Thus, the alleged analogies of the 566

energy efficiency function and the non-stationary combustion 567

FIGURE 10. Building a mathematical model of the carbon oxide

process become apparent. The first derivative of this function 568

distribution based on the data collected by the neural network. is shown in the block diagram (Fig. 12) a and is associated 569

with the choice of the type of fuel, its concentration, as well 570

as the load of the boiler unit. The second derivative is deter- 571

mined by the type of high-tech burner devices b. The third 572

reduces the role of deviations of the absolute temperature 573

of the flame and combustion products at the outlet of the 574

furnace c. The fourth eliminates emergency situations during 575

IEE

the operation of the boiler unit d. 576

The simulation based on the fourth derivative displays the 577

processes of minor adjustment of the regulatory subsystem to 578

obtain more accurate solutions in the conditions of sponta- 579

neous processes of increasing heat transfer (Fig. 12). Succes- 580

sive actions of the system of regulators in an interconnected 581

circuit and control loop with direct and feedback connec- 582

tions provide new properties of the energy-efficient operation 583

of the object under study. The diversity of the regulatory 584

FIGURE 11. Building a mathematical model of the nitric oxide parameters taken into account necessitates a transition to a 585
distribution based on the data collected by the neural network.
three-dimensional space for measuring interconnections, an 586

increase in the number of considered factors of the boundary 587

554 fields shown in Fig. 9, 10 and 11. Fig. 9 shows the velocity conditions, and leads to an increase in the multifactorial 588

555 fields for the lower tier of the burner devices with 16 percent nature and complexity of the functional and probabilistic 589

556 oxygen concentration in the air. Fig. 10 shows the concentra- effects on the indicator-property under study. 590

557 tion of carbon oxide, and Fig. 11 – the concentration of nitric

558 oxide. D. COMPARISON OF THE CALCULATION RESULTS 591

559 The authors presented the results of mathematical model- The authors emphasize that the obtained modeling results 592

560 ing in the ANSYS computer environment. The empirical data (Fig. 13) in ANSYS are very akin to the theoretical 593

8 VOLUME 8, 2020
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

REFERENCES 629

[1] S. V. Alyukov, ‘‘Approximation of step functions in problems of mathemat- 630

ical modeling,’’ Math. Models Comput. Simul., vol. 3, 2011, Art. no. 611, 631
doi: 10.1134/s2070048211050036. 632 AQ:5
[2] C.-W. Kok and W.-S. Tam, ‘‘Signal sampling,’’ in Digital Image Interpo- 633
lation in MATLAB. 2019, ch. 1, pp. 1–15, doi: 10.1002/9781119119623. 634 AQ:6
[3] C.-W. Kok and W.-S. Tam, ‘‘Partitioned fractal interpolation,’’ in Dig- 635
ital Image Interpolation in MATLAB. 2019, ch. 9, pp. 263–290, doi: 636
10.1002/9781119119623. 637
[4] B. Guler, A. Jayawant, A. S. Avestimehr, and A. Ortega, ‘‘Robust graph 638
signal sampling,’’ in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. 639
(ICASSP), May 2019, pp. 7520–7524. 640
[5] F. Gama, A. G. Marques, G. Leus, and A. Ribeiro, ‘‘Convolutional neural 641
network architectures for signals supported on graphs,’’ IEEE Trans. Signal 642
Process., vol. 67, no. 4, pp. 1034–1049, Feb. 2019. 643

of
[6] A. Ortega, P. Frossard, J. Kovacevic, J. M. F. Moura, and 644
P. Vandergheynst, ‘‘Graph signal processing: Overview, challenges, 645
and applications,’’ Proc. IEEE, vol. 106, no. 5, pp. 808–828, May 2018, 646
doi: 10.1109/jproc.2018.2820126. 647

FIGURE 13. Calculation errors during mathematical modeling a– ignoring

[7] B. Girault, A. Ortega, and S. S. Narayanan, ‘‘Irregularity-aware graph 648

the experimental data, b– taking into account the experimental data and Fourier transforms,’’ IEEE Trans. Signal Process., vol. 66, no. 21, 649

the developed signal sampling system, c– taking into account the pp. 5746–5761, Nov. 2018, doi: 10.1109/tsp.2018.2870386. 650

experimental data but ignoring the developed signal sampling system. [8] D. F. Mayers, G. H. Golub, and C. F. van Loan, ‘‘Matrix computations,’’ 651
Math. Comput., vol. 47, no. 175, p. 376, Jul. 1986, doi: 10.2307/2008107. 652

594

595

596

597

598

599

600
and experimental data. The differences are insignificant
since the error in the experimental measurements was leveled
by the use of modern tools certified by the European Union.
The error in the mathematical modeling was reduced due
to the use of the empirical data obtained by processing the
signals from analogous equipment and subsequent sampling
according to the technology developed by the authors. For
ro [9] K. Sozanski, ‘‘Analog signals conditioning and discretization,’’ in Digital
Signal Processing in Power Electronics Control Circuits (Power Systems).
2017, pp. 23–81, doi: 10.1007/978-1-4471-7332-8_2.
[10] K. Sozanski, ‘‘Signal-to-noise ratio in power electronic digital control
circuits,’’ in Proc. Signal Process., Algorithms, Archit., Arrangements,
Appl. (SPA), Sep. 2016, pp. 162–171, doi: 10.1109/spa.2016.7763606.
[11] R. Oshana, DSP Software Development Techniques for Embedded and
Real-Time Systems. Newnes, 2005, p. 608.
[12] A. Meyer-Base and V. J. Schmid, ‘‘Transformation and signal-separation
neural networks,’’ in Pattern Recognition and Signal Analysis in Medical
653
654
655
656
657
658
659
660
661
662
EP
Imaging. 2014, pp. 245–289, doi: 10.1016/B978-0-12-409545-8.00008-X. 663
601 example, this is evident by the velocity values v. During stan- [13] D. O’Shaughnessy, ‘‘Recognition and processing of speech signals 664

602 dard modeling without the use of sampled data, the velocity using neural networks,’’ Circuits Syst. Signal Process., vol. 38, no. 8, 665
pp. 3454–3481, 2019, doi: 10.1007/s00034-019-01081-6. 666
603 value ranged from 18-23 m/s, i.e., it averaged 20.5 m/s. At the
[14] A. G. Khobotov, A. I. Khilko, and A. A. Telnykh, ‘‘Studying the pos- 667
604 same time, when using the authors’ developments, the value sibilities and efficiency of operation with signals in neural networks,’’ 668
605 of the average velocity was 23.6 m/s. This value is the closest Radiophys. Quantum Electron., vol. 61, no. 2, pp. 69–76, 2018, doi: 669
10.1007/s11141-018-9871-x. 670
606 to the experimentally obtained values of 23.9 m/s. Thus, the
[15] M. Y. Malsagov, E. M. Khayrov, M. M. Pushkareva, and I. M. Karandashev, 671
607 maximum differences are obtained between the experimental ‘‘Exponential discretization of weights of neural network connections in 672
608 data and computer modeling data without the use of the pre-trained neural networks,’’ Opt. Memory Neural Netw., vol. 28, no. 4, 673

609 authors’ developments, while the maximum coincidence is pp. 262–270, Oct. 2019, doi: 10.3103/S1060992X19040106. 674
[16] B. Kryzhanovsky, M. V. Kryzhanovsky, and M. Y. Malsagov, 675
610 shown between the average values according to the model- ‘‘Discretization of a matrix in quadratic functional binary 676
ing results taking into account the sampled signals and the
IEE

611 optimization,’’ Doklady Math., vol. 83, no. 3, pp. 413–417, 2011, 677

612 experiment. doi: 10.1134/S1064562411030197. 678

[17] B. V. Kryzhanovsky and V. M. Kryzhanovsky, ‘‘An accelerated procedure 679
for solving binary optimization problems,’’ J. Comput. Syst. Sci. Int., 680

613 V. CONCLUSION vol. 48, no. 5, pp. 732–738, Oct. 2009, doi: 10.1134/s1064230709050074. 681
[18] B. V. Kryzhanovskii, V. M. Kryzhanovskii, and A. L. Mikaelyan, ‘‘Appli- 682
614 For the first time, we used a mathematical apparatus to cation of the clipping procedure to the binary minimization of a quadratic 683
615 approximate functions, which allows us to expand the signal functional,’’ Doklady Math., vol. 75, no. 2, pp. 310–313, Apr. 2007, doi: 684

616 sampling capabilities for more precise regulation of jump-like 10.1134/s1064562407020354. 685
[19] S. Haykin, Neural Networks: A Comprehensive Foundation. New York, 686
617 processes. NY, USA: Macmillan College Publishing Company, 1994. 687
618 To eliminate the Gibb’s effect and the sampling error, [20] P. K. A. Simpson, Review of Artificial Neural Systems II: Paradigms, Appli- 688

619 we proposed to approximate the initial step function by a cation and Implementations. San Diego, CA, USA: General Dynamics, 689
1988. 690
620 sequence of recursive periodic functions. [21] Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence. Berlin, 691
621 We compared the analytical calculations according to the Germany: Springer-Verlag, 1984. 692

622 developed algorithm with the experimental data. [22] A. Sudou, P. Hartono, R. Saegusa, and S. Hashimoto, ‘‘Signal recon- 693

623 When using systems to control process power engineer- struction from sampled data using neural network,’’ in Proc. 12th IEEE 694
Workshop Neural Netw. Signal Process., Sep. 2002, pp. 707–715, doi: 695
624 ing facilities through the use of neural networks and signal 10.1109/NNSP.2002.1030082. 696
625 sampling according to the developed methodology, the data [23] G. Starr, ‘‘Sampled data systems,’’ in Introduction to Applied Digital 697

626 transmission and processing error is significantly reduced. Controls. 2020, doi: 10.1007/978-3-030-42810-5_4. 698
[24] E. Toropov, K. Osintsev, and S. Aliukov, ‘‘New theoretical and method- 699
627 The developed mathematical models optimize the control ological approaches to the study of heat transfer in coal dust combustion,’’ 700
628 process and reduce the energy costs of the control object. Energies, vol. 12, no. 1, p. 136, Jan. 2019. 701

VOLUME 8, 2020 9
 K. Osintsev et al.: New Methods for Control System Signal Sampling in Neural Networks of Power Facilities

702 [25] E. V. Toropov, K. V. Osintsev, and S. V. Aliukov, ‘‘Analysis of the calcu- SERGEI ALIUKOV was born in Chelyabinsk, 722
703 lated and experimental dependencies of the combustion of coal dust on the Russia, in 1957. He received the M.S. degree 723
704 basis of a new methodological base of theoretical studies of heat exchange in mechanical engineering from South Ural State 724
705 processes,’’ Int. J. Heat Technol., vol. 36, no. 4, pp. 1240–1248, Dec. 2018. University, Chelyabinsk, in 1978, the M.S. degree 725
706 [26] S. Alyukov, ‘‘Approximations of generalized functions and their deriva- in mathematics from the Ural State University, 726
707 tives,’’ Que. At. Sci. Tech. Ser. Math. Model. Phys. Proc., vol. 2, pp. 5–62, Yekaterinburg, Russia, in 1984, and the Ph.D. 727
AQ:7 708 2013. degree in mechanical engineering and the doctor’s 728
degree in the field of mechanical engineering from 729
South Ural State University, in 1983 and 2005, 730
respectively. He is currently a Professor with South 731
Ural State University. He is the author of three books, more than 150 articles, 732
and ten inventions. His research interests include mathematical modeling 733
processes and applications, mathematical problems, and non-linear dynamic 734
systems. Some of his scientific research achievements have been applied in 735
engineering.

of
736

YURI PRIKHODKO received the M.S. degree 737

709 KONSTANTIN OSINTSEV was born in in heat and gas supply and ventilation engineer- 738
710 Chelyabinsk, Russia, in 1986. He received the ing from South Ural State University (SUSU), 739
711 M.S. degree in engineering and the Ph.D. degree in Chelyabinsk, Russia, in 2017, where he is cur- 740
industrial heat engineering from South Ural State rently pursuing the Ph.D. degree in industrial heat
712
713
714
715
716
717
718
719
720
721
ro
University (SUSU), Chelyabinsk, in 2008 and
2009, respectively. Since 2013, he has been the
Head of the Industrial Heat Department, SUSU.
He has published more than 100 high-quality
journal articles and conference papers and holds
more than ten patent rights. His several scientific
research achievements have been applied in industrial heat engineering.
His research interests include steam turbines of thermal power plants and
thermodynamics and heat transfer.
power engineering. From 2017 to 2019, he was a
Research Assistant with the Department of Heat
and Gas Supply and Ventilation, South Ural State
University, where he has been working as a Lec-
turer with the Department of Industrial Heat Power
Engineering, since 2019. His research interest includes the development
of heat processing and neuronet technologies based on ANSYS computer
program.
741
742
743
744
745
746
747
748
749
750
EP
IEE

10 VOLUME 8, 2020