: pp. 42-46
Lviv Polytechnic National University

The  current  article  describes  the  results  of  the  study  of  the  neural  networks  temperature  prediction  error dependence  on  measurement  errors,  which  are  random,  nonlinear  and  multiplicative  errors.  It  is  noted  applicability  of  the architecture of neural network for temperature prediction. The formula of temperature step response for ideal sensor is given. 

At  the very beginning an algorithm  for  calculating and creating  test sequences  for neural network  training  is developed. The  studies  described  in  this  article  are  implemented  in  the  computing  environment. There  are  given  formulas  and  figures  of measurement errors models. After considering  the measurement error, which neural networks were trained and verified with the training set.

The  results  of  the  study  of  the  temperature  prediction  error  dependence  on  the  multiplicative  measurement  error  and nonlinear measurement  error are presented. They  allow  conclude  that  raising  the measurement  errors with  the prediction errors increase. As the result, for the maximal measurement error (2.5 %) an absolute temperature prediction error is achieved at the level lower  than 5∙10-5 °C. The  results  of  the  similar  studies  of  dependence  on  the  random measurement  error  are  presented. They underline the mentioned errors increasing with the prediction error. For random measurement error (0.5 %) absolute temperature prediction error is of 0.5 °C and for 2.5 % random measurement error absolute temperature prediction error is of 1.5 °C.

It is described also the study of the temperature prediction error dependence on the aforesaid three types of measurement errors.

The major  conclusion  of  the  received  results  (the  dependences  of  the  temperature prediction  error  on  the measurement errors) consists in the next. The prediction temperature value slightly depends on multiplicative and nonlinear errors. In addition, the main impact on neural network temperature prediction error is caused by the random error.

[1]  F.  Bernhard,  Handbuch  der  technischen  temperaturmessung, Springer Vieweg, 2014. 

[2]  N.  Yaryshev,  Theoretical  basis  for  measuring  non-stationary temperature, Leningrad, USSR: Energoatomizdat, 1990.  

[3] D. Kriesel. A Brief,  Introduction  to Neural Networks, 2007.  [On-line],  Available: en/science/neural_networks. 

[4]  R.  Bordawekar,  B.  Blainey,  R.  Puri,  Analyzing Analytics. Morgan & Claypool Publishers, 2015. 

[5] O. Lopatko, I. Mykytyn, “Temperature value prediction errors  using  neural  networks  and  ideal  transition  process”, Measuring equipment and metrology, vol. 78, p. 20–24. 2017.

[6] M. Dorozhovets, V. Motalo, B. Stadnyk, Fundamentals of  metrology  and  measuring  technique,  vol.  1.  Lviv,  Ukraine:  Publ. House Lviv Pol. Nat. Univ., 2005.