Prediction of power loss and permeability with the use of an artificial neural network in wound toroidal cores

https://doi.org/10.1016/j.jmmm.2008.04.177Get rights and content

Abstract

This paper presents an analysis of use of artificial neural network algorithm for prediction of power loss and relative permeability in toroidal cores wound from grain-oriented electrical steel sheet and cobalt-based amorphous ribbon. The properties of the grain-oriented samples were measured at peak flux densities from 0.3 to 1.8 T and frequencies from 20 Hz to 1 kHz, and those of the cobalt-based samples were measured at peak flux densities from 0.1 to 0.5 T and over a frequency range from 20 Hz to 25 kHz. Measurements were carried out under sinusoidal flux density and pulse-width-modulated voltage supplies. In each case, 80% of the measured results were used for the training procedure and 20% for detection of over-training. It has been found that optimisation of training data significantly increases the accuracy of power loss prediction. The prediction errors of the range of measured results of power loss and permeability for the grain-oriented cores are lower than ±3% with 97% confidence level and ±4% with 83%, respectively. For the cobalt-amorphous cores, these values are ±10% with 95% confidence and ±10% with 85% confidence, respectively.

Introduction

The concept of artificial neural network algorithms (ANN) was first introduced in the middle of last century. Since then, the ANNs have been used for various tasks such as: optical pattern recognition, digital and adaptive filtering, control, prediction, estimation of credit conditions, stock market predictions, etc. [1], [2].

The prediction of power loss and permeability in soft magnetic materials is an important issue in many electromagnetic designs and performance assessments. Phenomenological and theoretical approaches have been investigated in the past. ANNs have also been used for this purpose with varying degrees of success [3], [4], [5], [6]. However, a modified formulation of the prediction problem used in this study leads to improved prediction accuracy.

Toroidal cores are widely used in applications where the supplying voltage and magnetic flux density vary sinusoidally with time, but in an increasing number of applications the voltage is generated electronically so it can be more conveniently controlled. In this case, the flux density waveform contains a series of time harmonics, which increase the total losses and reduce the effective permeability. A common case of this is pulse width modulation (PWM) generated voltage. In this study, magnetic properties were assessed under such conditions, as well as under pure sinusoidal magnetisation. However, the definition of effective permeability under PWM conditions is still to be formulated [7]. For this reason, the prediction of permeability for PWM magnetisation was not carried out in this investigation.

This paper presents an ANN-based prediction of power loss and relative permeability of a range of toroidal cores wound from grain-oriented electrical steel (GOS) and cobalt-amorphous ribbon (CAR).

The ANN software aNETka [8] used in this investigation was developed in LabVIEW 5.1. The measurement software was also developed in LabVIEW, which allowed seamless interfacing between the two packages.

Section snippets

Prediction of P and μ of GOS samples

A series of eight GOS cores of different dimensions, representing a range of industrial interest, was selected. The inner diameter, Din, varied from 34 to 79 mm, the outer diameter, Dout, from 60 to 146 mm and the strip width, W, from 10 to 30 mm (for definition of dimensions, see Fig. 1). The cores were wound from 3% silicon content, 0.27 mm thick, conventional grain-oriented steel sheet (grade M4).

The magnetic properties of the GOS samples were measured in a peak flux density (B) range from 0.3

Prediction of P and μ of CAR samples

A range of 28 cores wound from CAR were chosen, in which Din varied from 14.3 to 25.5 mm, Dout from 21.5 to 37.5 mm, and W from 3.7 to 10.0 mm. The cores were made from 0.03 mm thick cobalt-based amorphous ribbon (VITROVAC 6025). The cores were magnetised from 0.1 to 0.5 T over a frequency range from 20 Hz to 25 kHz, which resulted in 800 data sets available for the ANN training.

The target values of P were pre-processed in a similar way as for the GOS samples. The optimal power coefficients were found

Conclusions

Various ANN structures have been investigated, ranging from very small (4 active neurons) to relatively large (101 active neurons), distributed between 1 and 3 hidden layers. The analysis shows that the number of hidden layers does not affect significantly the prediction accuracy of power loss or permeability, but the number of neurons is critical.

It has been found that for prediction of power loss and relative permeability in toroidal cores made of GOS or cobalt-based amorphous ribbon the

Acknowledgements

This work was supported by EPSRC Grants GR/R82869/01 and EP/C518616/1. The authors would like to thank Wiltan Ltd. for providing samples for this investigation.

References (8)

  • G.K. Miti

    JMMM

    (2003)
  • D. Makaveev

    JMMM

    (2003)
  • D.T. Pham et al.

    Neural Networks for Identification, Prediction and Control

    (1999)
  • Qnet 2000, Neural Network Modeling for Windows 95/98/XP, Internet (August 2007)...
There are more references available in the full text version of this article.

Cited by (5)

  • Determination of the uranium elemental concentration in molten salt fuel using laser-induced breakdown spectroscopy with partial least squares-artificial neural network hybrid models

    2022, Spectrochimica Acta - Part B Atomic Spectroscopy
    Citation Excerpt :

    For the 60 data sets used in the training, with 9 input values and 1 output value, the configuration with 10 neurons is the best and optimum configuration. This result is in contrary to the results obtained by Zurek et al. who reported of improved accuracy with a larger number of neurons [34]. Greater numbers of neurons yield less accurate results, probably due to under fitting with the restriction of 10,000 iterations for the above study.

  • A review of modelling and optimization techniques in turning processes

    2018, International Journal of Mechanical Engineering and Technology
View full text