In order to recognize on-line different kind of defects where partial discharges take place it is necessary to employ an embedded acquisition system, possibly low-cost, near to the discharges site. Such acquisition system should be capable to analyze the acquired data in real time to recognize different defect sources. For this reason, a single hidden-layer feedforward neural network (SLFN) is presented in this work. Such neural network inherits the approach adopted by the Extreme Learning Machine (ELM), but its activation functions are based on hardware friendly step (hard-limit) function. This facilitates the hardware resources optimization. In order to show the effectiveness of this solution different kind of PDs signals have been acquired. Once the data have been available, they have been divided in two sets: training and test. The training data have been used off-line to construct the model. Therefore the test data have been used on-line to perform the neural network. Moreover a comparison, in term of generalization error, between the classic ELM (based on complex activation functions, e.g. sigmoid), the Support Vector Machine (SVM) and the digital realization has been analyzed and showed. The results demonstrate how a simple algorithm as a single layer feedforward neural network is able to distinguish different kind of PDs using limited resources. Eventually the occupation area is shown on different low-cost devices varying the number of neurons in the hidden layer. The new ELM digital implementation is able to achieve good results, in term of generalization error when involved in supervised problems recognizing different kind of PD activity. These results are comparable to the ones obtainable by mean of standard ELM and SVM.

Hardware Friendly Neural Network for the PD Classification

Gianoglio, C.;Guastavino, F.;Ragusa, E.;Bruzzone, A.;Torello, E.
2018-01-01

Abstract

In order to recognize on-line different kind of defects where partial discharges take place it is necessary to employ an embedded acquisition system, possibly low-cost, near to the discharges site. Such acquisition system should be capable to analyze the acquired data in real time to recognize different defect sources. For this reason, a single hidden-layer feedforward neural network (SLFN) is presented in this work. Such neural network inherits the approach adopted by the Extreme Learning Machine (ELM), but its activation functions are based on hardware friendly step (hard-limit) function. This facilitates the hardware resources optimization. In order to show the effectiveness of this solution different kind of PDs signals have been acquired. Once the data have been available, they have been divided in two sets: training and test. The training data have been used off-line to construct the model. Therefore the test data have been used on-line to perform the neural network. Moreover a comparison, in term of generalization error, between the classic ELM (based on complex activation functions, e.g. sigmoid), the Support Vector Machine (SVM) and the digital realization has been analyzed and showed. The results demonstrate how a simple algorithm as a single layer feedforward neural network is able to distinguish different kind of PDs using limited resources. Eventually the occupation area is shown on different low-cost devices varying the number of neurons in the hidden layer. The new ELM digital implementation is able to achieve good results, in term of generalization error when involved in supervised problems recognizing different kind of PD activity. These results are comparable to the ones obtainable by mean of standard ELM and SVM.
2018
978-1-5386-6192-5
File in questo prodotto:
File Dimensione Formato  
Hardware Friendly Neural Network for the PD Classification.pdf

accesso chiuso

Tipologia: Documento in versione editoriale
Dimensione 400.54 kB
Formato Adobe PDF
400.54 kB Adobe PDF   Visualizza/Apri   Richiedi una copia

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11567/948147
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 5
  • ???jsp.display-item.citation.isi??? ND
social impact