Nested SVDD in DAG SVM for induction motor condition monitoring
| Autor: | Slaheddine Zgarni, Ahmed Braham, Hassen Keskes |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
0209 industrial biotechnology
Training set business.industry Computer science Condition monitoring Pattern recognition 02 engineering and technology Support vector machine Multiclass classification ComputingMethodologies_PATTERNRECOGNITION 020901 industrial engineering & automation Hyperplane Artificial Intelligence Control and Systems Engineering Margin (machine learning) Outlier 0202 electrical engineering electronic engineering information engineering Decision boundary Embedding 020201 artificial intelligence & image processing Artificial intelligence Electrical and Electronic Engineering business Induction motor Optimal decision |
| Zdroj: | Engineering Applications of Artificial Intelligence. 71:210-215 |
| ISSN: | 0952-1976 |
| DOI: | 10.1016/j.engappai.2018.02.019 |
| Popis: | Nowadays, multiclass classification is considered as the leading technique on the issue of condition monitoring in induction motor which can be performed accurately and efficiently by support vectors machines. The standard multiclass SVM (MSVM) approaches consists in constructing an optimal decision hyperplane maximizing the margin of separation between the training data. However, relatively small number of outliers can obviously reduce the performance of the classical MSVMs and would have an impact on the decision boundary and the margin calculation. Support Vector Data Description (SVDD) based on hyper-spheres decision boundary is often performed to overcome this drawback. The originality of this paper is to introduce a new extension of the MSVM for broken rotor bar fault diagnosis by embedding the SVDD in the classical multiclass SVM. The experimental results prove the merit of the proposed approach with a classification rate of 100%, which is higher than any accuracy rate achieved by standard MSVM approaches. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full Text from ScienceDirect