Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
388538 | Expert Systems with Applications | 2011 | 9 Pages |
Indicator diagram plays an important role in the health monitoring and fault diagnosis of reciprocating compressors. Different shapes of indicator diagram indicate different faults of reciprocating compressor. A proper feature extraction and pattern recognition method for indicator diagram is significant for practical uses. In this paper, a novel approach is presented to handle the multi-class indicator diagrams recognition and novelty detection problems. When multi-class faults samples are available, this approach implements multi-class fault recognition; otherwise, the novelty detection is implemented. In this approach, the discrete 2D-Curvelet transform is adopted to extract the representative features of indicator diagram, nonlinear PCA is employed for multi-class recognition to reduce dimensionality, and PCA is used for novelty detection. Finally, multi-class and one-class support vector machines (SVMs) are used as the classifier and novelty detector respectively. Experimental results showed that the performance of the proposed approach is better than the traditional wavelet-based approach.
► A new recognition and novelty detection approach for indicator diagram of reciprocating compressor is proposed. ► This approach is a combination of Curvelet transform, nonlinear PCA and SVM. ► More accurate results than traditional wavelet-based approach were obtained. ► Indicator diagram diagnosis can be automatically done by our approach.