Improving one class support vector machine novelty detection scheme using nonlinear features

Novelty detection scheme in bearing vibration signals of rotating system is investigated in this article. One class support vector machine (OC-SVM) is used for novelty detection. It focuses on the required preprocessing steps including denoising, feature extraction, vectorization, normalization and dimensionality reduction, and a systematic method is proposed for each of them. A new scheme is used for denoising which presents the best combination of mother wavelet and thresholding rule for each signal. The required features are extracted from time and time-frequency domains, and the best mother wavelet for extracting features from each signal, is presented by means of energy-to-Shannon entropy ratio criterion. It is shown that the load factor is the most important factor to impose nonlinearity in vibration signals of bearings compared to fault type and fault intensity factors. Besides, for the first time, this paper demonstrates that by increasing the nonlinearity in the signals, the statistical traditional or combinations of statistical traditional and nonlinear features fail to classify data completely and only the nonlinear features have this capability. The proposed systematic preprocessing improves the efficiency of OC-SVM novelty detection upto 100% in some cases, when applied to three different data sets. Also, it yields a satisfying result compared to other similar works, in the field of classification.