Abstract
Rotor systems have been extensively used in a variety of industrial applications. However an unexpected failure may cause a break down of the rotational machinery, resulting in production and significant economic losses. Efficient incipient fault diagnosis is therefore critical to the machinery normal operation. Noise and vibration analysis is popular and effective for the rotor fault diagnosis. One of the key procedures in the fault diagnosis is feature extraction and selection. Literature review indicates that only limited research considered the nonlinear property of the feature space by the use of manifold learning algorithms in the field of mechanical fault diagnosis, and nonlinear feature extraction for rotor multi-fault detection has not been studied. This paper reports a new development based on a novel supervised manifold learning algorithm (adaptive locally linear embedding) applied to nonlinear feature extraction for rotor multiple defects identification. The adaptive locally linear embedding (ALLE) combines with the adaptive nearest neighbour algorithm and supervised locally linear embedding (LLE) to provide an adaptive supervised learning. Hence, distinct nonlinear features could be extracted from high-dimensional dataset effectively. Based on ALLE, a new fault diagnosis approach has been proposed. The independent component analysis (ICA) was firstly employed to separate the faulty components of the rotor vibration from the observation data. Then wavelet transform (WT) was used to decompose the recovered signals, and statistical features of frequency bands were hence calculated. Lastly, ALLE was applied to learn the low-dimensional intrinsic structure of the original feature space. The experiments on vibration data of single and coupled rotor faults have demonstrated that sensitive fault features can be extracted efficiently after the ICA-WT-ALLE processing, and the proposed diagnostic system is effective for the multi-fault identification of the rotor system. Furthermore, the proposed method achieves higher performance in terms of the classification rate than other feature extraction methods such as principal component analysis (PCA) and locally linear embedding (LLE).