Abstract
Feature extraction plays a key role in fault diagnosis of rotating machinery. Many methods reported in the literature are based on masses of labeled data and need much prior knowledge to select the most discriminating features or establish a complex deep-learning model. To solve the dilemma, a novel feature extraction method based on kernel principal component analysis (KPCA) and an autoencoder (AE), namely, SFS-KPCA-AE, is presented in this paper to automatically extract the most discriminative features from the frequency spectrum of vibration signals. First, fast Fourier transform is calculated on the entire vibration signal to get the frequency spectrum. Next, the spectrum is divided into several segments. Then, local-global feature extraction is performed by applying KPCA to these segments. Finally, an AE is employed to obtain the low-dimensional representations of the high-dimensional global feature. The proposed feature extraction method combined with a classifier achieves fault diagnosis for rotating machinery. A rotor dataset and a bearing dataset are utilized to validate the performance of the proposed method. Experimental results demonstrate that the proposed method achieved satisfactory performance in feature extraction when training samples or motor load changed. By comparing with other methods, the superiority of the proposed SFS-KPCA-AE is verified.
Highlights
Rotating machinery is widely used in the modern industry
Two principal components considered as local features of the spectrum are extracted from each spectrum segment using kernel principal component analysis (KPCA). en, these local features are connected in sequence to obtain the global feature vector
fast Fourier transform (FFT) is first used to acquire the frequency spectrum of vibration signal, later the frequency spectrum is divided into several segments
Summary
Rotating machinery is widely used in the modern industry. after long operating time under harsh working conditions, it will inevitably break down and even result in catastrophic accidents and huge economic losses [1, 2]. There is a lot of information in the frequency spectrum of vibration signal that is very critical for fault diagnosis, and the two unsupervised feature learning methods mentioned above can be used to extract the important information. E redundant information in the frequency spectrum of vibration signals, such as those frequency components with the same amplitude in different samples, can be removed by using KPCA as a data preprocessing method. In order to take advantage of the aforementioned unsupervised feature learning techniques and extract key information that can distinguish the state of the device from the frequency spectrum of vibration signals of rotating machinery, an unsupervised feature learning method based on the KPCA and AE is proposed. The superiority of the proposed method was validated by designing comparative experiments
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
