An image dimensionality reduction method for rolling bearing fault diagnosis based on singular value decomposition

Abstract

The fast kurtogram, a faint signal extraction method, has been regarded as an effective approach to detect and characterize faint transient features in vibration signals. However, the fast kurtogram, a band-pass filtering method, which extracts transient signals by optimal frequency band selection and leaves the noise in the selected frequency band unprocessed. Therefore, to overcome the shortcoming of the fast kurtogram method, a method which can wipe off the noise in the whole frequency band is necessary. This paper proposes a novel faint signal extraction method by timefrequency distribution image dimensionality reduction. Since time-frequency distribution image can reveal intrinsic feature of nonstationary signals and can make the weak impulses feature prominent, and besides, the transient impulse feature and the noise component lie in different dimensions, so using the dimensionality reduction method based on singular value decomposition to suppress the background noise in the raw time-frequency distribution image is motivated. A bearing outer race fault signal obtained from a test-to-failure experiment and a bearing inner race fault signal obtained from an experimental motor are employed to demonstrate the enhanced performance of the proposed method in faint signal extraction. The results indicate that the proposed method outperforms the fast kurtogram method and is effective in faint signal extraction.

A Review on Rolling Bearing Fault Signal Detection Methods Based on Different Sensors

“…The SPWVD method had the highest classification accuracy and sensitivity (Wang, Yaping et al, 2017) [131]. A singularvalue decomposition-dimension-reduction method was used to suppress background noise in original time-frequency images and diagnose bearing faults (Wang, Y et al, 2016) [132]. Support vector machines were used to identify smooth pseudo-Wigner-Ville distribution (SPWVD) feature vectors obtained by position-constrained linear coding (LLC) and the spatial pyramid method to diagnose bearing faults (Wei Gang Wang et al, 2014) [133].…”

Bearing Fault Diagnosis Based on Subband Time-Frequency Texture Tensor

“…Because of the non-stationary nature of the faulty bearing's vibration, the conventional time or frequency domain features based on statistical analysis have difficulty representing the fault states of the bearing, especially in a noisy environment. Texture feature can steadily describe the spatial changes in color, gradation or fine structure, and shape of the image with favorable rotation invariance and anti-interference capability [27], [28].…”

Rolling Bearing Incipient Degradation Monitoring and Performance Assessment Based on Signal Component Tracking

“…Thus one-dimensional and two-dimensional improved convolution neural networks are established based on the spectrum analysis and timefrequency analysis. The key work in conventional CNN is the calculation of weights between layers, which are trained by standard back-propagation procedures and gradient descent algorithm according to the loss function commonly using mean squared-error [12]. In this paper, two-layer architecture of CNN is established, whose forward mapping process is reserved and training phase is omitted.…”

“…Characteristic frequency of bearing incipient fault can be easily interrupted by noise or other masking sources in low-frequency band and meanwhile is found in a widely high-frequency band due to the modulation phenomenon [8]. Envelope analysis based on selection of optimal demodulation band has become an effective approach to recognize weak fault characteristic frequency, and thus different decomposition methods combined with component selection strategies were proposed, such as kurtogram and Protrugram [9], [10], minimum entropy de-convolution [11], singular value decomposition [12], ensemble empirical mode decomposition [13], [14], wavelet decomposition [15], intrinsic characteristic-scale decomposition [16] and local mean decomposition [17]. Moreover, multi-scale morphology filter [18] was also used lately for signal component extraction.…”