US 12,222,259 B1
Rolling bearing fault diagnosis method based on fast fourier transform coding and lightweight convolutional neural network
Mei Liu, Maoming (CN); Kun Cui, Maoming (CN); Huizi Han, Maoming (CN); and Shijie Liu, Maoming (CN)
Assigned to GUANGDONG UNIVERSITY OF PETROCHEMICAL TECHNOLOGY, Maoming (CN)
Filed by Guangdong University of Petrochemical Technology, Maoming (CN)
Filed on Sep. 27, 2024, as Appl. No. 18/898,744.
Claims priority of application No. 202311494001.1 (CN), filed on Nov. 10, 2023.
Int. Cl. G01M 13/045 (2019.01); G06N 3/0464 (2023.01)
CPC G01M 13/045 (2013.01) [G06N 3/0464 (2023.01)] 6 Claims
OG exemplary drawing
 
1. A rolling bearing fault diagnosis method based on FFT coding and L-CNN, comprising:
obtaining original bearing fault vibration data, extracting intrinsic mode components of different frequency bands in the original bearing fault vibration data, calculating a permutation entropy value corresponding to each of the intrinsic mode components, and performing wavelet threshold denoising according to the permutation entropy value to obtain a denoised reconstructed time domain signal;
performing fast Fourier transform to the denoised reconstructed time domain signal to obtain a frequency domain signal and a phase angle corresponding to the time domain signal, reconstructing the frequency domain signal according to a preset rule, retaining frequency domain data of features of the phase angle, and drawing FFT-x heat maps of different fault type data according to an amplitude range; wherein
the drawing FFT-x heat maps of different fault type data comprises:
decomposing the denoised reconstructed time domain signal into several single harmonic components by fast Fourier transform, and obtaining a relationship between amplitude, phase, power and frequency domain of each harmonic of the signal; and
when the phase angle is in a first quadrant and a second quadrant, a frequency domain value is positive, and when the phase angle is in a third quadrant and a fourth quadrant, a frequency domain value is negative; and according to this characteristic, reconstructing the frequency domain signal, and according to an amplitude characteristic of a reconstructed frequency domain signal, setting a boundary to (−150, 150) for heat map coding; and
constructing an improved lightweight convolutional neural network model L-CNN, and inputting coded data in the FFT-x heat maps into the L-CNN model for processing and diagnosis, and obtaining fault diagnosis results;
the inputting coded data in the FFT-x heat maps into the L-CNN model for processing and diagnosis, comprising:
capturing different levels of features of the coded data in the FFT-x from different scales by using convolution kernels with different sizes, and weighting captured features for importance of each channel by structures of global average pooling, one-dimensional convolution, channel multiplication and spatial replication through an ECA attention mechanism; and
processing a channel dimension and a spatial dimension of a color feature of an input FFT-x coded map by depth separable convolution to obtain a feature map; exchanging a channel sequence of the feature map between different depths and different groups by using a ChannelSplit module and a Channelshuffle module, and changing a channel arrangement mode; further extracting data features based on the residual and depth separable convolution; and finally, outputting the fault diagnosis results through average pooling and two fully connected layers.