Fault Diagnosis of Rotating Machinery Using an Optimal Blind Deconvolution Method and Hybrid Invertible Neural Network

Abstract

This paper proposes a novel approach to predicting the useful life of rotating machinery and making fault diagnoses using an optimal blind deconvolution and hybrid invertible neural network. First, a new optimal adaptive maximum second-order cyclostationarity blind deconvolution (OACYCBD) is developed for denoising vibration signals obtained from rotating machinery. This technique is obtained from the optimization of traditional adaptive maximum second-order cyclostationarity blind deconvolution (ACYCBD). To optimize the weights of conventional ACYCBD, the proposed method utilizes a probability density function (PDF) of Monte Carlo to assess fault-related incipient changes in the vibration signal. Cross-entropy is used as a convergence criterion for denoising. Because the denoised signal carries information related to the health of the rotating machinery, a novel health index is calculated in the second step using the peak value and square of the arithmetic mean of the signal. The novel health index can change according to the degradation of the health state of the rotating bearing. To predict the remaining useful life of the bearing in the final step, the health index is used as input for a newly developed hybrid invertible neural network (HINN), which combines an invertible neural network and long short-term memory (LSTM) to forecast trends in bearing degradation. The proposed approach outperforms SVM, CNN, and LSTM methods in predicting the remaining useful life of bearings, showcasing RMSE values of 0.799, 0.593, 0.53, and 0.485, respectively, when applied to a real-world industrial bearing dataset.

The Prediction of the Remaining Useful Life of Rotating Machinery Based on an Adaptive Maximum Second-Order Cyclostationarity Blind Deconvolution and a Convolutional LSTM Autoencoder

“…In the architecture of the ACYCBD, the envelope harmonic product spectrum (EHPS) can detect the cyclic frequency in the vibration signals for the processing of the CYCBD, as in the blind deconvolution theory [ 6 , 7 , 8 ]. The input signal is multiplied with the inverse FIR filter to compute the source signal .…”

“…Data are utilized to verify the proposed performance of the approach (from the Intelligent Maintenance System (IMS) center) [ 6 ]. Four Rexnord ZA-2115 double-row bearings are applied in the experiment with an AC motor working at a constant speed of 2000 revolutions per minute (RPM) and a sampling frequency of 20,000 Hz.…”

“…To resolve this, the maximum correlated kurtosis deconvolution (MCKD) is based on the correlated kurtosis to improve the filtering performance for vibration signals. From the cyclostationary view, the MCKD does not have the best performance for the statistical processing of signals [ 5 , 6 ]. The maximum second-order cyclostationarity blind deconvolution (CYCBD) is proposed to make use of the maximization of the cyclostationary to address the eigenvalue problem, which can improve the recovery of the denoising of the feature signals.…”

A review on convolutional neural network in rolling bearing fault diagnosis

“…Wang et al [49] proposed SVD-1DCNN, which was constructed by embedding an SVD layer after the first convolution or easy to get stuck at locally optimal value. Gao et al [53] proposed a bearing fault diagnosis method based on optimal blind deconvolution and hybrid invertible neural network.…”

An Adaptive Multi-D-Norm-Driven Sparse Unfolding Deconvolutional Network for Bearing Fault Diagnosis

“…In practical engineering applications, it is difficult to calculate or extract periodic information from noisy observation signals due to low-SNR conditions. Therefore, some stage-wise periodic detection methods have been developed and their main aim is to find the maximum peaks of the auto-correlation function of envelope signals [ 17 , 18 , 19 , 20 ].…”