82
X. Wang et al.
calculating the remaining life of bearings was proposed based on the L-P theory.
After continuous revision, ISO proposed a standardized formula for calculating the
life of bearings [1]. Using the formula to predict bearing life results only represented
the average bearing life, and it cannot reflect the extreme bearing life. On the other
hand, Weibull proposed the Weibull model for bearing life prediction, which was still
used in the field of life prediction. The prediction of bearing life using the Weibull
model required a large amount of testing data as support, and the prediction results
will be accurate [2–4]. In the past thirty years, the finite element analysis method
has been applied to the prediction of the RUL of the bearing. By the finite element
analysis, the stress and strain datasets for outer race, inner race and ball are obtained
and used to calculate the fatigue life [5]. In recent years, with the rise of artificial
intelligence, the combination of experiments and artificial intelligence algorithms to
solve bearing life problems has been a popular solution. A novel prediction method
based on MSPSO and MK-LSSVM was proposed for the life prediction of rolling
element bearing by Zhang. The analysis results of actual rolling bearing vibration
data verify the effectiveness of the method in rolling bearing life prediction [6].
The DPNN was used to predict bearing RUL. The experimental results show that
DPNN can predict effectively the RUL of bearing with high prediction accuracy
and strong robustness [7]. An online method of RUL prediction was proposed by
the Kolmogorov–Smirnov test theory and LSTM network [8]. In the above methods,
when the performance degradation index changes abruptly, the accuracy of prediction
results will decrease.
In the actual prediction process, the problems of unpredictability and decreased
prediction accuracy caused by sudden changes in performance degradation indicators, the paper proposed a denoising method of bearing performance index sequence
based on VAE, and the smooth bearing performance index sequence facilitated the
LSTM network to predict the RUL of the bearing. The RUL prediction results based
on the LSTM network were compared with the results using SVR and ANN, which
verified that the accuracy of the RUL prediction results based on the LSTM network
was higher than others. The methods in the paper not only improved the accuracy
of bearing RUL prediction, but also provided reference for solving other parts RUL
prediction problems and the reliability problem of mechanical system.
7.2 Theory
7.2.1 LSTM Network
LSTM network is a recurrent neural network for time series prediction. The structure
diagram of the LSTM unit was shown in Fig. 7.1. The unit includes input gate, output
gate and forget gate [9–11].
The new signal Xt and the previous unit output ht-1 were used as input to the
unit. After the forgetting gate operation, the information of the previous unit was
X. Wang et al.
calculating the remaining life of bearings was proposed based on the L-P theory.
After continuous revision, ISO proposed a standardized formula for calculating the
life of bearings [1]. Using the formula to predict bearing life results only represented
the average bearing life, and it cannot reflect the extreme bearing life. On the other
hand, Weibull proposed the Weibull model for bearing life prediction, which was still
used in the field of life prediction. The prediction of bearing life using the Weibull
model required a large amount of testing data as support, and the prediction results
will be accurate [2–4]. In the past thirty years, the finite element analysis method
has been applied to the prediction of the RUL of the bearing. By the finite element
analysis, the stress and strain datasets for outer race, inner race and ball are obtained
and used to calculate the fatigue life [5]. In recent years, with the rise of artificial
intelligence, the combination of experiments and artificial intelligence algorithms to
solve bearing life problems has been a popular solution. A novel prediction method
based on MSPSO and MK-LSSVM was proposed for the life prediction of rolling
element bearing by Zhang. The analysis results of actual rolling bearing vibration
data verify the effectiveness of the method in rolling bearing life prediction [6].
The DPNN was used to predict bearing RUL. The experimental results show that
DPNN can predict effectively the RUL of bearing with high prediction accuracy
and strong robustness [7]. An online method of RUL prediction was proposed by
the Kolmogorov–Smirnov test theory and LSTM network [8]. In the above methods,
when the performance degradation index changes abruptly, the accuracy of prediction
results will decrease.
In the actual prediction process, the problems of unpredictability and decreased
prediction accuracy caused by sudden changes in performance degradation indicators, the paper proposed a denoising method of bearing performance index sequence
based on VAE, and the smooth bearing performance index sequence facilitated the
LSTM network to predict the RUL of the bearing. The RUL prediction results based
on the LSTM network were compared with the results using SVR and ANN, which
verified that the accuracy of the RUL prediction results based on the LSTM network
was higher than others. The methods in the paper not only improved the accuracy
of bearing RUL prediction, but also provided reference for solving other parts RUL
prediction problems and the reliability problem of mechanical system.
7.2 Theory
7.2.1 LSTM Network
LSTM network is a recurrent neural network for time series prediction. The structure
diagram of the LSTM unit was shown in Fig. 7.1. The unit includes input gate, output
gate and forget gate [9–11].
The new signal Xt and the previous unit output ht-1 were used as input to the
unit. After the forgetting gate operation, the information of the previous unit was