7 Prediction of Bearing Remaining Useful Life Based on LSTM Network
85
Table 7.1 Sampling points at which bearings began to fail and the corresponding RUL
Conditions
No
Sample point
RUL (min)
Conditions 1
Bearing1_1
94
29
Bearing1_2
129
32
Bearing1_3
128
25
Bearing1_5
40
12
Conditions 2
Bearing2_1
467
24
Bearing2_3
500
21
Bearing2_4
32
10
Conditions 3
Bearing3_1
2519
19
Bearing3_3
345
8
Bearing3_5
95
19
Table 7.2 The training set and the testing set
Training set
Testing set
Bearing1_1, Bearing1_2, Bearing1_3, Bearing2_1, Bearing2_3,
Bearing2_4, Bearing3_3, Bearing3_5
Bearing1_5, Bearing3_1
not need to mark the training data variables. The K-Means algorithm divided all
sampling points into 4 categories. The first sampling point of the third category (the
expiration period) was used as the initial sampling point of each bearing failure
period. The results were shown in Table 7.1. In the paper, the starting sampling point
of each bearing failure period was taken as the starting point of prediction.
There were 209 sampling points in the dataset. The dataset was divided into
training set and testing set, including 176 sampling points in the training set and 33
sampling points in the testing set. The details were shown in Table 7.2.
7.4 Methods
7.4.1 The RUL Prediction Using LSTM Network Directly
The bearing RUL prediction method based on LSTM network was that the bearing
performance degradation index was input to the LSTM network to complete the
prediction of bearing RUL. The specific process was as follows:
a. The training set and the testing set were constructed and normalized. In the
experiment data, the training set and the testing set were normalized.
b. The network was trained. The network-related parameters were set, the ‘tanh’
function was chosen as the activation function, the ‘mse’ function was chosen as