e-RULENet: remaining useful life estimation with end-to-end learning from long run-to-failure data

Figures & data

Figure 1. Architecture of the Dual-estimator with multiple segment input. This example takes three segments as the reference measurements.

Figure 2. Leaky truncated RUL function.

Table 1. Summary on C-MAPSS dataset.

Property \ \Dataset	FD001	FD002	FD003	FD004
Train trajectories	100	260	100	248
Test trajectories	100	259	100	249
Operational modes	1	6	1	6
Fault modes	1	1	2	2
Min. RUL at test	7	6	6	6
Max. RUL at test	145	194	145	195
Min. length at test	31	21	38	19

Table 2. RMSE comparison on C-MAPSS dataset. The best performance is shown in bold.

Method \Dataset	FD001	FD002	FD003	FD004
LSTM	15.8	16.5	16.1	17.9
e-RULENet	15.4	19.7	15.5	20.8

Table 3. Score comparison on C-MAPSS dataset. The best performance is shown in bold.

Method \ \Dataset	FD001	FD002	FD003	FD004
LSTM	486	2604	867	2416
e-RULENet	303	1330	509	1554

Figure 3. Examples of trajectories of estimated RUL by the e-RULENet.

Table 4. RMSE comparison on HDD dataset. The best performance is shown in bold.

Method	RMSE
LSTM	0.089
e-RULENet	0.081

Table 5. RMSE comparison on the milling dataset. The best performance is shown in bold.

Method	RMSE
LSTM	5.45
e-RULENet	4.33

Figure 4. Normalized belt tension estimated by the e-RULENet.