Icon

RUL_​LAB_​Part3_​CMAPSS_​TrainedRUL

RUL LAB Part 3 - when a trained RUL model really improves (NASA C-MAPSS FD001, metanode version). Double-click a metanode to open it.
1 Prepare dataNASA C-MAPSS FD001: 20,631 rows, 100 turbofan engines, one row per flight cycle, every engine run to failure.Drop the 10 constant columns -> 14 sensors.rul_cap = min(rul, 125): early life is not predictable, so the target is capped.Look first (Row Filter -> Line Plot): engine 1, sensor s11 rises slowly until failure.
2 Model A: raw snapshotSplit by ENGINE, not by row: units 1-80 train, units 81-100 test. The 20 test engines are never seen in training, so nothing leaks.Random Forest, 200 trees, seed 0; features = the 14 sensor values of ONE cycle.Mean baseline (predict the training mean): RMSE 41.5 cycles.Model A: RMSE 18.96, R2 0.79.
3 + 4 Rolling features, then model BLoop over one engine at a time; inside each engine take the mean and std of every sensor over the last 10 cycles (28 new columns). The window never crosses two engines.Same split, same forest; features = cycle + 14 raw + 28 rolling = 43.Model B: RMSE 18.08, R2 0.81.
5 + 6 Health index, then model CEach engine's first 20 cycles are its healthy start (median of the 14 sensors). HI_mean = mean relative deviation |x - median| / |median| of the current cycle from that start; HI_m10 = its mean over the last 10 cycles.Same split, same forest; features = 43 + 2 = 45.Model C: RMSE 12.67, R2 0.91; rul <= 50: RMSE 16.32 -> 7.04.Not a leak: each engine is centred on its own past, and the gain is that centring. FD001 is friendlier than a real fleet (one operating condition, one fault mode, every run starts healthy).
Where the gain really isA -> B: overall RMSE moves only 4.6 % (18.96 -> 18.08); near failure, rows with rul <= 50 (n = 1,020): RMSE 19.35 -> 16.32 (-16 %), MAE 12.03 -> 10.45. The gain sits exactly where a maintenance decision is made.B -> C: two health-index columns do more than 28 rolling columns: RMSE 18.08 -> 12.67, rul <= 50: 16.32 -> 7.04.Numeric Scorer is a view node: the numbers are recomputed from out/pred_raw.csv, out/pred_fe.csv and out/pred_hi.csv (4,493 test rows each).
Read cmapss_fd001_train.csv20,631 rows, 100 engines
CSV Reader
Model A: RMSE 18.96, R2 0.79mean baseline: RMSE 41.5
Numeric Scorer
Model C: RMSE 12.67, R2 0.91rul <= 50: 16.32 -> 7.04
Numeric Scorer
out/pred_hi.csv4,493 test rows, 51 columns
CSV Writer
out/pred_raw.csv4,493 test rows
CSV Writer
1 Prepare data
3 Rolling features per engine
2 Model A: raw snapshot
Keep engine 1 only(192 cycles)
Row Filter
5 Health index per engine
s11 rises slowly until failure:the trend a model can learn
Line Plot
4 Model B: rolling features
Model B: RMSE 18.08, R2 0.81rul <= 50: 19.35 -> 16.32 (-16 %)
Numeric Scorer
Statistics
out/pred_fe.csv4,493 test rows, 49 columns
CSV Writer
6 Model C: + health index

Nodes

Extensions

Links

Download

To use this workflow in KNIME, download it from the below URL and open it in KNIME:

Download Workflow
Created by: anos
Created at: 2026-09-25
On NodePit since: 2026-09-25
Last update: 2026-09-27
Created with KNIME version: v5.12.0

Deploy, schedule, execute, and monitor your KNIME workflows locally, in the cloud or on-premises – with our brand new NodePit Runner.

Try NodePit Runner!