MODEL OPTIMIZATION

NASA’s C-MAPSS (dataset to assess the engine’s lifetime and engine degradation).

Optimization of the out-of-the-box model from the first regression task to improve prediction of RUL (Remaining Useful Life).

DENSITY OF UNITS OVER TIME CYCLE

We want also to understand the density of units over time cycle which affecting our prediction during the training of our models. The plots below are for the training set and test set. NOTE that the final RUL of the test set is the censored time cycle + RUL (given). Therefore, we have to measure our prediction metrics from these values for the event of failure and NOT from the censored time (based on the survival function). Below are encircled with area of interest where prediction with time-varying effects that need to be modeled explicitly.

TEST/TRANING SET

USING ENSEMBLE KALMAN FILTER AT PROCESS NOISE OF 0.0005 AND OBSERVATION NOISE OF 0.01 AT 500 ENSEMBLES OF HPC OUTLET PRESSURE PSIA
RMSE: 0.665901260430472

MAE: 0.535468065659605

R²: 0.36242445284651403

PARTICLE FILTER

RMSE: 0.7621426750670304

MAE: 0.6098987551872036

R²: 0.1648114372268109

STANDARD KALMAN FILTER STATE ESTIMATION

RMSE: 0.5325445075171134

MAE: 0.42798025301965287

R²: 0.59222199373269

HYPERPARAMETER SEARCH

Instead of using LSTM (also tested), we used GRU(Gated Recurrent Unit) for Temporal Fusion Transformer to train our models to search for hidden size, number of layers, number of attention heads, learning rate, weight decay and number of epochs. This is for the single model where FD001 was divided into quantile with slight treatment of outliers since our data is sensitive to smoothing.

at Fold 5

Test Loss (MSE): 1635.7650

Mean Absolute Error: 34.3026

R² Score: -0.0095

Mean Absolute Percentage Error: 126.0648%

Root Mean Squared Error: 40.4446

Mean of Residuals: -3.9302

Variance of Residuals: 1620.3184