Predict Jet Engine Thrust from Sensor Data with Random Forest

Download the C-MAPSS dataset from the NASA Prognostics Data Repository. Start with the FD001 subset — it contains data from 100 engines operating under a single flight condition with one fault mode (HPC degradation). Each engine runs from healthy to failure, with 21 sensor channels recorded at each cycle.

Read the dataset documentation carefully. The 21 sensors include: fan inlet temperature (T2), LPC outlet temperature (T24), HPC outlet temperature (T30), LPT outlet temperature (T50), fan speed (Nf), core speed (Nc), and several pressure measurements. Three operational setting columns define the flight condition. Understand what each sensor measures and where it sits in the gas path.

Pick 3–4 engines and plot each sensor reading vs. operating cycle. Some sensors trend clearly as the engine degrades (T30 and T50 increase, efficiency-related measurements decrease). Others are nearly flat and carry little information. Use correlation analysis to identify which sensors are most correlated with each other and with the operating settings.

Since C-MAPSS doesn't provide explicit thrust values, derive a thrust proxy: the fan speed corrected for inlet conditions (Nf_corrected = Nf / sqrt(T2/288.15)) is commonly used, or you can use the provided "fan efficiency" proxy. This corrected parameter is your regression target.

Select the sensor columns and operational settings as your features. Drop any constant or near-constant sensors (some channels in FD001 have zero variance because the operating condition is fixed). Normalize all features using StandardScaler.

Add rolling window features — for each sensor, compute the mean and standard deviation over the previous 5 cycles. These rolling statistics capture short-term trends and often improve model accuracy because they encode the rate of change, not just the instantaneous value.

Split the data by engine ID (not randomly!) — use engines 1–80 for training and 81–100 for testing. This prevents data leakage from the same engine appearing in both sets. Train a RandomForestRegressor with 200 trees:

from sklearn.ensemble import RandomForestRegressor
rf = RandomForestRegressor(n_estimators=200, max_depth=15, random_state=42)
rf.fit(X_train, y_train)

Use 5-fold cross-validation on the training set to tune max_depth (try 10, 15, 20, None) and min_samples_leaf (try 1, 5, 10). Track RMSE and R² at each fold.

Extract and plot the feature importances from the trained Random Forest. The top features should include core speed (Nc), HPC outlet temperature (T30), and fan speed (Nf) — these are the thermodynamic quantities most directly linked to thrust through the Brayton cycle.

Run a permutation importance analysis (scikit-learn's permutation_importance) to cross-check. Permutation importance is model-agnostic and avoids the bias that tree-based importance has toward high-cardinality features. Do the rankings agree? Discuss any differences in your report.

Divide the test data into early-life (first 30% of cycles), mid-life, and late-life segments. Evaluate model accuracy separately on each segment. Does the model perform worse on degraded engines? If so, this suggests the relationship between sensor readings and thrust changes as the engine degrades — a key insight for engine health management.

Plot predicted vs. actual for the full test set, color-coded by engine life stage. A well-performing model will show tight clustering around the diagonal regardless of degradation level.

Write a technical report covering the dataset, feature engineering decisions, model selection and tuning, and results. Include the feature importance plot and the degradation-stage analysis. Discuss how this approach could be deployed in a real engine health monitoring system — what would be needed to go from this offline analysis to a real-time thrust estimation model running on an aircraft?

Compare your ML approach to the physics-based alternative: a full engine performance model (like GasTurb or NPSS) that computes thrust from first principles. When would you use each approach?