Feature Selection: 17 Informative Features

Learning Objectives

By the end of this section, you will:

1. Understand the actual feature selection used in our research implementation
2. Identify the 14 selected sensors and why each was chosen
3. Know which sensors are excluded and the reasoning behind removal
4. Understand the 17-feature convention (3 settings + 14 sensors) used in literature
5. Apply the enforce_feature_set parameter for consistent benchmarking

Why This Matters: The C-MAPSS dataset has 21 sensors, but not all carry useful information. Our research implementation uses exactly 17 features—3 operating settings plus 14 carefully selected sensors. This section shows the actual code that selects these features and explains the selection criteria.

Why Feature Selection?

With deep learning, it's tempting to feed all available data into the model. However, the literature has converged on a specific 17-feature subset for C-MAPSS benchmarking:

Benefits of Curated Feature Selection

The Selection Criteria

For each of the 21 sensors, we ask:

1. Does it vary? Constant sensors (std < 0.01) provide no degradation information
2. Does it correlate with RUL? Variation should reflect degradation, not just noise
3. Is it physically meaningful? Domain knowledge validates the selection

Research Implementation

This is the actual feature selection code from our EnhancedNASACMAPSSDataset class. It defines exactly which sensors are used as model inputs.

setting_1: Altitude (0-42,000 ft)

T30 (sensor_7): Increases ~10°R over engine life

1# Define feature columns based on literature-recommended sensors
2self.setting_cols = ['setting_1', 'setting_2', 'setting_3']
3
4# 14 informative sensors selected based on literature analysis
5# These sensors show meaningful variance and degradation correlation
6sensor_keep = [
7    'sensor_7',   # Total temperature at HPC outlet
8    'sensor_8',   # Total temperature at LPT outlet
9    'sensor_9',   # Pressure at HPC outlet
10    'sensor_12',  # Pressure ratio (EPR)
11    'sensor_16',  # Bypass ratio (BPR)
12    'sensor_17',  # Bleed enthalpy
13    'sensor_20',  # HPT coolant bleed
14    'sensor_2',   # Total temperature at LPC outlet
15    'sensor_3',   # Total temperature at HPC outlet
16    'sensor_4',   # Total temperature at LPT outlet
17    'sensor_14',  # Corrected core speed
18    'sensor_11',  # Static pressure at HPC outlet
19    'sensor_13',  # Corrected fan speed
20    'sensor_6',   # Total pressure at HPC outlet
21]
22
23if self.enforce_feature_set:
24    # 17 features total: 3 settings + 14 sensors
25    self.feature_columns = self.setting_cols + sensor_keep
26else:
27    # Use all 24 features (3 settings + 21 sensors)
28    self.feature_columns = self.setting_cols + [f'sensor_{i}' for i in range(1, 22)]
29
30logger.info(f"Dataset features: {len(self.feature_columns)} features configured")
31logger.info(f"Using {'enforced 17-feature set' if self.enforce_feature_set else 'all features'}")

Feature Column Mapping

Constant Sensors (Removed)

Seven sensors are excluded because they show near-zero variance across all engines and cycles:

Excluded Sensors

Informative Sensors (Kept)

The 14 selected sensors show meaningful variation that correlates with engine degradation:

Temperature Sensors

The physics is straightforward:

Pressure and Speed Sensors

Correlation with RUL

Final 17 Features

After selection, our input data consists of 17 features in a specific order:

Feature Order (as implemented)

Input Tensor Shape

After feature selection and windowing:

• $N$: Number of sliding window samples
• $30$: Window size (timesteps)
• $17$: Selected features

Summary

In this section, we examined the actual feature selection from our research code:

1. 17 features total: 3 operating settings + 14 selected sensors
2. Selection criteria: Variance, RUL correlation, physical meaning
3. Excluded sensors: 7 sensors with near-zero variance (T2, P2, epr, etc.)
4. Key informative sensors: Temperature (T30, T50), pressure (P30), speeds (NRc, NRf)
5. enforce_feature_set: Parameter to ensure consistent benchmarking

Looking Ahead: With features selected, we need to define our prediction targets. The raw RUL is problematic—early-life engines all have high RUL with imperceptible degradation. Next, we introduce the piecewise linear RUL model that clips RUL at 125 cycles.

With informative features identified, we are ready to formulate the target variable for our regression task.