The 21 Sensors and 3 Operational Settings

A Tour of a Turbofan

A modern high-bypass turbofan moves air through five rotating stages: a fan that pushes most of the airflow around the core, a low-pressure compressor (LPC), a high-pressure compressor (HPC), a combustor, then a high-pressure turbine (HPT) and low-pressure turbine (LPT) that extract energy. Sensors are placed at standard gas-path stations — T2 at the fan inlet, T24 at LPC outlet, T30 at HPC outlet, T50 at LPT outlet, and so on — plus rotational-speed and bleed-flow probes.

The C-MAPSS simulator outputs 21 such gas-path readings on every cycle, alongside 3 operational settings that describe what regime the engine is currently in. Together they form the 24-column feature vector (plus engine_id and cycle = 26 columns total) that every paper in this space consumes.

Three Operational Settings

On FD001 / FD003 these three columns are essentially constant - the engine is run at sea-level idle. On FD002 / FD004 they jump between six discrete regimes (Section 2.2). For both cases, the op_set columns are NOT modelling features — they describe the current operating point, which downstream code uses to normalise sensor readings within their regime (Chapter 6).

Twenty-One Sensors, Catalogued

Why Each Sensor Drifts

Physical degradation in a turbofan shows up in predictable directions on each sensor:

Knowing the direction matters when reading attention maps later — the model spends more attention on sensors that drift consistently with the failure mode.

Python: Range and Variance Per Sensor

Quick statistical pass that confirms which sensors carry signal. Anything with std < 1e-6 is constant; the top-5 high-variance sensors will dominate downstream modelling.

1import numpy as np
2import pandas as pd
3
4COLUMNS = (
5    ["engine_id", "cycle"]
6    + [f"op_set_{i}" for i in range(1, 4)]
7    + [f"sensor_{i}" for i in range(1, 22)]
8)
9SENSOR_COLS = [f"sensor_{i}" for i in range(1, 22)]
10
11
12df = pd.read_csv("data/raw/train_FD001.txt", sep=r"\s+", header=None, names=COLUMNS)
13
14
15# ----- Per-sensor diagnostics -----
16mean   = df[SENSOR_COLS].mean()
17std    = df[SENSOR_COLS].std()
18range_ = df[SENSOR_COLS].max() - df[SENSOR_COLS].min()
19
20stats = pd.DataFrame({"mean": mean, "std": std, "range": range_}).round(3)
21print(stats)
22
23# Tag the constant ones
24constant = stats[stats["std"] < 1e-6].index.tolist()
25print(f"\nconstant sensors ({len(constant)}):", constant)
26
27# Tag the high-variance informative ones (top 5 by std)
28top5 = stats.sort_values("std", ascending=False).head(5)
29print("\ntop-5 high-variance sensors:")
30print(top5)
31
32# constant sensors (7): ['sensor_1', 'sensor_5', 'sensor_6',
33#                        'sensor_10', 'sensor_16', 'sensor_18', 'sensor_19']
34# top-5: sensor_9, sensor_14, sensor_2, sensor_3, sensor_4

PyTorch: Selecting Sensors as Channels

1import torch
2import numpy as np
3
4# Pretend X is loaded as (B, T, 21) with all 21 sensors
5torch.manual_seed(0)
6X = torch.randn(2, 30, 21)
7
8# ----- Drop constant sensors via index list -----
9INFORMATIVE_IDX = [1, 2, 3, 6, 7, 8, 10, 11, 12, 13, 14, 16, 19, 20]   # 0-based; 14 sensors
10
11X14 = X[:, :, INFORMATIVE_IDX]                    # PyTorch advanced indexing on last axis
12print("X.shape   :", tuple(X.shape))
13print("X14.shape :", tuple(X14.shape))            # (2, 30, 14)
14
15# Equivalently with torch.index_select
16idx_tensor = torch.tensor(INFORMATIVE_IDX, dtype=torch.long)
17X14b = torch.index_select(X, dim=2, index=idx_tensor)
18print("match     :", torch.equal(X14, X14b))     # True

What ‘Sensor’ Means in Other Domains

Two Sensor-Catalog Pitfalls

The bottom line. 21 sensors and 3 operational settings, of which 14 sensors are useful for FD001 modelling. The rest of the dataset processing pipeline (Section 5.3, Chapter 6, Chapter 7) ratchets through the obvious cleanup steps before any neural network sees the data.

Takeaway

• 3 op-settings + 21 sensors per row. 24 numeric columns total (plus engine_id, cycle = 26).
• Op-settings describe the regime. They are not modelling features; they tell us how to normalise the sensors.
• 14 informative sensors on FD001. The other 7 are constants - drop them in feature selection.
• Drift direction encodes failure mode. T30 up = HPC degradation; BPR down = fan wear. Useful for interpreting attention maps later.
• PyTorch advanced indexing handles selection. X[:, :, INFORMATIVE_IDX] reduces (B, T, 21) to (B, T, 14) in one line.