README.md

---
license: other
task_categories:
  - time-series-forecasting
  - tabular-regression
pretty_name: TwinnableAgent C-MAPSS processed parquet data
tags:
  - cmapss
  - predictive-maintenance
  - remaining-useful-life
  - twinnableagent
---

# TwinnableAgent C-MAPSS Processed Parquet Data

This dataset hosts the processed NASA C-MAPSS turbofan degradation files used by
TwinnableAgent. It includes FD001 through FD004 as parquet files plus the
conversion script used to derive them from the NASA raw training text files.

Raw NASA zip and text dumps are intentionally not mirrored here. The raw source
of record is the NASA Prognostics Center of Excellence C-MAPSS Turbofan Engine
Degradation dataset, currently distributed through NASA Open Data as
`CMAPSSData.zip`.

## Files

| File | Rows | Units | Operating conditions | Fault modes |
|---|---:|---:|---:|---:|
| `FD001_processed.parquet` | 20,631 | 100 | 1 | 1 |
| `FD002_processed.parquet` | 53,759 | 260 | 6 | 1 |
| `FD003_processed.parquet` | 24,720 | 100 | 1 | 2 |
| `FD004_processed.parquet` | 61,249 | 249 | 6 | 2 |
| `scripts/convert_cmapss_to_parquet.py` | - | - | - | - |

## Conversion

Each processed parquet is derived from the matching NASA training file:

```text
train_FD001.txt -> FD001_processed.parquet
train_FD002.txt -> FD002_processed.parquet
train_FD003.txt -> FD003_processed.parquet
train_FD004.txt -> FD004_processed.parquet
```

The conversion keeps the NASA column order, assigns explicit column names, and
adds a per-row remaining-useful-life label:

```text
rul = max(cycle for unit) - cycle
```

The conversion command used by TwinnableAgent is:

```bash
python scripts/convert_cmapss_to_parquet.py --datasets FD001 FD002 FD003 FD004
```

## Schema

All four parquet files use the same schema:

| Column | Type | Meaning |
|---|---|---|
| `unit` | integer | Engine trajectory identifier. |
| `cycle` | integer | Time-cycle index within the trajectory. |
| `op1`, `op2`, `op3` | float | NASA operating setting columns. |
| `s1` through `s21` | float | NASA sensor channels. Sensor units are anonymized by C-MAPSS. |
| `rul` | integer | Remaining useful life in cycles, derived during conversion. |

## Operating-condition clusters

TwinnableAgent PAL resolves FD002 and FD004 into six operating-condition groups
by rounding the three operating settings to the nearest integer and assigning a
stable label by sorted tuple order.

FD002 mapping:

| Cluster | Rounded `(op1, op2, op3)` | Rows | Units represented |
|---|---:|---:|---:|
| `op_cluster_0` | `(0, 0, 100)` | 8,044 | 260 |
| `op_cluster_1` | `(10, 0, 100)` | 8,096 | 260 |
| `op_cluster_2` | `(20, 1, 100)` | 8,122 | 260 |
| `op_cluster_3` | `(25, 1, 60)` | 8,002 | 260 |
| `op_cluster_4` | `(35, 1, 100)` | 8,037 | 260 |
| `op_cluster_5` | `(42, 1, 100)` | 13,458 | 260 |

FD004 mapping:

| Cluster | Rounded `(op1, op2, op3)` | Rows | Units represented |
|---|---:|---:|---:|
| `op_cluster_0` | `(0, 0, 100)` | 9,238 | 249 |
| `op_cluster_1` | `(10, 0, 100)` | 9,224 | 249 |
| `op_cluster_2` | `(20, 1, 100)` | 9,091 | 249 |
| `op_cluster_3` | `(25, 1, 60)` | 9,139 | 249 |
| `op_cluster_4` | `(35, 1, 100)` | 9,162 | 249 |
| `op_cluster_5` | `(42, 1, 100)` | 15,395 | 249 |

## Provenance note: FD004 unit count

The NASA readme describes FD004 training data as 248 trajectories, but the raw
`train_FD004.txt` file contains 249 distinct unit identifiers. This repository
preserves the raw file as authoritative and reports FD004 as 249 units.

## Intended use

These files support TwinnableAgent reproducibility, PAL adapter validation,
cross-dataset predictive-maintenance experiments, and C-MAPSS RUL benchmark
work. They are processed training trajectories, not raw NASA archives.