File size: 2,989 Bytes
2901574
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
---

license: mit
tags:
  - cbc-reference-model
  - mlops-100-day
  - predictive-maintenance
  - pytorch
  - lstm
---


# CBC Reference Model: Turbofan Remaining Useful Life (C-MAPSS FD001)

> Pre-trained reference model for the **CBC [MLOps 100-Day Track](https://github.com/careerbytecode/cbc-learning-hub/tree/main/100-days/mlops)** (Capstone 4). Published twin of ML Development Capstone 4. **This is the one deep-learning reference model** — it ships as a torch `state_dict` + `model.py`, NOT a single joblib.



## Model details

- **Type:** single-layer LSTM (hidden 64) over a 30-cycle window of 15 normalized sensors -> scalar RUL (capped at 125).

- **Framework:** pytorch 2.12.0+cpu · **Serialization:** `manufacturing_lstm.pt` (state_dict) + `manufacturing_meta.joblib` (normalization + arch). Reconstruct with the shipped `model.py`.
- The LSTM is the rare counterpoint to the classical models: run-to-failure multivariate sensor sequences are the data shape deep learning is for. It beats an XGBoost baseline (RMSE 18.45) overall and decisively near failure.

## Intended use
Decision-support estimate of operational cycles remaining, to prioritize inspection/maintenance. NOT an automated ground-or-fly authority. Teaching/reference artifact.

## Training data
NASA C-MAPSS FD001 (100 train + 100 test run-to-failure engines, single operating condition). 15 non-flat sensors used. Simulated, no PII. NASA Open Data (public domain).

## Metrics (test = last cycle of each test engine vs RUL_FD001, scored once)

| Model | Test RMSE | Near-failure RUL[0,50) |

|---|---|---|

| XGBoost baseline | 18.45 | — |

| **LSTM (deployed)** | **14.88** | **4.78** |



The LSTM wins overall and is decisively better in the operationally critical near-failure band.



## How to load and predict

```python

from huggingface_hub import snapshot_download

import sys, json

d = snapshot_download("careerbytecode/mlops-ref-manufacturing-rul")
sys.path.insert(0, d + "/model"); sys.path.insert(0, d)
from model import load_model, predict_rul       # needs torch installed
model, meta = load_model(d + "/model")

sample = json.load(open(d + "/sample_input.json"))
print(predict_rul(model, meta, sample["window"]))  # predicted RUL (cycles)

```

**Serving requires `torch`** and the shipped `model.py` (the class definition) — a joblib load alone will not work.



## Limitations

- Trained on FD001 (one operating condition); FD002/FD004 (six conditions) need condition-aware normalization.

- RUL capped at 125: cannot distinguish a very-healthy from a merely-healthy engine, by design.

- Needs the full 30-cycle window + the training normalization stats to serve; a single reading is not enough. Simulated data — expect drift on real telemetry. Reference/teaching artifact only.



---

© 2015-2026 CareerByteCode. All rights reserved. | CC BY-NC-SA 4.0 (docs), MIT (code) | Authored by Raghavendra R, Platform Owner CareerByteCode, Solution Architect