Machine Learning & AI / Independent exploration
Aerospace Remaining Useful Life
A C-MAPSS sequence-model comparison whose current entrypoint and evaluation protocol expose reproducibility gaps.
Problem
Remaining useful life prediction estimates degradation from multivariate sensor sequences. The repository explores recurrent, attention-based, convolutional, and temporal-convolution models.
Architecture
Separate modules load data, normalize features, build engine-wise windows, train models, and evaluate predictions. Model checkpoints and result files are present even though a top-level README is absent.
Implementation
The inspected loader uses FD001 files and constructs test RUL using the official remaining-life offsets. Scaling is fitted on training records. Targets are capped at 125 cycles, and 30-cycle windows predict the following row. Evaluation is over overlapping windows, not simply the last window of each engine.
Results
VERIFIED: sequence preparation, MSE/Adam training, and regression/thresholded classification metrics exist. The current main script imports load_FD001 while the loader defines load_dataset; it also requests lowercase metric keys while the evaluator returns uppercase keys. The current entrypoint is therefore not a reproduced experiment.
Limitations
The dataset argument is ignored by the inspected loader, so an FD004-named result does not establish FD004 reproduction. There is no NASA score in the inspected evaluator. Published model metrics are withheld pending a corrected, seeded, engine-aware protocol. The repository name does not imply Rolls-Royce collaboration or endorsement.
Reproducibility
Reviewed at commit 934d8de704e1 on 2026-11-09. Source inspection establishes implementation structure; repository-authored results are not independent measurements. The upstream application and benchmark suite were not rerun during this review.
- main.py
- src/data_loader.py
- src/preprocessing.py
- src/sequence_generator.py
- src/train.py
- src/evaluate.py
Next investigation
Resolve the limitations above, preserve the reviewed commit and exact experiment configuration, and reproduce correctness checks before promoting performance or deployment claims.