An AI agent skill that generates complete predictive maintenance (PdM) models end-to-end: from raw IoT sensor data in S3 to a deployed SageMaker inference endpoint or daily batch job.
Give Kiro access to your S3 data and ask it to build a failure prediction model. The skill guides the agent through a nine-phase workflow:
- Setup — project scaffolding, dependencies
- Data Exploration — discover schemas, joins, formulation strategy
- Raw Aggregation — load, join, label, temporal split
- Baselines — first working model per formulation
- Experimentation — hypothesis-driven feature engineering loop
- Save & Plan — artifacts to S3, post-training decisions
- Example Inference — score one day of data for validation
- Deploy — real-time endpoint and/or daily batch
- Documentation — auto-generated project README
| Formulation | Use When |
|---|---|
| RUL Regression | You have run-to-failure data and need precise remaining-life estimates |
| Failure Classification | You have failure labels and need binary/multi-label alerts |
| Survival Analysis | Devices are maintained before failure (censored data) |
| Anomaly Detection | You want unsupervised deviation detection without failure labels |
Results from running the full skill workflow (Phases 1–5: setup → data prep → baseline → experimentation) on four canonical PdM benchmarks.
Best results (after Phase 5 experimentation loop with feature engineering):
| Benchmark | Metric | Best Result | Published SOTA | Notes |
|---|---|---|---|---|
| C-MAPSS FD001 (RUL) | RMSE ↓ | 11.40 | 9.21 (AutoRUL) | +24% gap; feature selection + XGBoost HPO |
| AI4I 2020 (Classification) | F1 ↑ | 0.889 | ~0.99 (FL+SMOTE) | -10% gap; closes with SMOTE |
| Backblaze hdfail (Survival) | C-index ↑ | 0.88 | 0.958 (Cox, 21 feats) | 52K drives, 94% censoring |
| NASA SMAP (Anomaly Detection) | F1 ↑ | 0.97 | ~0.90 (THOC/TranAD) | Spectral Residual; exceeds SOTA |
Baselines (Phase 4 only — no experimentation, default model, no feature engineering):
| Benchmark | Metric | Baseline | Notes |
|---|---|---|---|
| C-MAPSS FD001 (RUL) | RMSE ↓ | 16.48 | RULPredictor(window_size=15, stride=5, presets=medium_quality) |
| AI4I 2020 (Classification) | F1 ↑ | 0.82 | FailureClassifier() with default features |
| Backblaze hdfail (Survival) | C-index ↑ | 0.88 | SurvivalPredictor() — 52K drives, 94% censoring |
| NASA SMAP (Anomaly Detection) | F1 ↑ | 0.97 | SpectralResidualDetector(sr_window=5, aggregation_percentile=95) |
The baselines in pdm/benchmarks/baselines.json are regression gates — they ensure code changes don't degrade the out-of-the-box model quality. The "Best Result" column reflects what the skill achieves when the full experimentation loop (Phase 5) runs with domain-informed feature engineering.
The library achieves competitive results with minimal configuration. The RUL gap is primarily a model diversity issue (AutoRUL searches over SVR/KNN in addition to tree ensembles). The classification gap closes with SMOTE and threshold tuning. The anomaly detection F1 improved from 0.73 to 0.97 (+33%) by switching from temporal PCA reconstruction (TemporalAnomalyDetector) to Spectral Residual frequency-domain analysis (SpectralResidualDetector). The SR method computes per-feature FFT saliency, z-scores against training distribution, and aggregates using a robust percentile — achieving results that exceed published SOTA on SMAP without requiring deep learning or GPU.
git clone git@github.com:aws-samples/sample-predictive-maintenance-skill.git ~/.kiro/skills/predictive-maintenanceCopy this folder into your Kiro skills directory:
- Global (all projects):
~/.kiro/skills/predictive-maintenance/ - Workspace (one project):
.kiro/skills/predictive-maintenance/
The predictive-maintenance skill can leverage the knowledge-acquisition skill to build a domain-specific research wiki before running experiments. This is used in Phase 5 (Experimentation) — specifically Step 0: "Build Domain Knowledge Base" — where the agent researches state-of-the-art techniques from academic literature (arXiv, Semantic Scholar, Papers With Code, etc.) to ground experiments in proven methods rather than ad-hoc intuition.
The wiki produces actionable hypotheses for feature engineering and model selection, tailored to the specific PdM formulation (RUL, classification, survival, anomaly detection). Experiments that draw from literature-backed techniques consistently outperform those based on intuition alone.
git clone https://github.com/aws-samples/sample-knowledge-acquisition-skill.git ~/.kiro/skills/knowledge-acquisitionOr for workspace-scoped installation:
git clone https://github.com/aws-samples/sample-knowledge-acquisition-skill.git .kiro/skills/knowledge-acquisitionWhen the experimentation phase begins, the agent uses the knowledge-acquisition skill to:
- Gather sources — searches arXiv, Semantic Scholar, OpenAlex, and Papers With Code for papers relevant to the formulation (e.g., "remaining useful life prediction", "unsupervised anomaly detection IoT")
- Distill findings — extracts key techniques, feature engineering strategies, and model architectures
- Write wiki pages under
./wiki/concepts/:feature-engineering-pdm.md— state-of-the-art feature engineering for the formulation<formulation>-techniques.md— best models, loss functions, evaluation metricsdomain-features.md— domain-specific features for the equipment type
- Generate experiment hypotheses — each wiki page ends with an "Implications for Experiments" section listing concrete, actionable ideas
Without this skill installed, the experimentation phase still works — the agent will propose experiments based on feature importance analysis and CAAFE-style semantic reasoning. The knowledge-acquisition skill adds literature grounding that improves experiment quality.
- Python 3.11+ with
uv(or pip) - AWS credentials with S3 read access + SageMaker (for deployment)
- Kiro CLI or Kiro IDE
Key Python dependencies (installed automatically by scripts/setup.sh):
autogluon.tabular— AutoML trainingboto3/pyarrow— S3 + Parquet I/Opandas/numpy/scikit-learn— data processinglifelines— survival analysisflask— inference container
Start Kiro and say:
Build a predictive maintenance model from my S3 data in s3://my-bucket/
Or for benchmark datasets:
Train a RUL model on the C-MAPSS dataset
The skill activates automatically based on keywords like "predictive maintenance", "failure prediction", "remaining useful life", "RUL", or "survival analysis".
predictive-maintenance/
├── SKILL.md # Agent instructions (9-phase workflow)
├── README.md # This file
├── references/ # Detailed phase guides (loaded on demand)
│ ├── data-exploration.md
│ ├── benchmark-datasets.md
│ ├── raw-aggregation.md
│ ├── baselines.md
│ ├── experimentation.md
│ ├── artifacts.md
│ ├── deployment.md
│ └── readme-template.md
├── scripts/ # Setup & utilities
│ ├── setup.sh # One-command project bootstrap
│ ├── save_model.sh # Upload artifacts to S3
│ └── pyproject.toml # Dependencies
├── pdm/ # Standalone PdM library (usable without the skill)
│ ├── README.md # Library documentation
│ ├── data/ # S3 exploration, EAV aggregation, feature utils
│ ├── anomaly_detection/ # Spectral Residual + Temporal PCA + Isolation Forest
│ ├── fault_prediction/ # AutoGluon classification + RUL
│ ├── rul/ # Sliding-window RUL regression
│ ├── survival/ # Cox PH, Weibull AFT, Random Survival Forest
│ ├── deployment/ # Batch inference + SageMaker handlers
│ └── remote/ # SageMaker Training Job submission
├── sagemaker_container/ # Custom Docker container for endpoints
│ ├── README.md
│ ├── Dockerfile
│ ├── serve.py
│ └── inference.py
└── infrastructure/ # CDK stack for batch scheduling
├── README.md
├── app.py
├── batch_inference_stack.py
└── lambda/trigger.py
The pdm/ directory is a standalone Python library usable independently of the Kiro skill. It provides:
- 5 model classes:
AnomalyDetector,TemporalAnomalyDetector,SpectralResidualDetector,FailureClassifier,RULPredictor,SurvivalPredictor - CLI tools for training, evaluation, and inference
- S3 utilities for exploring and loading partitioned parquet data
- Batch inference utilities for production scheduling
See pdm/README.md for full API documentation and usage.
Kiro loads skills through progressive disclosure:
- At startup, only the
nameanddescriptionfromSKILL.mdfrontmatter are loaded (minimal context cost) - When your request matches the skill description, the full
SKILL.mdinstructions load - The agent reads
references/files only when needed for the current phase
This means the skill has near-zero overhead when not activated.
This skill follows the open Agent Skills standard and works with:
- Kiro CLI / Kiro IDE (primary target)
- Any agent that supports the Agent Skills
SKILL.mdformat
See CONTRIBUTING for more information.
This library is licensed under the MIT-0 License. See the LICENSE file.