CESTA

Communication-Efficient Spatial-Temporal Aggregation for sensor-network fault diagnosis.

Official implementation of the CESTA paper.

Contents: Setup · Quickstart · Capabilities · Usage · Firmware · Citation · License

Setup

Requirements: Python 3.11+ and uv. uv sync creates the environment with all dependencies, including PyTorch.

git clone https://github.com/Sinner/CESTA.git
cd CESTA
uv sync
uv run cesta --help

Quickstart

Three steps from raw data to an evaluated model:

# 1. Transform raw Intel Lab data into a canonical fault-injected dataset
uv run cesta transform intel_lab data/raw/Intel/data.txt data/datasets/Intel_fault15 \
  --config config/datasets/intel-lab/fault-15.yaml

# 2. Train the communication-aware CESTA model
uv run cesta train config/training/cesta.yaml data/datasets/Intel_fault15

# 3. Evaluate a run (test split, or validation-only for tuning)
uv run cesta evaluate --model runs/cesta/<run_id> --data data/datasets/Intel_fault15
uv run cesta evaluate --model runs/cesta/<run_id> --data data/datasets/Intel_fault15 --split val

Capabilities

Models

• Temporal baselines: CNN1D, Transformer, Autoformer, Informer, PatchTST, ModernTCN, and Hydra.
• Spatial models: dynamic ST-GCN, HiFiNet, HMCT, DCRNN, and CESTA.

Data and evaluation

• Markov injection of SPIKE, DRIFT, and STUCK sensor faults with per-event randomization.
• Canonical graph datasets with node masks, dynamic edge masks, node positions, and edge distances.
• Classification, communication, and theoretical TX+RX radio-energy metrics.

Reproducibility

• Immutable per-run artifacts: manifests, configs, checkpoints, histories, and predictions.
• Optuna hyperparameter search and audit tooling for benchmark matrices.
• ESP32-S3 firmware for distributed on-device deployment (see Firmware).

Usage

The full experimental protocol is executable from the checked-in configs and scripts:

# Baseline matrix (inspect with --dry-run first; resumable, multi-GPU)
uv run python scripts/run_all_baselines.py --num-gpus 4 --keep-progress-log

# Audit run artifacts against the decisive-comparison specification
uv run python scripts/audit_decisive_comparison.py \
  --spec config/benchmarks/decisive-comparison.yaml \
  --runs-root runs --output runs/decisive-comparison-audit

# Paired summaries against a locked reference variant
uv run python scripts/summarize_decisive_comparison.py \
  --runs-csv runs/decisive-comparison-audit/runs.csv \
  --output runs/decisive-comparison-summary \
  --comparison <variant> <locked-reference>

# Budget-matched controls: tune on validation artifacts only, lock, then audit
uv run python scripts/generate_control_tuning.py \
  --spec config/benchmarks/control-tuning.yaml \
  --output runs/control-tuning/generated
uv run python scripts/derive_control_budgets.py \
  --runs-csv <validation-runs.csv> --source-variant <variant> \
  --output runs/control-tuning/control-budgets.yaml
uv run python scripts/lock_control_policies.py \
  --budgets runs/control-tuning/control-budgets.yaml \
  --validation-runs-csv <validation-runs.csv> \
  --controller <control> <variants...> \
  --output runs/control-tuning/control-lock.yaml
uv run python scripts/audit_locked_controls.py \
  --lock runs/control-tuning/control-lock.yaml \
  --test-runs-csv <test-runs.csv> --output runs/control-locked-audit

Every training invocation creates a new, never-overwritten run:

runs/<model>/<run_id>/
├── weight.pt
├── config.json
├── history.jsonl
├── manifest.json
├── eval_metrics.json
├── predictions.npz
└── communication_metrics.json  # communication-aware models

Firmware

See firmware/README.md for hardware requirements, configuration, model export, build/flash commands, and telemetry formats.

Citation

License

MIT