How duytrangiale/tower_watch works
TowerWatch is a prototype for detecting and localising lattice-tower damage from vibration, first using synthetic data and then transferring the same analysis to real LUMO measurements. Its main workflow models the tower and damage, simulates or loads sensor signals, extracts features, trains a healthy-data autoencoder, and evaluates anomaly scores and sensor-level localisation. A separate prioritisation stage ranks evaluated instances for maintenance. The graph reflects documented stages and indexed script dependencies; some feature-extraction and evaluation wiring is documented in the README rather than shown in sampled script excerpts.
Main entry points: truss.py Truss FEM · gcn_ae.py Graph autoencoder · metrics.py Detection and localisation · geometry.py Lattice geometry
Components 20
Structure and simulation
- Lattice geometry
src/fem/geometry.py - Truss FEM
src/fem/truss.py - Damage injection
src/fem/damage.py - Temperature model
src/simulate/environment.py - Vibration simulation
src/simulate/response.py - Modal energy analysis
src/fem/modal_energy.py
Data and features
- LUMO adapter
src/data/lumo.py - Time features
src/features/windows.py - Spectral features
src/features/spectral.py
Models and evaluation
- Sensor graph
src/graph/build.py - Graph autoencoder
src/models/gcn_ae.py - Per-sensor autoencoder
src/models/per_sensor_ae.py - Message-passing model
src/models/message_passing_ae.py - ARX model
src/models/arx.py - Baseline models
src/models/baselines.py - Autoencoder training
src/models/train_utils.py - Detection and localisation
src/evaluate/metrics.py - Tower visualisation
src/fem/visualize.py
Maintenance ranking
- Priority scoring
src/evaluate/priority.py
Other components
- TowerWatch user
Connections 22
- TowerWatch user → Lattice geometry: runs synthetic workflow
- Lattice geometry → Truss FEM: builds model
scripts/01_generate_synthetic.py - Damage injection → Truss FEM: alters stiffness
scripts/01_generate_synthetic.py - Temperature model → Truss FEM: adjusts material
- Truss FEM → Vibration simulation: provides structure
src/simulate/response.py - Vibration simulation → Time features: yields signal windows
- TowerWatch user → LUMO adapter: runs real-data workflow
- LUMO adapter → Time features: windows measurements
src/data/lumo.py - LUMO adapter → Spectral features: computes features
src/data/lumo.py - Time features → Graph autoencoder: provides features
- Spectral features → Graph autoencoder: provides features
- Lattice geometry → Sensor graph: defines mesh
scripts/03_train.py - Sensor graph → Graph autoencoder: provides adjacency
scripts/03_train.py - Autoencoder training → Graph autoencoder: trains model
scripts/03_train.py - Autoencoder training → Message-passing model: trains model
scripts/day3_gnn_architecture_experiment.py - Graph autoencoder → Detection and localisation: scores reconstruction
scripts/04_evaluate.py - Per-sensor autoencoder → Detection and localisation: evaluated by
scripts/day4_lumo_architecture_experiment.py - Baseline models → Detection and localisation: evaluated by
scripts/05_lumo_transfer.py - ARX model → Detection and localisation: evaluated by
scripts/day3_arx_localization_experiment.py - Detection and localisation → Priority scoring: supplies anomaly results
scripts/06_priority_ranking.py - Lattice geometry → Tower visualisation: renders structure
src/fem/visualize.py - Truss FEM → Modal energy analysis: supports modal analysis
src/fem/modal_energy.py