SubseaAI — Real-Time Pipeline Leak Detection

Full Technical Explanation The Physics FBG Sensors (Fiber Bragg Grating) When oil leaks from a hot pipeline (80°C fluid) into cold seawater (4°C), a thermal plume rises. FBG sensors detect this by measuring wavelength shift — when temperature rises, the fiber grating expands, shifting the reflected wavelength by ~10 picometers per °C. 64 sensors per node give spatial resolution of 1.5m per sensor. DTS (Distributed Temperature Sensing) Uses Raman scattering — a laser pulse fired down the fiber scatters light. The ratio of Stokes/Anti-Stokes backscattered light is temperature-dependent (Boltzmann statistics). This gives continuous temperature every 1 meter along 100m of fiber. The plume appears as a Gaussian anomaly above ambient. Hydrophones Turbulent oil flow through a crack generates broadband acoustic noise (0.1–50kHz). Leak rate scales with acoustic power as Q ∝ rate^1.5. Four hydrophones enable TDOA (time difference of arrival) triangulation to locate the source within ±2m. Pressure Bernoulli's equation — a leak creates a flow path, causing pressure to drop at rate dP/dt ∝ leak_rate². This is the fastest-responding signal (milliseconds) but least informative alone. The AI Dempster-Shafer Fusion — three independent evidence sources vote. Each sensor's reliability weight adapts to environment (high ocean current → thermal signal less reliable → acoustic gets more weight). This handles conflicting evidence better than simple averaging. Isolation Forest — self-trains on 500 normal readings, then scores anomalies. No labels needed. Feeds as 30% weight into fusion score. Kalman Filter — smooths noisy location estimates over time using state prediction + measurement correction. The System FastAPI async backend → 100 nodes read per second → WebSocket broadcasts to dashboard → SQLite WAL handles concurrent reads → alerts fire with 60s cooldown → tamper-evident audit log using SHA256 hash chaining. Proven: 85% F1, 88% recall, ~0.3ms per node inference, 100 nodes in <100ms.

Built With

• python