SUBSEA V7

Inspiration

Deep-sea pipeline integrity is a high-stakes game where current monitoring methods are often too slow or too localized. We wanted to build a "nervous system" for the ocean floor—a project that doesn't just watch for leaks, but understands the fundamental physics of the environment to prevent ecological disasters before they escalate.

What it does

SUBSEA V7 is an autonomous, multi-modal detection system that monitors subsea pipelines for leaks. By fusing thermal, acoustic, and pressure data, it detects anomalies (like a hot oil plume in 4°C water) and triangulates the exact source of a leak within a \pm 2\text{m} radius. It provides real-time alerts through a high-concurrency dashboard, ensuring that operators see the "why" and "where" of an incident in milliseconds.

How we built it

The system is built on a sophisticated hardware-software integration:

• The Sensory Layer: We utilized FBG Sensors to measure wavelength shifts (\sim 10\text{pm/°C}) and DTS (Distributed Temperature Sensing) using Raman scattering to map temperature gradients every meter.
• Acoustics & Pressure: Four hydrophones capture broadband noise (0.1\text{--}50\text{kHz}), while pressure sensors monitor drops via Bernoulli’s equation.
• The Intelligence: We implemented Dempster-Shafer Fusion to weigh sensor reliability dynamically. This is supported by an Isolation Forest for unsupervised anomaly scoring and Kalman Filters for smoothing location tracking.
• The Backend: A FastAPI async architecture with WebSockets for real-time broadcasting and SHA256 hash chaining for a tamper-evident audit log. ## Challenges we ran into The biggest hurdle was conflicting evidence. In high ocean currents, thermal plumes dissipate quickly, making FBG data less reliable. We solved this by creating an adaptive weighting system where the AI automatically shifts "trust" to acoustic signals when thermal variance exceeds a specific threshold. Additionally, managing the high data throughput of 100 nodes per second required optimizing our SQLite WAL (Write-Ahead Logging) to prevent database bottlenecks. ## Accomplishments that we're proud of
• Precision: Achieving a spatial resolution of 1.5\text{m} per sensor and source triangulation within \pm 2\text{m}.
• Speed: Clocking inference at \sim 0.3\text{ms} per node, allowing us to process 100 nodes in under 100\text{ms}.
• Accuracy: Reaching an 85% F1-score and 88% recall, which is critical in an environment where a missed leak is a catastrophe. ## What we learned We gained a deep appreciation for Multi-sensor Data Fusion. We learned that individual sensors are fallible—pressure is fast but vague, and thermal is precise but slow—but when fused via Dempster-Shafer logic, they provide a level of certainty that a single-stream AI simply cannot match. ## What's next for SUBSEA V7 The next evolution involves moving from detection to prediction. We are looking into integrating historical current data and structural stress modeling to predict "weak points" before a breach even occurs. We also aim to scale the network beyond 100 nodes to cover trans-continental pipeline distances.

Built With

• claude
• glm
• python