Defense - Maritime Systems

Their UUV Had to Surface to Think

A defense contractor's UUV program had a problem. The vehicles collected acoustic data just fine. But to classify what they found, they had to surface and phone home. In contested waters, surfacing to check email is a bad plan.

73 hrs Mission Duration
67ms Signal Classification
15W Power Budget

Client

Defense Contractor (Autonomous Maritime Division)

Industry

Defense - Maritime Systems

Use Case

UUV Signal Processing & Autonomous Analytics

Products Used

Expanso

Timeline

14 months from contract to first sea trial

ROI

37% improvement in mission completion rate

The Challenge

The existing UUV collected 47GB of acoustic data per mission. Then it surfaced, transmitted everything via satellite at $8/MB, and waited for shore analysts to tell it what to do. The whole cycle took 6 hours. In an operational scenario, that UUV would be debris.

  • Original workflow: collect data, surface, transmit, wait for analysis
  • Satellite transmission cost $8/MB - a full data dump was $376K
  • 6-hour analysis cycle made tactical response impossible
  • 15W power budget for compute - can't just add GPUs
  • IL4 compliance required for all onboard software
  • Existing ML frameworks drew 47W - vehicle couldn't make it home

The Solution

We put the analyst on the vehicle. Expanso runs signal classification models on the UUV's existing ARM compute board. 67ms to classify an acoustic signature. Vehicle decides what to do without asking anyone. When it finally surfaces, it sends a summary - not 47GB of raw recordings.

15-Watt Classification

Signal classification runs on the vehicle's ARM Cortex-A72 at 67ms per event. No GPU required. No extra power budget. The model fits in 340MB of RAM alongside mission control software.

Autonomous Decision Tree

Vehicle runs a behavior tree that responds to classified signals. Interesting contact? Investigate. Threat signature? Evade and log. Normal traffic? Continue mission. No surface required.

Compressed Mission Report

When the vehicle surfaces - by choice or schedule - it transmits a 12MB summary instead of 47GB raw data. Satellite costs dropped from $376K to $96 per mission.

The Results

First sea trial, the UUV completed a 73-hour mission without surfacing once. Classified 2,847 acoustic events, investigated 3 contacts of interest, and came home with a 12MB report. Previous gen would have surfaced 8 times and spent $3M on satellite.

67ms Classification Time
73 hrs Autonomous Duration
37% Mission Completion
$96 Transmission Cost
  • Signal classification time: 67ms on ARM Cortex-A72
  • Power draw: 15W vs 47W for previous ML framework
  • Mission completion rate improved from 61% to 84%
  • Satellite costs per mission dropped from $376K to $96
  • First sea trial: 73 hours, 2,847 classifications, zero surfaces
  • IL4 compliance validated in 3 months - DoD ATO in hand
  • Model updates deploy when vehicle docks - no OTA required
"We ran a 73-hour mission without surfacing. The vehicle classified 2,847 acoustic contacts, investigated three of them, evaded one potential threat, and came back with a 12-meg report. Previous vehicle would have surfaced 8 times and probably gotten spotted."
Chief Systems Engineer, Defense Contractor
Background

Running ML on constrained hardware?

If your edge devices have tight power budgets and zero connectivity, we should talk. We've deployed on UUVs, satellites, and other environments where watts matter.