Escape prevention is a frequency problem

A net tear does not wait for the quarterly dive. It happens on a particular night, in particular weather, and every day between then and the next inspection is a day of exposure. Which is why the useful question is not "how do we inspect nets better" but "how do we inspect them more often".

The maths of the problem

If damage can occur on any day and inspection happens quarterly, average exposure between occurrence and detection is around six weeks. Move to weekly inspection and it drops to three or four days. Move to daily and it is measured in hours.

Average time a tear goes undetected

Nothing about the quality of the inspection changed. The entire improvement came from frequency - and frequency is a cost problem, not a technology problem.

What makes daily inspection realistic

Three things have to be true, and they are all operational rather than technical.

Farm staff have to be able to run it

If a specialist is required, inspection becomes an event to be scheduled and you are back where you started. Modern compact ROVs are designed for single non-specialist operator deployment, which is precisely the change that made daily programmes possible. Fish farm operators run these systems without ROV pilot training.

It has to fit inside an existing shift

Deployment in three to five minutes from setup to water entry means a cage round fits into a normal working day rather than requiring one. If getting the vehicle in the water takes half an hour, the programme will not survive a busy week.

The record has to build itself

Manual documentation is the part that quietly consumes staff time and the part that lapses first. Automated logging and structured reporting are what turn a series of inspections into an audit-ready record without anyone typing it up.

Where AI genuinely helps here

Worth being specific, because the term covers both real capability and marketing. Net damage detection is a narrow, well-defined visual task with abundant training data - exactly the kind of problem machine learning does well. AI-assisted net damage detection flags candidate areas automatically, which matters when an operator is reviewing the same repetitive footage daily and attention inevitably drifts.

Treat it as a second pair of eyes that never gets bored, not as a replacement for the operator. Broader claims about automated condition assessment deserve more scepticism.

Choosing the platform

Three genuinely different starting points, and the deciding question is whether you need to do something as well as see it:

• Look only, daily, simple. A compact purpose-designed farm vehicle with a gripper option and water quality sensors. The realistic basis for a routine programme.
• Intervene as well. A heavier platform with manipulator capability for net repair support and mooring recovery - a different class of vehicle.
• Multi-site at scale. AI-assisted detection and remote operation, aimed at operations with enough sites that inspection has to scale without adding headcount at each one.

Video · QYSEAIntroducing: FIFISH E-GO Underwater Robot

The site conditions that change the answer

• Visibility through the year. Autumn blooms and sediment change what a camera can do - see sonar or camera.
• Current at the site. Station-keeping matters more at exposed sites than most specifications acknowledge.
• Cage depth. Usually well within any professional platform's rating, which is why depth should not drive this decision.
• Who is on site in winter. The programme has to work with the staffing you actually have in February.

Start by proving it on your own cages

Water clarity, current and staff confidence vary enough between sites that a demonstration elsewhere tells you very little. Rent or trial a platform on your own cages, with your own people, in conditions that are typical rather than favourable - and specifically test whether the person who will use it weekly is willing to. That, more than any specification, predicts whether the programme survives its first six months.