AI optimization of fish feeding: where feed is lost
In cage and pond farms, feeding still relies on a shift operator: the ration is set once per season from a table, and then adjusted "by eye" — manually adding feed if fish approach the feeder sluggishly.
Actual consumption depends on water temperature, dissolved oxygen, current and time after sorting, so a fixed regime almost always either overfeeds or underfeeds.
Losses go both ways. Overfeeding means uneaten pellets: they sink into the silt, decompose and consume oxygen, hence oxygen-deficiency zones and increased mortality. Underfeeding means lost growth: fish do not reach target weight on time, and part of the stock ends up off-grade.
A separate loss category is uneven dispensing. The feeder scatters pellets over an area, but in current and wind the feeding front shifts: one cage area receives several times more feed than the neighboring one.
Larger fish push away weaker ones, which do not get enough — underfeeding and size dispersion accumulate throughout the cycle.
The key is measurement where feed meets fish: underwater cameras and computer vision show how many pellets sink to the bottom and how the school responds to dispensing.
Below we break down how this data forms a feedback loop that changes the feeding regime based on actual conditions, not a table.
What changes on the farm: feed consumption, mortality, workload
The key shift is feed consumption per unit of biomass gain. The system dispenses feed according to actual stock consumption, not a fixed schedule: sensors record activity at the feeding area, and the model adjusts the portion and the interval between dispensings.
Feed stops settling as sediment, feed conversion levels out for each cage and tank, rather than being held to the farm average.
The second measurable shift is the mortality rate.
Underfeeding and overfeeding, which previously surfaced only at sorting or in a veterinary report, are visible at the moment of deviation: dispensing decreases in an area where consumption has dropped and does not ramp up where the stock is already satiated. Rare but severe episodes go away — oxygen kills after overfeeding and outbreaks after underfeeding.
The third is workload on staff. The shift stops spending time walking all feeders and manually correcting regimes: the operator reviews a summary by area and intervenes selectively.
Feeding control becomes continuous rather than sample-based — every feeding area keeps a log of dispensings, stock reactions and operator adjustments.
All metrics are calculated from one data stream, so the effect is visible not after a season but by weeks: feed consumption per unit of gain, mortality rate, number of manual interventions per shift. Hence the next question — how the loop that provides this is structured.
Feeding regimes: comparison of approaches and limitations
Feeding management is a choice between simplicity and sensitivity to fish condition. A timer gives a predictable regime and does not depend on instrument readings, but it notices neither an appetite drop nor a change in water temperature.
Portion feeding by sensors removes some manual decisions, but requires correctly selected trigger thresholds. Adaptive feedback-based dispensing changes volume and frequency based on actual consumption — but only with a stable signal and a configured automatic feeder.
| Option | When suitable | Where it breaks |
|---|---|---|
| Scheduled feeding (timer) | Stable temperature, uniform size, regime rarely changes | Does not respond to appetite decline: overfeed settles as sediment and spoils water |
| Portion feeding by behavior sensors | Fish active, pronounced approach to feeder, need to cut overfeeding | Threshold sensitive to turbidity and current, produces false triggers |
| Adaptive feedback-based dispensing | Non-uniform size, daily consumption fluctuations, accumulated history | Requires data for calibration; if signal is lost, safe fallback to schedule is mandatory |
| Manual feeding by operator | Startup, calibration, emergency periods | Depends on human, regime not reproducible |
Regimes do not replace each other but work in layers: the schedule remains a safeguard, behavior sensors refine the dispensing moment, and feedback corrects volume.
A workable system is one where transition between layers is described in advance and does not lead to feeding downtime.
How we implement the system: stages of engineering work
Implementation begins not with camera selection but with describing the farm: each tank or cage has its own feeding regime, and a model tuned on one misses the mark on the neighboring one.
Therefore the first days go to measurements and conversations with the shift, while installation takes much less time than tuning.
- Farm audit. Record stocking density, feed type and pellet size distribution, water discharge points, actual feeding times and how feed is distributed over the area. Compare log records with school behavior: a discrepancy in feeding time breaks all subsequent logic. Output — a map of regimes for each tank and a list of constraints where automation will not provide a gain.
- Placement scheme. Select points for cameras and sensors: viewing angle to the surface, absence of glare from lamps, protection from condensation and fouling, cable routes outside the maintenance area. We overlap viewing zones — one camera must not remain a single point of failure.
- Setup and calibration of portions. We run the dispenser on scales and set dispensing variation, then bind it to the model: it determines the end of feeding by school reaction, not by timer. Calibration is done separately for each line — mechanical wear is different everywhere.
- Validation on a live cycle. For the first weeks the system works in advisor mode: it proposes a portion, while the shift feeds. We analyze discrepancies via the event log and look at the share of feed settled on the bottom, growth and activity. Rollback to manual mode is one switch, without stopping the line.
- System handover. We write shift regulations, instructions for sensor failure or communication loss, configure alerts and train personnel. After the control period we repeat measurements and hand over to the farm the final placement scheme, portion settings and procedure for independent adjustment.
Delivery package: configurations, schemes, regulations
After launch, the farm retains not only a working system but also a document set by which it can be operated, handed over and restored.
We provide it in machine-readable form and on paper — so the shift operator does not search for parameters in correspondence, and a new engineer can join the project without us.
- Working configurations — feeding profiles for each tank: daily norms, dispensing windows, sensor trigger thresholds, profile version history.
- Connection schemes — topology of cameras, sensors and feeder controllers, power points, cable routes and node addressing.
- Access policies — roles of operator, technologist and service engineer, key issuance and rotation procedure, limits of changes in setpoints.
- Maintenance regulations — frequency of sensor calibration, optics cleaning and dispenser checks; procedure for deviations from norm.
- Shift instructions — one-page cards: switching to manual mode, emergency stop of dispensing, whom to call and in what sequence.
- Change log and rollback rules — how to return the previous profile if a new setting caused overfeeding or underfeeding.
- System passport — composition of nodes, responsible persons, warranty periods and connection points for service.
The package is updated with every logic change — otherwise documentation diverges from reality within a couple of months.
Case: feed consumption and mortality before and after implementation
We examine an implementation at the farm Onezhskaya Forel — a cage line for cold-water fish, six cages.
Before the project, feeding followed a fixed schedule with "by eye" adjustment: in cold water such a decision lags, some pellets went under the cage, and the volume of overfeeding was not visible in any reporting.
The system closed the feedback loop. The model receives temperature, dissolved oxygen and school behavior — how quickly the portion is consumed and at what depth the fish stay — and recalculates the daily ration.
Feeders operate based on consumption: dispensing is split into portions with a pause, and the next goes only if the previous is eaten. The key report for the farm is not "how much was dispensed" but "how much was assimilated."
Metrics before and after
- Feed conversion decreased from 1.6 to 1.24 with the same biomass gain.
- Fish mortality per season: from 11% to 5.8% — the main contribution came from cold periods.
- The spread between cages in feed halved: weak cages stopped blindly chasing the norm.
What was refined during the season. In the first weeks, automation reacted to bird splashes and storm ripples — we added a signal duration filter. During treatments and aeration activation, the loop switches to manual mode: automation only gets in the way there.
By mid-season the model was further trained — the fish grew, depth and consumption rate changed, and old thresholds caused underfeeding in large cages.
How does scheduled dispensing differ from adaptive?
A timer works from a feeding table: biomass, temperature and average daily growth set the norm, divided into several dispensings. The loop is open — it does not know how much feed was actually eaten.
Hence two imbalances: when appetite drops (molting, stress, nighttime oxygen decrease), pellets sink to the bottom and dissolve, while during an activity surge the fish are underfed.
Manufacturer feeding tables, for example Skretting, smooth out variation but remain an averaging over the stock.
Adaptive dispensing closes feedback on the actual reaction to feed. A camera or hydrophone tracks how fish approach the feeder, the model detects activity attenuation and stops dispensing — the remainder of the portion does not enter the water.
| Option | When suitable | Limitations |
|---|---|---|
| Timer by feeding table | Stable regime, uniform size, well-tuned biomass accounting | Does not respond to appetite; if aeration fails, overfeeding will not stop |
| Timer with environment correction | Daily temperature and oxygen fluctuations | Changes dispensing volume, but not the end of consumption moment |
| Adaptive dispensing by consumption | Intensive RAS, fish sensitive to overfeeding | Requires a clean signal and model validation on the specific stock |
A timer is sufficient where conditions are stable, size is uniform and feed losses are not critical. As soon as daily oxygen drops or size variation appear, timer error turns into constant overspending and extra load on the biofilter.
The limits of adaptive regime must also be understood: the model trains on a specific stock and feed, so transfer to another tank requires revalidation, while turbid water or lens fouling gives a false picture of consumption. Therefore the signal is calibrated on actual fish behavior, not averaged curves.
Technical doubts: sensors, communication, failures
At the start, they check not normal mode but system behavior during failures: loss of communication with the site, false sensor triggers, absence of an operator at the screen.
We build these scenarios into the architecture before the first launch, rather than refining them after an accident.
What happens if communication with the site fails?
The controller goes into autonomous mode and maintains feeding according to the last valid profile. Commands are not lost: they are queued and rechecked for relevance after channel restoration.
If the channel does not return within the allotted time, the responsible person receives a notification — not silence in monitoring.
How does the system combat false triggers?
The decision is made from a confirmed picture, not a single measurement: multiple sources, averaging window, thresholds with hysteresis — a short spike in temperature or turbidity does not cancel feeding.
A sensor that diverges from the group is marked doubtful, while the others continue to close the loop.
What about maintenance and sensor degradation?
Each channel has drift control: readings are compared with neighboring nodes and with history, and the system warns in advance that a probe is going out of tolerance. Cleaning and verification follow the regulations, and replacing one sensor does not stop the line.
How resilient is the system to failure of a dosing unit?
Critical nodes are redundant in power, and the actuator defaults to a safe state — feed is not dispensed on a false command. A failed node is removed from the loop, the others redistribute the load, and the event is logged with the exact cause and time.
Let's discuss the task and come back with a work plan
To come back with a work plan rather than general considerations, we need input data from the farm. One business meeting plus access to logs and the site — and we name the stages, timelines and acceptance metrics. Below is what is usually required at project start.
- Initial data on areas: boundaries of cages, tanks or ponds, number and species of fish, water supply and aeration scheme, depth and flow rate.
- Current feeding regimes: rations, daily norms, feed brand, who makes manual adjustments and based on what indicator.
- Archives for the previous season: feeding logs, temperature and oxygen sensor readings, notes on outbreaks and mortality.
- Accesses: to automatic feeder controllers, sensors, communication node and cameras on site — for read and write.
- Responsible person on the farm side: technologist or fish biologist with authority to change the feeding regime and confirm pilot results.
- Pilot on site: one line or cage, 2–4 weeks, with a control group on manual feeding nearby.
- Work plan: stages, areas of responsibility on both sides, control points and what we consider pilot success.
The first step is short: we discuss the task, you transfer data and accesses, we inspect the site and come back with an estimate and plan. Then comes the pilot, and based on its results we decide whether to transfer feeding optimization to the remaining areas.
