A hatchery that adds sensors usually discovers something uncomfortable in the first month: the machines were already telling it the truth, and nobody was in a position to act on it. Detection is the easy half. The half that decides whether the investment pays is how much warning the signal gives, and whether that warning arrives while the batch can still be saved.
Key takeaways
- Air temperature is a setpoint. Eggshell temperature is the outcome, and a machine can hold one perfectly while the other drifts.
- The same CO2 reading means opposite things on day 4 and on day 14. A threshold that ignores incubation day generates noise, not alarms.
- Analysis earns its keep by showing spread across machines and batches, not by producing more averages.
- On rotating equipment, ISO 17359 makes “keep monitoring and do nothing” a legitimate documented decision.
Why the reading is not the diagnosis
Every alarm needs a reference value, and in incubation the reference is not always the one displayed on the panel. The setter controls air temperature. What determines embryo development is eggshell temperature, with a working optimum near 37.8 °C inside a 37.5 to 38.3 °C band, measured at the equator of the egg and only on eggs containing a live embryo, because a clear egg generates no heat of its own.
That distinction is the whole reason a well-instrumented machine can still run a bad batch. Embryo heat production rises through incubation, so a fixed air temperature that was correct on day 6 can leave eggshell temperature above the band on day 16. The sensor is not wrong. It is answering a different question from the one that matters.
How much warning each signal actually buys
Signals differ mostly in the timescale on which they move, and that timescale is what determines whether a reading is an alarm or a trend. The table below sets out how we read them operationally; it is a framework for specifying a monitoring system, not a set of measured constants.
| Signal | What it is really telling you | What it cannot see |
|---|---|---|
| Air temperature | Control loop and valve behaviour, minute by minute | The embryo’s own heat production |
| Eggshell temperature | The thermal balance the embryo is actually experiencing | Anything, if the sampled egg is clear |
| CO2 | Air exchange against embryo metabolism | Which of the two changed |
| Humidity and weight loss | Moisture loss trajectory across the batch | Embryo viability |
| Differential pressure between rooms | Whether the sanitary cascade still holds | Contamination itself |
| Vibration and bearing temperature | Mechanical degradation, over weeks | Process quality |
CO2, the parameter with published numbers
Carbon dioxide is worth singling out because the guidance is specific enough to build rules on. Royal Pas Reform recommends that air entering the setter plenum carry no more than 0.09% CO2, and describes broiler embryos as tolerating a gradual rise to 1.5% at day 4, maintained to day 10, while work on white layers reported negative effects at 2% compared with 0.03 to 0.05% during the first four days. After day 10 to 12, ventilation is fine-tuned to hold a maximum of 0.4%, and gradual ventilation should begin no later than the third day.
Read that as an instruction for alarm design rather than as a setpoint. The same 1.2% reading is normal on day 6 and a genuine problem on day 15. A threshold that is not indexed to incubation day, and to the species in the machine, will either alarm constantly in early incubation or stay silent when it matters. The same source also notes that CO2 sensors calibrated at sea level need a correction factor at altitude, which is the kind of detail that quietly invalidates a year of trend data.
Three ways a monitoring programme fails
- No baseline. An alarm threshold without a reference distribution is somebody’s opinion written in a config file. Before setting limits, log a few normal batches and look at how much the parameter moves when nothing is wrong.
- Thresholds set at the point of certainty. Limits placed where the deviation is undeniable are limits placed after the useful window has closed. Limits placed too tight produce an alarm on every batch, which trains the team to clear alarms rather than read them.
- No named owner. A deviation that appears on a screen and in nobody’s job description is data, not detection. Each signal needs a person, a response and a place where the response is recorded.
Machines degrade on a different clock
Process parameters change within a batch. Mechanical condition changes over months, and it is governed by a different logic. ISO 17359:2011, on condition monitoring and diagnostics of machines, makes the monitoring interval depend on the type of fault and the speed at which it progresses, and explicitly provides for deciding that no action is required on a low-criticality machine while surveillance continues.
The practical consequence is the P-F interval, the gap between the point where a fault becomes detectable and the point where the machine can no longer do its job. Where that interval is long, monitoring buys planning time. Where it is short, the honest answer is often to hold a spare part rather than to instrument the asset.
What analysis adds, and what it quietly hides
Analysis earns its place by exposing spread, not by producing more averages. Comparing the same parameter across machines, across positions in the same machine and across consecutive batches is what turns a stream of readings into a finding. A single mean value for the week conceals exactly the variation you are paying the sensors to reveal.
It also has to account for what changes outside the machine. Petersime places normal early embryonic mortality at 2.5 to 5.5% depending on breeder flock age, and attributes around 65% of early deaths to incorrect incubation conditions. Both halves of that sentence matter: a rise in early mortality is a strong signal, and it is not automatically a machine fault. The day an embryo died tells you when, never why. The answer is read on the flock, the storage history and the settings together, and that is where breakout analysis still beats any dashboard.
Questions that come up when the system is specified
What should be instrumented first?
Eggshell temperature and CO2, followed by differential pressure between zones. The first two act on the batch in progress, the third protects everything downstream of a containment failure and is cheap to monitor continuously.
Is predictive analytics realistic in a hatchery?
It is realistic where a fault progresses slowly and leaves a signature, which is mostly true of rotating equipment and much less true of biological outcomes. Predicting a bearing failure is a solved engineering problem. Predicting a hatch result from environmental data alone remains an estimate, because breeder flock age, egg storage and fertility all sit outside the sensor’s field of view.
How many alarms is too many?
The workable test is behavioural rather than numerical. If the team clears alarms without reading them, or if an alarm has never once changed what anybody did that shift, the threshold is wrong regardless of how defensible it looked on paper.
What derails these projects most often?
Calibration drift and undocumented sensor moves. A probe relocated during maintenance, or a CO2 sensor left on its sea-level factory calibration, invalidates the trend without producing a single error message. Recording where each sensor sits and when it was last verified is unglamorous and it is what keeps the historical data usable.
From signals to decisions
Detection is only the first half. Turning a stream of readings into decisions that hold up across seasons is a separate discipline, with its own data structure and its own discipline of record keeping.
Sources: Royal Pas Reform technical knowledge base on managing carbon dioxide in the setter, for the 0.09% intake limit, the 1.5% tolerance from day 4 to day 10, the 0.4% maximum after day 10 to 12, the day 3 ventilation start and the altitude calibration of CO2 sensors; Royal Pas Reform and Petersime technical documentation on eggshell temperature, for the 37.8 °C optimum, the 37.5 to 38.3 °C band and the measurement method; Petersime, for early embryonic mortality of 2.5 to 5.5% by breeder flock age and the 65% share attributable to incubation conditions; ISO 17359:2011, Condition monitoring and diagnostics of machines, for the monitoring interval principle and the P-F interval. Consulted August 2026.
Published previously, fully revised on 13 August 2026. General technical guidance for hatchery professionals. Alarm thresholds, incubation profiles and any change to sanitary parameters should be validated with your equipment supplier and your veterinary adviser before implementation.