What the published case confirms

EXPAND describes Rayonics International, an Italian company that designs and manufactures food-quality inspection machinery, including X-ray-based systems. Through the SME support program, the company received training, an opportunity analysis, and work on an AI technology roadmap.

The case identifies two streams: analyzing data from installed machines for predictive maintenance, and using edge models to check whether equipment is operating correctly. At publication time, a first practical edge model was ready for field testing; a cloud model for machine data was still being analyzed. This distinction matters: the case confirms preparation and readiness for a test, not the outcome of a production deployment.

Why edge inference may fit

When inspection equipment sits on a production line, local processing can reduce latency, preserve continuity during unstable connectivity, and limit data transfer outside the site. But an edge model alone does not guarantee defect detection or savings. The team must define which signal is analyzed, how ground truth is established, and what happens when the result is uncertain.

For food inspection machinery, the model must fit the engineering workflow: telemetry, sensor state, operating mode, batch identifier, maintenance, and operator events. If the goal is machine health rather than product-image inspection, the project needs a sufficiently long history of operating states and labeled faults. Routine logs may not contain enough examples of rare failures.

How to test before putting a model on the line

1. **Define an event.** For example, a sensor deviation, a component stop, or suspected instability. Do not start with a general promise to “predict everything.”
2. **Record the baseline.** Capture current signals, thresholds, operator interventions, downtime, and scheduled checks.
3. **Run in shadow mode.** The model logs recommendations without changing settings or stopping equipment. Compare alerts with later diagnostics.
4. **Price errors.** A false alert distracts an operator; a missed failure may cause scrap or unplanned downtime.
5. **Test failure handling.** Specify what happens if the model, network, or sensor is unavailable; the standard control loop must keep working.

Limits and economics

The EXPAND case does not provide figures for accuracy, number of machines tested, downtime reduction, false alerts, or payback. It therefore cannot be treated as evidence of an economic effect. Even a result on one installation may not transfer to machines with different age, configuration, or production conditions.

In a pilot, measure event accuracy, warning lead time, false-positive share, and operator minutes spent investigating alerts separately. Calculate value from actual downtime hours and maintenance cost; do not attribute all line availability to the model.

A next step for an SME

Choose one machine and one event for which historical signals exist. Before development, agree who labels the data, how many weeks the shadow test will run, and what results would justify continuation. If historical examples are missing, start by logging events and collecting stable telemetry. For this type of project, that is often more useful than immediately buying hardware or promising autonomous control.