A manufacturing AI project should begin with a specific operational loss: an avoidable stoppage, an inspection problem or a planning constraint. The useful question is whether a system improves that decision under actual production conditions.
There is no universal reduction in downtime or defects that applies across factories. Machine behaviour, product mix, instrumentation and the way people respond to alerts determine the result.
Predictive maintenance: an alert needs a response
Condition monitoring uses measurements such as vibration, temperature or electrical current to identify changes. Some systems detect anomalies; others estimate a particular failure risk. These are different tasks. An unusual reading does not, by itself, establish that a component will fail or how long it has left.
Connect sensor readings with operating conditions, maintenance records and known events. A model trained on one operating regime may produce misleading alerts when load, tooling or material changes. Avoid presenting anomaly detection as a guaranteed advance warning of breakdowns.
Agree who receives an alert, how it is investigated and what evidence justifies stopping equipment. Measure false alarms, missed events and useful warning time. A technically accurate alert that nobody can act on has limited operational value.
Visual inspection: measure the defects that matter
Cameras and models can support inspection when lighting, positioning and the relevant defects are observable. Build the dataset from representative products, shifts and conditions. Include acceptable variation as well as defects.
| Evaluation question | Why it matters |
|---|---|
| Which defects were missed? | A high overall accuracy can hide rare but costly defects. |
| Which good parts were rejected? | False rejects create scrap, reinspection and delays. |
| Does performance vary by product or shift? | A combined score can conceal a weak production condition. |
| Can the line handle the inspection latency? | A good offline result may not fit the production cycle. |
Test on separate production batches or time periods where appropriate. Near-identical images in training and test sets can overstate generalisation. Keep a route for uncertain images and a way to audit the decision.
Planning and forecasting: compare with the existing method
Production planning combines constraints such as material availability, changeovers, due dates and staffing. Mathematical optimisation may be useful without a machine-learning model. Demand forecasting can inform a plan, but it cannot remove supply constraints or promise that additional capacity will be sold.
Compare a proposed approach with the current planner and a simple baseline. Evaluate on later periods that were unavailable during development. Include stockouts, schedule changes, work in progress and delivery performance rather than focusing only on forecast error.
An improvement in OEE is not automatically an equivalent increase in revenue. Demand, downstream bottlenecks and contribution margin affect the financial outcome.
Prepare a pilot around one production question
A hypothetical pilot might examine whether vibration alerts help maintenance staff identify a particular issue on a defined group of machines. Its design would specify the event being detected, available records, the validation method and the response procedure. It would not assume a predetermined saving or claim that an avoided failure has already been proved.
Before live operation, establish:
- Reliable timestamps and equipment identifiers across data sources.
- Appropriate access to machine, MES and maintenance records.
- A test plan covering faults, missing signals and sensor drift.
- A responsible engineer and an escalation procedure.
- A fallback if the model or connection becomes unavailable.
Introducing model output into machine control requires a separate engineering and safety assessment. An advisory pilot does not authorise automatic control of equipment.
Evaluate costs and operational value
Include sensors, installation, connectivity, compute, integration, labelling, evaluation, training and maintenance. Existing equipment may need adaptation; retrofitting feasibility must be checked for the specific machine and installation.
Separate observed benefits from estimates. Track actual downtime and maintenance interventions using comparable periods, and account for changes in production volume or product mix. Avoid counting the same loss reduction in both quality savings and additional throughput.
Use scenario ranges when failures are rare or the financial impact is uncertain. A project can be valuable without a credible short-term ROI estimate, but that uncertainty should be explicit in the decision.
Questions before a manufacturing pilot
How much historical data is required?
There is no fixed minimum that suits every task. Data must represent the operating conditions and events the system is expected to handle; rare failures may require a different validation approach from frequent defects.
Who should own the system after launch?
An operational owner should decide how outputs affect work, supported by the technical team responsible for data, monitoring and updates. Maintenance and quality specialists need a clear escalation and fallback procedure.
Connect the pilot with the wider process
Q-Journey starts from the organisation and its processes, with AiP as part of that journey. If a defined integration or data system is the immediate requirement, explore Q-Dev Data. Project scope and responsibilities are agreed individually.