Process Optimization with Machine Learning: Better Setpoints

In many plants, process setpoints such as oven temperature, line speed, pressure, or dosing rate come from commissioning values or a senior operator's habits. Each shift has its own "recipe". Results drift: energy per tonne, yield, and rework vary from day to day for the same product.

PLC data and quality records often already show that some periods ran better than others. What is missing is a systematic way to ask: under which conditions, and with which settings? That is where process optimization with machine learning comes in. This guide covers the data you need, recommendations versus closed-loop control, safety limits, and how to validate results with the process team.

What process optimization with machine learning means

Machine learning in a plant usually answers one of three questions:

• Prediction: what will happen to this batch? Defect risk is one example, covered in predicting production defects with machine learning.
• Forecasting: what will production load or energy demand be next week? See AI forecasting for production and energy.
• Optimization: which settings should we run now for the best result, without breaking quality or safety limits?

Optimization usually treats a predictive model as a simulator of the process. The model learns how setpoints and process conditions relate to outcomes. An optimization routine then tries combinations of setpoints inside the permitted ranges and picks the one expected to perform best under today's conditions.

Define the target and the constraints first

Choose one measurable target, such as energy per tonne, yield, cycle time, or rework rate. Quality specifications become constraints, not a second target. General examples, not case studies:

Process	Adjustable setpoints	Target	Constraint
Dryer or oven	Air temperature, conveyor speed, damper opening	Energy per tonne	Product moisture within spec
Air compressor	Header pressure	kWh per m³ of air	Minimum pressure at point of use
Batch mixing	Temperature, mixing time, dosing rate	Cycle time	Lab results within spec

The data you need

• Three groups of variables. What you can adjust (setpoints), disturbances you cannot (raw material properties, ambient temperature and humidity, load, product type), and outcomes (quality, energy, output). Without disturbance data, a model cannot tell why the same settings gave different results.
• Past variation in setpoints. If the temperature was always 180 °C, the data says nothing about 175 °C or 185 °C. Variation can come from differences between shifts, or from small planned trials inside the safe range.
• Time lags. A setpoint change shows up in the outcome only after the material's residence time, and lab results often arrive hours later. Align data by batch or residence time, not by clock hour.
• Clean periods. Flag start-ups, product changeovers, and stoppages so they are not treated as normal running.
• Context notes such as maintenance, raw material lots, and quality complaints. If they still live on paper, start with digitising production logsheets.

How to find better setpoints

1. Compare the best periods. Group similar conditions, such as the same product and raw material, then compare the settings in the best-performing periods with the rest. It is simple, explainable, and often enough as a first step.
2. Model plus optimization. A model estimates outcomes from setpoints and disturbances, and the optimizer searches for the best settings inside the permitted range. Recommendations follow current conditions, for example different settings when raw material arrives wetter.
3. Stay inside the data. Never recommend values the plant has not run; a model knows nothing beyond its data.

Beware of correlation

As an illustration, operators raise dryer temperature whenever raw material arrives wet. In the data, high temperature then appears alongside high product moisture. A naive model may conclude that heat makes the product wet and suggest lowering it. That is why disturbances such as raw material moisture must be included, and every finding needs judging by someone who knows the process.

Recommendations or closed-loop control

Aspect	Recommendation mode	Closed loop
How it works	Suggested setpoints appear on a dashboard; the operator decides	The system writes setpoints to the PLC at intervals
Risk	Low; read-only access to the PLC	Higher; needs layered safeguards
When to use	Early stages and sensitive processes	After recommendations prove themselves over time

Start with recommendations. Log every suggestion that is accepted or rejected, with the reason; that log is your best learning material. If you later move to closed loop, the PID loops in the PLC still hold the process while the model only shifts their targets within limits. The system must fall back to the last safe setpoint when data goes bad or the model stops, and operators must be able to switch it off.

Safety limits that must never be crossed

• Hard limits from the process team and equipment design, such as maximum temperature and pressure, are enforced in the PLC, not just in the model.
• Size and frequency of changes are capped, for example a maximum step per change and a minimum gap between changes.
• Interlocks and safety systems are out of scope. Optimization never touches them.
• Changes to the permitted ranges go through the same management of change procedure as any other process change.
• A data quality guard. If a key sensor freezes or reads outside a plausible range, no recommendation is issued.

Validating with the process team

1. Backtest on a period the model was not trained on.
2. Engineering review. Does the direction of each recommendation make physical and chemical sense?
3. Controlled trial. For example alternating days or shifts, with the target, baseline, duration, and stop criteria agreed beforehand.
4. Measure honestly. Compare like with like, such as product mix and weather, and record when advice was actually followed.
5. Monitor and retrain. Equipment wear, a new raw material supplier, or a new product can move the optimum.

Things to keep in mind

• The benefit depends on the process. A process already near its physical limits or under tight control leaves little room.
• A model only reads causes that are recorded. Unmeasured disturbances weaken recommendations.
• Savings cannot be promised before the data is reviewed.
• Involve operators from day one. Advice nobody trusts never gets used.

Frequently asked questions

How is this different from advanced process control (APC)?

APC, such as model predictive control, uses dynamic models to control several variables continuously. Machine learning setpoint optimization usually works at a slower layer, choosing the targets that the control system then holds. The two can complement each other.

Do we need new sensors?

Not necessarily. Many setpoints and process values are already logged in PLCs. New sensors are needed when an important disturbance, such as raw material moisture, is not measured yet.

How much historical data is needed?

Variety matters more than length: enough products, raw materials, weather, and settings on record. Sufficiency is judged in the data review, before any model is built.

Process optimization with INCLUDE

Process data from sensors and PLCs reaches the INCLUDE platform over MQTT, HTTP, or IncludeGateways for Modbus RTU devices. Historical data can be imported (Pro plan and above) without triggering alarms, so past production can be reviewed.

The model is built through our predictive analytics and machine learning services: starting from the business question and a data review, then testing against your historical data and going through the results with your operations team. Once running, recommendations and predictions appear on the dashboards and alarms of the same platform, accuracy is monitored, and the model is retrained when conditions change. The platform can change setpoints from a dashboard, but for optimization we recommend starting in recommendation mode with read-only access to the PLC.