RPDS AI · USER GUIDE

Technical Documentation

Understand the concepts, prepare your data, and follow a clear path from process measurements to useful insights.

Process analyticsModel developmentBatch optimisation

THE ESSENTIALS

Getting started

Begin with a specific process question. Decide what you want to estimate, compare, or improve before choosing a workflow.

Prepare your data

Explore measurements and resolve data quality issues.

Open Data Cleaning

Estimate a variable

Build a model using the process inputs you can measure.

Open Soft Sensor

Compare your batches

Establish a reference and investigate process deviations.

Open Golden Batch

A useful starting point

Define the target variable, its unit, and the outcome that would make the analysis useful. For example: estimate final product concentration from temperature, pH, and feed rate.

BEFORE YOU BUILD

Data preparation

Reliable analysis starts with consistent measurements. Check the dataset before selecting features or training a model.

Confirm the structure. Use meaningful column names and consistent units. For batch analysis, clearly define batch identifiers and the time or progress axis.
Review data quality. Inspect missing values, duplicate records, and unusual readings. Investigate outliers before deciding whether to remove them.
Select the target and inputs. Choose the output to estimate and the measurements available at the time a prediction will be made.
Keep evaluation data separate. Avoid using information from the test set during preprocessing or feature selection. For time-dependent data, preserve the time order.

Check for data leakage

An input recorded only after a batch finishes cannot support a prediction made during that batch. Choose inputs that match the intended prediction point.

PROCESS REPRESENTATION

Digital twin

A digital twin is a virtual representation of a physical system or process. It can combine measured data, mathematical models, and machine learning to describe process behaviour.

When to use it

Use process models to study relationships between operating conditions and outcomes, or to explore scenarios before making a plant change.

What it helps you understand

How process inputs affect performance.
How operating scenarios compare.
Where observed behaviour differs from the model.

PROCESS EXAMPLE

A reactor model can estimate how temperature and feed conditions influence reaction performance. Compare its predictions with measured results before using it to guide decisions.

The RPDS model workflow provides tools for exploring data and developing predictive models. A live digital twin also requires appropriate plant data integration and validation.

Explore the model workflow

FUTURE OUTCOMES

Forecasting

Forecasting uses historical observations to estimate future values. The forecast horizon defines how far ahead the prediction is intended to apply.

Common applications

Estimate future production or demand.
Anticipate energy consumption.
Predict process yield or batch duration.

Recommended approach

Define the prediction horizon, preserve the sequence of observations, and evaluate predictions against later periods that were excluded from training.

PROCESS EXAMPLE

Use historical operating conditions to estimate the yield of a future production cycle. Compare predictions with a simple baseline to establish whether the model adds value.

Match the model to the question

Estimating a value from current inputs and forecasting a future value are different tasks. Time-based forecasts require an appropriate time structure and evaluation strategy.

REFERENCE & COMPARISON

Golden batch

A golden batch is a reference production run, or a set of runs, selected for desirable quality and process performance. Its profile provides a benchmark for comparing other batches.

Follow the workflow

Upload and inspect. Review the batch data and choose the target variable.
Define the reference. Select and lock the golden batches using your process and quality criteria.
Review statistics. Examine reference profiles, variability, and relevant process variables.
Compare and investigate. Review similarity and deviations, then use process context to interpret the differences.

PROCESS EXAMPLE

For a fermentation process, compare temperature, pH, and feed profiles against successful reference batches. Investigate where a lower-yield batch departs from the reference.

A close match to a reference profile does not by itself establish product quality. Interpret comparisons alongside measured outcomes and operating conditions.

Open Golden Batch

INFERENTIAL MEASUREMENT

Soft sensors

A soft sensor is a computational model that estimates a process variable from other available measurements. It is useful when the target is difficult to measure continuously or requires laboratory analysis.

Choose the inputs

Select measurements related to the target, such as temperature, pressure, pH, or flow rate. Ensure they are available when you intend to make a prediction.

Validate the estimates

Compare predictions with measured target values on held-out data. Review the error across operating conditions, rather than relying on a single overall score.

PROCESS EXAMPLE

Estimate biomass concentration from pH, dissolved oxygen, and temperature. Laboratory measurements provide the reference values for training and evaluating the model.

Stay within validated conditions

Review model performance when feedstocks, equipment, or operating ranges change. Predictions outside the training range may be less reliable.

Open Soft Sensor

INTERPRETING RESULTS

Reading your graphs

Choose a graph that matches your question. Always check the variable names, units, and data range before interpreting a pattern.

Quick reference for common analytical charts
Graph	Use it to examine	What to look for
Histogram	The distribution of one numerical variable.	Spread, skewness, multiple peaks, and unusual values. Bin size can change the apparent shape.
Time-series plot	How a variable changes over time or batch progress.	Trends, shifts, cycles, and differences in timing. Compare batches on a consistent time basis.
Scatter plot	The relationship between two numerical variables.	Patterns, clusters, and outliers. A relationship does not establish cause and effect.
Correlation heatmap	Pairwise relationships among numerical variables.	Strong associations and redundant inputs. Correlation may miss nonlinear relationships.
Predicted vs. actual	Agreement between model estimates and measured values.	Distance from the agreement line, systematic bias, and errors at high or low target values.

Using a histogram

Choose a numerical variable, such as temperature or yield, and confirm its unit.
Review the frequency of observations in each interval, or bin.
Compare the distribution across relevant batches or operating conditions.
Investigate unusual values using the original records and process context.

Ready to work with your process data?

Explore RPDS workflows