IsanbaevaRR

1.

Digital Transformation of
Corrugated Cardboard
Production Based on
Artificial Intelligence
Isanbaeva Roxana
1

2.

Goals of Digital Transformation
Predictive control
Self-diagnosis and adaptation
Augment people
Shift from reactive automation to
anticipating problems before they
occur.
Build a production system capable
of identifying issues and adapting
its operation.
The goal is not to replace people,
but to extend their capabilities.
2

3.

Objectives Solved by AI
Quality
Automate visual quality control of surface defects, geometry, and glue
lines.
Resource use
Optimize glue and energy consumption.
Equipment
Predict equipment maintenance needs and reduce downtime.
Product grade
Reduce waste and improve product grade.
Business processes
Accelerate order evaluation, logistics, shipment consolidation, and
document search.
3

4.

What AI-Driven Transformation Actually Means
Not just installing sensors – changing the logic of control: data is converted into decisions.
4

5.

Quality Control and Defect Detection
Computer vision
Glue seams
Continuous inspection
Find scratches, spots, and thickness
deviations. Because corrugated-board
surfaces are non-uniform, conventional
algorithms can produce false positives;
hybrid neural networks with attention
mechanisms address this challenge.
IR spectroscopy and line-geometry
analysis can reveal defects in glue
seams that are invisible to the eye.
Detect defects in real time without
stopping the production line.
5

6.

Predictive Maintenance
Signals and records
Analyze equipment vibration and acoustic signals, using
repair histories and operator data for training.
Helios case
The Helios system detected a loose component before
failure, preventing six-figure damage.
Plan maintenance ahead
A model can warn days before a knife or bearing fails,
allowing repairs to be scheduled and avoiding hours of line
downtime.
6

7.

Machine Learning for Forecasting
Support Vector Machines (SVM)
Use for glue consumption and application-thickness control. SVM
suits regression with a small number of features, such as the
relationship between glue viscosity and flow rate.
Random Forest and ensembles
Apply ensemble models to predict board warping where
dependencies are more complex.
Neural networks
Predict paper strength and process parameters, including complex
relationships such as drying regimes.
8

8.

Computer Vision and Sensor Fusion
Surface capture
Hidden seam defects
Fuse sensor data
Line cameras and 3D structured light
inspect board surfaces and geometry.
InGaAs sensors provide infrared
analysis of glue seams.
Hybrid neural networks (MobileNetV3
plus attention) handle textured surfaces,
combining camera views of the surface
with infrared information about hidden
defects.
9

9.

Predictive Analytics and Digital Twins
Monitor equipment signals
Analyze vibration, acoustic, and current signals; use supervised
learning on historical repair data.
Model future behavior
A digital twin is a trained model–not merely a visualization–that
predicts equipment or product behavior from current parameters.
Adjust in advance
The model can predict how the board or machine will behave an
hour from now and adjust operating regimes in advance.
10

10.

Data Infrastructure
Data sources
Connectivity and storage
Bring together PLC, MES, and ERP Use unified exchange protocols
systems, sensors, and laboratory
such as OPC UA and centralized
data.
storage.
Start with what exists
Logs, reports, procurement records,
knife-purchase records, and repair
reports can already reveal patterns;
AI does not always require
expensive IoT sensors.
11

11.

Advantages of AI Adoption
Less waste
Early detection can reduce defects and waste.
Less downtime
Predictive maintenance can reduce equipment downtime.
Resource savings
Optimize glue and energy use.
Faster operations
Speed up order evaluation and logistics.
Retained expertise
Preserve experienced employees’ knowledge in the system.
12

12.

Disadvantages and Barriers
Data quality
People and skills
Cost and integration
Records may be “dirty” or Workforce resistance and Sensors and integration
fragmented.
a lack of competencies can create costs for pilot
hinder adoption.
projects.
Interpretability
Human context
A model may detect an
anomaly without
explaining its cause.
AI does not replace an
experienced operator’s
context: understanding
what is wrong and
deciding what to do
remain human tasks.
13

13.

Conclusions and Recommendations
Start with a real problem
Choose an operational challenge, not a technology purchase.
Pilot with existing data
Use available data to launch a focused pilot.
Keep people involved
Keep humans in the decision-making loop and involve operators.
Prove and scale
Measure results, prove the effect, and scale what works; the entire plant
does not need to be transformed in one step.
14

14.

Thank you for your attention!