DIGITAL TWINS FOR

MECHATRONICS
Introduction to Digital Twins
• Origin of Digital Twin

• The concept of the Digital Twin dates back to NASA's Apollo program in the 1960s.
NASA created identical physical twins of space systems on Earth to simulate and
monitor conditions during missions, especially for diagnostics and support during Apollo
13.

• The formal terminology “Digital Twin” was coined by Dr. Michael Grieves at the
University of Michigan around 2002 in the context of product lifecycle management
(PLM). Since then, it has evolved into a powerful tool in Industry 4.0, enabling real-time
monitoring, simulation, and predictive analytics.
What is a Digital Twin?
• A Digital Twin is a virtual replica of a physical system that is continuously updated with
real-time data and simulates the behavior and performance of its real-world counterpart.

• Definition (Altair, 2023):

“A digital twin is a dynamic virtual model of a physical system that uses simulation,
machine learning, and reasoning to help decision-making.”

• It integrates IoT (Internet of Things), data analytics, and AI/ML to provide insights
throughout the design, operation, and maintenance phases.
Core Components of a Digital Twin

Physical Entity

•The actual object or system (e.g., motor, robot arm, aircraft engine).
•Embedded with sensors for collecting operational data.

Digital Counterpart (Model)

•A virtual model that mirrors the physical system’s behavior.

•Can be built using:
•Physics-based modeling (e.g., finite element models)
•Data-driven or hybrid models

Data & Communication Layer

•Sensor data from the physical entity is streamed in real time.

•Data includes temperature, pressure, vibration, usage, etc.
•Enables feedback loops for control, prediction, or optimization.
DIGITAL MODEL, SHADOW and TWIN
Feature Digital Model Digital Shadow Digital Twin
A digital model that is A dynamic, bidirectional
A static digital
automatically updated model that both receives
Definition representation of a
based on physical system from and sends feedback
physical object or system
data to the physical system
One-way: Physical → Two-way: Physical
Data Flow No data flow
Digital Digital
Real-Time Update No Yes Yes
Full interaction and
Interactivity None Limited (read-only)
control
Virtual model that
Real-time dashboard of a
Example CAD design of a car predicts failure and
production line
triggers maintenance
Monitoring + simulation +
Use Case Design phase Monitoring phase
prediction + control
Feedback Loop None No feedback Closed feedback loop
DIGITAL MODEL, SHADOW and TWIN - Real-world Example
(Industrial Machinery)

• Digital Model: 3D CAD drawing of a machine.

• Digital Shadow: Live sensor dashboard showing temperature and speed.
• Digital Twin: Predictive engine that not only shows data but also simulates wear and
sends control signals to adjust the speed or notify maintenance.
3D and 5D Digital Twins
• A digital twin of an elevator:
Layer Description
Shows what the system is (e.g.,
1D – Descriptive CAD model or static
representation)
Shows what is happening
2D – Diagnostic using real-time data (sensor
dashboards)
Uses simulations and ML to
3D – Predictive
predict what will happen

• Descriptive: CAD model of the elevator

• Diagnostic: Real-time sensor feed showing speed, door status, weight
• Predictive: AI predicts that door motor may fail in 3 days due to abnormal
vibration pattern
3D and 5D Digital Twins
• 5D Digital Twin
• "Descriptive + Diagnostic + Predictive + Prescriptive + Autonomous"

Layer Description
Recommends actions to fix
4D – Prescriptive
or optimize the system
Automatically takes
5D – Autonomous
corrective actions
A digital twin of a smart factory conveyor:

• Prescriptive (4D): Recommends reducing motor speed to

prevent overheating
• Autonomous (5D): Adjusts the speed on its own and alerts
the technician with a report
3D and 5D Digital Twins

Feature 3D Digital Twin 5D Digital Twin

Real-time Monitoring

Predictive Analytics

Recommendations

Autonomous Control

Machine predicts Machine predicts,

Example
failure decides, and acts

Type What it does

3D Twin Predicts engine failure using sensor data and ML
Automatically slows the car, alerts service center,
5D Twin
schedules maintenance
Modeling enabling technologies
• Modeling plays an essential role in digital twins by creating a digital replica of the
physical entity of interest.
• According to the purpose and complexity of different modeling methods, we classify
them into the following five categories.
• Geometric modeling
• Physics-based modeling
Geometric modeling
• Geometric modeling refers to the process of mathematically describing the geometry
(shape and size) of physical objects. It is used to create 3D digital representations that
can be visualized, simulated, or used as part of a digital twin.
Type Description Example
Represents object edges Skeleton of a cube with no
Wireframe Modeling
only using lines and curves surfaces
Defines the surface
Surface Modeling geometry (curves, NURBS) Car body surface in CAD
without volume
Includes complete volume, 3D model of a piston with
Solid Modeling
mass, and material info weight and density
Uses mathematical
formulas to define shapes; A hole in a gear that resizes
Parametric Modeling
can be updated by when diameter is changed
changing parameters
Object represented as a
3D scanned object in
Mesh Modeling collection of polygons
gaming or VR
(usually triangles)
Geometric modeling
• Where Is It Used in Digital Twins?

• Design and CAD Modeling

• Create accurate models of mechanical parts or assemblies
• Tools: SolidWorks, AutoCAD, Fusion 360
• Simulation
• FEM, CFD simulations rely on accurate 3D geometry for stress, heat, and flow analysis
• Mesh generation is critical here
• AR/VR Integration
• Geometrically modeled assets are used for real-time visualization in immersive environments
• 3D Scanning and Reverse Engineering
• Geometric modeling from point clouds (e.g., using LiDAR) to recreate physical objects digitally
Geometric modeling- Laser Scanning
• Laser scanning (also known as LiDAR – Light Detection and Ranging) is a non-contact
3D measurement technique that uses laser beams to capture the geometry of physical
objects or spaces by measuring the time it takes for a laser to reflect off surfaces.

How Does It Work?

• Laser emitter sends out a pulse of light.
• The light hits a surface and reflects back.
• The scanner calculates the distance based on the time-of-flight.
• Thousands to millions of these distance points are collected to form a point cloud.
• This point cloud is converted into a 3D mesh or CAD model for use in digital twins.
Geometric modeling- Laser Scanning

Types of Laser Scanning Applications in Digital Twins

Type Use Case Application Role of Laser Scanning

Terrestrial Laser Scanning buildings, Scanning existing
Scanning (TLS) factory floors, bridges Smart Factory Twin production lines to
Mobile Laser Mounted on vehicles generate 3D models
Scanning (MLS) for road, tunnel scans Creating DTs of bridges,
Mounted on drones Infrastructure Twin tunnels, or urban
Aerial LiDAR or aircraft for environments
mapping terrain 3D capturing of
Portable, used for Heritage Preservation monuments and sites for
Handheld Scanners parts, interiors, virtual reconstruction
equipment Scanning buildings under
Construction Monitoring construction to compare
with BIM
Application Role of Laser Scanning
Physics-based modeling
• Physics-based modeling refers to the use of mathematical models derived from physical
laws (like Newton's laws, thermodynamics, fluid dynamics, etc.) to simulate and predict
the behavior of real-world systems.

• Common Physics-Based Modeling Techniques

Technique Description Example
FEA (Finite Element Simulates structural stress,
Bridge under load
Analysis) strain, vibration
Models airflow,
CFD (Computational Fluid
temperature, pressure in Airflow over car body
Dynamics)
fluids
Simulates movement of
Multibody Dynamics Robot arm simulation
interconnected parts
Models heat flow and
Thermal Modeling Laptop cooling design
dissipation
Physics-based modeling
Levels of Fidelity in Modeling
• Fidelity refers to how closely the model matches reality. There’s a trade-off between
accuracy, computational cost, and real-time usability.

Fidelity Level Description Use Case Pros Cons

Simplified physics,
Quick analysis or
Low Fidelity empirical models, Fast, lightweight Less accurate
real-time simulation
approximations
Combines
Control systems,
simplified physics Balance of speed Still may miss edge
Medium Fidelity digital twins for
with key real-world and realism cases
monitoring
behaviors
Detailed physics,
Structural failure
multiphysics High cost, slow
High Fidelity prediction, High accuracy
coupling (e.g., computation
aerospace, defense
FEA+thermal+fluid)
Data-Driven Modeling
• Data-driven modeling uses measured or simulated data — instead of physical equations
— to learn how a system behaves.
• It typically involves machine learning, statistical models, or deep learning to map input-
output relationships.
• Instead of “first principles”, it learns patterns directly from data — ideal when physics is
complex or unknown.

• Why Data-Driven?
• Suitable when:
• Physical laws are too complex or hard to model (e.g., corrosion, wear).
• Real-time performance is needed (fast inference).
• Rich sensor data is available (IoT environments).
 Data-Driven Modeling

Category Methods Use Case

Linear Regression, SVM, Decision Predict part failure from sensor
Classical ML
Trees readings
Anomaly detection in machine
Unsupervised PCA, K-means, Autoencoders
health
Fault detection from images, time
Deep Learning CNNs, RNNs, LSTMs, GNNs
series, graphs
System Identification ARX, NARX, Kalman Filter Modeling motor response to voltage

Fan Vibration Monitoring Robot Arm Inverse Kinematics

HVAC Energy Prediction
Data: Accelerometer readings Input: End-effector position
Goal: Predict failure Data: Temperature, occupancy, time Output: Joint angles
Model: Decision Tree or LSTM Model: Neural Network Model: Feedforward Neural
Output: Probability of imminent Output: Predicted energy usage Network (learns the inverse
bearing failure mapping)
Data-Driven Models

Challenges
Benefits of Data-Driven Models
Challenge Example
Advantage Detail Poor or missing sensor
Data Quality data leads to poor
Once trained, models
models
Fast execution are very fast (ideal
for real-time twins) Models may memorize
Overfitting training data but fail in
Can learn patterns
new conditions
Adaptable not captured by
physics Harder to explain than
Interpretability
physics-based models
Can handle real-
Works with noisy
world sensor noise Cannot predict outside
data
with proper training No extrapolation training range (unlike
physical models)
Degradation Modeling
• Degradation Modeling
• Degradation modeling is the process of predicting how a system or component deteriorates over
time due to wear, fatigue, corrosion, usage, etc. It’s central to predictive maintenance and
reliability estimation in Digital Twins.
Method Description Example
Use fatigue, wear, corrosion Crack propagation in metal
Physics-based
models under stress
Battery capacity
Use historical failure and
Data-driven degradation prediction
usage data
using ML
Probabilistic modeling (e.g., Estimate remaining useful
Stochastic Models
Wiener, Gamma process) life (RUL) with uncertainty
Convert sensor data into a Turbine vibration health
Health Index Models
“health score” over time degradation index

Jet engine degradation modeled using flight hours + sensor data (temperature, vibration) → RUL prediction →
maintenance alerts.
Surrogate Modeling
• Surrogate models (aka metamodels or emulators) are simplified models that
approximate the behavior of complex, expensive, or high-fidelity models (like CFD or
FEA), enabling faster simulation and optimization.
Method Description Example
Fit a curve to known data Approximating heat
Polynomial Regression
points transfer function
Probabilistic surrogate with Wind turbine performance
Gaussian Process (Kriging)
confidence bounds prediction
Aerofoil lift simulation
Neural Networks Learn complex mappings
under various angles
Compress high-
Reduced Order Models Simulating vehicle crash
dimensional simulation
(ROM) with 1% of original cost
data

Instead of running a full CFD simulation for each engine design variant, a neural net surrogate
model predicts temperature distribution in milliseconds.
Dynamic System Identification
• System identification involves creating mathematical models of dynamic systems based
on measured input-output data — without explicitly knowing the system’s internal
physics.

Category Technique Use Case

ARX, ARMAX, NARX, Identify transfer function of
Parametric
State Space Models DC motor
Impulse response, Audio speaker response
Non-parametric
frequency response testing
Nonlinear systems like
Black-box Neural networks, SVMs
robot arm behavior
Combine known physics + Mass-spring-damper with
Grey-box
learned parameters unknown friction

Steps in System ID
[Link]: Apply known input (e.g., voltage) and record output (e.g., motor speed)
[Link] Selection: Choose ARX, state space, etc.
[Link] Estimation: Fit model to data (least squares, Kalman filter)
[Link]: Simulate and compare with test data
Comparison

Concept Purpose Common Use Example

Battery aging, bearing
Degradation Modeling Predict component aging Predictive maintenance
wear
Replace expensive
Surrogate Modeling Optimization, control Aero drag prediction
simulations
Learn dynamic behavior Robot joint motion
System Identification Control, modeling
from data prediction
Physics-Informed Machine Learning (PIML)
• Physics-Informed ML integrates physical laws (like conservation of energy, Newton’s
laws, etc.) into machine learning models either by embedding them in the architecture,
loss function, or training process.

Why?
• Many real-world systems have scarce data but well-known physics.
• Pure ML may overfit or be unphysical.
• Physics-based models are interpretable but slow or incomplete. (PIML bridges the gap.)
Physics-Informed Machine Learning (PIML)
Categories & Approaches of PIML

• Physics-Informed Neural Networks (PINNs)

Feature Description
Embed PDEs (Partial Differential
Key Idea Equations) directly into the loss function of
a neural network.
Network learns both from data and from
Training
the physics it must obey.
Loss = Data Loss + PDE Residual Loss

Instead of just training on temperature data, the NN is forced to satisfy the heat equation ∂T/∂t =
α∇²T during training.
Physics-Informed Machine Learning (PIML)
Categories & Approaches of PIML

• Hybrid Models (Gray-Box)

Feature Description
Add physics-based constraints or penalties
Key Idea
to regularize the model.
Helps guide the learning process and
Usage
avoid physically impossible results.

Example: Learning a pendulum trajectory with a neural net, while penalizing deviations from
energy conservation (i.e., total energy should remain constant).
Physics-Informed Machine Learning (PIML)
Categories & Approaches of PIML

• Data Augmentation Using Simulations

Feature Description
Use synthetic data from physics-based
Key Idea simulators to augment small real-world
datasets.
Improves generalization where real data is
Benefit
scarce.

Example: Generate fluid flow data from CFD simulations to train a neural network that can
generalize to new shapes or conditions.
Physics-Informed Machine Learning (PIML)
Categories & Approaches of PIML

• Learning Residuals
Feature Description
Use physics-based model for coarse prediction
Key Idea and train an ML model to learn the error
(residual) between model output and real data.
“Model discrepancy correction” or “delta
Also Called
learning”

Example: A thermodynamic simulator predicts engine temperature with 90% accuracy. A neural
network learns the 10% residual from real-world sensor data.
Physics-Informed Machine Learning (PIML)
Transfer Learning

• Transfer learning is the process of reusing knowledge (like model weights or features) from one
task or domain to accelerate learning in another, related task — especially useful when target data
is scarce.

• Suppose you have:

• A deep neural network trained to predict turbine blade temperature in a large jet engine.
• You now want to predict the same for a smaller, similar engine with limited data.
• Instead of training from scratch, you fine-tune the pretrained model using the smaller dataset —
saving time and improving generalization.

Use in Digital Twins:

• Transferring learned representations from one physical twin to another
• Adapting models across machines, operating conditions, or even domains (e.g., from simulation to
reality)
Physics-Informed Machine Learning (PIML)
Delta Learning
• Delta learning refers to learning the "difference" or "residual" — either between models
or domains.

Delta Between Physics Model and Reality (Model Correction)

• What: Learn the error (Δ) between a physics-based model and real-world observations.
• Why: Physics models may miss certain dynamics or degrade over time.
• Example: A thermal simulation underestimates the heat in a motor by 5%. Train a model to predict
this residual error (delta) and correct the prediction.

Final output = Physics model + ML-predicted delta

Physics-Informed Machine Learning (PIML)
Delta Learning
• Delta learning refers to learning the "difference" or "residual" — either between models
or domains.

Delta Between Source and Target Domain (Domain Adaptation)

• What: When simulation and real-world data differ, learn the “delta” between them.
• Why: Simulation-trained models may not transfer directly to real-world (Sim2Real gap).
• Example: You train an AI model to detect faults using synthetic (simulated) data. Then, you use
delta learning to learn the difference between simulation and real sensor readings, making the
model robust to real-world noise.
Physics-Informed Machine Learning (PIML)
• ML-Assisted Prediction
A hybrid approach where Machine Learning (ML) is used to assist traditional first-principles
(physics-based) models by predicting their inputs or parameters, rather than replacing them entirely.
This method leverages the strengths of both ML (data-driven pattern recognition) and physics-based
models (interpretability and consistency with known laws).

How ML-Assisted Prediction Works

• Identify Predictive Variables:
Choose the inputs (x) or parameters (θ) of the physics model that are difficult to measure directly or
evolve with time.
• Train an ML Model:
Use historical or real-time sensor data to train a model to estimate or forecast these inputs or
parameters.
• Use ML Output in Physics Model:
Feed the ML-estimated values into the physics model to compute the desired output response (y), ,
e.g., stress, temperature, Remaining Useful Life (RUL), etc.
Physics-Informed Machine Learning (PIML)
Example

• Battery Health Prediction

• Challenge: Directly measuring battery degradation parameters (like internal resistance or lithium
diffusion coefficients) is hard in real-time.
• ML-Assisted Prediction:
• ML predicts degradation parameters using voltage/current data.
• The physics-based model uses these parameters to simulate the battery’s future behavior and
predict remaining capacity or failure time.
Physics-Informed Machine Learning (PIML)
Physics-Informed Machine Learning (PIML)
Why Use PIML in Digital Twins?

Benefit Description
Combines physics-based realism with data-
Accuracy
driven flexibility
Data Efficiency Learns well even from small datasets
Generalization Better extrapolation to unseen scenarios
Interpretability Models remain grounded in physical reality

Example- Modeling a falling object with air drag:

• Physics: F = mg - kv
• Neural network learns velocity over time
• PINN enforces that the learned function satisfies Newton's 2nd law throughout the training
domain
System modeling
• System modeling involves the representation of the functional and structural relationships
between various components of a complex system — such as machines, sensors, control systems,
and humans — using standardized modeling languages or conceptual frameworks.
• Unlike geometry (shape) or physics-based (behavior) models, system models define how different
parts interact, their information flow, dependencies, and logic — often without detailed math or
geometry.

Key Approaches to System Modeling

1. Unified Modeling Language (UML)
• A general-purpose visual modeling language, originally developed for software design.
Feature Description
Example: Model the
Type Graphical modeling language workflow of a factory
Captures logic, events, and data flow in assembly line using
Purpose UML activity diagrams
systems
to visualize steps and
Class diagrams, activity diagrams, decision logic.
Popular Diagrams
sequence diagrams
System modeling
Systems Modeling Language (SysML)
• A more specialized version of UML, tailored for engineering systems.

Feature Description
Type Extension of UML for systems engineering
Model structure, behavior, requirements,
Purpose
and constraints of physical systems
Block Definition Diagram (BDD), Internal
Popular Diagrams
Block Diagram (IBD), Parametric Diagram

• Example: In an autonomous car system, SysML is used to define sensors, controller logic,
communication with actuators, and environmental constraints.
System modeling
Ontology-Based Modeling
Used for semantic modeling and defining relationships in terms of concepts and logic rather than
behavior.
Feature Description
Knowledge
Type
modeling
Define taxonomy
Purpose and semantic
relationships
Healthcare DTs,
Application smart cities, IoT
networks

• Example: A hospital twin uses ontologies to relate patients, equipment, procedures, and locations
meaningfully for queries like “Which ICU beds are suitable for a patient with infection?”
System modeling
Why System Modeling Is Important in Digital Twins

• Defines the architecture of complex systems before detailed modeling.

• Ensures clarity in component interaction and traceability of requirements.
• Facilitates co-simulation, i.e., integrating multiple models across tools.
• Enables integration with MBSE (Model-Based Systems Engineering) workflows.
Physical-to-virtual (P2V) twinning enabling technologies
• P2V refers to the technologies and methods used to update and synchronize digital models using
data from the physical system. This is what enables a Digital Twin to reflect the real-time or near-
real-time state of the physical world.

1. Physical Measurements as Input

• Description: Use sensor measurements (e.g., temperature, GPS, pressure) as direct input to the
digital model.

Subtypes Example
In smart factories, IoT-based location data
a. Motion/Position Updating updates digital models of machine
movements.
Real sensor data drives simulations, e.g.,
b. Physics-based Input using in-flight data for aircraft structure
risk analysis.
Physical-to-virtual (P2V) twinning enabling technologies
Probabilistic Model Updating
• Description: Use uncertainty-aware methods (like Bayesian filters) to update internal digital states
which are not directly observable.

Key Techniques Use Case

Estimating battery State of Health (SoH),
Kalman Filter, Particle Filter
UAV structural health
Evolving system parameter tracking under
Dynamic Bayesian Networks
uncertainty

• Digital State: Internal parameters like resistance, stress, temperature that evolve over time and
require estimation.
Kalman Filter
• The Kalman Filter is an algorithm used to estimate the state of a system over time —
especially when the measurements are noisy or uncertain. It’s widely used in applications
like navigation, tracking, and robotics.

Example: Estimating the Position of a Car

Scenario:
You’re tracking a car driving in a straight line. You want to estimate its position over time.
You have:
• A GPS sensor that tells you the car's position every second.
• But the GPS is noisy, it doesn't give the exact position.
You know:
• The car moves at a constant speed (say, 60 km/h = 16.67 m/s).
• GPS gives a measurement every second, but it's off by ±5 meters sometimes.
Kalman Filter
Goal:
• Estimate the true position of the car using:
• Noisy GPS readings
• Your knowledge of how the car is supposed to move.

Step-by-Step Kalman Filter (simplified):

Predict the next position:

You know the car was at position X and moves at a constant speed.
• So, after 1 second, your predicted position is: Predicted Position = X + (Speed × Time)
Measure:
• You get a noisy position from the GPS, say: Measured Position = X_measured
Kalman Filter
Update (combine prediction and measurement):
• You take a weighted average of your prediction and the measurement:
Estimated Position = Prediction × (1 - K) + Measurement × K

Where K is the Kalman Gain:

• If GPS is very noisy, trust the prediction more → small K
• If GPS is accurate, trust the measurement more → large K

Kalman Gain is calculated based on:

• Uncertainty in your model (how much you trust your prediction)
• Uncertainty in your measurement (how noisy GPS is)
Kalman Filter
Particle Filter (Monte Carlo Localization)
• A Particle Filter (also known as Monte Carlo Localization) is another way to estimate the
state of a system over time. (better suited to nonlinear systems or non-Gaussian noise)

Revisited: Tracking a Car with a Particle Filter

Problem:
• You want to estimate the car's position over 30 seconds, using noisy GPS measurements.
But now, instead of assuming perfect models or Gaussian noise like in Kalman, you’ll use
a swarm of particles.

Basic Idea:
• A particle is just a guess of where the car might be.
• You simulate hundreds or thousands of particles.
• Each particle has:
• A position
• A weight (how likely it is to be correct)
Particle Filter Steps
• Initialize:
Generate a bunch of particles with random positions.
• Predict (Motion Update):
Move each particle based on the car's motion (e.g. speed + noise).
• Update (Measurement Update):
For each particle, compare its position to the GPS measurement.
Give it a higher weight if it matches well (i.e. it's close to the GPS).
• Resample:
Select particles with higher weights to create the next generation of particles.
Discard particles with low weights.
• Estimate:
The estimated position is the weighted average of all particles.
• Repeat for every second.
Particle Filter
Physical-to-virtual (P2V) twinning enabling technologies
• ML Model Updating
Description: Update or adapt data-driven models (like neural networks) based on new incoming
data.
Method Use
Adapting turbine RUL models or anomaly
Online learning, Transfer learning
detectors on-the-fly

Advantage: Keeps ML models aligned with time-varying system behavior (e.g., drift, sensor wear).

• Fault Diagnostics and Failure Prognostics

Description: Use AI or statistical models to detect faults (diagnostics) and predict failures
(prognostics).
Techniques Tools
Decision trees, CNNs, RNNs, SVMs Vibration data, current draw, thermal data

Example: Predicting when a pump or bearing might fail in an industrial machine.

 Physical-to-virtual (P2V) twinning enabling technologies
• Diagnostics vs. prognostics
Aspect Diagnostics Prognostics
Identifies if a fault has already Predicts when a fault will occur in
Definition
occurred the future
Estimates remaining useful life
Purpose Detects and classifies current faults
(RUL) or time to failure
Timing After or during fault onset Before fault actually occurs
Predicted time/date of failure or
Output Type, location, and severity of fault
degradation level
Rule-based systems, decision trees, Time series forecasting, RNN,
Tools
SVM, CNN LSTM, survival models

Diagnostics: Prognostics:

• Sensors detect abnormal vibration. • Based on degradation trends (vibration + temperature +

• A machine learning model analyzes this and identifies: past behavior), a predictive model estimates:
“Pump bearing failure – early-stage degradation detected.” “Bearing is likely to fail in 16 days with 90% confidence.”
Action: Flag maintenance need immediately. Action: Plan maintenance next week before failure.
Physical-to-virtual (P2V) twinning enabling technologies
• Ontology-Based Reasoning
Description: Use knowledge graphs and ontologies to represent the structure and logic of the system
semantically.
Benefits Applications
Semantic interoperability, dynamic Smart manufacturing, modular systems,
reconfiguration asset management

Example: In a factory, ontologies track which machines are connected, what data they produce, and how
they are linked, even when reconfigured.
P2V Method Purpose Example
GPS, pressure sensors in smart
1. Physical Measurements Real-time model input
cities
2. Probabilistic Updating Estimate hidden states Battery degradation state
3. ML Model Updating Adapt to new data Update NN after sensor drift
4. Fault Diagnostics/Prognostics Predict and prevent failures Turbine failure prediction
Factory reconfiguration,
5. Ontology Reasoning Semantic logic and structure
component linking
Virtual-to-Physical (V2P) Twinning Enabling Technologies
• These technologies enable a Digital Twin to influence its physical counterpart, closing
the feedback loop by sending optimized actions, maintenance schedules, or control
commands from the virtual model to the real system.
Virtual-to-Physical (V2P) Twinning Enabling Technologies
Model Predictive Control (MPC)

• Definition: MPC is an optimization-based control

technique that uses a digital model to predict the future
state of a system and make decisions to achieve an
objective under constraints.
• Working Principle:
A. Predict future behavior of a process using a model:
B. Optimize a cost function 𝑚𝑖𝑛 σ𝒌𝒕=𝟎 𝑱(𝒙 𝒕 , 𝒖 𝒕 ) over a
prediction horizon (e.g., minimize deviation from a
reference.
C. Only the first control action is implemented before
repeating the cycle — hence also called receding horizon
control.
Virtual-to-Physical (V2P) Twinning Enabling Technologies
Temperature Control with MPC
Imagine a room where we want to maintain the temperature at 22°C by controlling a heater.

Objective:
• Predict future temperature based on current heater setting
• Choose the best sequence of heater inputs (e.g., on/off levels) over a prediction horizon
• Apply only the first control action, then repeat at next step
Virtual-to-Physical (V2P) Twinning Enabling Technologies
Temperature Control with MPC
A simple thermal model: 𝑻𝒌+𝟏 = 𝑻𝒌 + 𝝈. 𝝁𝒌 − 𝑻𝒌 − 𝑻𝒂𝒎𝒃𝒊𝒆𝒏𝒕

𝑻𝒌 − 𝒄𝒖𝒓𝒓𝒆𝒏𝒕 𝒓𝒐𝒐𝒎 𝒕𝒆𝒎𝒑𝒆𝒓𝒂𝒕𝒖𝒓𝒆

𝝁𝒌 − heater power (control input, between 0 and 1)
𝝈 − heat transfer coefficient

MPC Key Parameters

• Prediction horizon: 10 steps
• Objective: Minimize the difference between predicted temperature and target
• Constraints 0 < 𝝁𝒌 < 𝟏
Virtual-to-Physical (V2P) Twinning Enabling Technologies
• Temperature Control with MPC

• Step 1: 𝑻𝟏 = 𝑻𝟎 + 𝟎. 𝟏. 𝟎. 𝟖 − 𝟏𝟓 − 𝟏𝟎 = 14.58 (𝝁𝒌 = 𝟎. 𝟖, Assuming 𝑻𝟎 = 𝟏𝟓)

• Step 2: 𝑻𝟐 = 𝟏𝟒. 𝟓𝟖 + 𝟎. 𝟏. 𝟏 − 𝟏𝟒. 𝟓𝟖 − 𝟏𝟎 = 14.2 (𝝁𝒌 = 𝟏, )
• Step 3: 𝑻𝟑 = 𝟏𝟒. 𝟐 + 𝟎. 𝟏. 𝟏 − 𝟏𝟒. 𝟐𝟐 − 𝟏𝟎 = 13.8 (𝝁𝒌 = 𝟏, )
Virtual-to-Physical (V2P) Twinning Enabling Technologies
• Applications:
• Refining, petrochemical, thermal building systems
• Integrated with ML models for robot manipulation (e.g., DL + MPC for trajectory optimization)

• Example:
In a smart HVAC system, MPC predicts future temperatures and adjusts fan speeds and cooling
intensities in advance to minimize energy consumption and maintain comfort.
• Enhanced with ML:
Future MPCs will likely use deep learning architectures as process models instead of physics-based
models — improving speed and adaptability
Virtual-to-Physical (V2P) Twinning Enabling Technologies
• Predictive Maintenance

• Definition: Predictive Maintenance anticipates when equipment will fail and prescribes timely
maintenance, reducing unplanned downtime.
• Steps:
1. Sensor Deployment (P2V): Vibration, temperature, etc.
2. Health Estimation: Use ML/physics models to evaluate system degradation
3. RUL Prediction: Estimate Remaining Useful Life
4. Action (V2P): Trigger alerts, schedule maintenance just before failure

• Examples:
• Baker Hughes uses a MATLAB-based neural network for pump health monitoring
• ANSYS & PTC ThingWorx predict electric motor failure using sensor data
Predictive Maintenance Demo
Virtual-to-Physical (V2P) Twinning Enabling Technologies

An overview of an ML pipeline for predictive maintenance

An IIoT platform for predictive maintenance of industrial assets

Virtual-to-Physical (V2P) Twinning Enabling Technologies
• Real-Time Perception and Decision Making

• Challenge: Real-time simulation and optimization are computationally heavy.

• Solution: Use ML to:
1. Speed up simulations (e.g., fire propagation modeling using ML for sub-model selection —
Zohdi, 2020)
2. Create optimization proxies: ML replaces slow solvers with fast approximations (predict-then-
repair)

• ML Advantages:
• Near real-time inference from sensor data
• Fast response for time-sensitive systems like autonomous robots or predictive control loops.

• Example:
An ML model predicts how a fire will spread inside a building using current sensor readings. This
informs the building's suppression system in real time
Digital Twin Architecture &
Workflow
Digital Twins for Mechatronics
5-Layer Digital Twin Architecture

Layer Description Example

Collects raw data from sensors, Vibration data from motors, GPS in
1. Data Layer
devices, and logs vehicles
Ensures real-time, secure, and
2. Integration Layer MQTT, OPC UA, REST APIs
standardized data transfer
Represents physical systems using
3. Modeling Layer physics-based, ML, or hybrid FEA of machines, battery SoC model
models
Performs predictive maintenance,
4. Analytics Layer RUL prediction using ML
anomaly detection, optimization
User interfaces for interacting with Dashboards, 3D models, AR/VR
5. Visualization Layer
the twin interfaces
 Common Protocols &
Layer Function Data Formats / Headers Tools / Platforms
Standards

Data collection from - Modbus (RTU, TCP)- CAN Sensor-specific headers:

Arduino, Raspberry Pi,
1. Data Layer sensors, machines, Bus- Zigbee / Bluetooth LE- time-stamp, device ID,
PLCs, NI LabVIEW
environments OPC-UA (sensor standard) units

- MQTT (lightweight pub-

Node-RED, HiveMQ,
Communication, data sub)- CoAP (constrained JSON, XML, Protocol
2. Integration Layer Mosquitto, Kafka, AWS
routing, streaming networks)- HTTP/HTTPS- Buffers (for gRPC)
IoT Core
WebSocket, gRPC

- Simulink (co-simulation)-
FMU/FMI (Functional Mock- ANSYS TwinBuilder,
Creating digital replicas FMU metadata headers,
3. Modeling Layer up Interface)- Python APIs COMSOL,
via simulation or ML ONNX model formats
(PyTorch, TensorFlow)- ROS OpenModelica, PyTorch
(for robotics)

Predictive models, - REST APIs for ML services- scikit-learn, MLflow,

CSV, Parquet, SQL
4. Analytics Layer anomaly detection, ONNX for model exchange- Azure ML, TensorFlow
queries, Inference APIs
optimization Python/R scripts Serving

Human-machine - HTTP/HTTPS- WebSocket

5. Visualization HTML/CSS/JS3D Grafana, Power BI, Unity
interface, AR/VR, (for live data)- Unity (for
Layer formats: glTF, FBX, OBJ 3D, Plotly Dash, [Link]
dashboards AR/VR)
Digital Twin Architecture & Workflow
• Digital Thread & Data Flow

• Digital Thread = Seamless flow of data across the lifecycle of a physical asset.
• Key Concepts:
1. Connects design → manufacturing → operation → maintenance → disposal
2. Allows continuous feedback loop between product and model
3. Enables traceability, version control, and decision-making at all stages

• Example: Design updates from simulation directly feed into manufacturing settings or predictive
algorithms.
Cloud Deployment
• Cloud deployment refers to running the Digital Twin’s components — data storage, modeling,
analytics, and visualization — entirely on remote cloud infrastructure provided by platforms like
AWS, Azure, or Google Cloud.

• Key Idea:
• The physical system sends data to the cloud, and all computation (modeling, simulation,
predictions) happens remotely.

• Core Concepts:
❑ Centralized processing: All data, modeling, and visualization happens in remote cloud servers.
❑ Scalable infrastructure: Auto-scaling compute and storage resources.
❑ Global access: Ideal for distributed assets or multi-site systems.

• Example:
• A wind farm sends turbine data to the Azure Cloud, where a Digital Twin simulates blade fatigue
and visualizes performance trends for remote operators.
Cloud Deployment- Technical Specifications

Feature Details
Data flow Sensor data → Gateway → Cloud via MQTT/HTTPS
Storage Cloud databases (e.g., AWS S3, Azure Blob, BigQuery)
AI/ML models on cloud GPUs/CPUs (e.g., SageMaker,
Model Hosting
Vertex AI)
Streaming Apache Kafka, AWS Kinesis, Azure Event Hubs
APIs RESTful APIs, gRPC for microservices
Security TLS encryption, IAM roles, data encryption at rest
Visualization Dashboards via Grafana, Power BI, or custom web UIs

[Device] → [Gateway/Edge] → [Cloud Ingestion Service] → [Cloud Storage + ML Model] →

[Dashboard or Control System]
Cloud Deployment - Limitations
• High Latency
1. Cloud servers are often geographically distant from the physical system.
2. Even small round-trip delays (100–300 ms) are too slow for real-time control.
3. Not suitable for robotics, autonomous vehicles, or manufacturing automation.

• Network Dependence
1. Requires constant, reliable internet connection.
2. In remote or hostile environments (e.g., oil rigs, rural factories, ships), connectivity is unstable or
expensive.

• Bandwidth Cost & Congestion

1. Continuously sending high-frequency data (e.g., 50 sensors at 1 Hz) to the cloud increases:
1. Operational costs
2. Network congestion
2. Also introduces a bottleneck when multiple devices compete for upload bandwidth.
Cloud Deployment - Limitations
• Data Privacy & Security Concerns
• Transmitting sensitive data (e.g., healthcare, defense, proprietary manufacturing) over the internet
increases exposure to cyber threats.

• Limited Offline Functionality

• If the cloud is unreachable (due to firewall, downtime, or outage), most cloud-only systems stop
functioning.
• Mission-critical systems (like smart grids or autonomous machines) cannot tolerate such failures.

• Latency in Actuation
• Cloud inference means slow response times for decisions → unsuitable for:
• Collision avoidance
• Instant machine shutdown
• Feedback control loops (<50 ms)
Emergence of Edge Deployment
Challenge Edge Response
Latency Compute locally near the device → faster reaction
Network failures Operates offline or with intermittent sync
Data privacy Keeps sensitive data on-site
Cost Reduces need to send all raw data to the cloud
Control Enables real-time decisions and actuation
Allows operation in isolated, mobile, or dangerous
Autonomy
environments

Autonomous Drone

• Cloud-only: Too slow to detect obstacles → crashes.

• Edge: Onboard camera + AI model detects obstacle →
instantly changes path.
Edge Deployment
• Edge deployment means hosting the Digital Twin (or part of it) locally on or near the physical
asset, such as on an embedded device, PLC, or industrial gateway.

• Key Idea:
• Instead of sending all data to the cloud, the system processes and reacts to data right at the edge,
enabling fast, offline, or bandwidth-constrained operations.

• Core Concepts:
• Local computing close to the physical system.
• Real-time processing with low latency.
• Reduced cloud dependency, useful in remote or offline environments.

• Example:
• A robot arm runs a real-time twin on a NVIDIA Jetson board, which immediately detects vibration
anomalies and stops the arm if needed — without needing cloud access.
Edge Deployment- Technical Specifications

Feature Details
Data flow Sensor data → Edge processor → Local inference
Industrial gateway, PLC, embedded PC, Raspberry Pi,
Compute device
NVIDIA Jetson
ML Inference TensorRT, OpenVINO, ONNX Runtime on-device
Communication Local MQTT broker, OPC-UA server, BLE/Zigbee
Storage SQLite, flat files, small time-series DBs
Security TPM chips, firewalled LANs, local certificates
Control Direct actuator control via GPIO, CAN, RS-485

[Sensor] → [Edge Device] → [Local Processing + Inference] → [Immediate Action or Optional Cloud
Sync]
Edge Computing Framework (Architecture Overview)

Layer Description
Sensors, embedded systems, or machines (e.g.,
1. Edge Devices
turbine sensors, Raspberry Pi)
Local compute node that runs analytics, filtering,
2. Edge Gateway / Node
ML inference
Optional: Near-edge servers for aggregation,
3. Fog Layer
buffering, orchestration
Central system for heavy storage, training,
4. Cloud
dashboards, and system updates
Edge Computing Framework

Cloud Platform
(model training and
Storage)

Fog Node/ Gateway Edge Node Collect

(message broker, data, runs of ml
Edge Devices
buffering, load (pre-trained model),
balancing) send alerts
Hybrid Deployment
Feature Details
Data collection, local ML inference,
Edge role
immediate response
Historical analysis, visualization,
Cloud role
retraining
Data sync MQTT + buffering (e.g., Redis, EdgeHub)
OTA (Over-the-Air) deployment from
Model updates
cloud to edge
Cloud-edge synchronization (eventual or
State consistency
strong)
Local & cloud IAM, encryption, version
Security
verification

[Sensor] → [Edge Inference] → [Control Action]

↓
[Periodic Sync] → [Cloud Analytics + Retraining]
↓
[Updated Model → Edge Device]
Comparison

Feature Cloud Edge Hybrid

Latency Medium to High Very Low Balanced
Bandwidth Use High Low Optimized
Scalability High Device-limited High
Reliability (offline) Low High Medium
Maintenance Easy (centralized) Harder (device-specific) Managed
Use Case Global analytics Real-time control Most industrial DTs