MoLe aims to facilitate the development and use of digital twins for smart factories, so factory stakeholders can enjoy all the benefits of digital twin technologies with little effort.

1. Speaker: Mauricio Fadel Argerich
NEC Laboratories Europe GmbH
Model Learning for Cloud-Edge
Digital Twin

2. MoLe: Model Learning for Cloud-Edge Digital Twin
• With MoLe we aim to simplify the
implementation and execution of Digital Twins
in digital factories by:
• Utilizing FIWARE technologies (NGSI-LD, Scorpio,
FogFlow) to dynamically orchestrate a setup to
generate a Digital Twin from data and execute it
on Edge and Cloud
• Reducing effort needed to develop prediction and
simulation models by using Knowledge Infusion

3. I4.0Lab – Manufacturing & Assembly process
• We have developed our solution for the MIDIH
Didactic Factory in Milan
• The factory implements the Manufacturing and
Assembly process of PCBs
• 7 steps, carried out by different stations

4. Translator and Scorpio
• Stations in didactic factory use NGSIv2, Scorpio
uses NGSI-LD
• Translator reads sensor data from the Factory
Information Bus and uses Kafka information (topic,
key, message schema) to transform it into NGSI-LD
entities
• Data in NGSI-LD is received by the Scorpio Broker,
where it is stored and sent to any subscribers
• Scorpio optimizations:
• We optimized data handling and serialization, as well as
error handling
• We optimized Scorpio’s vertical scaling and internal data
handling
Translator
DB
Subscriber
Subscriber
Subscriber
Subscriber
Scorpio

5. Results: Translator and Scorpio
Translator
• Worst case delay of 1ms between a message arrival
in the Kafka bus and the retrieval in the Python
library
• Below KPI of 2Hz (sensors sampling frequency)
Scorpio
• We made a comparison with the updated version of
the FIWARE GE Orion called Orion-LD
• Scorpio achieved in average half the latency of Orion
Scorpio
Orion

6. Digital Twin Models
• We implemented Digital Twins (DTs) for the
different stations in the M&A process
• DTs are programmable objects that can be
instantiated for real-time monitoring and
simulations
• DTs of Front Cover Magazine and the Press Station
• These DTs implement specific models to:
• detect the current status of the station based on its
current sensor and actuators data
• predict energy usage based on the same data

7. Knowledge Infusion
• Energy usage prediction à pure ML
• State inference model à Knowledge Infusion (KI) = ML + domain knowledge
• Domain knowledge is infused through Knowledge Functions (KFs)
• KFs output a single value: a label.
• Functions implement human provided logic and utilize facts derived from internal and
external knowledge bases
• Types of knowledge functions: Weak and Strong
• KI creates a Knowledge Model that serves two purposes:
• Data augmentation: improves data quality by creating new features or labelling data
• Robustness: allows us to correct some obviously wrong outputs of the ML model during runtime
ML
Model
Knowledge
Model
KI Model

8. FogFlow
• FogFlow was extended to serve ML models as serverless fog functions
• ML models are implemented based on FogFlow ML operators
• To implement a FogFlow ML model, we follow 3 steps:
1. Model registration: to register a ML through a web-based GUI
2. Model deployment: to create and deploy a serverless fog function in FogFlow to run the ML model
3. Model serving: to apply the ML model inside the deployed function instances to produce the
detection/prediction result of the input data and then update the state of the corresponding DT

9. Results: Digital Twin Models
• Model for the station’s state
• Small high quality dataset: manually labeled 10m of
data (2Hz, 1200 data points total) as “Idle” or
“Working”
• Larger noisy dataset: around 50m or 6000 data points,
labelled with 2 simple functions
• We used the small dataset to train a Random Forest
Classifier (RFC) using all the features (45)
• First 600 samples to train RFC and last 600 to evaluate it
• The RFC achieved an accuracy of 82.11% on its test set
• We also implemented a Knowledge Model (KM),
based on two programmable functions
• Each function took no more than 5 minutes, they check
variable values and return state of the machine
• Test accuracy is 70%
Performance of RFC on test set when trained
with 50% of manually labeled data
Performance of KM on test set with 2 simple
labelling functions
S
L
train
test
à

10. Results: KI for Digital Twin Models
• We can utilize KM to label larger unlabeled dataset and
train a ML model with it
• Hopefully, the ML will learn to filter out the noise
• The ML model is trained with 10x more data as
before
• We re-trained the RFC with these larger noisy dataset
• Slight improvement in its performance: 82.33%
• Without any manually and costly labeled data!
• KM took us about 10 minutes to implement
• We have also implemented a KI model that utilizes a
supervision function
• This function verifies the value of certain variables
and forces the output of the ML model
• This function represents a layer of safety
• Accuracy in our tests remained the same

11. Results: Energy Consumption Prediction
• We implemented a Random Forest
Regressor to deal with correlated and non-
informative variables
• Energy consumption is influenced by activity of
the station, logged by its sensors
• There are variables which are correlated and
others are nearly constant throughout the
activity, it seems we might have partial visibility
• We trained it using 80% of the full time
series data for Press Station and kept the
20% remaining as test set
• The Random Forest Regressor obtained a Mean
Absolute Error of 2.75. train
test
à
Results on test set

12. KPIs
KPI 1: Data velocity
1000 msgs per second >> 2 msgs per second (sampling frequency of
sensors in digital factory)
KPI 2: Generation of knowledge graph for the DT
The Translator is capable of generating a graph structure in the form
of NGSI-LD Entities. This is done automatically using provided meta
data from the factory message bus.
KPI 3: Accuracy of KI for DT
Accuracy of ML model trained with hand labeled data: 82.11%
Accuracy of Knowledge Model: 70.50%
Accuracy of ML model trained with data labeled by Knowledge
Model: 82.33%
KPI 4: Accuracy of DT Refinement (Strong KF)
Accuracy of KI model with strong KF: 82.33%
Note: Same accuracy as without refinement because Strong KF did
not find any necessary correction

13. Lessons learnt
NGSI Translator: Valuable tool for extracting knowledge from raw data, allows for more flexibility
(NGSIv2, NGSI-LD). You can find the translator at
https://github.com/ScorpioBroker/ScorpioBroker/tree/feature-82/NGSILDTools/NGSILDTranslator
KAFKA NGSI-LD Integration: Kafka is a good choice as it is wide spread and has excellent performance.
Recommendation: use key on Kafka messages so an identifier is attached to the data.
Digital Factories Data: Data heterogeneity between factories is still very high.
Opportunity: tools/techniques to join data from different factories are valuable and needed!
Recommendation: it’s beneficial to publish example data from MIDIH factories.
Knowledge Infusion: it enabled us to train a classifier with no manually labeled data, achieving high
accuracy. KI shows great potential to reduce effort of creating ML models.
KI, ML models’ performance and execution in FogFlow: Models achieved good accuracy but we
believe this can be further improved.

14. THANK
YOU!
NEC Laboratories Europe