BigML’s partner, A1 Digital, showcases in this presentation a specific use case on “Damage Detection in Rail with Embedded Machine Learning”.
Speaker: Dieter Myer, Senior Data Scientist and Machine Learning Consultant at A1 Digital.
*Intelligent Mobility 2021: Virtual Conference.
2. 2
IoT architecture: traditional ML approach
Dashboard
Reports
Notifications/
Alarms
Cloud-based Machine Learning
Development and Prediction
A1 Digital IoT Platform
Cloud
3. 3
IoT architecture: EdgeML approach
Dashboard
Reports
Notifications/
Alarms
EdgeML-
Prediction
A1 Digital IoT Platform
EdgeML:
Analyzing date already on
the IoT device
Cloud
Cloud-based Machine Learning
Development
5. 5
Information from data
is generated directly at the source (IoT Device),
by running ML-Algorithm directly on low-power and low-cost
microcontroller.
What is EmbeddedML or EdgeML?
9. 9
Key steps behind EdgeML development
Data capturing
Data preprocessing and
feature engineering
Model Training
Convert into
optimized code for MCU
Target Platform
Analysis
ML Model Creation Operating Mode
1
2
3
4
5
6
Connect to Cloud
12. 12
Damage Detection
Damage Detection for Freight Wagons
REQUIREMENTS
Detect damages based on recorded shocks
Identify customers that cause damages
OUR SOLUTION
Machine Learning model trained on 200k shocks from 6000 wagons
Geographic clustering of shocks identifies „high risk customers“
RESULTS
Detection and live monitoring of wagon condition and damages
Automated alarms to take immediate action
wagons without
alarms 19%
wagons with alarms
81%
118
Wagons without
alarms 91%
Wagons with alarms
9%
1410
Customer A: 254 Alarms
Customer B: 287 Alarms
13. 13
Experimental Setup for Damage Detection
Damage Detection
Balanced training set with 234 Wagons
inspected recently
equipped with IoT Device (acceleration sensors)
50 % heavy damages 50 % minor issues (e.g. missing signs)
= not damaged
ML Challenge: identify damaged wagons based on recoreded shocks of last weeks
14. 14
ML Workflow and Results
Damage Detection
Feature Engineering
Only shocks during
loading the wagon
Wagon type
Customer
ML Modelling
Automated testing of
150 machine learning
models
Best Model
Neural network detects 30%
of damages with
>95 % precision.
15. 15
Energy Supply
Challenge: Monitor and evaluate condition of
power supply infrastructure and devices
Solution: Energy independent sensors with
EdgeML application to define condition status
and alarm in case of critical events
Further use cases
Flat Spot
Challenge: detect flat spots by analyzing
vibration data.
Solution: Apply EdgeML with continuously
analyzing incoming ‘vibration’ data and send
warning
Weighting Sensor
Challenge: is the wagon overloaded?
Solution: Use infrared distance sensors on the
suspension and calibrate this with ML to detect
overloading
16. Get in Touch!
A1 Digital (Österreich & International)
Dieter Mayr
Data Scientist & EdgeML Consultant
+43 664 66 22570
Dieter.mayr@a1.digital
Thank you for
attention
Editor's Notes
Müssen wir Rechte an Bild kaufen`?
https://pixabay.com/de/vectors/amazon-alexa-echo-5241000/
https://pixabay.com/de/illustrations/intelligente-%C3%BCberwachung-apple-uhr-2845072/
Derive information from data directly at the source
Wieso gerade jetzt so großes Thema
https://pixabay.com/de/vectors/amazon-alexa-echo-5241000/
https://pixabay.com/de/illustrations/intelligente-%C3%BCberwachung-apple-uhr-2845072/
Derive information from data directly at the source
Solution: ML model (precision > 99 % -> 20 % of damages detected). For RCA it is not important to detect all damaged wagons. Precision is more important – >when they send someone to check the wagon, false positives should be kept minimal. For higher recall (detecting more than 20 % of damages), precision would go down to 60-70%.
„high risk customers“
Customer A: Acceleration alarms (shocks) for 81 % of 118 wagons -> results in higher damage probability
Customer B: Slightly more alarms within same period, but on much more wagons (1410) ->lower probability for wagon to get damaged at this customer!