Scribd
Upload
240 views

Time Series Anomaly Detection With DL

Slides for a talk given by Sergei Bobrovskyi @meetup.ai Hamburg Link to event: https://www.meetup.com/de-DE/meetupai-hamburg/events/266726633/

Uploaded by

Anonymous 9tJYscs
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
240 views

Time Series Anomaly Detection With DL

Slides for a talk given by Sergei Bobrovskyi @meetup.ai Hamburg Link to event: https://www.meetup.com/de-DE/meetupai-hamburg/events/266726633/

Uploaded by

Anonymous 9tJYscs
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 18

Time Series Anomaly

Detection with DL
The coming age

Unified data ecosystem.


Connected Aircraft
24,000 Parameters
100% available data

Operational Environment Data


In-Service Data
Industrial Data
Supply Chain Data
Time
Series
The beating
heart of an
airplane
Challenge

Open Dynamical Systems


Projection to 3 parameters

Pilot commands Outputs to other


Input from other Subsystem systems
systems System parameters
External conditions
(weather)

System boundary

~500 parameters
Opportunity

Anomaly Detection

Anomalies:
Patterns in data not conforming to
expected behaviour

Rules already find severe anomalies

Challenge:
small, context-dependent deviations

regular anomalous
Vision

Industrial solution
Semi-supervised human-level performance Domain Expert
on time series from industrial assets

Validation System Data Enrichment


(visualisation &
feedback)

Model Training Skywise


Train. set
Model Selection Other
Performance Platforms
Val. & Deployment
Test Sets estimation
Application of rules

Performance
Control
State of the art

Statistical approaches
Vectors from Dimensionality Anomalies from
time series reduction distance measure

Wavelets
Antoniadis et al. 2007 Clementine Barreyre. Statistiques en grande dimension
FPCA pour la détection d’anomalies dans les données
fonctionnelles issues des satellites. Statistiques
Ramsay et al. 2005 [math.ST]. INSA de Toulouse, 2018.
Français.

2018ISAT0009

Other approaches...
State of the art

Deep Learning approaches


Predictive Reconstructive

LSTM
Malhotra et al. 2015 Autoencoder
Hundman et al. 2018 Hawkins et al. 2002
CNN Malhotra et al. 2016 (LSTM)
Munir et al. 2018 Su et al. 2019 (stochastic RNN)
GAN
Schlegl et al. 2017
Li et al. 2019
Foundation

ADvISED
State of the art,
modular time
series processing
toolbox
Solution

Deep Learning architecture

Choices

Hundman, Kyle, et al. "Detecting spacecraft anomalies using lstms and nonparametric dynamic
thresholding." Proceedings of the 24th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining. ACM, 2018.
Solution

Investigations
Autoregressive Multivariate
next step 30 s ahead 30 s ahead

Cross-Correlation Cross-Correlation Cross-Correlation


Solution

Airbus approach
Multivariate anomaly detection: combining Deep Learning techniques with engineering knowledge

3 Knowledge
transfer

Co-Innovation

Subsystem level 1 2
understanding
System level
understanding
Co-Innovation
aigym.airbus.com

AI Gym
Collaborate with
the most
innovative
companies at
massive scale.
AI Gym

Time series challenge

Data Train Validation

Unlabeled data 645 ground tests 32 ground tests


355 flights 18 flights

81 context 17 anomalies 79 anomalies


parameters detected by labeled by
8 parameters of community engineers
interest
AI Gym

Challenge results
Hard problem ...
Team F Score Precision Recall Rank
Datapred 0.51 0.53 0.50 1
Industrial team 1 0.36 0.36 0.36 2
Industrial team 2 0.30 0.30 0.30 3 ● modeling engine for
continuous intelligence
Industrial team 3 0.06 0.06 0.06 4
Industrial team 4 0.04 0.06 0.04 5 ● ability to combine machine
Academic team 1 0.02 0.06 0.02 6 learning with other types of
models in real time
Individual team 1 0.02 0.02 0.02 7
Individual team 2 0.00 0.00 0.00 8 ● faster garage-to-factory cycle
Next steps

Mature Engineering AI for Time Series


infrastructure input
Optimised
Production level Co-development with performance of
infrastructure engineering end-users algorithms

Industrial Anomaly Detection Service


Software suite including visualisation tool for 1 aircraft subsystem
developed in a short cycle in collaboration with Datapred
Thank you
References
Antoniadis, Anestis. "Wavelet methods in statistics: Some recent developments and their applications." Statistics Surveys 1 (2007): 16-55.

Ramsay, J., and B. Silverman. "Functional data analysis. 2nd springer." New York (2005).

Clementine Barreyre. Statistiques en grande dimension pour la détection d’anomalies dans les données fonctionnelles issues des satellites. Statistiques [math.ST]. INSA de
Toulouse, 2018. France

Malhotra, Pankaj, et al. "Long short term memory networks for anomaly detection in time series." Proceedings. Presses universitaires de Louvain, 2015.

Hundman, Kyle, et al. "Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding." Proceedings of the 24th ACM SIGKDD International Conference on
Knowledge Discovery & Data Mining. ACM, 2018.

Munir, Mohsin, et al. "Deepant: A deep learning approach for unsupervised anomaly detection in time series." IEEE Access 7 (2018): 1991-2005.

Hawkins, Simon, et al. "Outlier detection using replicator neural networks." International Conference on Data Warehousing and Knowledge Discovery. Springer, Berlin, Heidelberg,
2002.

Malhotra, Pankaj, et al. "LSTM-based encoder-decoder for multi-sensor anomaly detection." arXiv preprint arXiv:1607.00148 (2016).

Ya Su, et al. “Robust Anomaly Detection for Multivariate Time Series through Stochastic Recurrent Neural Network”, Accepted for KDD 2019, 2019

Schlegl, Thomas, et al. "Unsupervised anomaly detection with generative adversarial networks to guide marker discovery." International Conference on Information Processing in
Medical Imaging. Springer, Cham, 2017.

Li, Dan, et al. "MAD-GAN: Multivariate anomaly detection for time series data with generative adversarial networks." arXiv preprint arXiv:1901.04997 (2019).

ADvISED. Anomalies Detection in Time Series Data: state of the art. Technical Report, Airbus, 2018.

You might also like