Deep learning workshop at Houston machine learning meetup. Video uploaded: https://youtu.be/bAkBGKF2s2I?t=2320

1. Deep Stacked Autoencoders
For Anomaly Detection
Deep learning workshop
Ziad Katrib

2. How important is equipment prognostics
• 80+ Plants, several thousand rotating equipment, translating to several thousand potential
failures
– Better monitor thy equipment
• Since 2013 Calpine has been using a 3rd party Anomaly Detection framework based on NASA’s
ORCA algorithm
– Blind spots
– No Fault Detection or Remaining Useful to speak of
• Improving on existing methods through a combination of DL and classical ML algorithms

3. Anomaly detection vs Fault Classification
Unknown unknowns Unknown knowns
Anomaly Detection, 2018 PHM Society Annual Conference Tutoria, Neil Eklund

4. Anomaly detection in context
Maslow Pyramid of needs
Sensor Data Collection
Maintenance Labeling
Remaining
Useful Life
Anomaly
Detection
Fault
Detection
Prognostics

5. Framework
for
equipment
monitoring

6. Framework overview
Design Build
Publish Model
architecture
Get
New Data
QC Data
(Prep Training Data)
Deploy Model
Read
Model
Retrieve / Receive
New Data
Score &
Alert / Store
Data Science Team
Analysts

7. Framework overview: Analyst Perspective

8. http://wasp-sweden.org/custom/uploads/2017/08/ViditSaxena-DenoisingAutoencoder-20170811.pd
KNN Regressor Model vs
actual
threshold
Existing Algorithm
New supplemental
Algorithm based
on Autoencoders
Model vs
actual
threshold
Isolation
Forest
Normal/
Abnormal
Improving existing Anomaly Detection Methodology
Reconstruction
Error

9. Hands on Machine Learning with Scikit-Learn and TensorFlow
Anomaly Detection and Fault Disambiguation in Large Flight Data:
A Multi-modal Deep Auto-encoder Approach
Kishore K. Reddy, Soumalya Sarkar, Vivek Venugopalan, Michael Giering
Denoised Stacked Autoencoders
Reconstruction
Error
Why Autoencoders:
- Semi Supervised
- Easy to train
- Can use on multi
modal and
heterogeneous
without feature
engineering

10. Isolation Forest for one class classification(OCC)
Isolation
Forest
Reconstruction
Error

11. Isolation
Forest
Normal/
Abnormal
Optimizing with OCC
Reconstruction
Error
Hyperparemeter tuning is done with entire pipeline in mind, i.e. optimizing for False
Positives, True positives and not just Autoencoder re-construction error

12. https://github.com/zikkat/DLWorkshop_AD/blob/master/01_Regressor_Classify.ipynb