This contains the slides for the presentation given on failure prediction at pune university's Smart India Hackathon and other conferences, This was presented to various tech professionals and esteemed academics as well as members of government bodies. This presentation was also used while presenting my thesis on predictive maintenance.

1. Failure prediction for
APU's on a Metro
System
Aaryadev Ghosalkar

2. Agenda
• Introduction
• Problem Statement
• Literature Review
• Existing Systems
• Modelling
• Results
• Conclusion
• Future Scope
Failure prediction for APU's on a Metro System 2

3. Introduction
Air production unit (APU) is component on most modern metro systems
which circulates compressed air through the metro, our research presents a
comprehensive comparison of PdM models on APU failure detection. Our
goal is to detect failures at least 2 hours in advance, the challenge with APU's
in particular is that APU failures are rare events
In the dataset there are 3 failure within 6 months of operation
Failure prediction for APU's on a Metro System 3

4. Problem Statement
• Detect failures at least 2 hours in advance
• Create a system that is customizable can expand to more than 2
hours if required (also keeping in mind the side effects that may
arise)
• Provide a way to deploy the model for real time inference
Failure prediction for APU's on a Metro System
4

5. Literature
Review
Taking a look at what other
researchers have done

6. Literature Review
• Veloso et al performed the initial data collection, which
involved converting data from the metro company
database into a CSV file
• We used some of the ideas for data preprocessing
presented by the AzureML team at microsoft
• A study on Davari et al motivated many of the algorithms
that we used in this study, there work has been a great
resource in our study
Failure prediction for APU's on a Metro System 6

7. Literature Review continued
• Chaudhuri et al used SVM to classify vehicles into 3
distinct risk levels, focusing on model interpretability,
model interpretability refers to how easy it is to
understand the choices made by the model and results of
the model
• Various other researchers have used CNN with GAF to
convert timeseries data into images and used CNNs for
classification most notable was done by Silva which
reported an accuracy of 93% in their study
Failure prediction for APU's on a Metro System 7

8. Literature Review continued
• Nguyen and Medjaher used LSTM to predicted the
probability of failure in a given time window, we drew a lot
of inspiration from their work in terms of balancing the
dataset and the model used
• In terms of early research into RUL estimation a majority
of the work revolves around using statistical models or
models which assume linear degradation pattern, it was
only after 2016 Deep learning models were used in this
field
Failure prediction for APU's on a Metro System 8

9. Existing
Systems
Predictive maintenance has been
used in a lot of domains such as
elevators and Jet engines.

10. The Infrastructure abroad
• Uses NLP on text fields where engineers
describe the problem and how the
problem was solved.
• Also predict if a failure is about to happen
and which component will fail
• TFL expects to save £3 million a year
Failure prediction for APU's on a Metro System
10
• TrainDNA collects and analyzes real-
time data from more than 200 trains
across Australia using IBM Maximo
• 51% Increase in realiablity after
introducing the TrainDNA system
• Considers multiple failures
• Large scale system able to process 30
Million message every hour
London Australia and NZL

11. Modelling
Taking a look at how the data
was processed and what ML
models were tested
Failure prediction for APU's on a Metro System 11

12. Data Preprocessing
Most of the data collection
and cleaning work on this
data set was done by
researchers that created
the data set
Thus this data set did not
have any null values
Failure prediction for APU's on a Metro System
• Linear Model are a bad
idea
• There is no obvious
pattern which we can
see to distinguish each
class, which further
motivated use of deep
learning
• Balance is key!
Collection EDA
• Discretizing the data to
make this more
suitable for
classification
• Generalizing the model
using a parameter to
control number of
hours before warning.
Preparation
12

13. Challenges in data processing
• The sheer volume of data presented a significant
challenge, there were more than 10 000 000 (1Cr) rows in
the dataset to solve this we performed carefully changed
the data types of the features to minimize the loss of
information this allowed us to significantly reduce size of
the data with a 91% decrease in storage space and 77%
decrease in RAM usage
Failure prediction for APU's on a Metro System 13

14. Challenges in Data processing
continued
• The highly imbalanced nature of the data also presented
a huge challenge when training the models, since APU
failures are a rare event and only 3% of the data
constitutes of APU failure a model that predicts all data
points as normal would mathematically have an accuracy
of 97%, to combat this we used Near miss under
sampling and a convenience sampling like strategy for
the LSTM, inspired by Chen et al
Failure prediction for APU's on a Metro System 14

15. Machine Learning Models
• SVMs take very long to
train
• Results are not
satisfactory and most
other models perform
better
Failure prediction for APU's on a Metro System
• Very quick training time
• The hyper parameters
such as the criterion do
not make any
difference to the
accuracy
SVM Decision Trees
• Provide greater
accuracy than decision
trees due to this being
an ensemble model
• Training time is similar
to decision trees
Random Forests
15

16. Deep Learning Models
• Used a small neural network of 25k
parameters with AdamW optimizer
• Best results from all tested model
• Does not take into account the time
series nature of the data
• Requires CUDA for efficient
deployment.
Failure prediction for APU's on a Metro System 16
• Can incorporate temporal patterns
• Large model almost 184K
parameters which makes training
difficult.
• Requires 3rd order tensors as input
for training thus balancing data is
hard
• Also requires CUDA for efficient
training and deployment
Neural Network LSTM Network

17. Results
Accuracy Precision (Class
1)
Recall (Class 1)
SVM 0.57 0.60 0.62
Decision Tree 0.66 0.69 0.70
Random Forest 0.70 0.69 0.78
Neural Network (Adam) 0.72 0.67 0.91
Neural Network (AdamW) 0.76 0.75 0.87
LSTM Network 0.76 0.64 0.85
Failure prediction for APU's on a Metro System 17

18. Conclusion
• Neural networks produce satisfactory results however the
Random forest model can be used when deploying on
low end hardware or considering an edge solution.
LSTMs perform nearly identical to neural networks with
less data
• Microservices can be considered when deploying as this
will allow different parts of the model to be deployed
separately
• Multi collinearity is present in the data which can makes
it hard to fit any kind of linear model
Failure prediction for APU's on a Metro System 18

19. Future Scope
• Transformer based architecture can be considered as
they have shown promising results with sequential data in
the case of NLP applications
• Perhaps due to the imbalanced nature and the rarity of
APU failures on a real metro systems an anomaly
detection approach would be better as this would not
need the data to be balanced and a lot more of the
existing data can be used
Failure prediction for APU's on a Metro System 19

20. Question and
Answer
Aaryadev Ghosalkar
aaryadevg@gmail.com
https://github.com/aaryadevg