Open Source Projects
These are some notable side projects I've undertaken with open source code repositories.
- Glance - Anomaly Detection on Industrial Sensor Data
Back in 2019 I participated in the Insight Data Science Fellowship. It was a great experience for me to focus on
machine learning for a couple months and really hone my skills. The cornerstone of the fellowship was a data
science project of my choosing, with the only requirements being that it covered the whole process from data
collection to ML modeling to cloud deployment and that it had actual business value. For my own requirements, I
really wanted to do something with cutting edge neural networks and I wanted to focus on an industrial use-case,
as opposed to many of the web service based projects others were tackling (e.g. AirBnB predictions).
What I created is a tool for anomaly detection on time-series industrial sensor data. It works by using a
variational auto-encoder architecture trained on normal operating data for the use-case, like the kind that may be
collected during equipment commissioning. The trained model is then dismantled and the center parameters at the
lowest dimensionality are used as a statistical distribution for normal operation. This works because the network
is forced to learn a meaningful latent space representation for the sensor data. Anything that fell too far
outside that distribution would then be flagged as an anomaly. In practice using a previously published dataset for a hydraulic system I
was able to achieve a better F1 classification score than any other method. The final model was an excellent fit
for the business case because it did not require actual failure data, normal operating data was enough. I digested
this work into a handful of high level slides for demonstration.
To my knowledge this approach was novel, although it was inspired by a paper that used a VAE model to detect anomalies in skin cancer
images. That paper used the recreation error of the VAE model as a metric for anomaly detection but provided
mediocre results (below 60% F1) for my problem. I kept the CNN layers because I saw good results using them over
RNN layers but I modified them to work on 1-dimensional time-series sensor data. The data was grouped by cycle
time and multiple sensors on the equipment became multiple channels in this representation. The most novel aspect
of my approach was to use the latent space of the VAE in a follow on clustering anomaly detection model. This
insight was what took this approach from a mediocre performance to over 87% in terms of F1 classification score.
- Submatter - Head Tracking with Computer Vision
...
- Crush Rig - Tissue Sample Data Collection
...