Machine Learning in Rust Detection
Machine learning is used on an autonomous UGV to detect rust in pipelines.
Toggle between list and paragraph view.
- [00:00:02.700]Good afternoon to whoever this may
concern. My name is Gabriel Clark.
- [00:00:07.590]I worked with the BIM lab located in
Nebraska hall to improve their autonomous
- [00:00:12.150]pipeline analysis project.
- [00:00:14.430]My main goal with this project was to
design and implement a rust detection
- [00:00:19.230]program that could run
on an autonomous Rover.
- [00:00:25.140]This project is large in scope and is
overshadowed by the mission to better the
- [00:00:30.030]safety of pipelines across America.
- [00:00:32.640]Through the through
better analysis methods.
- [00:00:36.510]This is done through automated
drone inspection and automated Rover
- [00:00:42.300]Several notable parts of this project
include an autonomous Rover that will
- [00:00:46.650]navigate pipelines and create high
resolution digital scans of these
- [00:00:51.060]pipelines. It also includes
an autonomous drone, um,
- [00:00:55.680]that can 3d model and scan
pipelines from above ground
- [00:00:59.940]autonomously. Um, and
then with these things,
- [00:01:04.080]it also includes software
to process these output from
- [00:01:08.730]the UAVs, which are unmanned
aerial vehicles and UGA EVs,
- [00:01:13.740]which are unmanned ground vehicles.
- [00:01:19.940]The piece of the project
that I worked on concerns,
- [00:01:22.640]the processing of the [inaudible] output,
- [00:01:26.030]this Rover generates a
geotagged point cloud, um,
- [00:01:30.230]of the inside of the
pipeline it's traversing,
- [00:01:32.780]which means every point in this point,
- [00:01:35.150]cloud of information has a, um,
- [00:01:42.320]A global position associated with it.
- [00:01:44.600]And it can be related back to anywhere
in the world based off of this
- [00:01:48.170]information. Um,
- [00:01:49.880]so not only does this include pictures,
- [00:01:53.240]color information about the sides
of the pipes that we're scanning.
- [00:01:57.200]It also includes depth information, um,
- [00:02:00.470]as well as heat coming
off of the pipe. Um,
- [00:02:04.610]cause we use IRR sensors
to accomplish this task.
- [00:02:08.840]I used common visual analysis
machine learning algorithms.
- [00:02:13.550]So after the rust is located by
these algorithms and the pipeline,
- [00:02:17.780]the locations, uh,
- [00:02:19.310]the rest are then related back into
3d space through these geotagged point
- [00:02:28.460]To gather this training data
for the machine learning model.
- [00:02:31.670]I use an automated web scraping
- [00:02:36.080]downloads images based on keywords
from multiple different languages.
- [00:02:41.150]I then send these images, um,
- [00:02:45.350]through a simple binary
classifier to help, um,
- [00:02:48.980]automate the separation of
this large dataset. I collected
- [00:02:53.720]to create a comprehensive data set because
this original data set only contained
- [00:02:58.400]images, rust images.
- [00:03:01.720]I supplemented my web findings with a
selection, uh, from the image net dataset,
- [00:03:07.450]and then use this combined data set to
train a convolutional neural network for
- [00:03:12.010]Russ localization,
- [00:03:13.630]whose architecture was based off
of the resonant 50 architecture.
- [00:03:18.490]I have not yet finished this task.
- [00:03:21.160]Although currently the model
accuracy is hovering around 89% in
- [00:03:25.990]binary classification of
rust and not rest images.
- [00:03:32.890]So as it could be seen in this second
slide with the down in the left of the
- [00:03:37.870]image here, um, we have a
- [00:03:45.910]cardboard simulated pipeline in
the basement of Nebraska hall.
- [00:03:50.800]Um, we ended up printing pictures of rust
and gluing them to the inside of this
- [00:03:54.760]pipeline in order to test the, um,
- [00:03:57.910]unmanned ground vehicle before it
is used in a real situation. Um,
- [00:04:02.230]after seeing this,
- [00:04:03.280]I recommended that we try
building a digital simulation
- [00:04:08.200]in unity, um, to speed
up testing as well as,
- [00:04:13.240]um, in the future training
of machine learning
- [00:04:19.960]And here's a image or video of the
- [00:05:11.370]Again, later in the research process,
- [00:05:13.500]the models that we generate with
this digital simulation, um,
- [00:05:17.310]machine learning models,
- [00:05:19.120]it could be used as the base of a
transfer learning algorithm for real world
- [00:05:27.060]So there's many things that have not
been accomplished yet in this process and
- [00:05:30.690]in this project, um,
- [00:05:33.000]we hope to integrate the machine learning
model into the unity simulation so
- [00:05:37.230]that we can train repeatedly overnight
without watching the Rover in the
- [00:05:42.420]physical world simulation of the pipeline.
- [00:05:45.900]I also hoped to improve the unity
simulation to more closely match what
- [00:05:50.790]actual pipeline rust
looks like currently. Um,
- [00:05:54.630]it's a simple repeated texture
in the unity just to help people
- [00:05:59.360]visualize what it will
look like eventually,
- [00:06:01.340]but I've not spent the
time to generate, um,
- [00:06:04.910]accurate rust for the inside
of that pipe. And then lastly,
- [00:06:09.950]we hope to implement transfer,
- [00:06:11.300]learning on the models made from
this unity simulation and then start
- [00:06:16.040]training our Rover in the actual physical
space in the basement of Nebraska
Log in to post comments