Space Apps Challenge 2018

INTRODUCTION

What happens when a spaceship suffers from MMOD "Micro Meteoroid and Orbital Debris " and can’t detect what the MMOD impact occurred outside of the ship looks like? And here comes the challenge!

To design and construct a free-flyer that detects and characterizes MMOD impacts occurred on spaceship on astronaut command. We designed and built“Pneu-UFO” , a UFO shaped free-flyer which maneuvers using a pneumatic system.

Our Solution is divided into two partitions, a hardware based solution and a software based solution. The software partition: Image Classification model using Convolutional Neural Networks and Transfer learning based approach to process and analyze the feed frames from the camera.The hardware partition: Autonomous free flyer using Pneumatic thrust force.The two partitions are not independent; they in fact complete each other where the free flyer holds the camera that feeds its output to the processing unit of the
software. However, that doesn’t completely define the entities, as they can be treated as two separate entities, where the free flyer can be used to feed into any other image processing software and the image processing software can be used to analyze any set of images collected by any other hardware.

Mechanical design

Frame

A composites are materials made of two (a matrix or binder and a reinforce) or more constituents with different physical or chemical properties. When these materials are combined, the new material has different characteristics from the individual components. Usually the load is carried out by the fiber (from 70-90% of the load) and the rigidity and shape is provided by the matrix, which transfers the load to the fibers and stops or slows the propagation of cracks by isolating the fibers so that individual elements can act separately. However, one of the most important characteristics to take into consideration when working with composites is that their mechanical properties, such as strength, usually depend upon the direction of the applied load. That’s why in most drones, the fibers are bi-directional; to withstand loads in all directions during impact. One of the most commonly used types of composites is epoxy reinforced carbon fiber. As a comparison, steel has a tensile modulus of about 29 million psi (200 million kPa). Thus, the strongest carbon fibers are ten times stronger than steel and eight times that of aluminum, not to mention much lighter than both materials, 5 and 1.5 times respectively.

Propulsion

An eight thrusters configuration will provide all possible degrees of freedom for an aircraft to maneuver in space. Four of the eight thrusters are used for levitation of the flyer, two for the upward movement and the other two for the downward movement. The other four are used to rotate around the spaceship and provide lateral motion.

Electrical Design

Control Systems:

a-Microcontroller:

Arduino boards were used because of their affordable prices as well as their open source AVR microcontroller-based development which can be programmed easily.

b- Actuators control:

For the DCVs (Directional control Valves) , Relay Modules were used. These Valves control the flow of pressurized air through the hoses to control the thrust force to moving the flyer.

Tether:

Communication and Power
The camera signal is transmitted from the flier to the space craft through one CAT6 Ethernet cables which have 8 cores for camera signal transmission. CAT6 was used
as it can provide serial communication at a rate of 250 Kbps.t
Pneumatic
The pneumatic hose has an inner diameter of 4 mm and an outer diameter of 6
mm. It’s able to withstand pressure up to 10 bars .

Vision system:

Main Camera

The main camera used is a low light blue robotics analogue camera. The camera was horizontally oriented on the top of the flyer to provide perfect vision of the space ship to detect any possible cracks or damages. We chose to use the low light analogue camera with a 2.1mm focal length; because it has a wider angle of view(128 degrees horizontally and 96 degrees vertically) , its compact size and small weight.

Software design

On Board Software

We designed the on board software using a simple approach where the dimensions of the spacecraft are entered as an input, and the algorithm then finds the best route for the free flyer to orbit around the spaceship./

Image Processing Software

Transfer learning is a machine learning method which utilizes a pre-trained model. With transfer learning, we basically try to exploit what has been learned in one task to improve generalization in another. We transfer the weights that a Network has learned at Task A to a new Task B.

The general idea is to use knowledge, that a model has learned from a task where a lot of labeled training data is available, in a new task where we don’t have a lot of data. Instead of starting the learning process from scratch, you start from patterns that have been learned from solving a related task.

For example, the image recognition model called Inception-v3 consists of many parts as the convolutional layers for feature extraction and the fully connected layers for the classification.

Our approach included using RESNE-T18 model and removing its last fully connected layer,then treat the rest of the convolutional neural network as a fixed feature extractor for the new data set,then add our custom fully connected layer based on our application and retrain the model.We added a layer that classifies 4 classes; no damage,low damage, medium damage and high damage.After that we retrained only the last layer of our model using 1200 images divided into 80% for training and 20% for validation.In the end ,we achieved 96.8% accuracy on the validation dataset.

To improve the performance of our model we needed to add more data to our training set because our data wasn't large enough. So, we used data augmentation to increase our dataset and enhance it by addding translational invariance to make the data representative of the many different positions, angles, lightings, and miscellaneous distortions that are of interest to the image classification task.

Language and Libraries used:

We used python as our programming language and pytorch as our main library to handle the datasets, the manipulation of the pretrained model, the creation of our custom layer and the training of it.

Challenges

1. We didn't have a data-set for training our classifier so we had to collect our own.
2. Our computation resources used for training the model were limited.

Future Upgrades

1-Improving the detection performance by increasing our data sets.

2-Performing the image processing onboard by adding a Raspberry pi to the free flier and sending the results to the space ships.

3-Adding a dynamic positioning feature to our free flier to maintain better stability .

4-Implementing a closed loop control system that takes its feedback from sensors to avoid mmods hitting the free flier.

Resources

1. NASA space apps: Design by nature
   https://2018.spaceappschallenge.org/challenges/can-you-build/design-based-nature-fusion/details
2. Pytorch resource
   https://pytorch.org/docs/stable/index.html
3. Pretrained models
   https://pytorch.org/docs/stable/torchvision/models.html?highlight=models
4. Carbon fiber material:
   https://www.unmannedsystemstechnology.com/technical-article/composite-materials-for-
   unmanned-systems/
   https://www.azom.com/article.aspx?ArticleID=12234
   http://zoltek.com/carbon-fiber/what-is-carbon-fibe..

SOURCE CODE

https://github.com/oswidan97/SurfaceDamageDetection/commits/master