1 of 8

Sensing fluids

By

Hunter Brown

Jackson Pontsler

2 of 8

Background

Newtonian Fluids are defined by:

​

This means that any forces cause shape change

To model a fluid FEA is often used but this can be computationally expensive

​

​

Some simplifying assumptions:

  1. Constant Volume (incompressible)
  2. No energy generation
  3. No mass generation

To model using a CFD often the navier-stokes equations are often used:

​

3 of 8

Current Work

The current work is sparse due to the complex nature of problem

Pouring is sometimes seen in

  1. Human Demonstration
  2. Skill Teaching (FSS)

Common simplification is Horizontal movement of liquid holding container

4 of 8

What is a CNN?

5 of 8

What is a Convolutional Layer

Subsampling or pooling layers

  • In layman's terms think of it has running a filter over the image.
  • This isn’t done only once but will be done with different filters and these separate outputs will be saved to be used later.
  • From these convolutional layers they will be subsampled or pooled to form new layers.

6 of 8

Why does this work

  • By pooling you may find features such as percentage of pixels used in a certain area of image.
  • Convolution layers to find features like symmetry over a certain decision boundary.
  • By repeating these layers you classifier will discover hidden features that you may not initially think of.

7 of 8

Basic LSTM

(Long-Short Term Memory)

3. Decide what to Output

    • Scale cell state (tanh)
    • Multiply by output gate

Normal RNN struggles to learn long-term memory

This builds long-term and short-term memory into the structure

Three main steps

  1. Forget Gate Layer
    1. Multiply cell state by forget gate layer
  2. Add to Gate Layer
    • Create input gate layer
    • Create Vector of new data
    • Multiply input gate layer and new data vector
    • Add this to hidden layer

8 of 8

Discussion Questions

Major misunderstandings?

Any initial thoughts or comments?

List the assumptions used in this model?

Are these valid assumptions?

How do these limit the model?

Possible ways around this?

​

Are neural networks the only way to deal with this in real time?

Is there structure in this problem that could be exploited?

How could we use these findings?

How could we expand on these finding?

​