Introducing 

Prezi AI.

Your new presentation assistant.

Refine, enhance, and tailor your content, source relevant images, and edit visuals quicker than ever before.

Loading…
Transcript

Electric Motor Temperature

Federico Moiraghi - 799735

Pranav Kasela - 846965

Roberto Berlucchi - 847939

Our

Problem

Our Problem

Predict temperature of an experimental Electric Motor

Motor is a prototype, tested only on a bench

Our Data

Anonymized real Data

Our Data

White noise to simulate

real-driving cycles

Target

2Hz

Forecasting

Forecasting

Sensors fall offline for long time:

this Model can replace NA or simple interpolations

high temperatures are more

important

Prediction

Prediction

Sensors are expensive and can get damaged:

this Model can replace them

Autos can adjust the cooling of their motor without a specific sensor

Our Approach

Our Approach

Testing different kind of

Recurrent Neural Networks

Testing new trends in Literature

Various Architectures

Various Architectures

Seq2Val training to shuffle data

and speed-up learning process

Using Mean Squared Error as Loss-function in a first trial

Using a weighted version of MSE to improve prediction on high temperatures

Recurrent Neural Network

Simplest approach: something in the past

is reused in the present

Gated Recurrent Unit

Improved version: the Model

learns what to remember

Long Short-Term Memory

A-sort-of generalized version

Convolutional Neural Network

Features are channels of the same

time series

Information is extracted and then

flows through a feed-forward NN

Auto-ML

Highly

Optimized

Each Model is optimized through

Auto-ML - GP with EI

Results

Results

After Auto-ML, models are evaluated to find the most fitting one

Ideal is with high performance

and low parameters

Forecasting

Forecasting

Prediction

Prediction

Using the Weighted MSE

Using the MSE

Learn more about creating dynamic, engaging presentations with Prezi