Understanding Applicability of Echo State Networks to Diverse Industrial Data
| Title | Understanding Applicability of Echo State Networks to Diverse Industrial Data |
|---|---|
| Summary | Understanding the Governing Dynamics of Echo State Networks (ESN) on time varying signals (in terms of properties) from different industry sources. |
| Keywords | |
| TimeFrame | Spring 2017 |
| References | |
| Prerequisites | |
| Author | |
| Supervisor | Shiraz Farouq, Sławomir Nowaczyk |
| Level | Master |
| Status | Open |
Background: In the age of IoT and various sensor technologies, data driven techniques compared to physics based models are now becoming the de-facto standard to understand and learn the underlying time series generated by these systems; Machine Learning (ML) is among one of the techniques used for the purpose. The Recurrent Neural Network (RNN) is machine a learning model that most closely resembles the working of biological brains. However there is a caveat here; these models are difficult to train. To overcome this, various alternatives within the RNN framework have been proposed, ESN being among one of those. Despite this, ESN’s still pose a challenge in terms of its hyper-parameter tuning: for instance, the size of the internal nodes (also called dynamical reservoir), spectral radius, connectivity of the node, needs to be set in an optimal way.
The Problem: We have access to data from various industries, for instance Heat Pumps, District Heating, operational data from buses (Volvo). Each data source will present its own challenges in terms of setting the hyper-parameters in the context of ESN’s.
Project Goals: The project is exploratory in nature. The students, therefore, will first need to understand the theory behind the ESN’s and the current state of the art. It must be noted that different signals with have different properties, for instance it terms of trend, seasonality, jumps, frequency etc. The major goal here will be to understand how the hyper-parameters of the ESN will vary according to these properties. Hence the students will implement the ESN’s to model the behaviors of various signals from various industry sources with varying properties and report the challenges and accomplishments. Moreover, the students are free to explore new ideas.
