Overview

The principal aim of this research project is to enable industrial machines with communication, deep-learning and decision making capabilities in order to improve failure prediction in a cost-effective way. Current prognostics frameworks usually rely on models that operate in a centralised architecture. However, industrial fleets of machines are highly heterogeneous and dynamic, conditions known to be conducive for distribution.

Methodology

This project combines survival-based deep learning techniques for machine prognostics with Multi-Agent System architectures to predict machine failures in real time. Prognostics is done locally in each asset using a LSTM Recurrent Neural Network designed for regression. This neural network can use a loss function based either in a Gaussian or a Weibull posterior. Learning between the agents in the machine fleet is done by sharing data among machines that are similar to each other. 

Supervisor

Dr Ajith Parlikad

Sponsor

"La Caixa" Fellowship, Siemens