MA: Feature Selection on Highly Redundant Multivariate Time Series in Deep Learning

[kauschke]

Diese Arbeit findet in Kooperation mit Compredict statt. Kenntnisse in Machine Learning und insbesondere Deep Learning sind eine Voraussetzung.

Themenbeschreibung
Feature selection is a critical component in Machine and Deep Learning workflows, where it provides less complex and cost-effective predictors. In addition, it exploits the underlying patterns in the given data as well as it illustrates how the features correlate with the target. In highly redundant signal data, for instance generated through automotive sensors, there are many irrelevant and similar features which have negative impact on the model. State-of-the-art feature selection techniques (such as Recursive Feature Elimination Algorithms) fail to exploit the importance of the features in such timeseries because they lose the correlation information among the features at different intervals.

Automotive sensor data will be used as a case study. Hence, the objective to be achieved in this work is to develop an algorithm that can reliably determine the most valuable features within a certain interval to reconstruct a given target value, e.g., a force or a torque acting on a component, using Deep Learning. In more detail, the data is collected from the CAN bus of a BMW M3 and the algorithm should be able to identify which timeseries variables have the strongest correlation with the side shaft torque, the suspension deflection and the steering force of the vehicle. The chosen variables from the algorithm will be fed to a Neural Network Model and the achieved result from the selection should perform better than the current model that considers many unrelated variables. Optionally, the final Neural Network model should be implemented on a Nvidia Jetson TX2 Microprocessor and tested live in the BMW M3 to prove the real time capabilities of the method.

Tasks:
• Study of existing feature selection techniques for multivariate time series.
• Development of an algorithm to determine the best features in a multivariate time series.
• Testing of the algorithm on CAN bus data to determine the variables that have an impact on the side-shaft torque, the suspension deflection, and the steering force.
• (Optional) Implementation of the algorithm on Nvidia Jetson TX2 Microprocessor for live testing in the BMW M3.

For more information, please send a mail to kauschke@ke.tu-darmstadt.de.