Masked Prediction of Time Series data Using Novel Machine Learning Models

Abstract

We are living in a time where data is the new oil, where the industries are data driven. This change was expedited by the enormous amount of data that we are producing every second and the increase in computational power. At present the automotive sector is thriving in this data-driven model by their ubiquitous need for common and industrial purposes and the various data they are collecting to improve the sector as a whole. We can now see an automobile as not just a mechanical prod uct but as a robot on wheels. Along with this data-driven model, the electrification of automobiles has revolutionized the industry. In the electrification process, the battery module is one of the key components that power the systems. To do this we must analyse the data from the battery modules for its efficient usage. However, due to certain hardware issues or if the vehicle is out of range and it cannot update the data we might lose data. This loss of data can obstruct the efficient usage of the data in machine learning models to optimize the system. There are several methods to impute missing data, for example, there are statisti cal methods such as the Auto-regressive methods, which are limited by their time and the high cost of their computations. This thesis focuses on this problem and designing a neural network model for masked prediction of the Time series data. In this thesis, a Transformer Network is implemented for the masked prediction of the missing time series data. In this thesis, we have built the machine learning model from scratch after weighing several factors. The data on which the model is trained is generated by the vehicle collected. This was led by pre-processing, later following the selection of the model. The model developed here is a variation of the transformer model, called the Time Series Transformer(TST), which predicts the missing values in the time series data. This model is then evaluated with suitable metrics by the model and the problem statement. The thesis aims to predict the missing values to improve the quality of the data collected and its quality usage to improve the performance of the vehicle.