Functional data preprocessing application to oil-transfer pumps vibration parameters forecasting

This work describes functional data preprocessing algorithm. This algorithm provides a way to reduce error in forecasting problems solution. The algorithm is a part of oil-transfer pumps vibration parameters forecasting system that enables pump failures dynamics forecasting.
The author analyses existing approaches to vibration monitoring and decides to solve failure-forecasting problem as a long-term forecasting problem although researchers usually solve such problems with classification methods. An insufficiency of labeled data is the main reason of such a decision. Main ideas of the problem solution are the following. Neural network model takes and calculates periodical metered values characteristics. Time is split into periods using scales with different periods. We use shorter periods for short-term forecasting and longer periods for long-term forecasting.
Functional data preprocessing provides a way to increase forecasting quality. Preprocessing key idea is following. Functions sequence transforms one of model’s inputs in order to increase correlation between input and output.
Metered values distributions and dependencies between values can be variant because of observed time series nonstationarity. Author decided to modify original preprocessing algorithm to solve a nonstationarity problem.
Idea of the modification is to add steps that provide preprocessing robustness i.e. allow to reduce difference between preprocessing results on different datasets. Preliminary preprocessing functions selection provides robustness. There are two variants of preliminary selection. The first one is following: function with the least difference between correlations between input and output in data subsets pass the selection. The second one is following: functions that increase correlation on both subsets pass the selection.

Experiments on two pumps data prove the hypothesis that data preprocessing in vast majority of cases allows to decrease forecasting error. Modified algorithm often has less test error than original one.