Causality-Based PCA Methods for Condition Modeling of Mechatronic Systems

Abstract

This paper investigates the feature dimensionality reduction problem of high-dimensional data in condition modelling of complex mechatronic systems, aiming at improving performance, robustness and interpretability of classical principle component analysis (PCA) methods. First, the causal discovery algorithms are adopted to construct the corresponding causal network among monitoring variables. Then, two causality-based PCA methods are proposed, based on the corresponding qualitative reachability matrix and quantitative reachability strength matrix, respectively. Average treatment effect (ATE) values are considered as the quantitative measure of the causalities in a reachability strength matrix. The robustness and effectiveness of the proposed methods are verified on a synthetic dataset and four public datasets in comparison with two benchmark methods, i.e. PCA and autoencoder (AE). The proposed methods are also adopted in a high-speed train braking system for extracting informative features from the monitoring variables.