Machine Learning-Based Embedding for Discontinuous Time Series Machine Data

This paper presents a machine learning-based dimension reduction framework (ML-framework). The ML-framework is designed to circumvent the challenges of high-dimensional discontinuous machine data applied in machine learning-based predictive maintenance analysis. To circumvent high-dimensionality and discontinuity in machine data, the ML-framework minimizes discontinuity by defining point clusters based on the dataset’s modality defined by a kernel density estimation (KDE). The bandwidth of the KDE is parameterized through a solution of the approximate mean integrated squared error (AMISE) obtained using the heat equation. Then, low-dimensional representations of each cluster are learned using Laplacian eigenmaps. Finally, the original time sequence of each observation across the low-dimensional clusters is used to re-index the disjointed low-dimension representations into a continuous low-dimension feature set. We demonstrate the ML-framework’s utility on common machine learning-based predictive maintenance analysis using machine data.