Device health estimation by combining contextual control information with sensor data and device health prognostics utilizing restricted Boltzmann machine

Abstract

The goal of this work is to bridge the gap between business decision-making and real-time factory data. Beyond real-time data collection, we aim to provide analysis capability to obtain insights from the data and converting the learnings into actionable recommendations. For device health estimation, we focus on analyzing device health conditions and propose a data fusion method that combines sensor data with limited diagnostic signals with the device's operating context. We propose a segmentation algorithm that provides a temporal representation of the device's operation context, which is combined with sensor data to facilitate device health estimation. Sensor data is decomposed into features by time-domain and frequency-domain analysis. Principal component analysis is used to project the high-dimensional feature space into a low-dimensional space followed by a linear discriminant analysis to search the optimal separation among different device health conditions. Our industrial experimental results show that by combining device operating context with sensor data, our proposed segmentation and linear transformation approach can accurately identify various device imbalance conditions even for limited sensor data which could not be used to diagnose imbalance on its own. For device health prediction, we propose a restricted Boltzmann machine based method to automatically generate features that can be used for remaining useful life prediction, which is performed by a random forest regression algorithm. The proposed method was validated through run-to-failure dataset of a machine tool spindle test-bed.