Remaining Useful Life Estimation Using CNN-XGB With Extended Time Window

Abstract

The remaining useful life estimation has been widely studied for engineering systems. A system commonly works under varying operating conditions, which may affect the system degradation trajectory differently and consequently reduce the accuracy of remaining useful life estimation. In this paper, we propose CNN-XGB with extended time window to tackle this issue. Firstly, the extended time window is created by feature extension and time window processing in data preprocessing. In feature extension, multiple degradation features are extracted by an improved differential method, and these features are appended to the raw data as additional features. To make the time window cover more information for better prognostic accuracy, a time window padding method is used considering the problem of missing data in some samples. Secondly, a convolutional neural network architecture with multichannel 1 * 1 filter kernel is proposed considering the effect of varying operating conditions. Furthermore, to improve the prognostic robustness and avoid the sensitivity to the abnormal data, convolutional neural network and extreme gradient boosting are fused by model averaging (CNN-XGB). The validity of the proposed method is verified using aero-engine datasets from NASA.INDEX TERMS Neural networks, lifetime estimation, time series analysis, prognostics and health management.

An Optimal Stacking Ensemble for Remaining Useful Life Estimation of Systems Under Multi-Operating Conditions

“…Gradient boosting decision tree (GBDT) is an integrated algorithm based on residual iterative tree [26], which has been widely applied to click-through rate prediction. Extreme gradient boosting (XGBoost) [22] and Light gradient boosting machine (LightGBM) [16] are variants of GBDT. they have achieved good results on other prediction problems.…”

“…Although a single model is easy, it usually suffers a setback for its one-sidedness, domain unity, and intrinsic property. Consider the range fluctuation in time series data caused by unstable environmental variability (e.g., temperature, relative humidity, pressure) [19], diversity of operating modes (e.g.,loading, usage, rotatory speed) [20], nonlinear degradation modes with noise [21], it is difficult to use a single model mapping from signal features to RUL values in a multi-noise and time-varying environment [22]. While the combination of multiple models has the potential for reducing the influence of range fluctuation in degradation data.…”

Automated Machine Learning for Remaining Useful Life Estimation of Aircraft Engines

“…In [11], Li et al addresses this issue by proposing to use 1-dimensional convolution filters in their CNN. Zhang et al [12] investigated the use of CNN with extended time window to tackle the RUL estimation problem under varying operating conditions. Furthermore, to improve the prognostic robustness and avoid the sensitivity to the abnormal data, CNN and extreme gradient boosting (XGB) are fused with model averaging (CNN-XGB).…”

Tool Wear Conditions Monitoring Based on Kernel Smoothing Density and Deep Learning

“…), with adjustment of scan conditions such as sampling frequency, then the recovered data will be stored for further processing which can be expensive in time [33], In a data-driven approach, there are two methods for estimating from historical data, which are direct and indirect. The indirect method relies on the construction of the HI to be able through a model to go back to RUL, by contrast, direct method allows to estimate directly RUL from data [34]. Now, move on to the second stage, which is Health Indicator (HI) construction, in this stage, seek to find an indicator reflecting cutter health stat from the acquired signal, several works have been done in this framework, always looking for an indicator that represents the most monotonicity and reflects the degradation of the component to be studied [35].…”