Just-in-time learning based probabilistic gradient boosting tree for valve failure prognostics

https://doi.org/10.1016/j.ymssp.2020.107253Get rights and content

Highlights

Abstract

Historical failure instances of a system with diversified degradation patterns will pose great challenge for prognostics. Consequently, it is challenging to accurately predict the remaining useful life (RUL) using a prognostic model trained from such data. To solve this problem, this paper proposes a just-in-time learning-based data-driven prognostic method for reciprocating compressors with diverse degradation patterns and operating modes. The proposed framework employs a just-in-time learning (JITL) scheme to deal with the stochastic nature of fault evolution and the diversity of degradation patterns. Moreover, a data-driven forecasting model that features a randomized and smoothed gradient boosting decision tree (RS-GBDT) is developed for RUL and uncertainty predictions. The effectiveness of the proposed approach was validated on temperature measurements collected from 13 valve failure cases of an industrial reciprocating compressor.

Introduction

Health prognostics is at the heart of any condition-based monitoring (CBM) task, which attempts to predict the remaining useful life (RUL) of a system based on the historical and in-process degradation data. The identification and prognostics of faults has received considerable attention from both researchers in academia and engineers in industry. A number of review papers have been written on the subject [1], [2], [3].

Reciprocating compressors play a crucial role in a number of industrial applications, such as gas refineries, refrigerators and compressed natural gas stations; to name a few. These compressors account for a large portion of energy consumption in machinery sectors, thus it is crucial to ensure high levels of availability and reliability in order to reduce production cost and improve operational efficiency. Despite their popularity, owing to the fact that reciprocating compressors constitute high number of moving parts, the maintenance cost of these machines can be several times higher than that of other compressor types, and they are likely to have more failures [4]. Reciprocating compressor failures are frequently caused by breakdown of valves. It has been reported that valves account for more than one third of unplanned shutdown cases and nearly half of the total maintenance cost, rendering them one of the weakest components in a compressor [5]. Valves typically work under adverse conditions, such as rapid opening and closing, high and low flow difference, high temperatures and corrosive gases and liquids, and are thus subject to performance deterioration and failure. Therefore, it is crucial to implement effective maintenance strategies that provide accident prevention and maintenance decision-making. Despite the fact that much research has been dedicated to fault detection of reciprocating compressors, their prognostics still receive little attention.

Owing to the advances in sensing and data storage technology, massive filed data collection from machinery becomes feasible. The collected life-cycle condition data has enabled the implementation of diagnostics and prognostics using data-driven algorithms. However, historical life-cycle instances of machinery often have both diversified degradation patterns and lifespans due to a number of reasons such as various operational modes, engineering variance and environmental conditions, etc. These failure instances, if mixed together, will pose great challenges to RUL prediction. This has encouraged the CBM community to look for advanced analytical techniques. One solution is to employ the just-in-time learning (JITL) methods and develop models that are based on finding historical instances demonstrating similar degradation patterns to the testing data. Attempts have been made [6], [7] to build a fault detection model using the JITL scheme which evaluates the similarity between the current data and the historical database. However, the engineering applications of these JITL-based models mainly focus on fault detection. On the other hand, instance-similarity based methods, which can be treated as another form of JITL scheme, have attracted considerable attention from researchers in the past decade [8], [9], [10]. In these methods, the RUL is estimated as a weighted-sum of the founded similar historical instances, and therefore there is a lack of model generalization and uncertainty quantification. Another solution is to employ the ensemble modelling techniques, such as boosting, to enhance the predictive performance. One approach along these lines which involves the ensemble of several predictors was proposed and tested in [11]. A GBDT-based prognostic model was proposed based on probabilistic mathematical models in [12]. Despite the fact that these methods are capable to approach the actual RUL with a high accuracy, they do not take the diversity of degradation patterns and lifespans into consideration.

In this paper, a novel JITL-data-driven prognostic model consisting of an improved JITL scheme and a randomized and smoothed gradient boosting decision tree (RS-GBDT) is proposed. The JITL scheme accounts for the diversified degradation patterns and various operational modes and facilitates the subsequent predictive model by grouping the instances with similar degradation patterns together. The RS-GBDT approach employs a randomized and smoothed gradient and inherits the merits of high predictive accuracy and narrow probabilistic output. temperature measurements obtained from an in-service reciprocating compressor are utilized to verify the effectiveness of the proposed model.

The main contributions of this work are summarized as follows:1) proposes a JITL-data-driven prognostic model for in-service reciprocating compressors with diversified degradation patterns and various operational modes; 2) develops a JITL scheme which realizes time-series clustering without knowing the number of cluster centers; 3) develops a RS-GBDT model with improved predictive performance and uncertainty quantification; 4) Presents the degradation data of a large-scale in-service compressor.

A brief outline of remaining parts of the paper is as follows: In Section 2, related algorithms and the proposed prognostic model are presented in details. In Section 3 the experimental data is explained, and experimental results as well as detailed discussions about the results are presented. Conclusion remarks are provided in Section 4.

Section snippets

Methodology

The proposed JITL-RS-GBDT prognostic scheme has been realized in two stages. The first stage is based on an improved JITL scheme, which takes the failure trajectories (both historical and up-to-date query data) as inputs and returns the similar failure cases as output. The second stage consists of an improved version of GBDT, which predicts the RUL of the query machine based on the similar trajectories found in the first stage. A brief description of the JITL method is provided in Subsection A.

Description of dataset

In this section, data captured from a two-stage, four-cylinder, double-acting operational reciprocating compressor are presented. The machine experienced thirteen valve failures within a period of 18 months with all failures took place at the fourth cylinder. The root cause of these failures was found to be improper sealing of the valve due to a missing piece from the outer structure of valve plate. These failures occurred at either the head end (HE) or the crank end (CE) discharge valve. Only

Conclusions

We developed a novel JITL-data-driven prognostic model to address the challenging problem of predicting RUL of compressors with diversified degradation patterns and various operational modes. Temperature measurements from the Head and Crank End discharge valves concerning thirteen failures were present. We detailed the theoretical advancement and practical implementation of the developed model. The proposed approach was compared with an existing GBDT model and an enhanced JITL-GBDT method. By

CRediT authorship contribution statement

Xiaochuan Li: Writing - original draft, Conceptualization, Methodology, Software. David Mba: Supervision, Writing - review & editing. Tianran Lin: Supervision, Writing - review & editing. Yingjie Yang: Supervision, Writing - review & editing. Panagiotis Loukopoulos: Data curation.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (30)

  • H.P. Bloch

    A Practical Guide to Compressor Technology

    (2006)
  • S. Yin et al.

    Fault Detection for Nonlinear Process with Deterministic Disturbances: A Just-In-Time Learning Based Data Driven Method

    IEEE Trans. Cybern.

    (2017)
  • L. Yingchao et al.

    Remaining useful life prediction based on health index similarity

    Reliab. Eng. Syst. Saf.

    (2019)
  • T. Wang, J. Yu, D. Siegel, and J. Lee, “A similarity-based prognostics approach for Remaining Useful Life estimation of...
  • T. Loutas N. Eleftheroglou G. Georgoulas P. Loukopoulos D. Mba I. Bennett Valve Failure Prognostics In Reciprocating...
  • Cited by (20)

    • Operation status monitoring of reciprocating compressors based on the fusion of spatio-temporal multiple information

      2022, Measurement: Journal of the International Measurement Confederation
      Citation Excerpt :

      The operation status monitoring can be realized through the analysis and processing of the temperature, pressure, vibration, liquid level and other information collected by sensors at different parts of the compressor. This kind of method contains more compressor operating information and reduces errors caused by vibration and noise [10,11]. Kim et al. [12] uses the temperature, pressure and vibration data of a company's compressor to propose a fault analysis method based on Bayesian networks.

    • An ensemble JITL method based on multi-weighted similarity measures for cold rolling force prediction

      2022, ISA Transactions
      Citation Excerpt :

      Therefore, it is impossible to describe all process characteristics by a fixed prediction model. Inspired by the local modeling [7], the Just-in-time-learning (JITL) modeling method was proposed for the time-variable and nonlinear processes which is based on database and local modeling techniques [8]. JITL method is often used in nonlinear process modeling, which is based on database and local modeling technology.

    • AI-based inspection of the axes of machine tools

      2024, International Journal of Advanced Manufacturing Technology
    View all citing articles on Scopus
    View full text