See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/373160794

Data-driven Predictive Maintenance: a paper making case

Preprint · August 2023

DOI: 10.13140/RG.2.2.23635.35360

CITATIONS READS
0 548

2 authors:

Davide Raffaele Guenter Roehrich

Mondi Mondi AG
5 PUBLICATIONS 8 CITATIONS 4 PUBLICATIONS 2 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Davide Raffaele on 16 August 2023.

The user has requested enhancement of the downloaded file.

 Data-driven Predictive Maintenance: a paper making case
Davide Raffaele Guenter Roehrich
Mondi Group Mondi Group
Marxergasse 4a Marxergasse 4a
1030, Vienna, Austria 1030, Vienna, Austria
davide.raffaele@mondigroup.com Guenter.Roehrich @mondigroup.com

ABSTRACT their break can cause significant production losses and even
Condition monitoring together with predictive maintenance of damages to the machine. The problem is how to predict the state of
bearings and other equipment used by the industry avoids severe a bearing using vibration data.
economic losses resulting from unexpected failures, greatly In predictive maintenance three kinds of approaches can be
improves the system reliability and allows a more efficient usage distinguished [1]: 1) Data-driven approach, 2) Model-based
of human experts’ time. This paper describes a predictive approach, and 3) Hybrid approach.
maintenance system, based on a data science approach. The system
was developed and tested on a single real paper machine, and then While the data-driven approach uses historical data to learn patterns
verified with multiple external validations. Results show a proper and system behaviour. The model-based approach is expert-driven
behaviour of the approach on predicting different machine states and has the ability to incorporate physical understanding of the
with high accuracy. target product, relying on the analytical model to represent the
behaviour of the system. A hybrid approach data combines data-
driven as week as model-based approach is also found in the
Keywords literature [5] . With the exponential increase of data being collected
Predictive maintenance, Data Science, Industry 4.0, Vibration and available in production environment, the use of data-driven
measurements, Data Analysis, Bearings, Condition Monitoring. approach is increasing [6], [7] and will be the focus of this paper.
The authors intend to build upon the aforementioned ideas and
1. INTRODUCTION present a method for early-stage detection of degradation in
Predictive maintenance, also known as "on-line asset monitoring", bearings. To increase efficiency and prevent downtimes it is crucial
or "smart condition-based maintenance", involves the intelligent to decide which bearings to replace at a given point in time. The
monitoring of equipment to prevent future failures and has received contributions in this paper primarily address the application of data
growing attention in research. In the past decade, predictive science approaches to real-world data related to paper machines on
maintenance has advanced from relying on visual inspection the field, the high level of accuracy on predictive state of the
methods to automated ones that leverage advanced signal bearings and, moreover, the specificity of the application for the
processing techniques such as pattern recognition, machine papermaking industry.
learning, neural networks, and fuzzy logic. These automated
methods offer viable solutions for various industries by detecting This paper is organized as follows: Section 2 gives an
and collecting sensitive information from equipment, where human overview of predictive maintenance approach used in literature.
observation may fall short [1] [2]. Integrated sensors and predictive Section 3 provides an introduction to domain knowledge of
maintenance can work together to prevent unnecessary equipment bearings and their faults modes. Section 4 provides a sound
replacement, reduce machine downtime, pinpoint the root cause of foundation for the methodology followed in this paper. Section 5
faults, and ultimately save costs and improve efficiency. Predictive demonstrates the experimental results based on the real use case
maintenance shares similarities with preventive maintenance, as under which predictive maintenance has been implemented and
both involve scheduling maintenance activities in advance to avoid tested. Finally, Section 6 provides a reflection on the results that
machine failures. However, unlike traditional preventive have been achieved and compares them to the current state of the
maintenance, predictive maintenance schedules activities based on literature.
data collected from sensors and analysed by algorithms. [3], [4] [1].
Traditionally, predictive maintenance aims to schedule 2. RELATED WORK
interventions on a machine based on health condition predictions Paper industry has an extensive literature on data-driven
derived from high frequency data collected by sensors. However, methodologies, but it rather focusses on virtual sensors [8] [9] [10]
in the paper industry, paper machines require to stop production at and not on predictive maintenance. For this reason, the authors
regular intervals for technical reasons (e.g. felts changes), which intend to hold the rest of the section industry-agnostic.
creates a great opportunity for preventive maintenance actions. In
Bearing degradation modelling methods can be categorized into
this context, predictive maintenance is no longer a regression
two groups: continuous degradation, focused on building a single
problem: when to schedule maintenance? It becomes a
model to capture the degradation, and discrete degradation stage
classification problem: what parts of a machine should be
models, typically classification models that predict degradation
exchanged at a given point in time?
changes over time.
We focus on one type of machine part: rolling-element bearings.
Bearing products are important components of paper machine as
Discrete degradation stage modelling is faced with the challenge of The term multiple-point defects refers to the occurrence of more
identifying multiple degradation stages. Methods such as hidden than one single point defect in a bearing. This type of fault results
Markov models [11] or Mahalanobis Taguchi System [12] are used in variations in the magnitude of the frequency in the frequency
to characterize bearing degradation stages based on observations. domain. The position of the defects can either reinforce or oppose
These approaches consider low dimensional feature vector inputs these frequency variations [24].
since model parameter estimation in high dimensional settings is
challenging [13]. Low dimensional inputs reduce model predictive
power, as information is lost when bearing vibration signal is
embedded intpo lower dimensions. Our work is different from these
models as we attempt to build a decision-support system that can
provide a health assessment of each bearing as continuum, which
serves as feature required for ranking-comparison of bearing status
at any given moment.
The continuous degradation is often referred as health index [14]
and is typically calculated from a single feature, or a combination
of features derived from bearing vibration data and fused together,
using e.g., RMS [15], kurtosis values [16], power densities [17], or
statistical methods [18]. A fault is detected when the health index Figure 1 - Structure of a rolling bearing
exceeds a predefined threshold value. This is a critical decision in Finally, distributed faults typically result from contamination, loss
the system, yet setting the threshold value is not a trivial task. The of lubrication, or coupling misalignment. These faults lead to a
work describes this body of literature and builds upon the roughening of the bearing surface, also known as "generalized
aforementioned methods. roughness". Unlike multiple-point defects, distributed faults cannot
be decomposed into distinct single-point defects. As a
3. DOMAIN KNOWLEDGE consequence, characteristic frequencies may not be readily
The Cross Industry Standard Process for Data Mining (CRISP-DM) detectable, or they may not exist at all [25].
is a widely recognized model for data mining, which emphasizes
the importance of having technical knowledge to effectively 4. METHODOLGY
classify the results of a data science approach and its underlying The basic scheme of a machinery health prognostic program is
process [19]. For instance, when dealing with bearing faults, having composed of following technical processes [26]: data acquisition,
an understanding of the technical background of bearings is crucial health indicator (HI) construction, health stage (HS) division.
for properly identifying, assessing and characterizing different
types of faults that may occur. Therefore, this section provides an 4.1 Data acquisition
in-depth exploration of the structural components of a bearing and Data acquisition is the procedure of acquiring and recording
the various types of faults that can manifest within the bearing various types of monitoring data from a multitude of sensors that
system are installed on the equipment under surveillance. This crucial
process serves as the initial step in machinery prognostics, and
Bearings consist of four essential components, namely the rolling
provides fundamental condition monitoring information that forms
elements, cage, inner ring, and outer ring. Rolling elements,
the basis for subsequent analysis. A data acquisition system
typically steel balls, are located between the outer and inner rings
consists of several components, including sensors that measure and
and are responsible for reducing friction between moving parts as
capture data, data transmission devices that transmit the data to a
shown in Figure 1. The cage maintains the relative positions of the
storage location, as well as data storage devices that store the
rolling elements, reducing friction and therefore preventing them
captured information. In order to build a predictive maintenance
from colliding or rubbing against each other. While a bearing fault
model, the most important data to acquire has to capture breakdown
can arise from any of the four components, approximately 90% of
events of the equipment. However, acquiring high-quality run-to-
faults are associated with the inner and outer rings [20]. This may
failure data for academic research on machinery remains a
be attributed to the permanent stress that the rings are subjected to,
challenge due to several reasons [26]. Firstly, machinery undergoes
whereas the rolling elements are in continuous motion, causing
a prolonged degradation process from a healthy state to failure,
their contact area to continually change. In addition, the cage does
which can last several months or even years. Gathering complete
not bear any load [21].
run-to-failure data during this extended period can be both time-
Fault can occur in three different forms [22]: A) Single-point consuming and expensive. Secondly, practical considerations
defects, B) Multiple-point defects, C) Distributed faults. prohibit allowing machinery to run to failure, as unexpected failure
can result in machine breakdown or catastrophic accidents. This
In the event of a defect occurring at a singular location on a bearing, makes it difficult to capture run-to-failure data in industrial settings.
all other components of the bearing remain in sound condition, with Thirdly, machinery, such as gearboxes, engines or bearings,
the exception of the specific point where the defect arises, resulting operates in harsh environments that introduce significant external
in an amplified magnitude of the characteristic frequencies interferences to monitoring data, thereby reducing its quality.
associated with the particular fault type. Examples of this fault Fourthly, much monitoring data are collected during out-of-service
classification include the emergence of pits, spalls, and cracks on periods, such as downtime or restart, which exhibit different
the outer and inner ring, balls, and cage of the bearing. The majority behaviours compared to in-service period measurements, further
of contemporary research articles are primarily on the investigation reducing the quality of the monitoring data. In our research, the
of single point defects [23]. authors reached out to all paper mills and plants within the Mondi
Group, to identify where the best data were available and
collaborated with local expert to ensure the highest possible quality
of those before the start of the work. Due to commercial system where the unhealthy stage begins when the health-index
competition, the original dataset cannot be shared, but the exceeds a pre-specified alarm threshold based on historical initial
publication should assist to find evidence that with the correct points of defects for bearings. More complex methods for a
approach a predictive maintenance system can be incredibility threshold-defined two-stage division are found in the literature
successful. This dataset is composed of 51 different bearing, and (Chebyshev inequality function [45], 3 Box-Cox transformation
for each the channels of acceleration, envelope and velocity were [46], Hotelling T2 statistic [47], locality preserving projection [48])
considered. Data were collected for 24 months, with 1h-sampled but complexity does not come at benefits of precision, as no method
data, providing a total of more than 850.000 data-points to analyse. outperforms the others. When variations in fault patterns or
operational conditions occur, the degradation trends of machinery
4.2 Health Indicator construction may change, leading to an additional layer of complexity. In such
The process of developing a Health Index (HI) holds a crucial role cases, it becomes challenging to accurately describe the
in the field of machinery prognostics. An appropriate HI can degradation processes using a signal degradation model.
significantly enhance the accuracy of prognostic modelling, Consequently, it is necessary to further subdivide the unhealthy
thereby improving the precision of the prediction results. Health stage into different stages based on the various degradation trends.
indexes can be classified into two categories based on their Some researchers have addressed this issue by dividing the
construction methods: physical HIs and virtual HIs. Physical HIs degradation processes into multiple stages through the analysis of
are associated with the physics of failure and are typically derived change points in health indexes or spectra [14]. The number of
from monitoring signals using statistical or signal processing identified levels and the methods used for their identification vary
techniques. In contrast, virtual HIs are often created by merging and sophistication raises compared to the two-stage case. From
multiple physical HIs or multi-sensor signals, resulting in a loss of using confidence levels to build a four-stage system [49], to analyse
physical significance and offering only a virtual representation of changes of frequency amplitudes in the power spectral density to
the machinery's degradation trends [27]. develop a five-stage model [50], and apply classification or
RMS is the most commonly employed physical Health Index in the clustering algorithms, such as K-nearest neighbour [51], fuzzy c-
prediction of remaining useful life for machinery. [28] predicted means [52] and K-means [53], to develop the multi-stage division
the remaining useful life of bearings using RMS applied to of machinery, the available methodologies for researchers are
vibration data. [29] extracted RMS and peak values from the numerous. For those industrial-oriented real applications though,
wavelet coefficients as a means of predicting the remaining useful semi-supervised learning has recently emerged as a prominent
life of bearings. Similarly, approaches set out it [30] [31] also methodology [54] [55] [56], as in practical applications, obtaining
applied RMS to predict bearing anomalies in machineries. Some accurate labels based on real-time bearing conditions can be far
researchers constructed new physical Health Indexes based on more challenging and semi-supervised approach allows for
statistical characteristics of signals in time domain. To evaluate the effective utilization of dataset when only a small subset of data have
evolution of cracks in gear teeth, [32] quantified the proportion of labels. In this paper, given the practical usage of the outcome, the
a residual error signal that exceeded a baseline threshold. [33] latter approach is further developed and applied.
developed a Physical Health Index for bearings by computing the
energy ratio between residual signals obtained through 5. RESULTS
autoregressive (AR) filtering and the original signals. On the side The dataset used in this study consisted of the complete data records
of virtual Health Index instead, the most popular technique is of approximately 50 bearings that were monitored over a period of
certainly PCA [14]. [34] used PCA to reduce the dimension of 2 years across 3 distinct paper machines situated in the same mill.
feature sets and further calculated the deviations between unknown In order to establish the methodology, data from 23 bearings on a
states and the healthy state as a virtual Health Index. [35] used PCA single machine were exclusively analysed. Subsequently, the
combined with isometric feature mapping to construct a VHI for developed methods were tested on the remaining machines to
cutting tools. When dimensionality is a challenge however, [36] validate their applicability and generalizability.
innovated with the usage of self-organizing map techniques, and
since then this technique has been widely used in the VHI
construction [37] [38] [26]. When faced with limited features, it is
very popular in the literature [39] [40] [41] [42] to use a statistical,
linear or non-linear, data transformation approach to create a virtual
Health Index by combining multiple features. This latter has a clear
advantage of understandability, which makes it more reliable to
validate. Given the practical objective of the health index proposed
in this paper, we chose to adopt the aforementioned method.

4.3 Health Stage division

The concept of Health Stage division is akin to the commonly used
terms "fault detection" or "fault diagnosis" in the field of
prognostics and health management [43]. However, their objectives
differ. Fault diagnosis aims to identify the specific fault patterns
and severity of a given machine at a single point in time. In contrast, Figure 2 - Bearing vibration signal. (a) raw (b) (c) filtered
Health Stage division seeks to partition the ongoing degradation To extract valuable insights from the vast amount of data in
process of a machine into distinct Health Stages based on the vibration-based Smart Maintenance, appropriate filtering
fluctuating trends of health indicators [14] techniques must be employed to mitigate the impact of noise on the
A simple but empirically-solid strategy for health stage division is analysis. Noise, which can arise from environmental factors or
presented by Wang [44], who developed a two-stage division intentional sources, can interfere with the accuracy of the system,
necessitating its active management. A reliable method for noise
reduction in the time domain is the application of Finite Impulse
Response (FIR) moving average and standard deviation filters.
These filters are not only simple and fast but also provide
remarkable results by eliminating noise while retaining essential
information and sharp step responses, as demonstrated in Figure 2

Figure 3 - Euclidean distance calculation (where p and q are

two points)
Figure 6 - Resulting threshold based on failure case. Red arrow
The process of developing a virtual health index commences with = moment of bearing failure; yellow/orange arrow = signal
the calculation of the Euclidean distance between the signal's quiet pattern change. Yellow/Orange threshold defined as warning
value of each bearing and that of the signal at time t, as depicted in and alarm stage division.
Figure 3, for each filtered component of the relative vibration
channel signal. Multiple features become available for each time t, remarkable. In addition to the case used for generalization, three
as shown in Figure 4. Following the methodology described in the more bearing cases were available, and all were correctly identified
by the developed methodology without any false positives or false
negatives (see Figure 7). To ensure correct fitting, the method was
also tested on the remaining two machines without recalibration.

Figure 4 - Features for VHI developed by raw signal filter and

Euclidean distance calculation
Figure 7 - Prediction result on development paper machine
Health Indicator construction chapter [39] [40] [41] [42], which
The results, as depicted in Figure 8, were equally remarkable, with
incorporates the use of a non-linear minimum function, the data can
no false positives or false negatives identified, even in instances
be aggregated to create a virtual Health Index, as illustrated in
when no failure cases were observed on the machine.
Figure 5.

Figure 5 - Virtual Health Index for a single bearing as derived

by signal displayed in Figure 5
The final step involves dividing the Health Index into various
health stages. As described in Chapter 4, a semi-supervised
approach is implemented due to the scarcity of labelled data
available in the dataset, which is limited by the practical run-to-
failure cases, as discussed in section 4.1. The threshold for health
stage division is determined based on the patterns observed in the
failure case depicted in Figure 5, and explained in Figure 6.
The availability of data from multiple paper machines allows for
cross-validation to assess the extent to which the developed
methodology can be generalized. The performance of the
methodology on the machine used for method development was

Figure 8 - Methodology testing on unseen/unbiased data

6. CONCLUSIONS AND FUTURE WORKS sensor for monitoring ash content at two-ply paperboard
In this study, we have presented a methodology for implementing manufacturing,“ Nordic Pulp & Paper Research Journal, pp.
predictive maintenance utilizing a vast array of vibration data, 175-183, 2022.
despite limited availability of failure labels. Our approach was [11] M. Yuwono, Y. Qin, J. Zhou, Y. Guo und B. Celler,
validated on multiple paper machines and demonstrated „Automatic bearingfault diagnosis using particle swarm
exceptional performance on various test and validation datasets. clustering and Hidden Markov Model,“ Eng.Appl. Artif.
Unfortunately, due to commercial confidentiality, we are unable to Intell., p. 88–100, 2016.
disclose the original dataset. Nonetheless, our findings serve as a [12] A. Soylemezoglu, S. Jagannathan und C. Saygin,
valuable contribution towards demonstrating the effectiveness of a „Mahalanobis taguchi system(MTS) as a prognostics tool for
well-designed predictive maintenance system. rolling element bearing failures,“ J. Manuf. Sci.Eng., 2010.
[13] N. Stadler und S. Mukherjee, „Penalized estimation in
7. ACKNOWLEDGMENTS highdimensional hidden markov models with state-specific
This work was made possible thanks to the support of Mondi Group graphical models,“ The Annals of Applied Statistics, pp. 2157
AG. A special thanks to the colleagues from the plants and the - 2179, 2013.
headquarter that committed their time and knowledge to the success
[14] Y. Lei, N. Li, L. Guo, N. Li, T. Yan und J. Lin, „Machinery
of this work. health prognostics: A systematic review from data
acquisition to rul prediction,“ Mechanical Systems and
8. REFERENCES Signal, pp. 799-834, 2018.
[1] M. Paolanti, L. Romeo, A. Felicetti, A. Mancini, E. Frontoni [15] Z. Huang, Z. Xu, X. Ke, W. Wang und Y. Sun, „Remaining
und J. Loncarski, „Machine Learning approach for Predictive useful life prediction for an adaptive skew-Wiener process
Maintenance in Industry 4.0,“ 14th IEEE/ASME model,“ Mech. Syst. Signal Process, pp. 294-306, 2017.
International Conference on Mechatronic and Embedded
[16] Z. Zhang, X. Si und C. Hu, „An age- and state-dependent
Systems and Applications (MESA), 2018.
nonlinear prognostic model for degrading systems,“ IEEE
[2] H. M. Hashemian und W. C. Bean, „State-of-the-art Trans. Reliab., pp. 1214-1228, 2015.
predictive maintenance techniques,“ EEE Transactions on
[17] M. Gašperin, Ð. Juricˇic´ , P. B. Boškoski und J. Vizˇin,
Instrumentation and measurement, vol. 60, no. 10, p. 3480–
„Model-based prognostics of gear health using stochastic
3492, 2011.
dynamical models,“ Mech. Syst. Signal Process., pp. 537-
[3] S.-j. Wu, N. Gebraeel, A. Lawley und Y. Yih, „A neural 548, 2011.
network integrated decision support system for condition-
[18] J. Hu und P. Tse, „A relevance vector machine-based
based optimal predictive maintenance policy,“ IEEE
approach with application to oil sand pump prognostics,“
Transactions on Systems, Man, and Cybernetics-Part A:
Sensors, pp. 12663 - 12686, 2013.
Systems and Humans, vol. 37, no. 2,, p. 226–236, 2007.
[19] W. Rüdiger und J. Hipp, „Crisp-dm: towards a standard
[4] E. Frontoni, R. Pollini, P. Russo, P. Zingaretti und G. Cerri,
process modell for data mining,“ Computer Science, 2000.
„Hdomo: Smart sensor integration for an active and
independent longevity of the elderly,“ Sensors, vol. 17, no. [20] D. Bently, Predictive Maintenance Through the Monitoring
11, p. 2610, 2017. and Diagnosticsof Rolling Element Bearings, Application
Note 44, 1989.
[5] C. Chen, G. Vachtsevanos und M. E. Orchard, „Machine
remaining useful life prediction: An integrated adaptive [21] Schaeffler Technologies A.G. & Co. KG, Manualfor the
neuro-fuzzy and high-order particle filtering approach,“ design and calculation of rolling bearings, 2015.
Mechanical Systems and Signal Processing, pp. 597-607, [22] S. Schwendemann, Z. Amjad und A. Sikora, „A survey of
2012. machine-learning techniques for condition monitoringand
[6] M. Paolanti, E. Frontoni, A. Mancini, R. Pierdicca und P. predictive maintenance of bearings in grinding machines,“
Zingaretti, „Automatic classification for anti mixup events in Computers in Industry, Bd. 125, p. 103380, 2021.
advanced manufacturing system,“ ASME 2015 International [23] Y. Li, H. Cao und X. Chen, „Modelling and vibration
Design Engineering Technical Conferences and Computers analysis of machine tool spin-dle system with bearing
and Information in Engineering Conference, 2015. defects,“ International Journal of Mechatronics and
[7] S. Naspetti, R. Pierdicca, S. Mandolesi, M. Paolanti, E. Manufacturing Systems Vol. 8, No. 1-2, pp. 33-48, 2015.
Frontoni und R. Zanoli, „Automatic analysis of eye-tracking [24] P. McFadden und J. Smith, „The vibration produced by
data for augmented reality applications: A prospective multiple point defectsin a rolling element bearing,“ Journal
outlook,“ International Conference on Augmented Reality, of Sound Vibration, Bd. 98, Nr. 2, p. 263–273, 1985.
Virtual Reality and Computer Graphics, 2016. [25] F. Dalvand, M. Kang, S. Dalvand und M. Pecht, „Detection
[8] D. Raffaele und T. Ondruch, „Data-driven soft sensor for of generalized-roughness and single-point bearing faults
continuous production monitoring: an application to paper using linear prediction-based currentnoise cancellation,“
strength,“ in 2020 25th IEEE International Conference on IEEE Transactions on Industrial Electronics, Bd. 65, Nr. 12,
Emerging Technologies and Factory Automation (ETFA), pp. 9728 - 9738, 2018.
Vienna, 2020. [26] Y. Lei, Intelligent Fault Diagnosis and Remaining Useful
[9] Z. He, J. Qian, J. Li, M. Hong und Y. Man, „Data-driven soft Life Prediction of Rotating Machinery, Oxford: Elsevier
sensors of papermaking process and its application to cleaner Butterworth-Heinemann, 2016.
production with multi-objective optimization,“ Journal of [27] C. Hu, B. D. Youn, P. Wang und J. T. Yoon, „Ensemble of
Cleaner Production, 2022. data-driven prognostic algorithms for robust prediction of
[10] T. Jani, O. Markku, M. Ritala, R. Risto, M. Markku, G. Jussi, remaining useful life,“ Reliability Engineering & System
V. Mikko, J. Teijo und R. Mika, „Mass-balance based soft Safety, Bd. 103, pp. 120-135, 2012.
[28] Z. Huang, Z. Xu, X. Ke, W. Wang und Y. Sun, „Remaining [43] J. Pellegrino, M. Justiniano und A. Raghunathan,
useful life prediction for an adaptive skew-Wiener process „Measurement Science Roadmap for Prognostics and Health
model,“ Mechanical Systems and Signal Processing, Bd. 87, Management for Smart Manufacturing Systems,“ National
pp. 294-306, 2017. Institute of Standards and Technology, , Gaithersburg, 2016.
[29] A. Malh, R. Yan und R. X. Gao, „Prognosis of Defect [44] W. Wang, „A model to predict the residual life of rolling
Propagation Based on Recurrent Neural Networks,“ IEEE element bearings given monitored condition information to
Transactions on Instrumentation and Measurement, Bd. 60, date,“ IMA Journal of Management Mathematics, Bd. 13, pp.
Nr. 3, pp. 703 - 711, 2011 . 3 - 16, 2022.
[30] L. Hu, N.-q. Hu, B. Fan, F.-s. Gu und X.-y. Zhang, [45] Y. Qian, R. Yan und S. Hu, „Bearing Degradation Evaluation
„Modeling the Relationship between Vibration Features and Using Recurrence Quantification Analysis and Kalman
Condition Parameters Using Relevance Vector Machines for Filter,“ IEEE Transactions on Instrumentation and
Health Monitoring of Rolling Element Bearings under Measurement , Bd. 63, Nr. 11, pp. 2599 - 2610, 2014.
Varying Operation Conditions,“ Mathematical Problems in [46] X. Jin, Y. Sun, Z. Que, Y. Wang und T. W. S. Chow,
Engineering, Bd. vol. 2015, 2015. „Anomaly detection and fault prognosis for bearings,“ IEEE
[31] H. Liao und Z. Tian, „A framework for predicting the Transactions on Instrumentation and Measurement, Bd. 65,
remaining useful life of a single unit under time-varying Nr. 9, pp. 2046-2054, 2016.
operating conditions,“ IIE Transactions, Bd. 45, Nr. 9, pp. [47] A. Ajami und M. Daneshvar, „Data driven approach for fault
964-980, 2013. detection and diagnosis of turbine in thermal power plant
[32] D. Lin, M. Wiseman, D. Banjevic und . A. K. Jardine, „An using Independent Component Analysis (ICA),“
approach to signal processing and condition-based International Journal of Electrical Power & Energy Systems,
maintenance for gearboxes subject to tooth failure,“ Bd. 43, Nr. 1, pp. 728-735, 2012.
Mechanical Systems and Signal Processing, Bd. 18, Nr. 5, [48] „Bearing performance degradation assessment using locality
pp. 993-1007, 2004. preserving projections and Gaussian mixture models,“
[33] R. Li, P. Sopon und D. He, „Fault features extraction for Mechanical Systems and Signal Processing, Bd. 25, Nr. 7,
bearing prognostics,“ Journal of Intelligent Manufacturing , pp. 2573-2588, 2011.
Bd. 23, p. 313–321, 2012. [49] Y. Hu, H. Li , X. Liao, E. Song, H. Liu und Z. Chen, „A
[34] A. Widodo und . B.-S. Yang, „Application of relevance probability evaluation method of early deterioration
vector machine and survival probability to machine condition for the critical components of wind turbine
degradation assessment,“ Expert Systems with Applications, generator systems,“ Mechanical Systems and Signal
Bd. 38, Nr. 3, pp. 2592-2599, 2011. Processing, Bde. %1 von %276-77, pp. 729-741, 2016.
[35] T. Benkedjouh, . K. Medjaher, N. Zerhouni und S. Rechak , [50] J. Kimotho,, C. Sondermann-Wölke, . T. Meyer und W.
„Health assessment and life prediction of cutting tools based Sextro , „Machinery Prognostic Method Based on Multi-
on support vector regression,“ Journal of Intelligent Class Support Vector Machines and Hybrid Differential
Manufacturing, Bd. 26, p. 213–223, 2013. Evolution – Particle Swarm Optimization,“ Chemical
[36] H. Qiu, J. Lee , J. Lin und G. Yu, „Robust performance engineering transactions, Bd. 33, pp. 619-624, 2013.
degradation assessment methods for enhanced rolling [51] E. Ramasso, M. Rombaut und N. Zerhouni , „Joint Prediction
element bearing prognostics,“ Advanced Engineering of Continuous and Discrete States in Time-Series Based on
Informatics, Bd. 17, Nr. 3-4, pp. 127-140, 2003. Belief Functions,“ IEEE Transactions on Cybernetics, Bd. 1,
[37] S. Hong , Z. Zhou , E. Zio und K. Hong, „Condition pp. 37-50, 2013.
assessment for the performance degradation of bearing based [52] K. Javed, R. Gouriveau und N. Zerhouni, „A New
on a combinatorial feature extraction method,“ Digital Signal Multivariate Approach for Prognostics Based on Extreme
Processing, Bd. 27, pp. 159-166, 2014. Learning Machine and Fuzzy Clustering,“ IEEE
[38] S. Hong, Z. Zhou, E. Zio und W. Wang, „An adaptive Transactions on Cybernetics , Bd. 45, Nr. 12, pp. 2626 -
method for health trend prediction of rotating bearings,“ 26394, 2015.
Digital Signal Processing, Bd. 35, pp. 117-123, 2014. [53] P. Scanlon, D. F. Kavanagh und F. Boland, „Residual Life
[39] Z. Xi , R. Jing, P. Wang und C. Hu, „A copula-based Prediction of Rotating Machines Using Acoustic Noise
sampling method for data-driven prognostics,“ Reliability Signals,“ IEEE Transactions on Instrumentation and
Engineering & System Safety, Bd. 132, pp. 72-82, 2014. Measurement, Bd. 62, Nr. 1, pp. 95-108, 2012.
[40] „A similarity-based prognostics approach for Remaining [54] C. Liu und K. Gryllias, „A semi-supervised Support Vector
Useful Life estimation of engineered systems,“ in 2008 Data Description-based fault detection method for rolling
International Conference on Prognostics and Health element bearings based on cyclic spectral analysis,“
Management, Denver, CO, USA, 2008. Mechanical Systems and Signal Processing, Bd. 140, 2020.
[41] P. Wang, . B. D. Youn und C. Hu, „A generic probabilistic [55] S. H. Wang, S. B. Xing, Y. G. Lei, N. Lu und N. P. Li,
framework for structural health prognostics and uncertainty „Vibration indicator-based graph convolutional network for
management,“ Mechanical Systems and Signal Processing, semi-supervised bearing fault diagnosis,“ in IOP Conference
Bd. 28, pp. 622-637, 2012. Series: Materials Science and Engineering, 2021 .
[42] C. Hu, . B. D. Youn, P. Wang und . J. T. Yoon, „Ensemble [56] D. N. Avendano, N. Vandermoortele, C. Soete , P. Moens,
of data-driven prognostic algorithms for robust prediction of A. P. Ompusunggu , D. Deschrijver und S. Van Hoecke, „A
remaining useful life,“ Reliability Engineering & System Semi-Supervised Approach with Monotonic Constraints for
Safety, Bd. 103, pp. 120-135, 2012. Improved Remaining Useful Life Estimation,“ Sensors , Bd.
22, 2022.
View publication stats