Beruflich Dokumente
Kultur Dokumente
Abstract The estimation of remaining useful life applied to monitoring systems based on computational intelligence are
industrial machinery and its components is one of the current being applied on multiple electromechanical systems. Several
trends in the advanced manufacturing field. In this context, this approaches to improve the estimation of the RUL have been
work presents a reliable methodology applied to ball bearings proposed in the literature.
health monitoring. First, the proposed methodology analyses the Numerous methods can be used to predict the RULs value
available vibration and temperature data by means of the with promising results [4-9]. These methods can be classified
Spearman coefficient. This step allows the identification of the into two principal approaches: model-based prognostics and
most significant monotonic relationship between features and the data-driven prognostics. Model based uses mathematical
evolution of the remaining useful life. The method is
models to represent bearing dynamic behavior and degradation
complemented by means of the application of one-class support
phenomenon. One drawback of these methods is that you must
vector machine in order to obtain the remaining useful life
indication trough the mapping of the classification scores. The have a reliable physical model for the fault degradation. For
proposed scheme shows a significant accuracy and reliability of most real-life signals and systems, including bearings, a
the degradation detection due to the coherent management of the reliable physical model for the degradation process is not
information. This fact is experimentally demonstrated by a run- available. The data-driven approach uses measured data
to-failure test bench and the comparison with classical extracted from the sensor as source for getting a better
approaches. understanding of the bearing phenomenon. Some of these
methods still lack of generalization capabilities to estimate the
Index TermsArtificial Intelligence; Remeaning Useful Life; RUL with test sets that are different from the training set, or
Machine Learning; One Class Support Vector Machines; rely on feature reduction techniques that does not exploit the
Classification Algorithms. appropriate characteristics to estimate the RUL[6, 7].
The originality of this work includes a complete remaining
I. INTRODUCTION useful life scheme applied to ball bearings in an
Condition Based Maintenance (CBM) has become electromechanical system. The method begins with the
accepted by the industry as a key factor to avoiding expensive management of the available raw data, in this case temperature
unplanned machine stoppages and maintaining high and vibration. Contrary to most of the classical approaches,
production ratios. The growing importance of computer this stage is focused taking into account the expected behavior
integrated manufacturing is increasing the necessity to of the component degradation profile. This fact allows a better
improve the performance of machine monitoring and fault management of the physical behavior of the component under
detection strategies. monitoring, from the numerical feature selection point of
One of the bases of an effective CBM strategy is the
view. Next, the implementation of the remaining useful life
accurate assessment of machine component condition. A
estimator is done by means of a one-class support vector
common strategy involves obtaining the condition profile of
the component to be monitored by the acquisition of some machine, which takes advantage of the previously selected
significant physical magnitude and, then the remaining useful information.
life time (RUL) of the component could be identified or This paper has been organized as follows. In section II the
predicted [1, 2]. In these cases, after a numerical feature One-Class Support Vector Machine basis and the output
calculations stage, a feature reduction technique is often used calibration strategy are explained. In section III each part of
to obtain a better representation of the data, aiming to obtain a the proposed methodology is shown, including the training
data distribution where the different conditions or classes of stage, which is composed by the following stages: feature
the monitored component could be clearly addressed. Then, a calculation, feature selection, training of the model, and the
classifier is trained with such training profile, in order to test stage, in which new measurements are evaluated. The
identify the corresponding percentage of degradation of the proposed strategy is tested with a bearing degradation dataset
component by the assessment of new measurements [3]. In and the results are presented in section IV. Moreover, the
regard with the industrial applications, these condition performance of the proposed methodology is compared with
can be formulated as
=
ns
a
i =1 j =1
j K ( xi , x j ) a (9)
1753
calibration of a classifier, which consist of the plot of the In this work, the RUL is defined as a percentage
classifier score versus the desired output (probabilistic, representing the remaining life-time of the bearing under
application specific parameter, etc.). If the classifier is well- operating conditions. Taking into account that the data used
calibrated the plot will be a line, however, for many datasets for experimental validation consist of run-to-failure bearing
this assumption is not always true and would lead to a miss- degradation experiments (called bearing profiles), the RUL
calibration. Binning is a histogram method for calibrating the would be 100% at the start of the experiment and decrease
output, and samples are analyzed according to the classifier linearly to 0% at the end of the experiment, which implies the
score that belongs to certain bin. The number of bins is usually bearing is no longer functional. A graphical explanation of the
RUL interpretation is shown in figure 2.
calculated by cross-validation which would often fail if the
training set is too small or unbalanced. The fixed size of bin is 100%
also a problem, which implies losing resolution in some zones. Estimated
Isotonic regression was proposed as an intermediary approach RUL
between sigmoid fitting and binning. This method is a non-
parametric form of regression with the restriction that the RUL Actual
mapping from scores in the application output is isotonic (i.e. RUL
non-decreasing). The pool or pair-adjacent violators (PAV)
algorithm is employed to perform the isotonic regression, 0%
which finds the stepwise-constant isotonic function g (x) Time
that fits the data according to a mean-squared-error criterion. Fig. 2. Definition of RUL on run-to-failure experiments. The time axis
represent the duration of the experiment until the bearing is no functional.
As a general explanation, the isotonic function fits the output
of the SVM classifier to obtain a monotonically output that
Taking into account that the objective of this work is the
could be associated with the application output. This is very RUL estimation, the presented methodology will consider the
useful in applications where the intrinsic behaviour of the properties of the RUL definition, which is monotonically
system is monotonic and a model with this property is decreasing over the experiment time. This fact is critical,
necessary to estimate a certain characteristic of the system. because will be used to emphasize this property during the
This method can be viewed as a binning algorithm, in which characterization of the bearing degradation profile.
the position and the size of the bins are chosen according to The presented methodology is divided in training phase
how well the classifier ranks the samples. Therefore, the and test phase. On the training phase, only a representative set
isotonic regression overcomes the drawbacks of the binning of a bearing degradation profile is used, this will be called
method. A more detailed explanation of the isotonic regression training set. The objective of this phase is to train a model
and the PAV algorithm can be found in [14]. capable to estimate the RUL of the bearing profile used as
training set with enough generalization to accurately estimate
III. REMAINING USEFUL LIFE ESTIMATION METHODOLOGY the RUL on different testing sets.
In this section the proposed methodology to estimate the Previous to the model training, a feature calculation step is
remaining useful life applied to a ball bearings is presented. required, which is an essential procedure in order to
The general idea for developing this methodology is to focus characterize each measurement and highlight the degradation
on the intrinsic characteristics of the phenomenon studied (in patterns. Numerical features can be calculated by using
this case the monotonicity of the degradation profile), and different signal processing methods based on time-domain,
select an appropriate set of features to characterize it, frequency-domain and/or time-frequency-domain.
implement a model with significant generalization capabilities Taking into consideration practical industrial applications,
to estimate the degradation for different cases and calibrate the usually, the electromechanical system works under specific
output to obtain a desired parameter or characteristic, in this stationary conditions among its useful life, this condition
case, the remaining useful life. Figure 1 shows a scheme of the allows the use of time domain features as condition indicators
proposed methodology. for the characterization of the degradation, which is the
domain used in this work. Among the possible choices, it is
Training
possible to distinguish the maximum value, the Root Mean
Square (RMS), the variance and the kurtosis as general
Bearing Feature Feature Modeling: physical magnitude descriptors used in different applications.
Profile Calculation Selection OC-SVM Once the feature calculation stage is done, and a set of
features is obtained for each measurement, a feature selection
stage is introduced. Feature selection is an important step in
Test the fields of pattern recognition and data mining technology. It
New
Feature RUL identifies a meaningful feature subset, k-feature set, from the
Bearing OC-SVM
Profiles
Calculation Estimation original one q-feature set, where q>k, by removing redundant
or non-significant information. This dimensionality reduction
allows reducing the training complexity, while simplifying the
Fig. 1. Proposed methodology for RUL estimation. A bearing degradation classification space.
profile is used to identify the relevant features and training the OC-SVM, then
different bearing degradation profiles are used to test the methodology.
1754
An important part of this methodology is the criteria to application), to directly obtain the RUL on future test sets, and
make the feature selection. Based on the assumption that a to avoid ambiguity on the interpretation of results, that is,
feature that monotonically increases over time is the ideal same associated RUL to several classification scores due to
degradation signal, Spearmans rank correlation coefficient training error.
was used to assess how strong the monotonic relationship was At this point, a model trained with the selected features is
between the set of features calculated and the time duration of obtained, and the classification scores are linearly scaled and
the experiment. then fitted with an isotopic regression. This model now can be
Spearman coefficient [15] is a non-parametric measure of tested with new bearing degradation profiles, which is the
statistical dependence between two observational stochastic second phase of the methodology. On this test part, there is no
sequences. It assesses the relationship among the sequences in need to calculate all features, but only the selected to be
which the coefficient can be depicted using a monotonic monotonically relevant identified in the training part, and
function as: posteriorly test the new profiles with the trained and calibrated
OC-SVM and obtain the RUL estimation. The performance of
6 d 2 the OC-SVM is evaluated by means of the Root Mean
s = 1 (10)
n(n 2 1) Squared Error (RMSE).
where s denotes the Spearman Rank correlation 1 N
(y y i )
2
RMSE = i
(12)
coefficient, d is the difference between the sequences, and N i =1
n is the number of the sequences. Basically, a high Spearman Where N is the size of the dataset, yi corresponds to the
coefficient of the feature analyzed will imply a strong
monotonic relationship of the feature and the time duration of estimated RUL, and yi to the real one.
the experiment, which is the ideal case for estimating the
RUL. Under this criterion the two features with highest
coefficient ranking will be selected to characterize the IV. EXPERIMENTAL RESULTS
degradation profile. In order to evaluate the performance of the proposed
It must be noticed that the capability of the model to methodology, experimental data from a laboratory test-bench
estimate the RUL of the bearing will be directly related to the has been used. This data corresponds to a so called run-to-
monotonic relationship over time of the features selected, failure experimental approach, where the elements are forced
which will be assessed with the Spearman coefficient. to work beyond their nominal values and then, the degradation
Once the feature set is reduced, a model is trained. In this profiles can be extracted. Next, the proposed methodology is
case OC-SVM with calibrated output is selected. OC-SVM is applied, the results are discussed and, additionally, some
usually used for one class classification problems, but with variants of the methodology are proposed and analyzed.
calibration of the classification score (output), the classifier
could be employed to characterize an incremental degradation A. Description of the experimental platform
3B
profile on the feature space with generalization capacities, The experimental datasets come from the bearing failure
which means, it will detect increment in all axis of the feature diagnostic and prognostics platform [16], which provides
space. access to accelerated bearings degradation tests. The test
It is important to recall that common classification bench, as shown in figure 3, is composed by the speed
algorithms delimit the feature space for classification and can variation, torque transmission and the load profile generation
work properly only on the regions in which the training was stages.
involved and also their performance is good only with tests in
which the behavior of the phenomenon is very similar to the
training. In this sense, OC-SVM with calibration of classifier
scores provides a more generalized point of view, assuming
the monotonicity of the features, the incremental degradation
profile will not always follow the same pattern over time, but
still, regardless of the pattern, an incremental in any feature
selected is detectable.
The classification scores, which are the output of the
classifier, are lineally escalated [14] using (9):
f ( x) b
s ( x) = (11)
ab
where f(x) is the classification score, s(x) the linear re-
scaled score to [0, 1] and [b, a] are the minimum and
Fig. 3. Overview of PRONOSTIA experimental platform based on
maximum classification score obtained from the training set. accelerated bearing degradation.
This will help the interpretation of results and help the future
association of the classification score with the RUL, but still A cyclic radial load is applied on the external bearing
an isotonic regression is employed to fit the degradation under test in order to simulate its mechanical stress conditions.
profile in an isotonic function (which is ideal to the The experiment starts at a fixed speed condition, and stops
1755
when the measured vibration at the bearing under test is higher selected due to its monotonic behavior. As can be seen the
than 20g (1g=9.81 m/s2). It should be noticed that, in order to coefficient is high (being 1 the highest value of the coefficient
speed up the degradation, the applied radial load exceeds the possible), which implies a strong monotonic relationship
maximal load supported by the bearing. During the between the features and the time of the experiment. These
experiments any kind of failure (inner race, outer race, ball or two selected features will be used to train the model, validate
cage) could occur. This fact allows better representation of a it and test it with the different tests sets. Figure 4 shows the
real industrial scenario. training set represented by the two selected features. To
Regarding the test bench instrumentation, two high include the time variable on the plot, each acquisition was
frequency accelerometers (DYTRAN 3035B), are mounted on labeled with a gray scale. As it was expected, it can be seen a
the bearing external race in order to measure the horizontal strong monotonic behavior over the time of the experiment in
and the vertical accelerations. In addition, the monitoring both axis.
system includes one PT100 to measure the bearing
temperature, which is placed near the external ring of the
bearing under test.
The signals are acquired by means of a NI DAQ card; the
acceleration signals are acquired in successive windows with
duration of 1/10 seconds, repeated every 10 seconds, with a
sampling frequency of 25.6Hz. Similarly, the temperature
signal is acquired every minute with a sampling frequency of
10Hz. One dataset from PRONOSTIA experiments under the
same operating conditions have been selected. The
characteristics of the selected datasets for this experimental
validation can be seen in table I: Fig. 4. Training set represented by the two features selected, with time
variable included in a gray scale plot of the acquisitions just for visualization
purposes.
TABLE I. CHARACTERISTICS OF THE DATASET.
1756
with monotonic properties on the features), the classifier will
estimate the RUL. Notice that the classification score is scaled
from 0 to 1 (as can be seen in the gray-scale values). The next
step is to fit the classification scores with an isotonic
regression and associate them with a RUL percentage, the
result is shown in figure 6.
Fig. 8. Comparison of the estimated RUL versus the real RUL of set
Bearing1_4.
configurations.
The methodology presented was limited to certain
configuration, for example, the feature selection associated to
the Spearman's coefficient and the features selected are
limited to two. In order to enhance the validation of the
methodology several test were performed with some variants,
including using Principal Component Analysis (PCA), as a
feature reduction technique and testing the methodology
proposed but selecting three features instead of two. The
Fig. 7. Comparison of the estimated RUL versus the real RUL of the results of adding a third feature to the methodology presented
validation test. are shown in table III.
As expected, an estimation error is obtained around 80%
TABLE III. RMSE ERROR PERCENTAGES FOR THREE FEATURES
to 50% of the RUL, but in the other parts of the validation set
the model had no trouble to estimate the RUL. These Bearing Bearing Bearing Bearing Bearing
percentages are related to the stationary part of the test where 1_1 1_4 1_5 1_6 1_7
the features calculated doesn't change, so it's not possible to
RMSE 6.33% 9.19% 22.25% 11.41% 22.42%
accurately identify changes on this part of the degradation
profile. The RMSE value is 5.99%. With the same trained OC-
SVM and the fit function four different test sets were used. The reslults in table III show that adding a third feature
The result of the estimation of set Bearing 1_4 is shown in does not significally improve the estimation of the RUL. The
figure 8. reason is that the third feature added does not provide
resolution in the percentages of the RUL estimation where the
OC-SVM fails to estimate.
1757
The results of chaning the feature reduction approach to R). The authors would like to express their thanks to the
PCA are shown in table IV. department of Automatic Control and Micro-Mechatronic
Systems (AS2M) of FEMTO-ST1 institute, for their
TABLE IV. RMSE ERROR PERCENTAGES FOR PCA disinterested contribution providing access to the
PRONOSTIA experimental platform analyzed in this work.
Bearing Bearing Bearing Bearing Bearing
1_1 1_4 1_5 1_6 1_7 REFERENCES
RMSE 19.03% 19.92% 28.23% 24.13% 19.93% [1] F. C. K. Medjaher, and N. Zerhouni, "Feature extraction and
evaluation for health assessment and failure prognostics,"
The results shown in table IV proves that using PCA as a presented at the European Conference of the Prognostics and
reduction technique does not exploit the characteristics of the Health Management Society, Dresden, Germany, 2012.
[2] L. Yi-Sheng, C. Yung-Chi, T. Shiao-Li, and T. Tzung-Cheng, "A
features that improve the estimation of the RUL for this novel search scheme for nonintrusive load monitoring systems," in
methodology. Industrial Technology (ICIT), 2012 IEEE International Conference
on, 2012, pp. 102-107.
V. CONCLUSION [3] N. Gebraeel, M. Lawley, R. Liu, and V. Parmeshwaran, "Residual
life predictions from vibration-based degradation signals: a neural
In this paper a methodology to estimate the RUL applied network approach," Industrial Electronics, IEEE Transactions on,
to bearing degradation has been presented. The proposed vol. 51, pp. 694-700, 2004.
method is based on the detection of monotonic properties of [4] L. Jianhui, K. R. Pattipati, Q. Liu, and S. Chigusa, "Model-Based
features calculated of the bearing degradation profiles to make Prognostic Techniques Applied to a Suspension System," Systems,
a selection of best features to characterize the phenomenon. Man and Cybernetics, Part A: Systems and Humans, IEEE
Transactions on, vol. 38, pp. 1156-1168, 2008.
Then, a OC-SVM is trained to learn the degradation profile [5] K. Medjaher, D. A. Tobon-Mejia, and N. Zerhouni, "Remaining
and the output of this model is calibrated to directly estimate Useful Life Estimation of Critical Components With Application
the RUL. to Bearings," Reliability, IEEE Transactions on, vol. 61, pp. 292-
Experiments performed with a dataset of bearing 302, 2012.
degradation indicate that the proposed strategy can estimate [6] T. Benkedjouh, K. Medjaher, N. Zerhouni, and S. Rechak, "Fault
prognostic of bearings by using support vector data description," in
the RUL of the bearing with a reasonable error and with
Prognostics and Health Management (PHM), 2012 IEEE
enough generalization to be applied to different test sets with Conference on, 2012, pp. 1-7.
different degradation profiles. Also a comparison between [7] E. Sutrisno, H. Oh, A. S. S. Vasan, and M. Pecht, "Estimation of
another feature reduction technique and increasing the number remaining useful life of ball bearings using data driven
of features selected were introduced to give more versatility to methodologies," in Prognostics and Health Management (PHM),
the study. It must be noticed that the performance of the 2012 IEEE Conference on, 2012, pp. 1-7.
[8] E. Akin, I. Aydin, and M. Karakose, "FPGA based intelligent
proposed methodology will be directly related to the condition monitoring of induction motors: Detection, diagnosis,
monotonic behavior of the selected features. and prognosis," in Industrial Technology (ICIT), 2011 IEEE
Several RUL estimation methodologies have been International Conference on, 2011, pp. 373-378.
presented previously achieving good results, but still some [9] P. Xinxing and B. Lee, "A comparison of support vector machines
drawbacks are detected, some of the methodologies doesnt and artificial neural networks for mid-term load forecasting," in
Industrial Technology (ICIT), 2012 IEEE International Conference
make a proper selection of features to ease the estimation but on, 2012, pp. 95-101.
just a general reduction, others over fit the method to certain [10] Sch, B. lkopf, J. Platt, J. Shawe-Taylor, A. Smola, and R.
profile and fail to estimate the RUL of different sets with Williamson, "Estimating the Support of a High-Dimensional
different patterns or they depend of complex models that Distribution," Neural Computation, vol. 13, pp. 1443-1471, 2001.
makes them difficult to implement on practical applications. [11] D. M. J. Tax and R. P. W. Duin, "Support vector domain
description," Pattern Recognition Letters, vol. 20, pp. 1191-1199,
The methodology presented in this paper tries to close the gap
11// 1999.
on these drawbacks detected by introducing a proper feature [12] L. Clifton, "Multi-Channel Novelty Detection and Classifier
reduction for this case, the use of OC-SVM, which would be Combination," Ph.D., Electrical and Electronic Engineering, Univ.
easy to implement, to generalize the RUL estimation in the Manchester, Manchester, U.K., 2007.
feature space and an isotonic fit to improve the estimation. An [13] H.-T. Lin, C.-J. Lin, and R. Weng, "A note on Platts probabilistic
important drawback of this method would be the monotonic outputs for support vector machines," Machine Learning, vol. 68,
pp. 267-276, 2007/10/01 2007.
properties of the features, if they dont present that property [14] L. Clifton, D. A. Clifton, Z. Yang, P. Watkinson, L. Tarassenko,
the RUL estimation would tend to fail, but taking into account and Y. Hujun, "Probabilistic Novelty Detection With Support
the phenomenon of degradation this property is expected to Vector Machines," Reliability, IEEE Transactions on, vol. 63, pp.
occur, if its difficult to detect, then new features should be 455-467, 2014.
calculated to identify this property. [15] L. Di, C. Siu-Yeung, S. Dong-mei, and Q. Zheng-ding, "A
Spearman correlation coefficient ranking for matching-score fusion
This method can be applied on other components, other on speaker recognition," in TENCON 2010 - 2010 IEEE Region 10
than bearings, but, in case of other components degradation, Conference, 2010, pp. 736-741.
an assessment of the monotonic properties of the features will [16] R. G. Patrick Nectoux, Kamal Medjaher, et al., "PRONOSTIA :
determinate the performance of the methodology. An experimental platform for bearings accelerated degradation
tests," presented at the IEEE International Conference on
ACKNOWLEDGMENT Prognostics and Health Management, Colorado, United States,
2012.
This work was supported in part by the Spanish Ministry
of Economy and Competitiveness, CICYT (TRA2013-46757-
1758