Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Fault detection in reciprocating compressor valves

under varying load conditions
Kurt Pichler a,n, Edwin Lughofer b, Markus Pichler a, Thomas Buchegger a,
Erich Peter Klement b, Matthias Huschenbett c
a
Linz Center of Mechatronics GmbH, Altenberger Str. 69, 4040 Linz, Austria
b
Johannes Kepler University Linz, Altenberger Str. 69, 4040 Linz, Austria
c
Hoerbiger Service America, Inc., 5613 DTC Parkway – Suite 1050, Greenwood Village, CO 80111, USA

a r t i c l e i n f o abstract

Article history: This paper presents a novel approach for detecting cracked or broken reciprocating
Received 27 April 2015 compressor valves under varying load conditions. The main idea is that the time fre-
Received in revised form quency representation of vibration measurement data will show typical patterns
18 August 2015
depending on the fault state. The problem is to detect these patterns reliably. For the
Accepted 1 September 2015
detection task, we make a detour via the two dimensional autocorrelation. The auto-
correlation emphasizes the patterns and reduces noise effects. This makes it easier to
Keywords: define appropriate features. After feature extraction, classification is done using logistic
Fault detection regression and support vector machines. The method's performance is validated by ana-
Compressor valves
lyzing real world measurement data. The results will show a very high detection accuracy
Varying load conditions
while keeping the false alarm rates at a very low level for different compressor loads, thus
Time–frequency analysis
Feature extraction achieving a load-independent method. The proposed approach is, to our best knowledge,
Feature engineering the first automated method for reciprocating compressor valve fault detection that can
handle varying load conditions.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction

Reciprocating compressors are heavily used in modern industry, for instance in chemical industry, refinery, gas trans-
portation and gas storage. Economic demands of the last decades have also affected the operation of reciprocating com-
pressors. In many cases, compressors run at full capacity without backup. Reliable performance is thus a key issue and
becoming more important than ever. Costumers expect reduction or even elimination of unscheduled shutdowns as well as
extended maintenance intervals. These challenges are addressed by the development of advanced materials and designs.
However, fatigue and wear cannot be avoided.
There is also an economic trend towards saving on labor costs by reducing the frequency of on-site inspections. Such
considerations mean that modern gas storage facilities are run by remote control stations and the compressors are mon-
itored by automated technical systems. In this case, the system must be able to retrieve and evaluate relevant information
automatically to detect faulty behavior.

n
Corresponding author. Tel.: þ 43 732 2468 6237; fax: þ 43 732 2468 6005.
E-mail addresses: kurt.pichler@lcm.at (K. Pichler), edwin.lughofer@jku.at (E. Lughofer), markus.pichler@lcm.at (M. Pichler),
thomas.buchegger@lcm.at (T. Buchegger), ep.klement@jku.at (E.P. Klement), matthias.huschenbett@hoerbiger.com (M. Huschenbett).

http://dx.doi.org/10.1016/j.ymssp.2015.09.005
0888-3270/& 2015 Elsevier Ltd. All rights reserved.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
2 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

For these reasons, monitoring and diagnostics are almost an economical and technical necessity. Firstly, monitoring
systems enable condition-based maintenance. Secondly, improved understanding of compressor behavior allows evaluation
and recommendations regarding efficient compressor operation.
Condition monitoring can be based on measurements of various physical states, for instance, vibrations, flow rate, power,
position, temperature, and pressure. The data required for diagnostic evaluation depend mainly on the types of faults
expected and observed. Broken valves, with a percentage of about 36%, are the most common reason for unscheduled
shutdowns [1], followed by faulty pressure packings (about 18%) and piston rings (about 7%). The objective of this paper is
the early detection of cracked or broken (i.e. leaking) valves by analyzing vibration measurements, which reflect upcoming
problems in the form of abnormal, changing patterns. This reduces the number of expensive in-cylinder pressure sensors
required, and thereby the cost. Furthermore, detecting a crack as early as possible prevents debris of the broken valve to be
sucked into the compression chamber and cause severe secondary damage.
Several papers have been published about valve fault detection in reciprocating compressors. In [2], pattern recognition
techniques are applied to time–frequency representations and non-stationary auto-regressive models of vibration mea-
surements. Lin et al. [3] combined time–frequency analysis of vibration data and an artificial neural network, which enabled
them to differentiate between new and worn valves. Applying their approach to extended test scenarios with 15 seeded
faults [4] did not lead to satisfactory validation results. However, by reducing the number of fault cases to 7, they finally
achieved good classification results. Cyclostationary modeling of reciprocating compressors is introduced in [5,6]. Zouari
et al. [5] decompose the vibrations into periodic and random parts and use the random part to construct an angular fre-
quency map of the vibration energy. This allows to extract indicators for detecting fault signatures. Tiwari and Yadav [7]
analyzed pressure pulsation with a back-propagation neural network. The pressure pulsation (peak to peak) is modeled in
relation to the leak percentage. In [8], a method using support vector machines was presented. The first four zero-lag sums
of sub-band signals of the intrinsic mode functions are extracted from vibration measurements and used as features. Then,
typical patterns are classified using support vector machines. Drewes [9] described the effect of a valve fault on the p–V
diagram. Additionally, the effects of some other faults such as piston ring wear and damages to crank gears and pistons were
discussed. Several different fault detection methods were introduced in [10]. They are divided into four main categories:
time domain analysis, frequency domain analysis, orbit analysis, and trend analysis. Yang et al. [11] focus particularly on
small reciprocating compressors for refrigerators at constant operation conditions. They use wavelet transform to extract
features from raw noise and vibration data and classify them using neural networks and support vector machines. The
changes in cylinder pressures and instantaneous angular speed for various leakage percentages were analyzed visually in
[12]. Based on the results, a decision table for valve faults is built. Wang et al. [13] introduce an automated evaluation of the
p–V diagram. They determine 7 invariant moments of the p–V diagram and classify them using support vector machines. In
[14], the valve motion is monitored using acoustic emission signals and simulated valve motion. As the authors state in the
paper, the method can easily distinguish between normal valve, valve flutter and delayed closing, but it is not sensitive
to leaks.
Some of the approaches are basically similar to the one proposed in this paper in the way that they analyze time–
frequency representations of vibration measurements, for instance [2,4,5]. However, none of the methods above, which are
sensitive to valve leaks, are designed explicitly for varying compressor load. Since modern reciprocating compressors are
controlled by reverse flow load control systems, load changes are not unusual. For some applications, changing the load is
even a necessity. Hence, an automated monitoring system has to cope with this fact.
In this paper, we propose a data-driven method based on time–frequency analysis of vibration data. In fact, it is an
extension of the approach presented in [15]. Therefore, we compare a spectrogram to a reference spectrogram by computing
the point-wise difference. In case of faults, this difference shows certain distinct patterns, indeed not appearing in the case
of fault-free states, but depending on the load levels of the spectrograms and on the fault state of the valve. The main task is
to identify these patterns on-line. However, due to some uncertainties, for instance noise levels and varying positions of the
patterns within the spectrogram, identifying the patterns directly did not lead to satisfying results. Taking a detour via two-
dimensional autocorrelation centers the patterns and reduces noise effects. This makes it easier to determine appropriate
features for valve fault detection. Once the features are extracted, a classifier is built up based on some pre-collected training
samples including different ’fault’ and ’no fault’ states with the usage of two machine learning techniques (logistic
regression and support vector machines). This is an important step in order to extract optimal decision boundaries between
the classes ’fault’ and ’no fault’, which are applied for classifying new on-line data. The big advantage of this approach is that
it is independent of the load levels of the two spectrograms involved in the analysis. Whatever the load is, the features will
indicate the fault state as showing the same trend over different loads. Furthermore, there is no need to scale or normalize
according to other measured values, which helps to avoid the risk of error propagation. The proposed method is, according
to our best knowledge, the first method for reciprocating compressor valve fault detection that can deal with changing load
levels. Another strong characteristics of our method is that classifiers can be established fully automatically (without any
manual redesign phases) on new scenarios or machines, once new training samples have been gathered there, thus opening
a wide range of applicability. The main novelty of the paper is the characterization of a leaking valve in the space of
spectrogram differences and the subsequent feature extraction.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 3

2. Problem statement

In [15] we introduced an approach for fault detection in reciprocating compressor valves. The method performs time–
frequency analysis of a vibration signal x. The reason for changing to frequency domain is indicated in Fig. 1. It shows the raw
vibration signals for three cases, the first two without a fault and at different load conditions, the last one with a cracked
valve. Of course there can be seen slight differences between the three cases, but due to noise effects, different operating
conditions and different valve types it is probably hard to discriminate reliably between the states based on raw
vibration data.
For changing to time–frequency domain, the short-time Fourier transformation [16,17] defined as
Z
Ψ Δx ðt; f Þ ¼ xðt 0 Þ  Δðt 0 t Þ  e  j2π f t  dt ;
0
0
ð1Þ
t0

with Δ denoting the window function of the short-time Fourier transformation and the overbar denoting the conjugate
complex, is computed. Subsequently, the spectrogram representation S of Ψ is computed by
 2
Sðt; f Þ ¼ Ψ ðt; f Þ : ð2Þ

An exemplary spectrogram can be seen in Fig. 2. Let throughout this paper be RA Rnf nt with entries
r i;j ; i ¼ 1; …; nf ; j ¼ 1; …; nt , the discrete matrix representation of a reference spectrogram from a measurement from faultless
valves with nf ticks in frequency dimension and nt ticks in time dimension. Furthermore, let T A Rnf nt with entries
t i;j ; i ¼ 1; …; nf ; j ¼ 1; …; nt , be the discrete matrix representation of a test spectrogram from an unknown test measurement.

No Fault
2

-2
No Fault, different load level
Sensor Voltage

-2
Cracked valve
2

-2
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Time [s]

Fig. 1. Raw vibration data over two cycles for faultless valves at different load levels and for a cracked valve.

Fig. 2. Time–frequency representation of a vibration measurement over two cycles.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
4 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Then the distance from the test spectrogram to the reference spectrogram is determined by the metric

1 Xnf
X nt  
dðT; RÞ ¼  t  r i;j  ð3Þ
nf  nt i ¼ 1 j ¼ 1 i;j

as defined in [15].
The test spectrogram T is detected to be similar to the reference spectrogram (and therefore from a measurement from
faultless valves) if it is located within an ε-sphere around the reference spectrogram, i.e. if
dðT; RÞ r ε: ð4Þ
Otherwise, the test measurement is detected as faulty. The threshold ε is chosen automatically by using the logistic
regression rule [18,19] in one dimension. As shown in [15] this approach does only work for a steady state operating mode of
the compressor. Since not only valve faults but also different load levels increase the distance (3) significantly, one cannot
always determine whether a broken valve or changing load causes the distance to exceed the threshold. The approach was
extended to a switching model in [20]. The switching model enables fault detection for various load levels. Unfortunately
there are some serious drawbacks. Firstly, the switching model gives reasonable results for different steady state load levels,
but not during load changes. That means, that the compressor valves are unmonitored during load changes. Secondly, the
effort for training the model is disproportionately high, especially for multi-stage compressors, and it has to be trained anew
for every valve type. To overcome those limitations, the method has to be extended to be independent of the load level. This
will be achieved by not only looking at the distance (3), but defining more features and thereby extending the problem to a
pattern recognition problem. Being independent of the load level is a huge advantage. Any measurement from faultless
valves, for instance after restarting the compressor, can serve as a reference spectrogram, without considering the load.

3. Factors influencing the spectrogram

Of course, numerous variables can influence the shape of the spectrogram. During development of the method, we found
two to be significant. One of them is a cracked or broken valve (which is to be detected). The other one is the load level.
Changing load is caused by the reverse flow capacity control system [21].

3.1. Cracked or broken valve

As shown in [22], a cracked or broken valve is accompanied by changing amplitudes in certain frequency bands of the
power spectrum. Unfortunately it was not possible to determine the frequency bands of interest a priori, i.e. without
measurements of a faulty valve. The frequency bands of interest are different for different types of valves and for different
compressor geometry. Furthermore, the amplitudes even change for changing load. These facts make it impossible to use
the power spectral density of a certain frequency band as an indicator for faulty valves. The effect of changed amplitudes can
of course also be seen in the spectrogram. Let D A Rnf nt with entries di;j ; i ¼ 1; …; nf ; j ¼ 1; …; nt , be the point-wise difference
of the reference spectrogram R and the test spectrogram T, i.e.
D ¼ T R ð5Þ
with the matrix entries
di;j ¼ t i;j r i;j : ð6Þ
We observe the point-wise difference because it obviously reveals the difference between the reference and the test
measurement.
Assume that the test spectrogram T is from a measurement from broken valves and with similar load level as the
reference spectrogram R. According to [22], there will be values significantly different from zero in certain frequency bands
of D. Since the frequency is plotted on the ordinate, this will correspond to significant horizontal patterns in the spectrogram
difference matrix D. Fig. 3 shows a spectrogram difference matrix in the case of a test spectrogram from a cracked valve. Of
course this figure shows only one example, but we observed similar patterns many times for different valve types. Moreover,
the valve design experts made some FEM-simulations to simulate the impact of the valves sealing elements on the seat and
the guard and their behavior between those opening and closing events. For confidentiality reasons, this work is not
published. The FE-models are not accurate enough to be used for monitoring, but the results suggest that the observable
frequencies are likely to change if the valve is damaged. Hence, there is also a theoretical background to observe the
frequency domain for abnormal frequency bands.

3.2. Load level

The load level is controlled by the reverse flow capacity control system. This system directly influences the closing time
of the suction valve. It keeps the suction valve open for a specified time at the beginning of the compression phase. This
allows a fraction of the gas to flow back through the suction valve into the suction chamber. The discharge valve is a passive

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 5

valve. This means that it opens when the pressure inside the cylinder is higher than the pressure outside. When the reverse
flow capacity control allows a smaller amount of gas to be compressed, the pressure for opening the valve is obtained later.
For that reason, the control system influences the timing of the suction valve as well as the discharge valve.
In [23] it is shown that the valve events (opening and closing times of the valves) can be identified uniquely in the
spectrogram. The experimental setup in that paper was a very simple compressor and a simple reed valve. For more
complicated compressors and advanced valves, it is not possible to identify the valve events uniquely due to disturbance
from other parts. Nevertheless, the spectrogram difference reveals different load levels of reference and test spectrogram as
vertical patterns (Fig. 4). These patterns are obviously induced by the delayed impact time of the sealing element due to the
load control. In comparison to Figs. 3 and 4, Fig. 5 shows the spectrogram difference matrix for similar load levels and
faultless valves. Since there is no valve fault and no delayed impact time, there are no significant patterns.
It comes as no surprise, that for different load levels and a valve fault, the spectrogram difference matrix D shows both
horizontal as well as vertical patterns, thus detecting faults by using a concept on dðT; RÞ is not appropriate.

4. Pattern recognition approach

To discriminate between the patterns mentioned in Section 3, we compute appropriate features and apply pattern
recognition techniques. Although the horizontal or vertical patterns in the spectrogram difference matrix D appear to be
significant, they are hard to detect directly. There are two main reasons: on one hand, the position of the patterns within the
spectrogram difference matrix D is varying with the load, with the valve type and with the compressor geometry. It is very
hard to determine a priori on which part of D the algorithm should concentrate. On the other hand, vibration measurement

Fig. 3. Spectrogram difference matrix for a test spectrogram with broken valves.

Fig. 4. Spectrogram difference matrix for a test spectrogram with a different load level but faultless valves.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
6 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Fig. 5. Spectrogram difference matrix for a test spectrogram with a similar load level and faultless valves.

data tend to be noisy which makes an automated detection even harder. Both problems can be solved by taking a detour via
the two-dimensional autocorrelation of spectrogram difference matrix D: the significant patterns are centered and the noise
effects are reduced. The main idea and advantage of using autocorrelation is the realignment of the characteristic patterns.
The noise reducing property of correlation is of course well known and a positive side aspect. It might also be possible to
develop an algorithm that detects the significant horizontal and vertical patterns in the spectrogram difference D directly.
But such an algorithm had to scan the whole (and noisy) spectrogram difference matrix. The autocorrelation works as an
auxiliary tool to map the decentralized horizontal and vertical patterns to some kind of 3-dimensional mountain structures
that run through the center and are aligned in certain directions. Thus, the detection algorithm only has to distinguish
whether the mountain crest is aligned in time- or in frequency-direction to decide if it is induced by a broken valve or by the
load control [24]. Therefore it is sufficient to consider the central lines of the autocorrelation matrix in time- and in
frequency-direction. This is finally done by extracting some features from these central lines.

4.1. Autocorrelation

The two-dimensional autocorrelation matrix A A Rð2nf  1Þð2nt  1Þ with entries ai;j ; i ¼ 1; …; 2nf 1; j ¼ 1; …; 2nt  1, of the
spectrogram difference matrix D is defined as [25]
X
nf
X
nt
akf þ nf ;kt þ nt ¼ di;j  di  kf ;j  kt ð7Þ
i¼1j¼1
 
for kf ¼  nf  1 ; …; nf  1 and kt ¼  ðnt  1Þ; …; nt  1 and with
8
>
> i o1
>
>
< i 4nf
di;j ≔0 for
>
> j o1
>
>
: j 4nt :

The effect of computing the autocorrelation can be seen in Figs. 6–8. The patterns are in the center and the noise is
reduced. This makes it easier to detect the patterns automatically. In the case of a faultless test spectrogram with similar
load level to the reference spectrogram, the autocorrelation shows a central peak and the rest is approximately zero (Fig. 6).
A similar shape would appear if each instance of the spectrogram is produced independently by a normal distribution. For
the faultless test spectrogram with a different load level (Fig. 8), the autocorrelation shows some kind of mountain structure
with the crest in frequency-direction. In comparison, for a faulty test spectrogram (Fig. 7) we recognize that the auto-
correlation shows a mountain structure with the crest in time-direction.
We can even strengthen this effect by computing the autocorrelation twice (the results section will show the benefit
when using a linear classifier). This means that we compute the autocorrelation of the autocorrelation matrix A analogously
to (7) which results in the matrix B A Rð4nf  3Þð4nt  3Þ with entries bi;j ; i ¼ 1; …; 4nf  3; j ¼ 1; …; 4nt 3.

4.2. Feature extraction

We extract features from the spectrogram difference matrix as well as from the autocorrelation matrices. The features are
motivated by visually comparing the shapes of the matrix plots and quantifying the obvious differences. In the results

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 7

Fig. 6. Autocorrelation of the spectrogram difference matrix in Fig. 5.

Fig. 7. Autocorrelation of the spectrogram difference matrix in Fig. 3.

Fig. 8. Autocorrelation of the spectrogram difference matrix in Fig. 4.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
8 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

section, the plausibility of the features will be validated experimentally. It has to be mentioned that originally many more
features were extracted. The proposed ones were selected using standard feature selection techniques.

4.2.1. Spectrogram difference matrix

The first feature, extracted from the spectrogram difference matrix D, is the already introduced metric (3):

1 Xnf
X nt  
f1 ¼  di;j : ð8Þ
nf  nt i ¼ 1 j ¼ 1

As it is shown in [15], feature f1 alone is sufficient for constant load conditions. It simply measures if there are any
significant patterns or not. For varying load, the features computed from the autocorrelation are necessary to decide
whether the patterns are aligned in horizontal or vertical direction.

4.2.2. Autocorrelation matrix

Since the patterns are centered in the autocorrelation matrix A, it is sufficient to observe the curves through the center of
A. This reduces the amount of data to be processed and therefore saves computational time. These curves are symmetric;
hence we only have to observe one half. By defining Nf ¼ ⌈ð2nf 1Þ=2⌉ and Nt ¼ ⌈ð2nt  1Þ=2⌉ the segments g A RNt and
h A RNf are given by the matrix entries:
g ¼ aNf ;1…Nt and h ¼ a1…Nf ;Nt : ð9Þ

For visualization, the curves g and h are depicted in Fig. 7.

The first feature of the autocorrelation matrix measures the ratio of the mean values of g and h. It is defined as
1 PNt
 i ¼ 1 gi
Nt
f2 ¼ : ð10Þ
1 PN f
 i ¼ 1 hi
Nf
If there is some kind of mountain structure in one direction, the mean value in this direction will be significantly higher
than in the other direction. For instance, patterns in frequency-direction (induced by the load control) result in a bigger
mean value of h than g, delivering a feature value of f 2 o1. In the case of patterns in time-direction (induced by a broken
valve), the mean value of g will be bigger than the mean value of h, resulting in a value of f 2 4 1.
The next feature is somewhat similar to the previous one in the way that it measures the ratio of the degree of curvature
of the segments g and h. Between the end points g1 and g Nt respectively h1 and hNf the curves g and h are linearly inter-
polated, i.e.
g Nt  g 1
g^ i ¼ g 1 þ  ði  1Þ for i ¼ 1; …; Nt ð11Þ
N t 1
respectively
hN  h1
h^ i ¼ h1 þ f  ði  1Þ for i ¼ 1; …; Nf : ð12Þ
Nf  1
The rational beyond that definition is that if there is a mountain crest in frequency-direction, then h will be quite close to
^ If not, it will be further away. The same applies for g in the case of a mountain crest in time-direction. Finally, we define
h.
the feature as
PN f  ^
2
i ¼ 1 hi  h i
f3 ¼ P   : ð13Þ
Nt
^ 2
i ¼ 1 gi  g i

Obviously, if there is a mountain crest in frequency-direction (induced by the load control), h will be closer to h^ than g to
^g. In the other case, a mountain crest in time-direction (induced by a broken valve), g will be closer to g^ than h to h.^ Hence,
different load conditions will result in a much smaller value of f3 than a broken valve.
The last feature of the autocorrelation matrix measures the ratio of deviation of the segments g and h from their
regression lines. With g and h denoting the mean values of g and h, the coefficients for linear regression [26] are given by
 
PN t 1 þN t  
i ¼ 1 i  gi  g
1 þ Nt
2
βg ¼ PN t
  ; αg ¼ g  β g  ;
1 þN t 2
i ¼ 1 i
2
   
PNf 1 þ Nf
i  hi  h
i¼1 2 1 þ Nf
βh ¼   ; αh ¼ h  β h  : ð14Þ
PN f 1 þ Nf 2
i ¼ 1 i 2

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 9

Then, we define the feature as the ratio of the sum of squares of the residuals:
PNf  2
hi  αh  i  βh
f 4 ¼ i ¼ 1  : ð15Þ
PN t 2
i ¼ 1 g i  αg  i  β g

Feature f4 is especially helpful if the mountain crest is not a, more or less, straight line, but has a changing slope. If, for
instance, h is a straight mountain crest or no mountain crest at all, h will be quite close to its regression line. But if it is a
mountain crest with changing slope, h will be on average farther from its regression line. The same applies for g. Feature f4
was intentionally just a variation of f3 and we did not necessarily expect it to work as good as f3. But experiments have
shown that it works well, and outperforms feature f3 if the mountain crest is not a straight line.
As already mentioned, we initially defined many more features, including features measuring the slope of g and h
directly, but they were not selected by the feature selection algorithms.

4.2.3. Twofold autocorrelation matrix

The features of the twofold autocorrelation matrix are similar to the features above. So the features f5, f6 and f7 are
computed according to (10), (13) and (15), but using matrix B as input instead of matrix A and of course adapting Nf and Nt to
the size of the twofold autocorrelation matrix as Nf ¼ ⌈ð4nf  3Þ=2⌉ and Nt ¼ ⌈ð4nt  3=2Þ⌉.

5. Experimental setup

For development and testing, we acquired real world data from a reciprocating compressor test bench (a schematic
sketch of the test bench is shown in Fig. 9). Therefore, a double acting two-stage reciprocating compressor, more precisely
an Ariel JG2, was instrumented with vibration sensors at the valve covers. It was also instrumented with several other
sensors, such as in-cylinder pressure or discharge temperatures. Since vibration sensors are usually cheaper and easier to
mount, we concentrate on the evaluation of vibration data. Furthermore, there was a sensor to determine when the piston
reaches its top dead center. This sensor is needed to synchronize the spectrograms. A spectrogram starts when the piston
reaches its top dead center. The data were recorded with a sampling rate of 100 kHz. As common today, the reciprocating
compressor operates at a constant revolution speed – in our case 740 rpm. Together with the top dead center measurement,
the constant revolution speed guarantees the comparability of the spectrograms in time dimension. The load is controlled
by a reverse flow capacity control system, which allowed us to make test runs with constant as well as changing loads.
Faults were simulated by taking a faultless valve and manipulating it by cracking or braking the valve plate or ring. In
most cases, we made very small fissures to evaluate the method's ability to detect small faults at an early stage. The faulty
valve was deployed as the discharge valve at the head end of stage two. The faults were detectable by analyzing the
vibration measurements recorded at the valve cover of the cracked valve.
We recorded multiple test runs with different measurement configurations. The first measurement configurations,
denoted by v1 ; …; v6 , are measured from different types of valves. The valve types differ in their material and the valve
geometry. For confidentiality reasons, we are not allowed to specify the exact valve type of each configuration. As basic valve
designs, ported plate valves and concentric ring valves were used. As materials, synthetic Hoerbiger materials known as
MTX, HTCX, PC and PK (for details see [27]) were used. Furthermore, the configurations were measured at different times,
for instance v6 was measured two years later than v1. The next configuration, v7, is in fact not a new configuration. It is the
same measurement as v6, but it was measured with a different vibration sensor. For the measurement configurations
v1 ; …; v6 , a Wilcoxon Research accelerometer was used. For measurement configuration v7, a Metrix Vibrations accel-
erometer was used. The two sensors were mounted side by side. Fig. 2 shows a spectrogram over two cycles from

Fig. 9. A sketch of the reciprocating compressor test bench.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
10 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

measurement configuration v6. In contrast, Fig. 10 shows the spectrogram over the same two cycles from measurement
configuration v7. We can clearly see that the sensors have different sensitivities. Nevertheless, we will see that they produce
comparable patterns. The measurement configurations v8 and v9 are different valve types again, both measured with only
one vibration sensor (the Metrix Vibrations accelerometer).
For each configuration, measurement runs with the fault states 0 (i.e. faultless) and 1 (i.e. cracked or broken) were
recorded. Furthermore, for each valve type and each fault state we did a couple of measurement runs, each lasted about 3–
5 min. The measurements runs include dynamic load as well as constant load at different load levels.
In order to provide most arbitrary test conditions, the load has not been controlled in a fixed routine, but arbitrarily.
Fig. 11 shows the load level of compressor stage two for certain parts of measurement configurations v2, v4 and v7. The load
value represents the fraction of the gas that is actually being compressed. A load of 1 means that the whole gas in the
compression cylinder is compressed, while a load of 0.75 means that the load control allows 25% of the gas to flow back into
the suction chamber, and the rest of the gas is compressed and discharged. At normal compressor operation, the load is
usually between 0.65 and 1.
The amount of all measurement runs can be written as mi;j;k where i ¼ 1; …; 9 indicates the configuration, j A f0; 1g
indicates the fault state and k ¼ 1; …; K ði; jÞ indicates the index of a measurement run with valve type i and fault state j (from
altogether K ði; jÞ measurement runs). We compute spectrograms in a sliding window approach. Each sliding window starts
when the piston reaches its top dead center and lasts for a certain number of cycles. Together with the constant revolution
speed, this guarantees that the spectrograms are comparable in time. The second important factor for obtaining comparable
spectrograms is the resolution. The resolution has to be the same for every spectrogram. We chose a resolution of nt ¼ 27
and nf ¼ 28 which worked well for our purpose. Using this synchronized sliding window approach we compute the spec-
trograms for each measurement run and denote them by Spi;j;k . Here, p ¼ 1; …; P ði; j; kÞ denotes the index of the sliding
window and P ði; j; kÞ denotes the number of sliding windows for the measurement run mi;j;k . Subsequently, we compute the

Fig. 10. Time–frequency representation of the vibration measurement v6 over two cycles.

1
v2
0.95 v
4
v7
0.9

0.85
Load

0.8

0.75

0.7

0.65

0.6
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Cycle

Fig. 11. The load for three different measurement configurations.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 11

Faultless
30 Faulty

25

20

3
f
15

10

0
2 4 6 8 10 12 14
f1

Fig. 12. Scatterplot showing the features f1 and f3 of all samples of all measurement configurations.

4
v , Faultless
6
3.5 v6, Faulty
3 v7, Faultless
v7, Faulty
2.5
f2

1.5

0.5

0
3 4 5 6 7 8
f1

Fig. 13. Scatterplot showing the features f1 and f2 of all samples of the measurement configurations v6 and v7.

spectrogram difference matrix:

p p
D ¼ Si22;j  Si11;j ð16Þ
2 ;k2 1 ;k1

 
for all possible tuples i1 ; j1 ; k1 ; p1 ; i2 ; j2 ; k2 ; p2 with i1 ¼ i2 (i.e. same measurement configuration) and j1 ¼ 0 (i.e. reference
spectrogram with fault state 0). Then we estimate the features of the spectrogram difference. This procedure delivers a
feature vector for every possible combination of faultless reference spectrogram and test spectrogram for each measure-
ment run. Since we know the fault states of the measurement runs and their point of time, we can annotate the feature
vectors (’fault’ versus ’no fault’) and apply the training methods of the classification approaches.
For testing logistic regression, we set the threshold to its natural value of T ¼ 0:5. This will demonstrate that the method
delivers good results with an a priori chosen threshold. We will also report the results when varying its value within ½0; 1 in
order to obtain an impression of its sensitivity.
To obtain good parameters C and γ for support vector classification, we employed a grid search as proposed in [28]. With
each pair of parameters, we computed the classification accuracy for all measurement configurations. Then we chose the
pair of parameters with the highest mean accuracy over all measurement configurations. This results in the parameters
C ¼ 213 and γ ¼ 2  5 . For more detailed results see Section 6.4.

6. Results

In this section, we will present the validation results using the real world data described in Section 5. For classification
we use two well-known algorithms: the rather simple logistic regression model [18,19], that produces linear classification
boundaries, and the more advanced and very flexible support vector machines (SVMs) [29]

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
12 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

6.1. Scatterplot

At the beginning of the results section, we provide a visual impression of the feature space. Therefore, we show in Fig. 12
a scatterplot containing all samples of all measurement configurations. The plot shows the features f1 and f3. We can already
see quite a good separability. To demonstrate the case of different vibration sensors, the features f1 and f2 of the mea-
surement configurations v6 and v7 are shown in Fig. 13. Of course, the features are not equal. Nevertheless, the sensors (with
different sensitivity as shown in Figs. 2 and 10) deliver comparable features and the separability is preserved.

6.2. Cross-validation

First, we performed cross-validation using all data samples. That means that we put together the feature matrices of all
measurement configurations and performed 10-fold cross-validation. This results in a validation quality of 99.991% for
logistic regression and 99.998% for SVM. Both values are very high and indicate the good separability of the classes.
The second type of cross-validation gives more information about the comparability of different measurement config-
urations. We trained the classifier with eight of the nine measurement configurations and validated it with the ninth. Every
configuration served as validation configuration once. The results can be seen in Table 1. When using all seven features,
logistic regression (validation rates between 97.35% and 100%) and SVM (validation rates between 99.51% and 100%) deliver
very good results. The table contains not only the validation results using all features, but also for subsets using just the
features f 1 ; f 2 ; f 3 ; f 4 and f 1 ; f 5 ; f 6 ; f 7 respectively. This is to show that both the features of the autocorrelation matrix and the
features of the twofold autocorrelation matrix are necessary to get satisfying results. The necessity of the features f 2 ; f 3 ; f 4
can mainly bee seen for validation of v7. Logistic regression and SVM validation quality decrease significantly when skipping
those features. While SVM delivers very good results when using just the features f 1 ; f 2 ; f 3 ; f 4 (lowest rate 99.42%), logistic
regressions needs the additional features f 5 ; f 6 ; f 7 . This is shown by the very low validation rates of the configurations v2
(7.61%), v4 (1.16%) and v6 (41.05%).

6.3. ROC analysis

For ROC (receiver operating characteristic) analysis [30] of the logistic regression classifier, we compute the confusion
matrix for every combination of eight training configurations and one test configuration (compare Table 1) using all fea-
tures. Then we sum up the nine confusion matrices. This results in an overall confusion matrix (see the confusion matrix of
logistic regression in Table 2 for T¼0.5). To compute ROC curves, the threshold value T A ½0; 1 is varied over its whole range.
This delivers a confusion matrix for each value of T. The confusion matrix enables determining true positive rate and false
positive rate for certain parameter values. In the case of Table 2, the true positive rate is 28230þ 30
918
918 and the false positive rate
101
is 11 679 þ 101. These two values define one point of the ROC curve. The ROC curve for T varying between 0 and 1 can be seen in
Fig. 14. The area under the curve (AUC) is a good measure for the method's performance. The perfect AUC value would be 1,
while 0.5 represents a classifier based on random guessing. For the shown curve, we obtain a very high value of
AUC ¼0.9997.

6.4. Grid search

To compute the classification accuracy of SVM for a pair of parameters C and γ, we again compute the confusion matrix
for every combination of eight training configurations and one test configuration (compare Table 1) using all features. To
obtain an overall confusion matrix, we sum up the nine confusion matrices. The overall confusion matrix for C ¼ 213 and
γ ¼ 2  5 can be seen in Table 3. The confusion matrix allows us to compute the classification accuracy by dividing the
number of correctly classified samples by the number of all samples. In the case of Table 3 the accuracy is given by
11 746 þ 34 þ 25 þ 31 175. For performing a grid search, we compute the classification accuracy on a grid of pairs of C and γ. As
11 746 þ 31 175

suggested in [28], we chose the grid C ¼ 2  5 ; 2  4 ; …; 214 ; 215 and γ ¼ 2  15 ; 2  14 ; …; 22 ; 23 . The accuracy values on the grid
can be seen in Fig. 15. We can see that a diagonal in the log C–log γ -plane splits the accuracy values into a lower and a higher
part. As already mentioned before, the best accuracy value is obtained for C ¼ 213 and γ ¼ 2  5 . A large region of the surface
is very smooth and at a level of high accuracy, especially around the optimal combination. This is an indication that the
method is robust against the final parameter choice.

6.5. Computational time

Finally, we provide some information about the computational effort based on our experiments. All computations were
s
executed on a Quad Core CPU with 2.83 GHz and 4 GB of RAM. For data processing we used MATLAB . All of the following
computational times are median values of a large number of test runs.
First of all, we have a look at the feature extraction, whose computational times can be seen in Table 4. Obviously, the
time consumption for computing the spectrograms is insignificant compared to the feature extraction itself. This is because
computing the autocorrelation is the most time consuming part of the whole procedure. Having in mind that one cycle lasts
about 0.08 s, one could argue that more than 1 s for computing the features is too much for on-line monitoring. To counter

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying

Table 1
Validation results and training time consumption for different combinations of training and test configurations and different selected features.

Training configurations Test config- Test Features Log Time SVM (%) Time
urations samples (%) (ms) (ms)

v2 ; v3 ; v4 ; v5 ; v6 ; v7 ; v8 ; v9 v1 6570 f 1; f 2; f 3; f 4; f 5; f 6; f 7 99.82 85.57 99.97 130.22

f 1; f 2; f 3; f 4 100 62.28 99.95 172.43
f 1; f 5; f 6; f 7 99.65 100.87 99.51 861.81

v1 ; v3 ; v4 ; v5 ; v6 ; v7 ; v8 ; v9 v2 2380 f 1; f 2; f 3; f 4; f 5; f 6; f 7 100 96.14 100 203.34

f 1; f 2; f 3; f 4 7.61 69.15 100 306.36
f 1; f 5; f 6; f 7 99.79 114.04 98.36 1323.95

K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

v1 , v2 , v4 , v5 , v6 , v7 , v8 , v9 v3 4780 f 1; f 2; f 3; f 4; f 5; f 6; f 7 98.18 86.48 100 154.59
f 1; f 2; f 3; f 4 92.74 62.71 100 171.07
f 1; f 5; f 6; f 7 98.97 102.83 98.95 1213.94

v1 , v2 , v3 , v5 , v6 , v7 , v8 , v9 v4 3180 f 1; f 2; f 3; f 4; f 5; f 6; f 7 100 93.85 100 167.26

f 1; f 2; f 3; f 4 1.16 66.82 100 284.12
f 1; f 5; f 6; f 7 100 110.83 100 1085.81

v1 , v2 , v3 , v4 , v6 , v7 , v8 , v9 v5 3980 f 1; f 2; f 3; f 4; f 5; f 6; f 7 100 109.14 100 181.15

f 1; f 2; f 3; f 4 99.05 65.09 100 208.32
f 1; f 5; f 6; f 7 99.90 106.83 99.97 1040.33

v1 , v2 , v3 , v4 , v5 , v7 , v8 , v9 v6 6570 f 1; f 2; f 3; f 4; f 5; f 6; f 7 98.39 83.66 99.51 116.24

f 1; f 2; f 3; f 4 41.05 62.25 99.42 74.21
f 1; f 5; f 6; f 7 99.60 100.72 99.51 771.55

v1 , v2 , v3 , v4 , v5 , v6 , v8 , v9 v7 6570 f 1; f 2; f 3; f 4; f 5; f 6; f 7 97.35 119.96 99.62 154.91

f 1; f 2; f 3; f 4 94.52 84.73 99.97 106.31
f 1; f 5; f 6; f 7 86.32 99.49 84.14 120.00

v1 , v2 , v3 , v4 , v5 , v6 , v7 , v9 v8 7960 f 1; f 2; f 3; f 4; f 5; f 6; f 7 100 78.26 100 156.01

f 1; f 2; f 3; f 4 100 57.96 100 248.19
f 1; f 5; f 6; f 7 100 94.00 100 945.06

v1 , v2 , v3 , v4 , v5 , v6 , v7 , v8 v9 990 f 1; f 2; f 3; f 4; f 5; f 6; f 7 100 97.77 100 411.66

f 1; f 2; f 3; f 4 100 68.57 100 179.05
f 1; f 5; f 6; f 7 100 112.77 100 1134.49

13
14 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Table 2
Overall logistic regression confusion matrix for T ¼ 0.5.

Estimated class

0 1

True class 0 11 679 101

1 282 30 918

ROC curve for logistic regression

1

0.9

0.8

0.7
true positive rate

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1
false positive rate
Fig. 14. ROC curve for logistic regression.

Table 3
Overall SVM confusion matrix for C ¼ 213 and γ ¼ 2  5 .

Estimated class

1 1

True class 1 11 746 34

1 25 31 175

Fig. 15. Grid search results, the best pair of parameters is marked.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 15

Table 4
Computational time for feature extraction including and excluding spectrogram computation.

Features Incl. spec. (s) Excl. spec. (s)

f 1; f 2; f 3; f 4; f 5; f 6; f 7 2.303 2.279
f 1; f 2; f 3; f 4 1.166 1.138
f 1; f 5; f 6; f 7 2.304 2.276

this, we have a look at the typical fault development. Usually, a valve fault starts with a tiny crack. The crack will increase
during the next couple of minutes, but at this stage it will not cause damage to the compressor. There is at least 1 min time
to detect the crack and give the alarm until a severe fault develops. Hence, it is sufficient to monitor the fault state every
1.5 s (for instance).
Let us now examine the classifier training. The detailed computational times can be seen in Table 1. We notice that SVM
is strongly dependent on the size of the training data. More training instances (e.g. at test configuration v9) lead to a
significantly increased computation time. In contrast, the effect on computation time is less for logistic regression. We also
see that using just the features f 1 ; f 5 ; f 6 ; f 7 is more time consuming than using the features f 1 ; f 2 ; f 3 ; f 4 or even
f 1 ; f 2 ; f 3 ; f 4 ; f 5 ; f 6 ; f 7 . Moreover, training time can be seen as not critical, as training is usually conducted during off-line
processing stages.
Unsurprisingly, the consumption for evaluating the classifiers is insignificant. Classifying one test instance takes about
4.3  10  5 s with logistic regression and 1.1  10  4 s with SVM.

7. Conclusions

We have introduced a novel method to detect cracked or broken valves in reciprocating compressors under varying load
conditions. The method's ability of delivering good results was validated with real world measurement data. We have seen
that classification bounds from certain valve types can be used to classify other valve types. Furthermore, the method can be
applied using different vibration sensors: different vibration sensors (with different sensitivities) produce comparable
features. In our opinion, the most important advantage is the method's independence of the load: for any load levels of the
reference and test spectrograms, the features give an indication of the fault state without normalizing or scaling, and this
with an accuracy very close to 100%, while keeping the false alarm rate at a very low level.
Of course, a very important question for an industrial application is reproducibility. Some issues regarding reproducibility
are already covered with the experiments, others are yet to be investigated. The data with configurations v6 ; …; v9 were
recorded two years later than the data with v1 ; …; v5 , but on the very same compressor. In the meantime, the compressor
has been used for many other tests, and the instrumentation has been dismantled and reinstalled many times. During the
test runs, the valve support had to be opened and closed many times to exchange the different valve types. Nevertheless, the
proposed method delivers comparable features for every test run, which is represented in the high cross validation accuracy.
Hence, this kind of reproducibility is tested and verified. However, there are certain aspects of reproducibility that have not
been tested yet. In the test runs, we compressed air at indoor climate. Compressing other gases at different temperatures
might affect the vibration patterns as well. Furthermore, all tests have been performed on the very same compressor. For
general applicability in industry, the method has to work on different compressors without causing too much training effort.
It is questionable if a pre-trained classifier can be easily (or even at all) adapted to be used at a different compressor, since it
might probably produce different vibration patterns. Most likely, the learning phase of the classifier has to be repeated for a
new compressor. In this paper, the learning is done with data from faultless and faulty valves, which is in fact not possible
for every new compressor. It would be a first step to learn the classifier in a one-class approach, or to adopt some kind of
adaptive online learning. These are tasks for future development, but the general capability of the proposed approach is
shown in this paper.

Acknowledgment

This work has been supported by the Austrian COMET-K2 programme of the Linz Center of Mechatronics (LCM), and was
funded by the Austrian federal government and the federal state of Upper Austria.

References

[1] M. Huschenbett, G. Will, Thermodynamic simulation of reciprocating compressors to enable diagnostics based on measured temperatures and
pressures, In: Proceedings of the 4th Conference of the European Forum of Reciprocating Compressors, Antwerp, Belgium, 2005.
[2] O. Bardou, M. Sidahmed, Early detection of leakages in the exhaust and discharge systems of reciprocating machines by vibration analysis, Mech. Syst.
Signal Process. 8 (September (5)) (1994) 551–570.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i
16 K. Pichler et al. / Mechanical Systems and Signal Processing ∎ (∎∎∎∎) ∎∎∎–∎∎∎

[3] Y.H. Lin, H.C. Wu, C.Y. Wu, Automated condition classification of a reciprocating compressor using time–frequency analysis and an artificial neural
network, Smart Mater. Struct. 15 (2006) 1576–1584.
[4] Y.H. Lin, H.S. Liu, C.Y. Wu, Automated valve condition classification of a reciprocating compressor with seeded faults: experimentation and validation
of classification strategy, Smart Mater. Struct. 18 (2009).
[5] R. Zouari, J. Antoni, J.L. Ille, M. Sidahmed, M. Willaert, M. Watremetz, Cyclostationary modelling of reciprocating compressors and application to valve
fault detection, Int. J. Acoust. Vib. 12 (September (3)) (2007) 116–124.
[6] J. Antoni, Cyclostationarity by examples, Mech. Syst. Signal Process. 23 (4) (2009) 987–1036.
[7] A. Tiwari, P. Yadav, Application of ANN in condition monitoring of a defective reciprocating air compressor, J. Instrum. Soc. India 38 (1) (2008) 13–20.
[8] Q. Ren, X. Ma, G. Miao, Application of support vector machines in reciprocating compressor valve fault diagnosis, In: L. Wang, K. Chen, Y.S. Ong (Eds.),
Advances in Natural Computation, Springer-Varlag, Berlin, Heidelberg, 2005, pp. 81–84.
[9] E. Drewes, Condition monitoring for reciprocating compressors, Hydrocarb. Process. (September (77–79)) .
[10] Y. Zhang, J. Jiang, M. Flatley, B. Hill, Condition monitoring and fault detection of a compressor using signal processing techniques, In: Proceedings of
the American Control Conference, Arlington, VA, 2001, pp. 4460–4465.
[11] B.S. Yang, W.W. Hwang, D.J. Kim, A. Chit Tan, Condition classification of small reciprocating compressor for refrigerators using artificial neural net-
works and support vector machines, Mech. Syst. Signal Process. 19 (2005) 371–390.
[12] M. Elhaj, M. Almrabet, M. Rgeai, I. Ehtiwesh, A combined practical approach to condition monitoring of reciprocating compressors using IAS and
dynamic pressure, World Acad. Sci. Eng. Technol. 63 (2010) 186–192.
[13] F. Wang, L. Song, L. Zhang, H. Li, Fault diagnosis for reciprocating air compressor valve using p–V indicator diagram and SVM, In: Proceedings of the 3rd
International Symposium on Information Science and Engineering, 2010, pp. 255–258.
[14] Y. Wang, C. Xue, X. Jia, X. Peng, Fault diagnosis of reciprocating compressor valve with the method integrating acoustic emission signal and simulated
valve motion, Mech. Syst. Signal Process. 56–57 (2015) 197–212.
[15] K. Pichler, A. Schrems, T. Buchegger, M. Huschenbett, M. Pichler, Fault detection in reciprocating compressor valves for steady-state load conditions, In:
Proceedings of the 11th IEEE International Symposium on Signal Processing and Information Theory, Bilbao, Spain, 2011, pp. 224–229.
[16] F. Hlawatsch, G.F. Boudreaux-Bartels, Linear and quadratic time–frequency signal representations, IEEE Signal Process. Mag. 9 (1992) 21–67.
[17] L. Cohen, Time–Frequency Analysis, Prentice-Hall, Engelwood Cliffs, NJ, 1994.
[18] A. Anderson, Logistic discrimination, In: P.R. Krishnaiah, L.N. Kanal (Eds.), Handbook of Statistics, Volume 2: Classification, Pattern Recognition and
Reduction of Dimensionality, North Holland, Amsterdam; Gulf Publishing Company, Houston, TX, 1982, pp. 169–191.
[19] A. Webb, Statistical Pattern Recognition, John Wiley& Sons, New York, NJ, 2002.
[20] K. Pichler, T. Buchegger, Huschenbett, A switching model for fault detection in reciprocating compressor valves under varying load conditions, In:
Proceedings of the 13th IASTED International Conference on Signal and Image Processing, Dallas, TX, 2011, pp. 86–92.
[21] H.P. Bloch, J.J. Hoefner, Reciprocating Compressors: Operation and Maintenance, Gulf Publishing Company, Houston, TX, 1996.
[22] K. Pichler, A. Schrems, M. Huschenbett, Fault detection for a reciprocating compressor by combining vibration analysis and transformation techniques,
In: Proceedings of BINDT CM 2010 and MFPT 2010, Stratford-upon-Avon, England, 2010.
[23] Spectra Quest Inc., Vibration Signatures of Reciprocating Compressors, Spectra Quest Technical Note, Richmond, VA, 2007.
[24] K. Pichler, E. Lughofer, T. Buchegger, E.P. Klement, M. Pichler, M. Huschenbett, A visual method to detect broken reciprocating compressor valves under
varying load conditions, Mechatronics Forum, Linz, Austria, 2012.
[25] J.P. Lewis, Fast template matching, In: Vision Interface, vol. 95, Canadian Image Processing and Pattern Recognition Society, Quebec City, Canada, 1995,
pp. 120–123.
[26] N.R. Draper, H. Smith, Applied Regression Analysis, John Wiley& Sons, New York, NJ, 1998.
[27] Hoerbiger Compression Technology, Compressor Valves—Engineered for Better Efficiency and Reliability, Product Brochure, 2015 〈http://www.hoer
biger.com/upload/file/valves_overview_service_en.pdf〉.
[28] C.W. Hsu, C.C. Chang, C.J. Lin, A Practical Guide to Support Vector Classification, Technical Report, Department of Computer Science, National Taiwan
University, 2003 〈http://www.csie.ntu.edu.tw/cjlin/papers/guide/guide.pdf〉.
[29] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer Verlag, New York, NJ, 1995.
[30] K.A. Spackman, Signal detection theory: valuable tools for evaluating inductive learning, In: Proceedings of the 6th International Workshop on
Machine Learning, San Mateo, CA, 1998, pp. 160–163.

Please cite this article as: K. Pichler, et al., Fault detection in reciprocating compressor valves under varying
load conditions, Mech. Syst. Signal Process. (2015), http://dx.doi.org/10.1016/j.ymssp.2015.09.005i