Int J Adv Manuf Technol (2014) 75:693–703

DOI 10.1007/s00170-014-6138-3

ORIGINAL ARTICLE

Managing measured vibration data for malfunction

detection of an assembled mechanical coupling
Gabor Manhertz · Gabor Gardonyi · Gabor Por

Received: 22 December 2013 / Accepted: 30 June 2014 / Published online: 1 August 2014
© Springer-Verlag London 2014

Abstract Vibration data recorded by accelerometer fixed 1 Introduction

on a mechanical coupling were treated using different
data processing techniques as estimating general root mean Electromotive drive trains are applied in many fields of
square, partial root means square, mean, variance, partial contemporary industry, for example in automotive indus-
variance, kurtosis, full Fourier spectra, and partial spectra, try. They may have different types and sizes and they have
as well as using sequential probability ratio test (SPRT). The typically some coupling mechanism, e.g., cogwheels and/or
numerical results of these methods measured on an electro- worm-gears. It is very important to select those products,
motive drive train-coupling were collected into a common which cannot or may not serve the expected service life. We
data set, and a search method of deviation from the nor- need a fast, nonintrusive method to classify the mechanism
mal behavior was elaborated. This selects the bad couplings before it would be built in.
if at least three of the presented qualification parameters In order to improve the manufacturing process ensur-
crosses either theoretically or experimentally settled limits. ing the quality of the products and reducing the costs, it
It was shown that the confidence level of such classifica- is necessary to identify the defective automotive parts. For
tion of good and bad species was better than 95 % compared manufacturing processes, there are several methods which
with the results of thoroughly tested ones by conventional can increase the quality of the products and the reliabil-
methods. ity of the process. A kind of the quality assurance is the
total quality maintenance method discussed in [1]. It claims
Keywords Drive train · Vibration diagnostics · that the implementation of vibration-based maintenance
Manufacturing quality · SPRT · RMS strategy provides possibilities for acquiring early indica-
tions of changes of machinery state and the vibration-based
condition monitoring programs became popular in many
G. Manhertz () · G. Gardonyi industries.
Department of Mechatronics, Optics and Mechanical Engineering There are several methods which can disclose defects
Informatics, Budapest University of Technology and Economics,
H-1111, Műegyetem rkp. 1-3., Budapest, Hungary
during or after the manufacturing and assembling pro-
e-mail: [email protected] cess, such as online monitoring systems, off-line production
G. Gardonyi
parameter optimization, or quality assurance procedures [2–
e-mail: [email protected] 4]. A specialized measurement system can be found in [5]
where the combination of the measurements are based on
G. Por measurement of vibration and acoustic emission. Some of
Hungarian Acoustics and Industrial Diagnostics Laboratory,
College of Dunaújváros, H-2400, Táncsics str. 1/A, Dunaújváros,
the condition monitoring systems uses the measurement of
Hungary vibration and/or emitted noise to classify good and bad
e-mail: [email protected] products with high efficiency [6].
694 Int J Adv Manuf Technol (2014) 75:693–703

Vibration analysis is applied frequently as the basis of have the smallest deviation from the normal), different
expert systems. Some of these systems are used for con- noises are produced that opens the possibility to recog-
dition monitoring and fault diagnosis of electrical motors nize deviation from normality. The problems mentioned
[7]. Frequency-based failure analysis is often used at the above that are causing abnormal vibration and sound effects
investigation of both motors [8] and the attached cou- can be detected using our extended analytical software
plings [9]. A higher level of computer intelligence could package.
be implemented in these systems with the use of the neu- Before starting the main part of the research, preliminary
ral networks [10, 11]. By combining the statistical quality measurements have been carried out with the aim to find
control, the vibration analysis [12] can result in a highly- the best place for vibration measurement, since we were
automated and controlled manufacturing process in the limited in the number of accelerometers to reduce the mea-
future. Vibration-based measurement systems became very surement time. We tested different places on the product
common in industrial environment nowadays. In [13], the (see Fig. 1).
generic methodology to develop intelligent monitoring sys- For the measurements finally, a B&K DeltaTron
tem for machining processes is described, presenting also Type 4397 piezoelectric accelerometer (with sensitivity of
the benefits of the usage of the vibration measurement in 9,688 mV/g) was attached to the drive train with beeswax.
such a system. Yu et al. [14] represents a vibration measure- The sensor was connected to a NI-9234 data acquisition card
ment experiment based on Gaussian mixture model to detect manufactured by National Instruments which was docked
bearing faults after assembly. In [15], a mathematical model into an eight-slot NI-9172 cDAQ USB chassis also from
based on the response surface-based D-optimal design is National Instruments. The chassis communicates with a
proposed for modeling and analyzing tool vibration and personal computer via USB connection. The signal of the
surface roughness in the precision end-milling process. A sensor, measured during the forward and reverse actua-
real-time implementation of common used vibration fea- tion, was stored on the computer. This was followed by the
tures (peak amplitude, root mean square, etc.) can be found analysis of the measured vibration data with the software
in [16]. developed for classification.
Our task was to disqualify electromotive drive train The results of the recorded time signals from different
couplings (Fig. 1), which exhibit unusual noises, i.e., accelerometers showed that the accelerometer fixed on the
which are suspicious for malfunctions or construction coupling (right side on Fig. 1) had the best signal-to-noise
failures. In these mechanical couplings, the cog wheels, ratio and the highest amplitudes compared to the others.
worm gears change several times the direction of their Accordingly, further measurements and classifications were
operation, while big load variations may occur. This carried out in this position.
means that all elements of the drive train is subjected During our research work, we developed two software:
to different load. A possible way to investigate stochas- one for calibration and test purposes of the implemented
tic audible noises to measure vibration during normal diagnostic methods, and one for measuring and qualifying
operation of the device. We are presenting such method the products. The measuring system was based mainly on
in this paper. Drive train noises can be recorded and stochastic parts of the measured vibration data. The main
processed during normal, regular operation. If the cou- objective of the development was that our new classification
pling has some deviation from the usual normal oper- system should determine the defective drives before instal-
ation (e.g., the surfaces of the cogwheels do not have lation with 95 % of confidence level compared to results of
the same smoothness, or the distances between them the existing complicated, thorough and detailed industrial
system.

2 Selected methods

First, mathematical vibration diagnostic methods was

quested for the qualification process. Due to the diversity of
the possible malfunctions, the correct classification can be
guaranteed only with the combination of several different
diagnostic methods. It should be noted that several methods
using basic statistical definitions (e.g., mean, variance, and
partial variance) are commonly well-known estimators from
Fig. 1 The drive and the mechanical coupling with mounted B&K text books. Therefore, these mathematical formulas will not
DeltaTron Type 4397 piezoelectric accelerometers be presented in this paper.
Int J Adv Manuf Technol (2014) 75:693–703 695

2.1 Time-domain methods parameters. More, in-depth information about these steps
can be found in references [20] and [21].
Three time-domain analysis were involved in the common One example is demonstrated on Fig. 2, where the posi-
qualifying methods: tive decision making level was set very high, while the nor-
mal state (negative side) was only at the 95 % of confidence
• Observing outlier acceleration values
level.
• Using sequential probability ratio test (SPRT) which
The method is sensitive for the stochastic part of the sig-
was relying on the signal filtered by autoregressive
nal. Therefore, it is advisable to remove all deterministic
model
parts of vibration signal and it is also advisable to set the
• The calculation of the Kurtosis value
mean value to zero. This was achieved in our method using
In order to register the peaks of the time signal, two quali- an adaptive autoregressive (AR) [22] filter before applying
fying methods were created. They were based on a common SPRT.
ground using thresholds, but calculated with slightly dif-
ferent methods. Both methods returned the number of the

M
peaks which were higher than a calculated threshold level. yk = ai yk−i + wk (2)
The difference between the two methods was in the esti- i=1
mation of the threshold level. The first method uses the
where yk are the sampled signal, ai are the autoregressive
standard deviation of the signal multiplied by a parame-
coefficients, wk is the stochastic noise contribution to the
ter, which was selected in experimental way (see details
signal, and M is the model order. Based on this formula,
about the parameter determination in a further section). The
autoregressive coefficients can be estimated for the given
other one calculates the average distance of data from zero
sampled signals and an SPRT pre-filter can be built:
producing another threshold level. This was calculated by
multiplying the average distance with another experimen-
tally determined parameter given as an input parameter as 
M

well. The number of the detected peaks was used for the x̃j = xj − ai xk−i (3)
later qualification. i=1
The sequential probability ratio test (SPRT) invented by where xj are the actual samples of the signal.
Wald [17] and applied in [18, 19] is a method which com- The Kurtosis (fourth moment of a stochastic process) can
pares the probability density functions using the ratio of two be estimated as [23]:
conditioned probabilities. In the denominator, there is the
probability which samples belong to a process with prob- 1 ı=1
n (xi − x̃)
4
ability density of hypothesis H1 , while in the nominator, α4 = n
 (4)
we have the same probability, which samples belong to ( n1 ı=1
n (xi − x̃) )
2 2

probability density of H0 hypothesis.

It shows if the signal distribution was more or less tighter
than the normal distribution and possesses the same mean
p(x1 , x2 , ..., xn|H1 ) value and standard deviation. During the analysis, the fourth
λn = ln( ) (1)
p(x1 , x2 , ..., xn|H0 ) moment values of the signals sections were calculated with
an overlapped windowing process. In this case, the qualifi-
If the set of n samples belongs to stochastic process
cation was based on the proportion of the Kurtosis values
characterized with H1 parameters, then the denominator is
that exceeded the defined threshold level. However, typi-
larger, if the opposite they belongs to H0 , then the nomina-
cally, a level of ten is a sensitive limit to find if there is a
tor is larger. Consequently, the lambda function, which is the
sudden peak in the time signal.
logarithm of that ratio, either grows (H1 case) or decreases
(H0 case) at each new sample. We start from zero lambda
2.2 Frequency-domain methods
value and taking the samples one by one. The lambda func-
tion is decreasing if the probability density function belongs
In frequency-domain, the signal spectral composition was
to H0 until it reaches a certain decision making level. This
investigated by using different methods based on the
is why the method is called sequential. If the level is set to
autopower spectral density (APSD) function.
−4.6, then the confidence level of the decision is 95 %. The
same is true for H1 : if the growing value of lambda func- • Operating frequency search
tion exceeds the +4.6 value, then we can state with 95 % of • APSD-masking
confidence level that the stochastic noise belongs to a pro- • Calculations based on spectral RMS (rms(ω),
cess with probability density function characterized by H1 RMSrat io , windowed rms(ω).
696 Int J Adv Manuf Technol (2014) 75:693–703

Fig. 2 The result of the SPRT

method with the following
settings: upper decision limit:
30, lower decision limit: −6.9,
autoregressive model order: 50
(black lambda function, red
upper decision limit)

The first method in frequency-domain finds the main that typically these curves also exhibit a sharp increase at the
peak in the APSD, which corresponds to the operating (rota- beginning of the high-frequency components. On the func-
tion) frequency. Typically, this can be found in a given tion in Fig. 3, it can be seen that there are jumps at 900 and
frequency interval. Deviation from the nominal frequency 1,350 Hz, which characterize the power distribution.
signalizes malfunction, since in that case the rotation speed The next RMS method (RMSrat io) is similar to the first
changes typically due to friction or cogwheel problems. one. The difference between them is that the whole fre-
The second method was the APSD-masking, in which quency interval is divided by a frequency value ωmiddle to
we fitted a predefined mask on the calculated spectra find- a lower and an upper section. After this, a ratio was formed
ing those points of the spectra, which were popping outside by the RMS value of dividing the RMS value of the upper
of the mask. Mask was defined by calculating the vibra- range by the RMS value of the lower range.
tion spectra of hundred of nondefective couplings. Thus, we
gained an averaged spectrum with estimated uncertainty .  ωmax
ω AP SD(ω)dω
Further frequency-domain analysis methods were based RMSrat io (ωmiddle ) =  ωmiddle
middle
(9)
AP SD(ω)dω
on the RMS calculation. It can be proved that, theoretically, ω0
the RMS can be estimated as the integral of the APSD [24]. The fourth type of the frequency-domain methods was
based on a time-windowing technique by dividing the total
 ∞  ∞
time signal into time-segments with overlapping, and esti-
AP SD(ω)dω = F (ω)F ∗ (ω)dω = (5)
0 0 mating the RMS values in segments. From this, a time-
 dependent RMS value has been constructed. Both the
∞ ∞  ∞
f (t).e−iωt dt f (t)e+iωt dtdω = (6) section length and the time steps between the beginnings
0 −∞ −∞ of the sections could be modified via input parameters (cf.
 ∞ Fig. 4). The maximum and the mean value of the func-
= f (t)2 dt = RMS 2 (7) tion were compared with gained qualification limits. This
−∞ kind of function representation resembles us to the so-called
Since for us only the components in audible range are short-time RMS methods [25].
important, the upper frequency limit of the integration can Finally, because of the design and the manufacturing pro-
be given as an input parameter. In this case, we can have a cess of the drive train-couplings, differences of acoustic
frequency-dependent RMS: and vibration phenomena were observed at the two differ-
 ent operating directions. All the abovementioned diagnostic
ω techniques were estimated for both directions. Comparison
rms(ω) = AP SD(ω)dω (8) (ratio and/or difference) of these results obtained during the
0
operation in different directions opened the possibility to
Figure 3 demonstrates such function in a practical case. detect deviations. During a comparison of the predefined
Naturally, the curve has a monotone-increasing nature qualification threshold limits and the results of difference or
because the spectral values are positive definite. The end- ratio calculation, a logical qualification conclusion could be
value of the curve equals to RMS. However, it is remarkable drawn.

Fig. 3 The result of the rms(ω)

calculation method with the
following settings: lower
frequency limit 0 Hz, upper
frequency limit 3 kHz, alarm
level 3.5 g (black the rms(ω)
function, red the prohibited
range of the maximum value)
Int J Adv Manuf Technol (2014) 75:693–703 697

Fig. 4 The result of the windowed RMS method with the following RMS function, blue the average value of the function, red the prohib-
settings: lower frequency limit 0 Hz, upper frequency limit 3 kHz, win- ited range of the maximum value, orange the prohibited value of the
dows width 0.01 s, degree of the overlap 0.005 s (black the windowed average)

3 Data pool and search for deviation cells in the first row of a column could contain the following
strings presented in Table 1.
In order to achieve correct qualification, we had to On Fig. 5, the following methods can be seen with the
build a properly tuned parameter set to the given prob- following input parameters:
lem. Collecting the data resulted in a rather big size
Excel table. Fragments of that table are shown on • Time signal peak search (based on the average dis-
Figs. 5 and 6. tance of data from zero) with input parameters (Peaks
The first column is an ID column which contained the m “percent”)
ID names of the drive trains and the operational direction ◦ “percent”: 400
information as well. The methods and the associated prede- ◦ “percent”: 450
fined parameters were represented in the first row. In details, ◦ “percent”: 500

Fig. 5 A fragment of the calibration set Excel table

698 Int J Adv Manuf Technol (2014) 75:693–703

Fig. 6 A fragment of the calibration set Excel table

Table 1 The first row of the Excel table generated for input parameter setup contained the following strings

String Description

Peaks m “percent” The number of the detected time signal peaks above the threshold. In this case, the threshold was
determined by multiplying the mean (m) of the signal with a percentage (“percent”).
Peaks d “percent” The number of the detected time signal peaks above the threshold. In this case, the threshold was
determined by multiplying the standard deviation (d) of the signal with a percentage (“percent”).
RMS Hzlower -Hzupper The RMS value between the frequency range given as input parameter
(lower frequency limit-Hzlower , upper frequency limit-Hzupper )
RMS-r Hzlower -Hzmiddle -Hzupper The RMS-ratio value between the frequency range given as input parameter
(lower frequency limit - Hzlower , the frequency value, where the RMS function was
divided-Hzmiddle , upper frequency limit-Hzupper )
AS peak freq (Hzlower -Hzupper ) The frequency value of the largest spectral component in the
specified interval (lower frequency limit - Hzlower , upper frequency limit - Hzupper )
AS peak value (Hzlower -Hzupper ) The amplitude value of the largest spectral component in the specified
interval (lower frequency limit - Hzlower , upper frequency limit - Hzupper )
SPRT max The maximum value of the calculated SPRT lambda function
Int J Adv Manuf Technol (2014) 75:693–703 699

• RMS calculation with input parameters (RMS Hzlower - parameters was a comparison of the aforementioned Excel
Hzupper ) table with sampled reality. We selected about 100 pieces
from the production line, which has been qualified by
◦ Hzlower : 0 kHz - Hzupper : 3 kHz
manufacturer as good and bad samples by detailed inves-
◦ Hzlower : 0 kHz - Hzupper : 7 kHz
tigations, which included other selection programs, as well
◦ Hzlower : 0 kHz - Hzupper : 12 kHz
as human audible tests and thorough parameter measure-
In details, on Fig. 6, the following methods can be seen ments. This set was regarded as a calibration set. We built
with the following input parameters: our Excel table of the calibration set on the basic of afore-
• RMS-ratio calculation with input parameters (RMS-r mentioned data evaluation methods. For each column, we
Hzlower Hzmiddle -Hzupper ) selected input threshold values manually in such a way,
that the cells for good samples were green (value was under
◦ Hzlower : 0 kHz - Hzmiddle : 3 kHz - Hzupper : the threshold level) and red for bad samples (value is over
7 kHz the threshold level) as it is shown on Fig. 5. In most of the
◦ Hzlower : 0 kHz - Hzmiddle : 3 kHz - Hzupper : cases, it was an easy task (for example for RMS and for
12 kHz Kurtosis), generally speaking for all parameters, which had
• Operating frequency search with input parameters (AS noninteger value. In these parameters, we usually could
peak value (Hzlower - Hzupper ) estimate such input value that only occurred 1–2 deviation
from good and bad, in any case: 95 % of the judgement
◦ Hzlower : 500 Hz - Hzupper : 600 Hz
were correct. However, in the case of cells filled in by
◦ Hzlower : 1 kHz - Hzupper : 12 kHz
integers, which were mainly the number of occurrences for
In the previous chapter, it was stated that input param- spikes, or events, we had to set the limit about zero.
eters are necessary for the calculation of the diagnostic Example given from Fig. 5, it can be seen that the time
methods. The simplest way for the determination of these signal peak search method based on the average distance

Fig. 7 The determination of the

qualifying parameters
700 Int J Adv Manuf Technol (2014) 75:693–703

of data from zero (peak m) was executed with three differ- classification, the measuring and qualifying system uses the
ent input percentage parameters (see also above). The worst following numerical results:
input parameter for this method was 400 % = 4.0 because
• Number of the peaks detected in the time signal (with
in the section of the good motors, one motor had been
the two methods—mean and variance calculation)
announced bad (one red cell). To choose between the other
• Operating frequency
two options, the defective motor section had to be investi-
• RMS value
gated. The question was which input parameter would result
• RMS-ratio value
the most red cell in the section of the defective motors.
• Mean and maximum value of the RMS(t) function
The answer was: the second setup with input parameter
• Number of deviances from the APSD-mask
450 % = 4.5. This means that the best input parameter for
• Maximum value of the Kurtosis
the time signal peak search method based on the average
• Comparing the ratios and the differences between the
distance of data from zero was 450 % = 4.5.
results of the two operation directions
3.1 A handy program for estimating qualifying threshold
level

After predefining all input parameters for the diagnostic

methods, the basis of the qualification process needed to
be determined by the qualification threshold values. For

Fig. 9 The process of determination of the qualification threshold

Fig. 8 The flow chart of setting the input parameter value
Int J Adv Manuf Technol (2014) 75:693–703 701

Fig. 10 The process of the qualification (the continuous lines are decision lines, the dotted lines are measured signals)
702 Int J Adv Manuf Technol (2014) 75:693–703

To enhance the estimation of the threshold levels of the or the operating frequency differs from the nominal fre-
qualification, a special utility software had been developed quency. If the number of the alarms was between 1 and
and implemented (Fig. 7). 3, then the software gave the operator the opportunity to
After loading the exported data set into the utility pro- decide. At this point, the following actions should be carried
gram, and selecting a diagnostic method in the list box out:
on the left side, the results of measurements of all drive
• The operator could listen to the measured signal and
trains were displayed graphically. The list box contains all
make a manual decision
of the diagnostic methods operating with the best input
• The operator could restart the qualification process with
parameter set defined in the previous step. To help the
the data acquisition on the given coupling
threshold limit setting, numerical indicators were placed.
These indicators display how many percent of the values The number of the alarms, based on which the system
are under the set threshold level. With the adjustment of the makes the decision, could be adjusted as well. The values
qualification threshold (blue line on the Fig. 7), it could mentioned before were the default. Next step was to test the
be seen how many elements of the data set were above or software in parallel with the system on the production line
below this, which could give a better chance to set the under industrial conditions. The program had fully repro-
threshold for each diagnostic method as tight as possible duced the expectations and proved to be even more reliable
to meet the 95 % confidence level, i.e., the qualification than the industrial system. After the in-line analysis, the
threshold on Fig. 7 for the operating frequency search analysis was carried out on devices with artificial defects
method was set at that level (70), where less than 5 % (rasped cog-wheel teeth, bent shaft, or scrapped sliding sur-
of the measurement results executed on calibration set face of worm). In order to obtain comparable results, these
were below the limit. This threshold limit then hold dur- devices were analyzed with the industrial software and with
ing real analyzing work, when manufactured coupling were the newly developed qualification software as well. These
tested. defects were recognized by the industrial system with nearly
Two flow chart diagrams can be seen on Figs. 8 and 9 50 % accuracy, while our software, with the newly devel-
which represent the processes of the input parameter set’s oped qualification algorithms, always detected the failure.
and the determination of the qualification threshold limit’s With this, the system topped the previously appointed 95 %
determination. accuracy objective.

4 Demonstration of the method on actual task 5 Conclusion

The final qualification process (Fig. 10) of the product It had been well known beforehand that analysis of vibration
was based on the previously described qualification thresh- data can lead to diagnostics of malfunction of mechanical
old set. The whole qualification process, starting from the coupling, since vibration and/or audible noises from bad
measurement and ending with the qualification result, took coupling are different from those of the good couplings.
only approximately 10 s per piece. The sampling period The task trusted by manufacturer was to build a system for
was the most time-consuming part because the coupling selecting bad and good couplings with better than 95 % of
had to be actuated for a few seconds to reach the steady- confidence level automatically, which is simple, fast enough
state before real sampling took place in both directions. to be used on all product manufactured on the production
After the measurement, the qualification result appeared in line.
a second. Nearly 200 pieces of electromotive drive train- We found that only a combination of several diagnos-
couplings were tested under laboratory conditions and later tics method was capable to ensure the desired level of
in the factory as well. The products were qualified by confidence. Besides the conventional parameters like mean,
the industrial software for control purposes. Compared to variance, root mean square, and main spectral compo-
those qualifications, the newly developed software with the nents, we applied also less frequently use one as kurtosis,
diagnostic algorithms produced approximately eight to ten and brand new ones like sequential probability ratio test
different results. But the later subjective and more care- and the modification of the RMS as a frequency-dependent
ful examination showed that our new qualification had RMS parameter. Consequently, typical deviation phenom-
the right in those cases as well. The results were corre- ena were detected like periodic clacking, squeaking, or
sponding to the 95 % conformity objective, which was the low-frequency buzzing. According to the type of devi-
manufacturer’s main requirement. The final qualification of ations, different methods are suitable, e.g., for clacking
the coupling was good, if none of the diagnostic methods the SPRT and Kurtosis, for squeaking the RMS, and
results an alarm; bad, if more than three alarms occurred APSD-mask have indicated the deviation. This means that
Int J Adv Manuf Technol (2014) 75:693–703 703

any malfunction of the analyzed device can be detected by 7. Nandi S, Toliyat HA, Li X (2005) Condition monitoring and fault
one or more methods. diagnosis of electrical motors–a review. IEEE Trans Energy Conv
20:719–729
All sensitive input parameters for the diagnostic meth-
8. Ahmadi H, Khaksar Z (2011) Power spectral density technique
ods were determined based on experimental results. It was for fault diagnosis of an electromotor. In: Hruschka ER Jr
established that if three of the selected parameters indicated et al (eds) INTECH 2011, CCIS 165, pp 105–112
bad couplings, then it can be stated that they are really 9. Lin D, Wiseman M, Banjevic D, Jardina AKS (2004) An approach
to signal processing and condition-based maintenance for gear-
bad. Later test on production line proved that our system
boxes subject to tooth failure. Mech Sys Sign Proc 18:993–
trained on this principle could identify bad products with the 1007
requested (95 %) confidence level. 10. Li L, Mechefske CK, Li W (2004) Electric motor faults diagnosis
The introduced two new methods, namely the SPRT and using artificial neural networks. Insight 46:1–6
11. Liu X-Q, Zhang H-Y, Liu J, Yang J (2000) Fault detection
the frequency-dependent RMS values, and also new thresh-
and diagnosis of permanent-magnet DC motor based on param-
old methods described in the paper contributed very much eter estimation and neural network. IEEE Trans Energy Conv
to the high confidence level. 47:1021–1030
In the future, we wish to substitute the manual training by 12. López-Escobar C, González-Palma R, Almorza D, Mayorga P,
Carnero MC (2012) Statistical quality control through process
some more advance training (for example by neural network
self-induced vibration spectrum analysis. Int J Adv Manuf Tech-
method), but for the manufacturer, this simpler version was nol 58:1243–1259
also good to serve the production line. 13. Abellan-Nebot JV, Subirón FR (2010) A review of machining
monitoring systems based on artifical intelligence process models.
Int J Adv Manuf Technol 47:237–257
14. Yu G, Li C, Sun J (2010) Machine fault diagnosis based on Gaus-
Acknowledgements We wish to say thank you to Hungarian Acous-
sian mixture model and its application. Int J Adv Manuf Technol
tic and Industrial Diagnostic Testing Laboratory (MAIDLab) of the
48:205–212
College of Dunaújváros for providing us with the measurement instru-
15. Chen C-C, Liu N-M, Chiang K-T, Chen H-L (2012) Experimental
ments and for providing also the original idea for this work. The work
investigation of tool vibration and surface roughness in the preci-
is supported by the TÁMOP-4.2.2.A-11/1/KONV-2012-0027 project.
sion end-milling process using the singular spectrum analysis. Int
The project is co-financed by the European Union and the European
J Adv Manuf Technol 63:797–815
Social Fund.
16. Lie T-I, Lee J, Singh P, Liu G (2014) Real-time recognition of ball
bearing states for the enhancement of precision, quality, efficiency,
safety, and automation of manufacturing. Int J Adv Manuf Technol
71:809–816
17. Wald A (1945) Sequential tests of statistical hypotheses. Ann
References Math Stat 16:117–186
18. Ou Y, Wu Z, Chen S, Lee KM (2010) An improved SPRT con-
1. Al-Najjar (1996) Total quality maintenance—an approach for con- trol chart for monitoring process mean. Int J Adv Manuf Technol
tinous reduction in costs of quality products. J Qual Maint Eng 51:1045–1054
2:4–20 19. Ou Y, Wu Z, Yu F-J, Shamsuzzaman M (2011) An SPRT control
2. Zhang J, Li Z, Wang Z (2012) A new adaptive control chart for chart with variable sampling intervals. Int J Adv Manuf Technol
monitoring process mean and variability. Int J Adv Manuf Technol 56:1149–1158
60:1031–1038 20. Deak T (2009) The usage of the SPRT algorithm in real-time,
3. Shi D, Axinte DA, Gindy NN (2007) Development of an paper for the student research conference. College of Dunajvros
online machining process monitoring system: a case study 21. Szappanos G, Por G (1998) Improvement in the theory of identi-
of the broaching process. Int J Adv Manuf Technol 34: fication of Burst-shaped events for fault diagnosis. Nucl Sci Eng
34–46 130:261–267
4. Ganesan R (2008) Real-time monitoring of complex sensor data 22. Box GEP, Jenkins GM (2008) Time series analysis: forecasting
using wavelet-based multiresolution analysis. Int J Adv Manuf and control, 4th edn. Wiley, New York
Technol 39:543–558 23. Hunyadi L (2009) About the fourth moment. Hung Stat Rev
5. Sun Q, Mechefske CK (2010) Failure detection in automotive light 87:262–286
assemblies during vibration endurance testing. Int J Adv Manuf 24. Davenport WB, Root WL (1987) An introduction to the theory of
Technol 51:799–810 random signals and noise. IEEE Press, Wiley, New York
6. Hu W, Sun Q, Mechefske CK (2013) Condition monitoring for the 25. Hegeduš H, Mostarac P, Malaric R (2011) Comparison of RMS
endurance test of automotive light assemblies. Int J Adv Manuf value measuerement algorithms of non-coherent sampled signals.
Technol 66:1087–1095 Meas Sci Rev 11:79–84