Sie sind auf Seite 1von 10

182 M. Tiboni et al.

A PCB uniaxial piezoelectric accelerometer, model 352C03, was used. The


acquisition system was a NI 9234 module mounted on a NI CDaq 9174 (manufactured
by National Instruments) and controlled with a dedicated software developed in Lab-
view. The sampling rate was fixed at 1 kHz and the data acquisition hardware
employed an anti-aliasing filter on board. The recorded data contain the sampled
acceleration in the time domain, acquired for a pre-established time. In Fig. 2 an
example of acquired acceleration is shown. Different working conditions were con-
sidered: velocity, loads and lubrication. The experimental parameters were: motor
power frequency (30–40–50 Hz), load (no load, eccentric load with one mass, sym-
metric load with two masses), and lubricant (Spartan EP 68-220-320-460) with dif-
ferent viscosity. The acquisitions were repeated thirty times for each case. The working
velocity is not an interesting condition for monitoring purpose, but it has been con-
sidered for the identification of the optimal values of the parameters of acquisition and
of the different pre-processing techniques. In fact, acceleration signals at different
speeds are obviously very dissimilar from each other and consequently easy to classify.
In the following we will refer to the case of speed variation as to the “calibration case”.
The loads, also easy to identify, are instead considered as a “verification case”. The
viscosity, finally, is the case used to compare the pre-processing methods.

Fig. 2. Experimental device (picture of the mechanical system and complete system scheme).

5 Results

The ANN architecture used in all the cases is a feed-forward with one hidden layer and
log-sigmoid activation functions. The number of hidden units varies for different
classifications, but for all it remains in the 20–40 range. The number of output neurons
is one for the classification of one condition and two or three if the network is destined
to the contemporary classification of two or three operative conditions. The values of
the output neurons are standardized between 0 and 1. Training uses the back-
propagation learning algorithm with momentum and validation is used to improve the
ANN capability of generalization. The samples are divided in three subsets: the training
set, the validation set and the test set. For each set the error of the neural network is
calculated. During the training the error on the training set is used for back-propagation
algorithm; the validation-error is monitored during training and when it starts to grow
the learning is interrupted. The test-error is used to verify the ANN generalization
ability on a still different data set.
Comparison of Signal Processing Techniques for Condition Monitoring 183

For the FFT pre-processing technique parameters to be correctly chosen are the
sampling frequency, the number of points sampled and the type of windowing. The first
two are common for all the pre-processing methods considered. From preliminary
classification tests performed for the calibration condition emerged that good values for
this parameters are: 1 kHz as sampling frequency, 8192 sampled points (correspondent
to about eight seconds of motion) and Hamming windowing. Figure 3 shows the result
of classification with FFT for the velocity case.

Fig. 3. Classification with: FFT, speed parameter. (a) Training error (blue, b), validation error
(green, g), test-error (red, r); (b) Speed classification.

Figure 3a shows the behavior of the training-error (blue line, b), the validation-error
(green line, g) and the test-error (red line, r). In this case, for the ANN 20 hidden
neurons were used, 300 epochs of trainings lasting about one minute. The output is
normalized between 0 and 1 and the expected value is shown on the abscissas of
Fig. 3b, while on the ordinate it is observed how the outputs are distributed around this
value (observe that there are no units of measurement in the chart because they are pure
numbers). The classification and generalization are almost perfect, as expected.
Figures 4 and 5 displays the classification results obtained for loads and oil viscosity.
While loads obtain results comparable with those of speed, viscosity is classified worse
but still acceptable. The possibility to use a single ANN to classify at the same time the
different operating conditions was investigated, by using an ANN with three output
neurons.

Fig. 4. Classification with: FFT, load parameter. (a) Training error (blue, b), validation error
(green, g), test-error (red, r); (b) Loads classification.
184 M. Tiboni et al.

Fig. 5. Classification with: FFT, viscosity parameter. (a) Training error (blue, b), validation
error (green, g), test-error (red, r); (b) Viscosity classification.

The classification results were much worse than in the case of three different ANN
dedicated to each condition. The difficulty on classifying viscosity has significantly
influenced this result. The ability to recognize viscosity was considered as a compar-
ison term for the different methods for signals pre-processing adopted.
The discrete Wavelet transform has been used to clean up from the noise the
acceleration signals, previously phased in the time domain. Two tests were conducted
passing directly into the ANN input the sampled values of the acceleration in the time
domain or the acceleration values always in the time domain but purified by the noise at
high frequencies with the wavelets. The results obtained with the noise elimination did
not significantly improve. The results of the classification in both the cases get worse
compared to the use of FFT, as shown in Fig. 6 (compared with Fig. 5). In the fol-
lowing we will include the wavelet case, compared with the others, because it is a
situation with the acceleration given directly in input to the ANN.
The PSD envelope was performed on a window of 1024 points sampled at
1000 Hz. The classification results with PSD were analogous but slightly worse than
those obtained with FFT.
For the use of the HOS pre-elaboration technique, based on the bicoherence estimated
via the direct (FFT) method, there are some parameters to be optimized. After making
some classification tests for the calibration case, the optimal choices resulted: for the FFT-
compliant length an aperture of 512; the overlap equal to 0; length of input vectors more
than 1000; the range of frequencies surrounded by the red rectangle in the Fig. 7a.

Fig. 6. Classification with: Wavelet, viscosity parameter. (a) Training error (blue, b), validation
error (green, g), test-error (red, r); (b) Viscosity classification.
Comparison of Signal Processing Techniques for Condition Monitoring 185

The classification results in the viscosity case by passing as input to the ANN the
HOS Bicoherence are reported in Fig. 8. The trend of the average values obtained (red
line in Fig. 8b) follows quite well that of the expected average values (blue line) even if
some cases are not classified correctly, as was also the case for the other indicators
considered.
For the PDF of the acceleration signals a parameter to be settled is the amplitude of
the abscissa classes. From classification tests performed for the calibration case the
optimal value was identified as 0.1 m/s2. Classification of viscosity is shown in Fig. 9.

Fig. 7. (a) Selected diagram of the output values divided into the classification bands.
(b) Average trend of the bicoherence for the three different speeds within the red box.

Table 1 specify the size and the content of the y vector of Fig. 1 for each method
considered. The ANN output (z) is a single value between 0 and 1. For the load
identification z values are 0–0.5–1; for the velocity 0–0.5–1 and for the viscosity 0.05–
0.333–0.667–0.95.

Table 1. Size and content of the ANN input vector (y) as the pre-processing technique changes.
Pre-processing Size Content
Wavelet 8192 Acceleration sampled in the time domain
FFT 80 Amplitude of the first 80 FFT components
PSD 80 Amplitude of the first 80 PSD components
HOS (bicoherence) 900 Bicoherence amplitude for a 30  30 square of frequencies
PDF (amplitude 0.1) 70 Number of classes of acceleration discretization

To obtain numerical values for the goodness of the recognition, the statistical
indicators mean and standard deviation were calculated. Table 2 reports results
obtained for the various pre-processing techniques.
186 M. Tiboni et al.

Fig. 8. Classification with: HOS bicoherence, viscosity parameter. (a) Training error (blue, b),
validation error (green, g), test-error (red, r); (b) Viscosity classification.

Fig. 9. Classification with: PDF, viscosity parameter. (a) Training error (blue, b), validation
error (green, g), test-error (red, r); (b) Viscosity classification.

The percentage of correct classification, that is the accuracy of classification, has


been also calculated. The range of the ANN output values has been divided into
horizontal bands of equal amplitude, in number equal to the cases to identify. Each
band will be assigned to a specific mode of operation of the equipment.

Table 2. Δ Mean (mean output value vs. expected output value) and standard deviation of the
ANN output for the different pre-processing techniques, in the viscosity case.
Oil identifier EP 68 EP 220 EP 320 EP 460
Expected output value 0.95 0.667 0.333 0.05
Δ mean (wavelet) −0.05 0.053 −0.083 0.06
Standard deviation (wavelet) 0.12 0.08 0.057 0.056
Δ mean (FFT) −0.07 −0.037 −0.013 0.01
Standard deviation (FFT) 0.11 0.12 0.06 0.02
Δ mean (PSD) −0.12 0.033 −0.033 0.07
Standard deviation (PSD) 0.09 0.075 0.065 0.085
Δ mean (HOS) −0.06 −0.027 −0.023 −0.01
Standard deviation (HOS) 0.09 0.11 0.07 0.02
Δ mean (PDF) −0.12 0.033 0.007 0.06
Standard deviation (PDF) 0.16 0.12 0.11 0.08
Comparison of Signal Processing Techniques for Condition Monitoring 187

The values related to each condition of motion should all fall within their respective
range, this does not mean that in case of inaccurate classifications some weights may
fall outside their respective band. A simple indicator of the goodness of a classification
can be identified by the percentage of points that fall accurately within the respective
range. Table 3 shows the results obtained from the application of this last indicator to
the viscosity parameter when the different pre-processing methods adopted change.

Table 3. Classification accuracy - CA (percentage of correct classifications) in the viscosity


case, as the pre-processing technique changes.
Oil identifier EP 68 EP 220 EP 320 EP 460
CA with wavelet 65 62 75 82
CA with FFT 84 71 85 99
CA with PSD 80 67 78 97
CA with HOS (bicoherence) 85 72 81 99
CA with PDF (amplitude 0.1) 84 70 73 93

6 Conclusion

The use of ANN as classifier, used in most cases of condition monitoring of mechanical
systems, give good results also in this situation. On the other hand, to identify the best
operator to be used in the pre-processing of signals is important, as well as to identify
the optimal values of the specific parameters. All the operators considered in the pre-
processing phase, once the parameters that characterize the system have been opti-
mized, permit to obtain acceptable classifications, but two of them (FFT and HOS
bicoherence) allow better results.

References
1. Fitch EC (1992) Proactive maintenance for mechanical system. Elsevier, Abingdon
2. Rai A, Upadhyay SH (2016) A review on signal processing techniques utilized in the fault
diagnosis of rolling element bearings. Tribol Int 96:289–306
3. Unal M, Onat M, Demetgul M, Kucuk H (2014) Fault diagnosis of rolling bearings using a
genetic algorithm optimized neural network. Measurement 58:187–196
4. Rao VV, Ratnam C (2015) Estimation of defect severity in rolling element bearings using
vibration signals with artificial neural network. Jordan J Mech Ind Eng 9(2):113–120
5. Strdczkiewicz M, Barszcz T (2016) Application of artificial neural network for damage
detection in planetary gearbox of wind turbine. Shock Vibr 2016. Article ID 4086324
6. Huang S, Wu X, Liu X, Gao J, He Y (2017) Overview of condition monitoring and
operation control of electric power conversion systems in direct-drive wind turbines under
faults. Front Mech Eng 12(3):281–302
7. Lee J, Wu F, Zhao W, Ghaffari M, Liao L, Siegel D (2014) Prognostics and health
management design for rotary machinery systems—reviews, methodology and applications.
Mech Syst Signal Process 42:314–334
188 M. Tiboni et al.

8. Kateris D, Moshou D, Pantazi X-E, Gravalos I, Sawalhi N, Loutridis S (2014) A machine


learning approach for the condition monitoring of rotating machinery. J Mech Sci Technol
28(1):61–71
9. Tiboni M, Remino C (2017) Condition monitoring of a mechanical indexing system with
artificial neural networks. In: WCCM 2017 – 1st world conference on condition monitoring
2017, London
10. Kocur D, Stanko R (2000) Order bispectrum: a new tool for reciprocated machine condition
monitoring. Mech Syst Signal Process 14(6):871–890 ISSN 0888-3270
Default Detection in a Back-to-Back Planetary
Gearbox Through Current
and Vibration Signals

Safa Boudhraa1,2(&), Alfonso Fernández del Rincón2,


Fakher Chaari1, Mohamed Haddar1, and Fernando Viadero Rueda2
1
Laboratory of Mechanics, Modelling and Production (LA2MP),
National School of Engineers of Sfax, BP1173, 3038 Sfax, Tunisia
boudhraa.safaa@gmail.com, fakher.chaari@gmail.com,
mohamed.haddar@enis.rnu.tn
2
Department of Structural and Mechanical Engineering,
Faculty of Industrial and Telecommunications Engineering,
University of Cantabria, Avda de los Castros s/n, 39005 Santander, Spain
{alfonso.fernandez,fernando.viadero}@unican.es

Abstract. Until the last century, vibration analyses have been the most used
method in monitoring the rotating machinery. But nowadays the electrome-
chanical interaction in machinery is being the new scientific trend, mainly for its
accuracy and facility.
Several research papers have shown that investigations done on the motor can
be used to describe the machines dynamic behaviour and give a deep overview
of any anomaly.
Within this context, this work reports the impact of an introduced pitting in
the sun of one of the gearboxes of a back-to-back planetary gearbox configu-
ration on the stator current signal. The gearbox monitoring was investigated
through analysing the stator current measured experimentally by clamp meter.
Later, the results presented in the frequency spectrum are obtained using Fast
Fourier Transform (FFT) to distinguish the frequency of the defect and its
harmonics besides to the mesh frequency which highlight the electromechanical
coupling.
Finally, the stator current signals were followed by the spectrum of acceler-
ation signals. These signals were registered using an accelerometer mounted on
the fixed ring in order to validate results foreseen in the current spectrum.

Keywords: Back-to-back configuration  Stator current  Planetary gear 


Stationary condition  Pitting defect  Vibration signals

1 Introduction

For as long as humans have been using mechanical mechanism in power transmission,
this revolutionary field, has been in permanent progress optimizing these components
design and especially working on its condition monitoring using non-destructive

© Springer Nature Switzerland AG 2019


A. Fernandez Del Rincon et al. (Eds.): CMMNO 2018, ACM 15, pp. 189–197, 2019.
https://doi.org/10.1007/978-3-030-11220-2_20
190 S. Boudhraa et al.

techniques. Therefore, this domain has been one of the biggest scientific and industrial
concern.
According the Chaari et al. [1], gears defects can be classified into geometry which
are regrouping the manufacturing defects and assembly defects and tooth failure that
has been an important affair in many scientific researches (like pitting, cracks or
breakages…).
This field has been an important scientific problematic since the 90th, Alfonso et al.
[2] has investigate the effect of the crack and pitting on the mesh frequency. Within the
same context, Choy et al. [3] had studied also the influence of pitting on the vibration
of a gear transmission system.
On the other hand, many other researches have turned to correlate between the
dynamic behaviour of gearbox and introducing the electric characteristics of the motor.
The claim of these proposes is weather to the influence of the electromagnetic effect
on the natural vibration characteristics of the gearbox like in the paper of Yi et al. [4]
and to monitor the mechanical system by analysing the electric inputs or to register the
mechanical unbalance presented in any component of the motor as shown in the work
of Balan et al. [5].
Meanwhile, Nabih et al. have reported in their paper [6], the impact of the presence
of any defects on the current signal using both numerical modelling. In the same paper,
that the most outstanding peaks in the frequency spectrum are related to the frequency
of the defect modulated by the sector frequency’s and of course, other frequencies can
co-exist that relate the mesh frequency and the defect frequency modulated by the
sector frequency.
In a more advanced work Ottewill et al. [7] in their paper had worked on moni-
toring a tooth defect in epicyclic gearboxes using numerical modelling pursued by
experimental validation.
However, this work presents experimental investigations done a back to back
configuration of planetary gearboxes. These studies show the electric response of the
asynchronous motor to the presence of a pitting in the mechanic system. Later inter-
pretations will be validated by recording the vibration system of the gearboxes.

2 Experimental Rig

Test Bench
The configuration of the bench is compacted with a mechanical power circulation for
economic energy efficiency [8].
Figure 1 shows in details the experimental rig which is consisted of a mechanical
system and an electric part.
As for the first, it is a back to back planetary gearbox. In such configuration the two
gearboxes are identic and connected through a shaft from the suns and test carrier and
reaction carrier are connected through a rigid hollow shaft.
The workbench is driven by an asynchronous motor, SIMENS Motor Type D-
91056 ELRLANGEN. This motor was controlled itself by a frequency inverter
“MICROMASTER 440” which is located on the electric cabinet.
Default Detection in a Back-to-Back Planetary Gearbox 191

Fig. 1. Experimental test bench

The motor was connected to the sun via a shaft, and each gearbox was composed of
16 teeth sun gear, 65 teeth ring gear and three planets, each with 24 teeth. The load is
applied by the adding mass on the arm connected to the test ring.
The mesh frequency of planetary gears depends on the fixed component. For this
configuration the ring is maintained fixed for that the mesh frequency is given by the
expression below:

Zs þ Zr
fm ¼ fs
Zs Zr ð1Þ

N
Where fs ¼ 60 : sun frequency and N is the rotating speed (rpm).
Instrumentations
For the current measurements, clamp meters of type fluke i200s ac current clamp
(Fig. 2a) were mounted around the input wires of the motor (Fig. 1) to measure the
current in each phase of the motor.
Additionally, the vibration signal was registered by the mean of a tri axial
accelerometer (Fig. 2b).
Before mounting it on the ring of the test gearbox, this accelerometer was calibrated
in the X direction using A “Brüel & Kjær” calibration exciter.
Simultaneously, we had used an optic tachometer type Compact VLS7 with pulse
tapes (Fig. 2c) on the shaft connecting the motor to the gearbox so that the angular
velocity is recorded.

Das könnte Ihnen auch gefallen