SPIE Digital Library Get updates from SPIE Newsroom
  • Newsroom Home
  • Astronomy
  • Biomedical Optics & Medical Imaging
  • Defense & Security
  • Electronic Imaging & Signal Processing
  • Illumination & Displays
  • Lasers & Sources
  • Micro/Nano Lithography
  • Nanotechnology
  • Optical Design & Engineering
  • Optoelectronics & Communications
  • Remote Sensing
  • Sensing & Measurement
  • Solar & Alternative Energy
  • Sign up for Newsroom E-Alerts
  • Information for:
    Advertisers
SPIE Photonics West 2017 | Register Today

SPIE Defense + Commercial Sensing 2017 | Call for Papers

Journal of Medical Imaging | Learn more

SPIE PRESS




Print PageEmail PageView PDF

Sensing & Measurement

Reducing speckle noise in vibration signals measured with laser doppler vibromety

A new method of processing vibration data can reveal mechanical faults in electric motors on the production line.
5 October 2006, SPIE Newsroom. DOI: 10.1117/2.1200609.0390

In industry, engineers often use vibration measurements to detect mechanical faults, or defects, in a device.1 In fact, faults can develop during manufacturing, so vibration-based diagnostics provide quality control for commercial products such as electric motors. In fault diagnostics, piezo-electric accelerometers are used most for sensing vibrations. For on-line quality control, however, accelerometers have several limitations, including invasiveness and problematic installation.2 For this reason, laser doppler vibrometry (LDV)—a noncontact approach that measures the velocity and displacement of a vibrating device—is growing in popularity.2–7

Still, LDV measurements from rough surfaces can be distorted, which results in speckle noise. Here, we present a new method that avoids this problem by either confirming that a measurement is undistorted, or selecting an unaffected portion of a signal even when speckle noise is present. As an example, we have applied this method to bearing-fault detection.

Figure 1 presents an on-line station that uses vibrometers for quality control of electric motors.4,5 A typical measurement process accelerates the motor, keeps it rotating at a constant speed for few a seconds, and then stops it. A 16bit data-acquisition board collects the vibration signals with a sampling frequency of 50kHz. During the measurement, the motor stands—with its axis in the vertical position—on a typical pallet for assembly lines. Motors have two bearings: one in the lower part of the rotor, and one in the upper part. The most important and frequent faults of electric motors come from the rotating parts—namely, the bearings and the rotor. The diagnostic station should identify these faults, so vibration measurements are taken on both bearing housings to maximise the signal-to-noise ratio and detect anomalous vibrations. Two industrial, single-point vibrometers—one for each bearing—are used for this measurement.


Figure 1. This on-line station tests electric motors for faults by analyzing signals collected with two laser vibrometers.
 

Laser vibrometers operate reliably on clean and smooth surfaces where the amount of backscattered light is sufficient for subsequent signal processing.7 In contrast, undesired interference occurs in measurements from a rough surface, such as the bearing housing, and this causes signal distortion, which is referred to as the speckle noise. This phenomenon forms a characteristic image, which is shown in Figure 2.


Figure 2. With laser doppler vibrometry (LDV), rough surfaces can create a speckle pattern. (Image courtesy of Vadim Makarov, Norwegian University of Science and Technology, Trondheim, Norway)
 

Speckle noise includes randomly occurring impulses that increase the amplitude of the effected samples. In some samples, this creates extreme amplitudes, or outliers that can be detected with the kurtosis ratio (KR):

where β2(x) is the kurtosis of the the raw vibration signal x, and β2(xt) is the kurtosis of the trimmed signal xt. Kurtosis—a measurement of impulsiveness—is defined as the fourth-central statistical moment normalized by the fourth power of the standard deviation. The KR quantifies speckle noise. The trimmed signal is obtained by thresholding the raw signal to remove the outlying samples.8,9 As a result, β2(xt) represents a robust estimate of kurtosis, which is unaffected by the presence of speckle noise.

Detection of speckle noise operates as follows. When no speckle noise is present in x, values of β2(x) and β2(xt) are comparable, thus KR(x) ≈ 1. KR(x) for a distorted signal is much greater than one, because outliers increase only β2(x), while β2(xt) is unaffected. The signal is regarded as distorted when KR(x) exceeds 2.

Since multiple measurements are time-consuming and quality control on production lines operates under strict time constraints, signals distorted by speckle noise cannot be rejected and measured again. Therefore, we use an algorthm to obtain a signal region without speckle noise. The algorithm's input is a raw LDV signal, and the output corresponds to the chosen undistorted region in the signal. When no errors are detected, the output equals the input. The algorithm operates solely in the time domain and consists of three steps: band-pass filtering in the range 5–10kHz; segmenting the filtered signal; and calculating the kurtosis ratio of each signal segment.9

We made measurements on good (healthy) bearings, so no fault-related impulses should be present in the LDV signals. Figure 3 represents a successful case, in which the selected region is undistorted and the region length is satisfactory. As can be observed, band-pass filtering highlights the impulsive components, although the full-band signal also exhibits several random impulses. On the contrary, the second example (Figure 4) required filtering. In this case, structural resonance frequencies dominated the wide-band spectrum, and the impulses can be seen only in Figure 4(c). Here, the undistorted region is very short, and the laser beam would need to be repositioned for a second measurement. Otherwise the human operator should be informed about an excessively small amount of correct data. Nevertheless, the selected region is undistorted.


Figure 3. Using good bearings and the LDV algorithm, this example show a successful selection of an undistorted region.
 

Figure 4. In this example on good bearings, the LDV measurement and our new algorithm reveal an undistorted region that is excessively short.
 

Overall, these examples on bearings in electric motors give promise to the combination of LDV and our algorithm. Our experimental results demonstrate that that speckle noise can be effectively avoided even in severely distorted signals.

This work was supported by GA ČR grant 102/03/H085 ‘Biological and Speech Signal Modelling’ and research program MSM6840770015‘Research of Methods and Systems for Measurement of Physical Quantities and Measured Data Processing’, which is run by the Czech Technical University in Prague and sponsored by the Ministry of Education, Youth, and Sports of the Czech Republic.


Authors
Jiří Vass, Pavel Sovka, Radislav Šmíd  
Faculty of Electrical Engineering, Czech Technical University
Prague, Czech Republic

Jiří Vass is a PhD student in the Department of Circuit Theory. He studied signal processing at Tampere University of Technology in Finland and at Technion, Israel Institute of Technology. Since 2004, he has worked on the development of diagnostic systems for household appliances using laser vibrometers. He is currently on a research visit to the School of Mechanical and Manufacturing Engineering, University of New South Wales, Sydney, Australia.

Cristina Cristalli, Barbara Torcianti  
Applicazioni Elettroniche Avondale (AEA), Loccioni Group
Angeli di Rosora (Ancona), Italy

Cristina Cristalli earned her electronics engineering degree in 1990 at the University of Ancona and her PhD in bioengineering in 1995. From 1994 to 1995, Cristalli was a fellow at the Electronic Design Center, Case Western Reserve University, Cleveland, OH, where she worked on thin-film technology for sensor fabrication. From 1996 to 1998, she worked at the Instrumentation Laboratory SpA, and since 1998 she has worked at AEA, Loccioni Group where she is the responsible for the research and development group.

Barbara Torcianti earned her electronics engineering degree in 2003 at the University of Ancona. From November 2003 to February 2004, she worked with the Pattern Recognition Group at the Technical University of Delft, The Netherlands. Since 2004, she has worked in the research and development group at AEA, Gruppo Loccioni. Her research focuses on the selection and characterization of different pattern-recognition techniques based on the most recent applications of artificial intelligence for on-line quality control.

References:
5. J. Vass, C. Cristalli, Bearing fault detection for on-line quality control of electric motors,
Proc. 10th IMEKO TC10 Int'l Conf. on Tech. Diagnostics,
pp. 93-97, 2005.
6. Principles of vibrometry, Polytec,
LMInfo Special,
Vol: 1, pp. E2-E4, 2003.