A NEW METHOD OF PROCESSING ROLLING ELEMENT BEARING SIGNALS

by Duncan L. Carter

Abstract

Antifriction or rolling element bearing diagnostics using demodulated high frequency acceleration signals is a technique introduced roughly thirty years ago that has become widely available in recent years. Currently, bearing diagnostic systems are available using hardware or software demodulators along with high pass or band pass filters used to select the portion of the frequency spectrum processed for analysis in simple devices such as current data collectors and multiple high pass or bandpass filters in more sophisticated test systems [1]. Continuous monitoring systems for bearings using peak and average acceleration levels are also widely available. When the data contains only signals from the bearing under test, good accuracy can result with existing systems. However, if diagnostics are made when vibrations from other machine sources such as shaft harmonic series, motor excitation frequencies, or gear mesh harmonic series are present in the data, the accuracy of the diagnostics and monitor levels can be seriously degraded [2]. This paper presents a newly patented [3] set of techniques for eliminating interfering signals from the bearing diagnostic process. The process also allows for concurrent high pass or band pass filter implementation as needed for different diagnostic and monitor systems.

Character of the bearing signals

Conventional vibration analysis is based on the assumption that vibrations can be described as harmonic series of frequencies corresponding to the motion of the components of the machinery being analyzed [4]. The signals commonly used for high frequency bearing analysis result from the response of the bearing materials and their surroundings to non-linear sources caused by impacts and friction between the bearing elements. For conventional vibration analysis, vibration signals are assumed to be series of discrete frequencies while the flaw and friction induced bearing signals are, instead, a continuous distribution of frequencies. The differences in the basic character of the bearing signals and other machine vibrations make practical the separation of the bearing signals from other machine signals using digital signal processing techniques. This paper presents two variations of a basic approach, each useful for different applications. One variation separates the bearing signals from other machine vibration sources before demodulation and the other variation removes specific demodulated signal components.

Basic method

In both variations of the method, data is sampled in the time domain either as the output of a transducer with optional preamplifier/high pass filter or as the output of a transducer/demodulator/high pass or band pass filter system. Next, the data is transformed to the frequency domain, typically with a fast Fourier transform, with the output of the transform being an array or arrays of complex numbers. The complex frequency domain data is then multiplied in appropriate order by a filter array of appropriate shape and inverse transformed back to the time domain for further processing as desired. Filtering by this process produces errors that are concentrated at the ends of the output time domain data blocks requiring that data at the ends of the arrays where data errors exceed design limits be discarded. If desired output data lengths are greater than available transform size, data can be processed in overlapping blocks, allowing the data to be streamed continuously, especially if the filter process is integrated with the raw data collection process.

Pre-detection filter example.

This example uses data sampled with a Vibration Specialties Corporation SpectraVib analog to digital converter board which VSC uses in both their portable data collector and their PC based continuous monitor systems. The data in this example was sampled and processed for export to a bearing diagnostic software package, DREAM, an acronym for Diagnostic Rolling Element Analysis Module, produced by VibroAcoustical Systems and Technology, Incorporated of Saint Petersburg, Russia. DREAM uses the envelope detected or demodulated frequency spectrum of bearing signals in a portion of the direct frequency spectrum selected using criteria based on the physical dimensions of the bearing being tested and the relative rotating frequencies of the inner and outer races of the bearing. The bearing used for this example has a single row of balls and is SKF part # 6226/C3 with the following dimensions: inner race I. D. = 130 mm, outer race o. d. = 230 mm, overall width = 40 mm, ball diameter = 1.25 inches, 9 balls, & pitch diameter = 180 mm. It is installed in a Reliance 1250 horsepower induction motor operated at approximately 15 Hertz or 900 RPM. In this example, DREAM calls for a 1/3 octave band pass filter with a center frequency of 5000 Hertz to filter the data before demodulation. The interference consists of the motor slot pass frequencies modulated by the harmonic series of twice the motor power supply frequency, 60 Hertz. In this and the following example, the accelerometer is mounted roughly 40 centimeters from the bearing, producing substantial attenuation of the signal at the pass band frequencies of the filters used.

First, the data is collected. Figures 1 is the first 0.2 seconds of a 2.4 second time domain sample.

Figure 2 is the spectrum of the same 2.4 second time domain sample. The y axis of figures 2, 7, and 8 are decibel scales with 90 dB equal to 1 G RMS.

Simultaneously with the data collection, the basic band pass filter, shown in figure 3, is created. The actual filter used is a variant of a 1/3 octave filter that has significantly better impulse response than standard 1/3 octave filters[6].

Next, the data is transformed to the frequency domain as an array of complex numbers. The data is then analyzed to the determine the frequencies of the interference to be removed. Next, the original band pass filter is modified by the addition of very narrow notches to the basic filter response at the frequencies of the interference, after which the overall filter gain is increased to compensate for the portion of the bearing signal removed with the interference, a process that is done automatically without operator input. Figure 4a is the plot of the modified band pass filter and figure 4b is an expanded or zoomed section of the modified filter.

Next the complex frequency domain data is multiplied by the modified filter array of figure 4 , inverse transformed back to the time domain and then demodulated as shown in figure 5.

Figure 6 is the demodulated time domain data with only the band pass filter operations having been performed.

Figure 7 is the average envelope spectrum for sixteen blocks and is only band pass filtered with the filter in figure 3 without the motor excitation interference removed.

Figure 8 is the spectrum of the same band pass filtered data as in figure 7 except that the interference frequencies have been removed . Note that, not only is the 120 Hertz harmonic series of the motor excitation removed, but also lower level distortion products are eliminated which considerably increases the accuracy of the diagnostic program, allowing the bearing to be tested as though the interference did not exist.

486 class micro-processors are adequate to process applications like this example at a "real time" rate. Since this is an "all software" process, no special filter or demodulator hardware is required, allowing the construction of an integrated data acquisition, filter, and demodulator software module. If desired, raw data can be collected and stored for later analysis. In addition, the software filters can be automatically optimized in each application for parameters like data signal to noise ratio.

Post-detection filter example.

While the pre-detection filter can automatically separate continuous signal interference source from bearing sources, the post-detection filter is more useful for continuous monitoring and other screening applications where time domain peak and average values are used and especially where specific, known or measurable frequency components are desired to be removed from a demodulated signal. One such screening example is the separation of rail car wheel surface flaw signals from rail car bearing signals. For a given data size, processing requirements are roughly 20 percent of the pre-demodulation filter.

Data for this example is from the same motor application used for the pre-demodulation filter example. In this case, it is desired to remove the 120 Hertz harmonic series produced by the motor excitation. Data was demodulated after being filtered with a two pole, 5 KHz high pass filter, emulating the basic method used by several instrumentation manufacturers. The input time domain data from the demodulator is transformed to the frequency domain as a complex number array. Next, the data points around the calculated interference frequencies are multiplied by a notch filter shape as shown in figure 9.

Then, a new value for the average value of modified complex frequency domain data is calculated, in this case by scaling the average value of the input data by the ratio of the dynamic parts of the filtered frequency domain output to the dynamic parts of the input frequency domain data. Additionally, a gain compensation value is calculated. Next, the filtered frequency domain data is transformed back to the time domain so that peak and average values can be calculated. Figure 10 is the plot of the 5 KHz high pass filtered envelope signal of figure 1 and figure 11 is its spectrum.

The peak and average acceleration values in figure 10 of 0.166 Gs and 0.078 Gs produce a crest factor of 2.13. Figure 12 plots the same data except the that the 120 Hertz harmonic series caused by motor excitation is removed, producing the peak, average, and crest factor values of 0.095 Gs, 0.029 Gs and 3.28.

Summary

The filter methods shown in this paper provide for substantial improvements in the state of the art of bearing testing. For diagnostic purposes, the pre-detection filter provides automatic separation of bearing signals from other machine sources with the flexibility to accommodate very sophisticated automatic diagnostic systems without need for operator input. Since it is an all software process, no special hardware is required once the basic time domain data is collected. Similarly, the post-detection filter furnishes the ability to provide time domain data values for monitoring and screening where known or measurable interference frequencies exist in the envelope detected data. In unusual measurement and diagnostic situations, both processes may be applied.

Return to Inteltek home page Return to bibliography page

1. Azovtsev Yu., Barkov A., Yudin I. "Automatic Diagnostics and Condition Prediction of Rolling Element Bearings Using Enveloping Methods", presented at the 18th annual meeting of the Vibration Institute, June, 1994.

2. A. Y. Azovtsev, A. N. Barkov, and D. L. Carter, "Improving The Accuracy of Rolling Element Bearing Condition Assessment", presented at the 20th annual meeting of the Vibration Institute.

3. Duncan L. Carter, U. S. Patent Number 5,477,730, “Rolling Element Bearing Condition Testing Method and Apparatus” issued December 26, 1995.

4. Duncan L. Carter, “Some Instrumentation Considerations In Anti-friction Bearing Condition Analysis”, presented at the 17th annual meeting of the Vibration Institute, June, 1993.

5. Angelo, Martin, "Vibration Monitoring of Machines," Bruel & Kjaer Technical Review No. 1-1987, pp. 1-36.

6. R. B. Randall, "Frequency Analysis", third edition, pp. 83-87, published by Bruel and Kjaer, 1987.

Return to Inteltek home page Return to bibliography page