Time Domain Processing Method for acoustic signals of faulty Bearings

Acoustic Signal Detection and processing methods for rolling bearing faults have been a hot research area for decades. Thousands of research papers have been published. Generally, there are two types of processing methods: the first is to process the collected acoustic signals directly in the time domain, and the second is the frequency domain processing method, of course, some processing methods are time-frequency combinations.

This article summarizes common time domain processing methods. Since the documents I collected are incomplete, this summary only introduces some traditional methods. Due to my limited level, the popular Acoustic Emission (AE) Technology and Its Related time domain processing methods have not been involved in this article.

Time Domain Processing Method

In the Time Domain Analysis and judgment bearing acoustic signal is called the bearing fault diagnosis time domain processing method. The simplest method is to draw a waveform map of the acoustic signal of the bearing to be viewed by the eyes and analyzed by the brain. Other complex methods include using various statistical parameters to evaluate the bearing status.

Don't underestimate the original method of Bearing Fault Analysis by looking at the waveform. In fact, the human brain is a very powerful supercomputer and runs Complex Adaptive algorithms. With this seemingly simple method, we can determine whether there is amplitude modulation in the waveform, whether there is a phenomenon of eccentric axis loading, and whether there is abnormal high-frequency vibration.

1. Sound waveform of a typical faulty Bearing

Figure 1 shows the sound signal of a faulty bearing in the inner ring. The sampling frequency is 100 kHz, and the waveform of 0.1 s is shown. By observing the waveform, we can obtain a large amount of information. For example, we can see that the sound signal contains a periodic impact Signal component. This impact signal is generated by the roller hitting the fault point through the inner ring fault point. It will be further zoomed in to make it clearer.

2. Local amplification of sound Waveforms

Figure 2 is a partial enlargement of Figure 1. The figure shows the strong impact vibration at three places. The repetition frequency is exactly consistent with the passing frequency of the roller through the inner ring, therefore, it can be regarded as a single fault in the inner ring. When the roller passes through the fault point, it will arouse the structural resonance of the bearing system. As the roller passes through the fault point, this resonance will quickly decay. Generally, experienced engineers can take a look at the sound waveform to determine whether the bearing is faulty. Simply measure the interval between two adjacent Impact Vibrations to determine the bearing fault type.

The sound of faulty bearings is very different from that of normal bearings, these differences lead to a large difference between the range of the statistical parameters of faulty bearing sound and that of normal bearing sound. Based on these differences, we can determine whether the bearing is faulty.

Common important statistical parameters include:

Peak value:

The peak value reflects the maximum amplitude of the vibration waveform and is suitable for Surface Peeling faults, when such a fault occurs, the rolling element and fault point will have a strong impact, causing sudden changes in the sound signal to generate a large value signal in a short period of time. However, the peak value is vulnerable to external noise interference and is rarely used as a judgment criterion.

Mean Value:

For sound signals, the average value is 0. Generally, this parameter is not required. However, this value is sometimes calculated to determine whether the signal collection system works normally.

Root mean square value (rms value ):

The root-mean-square value is also called a valid value, which is a widely used statistical parameter. This parameter represents the energy of the sound signal produced by the bearing, and is an important indicator for determining whether the bearing operates normally. When the bearing is not faulty, the operation is stable, the sound is small, and the corresponding rms value is relatively small. As faults gradually increase, RMS will also increase. The RMS value is very effective for analyzing the trend of diagnosing wear faults or bearing oil faults. However, it is not very sensitive to faults with impact vibration, such as small-scale strip or scars on the surface.

Crest factor ):

The peak factor reflects the ratio of the maximum value of the sound signal to the valid value. The larger the ratio, the instantaneous vibration of the large value in the sound signal is about severe. For non-Faulty bearings, the peak factor is close to 3.5.

Kurtosis ):

The kurtosis factor reflects the degree of waveform deviation from normal distribution. The kurtosis of white noise is 3. Sometimes we will see another definition of kurtosis:

This definition does not introduce a new concept, but sets the kurtosis factor of the Gaussian distribution signal to 0. The larger the kurtosis value, the longer the signal deviation from the Gaussian distribution.

There are usually two ways to calculate these statistical parameters. The first method is to directly calculate the original sound signal. The second method is to filter the original signal, extract the information of different frequencies from the original signal, and calculate the statistical parameters respectively.

Usually all kinds of noise are concentrated in the low frequency section, and the high frequency section can better reflect the fault characteristics. For example, the result of filtering the sound waveform in Figure 1 with a 20-40 kHz band-pass filter.

Figure 3 audio signal after filtering (20-40 kHz)

It can be seen from the figure that after filtering, the "vibration" of the fault point becomes very obvious. The peak value also increased to 25.3, and the peak factor increased to 10.3.

The RMS Value and peak value can be used to track the trend of the bearing running status, but it is not suitable for a single determination. These two values vary greatly because of different working environments and bearings. We cannot select a proper threshold to determine whether the bearing is faulty. The kurtosis value and wave crest factor are irrelevant to the sound signal size of the bearing. Therefore, the two parameters are used to determine the bearing fault, especially the early failure of the bearing, which is more accurate. However, as bearing faults increase, the vibration characteristics of the bearing become more and more random, and the calculation results of these two parameters also decrease. Therefore, the failure severity of the bearing cannot be determined only through these two parameters.

The selection of the filtering frequency band is also very important. The following table shows the calculation results of the kurtosis and wave crest factor after filtering the bearing sound waveforms in different frequencies.

The figure shows that the frequency segment is relatively higher, and the result is better. That is to say, the fault feature sound is more obvious in the high-frequency section. However, this is not absolute. For a specific problem, experiment tests are required to obtain the optimal filtering frequency band.

This problem can also be studied from the distribution of bearing sound signal amplitude. For good bearings, the sound is similar to white noise, and the amplitude distribution is close to normal distribution. Once the bearing fails, the amplitude distribution may deviate.

The amplitude distribution of A New bearing and a faulty bearing sound is given, and the original sound signal is filtered by band-pass. The abscissa is normalized according to the standard deviation of the signal.

Figure 4 amplitude distribution of new Axle Bearing and faulty bearing sound

The sound signal points of the new bearing are mainly concentrated near 0, and there are very few signal points that deviate from the three standard deviations, with almost no signal points exceeding 5 σ. Due to the existence of faulty bearing through vibration, a considerable proportion of signal points deviate far from 0 points. Therefore, by calculating the proportion of signal points that deviate from a certain threshold value, you can also determine whether the bearing is faulty.