Condition Based Monitoring Statistical Functions

Ver 0.0.1

Created: 1/7/2019
Author: Sharkey, Tom
Last Modified: 7/1/2021
Modified by: Sharkey, Tom
Modes:
Terms used:
Data frame – Used to describe a set of 512 elements of “raw” time domain data. This is the amount of
data seen on the time series plot and so a data frame can be considered a basic unit, to which each
statistical formula is applied.

Data point – One element of “raw” time domain data, sent from a mote.

RPM – Revolutions per minute. Describes the rotational speed of the motor or vibration source being
measured. Converted to Hertz by dividing by 60.

PDF – Probability Density Function. A function of a continuous random variable whose value at any given
point can be interpreted as providing a relative likelihood that the value would equal that sample. Useful
visualisation of the spread of time series data values – also necessary to describe the effects of kurtosis
and skew factor on the data. See Figure 6.

g – Measurement of acceleration used in applications describing impact and vibration of mechanical

systems. 1g = 9.8m/s2 .

Mathematical notation:

• xi - Current data point

• 𝒙̅ / µ - Mean
• σ - Standard deviation.
• N - Number of data points in calculation
• ││ - Modulus or positive value.
• μ3 - Third central moment.
• μ4 - Fourth central moment
DFT – Discrete Fourier Transform
Explanation DFT provides spectral analysis of the time domain signal. It contains the same
information as the time domain plot but provides a much clearer visualisation of the
motor / imbalanced load acting on the mote. Clear spikes in frequency are seen at the
frequency of motor rotation (RPM/60).

Measures the dominant frequency of the repetitive impulse period of certain faults
due to the contact between rolling elements and defective spot. Therefore, the type of
bearing defects such as outer race, inner race or ball defect can be detected [1] .
“Discrete” simply describes a computationally-efficient algorithm which involves
breaking up the time data into equally-spaced samples
Calculation The Fourier transform used is a fast Fourier transform performed on the embedded
used system i.e. the mote (C code).
Typical Figure 1 below shows typical values for DFT for an imbalanced motor at 2000 RPM.
values
Application The DFT plot is the most effective method of showing what source(s) of vibration are
acting on the mote in question.

The rotation of the motor is seen at lower frequencies, at the fundamental frequency
of the motor and integer multiples of this frequency.
Bearing faults constitute small spikes at high frequencies, which should also be visible
on this graph, if any are present.
Plot shown In Figure 1, the GUI plots the time and frequency data for a motor acting on an
imbalanced load at 2000 RPM. This evaluates to about 33.3Hz, which is the first large
spike seen in the figure. Additionally, the imbalance creates small frequency spikes at
integer multiples of this fundamental frequency.
Code See ARM math for extra info
https://www.keil.com/pack/doc/CMSIS/DSP/html/group__RealFFT.html#ga180d8b76
4d59cbb85d37a2d5f7cd9799
Figure 1: DFT data for 2000 RPM motor with imbalance

Figure 2: Explanation of DFT data

Peak
Explanation Provides useful indication for the max value of the time domain amplitude of the signal.
Not used in characterisation of motor or bearing faults, but can be used to calculate
some other useful statistics for bearing fault characterisation, such as crest factor.

Included in program to give user more flexibility with calculation of statistics, and a
means of setting an alarm on the max value of acceleration that a mote should reach.
Calculation Compares largest value in each raw data frame to the current max. If it is larger, sets a
used new max.
Typical Vibration source dependent. Test motor did not exceed 0.6 g.
values
Application For user visualisation and setting of alarms.
Plot shown Figure 3 shows peak values for almost 40 data frames, quite steady around the 0.53g
value. Label on peak graph shows max value achieved over all data frames collected.
Code

Figure 3: Peak
P2P - Peak-to-Peak
Explanation Similar to peak value, but calculates amplitude from max to min, rather than
from zero value to max.
Calculation used Subtracts max time series value from min time series value achieved.
Typical values ~1g when measuring an imbalanced motor at 2000 RPM – will vary based on
motor used – should be approximately double peak value.
Application Not used in further statistical analysis. Useful for setting alarms on mote
values i.e. Set alarm to trigger if P2P exceeds 1g acceleration.
Plot shown Figure 4 shows peak-to-peak values for 25 data frames. Approximately 1.020
in value.
Code

Figure 4: Peak to Peak

RMS – Root Mean Square
Explanation Root mean square (RMS) of the time series data. Denotes the average value
of the signal if all values have been made positive.
Calculation used Sums the squares of all data points in data frame, divide by length of frame,
get square root of answer. Basically the mean if all values are first made
positive.

𝑥𝑟𝑚𝑠

Typical values 𝑃𝑒𝑎𝑘

For a sin wave, RMS is given by or 0.707 of the peak value.
√2
Application The RMS value increases gradually as a bearing fault develops. It can be used
to characterise when a bearing fault might have occurred [1].
Plot shown Figure 5 shows RMS value for over 20 data frames. It is a very steady value for
the waveform shown, at 0.34 acceleration.
Code

Figure 5: Root Mean Square

Standard Deviation
Explanation Standard Deviation is a measure of the average distance a single data point is
from the mean value (0). A large standard deviation indicates that all data
points are spread over a wide range, not clustered around the mean value.
Calculation used Sums the square of the difference between each data point in a frame and
the mean. Calculates the square root of this sum averaged.

𝑁
1
𝑠= √ ∑(𝑥𝑖 − 𝑥 )2
𝑁−1
𝑖=1

Where xi is the current data point, 𝑥 is the mean, N is the number of data
points summed.
Typical values From testing – standard deviation values for our specific motor were around
0.335. This implies that most of the data lies between ±0.67 of the mean (0).
See “Chebyshev’s Inequality” for greater detail on this.
Application Standard deviation is used in conjunction with kurtosis to assess bearing
faults. Kurtosis is known to increase as a bearing fault develops, but this only
tells half of the story. Standard deviation describes how far points spread out
from the mean, and kurtosis indicates whether outliers are infrequent and
extreme (high kurtosis), or frequent and moderately sized (low kurtosis).
Plot shown Figure 6 describes the relationship between standard deviation in the normal
distribution and the percentage of data contained within each section. This
gives an approximation of the percentage of data contained within 1, 2 and 3
standard deviations.

Figure 7 is a typical example of standard deviations taken from an

unbalanced motor rotating at 2000 RPM. See “Typical Values” for relevance.

Code
 Figure 6: Standard Deviation for Normal Distribution [2]

Figure 7: Standard Deviation

Note: The standard deviation values shown in Fig.7 correspond to the width of the σ intervals in Fig.6. A
larger value of standard deviation would indicate a data distribution with values which are more spread
out from the mean.
Kurtosis
Explanation The “tailedness” of the data, kurtosis describes the “mass” of data contained
in the tails of the probability density function. Kurtosis indicates whether
outliers are infrequent and extreme (high kurtosis), or frequent and
moderately sized (low kurtosis).

The kurtosis displayed is the “excess kurtosis” which is the actual kurtosis
value minus 3. This causes the kurtosis value of the normal distribution to be
0.0. A distribution with larger and more infrequent outliers will have a higher
kurtosis. Such a distribution is “leptokurtic” – see Figure 8.
Calculation used Scipy.kurtosis given here:
https://github.com/scipy/scipy/blob/v1.3.0/scipy/stats/stats.py#L1137-L1214
Uses the second and fourth moments to calculate kurtosis

𝑋− 𝜇 4 𝜇4
𝐾𝑢𝑟𝑡[𝑋] = 𝐸 [( ) ]= 4
𝜎 𝜎
Where μ4 is the fourth central moment and σ is the standard deviation.
Typical values Kurtosis’ value for normal rolling element bearing (using the Fisher definition)
is 0.
Application Deviations from a zero value for kurtosis can be used to indicate a bearing
fault.
Plot shown Figure 8 shows the difference between positive and negative kurtosis, as
compared to a normal distribution which has 0 excess kurtosis

Figure 9 shows 30 data frames of typical values for excess kurtosis.

Code Function taken from scipy.stats, see “Calculation used”.

Figure 8: Kurtosis Explanation [3]

Figure 9: Kurtosis
Skew Factor
Explanation Quantifies asymmetry behaviour of vibration signal through its probability
density function. A function which is negatively skewed will have a long tail to
the left, the data is more concentrated to the right. A function with a positive
skew has a long tail to the right.
Calculation used The method employed is the Fisher Pearson coefficient of skewness.
𝑋− 𝜇 3 𝜇3
𝛾1 = 𝐸 [( 𝜎
) ] = 𝜎3
Where μ is the mean, σ is the standard deviation, μ3 is the third central
moment.
Typical values Normal distribution – 0
Exponential distribution – 2
Application It is known the skewness value for a normal bearing is approximately zero.
This will shift either negative or positive as a bearing fault develops.
Plot shown Figure 10 shows the difference between positive and negative skewness as
compared to the normal distribution.
Code Function used is from scipy stats:
https://github.com/scipy/scipy/blob/v1.3.0/scipy/stats/stats.py#L1137-L1214

Figure 10: Skew Explanation [4]

Figure 11: Skew Factor
Crest Factor
Explanation Indicates how extreme the peaks are in a waveform. Describes the ratio of
the peak of a waveform to the effective value (rms).
Calculation used
|𝑥𝑝𝑒𝑎𝑘|
𝐶=
𝑥𝑟𝑚𝑠

Typical values Some default values of crest factor are

Sin wave – 1.414
Triangle wave - 1.732
Gaussian noise - ∞
The values for an unbalanced motor at 2000 RPM were ~1.63.
Reason used The greater the crest factor, the greater the energy contained in higher
frequency harmonics. As higher frequency harmonics are often used to mark
bearing issues, the crest factor can be applied to find these faults.
Plot shown Figure 12 shows typical values for crest factor for an unbalanced motor at
2000 RPM.

Code

Figure 12: Crest Factor

References
[1]
https://www.researchgate.net/publication/320074126_A_Review_of_Feature_Extraction_Methods_in_
Vibration-
Based_Condition_Monitoring_and_Its_Application_for_Degradation_Trend_Estimation_of_LowSpeed_Sl
ew_Bearing

[2] https://commons.wikimedia.org/wiki/File:Standard_deviation_diagram.svg

[3] http://www.columbia.edu/~ld208/psymeth97.pdf

[4] https://commons.wikimedia.org/wiki/File:Negative_and_positive_skew_diagrams_(English).svg