FEATURE EXTRACTION FROM

SENSOR
DATA (VIBRATION OR MOTION)
Mannar Mannan T
Student of Electronics and communication, Anna university RO Tiruchirappalli.

INTRODUCTION
A characteristic of these large data sets is a large number of variables that require a lot of
computing resources to process. That’s why we use Feature extraction. it refers to the process
of transforming and reduce the raw data into a smaller number of features that can be processed
while preserving the information in the original data set. It yields better results than applying
machine learning directly to the raw data .

Signal features and time-frequency transformations

When analysing signals and sensor data, we can use Signal Processing
Toolbox™ and Wavelet Toolbox™ to provide functions that let us measure common
distinctive features of a signal in the time, frequency, and time-frequency domains. You can
apply pulse and transition metrics, measure signal-to-noise ratio (SNR), estimate spectral
entropy and kurtosis, and compute power spectra.

Time-frequency transformations, such as the short-time Fourier transform (STFT) can be used
as signal representations for training data in machine learning and deep learning models. For
example, convolutional neural networks (CNNs) are commonly used on image data and can
successfully learn from the 2D signal representations returned by time frequency
transformations.

Spectrogram of a signal using short-time Fourier transform. Spectrogram shows variation of frequency content
over time.
Features for audio applications and predictive maintenance
Audio Toolbox™ provides a collection of time-frequency transformations including Mel
spectrograms, octave and gammatone filter banks, and discrete cosine transform (DCT), that
are often used for audio, speech, and acoustics. Other popular feature extraction methods for
these types of signals include Mel frequency cepstral coefficients (MFCC), gammatone
cepstral coefficients (GTCC), pitch, harmonicity, and different types of audio spectral
descriptors. The Audio Feature Extractor tool can help select and extract different audio
features from the same source signal while reusing any intermediate computations for
efficiency.
Extract Regions of Interest from Whale Song
We have audio data from a Pacific blue whale. The file is from the library of animal
vocalizations maintained by the Cornell University Bioacoustics Research Program. The raw
audio signal is like the below one. its look like complicated doesn’t. if we can visually
recognize and understand it. It will be easy to machine too.

so, we get 2 seconds of data and do spectrum analysis and in the result the
spectrum shows a noticeable peak at around 900 Hz.
As like above we get 3 samples in size of 2 seconds and we do extraction .so we
can compare their spectra of three moans. the three regions of interest you just
extracted is displayed below. Their spectra lie approximately on top of each
other.

our focus
we specifically going to do feature extraction from sensor (accelerometer) data. That produce 3 data about
acceleration in 3 direction like x, y and z. I will show how to process and generate feature from the data
Literature survey
Average
The arithmetic mean is a data set's most commonly used and readily understood
measure of central tendency. In statistics, the term average refers to any
measurement of central tendency. The arithmetic mean of a set of observed data
is equal to the sum of the numerical values of each observation, divided by the
total number of observations. Symbolically, for a data set consisting of the
values x1,x2,………xn, the arithmetic mean is defined by the formula:

Root mean square

The RMS value of a set of values (or a continuous-time waveform) is the
square root of the arithmetic mean of the squares of the values, or the square of
the function that defines the continuous waveform. In physics, the RMS current
value can also be defined as the "value of the direct current that dissipates the
same power in a resistor."

The RMS over all time of a periodic function is equal to the RMS of one period of the function. The RMS
value of a continuous function or signal can be approximated by taking the RMS of a sample consisting of
equally spaced observations. Additionally, the RMS value of various waveforms can also be determined
without calculus, as shown by Cartwright.[4]
we use root mean square because if we take average of a sine or cosine signal it
will be zero because of positive and negative phase of the signal. but in root
mean square the values will be squared so the negative became positive so we
get the mean value

FOURIER TRANSFORM
The Fourier transform is an extension of the Fourier series, which in its most
general form introduces the use of complex exponential functions.

The Fourier transform is some important mathematical function. I still think that
there's something fun and enriching about seeing what all of its components
actually look like. The central example, to start, is goanna be the classic one:
Decomposing frequencies from sound. Let’s dive in.

This first sound right here is a pure A 440 beats per second .2nd 440 oscillations
each second. A lower-pitched note, like a D, has the same structure, just fewer
beats per second. And when both of them are played at once, what do you think
the resulting pressure vs. time graph looks like?
And as you add in other notes, the wave gets more and more complicated.

But right now, all it is, is a combination of four pure frequencies.

a signal like this,

AND DECOMPOSE IT
INTO THE PURE FREQUENCIES THAT MAKE IT UP. PRETTY INTERESTING, RIGHT? ADDING UP
THOSE SIGNALS REALLY MIXES THEM ALL TOGETHER.SO PULLING THEM BACK APART...FEELS
AKIN TO UNMIXING MULTIPLE PAINT COLORS THAT HAVE ALL BEEN STIRRED UP TOGETHER.

NOW WHAT'S GOING ON HERE WITH THE TWO DIFFERENT SPIKES, IS THAT IF YOU WERE TO
TAKE
TWO SIGNALS, AND THEN APPLY THIS ALMOST-FOURIER TRANSFORM TO EACH OF THEM
INDIVIDUALLY,

AND THEN ADD UP THE RESULTS,

WHAT YOU GET IS THE SAME AS IF YOU FIRST ADDED UP THE SIGNALS, AND THEN APPLIED
THIS ALMOSTFOURIER TRANSORM.

THIS PROPERTY MAKES THINGS REALLY USEFUL TO US, BECAUSE THE TRANSFORM OF A PURE
FREQUENCY IS CLOSE TO 0 EVERYWHERE EXCEPT FOR A SPIKE AROUND THAT FREQUENCY.
SO, WHEN YOU ADD TOGETHER TWO PURE FREQUENCIES, THE TRANSFORM GRAPH JUST HAS
THESE LITTLE PEAKS ABOVE THE FREQUENCIES THAT WENT INTO IT.SO THIS LITTLE
MATHEMATICAL MACHINE DOES EXACTLY WHAT WE WANTED.IT PULLS OUT THE ORIGINAL
FREQUENCIES FROM THEIR JUMBLED-UP SUMS, LIKE UNMIXING THE MIXED BUCKET OF
PAINT.SO THANKS TO
THIS MARVELLOUS ONE WE CAN GET FRQUENCY COMPONETS OF A SIGNAL

What is a Power Spectral Density (PSD)?

Vibration in the real world is often "random" with many different frequency
components. Power spectral densities (PSD or, as they are often called,
acceleration spectral densities or ASD for vibration) are used to quantify and
compare different vibration environments. Engineers will see PSDs used in test
standards such as MIL-STD-810 and others that provide guidance on how to
qualify new products & systems for various operational and transportation
environments. I don't want to belabour the math that goes into a PSD (we have
our handbook for that!). To illustrate what a PSD is, let's dive into some data!
Here are vibration profiles in the time domain over 1 minute for a few different
environments:

• NAVMAT (temperature and random vibration screening for Navy

contractors)
• Jet Aircraft from MIL-STD-810 (another test standard for military
applications)
• Highway Truck from MIL-STD-810
• Rail Cargo from MIL-STD-810
• Shipboard from MIL-STD-810

Looking at this data in the time domain can tell us that the vibration in a jet is
likely more severe than rail. But that's only determined from looking at the peak
amplitude which isn't always a good indicator. And it doesn't tell us where that
energy lies... or what the power density is - that's why we have the power
spectral density!
 here
2
you can see that the PSD plots power (g / Hz) in the y axis as a function of
frequency (Hz) in the x axis. In this format, we can clearly see the distinction
between the vibration environments. So, what is this g2 / Hz thing? To answer
that question, let's define how a PSD is calculated. The power spectral density
function XPSD(f) is calculated from the discrete Fourier transform X(f) as

The one-half factor is needed to convert the amplitude from peak2/Hz to rms2/Hz
which highlights another benefit of PSDs we'll explore later on. The frequency
step is finite in practice and is the inverse of the total measured duration.

This frequency step is the smallest sine wave frequency which can be resolved.
A wider Δf gives greater PSD confidence in terms of smoothing the spectral
components.

Difference Between PSD and FFT

Although engineers are tempted to use FFTs (fast Fourier transforms) for
spectrum analysis, they should really be using (PSDs) power spectral densities.
The reason is that PSDs are normalized to the frequency bin width preventing
the duration of the data set (and corresponding frequency step) from changing
the amplitude of the result. FFTs don't do this!
Let's take that same data from the jet standard and calculate and plot FFTs for
different time lengths. Check out how the longer the time series is (and the finer
the frequency bin width/step is), the lower the amplitude of the FFT becomes!
This means you can't really use FFTs to compare and quantify vibration
environments
.

Now let's do that same calculation for the PSD. In this plot I've kept the
frequency bin width the same at 8 Hz. These different time lengths do nothing to
the amplitude of the resulting PSD because they are normalized to the bin width.

Benefits and Features of PSDs

Adjustable Width to Smooth Bin

 You probably noticed in the last section how amazingly clean those PSD were!
This is another major benefit of PSDs because they are dependent on that
frequency bin width. You can manipulate this to smooth a PSD. The
VibrationData Toolbox has a great feature that I highlight in our article on
generating a PSD test standard. Here I highlight how changing that bin width
from a very fine 0.02 Hz to 4 Hz will drastically smooth the PSD while keeping
the overall energy in each frequency accurate and consistent across the whole
range.

METHODOLOGY
One of the most important things we can do in machine learning, is figure out
which features to extract from our data to send to our model for training and,
ultimately, inference. Many modern deep learning techniques do allow us to
send raw data right to the model, and have the model figure out which features
it cares about. But that often requires a lot more processing power than we have
available in some of our embedded systems. At this point, you may be
wondering, what is a feature? For machine learning, a feature is an individual
measurable property or characteristic of a phenomenon being observed. Let's
start with a simple example.
Here is a plot of one of our ten-second motion examples we gathered earlier. By
looking at it, can you tell which motion it is? Look at the legend to see what's
moving the x, y, or z axis. This is the left to right motion as the x-axis plot tells
us it was moving back and forth. Now, let's start by looking at a very simple
example of a feature set. We'll take a single point in time and look at the x, y,
and z acceleration data at that point.

We could, potentially, use this as a simple feature set. Each value in the feature
set makes up a dimension in the input to our machine learning model. So, we'd
say that this feature set has three dimensions.
Our machine learning model would then, always, expect three inputs. We
would use these features and associated labels to train the model. When it's
done training, the model would, once again, look for three features as inputs,
and use those three features to try to predict the class that they belong to. Do
you think that these features make for a good feature set for our model to
predict these particular motions? Why or why not? These are actually pretty
poor features, because they do not take a time sequence into account. It's just a
snapshot in time of what the accelerometer is seeing. The input to the model is
just this snapshot. There's no information in these features about how the
acceleration changes over time. The nice thing about three dimensions is that
we can plot that point pretty easily and visualize what's going on. Now let's plot
similar points taken from different samples from each of our four classes.
I've made each of the classes a different colour. It's a little messy, but you can
see that there are distinct shapes. Most of the time, with classification problems
in machine learning, if you can visually separate the clusters of samples, you
can train a machine to do the same. However, if you look at the centre of the
group, you can see that all of the samples merged together in one clump. This is
where the x, y, and z acceleration in each of the four motions, looks about the
same. So, there are some snapshots in time where all of the classes look about
the same, and the machine learning model will have a hard time discerning the
difference among them. One way to fix this is to use the points in a longer
window of time as inputs. Let's make our window long enough to capture one
or two cycles of the movement, which would be about two seconds in each axis.

With a sampling rate of 62.5 hertz, that gives us 125 samples for each axis. So,
in total, that's 375 data points being fed into the machine learning model.
Unfortunately, this means our input matrix to the model is now 375 dimensions,
which is much harder for humans to visualize. Many machine learning
models, like deep neural networks, are capable of extracting the necessary
features from large inputs like this, automatically. They can take large input
arrays and learn what to look for to make good decisions. However, there are
two main problems with this approach. Here is an example of a deep learning
model.

These networks can take images which consists of hundreds or thousands of

pixels corresponding to hundreds or thousands of dimensions, figure out what
features to look for in those images, and then classify the image based on those
features. The problem is that, as the machine learning model grows in size and
complexity, so, do the computational requirements. It takes a lot of processing
power to automatically learn what features are useful for the classifier, and then
use that feature extractor in real time. With a 1,000-dimension input, you're
going to need to store, at least, that amount of data in memory, and be able to
perform mathematical operations on each of those values.
Feature extraction sections in such a model can often be several layers deep,
multiplying the computational complexity required to perform inference.
Additionally, as we make machine learning models larger, then generally
require a lot more data to train. Often, we don't have an endless supply of
training data and an endless amount of time for training. While such deep
learning models may be great for huge server farms, doing things like natural
language processing, we don't have that luxury in embedded systems. Even
though we can, and will, use deep neural networks in this class, you will benefit
greatly by choosing appropriate features from the raw data. You ultimately want
to keep your machine learning model as small and fast as possible. As most
machine learning algorithms require a lot of memory and processing power, and
those are in limited supply in most embedded systems. One way we can do this,
is by choosing the features for the model, manually. Rather than picking
and choosing some of the raw samples, we can combine the samples in a variety
of ways to generate unique features that help describe what's going on in the
system. For example, maybe we calculate the root mean square for all 125
samples in each axis.

This is a straightforward calculation that gives us a single number per axis, and
gives us something like an average or mean, for all these numbers in each axis.
We're back to having only three dimensions as the input for our model, and they
take into account two seconds of data. One thing you might want to do is filter
out the mean of each set of data before calculating the root mean square, so that
gravity is taken out of the equation
.
We could probably create a classifier model that works very well on just RMS
data, as there is a lot of separation between these feature groups.

But we can take this a step further. One common technique when looking at
vibration or motion data, is to take the Fourier transform of that data to get
information about it in the frequency domain. We won't go into detail about
how to calculate the Fourier transform, but know that it's a way to break apart
a signal into its various frequency components. The Fast Fourier Transform or
FFT is optimized for discrete sampled data like what we have here.
This is an example of the left to right motion. If you look closely at the
repeating left to right plot on the top graph, you can see that it takes about one
second to complete one cycle. That means it has a frequency of one hertz.
Because this is a prominent feature, it will stand out in the frequency domain.
Take a look at the bottom graph and you'll see a peak at one hertz, noting that
there is a large one hertz component. Similarly, if you were to move the phone
back and forth faster at, say, three hertz, this peak would then be at three hertz.
Edge Impulse takes this even another step further and computes the Power
Spectral Density or PSD from the FFT.
The FFT can be somewhat misleading sometimes, as the amplitude of the FFT
in each bin can vary based on the frequency width of that bin. The PSD helps
by normalizing the amplitude to the width of each bin. Rather than feed all of
the values in the PSD to the model, Edge Impulse extracts a few important
characteristics from it.

The first, is identifying the amplitude and frequency location of the highest
peaks of each axis. The second, is they sum all of the values between certain
ranges in the PSD. These two actions help describe the shape of the PSD
without needing to use all the values in it, which would require a higher
dimension input to the model, and we want to avoid that. Just know that the
power spectral density is another good set of features, especially, when it comes
to motion and vibration data. Edge Impulse is taking 375 raw samples from our
2 second window and computing a number of features.
From each axis, we'll send the RMS value, three peak amplitudes from the PSD,
three peak frequency locations from the PSD, and four summed bins from the
PSD to the model. So rather than have a 375-dimensional input to our model,
we have a 33-dimensional input, which is much better.

Result
By doing feature extraction we particularly extract the parametrical data from the
original data. because unlike like see from single side we choose the best view
parameter and specifically extract that and look from that way so we can easily
understand and also machine can also easily understand too and get the best result

Parameter of interest
As you can see from the above details that power spectral density is best for
vibration data analysis .and look upon our data is acceleration and it is just
like vibration so PSD is best. The reason is that PSDs are normalized to the
frequency bin width preventing the duration of the data set
(And corresponding frequency step) from changing the amplitude of the result
CONCLUSION

So, this is the end of this, we discussed about the feature extraction from the
acceleration data. And the steps to do that and benefits and how its benefit and
what are the difference like wise. And that all now