Instrumentation & Measurement Magazine 24-2 - 97

Mode 1: The combine's
engine is running, and
mechanized threshing is
not engaged ( " Empty " ); total recorded time was nine
seconds.
Mode 2: The combine engine is running, and
m e c h a n i z e d t h re s h i n g
is engaged with no actual threshing being
performed ( " Engaged " );
total recorded time was
Fig. 4. Possible hardware and software needed for an automatic decision-making instrument.
five seconds.
Mode 3: The combine's enIf 1D input data is used and features are to be calculated gine is running, and mechanized threshing is engaged and
from the frequency signatures of the data, the FFT is such utilized at approximately 80% capacity ( " Full " ); total recorded
a fast algorithm that it has been even successfully imple- time was ten seconds.
mented in real time on mobile phone platforms [41]. Even
though features are to be calculated, the computational cost Classification Based on a Conventional
can be minimal. This time-consuming feature extraction is NN
one good reason why only a few features are used and also Extracting features from the frequency information of a
highlights the importance of selecting the ones that are the signal via the absolute value of the Fourier transform,
most useful [31]. For that selection of features, there are very Periodogram (PG), can be an effective way to extract inwell-known statistical tools that can reduce the number of formation to be used in a classifier, as was used in [46] for
features, for example, selection of features corresponding to roller bearing fault diagnosis. For the harvester sounds, the
the maximal statistical dependency criterion based on mu- position of the highest peaks differs and the energy of diftual information [42]. Sequential forward selection is widely ferent frequency bands also differs between the different
used due to its simplicity and efficiency [43], [44]. Another three modes. With this in mind, a series of features were expopular feature reduction method, principal component tracted from the location of the three strongest peaks, the
analysis [45], is an alternative feature selection method that ratio of the amplitudes of the first and second strongest
transforms the original variables into another feature space peaks, the distance between the two strongest peaks and
the PG center of gravity. These features can be easily norbased on principal components.
Fig. 4 summarizes what has been discussed and shows how malized and be robust to the size of the FFT used as well as
the instrumentation can add many steps to an effective solu- the amount of noise, since noise tends to spread over a large
tion before data is to be used for training and decision making frequency band and does not considerably affect these feausing a neural network or DL network. Thus, we note that it is ture values.
Defining the PG as X(f), the calculated features are:
not just a matter of inputs-classifier-output, but several other
considerations can yield a more efficient device.


f1  X  f1  where X  f1   X  fi 

Proof of Concept Example

The work presented here formulates a proof of concept approach for the classification of harvester sounds that will
facilitate the creation of an important part of the automatization of this farm vehicle. Sound recordings were taken from
harvest video feed during the canola harvest in East Selkirk,
Manitoba. The canola was harvested with an S680 John Deere
combine and audio was captured with a GoPro Hero Session.
The recordings were taken from the rear of the combine near
the straw chopper. All recordings were taken in the same field
on the same day from the same machine. Sound sampled at a
rate of 48 kHz with AAC compression and automatic gain control was converted to Waveform Audio Format (.wav files) for
analysis.
Sound recordings were isolated into three different operating modes (classes) of the combine:
April 2021


f2  X  f2  where X  f 2   X  fi  i  i  1


f3  X  f3  where X  f3   X  fi  i  i  1, 2


f4  f1  f 2
N


f5 




iX  fi 

i1
N

X  fi 

i1

X  f1 

f6 

X  f2 

These features become the inputs to a classifier, and as it
is well known that the normalization process for the inputs

IEEE Instrumentation & Measurement Magazine 97

Instrumentation & Measurement Magazine 24-2

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-2