Instrumentation & Measurement Magazine 25-9 - 46

Table 3 - The prediction accuracy results (%) for the training and test sets of ten pepper samples.
Test
time
1
2
Average
Training
Testing
Training
Testing
Training
Testing
1
99.3
100
100
100
99.65
100
2
100
100
100
100
100
100
3
100
100
100
98.57
100
99.29
4
100
100
100
100
100
100
5
100
97.14
99.64
100
99.82
98.57
6
100
94.28
100
100
100
97.14
7
100
100
100
100
100
100
8
100
100
100
100
100
100
Table 4 - The prediction accuracy of SVM with and without normalization
Number of
experiments
Sample set
1
Without normalization process
2
1
With normalization process
2
Training
Testing
Training
Testing
Training
Testing
Training
Testing
to 120 s and set 2 was from 70 to 120 s. Data set 3 was also the
results from 70 to 120 s, but the samples were extracted from
set 2 with the time interval of 10 s. All tests were run twice. It is
seen from Table 4, the accuracy of the test set without normalization
treatment was only 10% for sample set 1 and 2. With
the normalization process before training and testing, the accuracy
of the test set was significantly improved. For sample,
set 1, the average prediction accuracy was enhanced from
10% to 80% after normalizing the data set. For sample set 2,
where part of stable response data was selected to build the
model, the prediction accuracy was also improved. However,
the prediction accuracy based on sample set 3 reached as high
as 99%, which is higher than that on sample set 2 and set 3. It
may be because the extracted data from the original responses
reduced the noise in the data. Therefore, the normalization
process improved the prediction accuracy of machine learning
model due to that it shrinks the difference among the
features from original data.
1
45%
10%
45%
10%
90%
90%
85%
70%
2
45%
10%
45%
10%
100%
70%
100%
60%
9
100
98.57
100
100
100
99.29
10
100
100
100
100
100
100
3
100%
35.71%
100%
34.28%
100%
100%
100%
98.57%
Effect of Kernel Functions
To evaluate the effect of different kernel functions on the performance
of the SVM model, the model with different kernel
functions were built to evaluate the quality of pepper powder.
All parameters of kernel functions were optimized with crossvalidation
method. The prediction accuracy of the training and
test sets with the samples of just one type of pepper (No. 10) and
all types together are both listed in Table 5. Four kernel functions
including linear, polynomial, RFF and sigmoid kernel
function were used for modeling. The results indicate that the
accuracy of the test set with the RBF kernel function performed
better than that with the linear kernel, polynomial, and sigmoid
kernel functions. However, the prediction accuracy using the
model trained with all data collected from ten types of pepper
was slightly lower than that with just one type of pepper. It may
be because the number of categories for all samples was 50, and
it was five for just one type of pepper. Too many labels may take
more deviations in the training and testing process.
Table 5 - The prediction accuracy of SVM with different kernel functions
Data set
No. 10 pepper
All samples
46
Kernel function
type
Training set
Test set
Training set
Test set
Linear kernel
98.75%
85%
82.93%
82.02%
Polynomial
kernel
100%
95%
24.08%
23.03%
IEEE Instrumentation & Measurement Magazine
RBF
100%
100%
100%
98.66%
Sigmoid
100%
85%
23.85%
23.45%
December 2022

Instrumentation & Measurement Magazine 25-9

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-9