Instrumentation & Measurement Magazine 25-9 - 45

6
5
4
3
2
1
6
7
5
4
3
2
1
(a)
6
7
5
4
3
2
1
(d)
20
40
60
Time (s)
(e)
Fig. 3. The response curves of No. 1 pepper with different adulterated powders. (a) Pure No. 1 pepper; (b) No. 1 with 20% wheat bran flour; (c) No.1 with 20%
rosin powder; (d) No. 1 with 20% cornstarch; and (e) No. 1 with 20% rice bran powder.
each run were not quite the same. After adopting the normalization
process and the cross-validation method to find the
optimal parameters (δ=0.5, c=210
), the prediction accuracy of
the training set and test set were both 100% for 40 runs. The
prediction results of one random run test set for No. 10 pepper
with different adulteration powders are shown in Fig. 4.
The samples of ten types of pepper including pure and
adulteration data were trained and tested separately with
SVM model. The prediction accuracy of training and test set
for each type of pepper with two random runs are listed in Table
3. Here, the class number for each type of the pepper is 5
(corresponding to the number of the adulterated powder).
It can be seen that the SVM model built here performed excellently
to identify ten pepper species and the adulteration
components in them. However, the prediction accuracy of
SVM varied for different types of peppers. The prediction accuracy
for No. 5 and No. 6 were lower than that for others.
Then, all samples formed a data set (5190 samples). The
training and test set were generated randomly (4000 sets for
training and 1190 sets for test). The classification category is 50
classes, which can be recognized as a specific adulterated ingredient
for certain type of pepper powder. The training was
performed by a normalization process using a parameter-optimized
SVM algorithm. In this case, the prediction accuracy
for training set was 100%, while it was 98.66% for the test set.
Hence, the results proved that it is feasible and effective to
identify the status of adulteration for pepper powder with
AOS based on sensor array and SVM model.
Fig. 4. Confusion matrix of SVM prediction results for one random test set of
No. 10 pepper powder.
December 2022
Effect of Normalization Process
To investigate the influence of the normalization method on
the prediction results, the samples with and without normalization
were compared to the train and test SVM models.
Here, the data set of No. 3 pepper powder were taken, for example,
and the experimental results are listed in Table 4. The
length of the samples was selected with different time scales
from original response data. Data set 1 was the results from 1
IEEE Instrumentation & Measurement Magazine
45
80
100 120
20
40
60
Time (s)
(b)
6
7
8
9
5
4
3
2
1
20
40
60
Time (s)
80
100 120
80
100
120
20
40
60
Time (s)
(c)
Sensor1
Sensor2
Sensor3
Sensor4
Sensor5
Sensor6
Sensor7
Sensor8
Sensor9
Sensor10
80
100 120
12
10
8
6
4
2
20
40
60
Time (s)
80
100 120
Parameter of sensitivity (G/G0
)
Parameter of sensitivity (G/G0
)
Parameter of sensitivity (G/G0
)
Parameter of sensitivity (G/G0
)
Parameter of sensitivity (G/G0
)
Instrumentation & Measurement Magazine 25-9

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