Instrumentation & Measurement Magazine 25-9 - 49

Table 10 - Prediction results for No. 10 pepper powder with random forest algorithm
Number
1
2
3
Average
Pure pepper
powder
100%
100%
100%
100%
Bran mixed
powder
100%
100%
100%
100%
Rosin powder
mixed
95.24%
100%
100%
98.41%
random forest algorithm had excellent classification results for
pepper powder No. 9.
As listed in Table 10, the random forest model also performed
excellently for No. 10 pepper sample. The recognition
accuracy of pure pepper, adulterated wheat bran and mixedrosin
powder were higher than other adulteration samples.
Moreover, the overall average accuracy reached 95%.
Compared with the results of the decision tree model (Table
8 and Fig. 7), the recognition accuracy was improved with
the random forest method (Fig. 8). The pure pepper could be
classified successfully from the samples with adulterations.
Moreover, the recognition accuracy of the powders mixed with
rosin and wheat bran were higher than that of rice bran and
Corn flour
mixed
85.71%
76.47%
100%
87.39%
Mixed with
rice bran
powder
91.67%
83.33%
100%
91.67%
Average
94.52%
91.96%
100%
95.49%
corn flour. However, all specific adulterations could be identified
with high recognition accuracy with machine learning
models including SVM, DT and RFT.
The recognition results for all ten types of pepper samples
with the RFT method are listed in Table 11. The recognition accuracy
of most types of samples was higher than 95% with the
RFT method. Hence, it is a potentially good method to recognize
adulterations in pepper powders.
Fig. 8. Confusion matrix of No. 10 pepper powder with random forest
classification results where labels 1 to 5 indicate the pure pepper, mixture with
bran powder, rosin powder, corn flour and with rice bran powder, respectively.
Comparison of Different Machine Learning Models
To compare the performance of different machine learning
models to recognize the adulterations in pepper powder, the
recognition accuracy for different adulterations were calculated
with SVM, DT and RFT method, as shown in Fig. 9.
All three algorithms could recognize the adulterations in the
pepper powders with high accuracy. The pure pepper was
identified with the highest accuracy among adulterated powder
samples with different machine learning models. On the
other hand, SVM performed better than RFT and DT in recognizing
the adulterations in the pepper samples both for
the specific species and overall results. Moreover, DT had the
lowest recognition accuracy among three machine learning
models in our test. Hence, the AOS with SVM model seems to
be a good choice to measure the quality of the pepper samples
to identify the adulations in the peppers.
Other ML models including decision tree, discriminant
analysis, k-nearest neighbor, and Ensemble Classifiers were
also tested to identify the adulterations in the pepper. A MATLAB
ML tool was used in this case. Hence, all samples (4000
samples for training and 1190 for testing) were utilized to classify
the adulterations in pepper with MATLAB ML tool. The
results of recognition accuracy for the test sets with different
machine learning models are listed in Table 12.
Table 11 - Training and testing results of all samples of ten pepper species (%)
Number of
experiments
1
2
Average
December 2022
Training
Testing
Training
Testing
Training
Testing
1
100
91.05
100
94.74
100
92.9
2
100
97.5
100
100
100
98.8
3
100
96
100
96
100
96
4
100
99
100
96
100
97.5
5
100
98
100
95
100
96.5
6
100
95
100
95
100
95
IEEE Instrumentation & Measurement Magazine
7
100
99
100
98
100
98.5
8
100
98
100
99
100
98.5
9
100
93
100
97
100
95
10
100
96
100
95
100
95.5
49
Instrumentation & Measurement Magazine 25-9

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