A Low-Cost Antipodal Vivaldi Antenna-Based Peanut Defect Rate Detection System

Abstract

Peanut quality, with the defect rate as a critical determinant, has a profound impact on its market value. In this study, we introduce an innovative non-destructive evaluation method for peanut defects. Differing from traditional and often expensive or complex detection methods, our approach utilizes a low-cost antipodal Vivaldi antenna, complemented by a custom-designed defect rate detection system. Prior to experimentation, we simulated the antenna and system architecture to ensure their operational efficiency, a step that not only conserves resources but also validates the reliability of subsequent results. We conducted experimental tests on fresh peanut pods, obtaining electromagnetic scattering parameters (S₁₁ and S₂₁ magnitudes/phases within 1–2 GHz) through non-destructive measurements. These parameters were used as input features, while the defect rate served as the output variable. By implementing the XGBoost algorithm, we established predictive models for defect rate quantification (regression) and defect grade classification. In comparison to some traditional statistical models, such as linear regression, which may struggle with non-linear data patterns, XGBoost effectively modeled the complex relationship between the scattering parameters and the defect rate. Experimentally, the regression model achieved an R² value of 0.8113 for defect rate prediction, and the classification model reached an accuracy of 0.7526 in grading defect severity. The entire device, costing less than USD 50, provides a significant cost advantage over many commercial systems. This low-cost setup enables real-time evaluation of peanut pod defects and efficiently categorizes the defect rate without the time-consuming sample preparation and tiling operations required by traditional image-based inspection methods. As a result, it offers an affordable and practical solution for quality control in peanut production, showing great potential for wide application in the peanut industry.

1. Introduction

Peanuts hold a pivotal position in the global agricultural economy as a major oilseed crop. However, their cultivation is frequently marred by pod rot, which can be attributed to a confluence of factors, including suboptimal field management practices and the onslaught of disease infections [1]. This pod rot phenomenon not only leads to a significant decline in the yield of peanuts but also severely compromises their quality. A particularly notable scenario is when the peanut husks deteriorate while the internal kernels remain unblemished. In the commercial marketplace, such pods are either flat-out rejected or procured at meager prices, inflicting substantial economic losses upon farmers. Thus, the accurate identification of these partially damaged peanuts emerges as a linchpin in optimizing pod utilization, curbing waste, and alleviating financial hardships within the peanut-farming community. This pressing need underscores the urgency of developing a cost-effective and efficient peanut defect detection system.

In recent years, the field of peanut quality assessment has witnessed a preponderance of image-recognition-based techniques [2,3,4,5]. For instance, the peanut pod quality detection algorithm (PQDA) built upon modified ResNet-derived convolutional neural networks (CNNs) meticulously compared the performance of architectures such as ResNet18, AlexNet, and VGG16. Through a series of iterative optimizations of these three algorithmic models, the refined PQDA achieved an impressive classification accuracy of 98.1%. Similarly, another machine-vision-based approach introduced a hierarchical grading system for pod rot. This system effectively circumvented the labor-intensive and error-prone nature of manual inspections. By integrating a tiered classification module, it managed to enhance the recognition accuracy by an average of 6.26 percentage points when benchmarked against YOLOv5, attaining an accuracy of 95.7% for intact pods and 90.8% for decayed pods.

Despite the intuitive nature and initial success of these vision-based methods, their practical utility is severely circumscribed by the prevalent storage practice of multi-layer peanut pod stacking. In this stacked configuration, pods often overlap, leading to partial surface occlusion. As a result, traditional imaging techniques fall short of comprehensively evaluating rot in obscured specimens. Furthermore, decayed peanut pods exhibit not only structural degradation in both husks and kernels but also measurable moisture variations [6,7,8,9,10]. Samir Trabelsi et al. [11] developed a microwave moisture meter using non-off-the-shelf components. Operating at a single frequency of 5.8 GHz, this meter applied the free-space transmission measurement principle to determine the dielectric properties of peanut pods. Field experiments demonstrated its ability to measure the moisture content of peanut kernels, with a standard error of 0.82% when compared to the official moisture meter. The instrument had a measurement range of 6.2–21% for peanut pod water content. Microwave-based sensing, in contrast, boasts superior penetrative capabilities and heightened dielectric sensitivity to moisture fluctuations. These attributes endow it with distinct advantages over optical methods, especially when inspecting bulk pods under stacked conditions.

In light of the aforementioned limitations, this study sets out to introduce an innovative peanut pod rot detection system. Our approach is centered around antipodal Vivaldi antennas, leveraging the free-space method to explore the intricate correlation between defective pod ratios and scattering parameters (S parameters). A miniaturized vector network analyzer (Nano VNA) is deployed to conduct multi-frequency, multi-point S parameter measurements. This novel experimental setup allows for a more comprehensive and detailed data acquisition compared to single-point or single-frequency measurement methods employed in previous studies. Subsequently, the acquired data are processed using the extreme gradient boosting (XGBoost) algorithm. This enables the establishment of predictive models for accurately estimating rot incidence and classifying defect severity. In comparison to some of the simple statistical models utilized in earlier peanut quality studies, XGBoost excels in handling the complex relationship between scattering parameters and the defect rate, effectively capturing non-linear patterns within the data. One of the most compelling aspects of our study is the cost-effectiveness of the proposed system. The total cost of the entire device is less than USD 50, presenting a time-efficient, labor-saving, and highly cost-effective solution for low-budget defective pod evaluation systems.

2. Materials and Methods

2.1. Design of Detection Device

2.1.1. Design of Antenna

The experimental system requires antennas operating in the 1–3 GHz frequency band with a voltage standing wave ratio (VSWR) below 2 across the entire operational range. Figure 1 shows the designed antipodal Vivaldi antenna for transmitting and receiving electromagnetic signals. This antenna features a relatively simple structure composed of two metallic layers fabricated on a dielectric substrate through etching or printing processes [12,13,14,15,16]. The substrate employs FR4 material with a dielectric constant of 4.4, thickness of 2 mm, and loss tangent of 0.02, ensuring low-cost and straightforward manufacturing. Owing to its planar configuration, the antenna reduces system volume and complexity while maintaining excellent radiation performance over a wide frequency range. It achieves high gain, stable phase characteristics, and minimal signal distortion within the operational band [17,18].

The specific parameters of the antenna are shown in Figure 2 and

Table 1. Antenna specification parameters.

Parameters	Value (mm)
Sl	116.54
Sw	152.42
Hfi	53.77
Hfo	123.7
Lfi	97.31
Lfo	55.03
Wf	1.446
Hti	27.61
Hto	61.12

.

The different frequency direction diagrams are shown in Figure 3.

As shown in Figure 4, the antenna exhibits radiation angles of 160° at 1 GHz, 90° at 2 GHz, and 85° at 3 GHz. The voltage standing wave ratio (VSWR) ranges from a minimum of 1.1711 to a maximum of 1.901. From the above data, the designed antenna meets the intended objectives.

2.1.2. Overall Structural Design of the Device

The Nano VNA V2 (SAA-2N) used in the test device is a product developed based on the Nano VNA open source project [19]. The Nano VNA is a small-sized handheld vector network analyzer. Its price is only USD 37, while the common vector network analyzer on the market requires more than USD 10,000. In this study, the 500 MHz–2 GHz range was explored. Nano VNA is connected to the computer through USB and then connected to the antenna through the SMA connection line. Based on the designed antenna, the overall detection device is designed. The main structure includes two antipodal Vivaldi antennas and one acrylic sample box with a specification of 200 × 200 × 100 mm. In addition, Nano VNA is used to detect scattering parameters, and several supporting structures are added. The main components are shown in Figure 5.

2.2. Device Simulation

To verify the usability of the device, simulations of the main structure were conducted using CST STUDIO SUITE 2019 before formal experiments. The simulation tests involved filling the sample container with air, wood, dry sandy soil, and iron blocks to investigate the effects of different materials on the scattering parameters. The electric field distribution diagrams for each material visually demonstrate the propagation of electromagnetic waves in different media. The electric field distribution diagrams obtained during the simulations are shown in Figure 6.

Taking 1 GHz as an example, distinct differences in electromagnetic wave transmission through various materials can be clearly observed from the electric field distribution diagrams: electromagnetic waves propagate through air with negligible attenuation, while partial energy loss occurs in wood and dry soil. The iron element completely reflects electromagnetic waves, with the received signal at the receiving end being nearly negligible. This phenomenon provides a basis for distinguishing different media in the sample box through variations in received signals.

Although electric field diagrams visually demonstrate energy transmission, practical measurement of such images presents challenges. Fortunately, quantitative analysis of reflection loss and other electromagnetic wave transmission characteristics can be achieved through scattering parameter measurements. Therefore, during validation simulations and subsequent experiments, the scattering parameters S₁₁ (reflection coefficient) and S₂₁ (transmission coefficient) were primarily recorded as discriminative indicators for characterizing different materials in the sample box. The simulation results are presented in Figure 7.

2.3. Experimental Methods

2.3.1. Measurement of Peanut Defect Rate and Grading of Defective Peanuts

First, the defect rate of peanut samples should be statistically calculated according to Formula (1).

$$Defectrate={\displaystyle \frac{Defectivepods}{Totalpods}}\times 100\%$$

(1)

The grading of defective peanuts is based on the defect rate: Grade 1, no defective peanuts (defect rate = 0); Grade 3, 0 < defect rate ≤ 10%; Grade 5, 10% < defect rate ≤ 25%; Grade 7, 25% < defect rate ≤ 50%; Grade 9, defect rate > 50%.

2.3.2. Sample Preparation

The selected peanut variety was Huayu 60, purchased from Qingdao Lufenghua Seed Sales Co., Ltd. (Pingdu, China). As shown in Figure 8, the pods with no damaged and full peanut shells were selected as sound pods, and the pods with damaged, mildewed, scratched and other defective pods on the surface of peanut pods were listed as defective pods. Defective peanuts were manually selected and reserved, then mixed with sound peanuts according to the defect rate standards outlined in the grading system. Samples were prepared in 5% increments of defect rate, with grading determined by the corresponding defect rate.

Statistical results of the prepared samples’ defect rates and grading are shown in

Table 2. Details of defect rate and grading in prepared peanut samples.

Sound Pods	Defective Pods	Total Pods	Defect Rate	Grade
200	0	200	0%	1
190	10	200	5%	3
180	20	200	10%
170	30	200	15%	5
160	40	200	20%
150	50	200	25%
140	60	200	30%	7
130	70	200	35%
150	100	250	40%
140	115	255	45%
140	140	280	50%
130	160	290	55%	9
120	180	300	60%
120	220	340	65%
100	300	400	75%
75	300	375	80%
60	340	400	85%
40	360	400	90%
20	380	400	95%
0	400	400	100%

.

2.3.3. Measurement of Scattering Parameters

Prior to measurement, calibrate the Nano VNA using a calibration kit. The SOLT (short-open-load-through) calibration method, a traditional two-port calibration technique, employs one-port calibration standards including a short circuit, open circuit, and load. Four standard parts are connected in turn and measured in the Nano VNA program, respectively. The relatively simplified equation is obtained by the built-in program. The system error term is solved by derivation and simplification. Finally, these error terms are brought into the error correction formula to complete the error correction of the DUT measurement value, and finally, a more accurate measurement result is obtained. By performing SOLT calibration [20] measuring these known standards (short, open, load, through), 12 equations containing 12 error terms are established. Solving these equations yields the specific values of the 12 error terms, enabling the derivation of the true S parameters of the device under test (DUT) from the measured four S parameters.

It is known that when water molecules vibrate in a microwave field, polarization loss occurs. The polarized molecules change from a disorderly arrangement to an oriented alignment. During the re-arrangement process, mutual collisions and frictions occur between polar molecules. Energy is consumed during these collisions and frictions, thus changing the electric field strength of the microwaves transmitted through the material. When the transmitted microwave signal passes through the peanuts being tested, the peanuts will consume a part of the energy, with the water in the peanuts consuming the vast majority of it. Due to the different sizes of the kernels in high-quality pods and defective pods, their moisture contents also vary. Therefore, this difference can be used to distinguish between the two.

During measurement, the peanut sample is placed between the transmitting and receiving antennas. The S parameters of peanuts with varying defect rates are measured across the microwave frequency range using the experimental setup. For each defect rate grade (gradient), three samples are sequentially measured, and their S parameters are recorded. The final S parameters for each gradient are obtained by averaging the results from the three corresponding samples.

2.4. Modeling Method

2.4.1. Data Pre-Processing

First, 5 feature columns and 1 target variable column were extracted from the dataset. Subsequently, polynomial feature engineering was performed using PolynomialFeatures [21] to generate second-order polynomial feature combinations (e.g., interaction terms and squared terms), thereby increasing the complexity of the feature space. Following this, feature scaling was applied via StandardScaler [22,23] to standardize the polynomial features, ensuring zero mean and unit variance. The final feature set comprised 21 engineered features, as summarized in

Table 3. Feature combination details.

No.	Feature	No.	Feature
1	1	12	S11 Magnitude (dB)2
2	Frequency (Hz)	13	[S11 Magnitude (dB), S11 Phase (deg)]
3	S11 Magnitude (dB)	14	[S11 Magnitude (dB), S21 Magnitude (dB)]
4	S11 Phase (deg)	15	[S11 Magnitude (dB), S21 Phase (deg)]
5	S21 Magnitude (dB)	16	S11 Phase (deg)2
6	S21 Phase (deg)	17	[S11 Phase (deg), S21 Magnitude (dB)]
7	Frequency (Hz)2	18	[S11 Phase (deg), S21 Phase (deg)]
8	[Frequency (Hz), S11 Magnitude (dB)]	19	S21 Magnitude (dB)2
9	[Frequency (Hz), S11 Phase (deg)]	20	[S21 Magnitude (dB), S21 Phase (deg)]
10	[Frequency (Hz), S21 Magnitude (dB)]	21	S21 Phase (deg)2
11	[Frequency (Hz), S21 Phase (deg)]

.

Finally, the training set and the test set are divided, and the data are divided into training sets and test sets according to the ratio of 8:2.

2.4.2. XGBoost

RandomizedSearchCV is a highly efficient hyperparameter optimization method designed to accelerate the tuning process by randomly sampling a fixed number of combinations from the hyperparameter space for evaluation, thereby significantly reducing computational time [24]. By leveraging RandomizedSearchCV, optimal hyperparameters can be systematically identified and validated. The candidate values of hyperparameters are shown in

Table 4. Hyperparameter candidate values.

Hyperparameters	Value
subsample	0.6, 0.7, 0.8,
n_estimators	1000, 2000, 3000, 4000
max_depth	5, 6, 7, 8
learning_rate	0.01, 0.05, 0.1, 0.2
colsample_bytree	0.6, 0.7, 0.8

.

To develop practical prediction and classification models for peanut defective pod rate, an XGBoost algorithm was employed to train the relevant data, and the model has achieved good results in various studies [25,26,27,28].

The mathematical formulation of the XGBoost model, as described in reference [29], is given as follows:

$${\widehat{y}}_i={\displaystyle \sum _{k=1}^K{f}_k(x_i)}$$

(2)

where ${\widehat{y}}_i$ represents the predicted value of the sample to be tested, $f_k\in \{f(x)=w_{q(x)}\}$; $f_k\left(x_i\right)$ represents the score of the ith sample in the kth regression tree, K denotes the total number of samples, $x_i$ is the ith input data, and $w_{q\left(x\right)}$ is the weight of each leaf node of the tree q.

The objective function is formulated as:

$$L={\displaystyle \sum _{i=1}^{n}l\left({\widehat{y}}_i,y_i\right)}+{\displaystyle \sum _{k=1}^K\Omega \left(f_k\right)}$$

(3)

$$\Omega \left(f\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda {‖w‖}^2$$

(4)

where $y_i$ is the actual value of the defect rate of the sample to be tested, $\gamma$ is the penalty regularization term of the leaf tree; $T$ is the number of leaf nodes of the regression tree, $\lambda$ represents the penalty regularization term for the leaf weights; and $‖w‖$ is the modulus of the leaf node vector.

The final objective function is:

$$L^{\left(t\right)}={\displaystyle \sum _{i=1}^{n}l\left[y_i,{\widehat{y}}^{\left(t-1\right)}+f_t\left(x_i\right)\right]}+\Omega \left(f_t\right)$$

(5)

where l represents the loss function, ${\widehat{y}}^{\left(t-1\right)}$ represents the predicted value of the defect rate of the sample to be tested given by the model at the step $\left(t-1\right)$, $f_t\left(x_i\right)$ is the predicted value of the new model to be added at step t, and $\Omega \left(f_t\right)$ is the regularization term that suppresses the complexity of the newly added model f_t.

Define $I_j=\left\{i|q\left(x_i\right)=j\right\}$ as the set of all sample serial numbers falling into the leaf node j under the tree structure, where $q\left(x_i\right)$ indicates the node that the sample falls into after being predicted by the tree q.

Through the above expression, we can obtain:

$${\widehat{L}}^{(t)}={\displaystyle \frac{1}{2}}{\displaystyle \sum _{j=1}^T(H_j+\lambda )(w_j+\frac{G_j}{H_j+\lambda }}{)}^2+\gamma T-{\displaystyle \frac{1}{2}}{\displaystyle \sum _{j=1}^T\frac{G_j^2}{H_j+\lambda }}$$

(6)

where $G_j={\displaystyle \sum _{i\in I_j}{g}_i}$ is the sum of the first-order gradient statistics of all characteristic data samples falling into leaf i, $H_j={\displaystyle \sum _{i\in I_j}{h}_i}$ is the sum of the second-order gradient statistics of all water characteristic data samples falling into leaf i; $g_i$ represents the first-order partial derivative of the samples contained in the leaf node j, $h_i$ represents the second-order partial derivative of the samples contained in the leaf node j, and $I_i$ represents the set of all sample serial numbers falling into the leaf node j; $w_j$ represents the leaf weight.

The final leaf weight is calculated by Equation (7):

$$w_j^{\ast }=-{\displaystyle \frac{G_j^2}{H_j+\lambda }}$$

(7)

After substituting the optimal leaf weight, the objective function is simplified to Equation (8):

$$\widehat{y}=-{\displaystyle \frac{1}{2}}{\displaystyle \sum _{j=1}^T\frac{G_j^2}{H_j+\lambda }}+\gamma T$$

(8)

2.4.3. Model Evaluation Metrics

(1) Regression Model Evaluation Criteria

This study employs two regression model evaluation metrics: the coefficient of determination (R²) and mean squared error (MSE). The R² metric quantifies the linear correlation between predicted values (${\widehat{y}}_i$) and ground-truth values ($y_i$), with its value range bounded in [0, 1]. An R² value approaching 1 indicates stronger linear correlation and better model fitting performance. The MSE measures prediction error dispersion by computing the squared standard deviation between predictions and actual values, where values closer to 0 signify smaller model errors. Superior prediction performance is achieved with higher R² and lower MSE values, while the converse indicates degraded performance. The mathematical formulations are given in Equations (1) and (2):

$$MSE={\displaystyle \frac{1}{n}}{{\displaystyle {\sum }_{i=1}^n\left(y_i-{\widehat{y}}_i\right)}}^2$$

(9)

$$R^2=1-{\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}}{{\displaystyle {\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}}$$

(10)

where n denotes the sample size, $y_i$ the ground-truth value, ${\widehat{y}}_i$ the predicted value, and $\overline{y}$ is the mean value of the real value $y_i$.

(2) Classification Model Evaluation Criteria

Accuracy measures the proportion of correctly classified samples relative to the total sample size, reflecting the model’s overall prediction correctness. This metric provides an intuitive assessment of classification performance across all samples and serves as a fundamental evaluation criterion.

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(11)

where TP (true positive) denotes the number of samples that are actually positive and correctly predicted as positive; TN (true negative) represents the number of samples that are actually negative and correctly predicted as negative; FP (false positive) indicates the number of samples that are actually negative but incorrectly predicted as positive; FN (false negative) refers to the number of samples that are actually positive but erroneously predicted as negative.

In addition, there is a classification report, which includes indicators such as precision, recall, F1 score and support, and can evaluate the performance of the classification model in detail from multiple perspectives. The accuracy rate focuses on the proportion of samples predicted as a class that actually belong to the class; the recall rate focuses on the proportion of samples that actually belong to a certain class that are correctly predicted as such; the F1 value is the harmonic average of the precision rate and the recall rate, and the two are considered comprehensively. Support represents the number of samples in each category in the data set.

3. Results and Analysis

3.1. Effect of Peanut Defective Rate on Scattering Parameters

From Figure 9, it can be seen that under the same conditions, the S₁₁ parameter of peanut is less affected by the defect rate, only a very few frequency points are different, and it is difficult to see obvious changes from the picture. The relationship between S₂₁ and the defect rate is particularly obvious, especially between 1 and 2 GHz. With the change in the defect rate, S₂₁ also has a large change trend. However, it is unfortunate that it is difficult to intuitively summarize the appropriate rules from the pictures to describe the relationship between S parameters and peanut defect rate. Therefore, it is necessary to construct a prediction model of peanut defect rate that can be applied in practice.

3.2. Prediction Model of Defect Rate

The parameters of the trained defect rate prediction model are as follows:

It can be seen from Figure 10 that there are differences in the value of feature importance. The top three eigenvalues are S₂₁ Magnitude (dB)², [Frequency (Hz), S₂₁ Magnitude (dB)], [S₁₁ Magnitude (dB), S₂₁ Magnitude (dB)]. Among them, S₂₁ Magnitude (dB)² has the highest feature importance value, which is close to 0.12, indicating that this feature plays a relatively key role in the prediction of the defect rate model.

Figure 11 shows the relationship between the different hyperparameters of the XGBoost model and the mean test score. The specific analysis is as follows:

Figure 11a shows the average test score distribution when n_estimators are 2000, 3000, and 4000. It can be seen that as n_estimators increase, the median of the average test score increases (from close to −0.028 to close to −0.022), indicating that an appropriate increase in the number of trees may improve the performance of the model.

Figure 11b shows the average test scores when learning_rate is 0.01, 0.05, 0.1, and 0.2. When learning_rate is 0.05, the median of the average test score is higher and the distribution is more concentrated, indicating that the model performs better and has higher stability under this value.

Figure 11c shows the average test scores when max_depth is 5, 6, 7, and 8. When max_depth is 7, the median of the average test score is higher and the distribution is relatively concentrated, indicating that the model performance is better at this depth. When max_depth is too large (e.g., 8), the average test score is low, and over-fitting may occur.

Figure 11d shows the average test scores when the subsample values are 0.6, 0.7, and 0.8. When the subsample is 0.7, the median of the average test score is slightly higher, and the data distribution is relatively concentrated, which means that the model performs relatively well under this sampling ratio.

Figure 11e shows the average test scores when colsample_bytree is 0.6, 0.7 and 0.8. When colsample_bytree is 0.6, the median of the average test score is higher, indicating that the average test performance of the model is relatively good under this column sampling ratio.

The hyperparameters of the final optimal model based on the above analysis are shown in

Table 5. The best value of hyperparameters.

Hyperparameters	Value
subsample	0.7
n_estimators	4000
max_depth	7
learning_rate	0.05
colsample_bytree	0.6

.

Finally, for the training set, a mean squared error (MSE) of 0.000178 and a coefficient of determination (R²) of 0.998075 are obtained. For the test set, the mean squared error (MSE) is 0.017245 and the coefficient of determination (R²) is 0.811373. The results are shown in Figure 12.

3.3. Defect Grade Evaluation Model

It can be seen from Figure 13 that there are differences in the value of feature importance. The top three eigenvalues are frequency (Hz), frequency (Hz)², and S₂₁ Magnitude (dB)². Among them, the frequency (Hz) feature importance value is the highest, close to 0.9, indicating that this feature plays a relatively key role in the defect grading model.

According to Figure 14, the confusion matrix of the training set of the hierarchical model, the model on the training set has a higher number of correct classifications in most categories, indicating that the model has a better fitting effect on the training data. Compared with the training set, the number of correct samples of each classification on the test set is generally reduced, and there are more misclassifications, indicating that the generalization ability of the model on the new data (test set) is lower than that of the training set. The accuracy of the training set was 0.9999, and the accuracy of the test set was 0.7526.

The classification report of the test set of the grading model is shown in

Table 6. Grading model test set classification report.

Table	Precision	Recall	F1 Score	Support
1	0.54	0.48	0.50	181
3	0.48	0.41	0.44	397
5	0.66	0.61	0.63	632
7	0.70	0.67	0.69	1073
9	0.86	0.94	0.90	2018
accuracy	0.75			4301
macro avg	0.65	0.62	0.63	4301
weighted avg	0.74	0.75	0.75	4301

.

It can be seen from the classification report that the accuracy rate of grade 3 is 0.48, the recall rate is 0.41, and the F1 score is 0.44, indicating that the model also has many errors in the prediction of grade 3, and the overall performance is not good. The accuracy rate of grade 9 is 0.86, the recall rate is 0.94, and the F1 score is 0.90, indicating that the model performs better when predicting grade 9.

The largest number of samples, 2018, was registered at grade 9, while the smallest number of samples, 181, was registered at grade 1. The imbalance in the number of samples may affect the performance of the model, resulting in better performance of the model in categories with a large number of samples. The overall accuracy of the model is 0.75, indicating that 75% of the prediction of the model is correct on the whole, and there is still room for improvement.

4. Discussion

In the context of the continuous progress of science and technology, research on the dielectric properties of agricultural products has been growing in depth. Nevertheless, the absence of well-developed detection devices persists. This study centers on the peanut defect rate, an essential yet under-explored aspect of peanut quality evaluation. By leveraging Nano VNA to measure the scattering parameters of peanuts with varying defect rates, we constructed a unique peanut defect rate detection platform. This approach enables the acquisition of S parameters for peanuts under different defect rate conditions, presenting a fresh perspective on peanut quality inspection. In contrast to traditional methods that predominantly rely on visual inspection or basic physical property measurements, our dielectric-property-based method can identify internal defects that may not be apparent on the surface. We employed XGBoost to formulate a prediction model for the peanut defect rate, using different S parameters as inputs and the defect rate as the output. The XGBoost model achieved a determination coefficient (R²) of 0.81137, signifying relatively high goodness of fit. XGBoost can effectively handle complex non-linear relationships among variables, making it more suitable for the intricate dielectric property data of peanuts. Our proposed method offers distinct application advantages. Unlike traditional image recognition methods that demand tiling and other intricate operations, our method allows for stacking measurement, which is notably more convenient and efficient.

In the current study, only a single peanut variety with prominent characteristics was selected. The experimental outcomes indicate the presence of certain errors, particularly in the defect rate classification model, where the overall accuracy rate stands at 0.75, leaving room for improvement. Additionally, there are substantial size and shape disparities between defective and normal pods. In the subsequent research, it is necessary to further refine the sample gradient, categorize defect types more precisely, diversify the peanut varieties under study, broaden the application scope, establish a peanut defect rate database, and develop a universal prediction and classification model for peanut defect rates.

The experimental device measurement system established in this study is relatively basic. To better apply it to defect rate detection, the measurement system should be miniaturized. Integration of the display measurement interface and miniaturization of the microwave signal source and processing circuit are feasible steps. The ultimate goal is to develop convenient and rapid online non-destructive testing equipment for peanut defect rates.

5. Conclusions

In this study, the peanut defect rate was taken as the research object, and the scattering parameters of peanuts with different defect rates were measured. A peanut defect rate prediction model and a defect rate grading model were established, respectively. The main conclusions are as follows: A peanut defect rate detection platform was established in combination with Nano VNA, and the S parameters of peanuts under different defect rates were measured. In order to accurately predict the change rule of different peanut defect rates and S parameters, XGBoost was used to establish a prediction model of peanut defect rates with different S parameters as input and defect rates as output. The results show that the determination coefficient R² of XGBoost is 0.81137, which has a certain reference value. Compared with traditional image recognition, this method can be used for stacking measurement, without the need for tiling and other operations, and the operation is convenient. It has application value for the measurement and classification of peanut defect rate.