Non-Contact Detection of Wine Grape Load Volume in Hopper During Mechanical Harvesting

Abstract

Issues of poor real-time performance and low accuracy in the detection of load volume in the hopper during the mechanized harvesting of wine grapes are addressed in this study through the development of a proposed volume detection method based on ultrasonic sensors. First, the ultrasonic sensor beamwidth and detection height were determined through calibration tests. Next, a test bench was used to explore the influence of the number of ultrasonic sensors and conveying speed on the detected grape pile height. Data-based regression and hopper configuration-based geometric models correlating grape load volume with detected pile height were subsequently constructed; their accuracies were compared using test bench experiments to identify the optimal detection scheme. The regression model was more accurate than the geometric model under the considered conveying speeds with a maximum relative error of 8.0% for the former. Finally, field tests determined that the average grape load volume detection error during actual harvesting was 14.4%. Therefore, this study provides an effective solution for the detection of grape load volume in the hopper during mechanized harvesting and establishes a theoretical basis for the development of intelligent grape harvesting methods.

1. Introduction

The wine industry is developing rapidly in China with the planted area of wine grapes increasing continuously [1,2]. Consequently, the need for mechanized harvesting is becoming urgent. However, mechanically harvested grapes are prone to damage, oxidation, and deterioration, all of which can seriously affect wine quality. Therefore, harvested grapes must be rapidly transported from vineyards to wineries [3,4]. Traditional grape harvesters rely primarily on the manual visual inspection of the load volume during harvesting and transfer, an approach that is subject to problems such as poor real-time performance and low accuracy [5]. These issues lead to situations in which the harvester’s hopper overflows or is not fully loaded and multiple transport vehicles are idle and waiting, wasting transportation resources and increasing the risk of grape deterioration.

Current methods for agricultural yield detection primarily focus on traditional crops, such as wheat and corn. Xiao [6] developed a yield detection system for corn combine harvesters using weight sensors to plot yield distribution maps, and Veal [7] developed crop mass-flow detection technology using tension sensors to improve the accuracy of the obtained yield data. However, these weight-based detection methods must be integrated into the mechanical structure of the harvesting equipment, a process that incurs high modification costs and leaves equipment susceptible to vibration interference [8,9,10].

Therefore, other researchers have studied non-contact detection methods [11,12]. Zhao [13] developed a method for measuring grain pile height based on the infrared photoelectric effect and established a fitting equation expressing the relationship between the output voltage and pile height. Fu [14] developed a grain yield metering system based on photoelectric diffuse reflection and verified its measurement accuracy through field tests. However, photoelectric sensors rely primarily on light for detection and can be easily affected by insufficient light intensity or the gloss on grape surfaces during harvest [15]. Yang [16] designed a grain yield measurement system based on machine vision, established a geometric model of grain piles, and calculated their volumes accordingly. Liu [17] used machine vision to extract the contours of forage piles and determined the angle of repose of these piles using linear fitting of their contours to construct an accumulated forage model. Although such image sensor-based systems can obtain visual information describing grapes, they are easily affected by factors such as dust and water mist, and the processing of image data requires complex algorithmic support [18,19,20].

Compared with weighing sensors, photoelectric sensors and image sensors, which have high retrofitting cost and are susceptible to vibration, light, dust, and other factors, ultrasonic sensors provide advantages such as non-contact measurement, strong environmental adaptability, and rapid detection speed [21,22,23,24,25]; they have been widely used in the agricultural field as a result [26,27,28]. However, as wine grapes are small, easily deformable, and stack in complex forms, the adaptability of the existing ultrasonic sensor detection methods to grape harvesting scenarios requires further verification. Indeed, little research has been conducted on grape load volume detection technology, with the complex stacking characteristics of grapes imposing particularly onerous requirements for detection methods.

This study addressed these issues by proposing a method for the detection of wine grape load volume in the hopper based on ultrasonic sensors. The effects of the number of sensors and conveying speed were explored using a test bench, then regression and geometric models were, respectively, constructed based on the collected grape pile height data and dimensions of the employed hopper. The accuracy of these two models were experimentally determined to identify the optimal detection scheme. Finally, the optimal scheme was applied in a field test to demonstrate its capabilities during actual wine grape harvesting.

2. Materials and Methods

2.1. Grape Load Volume Detection System

This study realized an experimental setup close to the actual wine grape harvesting scenario by applying previous research on pepper loading detection [29] to design a grape load detection test bench (Figure 1) based on the type 4P-1 wine grape harvesting system developed by the Chinese Academy of Agricultural Mechanization Sciences, Beijing, China. The test bench consisted of a conveying system and grape load detection system. The conveying system comprised a feeding device, lifting device, hopper, screw conveyor, and control cabinet. The feeding device, lifting device, and screw conveyor were driven by a servo motor (SD80AEA07530-SC3-AP-CB, Times Chaoqun Electrical Appliance Technology Co., Ltd., Beijing, China) with a power of 0.75 KW, applied transmission ratio of 1:1, and speed range of 0–3000 rpm set using the control cabinet. The screw conveyor was used to distribute the grapes evenly throughout the 1100 L hopper, which was a 1:1 copy of the hopper onboard the type 4P-1 wine grape harvester.

The loading detection system comprised LGUB1000-18GM55-I-V15 ultrasonic sensors (Luoge Intelligent Technology Co., Ltd., Leqing, China), a JY-DAM0888 data acquisition module (Beijing Juying Aoxiang Electronic Co., Ltd., Beijing, China), and a computer (Windows operating system). The detection range provided by each ultrasonic sensor was 60–1000 mm with a blind area of 0–60 mm; the analog current output was 4–20 mA and operating voltage was 10–30 V. The data acquisition module employed RS485 serial communication for eight simultaneous channels of analog input with a communication baud rate of 9600 and an operating voltage of 7–30 V. The computer was equipped with software supporting the data acquisition module for real-time acquisition and display of the current values (in μA) detected by the ultrasonic sensors.

2.2. Ultrasonic Sensor Evaluation

An ultrasonic sensor operates by using its transmitter to generate high-frequency ultrasonic pulse signals that propagate through air in the form of sound waves. When these waves encounter an object, some of them are reflected and captured by the receiver, which converts the reflected acoustic signals into electrical current signals. The time difference between signal emission and reception considering the speed of sound in air accurately indicates the distance between the sensor and target object.

An ultrasonic sensor calibration test bench was designed to determine the beamwidth range and detection height of the ultrasonic sensor employed in this study (Figure 2). This test bench consisted of an ultrasonic sensor, cantilever rod, adjustable height stand, bi-directional ruler, data acquisition module, switching power supply, computer, plumb bob, and measurement object.

2.2.1. The Method of Beamwidth Calibration

Because this study used multiple ultrasonic sensors to detect the grape load volume, potential ultrasonic cross-interference (Figure 3) could lead to erroneous signal reception and subsequent miscalculations. Therefore, the beamwidth range of a single ultrasonic sensor was determined to calibrate the sensor spacing and thereby avoid cross-interference during use.

A 25 mm cubic wooden block was used as the measurement object for beamwidth calibration. The current signal output line from the ultrasonic sensor was connected to the input port of the data acquisition module, and the data acquisition module was connected to a computer through the RS485 serial communication module. The JYDAM v 1.0 serial port debugging software, which was supplied with the data acquisition module, was run on the computer to receive and display the current signal values output by the ultrasonic sensor in real time.

The ultrasonic sensor was vertically fixed to the cantilever rod of the adjustable height stand with its detection centerline perpendicular to the ground. The wooden block was placed on flat ground, and a plumb line was dropped from the center point of the sensor detection surface to align it with the center point of the top surface of the block; this point was considered to be the origin for the subsequent tests.

Once the plumb line was removed, the current signal output by the ultrasonic sensor at the initial height was recorded and taken as the original signal value. The wooden block was subsequently moved horizontally to one side in 20 mm increments. The current signal output by the ultrasonic sensor at each block location was compared with the original current signal until a significant change in value occurred, indicating that the object was no longer detected by the sensor. The location of the center point of the block top surface when this occurred was considered the estimated beam boundary.

Next, the wooden block was moved back towards the origin along the original path in 2 mm increments until the current signal output by the ultrasonic sensor fluctuated repeatedly between values corresponding to the block and ground. The location of the center point of the block top surface at this time was set as the final beam boundary point, and the distance between the beam boundary point and origin was recorded.

The beam boundary point was determined three times at ten different vertical distances between the ultrasonic sensor and top face of the block ranging from 100 mm to 1000 mm by manually moving the cantilever rod along the stand in 100 mm increments. The average of each set of three measurements was taken as the beamwidth of the ultrasonic sensor at that distance.

2.2.2. The Method of Detection Height Calibration

The detection height of the ultrasonic sensor was calibrated using the ground as the detection object. Again, the JYDAM v 1.0 serial port debugging software was used to receive and display the current signal values output by the ultrasonic sensor in real time as its height above the ground was increased from 100 to 1000 mm in 100 mm increments. The measurements were repeated three times at each height and the actual height was determined using a tape measure. The average current signal value output by the ultrasonic sensor at each height was taken as the abscissa and the actual height of the ultrasonic sensor above the ground was taken as the ordinate to plot their relationship. Regression fitting was performed on the resulting curve to obtain the detection regression equation for the ultrasonic sensor in the following form:

$$\mathrm{y}=a{x}^b+c,$$

(1)

where $y$ is the height measured by the ultrasonic sensor in mm, $x$ is the current value output by the ultrasonic sensor in μA, $a$ and $b$ are regression coefficients, and $c$ is the height intercept in mm.

2.2.3. Accuracy of Grape Pile Height Detection

The ability of the ultrasonic sensor to meet the experimental requirements for detecting the grape pile height in the hopper under complex surface conditions was evaluated in this study using the ultrasonic sensor calibration test bench setup shown in Figure 4.

Grapes were randomly piled in a tray located immediately below the ultrasonic sensor, the sensor current signal was recorded, and a tape measure was used to measure the actual vertical distance (height) from the ultrasonic sensor to the highest point on the grape pile surface. The measured current signal value was substituted into the obtained detection regression equation for the ultrasonic sensor to calculate the detected height, then the calculated height was compared with the actual height. The vertical distance from the ultrasonic sensor to the highest point on the grape pile surface was increased from 150 to 850 mm in 100 mm increments, and the measurements were repeated three times at each value.

2.3. Grape Pile Height Tests

When the wine grape harvester is operating normally, its conveying speed is typically set equal to the driving speed in magnitude but opposite in direction to ensure that the conveying device does not scratch or damage the grape plants; these speeds typically range from 1 to 3 km·h⁻¹. Therefore, the conveying speed applied in the load volume detection test bench was also maintained within 1–3 km·h⁻¹, corresponding to servo-motor speeds between 53 and 160 rpm. Test costs were controlled by using simulated grapes as the test materials in this experiment.

The ultrasonic sensors (the total number of which was determined based on the results of the ultrasonic sensor beamwidth calibration test and size of the hopper) were evenly fixed 1000 mm above the bottom of the hopper using mounting brackets. Conveying speeds of 1, 2, and 3 km·h⁻¹ were applied as 100–800 L volumes of grapes were successively fed into the feeding device in 100 L increments. The grapes were transported to the top of the lifting device by the conveying system, where they fell into the hopper under the influence of gravity. The mounted ultrasonic sensors detected the height of the grapes as they piled in the hopper in real time. Each conveying speed test was repeated four times; three datasets were used to construct the detection dataset and the remaining dataset was used to verify the accuracy of the subsequently derived models.

2.4. Derivation of Grape Load Volume Detection Models

A regression model for grape load volume detection was derived using the detection dataset, and a geometric model for grape load volume detection was derived according to the dimensions of the hopper. Finally, the verification dataset was used to compare the errors between the volumes determined by the two constructed models and the actual values, and the model with the highest detection accuracy was identified as optimal.

2.4.1. Construction Method of Regression Model

The regression model was derived from the grape pile height detection dataset by setting the ultrasonic sensor-detected height at each grape load volume as the abscissa and the corresponding actual grape load volume as the ordinate to plot a correlation curve illustrating the relationship between pile height and volume. A regression analysis was subsequently applied to fit these data, ultimately establishing a regression model on the basis of the following:

$$V_1=\alpha H^{\beta }+\gamma ,$$

(2)

$$H=h-y,$$

(3)

where $V_1$ is the volume of grapes in L, $H$ is the average sensor-measured grape pile height in mm, $h$ is the distance between the sensor and bottom of the hopper (1000 mm in this study), $y$ is the sensor detection distance in mm, $\alpha$ and $\beta$ are regression coefficients, and $\gamma$ is the volume intercept in L.

2.4.2. Construction Method of Geometric Model

The geometric model for grape load volume detection was constructed by dividing the hopper into upper and lower trapezoidal prisms based on the structural characteristics of its cross section; these prisms are separated by the red line in Figure 5. Thus, the grape load volume can be expressed as

$$V_2=\left\{\begin{array}{ll}{\displaystyle \frac{W}{2}}\left[\left(tan{\theta }_1+tan{\theta }_2\right)H^2+2LH\right]& (0<H\le B)\\ {\displaystyle \frac{W}{2}} \left[\left(tan{\theta }_3+tan{\theta }_4\right){\left(H-B\right)}^2+2A\left(H-B\right)+\left(tan{\theta }_1+tan{\theta }_2\right)B^2+2LB\right]& (H>B)\end{array}\right.,$$

(4)

where $V_2$ is the grape load volume according to pile height in mL, $H$ is the average sensor-measured grape pile height (comprising H₁, H₂, and H₃ in Figure 5) in mm, $L$ is the length of the hopper in mm, $W$ is the width of the hopper in mm, $B$ is the vertical distance from the dividing line to the bottom of the hopper in mm, $A$ is the length of the red line dividing the prisms in mm, and ${\theta }_1$, ${\theta }_2$, ${\theta }_3$, and ${\theta }_4$ are the various angles between the inclined sides of the hopper relative to the vertical direction.

2.5. Field Validation

The effectiveness of the optimal model during actual mechanical grape harvesting was validated using a field experiment conducted at the Pulagu Wine Industrial Park in Laishan District, Yantai City, Shandong Province on 20 September 2023. This experiment used a wine grape harvester comprising a ZN-1154CH orchard multi-functional chassis and type 4P-1 wine grape harvesting system (Chinese Academy of Agricultural Mechanization Sciences, Beijing, China), shown in Figure 6. The picking height was set to 200–2000 mm, one working row was used, the operating speed was 1–3 km·h⁻¹, and the grape variety was Cabernet Sauvignon.

Before the field validation experiment, three ultrasonic sensors were evenly installed directly above the harvester hopper with the detection direction set perpendicular to the bottom surface of the hopper. During the harvest process, the current value of the output of the ultrasonic sensor when the grapes are continuously transferred to the hopper is randomly recorded. The actual volume of the grapes in the hopper at these measurement points was manually determined by stopping the equipment after sensor measurement and removing the grapes. After manual measurement, the grapes were poured back into the hopper. Each measurement was repeated three times. After the experiment was completed, the collected detection data were filtered and substituted into the optimal detection model identified using the bench tests. Finally, the calculated grape load volumes were compared with the actual volumes to determine the detection error.

3. Results and Discussion

3.1. Ultrasonic Sensor Calibration

3.1.1. The Result of Beamwidth Calibration

The beamwidths of the ultrasonic sensor were obtained at different heights using the beamwidth calibration test described in Section 2.2.1; the results are shown in Figure 7. When the distance between the ultrasonic sensor and measured object increased from 100 to 800 mm, the beamwidth slowly increased from 160 to 396 mm; when this distance increased from 800 to 1000 mm, the beamwidth decreased rapidly from 396 to 106 mm. Thus, the results indicated a maximum beamwidth of 396 mm at a height of 800 mm. Considering the length of the bottom surface of the hopper (1230 mm), this result indicates that a maximum of three ultrasonic sensors could be installed in the grape load volume detection system. If the detection method is used in the wine grape harvester, that is, the geometry and size of the hopper are different, the number of ultrasonic sensors installed needs to be re-determined.

3.1.2. The Result of Detection Height Calibration

The ultrasonic sensor detection height was calibrated as described in Section 2.2.2 with the results shown in Figure 8.

The coefficient of determination (R²) of the linear regression equation shown in Figure 8 was 0.9999, indicating a strong linear correlation between the current signal output and applied detection height.

3.1.3. Accuracy of Ultrasonic Sensor-Detected Grape Pile Height

The accuracy of grape pile height detection using the ultrasonic sensor was evaluated as described in Section 2.2.3 with the results shown in

Table 1. Accuracy of ultrasonic sensor-detected grape pile height.

Actual Height (mm)	Detected Height (mm)	Absolute Error (mm)	Relative Error (%)
150	143.1	6.9	4.6
250	246.3	3.7	1.5
350	348.9	1.1	0.3
450	448.5	1.5	0.3
550	551.3	1.3	0.2
650	663.8	13.8	2.1
750	777.1	27.1	3.6
850	874.9	24.9	2.9

. Both the absolute error and relative error of the grape pile height measured by the ultrasonic sensor initially decreased, then increased with increasing height. The maximum relative error was only 4.6%.

The observed change in accuracy with height can be attributed to interference effects caused by strong ultrasonic echoes off grapes at varying heights and positions during the close-range detection of piled grape surfaces. Thus, although the echo intensity gradually diminished as the height increased, the grape pile surface became more complex, introducing additional interference into the ultrasonic reflections and increasing the error.

3.2. Results of Grape Pile Height Tests

The grape pile heights were measured at three different conveying speeds to obtain the ultrasonic sensor detection values under different load volumes as described in Section 2.3; the results are shown in

Table 2. Grape pile height test results.

Conveyor Speed (km·h−1)	Supplied Grape Load Volume (L)	Grape Pile Height (mm)
		Sensor-1	Sensor-2	Sensor-3
1	100	183.6	93.9	37.4
	200	347.1	195.0	120.7
	300	421.5	296.0	241.5
	400	548.4	402.3	317.9
	500	589.7	480.3	394.1
	600	705.1	579.9	487.4
	700	779.7	642.8	576.6
	800	873.6	732.3	659.6
2	100	188.0	90.5	31.7
	200	331.7	205.2	120.5
	300	426.9	293.4	240.9
	400	559.7	384.4	318.9
	500	600.3	465.3	394.7
	600	714.6	564.7	483.7
	700	784.8	638.3	570.6
	800	868.0	729.7	676.8
3	100	187.4	89.1	31.0
	200	351.1	188.2	120.8
	300	426.0	283.7	241.6
	400	555.4	389.6	317.1
	500	593.8	474.5	390.5
	600	711.2	571.4	484.7
	700	792.6	646.5	574.9
	800	876.4	739.0	678.4

. Regardless of which of the three ultrasonic sensor was used, the detected height increased with the grape load volume given a constant conveying speed. Furthermore, the heights detected by the three ultrasonic sensors did not change significantly with the conveying speed given a constant grape load volume because the change in conveying speed was insufficient to shift the point from which the grapes fell into the hopper. However, the heights detected by the ultrasonic sensors decreased as their distance from the lifting device increased, indicating that the screw conveyor did not significantly disrupt the grape piling point; only when the hopper was nearly full (i.e., once the grape pile surface contacted the screw conveyor) did the screw conveyor evenly spread the pile.

These results indicate that if only one ultrasonic sensor is used, the obtained measurement data can only reflect the grape pile height at a specific location; the inherent randomness with which the grapes fall into the hopper limits the utility of such data and prevents a comprehensive understanding of the grape pile state within the entire hopper. This makes the overall grape pile height easy to misjudge when relying on the local grape accumulation.

Although the use of two ultrasonic sensors could provide more information describing the grape pile height distribution, detection blind spots could still exist, and some irregular accumulation patterns might not be captured accurately.

The beamwidth test results indicated that a maximum of three sensors could be installed above the hopper. Therefore, three ultrasonic sensors were supplied above the hopper to provide more comprehensive detection and accurately reflect the grape pile heights at different areas within the hopper. This can reduce the occurrence of errors caused by local differences in grape accumulation and improve the integrity and accuracy of the collected data.

3.3. Evaluation of Derived Models

3.3.1. The Construction Result of the Regression Model

The average accumulated height simultaneously detected by the three ultrasonic sensors at each conveying speed was correlated with the grape load volume through data analysis and fitting of the regression model in Equation (2) as follows (Figure 9):

$$V_1=0.0003H^2+0.8624H+2.4994.$$

(5)

This regression model yielded a coefficient of determination (R²) of 0.9985, indicating an exceptionally high degree of fit between the model and experimental data. These results demonstrate the excellent ability of the regression model to characterize the relationship between the grape load volume and average pile height detected by the ultrasonic sensors.

3.3.2. The Construction Result of Geometric Model

The following geometric model for grape load volume was obtained by inserting the measured hopper L of 1230 mm, $W$ of 650 mm, $B$ of 130 mm, $A$ of 1543 mm, and ${\theta }_1$, ${\theta }_2$, ${\theta }_3$, and ${\theta }_4$ of 18.27°, 64.16°, 26.59°, and 18.27°, respectively, into Equation (4) as follows:

$$V_2=\left\{\begin{array}{c}778.4H^2+799500H (0<H<130 \mathrm{mm})\hfill \\ 269.9H^2+932800H-8.761\times {10}^6 (130<H<1130 \mathrm{mm})\hfill \end{array}\right..$$

(6)

The geometric model given by Equation (6) provides an alternative solution for grape load volume determination. However, because this model is based on the dimensions of the hopper and does not account for the random distribution of grapes as they are piled within the hopper, the determined volume may deviate from the actual condition. Therefore, the regression and geometric models were compared to select the model that best aligns with the practical requirements for grape load volume measurement.

3.3.3. Comparison of Regression and Geometric Model Results

The verification data were substituted into the regression and geometric models, and the errors between the volumes output by each model and the actual volumes were compared to evaluate their accuracies; the results are shown in

Table 3. Comparison of model validation results.

Conveyor Speed (km·h−1)	Actual Load Volume (L)	Regression Model		Geometric Model
		Detected Load Volume (L)	Relative Error (%)	Detected Load Volume (L)	Relative Error (%)
1	100	92.0	8.0	88.0	12.0
	200	209.9	4.9	212.9	6.4
	300	307.9	2.6	316.0	5.3
	400	423.1	5.7	436.3	9.0
	500	497.7	0.4	513.7	2.7
	600	618.8	3.1	638.7	6.4
	700	717.0	2.4	739.4	5.6
	800	843.5	5.4	868.6	8.5
2	100	93.2	6.8	88.8	11.2
	200	210.6	5.3	213.6	6.8
	300	310.2	3.4	318.5	6.1
	400	418.3	4.5	431.3	7.8
	500	493.4	1.3	509.3	1.8
	600	613.3	2.2	633.0	5.5
	700	704.9	0.6	727.0	3.8
	800	829.2	3.6	854.0	6.7
3	100	94.9	5.0	90.6	9.4
	200	201.3	0.6	203.7	1.8
	300	304.6	1.5	312.5	4.1
	400	420.2	5.0	433.3	8.3
	500	495.3	0.9	511.2	2.2
	600	615.4	2.5	635.2	5.8
	700	719.3	2.7	741.8	5.9
	800	834.5	4.3	859.4	7.4

. When the conveying speed was 1 km·h⁻¹, the relative errors of the regression model were between 0.4% and 8.0%, whereas those of the geometric model were between 2.7% and 12.0%. When the conveying speed was 2 km·h⁻¹, the relative errors of the regression model were between 0.6% and 6.8%, whereas those of the geometric model were between 1.8% and 11.2%. When the conveying speed was 3 km·h⁻¹, the relative errors of the regression model were between 0.5% and 5.4%, whereas those of the geometric model were between 1.8% and 9.4%. Thus, the maximum relative errors of the regression and geometric models were determined to be 8.0% and 12.0%, respectively.

Notably, the regression model was constructed using actual experimental data describing the grape pile height and volume and as such fully considered the actual situation of the grapes as they piled in the hopper, including the influence of factors such as the grape pile form and falling point. Clearly, the large quantity of measurement data allowed the regression model to accurately reflect the relationship between the grape pile height and load volume, resulting in a relatively small error.

In contrast, the geometric model was constructed considering only the configuration of the hopper and could not account for the randomness of the grape piling process, neglecting the uneven distribution of grapes in actual operation. This led to relatively large detection errors when using this model. Thus, the comparison results indicated that the regression model provided more accurate grape load volume detection than the geometric model regardless of the applied conveying speed or grape load volume.

Although the regression model takes into account the actual situation of the accumulation of grapes in the hopper, there will still be the situation that the accumulation of grapes in the corner cannot be detected (the accumulation of grapes in the detection blind area of three ultrasonic sensors), which is also one of the reasons for the error of the regression model. The reason for the blind spot of ultrasonic detection is that the cross section of ultrasonic detection beam is round. In order to avoid cross interference between ultrasonic waves, there will be blind spots between adjacent ultrasonic sensors. And the three ultrasonic sensors cannot completely cover the entire hopper, ultrasonic sensors and the edge of the hopper will also have detection blind areas. Therefore, the use of three ultrasonic sensors maximizes the detection range covering the hopper and averages it to reduce the impact of extreme grape accumulation.

3.4. Field Test Results

Since the vibration generated by the harvester in the field will affect the detection of the ultrasonic sensor, the detection frequency of the ultrasonic sensor is set to 0.1 s, and the detection period is 10 times, and the influence of the vibration is reduced by calculating the moving average, as shown in Figure 10.

It can be intuitively seen from Figure 10 that the stack height data detected by the ultrasonic sensor during field operation of the harvester has significant high-frequency fluctuations due to mechanical vibration interference, and the original detection value may show an instantaneous jump of tens of millimeters at a sampling interval of 0.1 s. The moving average method is used to smooth the detection results every 10 times, and the results show that this kind of noise can be effectively suppressed.

The accuracy of the regression-based grape load volume detection model during actual harvesting operations was verified by substituting the current values output by the ultrasonic sensors during the field experiments into Equation (5) to obtain the detected grape load volumes reported in

Table 4. Field test results.

Detected Pile Height (mm)	Actual Load Volume (L)	Detected Load Volume (L)	Absolute Error (L)	Average Error (%)
49.0	53	45.5	6.5	14.4
56.1	60	51.8	8.2
83.1	92	76.2	15.8

. These results indicate that the regression model for grape load volume detection achieved an average error of 14.4% under actual grape harvesting conditions.

Various factors contributed to this error under the harvesting conditions found in a vineyard. First, the detection accuracy of an ultrasonic sensor can be compromised by environmental interference factors such as dust particles or water mist aerosols. Second, the dynamic grape cluster piling morphology continuously evolves with the movement of the harvester, resulting in significant disparities compared with the morphology realized by the controlled indoor experiment setup. Finally, the vibrations induced by changes in the operating speeds of the harvesting machinery have different impacts on the stability of the ultrasonic sensor outputs.

4. Conclusions

This study first conducted beamwidth and detection height calibration of an ultrasonic sensor to determine that three sensors could be used for grape pile height detection in the considered hopper. The grape load volume detection accuracy provided by three ultrasonic sensors was subsequently verified through laboratory tests indicating a maximum error of 4.6%, meeting the requirements for the subsequent grape load volume detection models. Notably, the results of the grape pile height detection test confirmed that simultaneous detection using three sensors more accurately reflected the state of the grape pile in the hopper and indicated that the conveying speed had a relatively small impact on the detection accuracy. Next, regression and geometric models correlating the detected pile height to the grape load volume were, respectively, constructed based on the collected height measurement data and configuration of the hopper. Experimental results indicated that the accuracy of the regression model was superior to that of the geometric model with a maximum error of 8.0% for the former. Finally, field tests indicated that the average detection error provided by the regression model using data from three ultrasonic sensors was 14.4%. Therefore, the proposed grape load volume measurement method provides a promising basis for the development of intelligent grape harvesting methods. The applicability and accuracy of the proposed detection method will be improved by considering the influence of environmental factors and driving speed in future studies and combine visual or liDAR data fusion to improve the robustness of the model.