A real-time detection device and method for cleaning loss cycle sampling of a rapeseed harvester

Abstract

This invention discloses a real-time detection device and method for cyclic sampling of cleaning losses in rapeseed harvesters. The detection device includes a sampling and detection unit installed at the tail of the cleaning screen, and an inference detection system for analyzing the data collected by the sampling and detection unit. The sampling and detection unit includes a mounting base, on which a receiving hopper, a detection channel, and a camera are mounted. It also includes a drive mechanism for driving the mounting base to reciprocate. The width of the receiving hopper is smaller than the width of the cleaning screen. A conveyor wheel is located between the receiving hopper and the detection channel, and the conveyor wheel is driven by a first motor. The detection channel is inclined. This invention drives the sampling and detection unit to reciprocate within the width range of the cleaning screen through a drive mechanism, realizing cyclic sampling and detection across the entire width range. This avoids measurement deviations caused by uneven lateral distribution of materials when sampling at fixed locations, improves the representativeness and accuracy of cleaning loss detection results, and significantly reduces the hardware cost of full-width detection while ensuring sampling coverage.

Description

Technical Field

[0001] This invention relates to the field of rapeseed cleaning and grain loss detection technology, and in particular to a real-time detection device and method for cyclic sampling of rapeseed harvester cleaning loss. Background Technology

[0002] Rapeseed is an important oilseed crop, and high harvesting losses are a prominent problem hindering the industry's development. Accurate detection of cleaning losses is not only a key indicator of machinery performance but also an important basis for operators to adjust working parameters in real time. However, currently, there is a lack of relatively accurate and rapid real-time detection methods in production. Existing harvest loss detection technologies are mainly divided into three categories, but all have significant technical shortcomings, which are explained in detail below:

[0003] First, there is the sensitive element detection method. As shown in CN108370712A, this method uses sensitive elements such as piezoelectric films or piezoelectric ceramics to collect and analyze the impact signals of the removed material. However, due to the complex and variable working conditions in the field, coupled with the small size of rapeseed grains and the fragmentation of the removed material, the impact signal characteristics are easily affected by environmental and crop conditions. The laboratory calibration conditions are difficult to cover the harsh working conditions in the actual field, resulting in low overall detection accuracy. Currently, it can only serve as a reference for the trend of loss changes and cannot achieve high-precision quantitative detection.

[0004] Second, the manual sampling and separation detection method. This method involves obtaining a portion of the waste material ejected during cleaning using a receiving device, followed by manual sieving and weighing. While this method can relatively accurately reflect actual operational losses, its sampling and separation processes heavily rely on manual assistance, making it time-consuming, labor-intensive, and inefficient in loss measurement, and completely lacking in real-time capability. In the complex and variable rapeseed harvesting scenarios in my country, this method cannot provide operators with immediate parameter adjustment feedback, making it difficult to scale up.

[0005] Third, the direct image processing detection method. This method uses a high-speed camera to photograph the material discharged from the cleaning screen and then employs a target detection algorithm to identify lost grains. However, the material flow rate and projection velocity at the tail end of the cleaning screen in a combine harvester are large. Direct in-situ imaging would typically require multiple high-cost, high-performance cameras, resulting in a complex and expensive system. More critically, the rapeseed harvesting site is extremely dusty, and small lost grains are easily obscured by impurities and other materials, severely impacting the clarity of the image acquisition and the final detection accuracy.

[0006] In summary, existing rapeseed harvester cleaning loss detection technologies have significant limitations in terms of detection accuracy, real-time performance, automation level, and ability to overcome complex field interferences (such as dust and material obstruction). Achieving high-precision, low-cost, real-time online detection of rapeseed harvester cleaning losses under harsh working conditions has become a pressing technical challenge in this field. Summary of the Invention

[0007] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides a real-time detection device and method for cyclic sampling of cleaning loss of rapeseed harvesters, which can achieve high-precision, low-cost real-time online detection of cleaning loss of rapeseed harvesters.

[0008] Technical solution: To achieve the above objectives, the present invention provides a real-time detection device for cyclic sampling of cleaning losses in a rapeseed harvester, which includes a sampling and detection unit installed at the tail of the cleaning screen, and an inference detection system for analyzing the data collected by the sampling and detection unit; the sampling and detection unit includes a mounting base, on which a receiving hopper, a detection channel and a camera are installed, the camera is capable of acquiring an image located at the bottom plate of the detection channel, and the camera has a light source for use in conjunction with it;

[0009] It also includes a drive mechanism for driving the mounting base to reciprocate, wherein the width of the receiving hopper is smaller than the width of the cleaning screen;

[0010] A feeding wheel is provided between the receiving hopper and the detection channel, and the feeding wheel is driven by a first motor; the detection channel is inclined.

[0011] During operation, the drive mechanism can drive the sampling and detection unit to move back and forth within the width range of the cleaning screen. In this way, the receiving hopper can move within the full width range to obtain the cleaning material discharged from the cleaning screen. The cleaning material entering the receiving hopper enters the detection channel via the conveyor wheel, realizing the thin layer of cleaning material and sliding down from the bottom plate of the detection channel. The camera captures the image of the bottom plate, and the inference detection system can obtain the grain loss data through the detection image.

[0012] Furthermore, the driving mechanism includes a driving screw and a driving sleeve, and also includes a second motor that drives the driving screw to rotate; the driving screw has a driving groove formed thereon, the driving groove including a forward helical groove and a reverse helical groove; the two ends of the forward helical groove and the two ends of the reverse helical groove are respectively connected; the driving sleeve is sleeved on the outside of the driving screw, and the driving sleeve has an insertion part that fits into the driving groove. In this invention, the second motor is a stepper motor.

[0013] Furthermore, the drive sleeve has a main sleeve body sleeved on the outside of the drive screw, and also includes a secondary sleeve body perpendicular to the main sleeve body. The internal spaces of the main sleeve body and the secondary sleeve body are connected. A rotating member is installed in the secondary sleeve body, and the insertion part is formed at the end of the rotating member. The insertion part is strip-shaped and can cross the intersection of the forward spiral groove and the reverse spiral groove. That is, the length of the insertion part is greater than the maximum diagonal length at the intersection, thus preventing the insertion part from getting stuck when passing through the intersection.

[0014] Furthermore, the feed wheel has a plurality of recessed grooves arranged in a circumferential array, and the extending direction of the recessed grooves is parallel to the axial direction of the feed wheel.

[0015] Furthermore, the inference detection system includes a deep learning development board and a display screen; the camera communicates with the deep learning development board through an SDK integrated development interface to transmit the acquired images to the deep learning development board in real time; the deep learning development board is equipped with an integrated target detection model to identify and count rapeseed grains in the image, and displays the cleaning loss data in real time through the display screen.

[0016] A method for real-time detection of cleaning loss cyclic sampling in rapeseed harvesters, based on the aforementioned real-time detection device for cleaning loss cyclic sampling in rapeseed harvesters, is implemented by the inference detection system and includes the following steps:

[0017] Step S1: Turn on the camera and load the target detection model; the target detection model can be a weighted model selected by the user on the host computer interface.

[0018] Step S2: Start the drive mechanism to drive the sampling and detection unit to reciprocate and circulate within the width of the cleaning sieve for sampling; and start the first motor so that the sample in the receiving hopper is quantitatively transferred to the detection channel and forms a thin layer distribution on the bottom plate of the detection channel.

[0019] Step S3: Obtain the image of the base plate position of the detection channel captured by the camera through frame-locked shooting, and call the target detection model to perform inference detection on the image, identify and count the number of lost seeds.

[0020] Furthermore, after step S3, the method further includes:

[0021] Step S4: The detection results are overlaid on the original image, that is, the identified seeds are outlined in the original image, and the total number of lost seeds in a single sampling is visualized and output in real time on the display screen.

[0022] Furthermore, the target detection model employs the YOLO v8n deep learning model.

[0023] Furthermore, in step S2, the rotation cycle of the conveying wheel is adjusted by controlling the speed of the first motor, thereby adjusting the thickness of the cleaning material laid on the bottom plate of the detection channel.

[0024] Further, in step S3, the camera continuously takes pictures of the detection channel base plate at a preset frame rate to form a continuous image stream; the inference detection system identifies the rapeseed grains in each frame image, obtains the detection box of each grain, and calculates the centroid spatial distance and regional overlap of the detection boxes of each grain between adjacent frames based on a multi-target tracking algorithm, assigning and maintaining a unique ID identifier for the same grain that appears consecutively; at the same time, a virtual detection line is drawn within the image frame, and when the trajectory of a grain with an ID identifier crosses the virtual detection line, the system includes it in the total loss of this period.

[0025] Beneficial Effects: The rapeseed harvester cleaning loss cyclic sampling real-time detection device and method of the present invention has the following beneficial effects:

[0026] (1) The detection device of the present invention drives the sampling and detection unit to reciprocate within the width range of the cleaning sieve through a drive mechanism, realizing cyclic sampling and detection within the full width range. This avoids measurement deviations caused by uneven lateral distribution of materials when sampling at fixed positions, improves the representativeness and accuracy of the cleaning loss detection results, and significantly reduces the hardware cost of full-width detection while ensuring the sampling coverage. The cleaning material obtained by the receiving hopper is evenly conveyed to the inclined detection channel by the conveying wheel, effectively spreading the cleaning material into a thin layer and sliding it down. This avoids the influence of material overlap on visual recognition, allowing the camera to obtain high-quality bottom plate images, making the real-time detection of grain loss by the inference detection system more accurate and reliable.

[0027] (2) Compared with existing loss detection methods that rely on manual sampling, sensitive element method and direct image processing method, the present invention effectively avoids complex and unfavorable working conditions such as large dust, many impurities and serious material obstruction at the cleaning outlet by using cyclic sampling detection method. It avoids the problem of image quality degradation caused by directly shooting the cleaning outlet. The sampling detection method is not affected by complex working conditions in the field and can detect loss in real time with high precision.

[0028] (3) In the structural design of the drive mechanism, by setting a drive screw with a bidirectional drive groove and a drive sleeve, the drive sleeve can first move in the forward direction under the action of the forward spiral groove. After the forward stroke ends, the insertion part transitions to the reverse spiral groove, so that the drive sleeve moves in the reverse direction again. This cycle is repeated. By using a unidirectional power source, the sampling and detection part can reciprocate to multiple points for sampling and detection. Attached Figure Description

[0029] Figure 1 Schematic diagram of the installation location of the real-time detection device for cyclic sampling of cleaning losses in rapeseed harvesters;

[0030] Figure 2 A first-view structural diagram of a real-time detection device for cyclic sampling of cleaning losses in a rapeseed harvester.

[0031] Figure 3 A second-view structural diagram of a real-time detection device for cyclic sampling of cleaning losses in rapeseed harvesters;

[0032] Figure 4 Cross-sectional structural diagram of a real-time detection device for cyclic sampling of cleaning losses in rapeseed harvesters;

[0033] Figure 5 for Figure 3 Enlarged structural diagram of section C;

[0034] Figure 6 This is a cross-sectional view of the drive mechanism;

[0035] Figure 7 This is a structural diagram of the rotating component;

[0036] Figure 8 A flowchart illustrating the real-time detection method for cyclic sampling of cleaning losses in rapeseed harvesters;

[0037] Figure 9 This is a diagram of the host computer interface displayed on the screen.

[0038] In the diagram: 1-Mounting base; 2-Receiving hopper; 3-Detection channel; 4-Camera; 5-Drive mechanism; 51-Drive screw; 51a-Forward spiral groove; 51b-Reverse spiral groove; 52-Drive sleeve; 52a-Main sleeve body; 52b-Secondary sleeve body; 52c-Rotating component; 53-Second motor; 54-Guide rod; 55-Slide seat; a-Insertion part; 6-Feeding wheel; 6a-Recessed groove; 7-First motor; 8-Inference detection system; 9-Harvester; A-Sampling and detection unit; B-Cleaning screen. Detailed Implementation

[0039] The invention will now be further described with reference to the accompanying drawings.

[0040] like Figure 1 The rapeseed harvester cleaning loss cyclic sampling real-time detection device shown is installed at the tail of the cleaning screen B in the harvester 9. It includes a sampling and detection unit A and a reasoning detection system 8 for analyzing the data collected by the sampling and detection unit A.

[0041] like Figure 2 and Figure 3 As shown, the sampling and detection unit A includes a mounting base 1, on which a receiving hopper 2, a detection channel 3, and a camera 4 are mounted. The camera 4 can acquire images located at the bottom plate of the detection channel 3, and the camera 4 has a light source for use. The inference and detection system 8 includes a deep learning development board and a display screen, such as... Figure 1As shown, the display screen is installed in the cockpit; the camera 4 is connected to the deep learning development board via the SDK integrated development interface to transmit the acquired images to the deep learning development board in real time; the deep learning development board is equipped with a target detection model to identify and count rapeseed grains in the image, and displays the cleaning loss data in real time through the display screen.

[0042] The detection device also includes a drive mechanism 5 that drives the mounting base 1 to reciprocate, and the width of the receiving hopper 2 is smaller than the width of the cleaning screen;

[0043] like Figure 4 As shown, a conveying wheel 6 is located between the receiving hopper 2 and the detection channel 3, and the conveying wheel 6 is driven by a first motor 7; the detection channel 3 is inclined. Multiple recessed grooves 6a arranged in a circular array are formed on the conveying wheel 6, and the extending direction of the recessed grooves 6a is parallel to the axial direction of the conveying wheel 6. The use of a grooved wheel structure ensures the continuity and uniformity of material conveyed from the receiving hopper 2 to the detection channel 3.

[0044] During operation, the drive mechanism 5 can drive the sampling and detection unit A to move back and forth within the width range of the cleaning screen B. In this way, the receiving hopper 2 can move within the full width range to obtain the cleaning material discharged from the cleaning screen B. The cleaning material entering the receiving hopper 2 enters the detection channel 3 via the conveyor wheel 6, causing the cleaning material to spread into a thin layer and slide down from the bottom plate of the detection channel 3. The camera 4 captures the image of the bottom plate, and the inference detection system can obtain the grain loss data by analyzing the image.

[0045] The detection device of this invention drives the sampling and detection unit A to reciprocate within the width range of the cleaning sieve B via the drive mechanism 5, achieving cyclic sampling and detection across the entire width range. This avoids measurement deviations caused by uneven lateral distribution of materials when sampling at fixed locations, improving the representativeness and accuracy of cleaning loss detection results. It also ensures sampling coverage while significantly reducing the hardware cost of full-width detection. The cleaning material acquired by the receiving hopper 2 is evenly conveyed to the inclined detection channel 3 by the conveyor wheel 6, effectively spreading the cleaning material into a thin layer before it slides down. This avoids the impact of material overlap on visual recognition, allowing the camera 4 to acquire high-quality images of the substrate, making the real-time detection of grain loss by the inference detection system more accurate and reliable.

[0046] Compared to existing loss detection methods that rely on manual sampling, sensitive element methods, and image processing methods, this invention effectively avoids complex and unfavorable working conditions such as high dust levels, numerous impurities, and severe material obstruction at the cleaning outlet by using a cyclic sampling detection method. It also avoids the image quality degradation caused by directly photographing the cleaning outlet. The sampling detection method is not affected by complex field conditions and can detect losses in real time with high precision.

[0047] like Figure 3 As shown, the drive mechanism 5 includes a drive screw 51 and a drive sleeve 52, and also includes a second motor 53 that drives the drive screw 51 to rotate; Figure 5 As shown, a driving groove is formed on the driving screw 51, the driving groove including a forward spiral groove 51a and a reverse spiral groove 51b; the two ends of the forward spiral groove 51a and the two ends of the reverse spiral groove 51b are respectively connected; the driving sleeve 52 is sleeved on the outside of the driving screw 51, and the driving sleeve 52 has an insertion part a that inserts into the driving groove, as shown. Figure 6 As shown. In this invention, the second motor 53 is a stepper motor.

[0048] In the structural design of the drive mechanism 5, by setting a drive screw 51 with a bidirectional drive groove and a drive sleeve 52, the drive sleeve 52 can first move in the forward direction under the action of the forward spiral groove 51a. After the forward stroke ends, the insertion part 52a transitions to the reverse spiral groove 51b, so that the drive sleeve 52 moves in the reverse direction. This cycle is repeated. By using a unidirectional power source, the sampling and detection part A can reciprocate to perform multi-point sampling and detection.

[0049] Preferably, the drive sleeve 52 has a main sleeve body 52a sleeved on the outside of the drive screw 51, and further includes a secondary sleeve body 52b perpendicular to the main sleeve body 52a. The internal spaces of the main sleeve body 52a and the secondary sleeve body 52b are connected. A rotating member 52c is installed inside the secondary sleeve body 52b, and the insertion part a is formed at the end of the rotating member 52c. Figure 7 As shown, the insertion part a is strip-shaped and can cross the intersection of the forward spiral groove 51a and the reverse spiral groove 51b. That is, the length of the insertion part a is greater than the maximum diagonal length at the intersection, which can prevent the insertion part a from getting stuck when passing through the intersection.

[0050] In addition, the drive mechanism 5 also includes a guide rod 54 arranged parallel to the drive screw 51, and a slide block 55 that slides with the guide rod 54 is fixed on the sampling and detection unit A.

[0051] A real-time detection method for cyclic sampling of cleaning losses in rapeseed harvesters, based on the aforementioned real-time detection device for cyclic sampling of cleaning losses in rapeseed harvesters, is implemented by the inference detection system 8, such as... Figure 8 As shown, it includes the following steps:

[0052] Step S1: Turn on the camera 4 and load the target detection model; the target detection model can be a weighted model selected by the user on the host computer interface.

[0053] Step S2: Start the drive mechanism 5 to drive the sampling and detection unit A to reciprocate and circulate within the wide range of the cleaning sieve B for sampling; and start the first motor 7 so that the sample in the receiving hopper 2 is quantitatively transferred to the detection channel 3 and forms a thin layer distribution on the bottom plate of the detection channel 3.

[0054] Step S3: Obtain the image of the base plate position of the detection channel 3 captured by the camera 4 through frame-locked shooting, and call the target detection model to perform inference detection on the image, identify and count the number of lost seeds.

[0055] During operation, the drive mechanism 5 drives the sampling and detection unit A to move back and forth at a uniform speed within the wide range of the cleaning screen B. Simultaneously, the first motor 7 operates synchronously, using the conveyor wheel 6 to promptly transfer the cleaning material entering the receiving hopper 2 to the detection channel 3, ensuring minimal material accumulation in the receiving hopper 2. The camera 4 acquires images in real time for the inference detection system 8 to perform grain identification and statistics. During a preset cycle of movement of the sampling and detection unit A, the total number of grains within that cycle is counted. For example, moving the sampling and detection unit A from one side of the cleaning screen B to the other (from left to right or from right to left) counts as one cycle. Alternatively, the sampling and detection unit A can be set to move back and forth multiple times to count as one cycle. This ensures that the total number of grains counted in each cycle covers the full width loss of the cleaning screen B, eliminating the influence of uneven distribution of the cleaning material in the width direction.

[0056] The detection method of this invention effectively solves the technical problems of uneven material distribution, large flow rate and severe dust obstruction at the cleaning outlet by combining transverse full-width reciprocating sampling, longitudinal quantitative feeding and thin-layer visual inspection. By locking frame shooting and continuous sampling inspection, the representativeness of the detection results in spatial distribution is ensured, providing high-precision real-time feedback for operators to adjust the machine parameters.

[0057] Preferably, after step S3, the method further includes:

[0058] Step S4: The detection results are overlaid on the original image, that is, the identified seeds are outlined in the original image, and the total number of lost seeds in a single sampling is visualized and output in real time on the display screen, such as... Figure 9 As shown.

[0059] Preferably, the target detection model adopts the YOLO v8n deep learning model. Based on this learning model, the detection and inference time for a single image is controlled within 40 milliseconds, which can ensure extremely high inference speed under the limited computing power of edge devices, meet the high real-time requirements of rapeseed harvesting operations, ensure that loss detection is synchronized with harvesting operations, and avoid the impact of data lag on machine parameter adjustment.

[0060] Preferably, in step S2, the rotation cycle of the conveying wheel 6 is adjusted by controlling the rotation speed of the first motor 7, thereby adjusting the thickness of the cleaning material on the bottom plate of the detection channel 3. By adjusting the rotation speed of the first motor 7, the conveying frequency can be dynamically adjusted according to the total amount of material discharged from the cleaning screen B, ensuring that the material enters the field of view of the camera 4 with a suitable thickness, reducing the missed detection error caused by material stacking, and improving the detection limit.

[0061] In step S3, the camera 4 continuously takes pictures of the detection channel 3 base plate at a preset frame rate to form a continuous image stream; the inference detection system 8 identifies rapeseed grains in each frame image based on the YOLO v8n deep learning model, obtains the detection box of each grain, and calculates the centroid spatial distance and regional overlap of the detection boxes of each grain between adjacent frames based on a multi-target tracking algorithm, assigning and maintaining a unique ID identifier for the same grain that appears consecutively; at the same time, a virtual detection line is drawn within the image frame, and when the trajectory of a grain with an ID identifier crosses the virtual detection line, the system includes it in the total loss of this period.

[0062] In the above method, since the seeds slide down in one direction and their trajectory is determined, the algorithm introduces a trajectory preservation mechanism to address the situation where seeds may be temporarily obscured by impurities during their descent. This mechanism predicts the position and maintains the state of temporarily lost seeds, seamlessly restoring their original ID when they reappear, or counting and stopping tracking seeds with that ID when their predicted trajectory crosses the virtual detection line. Because the tracking only occurs within the field of view of camera 4 and the number of seeds is relatively small, the computational cost of this method is controllable.

[0063] The algorithm mechanism of dynamic seed tracking and seed crossing the line settlement is adopted. The seed is included in the total loss of the current period at the moment the seed crosses the line. The computing power is low, which can effectively reduce duplicate counting and missed detection, and improve the accuracy of lost seed counting.

[0064] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A real-time detection device for cyclic sampling of cleaning loss in a rapeseed harvester, comprising a sampling and detection unit (A) installed at the tail of a cleaning screen (B), and a reasoning detection system (8) for analyzing the data collected by the sampling and detection unit (A); the sampling and detection unit (A) comprises a mounting base (1), on which a receiving hopper (2), a detection channel (3), and a camera (4) are mounted; characterized in that: It also includes a drive mechanism (5) for reciprocating the mounting base (1), wherein the width of the receiving hopper (2) is smaller than the width of the cleaning screen; There is a feeding wheel (6) between the receiving hopper (2) and the detection channel (3), and the feeding wheel (6) is driven by the first motor (7); the detection channel (3) is inclined.

2. The real-time detection device for cyclic sampling of rapeseed harvester cleaning loss according to claim 1, characterized in that, The drive mechanism (5) includes a drive screw (51) and a drive sleeve (52), and also includes a second motor (53) that drives the drive screw (51) to operate; a drive groove is formed on the drive screw (51), the drive groove includes a forward spiral groove (51a) and a reverse spiral groove (51b); the two ends of the forward spiral groove (51a) and the two ends of the reverse spiral groove (51b) are respectively connected; the drive sleeve (52) is sleeved on the outside of the drive screw (51), and the drive sleeve (52) has an insertion part (a) that is inserted into the drive groove.

3. The real-time detection device for cyclic sampling of cleaning losses in rapeseed harvesters according to claim 2, characterized in that, The drive sleeve (52) has a main sleeve body (52a) sleeved on the outside of the drive screw (51), and also includes a secondary sleeve body (52b) perpendicular to the main sleeve body (52a). The internal spaces of the main sleeve body (52a) and the secondary sleeve body (52b) are connected. A rotating member (52c) is installed inside the secondary sleeve body (52b), and the insertion part (a) is formed at the end of the rotating member (52c).

4. The real-time detection device for cyclic sampling of cleaning losses in rapeseed harvesters according to claim 1, characterized in that, The feed wheel (6) has a plurality of recessed grooves (6a) arranged in a circular array, and the extending direction of the recessed grooves (6a) is parallel to the axial direction of the feed wheel (6).

5. The real-time detection device for cyclic sampling of cleaning losses in rapeseed harvesters according to claim 1, characterized in that, The inference detection system (8) includes a deep learning development board and a display screen; the camera (4) is connected to the deep learning development board; the deep learning development board is equipped with a target detection model for identifying and counting rapeseed grains in the image, and displays the cleaning loss data in real time through the display screen.

6. A method for real-time detection of cleaning loss cyclic sampling in a rapeseed harvester, based on the real-time detection device for cleaning loss cyclic sampling in a rapeseed harvester according to any one of claims 1-5, characterized in that, The method is implemented by the inference detection system (8) and includes the following steps: Step S1: Turn on the camera (4) and load the target detection model; Step S2: Start the drive mechanism (5) to drive the sampling and detection unit (A) to perform reciprocating cyclic movement sampling within the wide range of the cleaning sieve (B); and start the first motor (7) so that the sample in the receiving hopper (2) is quantitatively transferred to the detection channel (3). Step S3: Obtain the image captured by the camera (4), and call the target detection model to perform inference detection on the image, identify and count the number of lost seeds.

7. The method for real-time detection of cleaning loss in rapeseed harvesters by cyclic sampling according to claim 6, characterized in that, After step S3, the method further includes: Step S4: The detection results are overlaid on the original image and the total number of lost seeds in a single sampling is visualized and output in real time on the display screen.

8. The method for real-time detection of cleaning loss in rapeseed harvesters by cyclic sampling according to claim 6, characterized in that, The target detection model uses the YOLO v8n deep learning model.

9. The method for real-time detection of cleaning loss in rapeseed harvesters by cyclic sampling according to claim 6, characterized in that, In step S2, the rotation cycle of the conveying wheel (6) is adjusted by controlling the speed of the first motor (7), thereby adjusting the thickness of the cleaning material on the bottom plate of the detection channel (3).

10. The method for real-time detection of cleaning loss cyclic sampling in rapeseed harvester according to claim 6, characterized in that, In step S3, the camera (4) takes continuous pictures of the detection channel (3) base plate at a preset frame rate to form a continuous image stream; the inference detection system (8) identifies the rapeseed grains in each frame image, obtains the detection box of each grain, and calculates the centroid spatial distance and regional overlap of the detection boxes of each grain between adjacent frames based on the multi-target tracking algorithm, assigns and maintains a unique ID for the same grain that appears continuously; at the same time, a virtual detection line is drawn in the image screen, and when the trajectory of the grain with the ID crosses the virtual detection line, the system includes it in the total loss of this period.

Images