Harvester cleaning flow field analysis method based on target detection and rule reasoning

Abstract

The invention belongs to the technical field of combine harvester cleaning flow field analysis, and particularly relates to a harvester cleaning flow field analysis method based on target detection and rule reasoning. A combined harvester cleaning chamber three-dimensional model is constructed for fluid simulation analysis, a cleaning chamber flow field key section scatter diagram is drawn by using simulation result data, and YOLOv8 based on a deep learning target detection algorithm is selected as an identification algorithm of key points of the cleaning chamber flow field key section scatter diagram. And through the cleaning chamber flow field key points and a rule-based reasoning mechanism, cleaning chamber flow field key region division, region internal reasonable wind speed analysis and cleaning effect evaluation are completed. According to the method, under the condition that the types of harvested crops and the main parameters of the harvester cleaning room are selected, the cleaning room flow field key area analysis result is obtained through the knowledge base and the reasoning method, and the problems that after simulation analysis, processing depends on experience, efficiency is limited, and labor cost is high can be solved.

Description

Technical Field

[0001] The invention belongs to the technical field of combine harvester cleaning flow field analysis, and in particular relates to a combine harvester cleaning flow field analysis method based on target detection and rule reasoning. Background Art

[0002] The cleaning device is the main working part of the combine harvester. Its main function is to use the different movement characteristics of crops under the same wind field to realize the cleaning, classification and grading of crops. Its cleaning effect directly affects the performance of the whole machine. Therefore, when designing the parts of the cleaning device, the flow field simulation of the cleaning chamber has always been a research hotspot in this field. The analysis and judgment of the simulation data results are the basis for improving the design, so it is very important to analyze the simulation data reasonably and accurately in combination with the working status of the device. The efficiency of simulation pre-processing and solvers is constantly improving, but in the conventional simulation post-processing part, it still mainly depends on the actual engineering experience of simulation engineers.

[0003] Although the efficiency of simulation pre-processing and solvers in the existing technology is constantly improving, in the conventional simulation post-processing part, simulation post-processing data analysis is inseparable from the experience of engineers. Engineer experience is the result of long-term cultivation and does not currently meet the conditions for trading, circulation or storage transfer. This experience, as "tacit knowledge", is difficult to express and disseminate. When employees leave or change jobs, the post-processing analysis experience is also lost. In the absence of this kind of "experience", it will be difficult for the successor to the position to take over perfectly. In the context of the information and digital era, research on algorithm-based intelligent analysis methods for simulation data has been carried out to make automatic analysis of simulation post-processing possible.

[0004] The study shows that the simulation data of the flow field in the cleaning chamber of the combine harvester is used as the research object to explore the methods of identifying the key points of the upper and lower screens and dividing the key areas of the wind field without the assistance of engineers; combined with the wind field type, wind field intensity judgment algorithm and rule-based or instance-based reasoning mechanism, a flow field analysis method for the cleaning chamber of the combine harvester based on target detection algorithm and reasoning mechanism is proposed to realize the flow field analysis of the cleaning chamber of the combine harvester. To a certain extent, it alleviates the problems that post-processing experience cannot be efficiently passed on and post-processing efficiency is limited by labor costs. Summary of the invention

[0005] The purpose of the present invention is to provide a harvester cleaning flow field analysis method based on target detection and rule reasoning, so as to alleviate the problems that post-processing experience cannot be efficiently inherited and post-processing efficiency is limited by labor costs, so as to realize intelligent analysis of post-processing simulation in the cleaning room of a combine harvester.

[0006] In order to achieve the above-mentioned invention purpose and solve the problems existing in the prior art such as the difficulty in expressing and disseminating post-processing experience, the present invention proposes a harvester cleaning flow field analysis method based on target detection and rule reasoning. The specific steps are as follows:

[0007] S1 obtains the harvester cleaning flow field simulation result data;

[0008] S2 draws a scatter plot of the key cross-section of the flow field in the cleaning chamber based on the Python data analysis and visualization tool library and simulation result data to construct a training data set;

[0009] S3 selects YOLOv8 as the recognition model for key points of the scatter plot of the key cross-section of the flow field in the cleaning chamber, and uses the training data set to train and optimize the recognition model to obtain the key point recognition model;

[0010] S4 combines the key points of the key cross-section scatter diagram of the cleaning chamber flow field identified by the key point recognition model, and divides the harvester cleaning chamber into regions through division rules;

[0011] S5 analyzes the wind field intensity and wind field type in the divided area based on rule reasoning to obtain the post-processing conclusion reasoning result.

[0012] Furthermore, in the above step S1, multiple models of harvester cleaning chamber models are preferably selected, and model parameters are adjusted, wherein the parameters include: screen type, model size, and distance between upper and lower screens; SOLIDWORKS is used to establish a three-dimensional model of the harvester cleaning chamber, and the flow field of the cleaning chamber is simulated and analyzed through the fluid simulation software ANSYS FLUENT to obtain simulation result data.

[0013] Furthermore, the above step S2 includes:

[0014] S2.1 uses the Python data analysis and visualization tool library to filter out key cross-section data from the simulation result data, and processes the filtered key cross-section data into a key cross-section scatter plot to clearly show the upper and lower screen positions and lengths;

[0015] S2.2 By setting the thickness of the key section and controlling the sparsity of the scatter plot, the key section scatter plots under different slice thicknesses are obtained. At the same time, the position coordinates of the key points can be determined according to the sparsity of the points and the existence of the points;

[0016] S2.3 performs data enhancement on the key cross-section scatter plot, including random scaling, random offset, horizontal mirroring, vertical mirroring, and random rotation of the scatter plot to obtain a training data set.

[0017] Furthermore, the above step S3 includes:

[0018] S3.1 Label the training data set. Drag the mouse in labelme to get a key point labeling box. Then add classification labels to the boxed area, and then mark the area where the key points are located in the box.

[0019] S3.2 randomly divide the labeled training data set into a training set and a validation set in a ratio of 8:2, wherein the training set is used to train the target detection model, and the validation set is used to verify and test the key point recognition model;

[0020] S3.3 Configure the experimental environment, train the key point recognition model with the training set, and verify the trained key point recognition model with the verification set;

[0021] S3.4 uses intersection over union (IoU), recall (R), precision (P), average precision (AP) and mean average precision (mAP) as evaluation indicators of the key point recognition model, evaluates the key point recognition performance through the evaluation indicators, and analyzes the convergence of the key point recognition model through the convergence curve of the evaluation indicators and the loss function curve of the model;

[0022] S3.5 uses the gradient descent optimizer to adjust relevant parameters to optimize the key point recognition model and obtain a key point recognition model with better performance.

[0023] Furthermore, the division of the cleaning room area in the above step S4 includes:

[0024] S4.1 predict the key point coordinates using the key point recognition model, and determine the region division rules based on the prediction results;

[0025] S4.2 Use the regional division rules to divide the harvester cleaning room into areas, including:

[0026] Area ① is located above the upper screen surface of the front section, with a length of 30% of the total length of the upper screen, and the left side of the area is aligned with the left side of the upper screen surface;

[0027] Area ② is located above the middle upper screen surface, and its length accounts for 60% of the total length of the upper screen. The left side of the area is adjacent to the right side of area ①.

[0028] Area ③ is located above the upper screen surface of the tail section, and its length accounts for 10% of the total length of the upper screen, and the right side of the area is aligned with the right side of the upper screen surface;

[0029] Area ④ is located below the lower screen surface, with the same length as the lower screen surface, and the left side of the area is aligned with the left side of the lower screen surface;

[0030] Area ⑤ is located directly above Area ①, with its left side aligned with the leftmost end of the cleaning room, and its right side aligned with the right side of Area ①;

[0031] Area ⑥ is located below area ④, with its left side aligned with the left side of area ② and its right side aligned with the right side of area ②.

[0032] Furthermore, the rule reasoning in the above step S5 includes:

[0033] S5.1 Obtain the minimum floating velocity of grains according to the knowledge of suspension velocity parameters of rice and wheat components stored in the post-processing knowledge base;

[0034] S5.2 compare the maximum wind speed in the divided area with the minimum grain grain floating speed;

[0035] S5.3 Analyze the wind farm intensity and wind farm type based on the wind farm type judgment rule knowledge, wind farm intensity judgment rule knowledge, and unreasonable wind condition modification suggestion knowledge stored in the post-processing knowledge base.

[0036] S5.4 Determine whether the wind speed is reasonable. If it is unreasonable, analyze the reason. Determine whether there is an eddy current. If there is an eddy current, analyze the strength of the eddy current. Determine whether the eddy current has an adverse effect. If so, analyze the adverse effect to obtain the grain cleaning reasoning result.

[0037] Furthermore, the above S4.1 includes:

[0038] S4.1.1 Load the key section cross-section scatter plot and pre-process the input data;

[0039] The backbone network in the S4.1.2YOLOv8 model extracts key cross-section scatter plot features and locates and classifies key points through the detection head;

[0040] S4.1.3 The YOLOv8 model performs non-maximum suppression (NMS) on prediction boxes with a high degree of overlap, removes duplicate prediction boxes, and retains the best prediction results;

[0041] S4.1.4 There are four prediction results of the key point coordinates in the key section scatter diagram, which are located at the front end of the upper screen surface, the rear end of the upper screen surface, the front end of the lower screen surface, and the rear end of the lower screen surface in the cleaning chamber; the area division rules based on key points are determined by analyzing the fluid simulation result data in the cleaning chamber and the distribution of airflow in the cleaning chamber.

[0042] The intentional effects of the present invention are as follows:

[0043] The present invention combines the wind field type, wind field intensity determination algorithm and rule-based reasoning mechanism to realize the airflow analysis and cleaning effect analysis in the cleaning room and generate a result report. It alleviates the problems of slow simulation post-processing speed, high manual analysis cost and the fact that simulation post-processing data analysis cannot be separated from the simulation engineer's own experience, and has greater economy and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The figure is a flow chart of the method of the present invention.

[0045] Figure 2 Scatter plot of key cross sections of the cleaning chamber at different slice thicknesses.

[0046] Figure 3 This is a schematic diagram of the key point locations of the scatter plot of key sections in the cleaning room.

[0047] Figure 4 This is a schematic diagram of the area division location.

[0048] Figure 5 Reasoning flow chart for the division of flow field detection areas in the cleaning chamber.

[0049] Figure 6 Flow chart for flow field analysis and reasoning.

[0050] Figure 7 A flow chart for the conclusion reasoning of the post-processing process of the cleaning flow field. DETAILED DESCRIPTION

[0051] The present invention provides a harvester cleaning flow field analysis method based on target detection and rule reasoning. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0052] like Figure 1 As shown, the present invention is a harvester cleaning flow field analysis method based on target detection and rule reasoning, comprising the following steps:

[0053] S1. Obtaining harvester cleaning flow field simulation result data;

[0054] S2, based on the Python data analysis and visualization tool library and simulation result data, draw the scatter plot of the key cross-section of the flow field in the cleaning chamber and construct a training data set;

[0055] S3, select YOLOv8 as the recognition model of key points of the scatter plot of the key cross-section of the flow field in the cleaning chamber, use the training data set to train and optimize the recognition model, and obtain the key point recognition model;

[0056] S4, combining the key points of the key cross-section scatter diagram of the cleaning chamber flow field identified by the key point recognition model, and dividing the harvester cleaning chamber into regions according to the division rules;

[0057] S5. Analyze the wind field intensity and wind field type in the divided area based on rule reasoning to obtain the post-processing conclusion reasoning result.

[0058] As a preferred embodiment of the present invention, the present invention preferably takes the existing 12 types of cleaning chamber models, including 4LBZ-145, 4LZ-5G, 4LZY-1.5, DFQX-3, ZKB-5, 4LZ-7N, Fengshou-3.0, 4LZB-105, 4LZ-2.5, 4LZ-1.0, 4LQ-2.5, and DF-1.5, as the basis, and uses SOLIDWORKS to establish a three-dimensional model of a cleaning chamber of a combine harvester; and adjusts the parameters of the 12 models, and the parameters include: screen type, model size, and distance between upper and lower screens.

[0059] As a specific embodiment of the present invention, parameter adjustment specifically includes:

[0060] 1. To expand the screen types, half of the model lower screens were replaced from fish scale screens to mesh screens;

[0061] 2. Adjust the size, enlarge or reduce the original model by different proportions, so that the length range of the adjusted model is 800-1500mm, the width range is 500-1000mm, and the height range is 600-1000mm, and re-model the cleaning screen according to the changed cleaning chamber structural parameters;

[0062] 3. Adjust the distance between the upper and lower screens, and shift the upper and lower screens upward and downward from their original positions by a certain distance, with a shift range of 0 to 150 mm, to form the final model. Through the above process of expanding the number of simulation models, a total of 216 models were finally established, which were stored in 12 groups, which can enhance the practicality and reliability of the recognition model.

[0063] The fluid simulation software ANSYS FLUENT is used to simulate and analyze the flow field in the cleaning chamber to obtain simulation result data.

[0064] As a specific embodiment of the present invention, the specific steps of the cleaning chamber flow field simulation test are as follows:

[0065] 1. Preprocessing of geometric models: In the modeling process, in order to ensure the smooth operation of the computer in the simulation stage, a 1 / 2 model is established to reduce the subsequent calculation tasks. All the faces in the model are divided into three categories according to their functions: outlet, inlet, and cleaning chamber wall.

[0066] 2. Computational domain meshing: Open ANSYS Fluent, select the Meshing module, and select "Watertight Geometry" as the workflow mode; after importing the established cleaning chamber geometry model, select "Geometry type consists only of fluid areas without voids" as the description geometry type, select polyhedra as the filler mesh type, and set the maximum unit length to 10 mm.

[0067] 3. Solver settings: During the cleaning process, the gas density inside the cleaning chamber remains basically unchanged and can be regarded as an incompressible fluid, so the pressure-based solver is selected; the physical model selects the RNG k-ε turbulence model, and the wall function is the standard wall function SWF; the inlet of the cleaning chamber is set as the velocity inlet, the wind speed is set to 6-8m / s, and the turbulence is set to K and Epsilon; the outlet of the cleaning chamber is set as the pressure outlet, and the outlet pressure is 0Pa; the SIMPLEC algorithm is selected as the solution method, and the calculation is stopped after the custom iteration reaches 500 steps after the standard initialization.

[0068] Exporting the simulation result data of the flow field in the cleaning chamber provides a basis for the subsequent wind speed evaluation in the divided area.

[0069] As a preferred embodiment of the present invention, step S2 specifically includes:

[0070] S2.1 Based on the Python data analysis and visualization tool library, the key cross-section data are screened out from the simulation result data, and the screened key cross-section data are processed into key cross-section scatter plot data, that is, the key cross-section scatter plot of the cleaning chamber flow field is drawn using the simulation result data to clearly show the position and length of the upper and lower screens;

[0071] S2.2 controls the sparsity of the scatter plot by setting the thickness of the key section, and obtains the key section scatter plot under different slice thicknesses. At the same time, the position coordinates of the key points can be determined according to the sparsity of the points and the existence of the points.

[0072] The sparseness of the scatter plot can be controlled by changing the program parameters to set the critical section thickness, such as Figure 2 This is the result of drawing the scatter plot of key cross sections under different slice thicknesses. It can be seen that within the range of 2 to 100 mm, the slices can clearly show the basic characteristics such as the length and position of the upper and lower screens.

[0073] When the cleaning chamber network is divided, the surface size of the screen or screen mesh is quite different from that of the box body, so the mesh size near the screen or screen mesh will be significantly smaller than other areas. Therefore, the front and rear ends of the upper screen and the front and rear ends of the lower screen can be distinguished from the scatter plot of the key cross-section of the cleaning chamber according to the density of the scatter plot. This feature of the flow field node distribution provides a basis for key point identification. The key point position is as follows: Figure 3 shown.

[0074] S2.3 performs data enhancement on the key cross-section scatter plot, including random scaling, random offset, horizontal mirroring, vertical mirroring, and random rotation of the scatter plot to obtain a training data set.

[0075] As a preferred embodiment of the present invention, step S3 specifically includes:

[0076] S3.1 Label the training data set. Drag the mouse in labelme to get a key point labeling box. Then add classification labels to the boxed area, and then mark the area where the key points are located in the box.

[0077] S3.2 randomly divide the labeled training data set into a training set and a validation set in a ratio of 8:2, wherein the training set is used to train the target detection model, and the validation set is used to verify and test the key point recognition model;

[0078] S3.3 Configure the experimental environment, train the key point recognition model with the training set, and verify the trained key point recognition model with the verification set;

[0079] As a specific embodiment of the present invention, the model input size can be 640×640 pixels, the training batch (batch_size) can be set to 16, the initial learning rate can be set to 0.01, the stochastic gradient descent (SGD) method is used to optimize the model parameters, and the number of iterations can be 150.

[0080] S3.4 uses intersection over union (IoU), recall (R), precision (P), average precision (AP) and mean average precision (mAP) as evaluation indicators of the key point recognition model, evaluates the key point recognition performance through the evaluation indicators, and analyzes the convergence of the key point recognition model through the convergence curve of the evaluation indicators and the loss function curve of the model;

[0081] S3.5 uses the gradient descent optimizer to adjust relevant parameters to optimize the key point recognition model and obtain a key point recognition model with better performance.

[0082] As a preferred embodiment of the present invention, step S4 specifically includes:

[0083] S4.1 predict the key point coordinates using the key point recognition model, and determine the region division rules based on the prediction results;

[0084] First, the key section cross-section scatter plot is loaded, and the input data is preprocessed by resizing. Second, the backbone network in the YOLOv8 model extracts the key section scatter plot features, and the key points are located and classified through the detection head. Finally, the YOLOv8 model performs non-maximum suppression (NMS) on the prediction boxes with a high degree of overlap, removes duplicate prediction boxes, and retains the best prediction results.

[0085] There are four prediction results of the key point coordinates in the key section scatter plot, which are located at the inner end of the upper screen surface, the rear end of the upper screen surface, the inner end of the lower screen surface, and the rear end of the lower screen surface in the cleaning chamber. By referring to the fluid simulation result data of the cleaning chamber, the experience knowledge and literature of the screen surface airflow analysis, the distribution of the airflow in the cleaning chamber is analyzed, and the regional division rules based on key points are determined.

[0086] S4.2 Use the regional division rules to divide the harvester cleaning room into areas, including:

[0087] According to the key points of the key cross-section scatter diagram of the cleaning chamber flow field identified by the YOLOv8 model, the harvester cleaning chamber is divided into regions ①, ②, ③, ④, ⑤, and ⑥ using the regional division rule. Figure 4 As shown in the figure; combined with the key point recognition model's detection results of the key point coordinates, the division reasoning process of the six division areas of the sorting room is as follows Figure 5 shown.

[0088] The basis for determining the location of each area and the purpose of detection are as follows:

[0089] Area ① is located above the upper screen surface in the front section, with a length of 30% of the total length of the upper screen, and the left side of the area is aligned with the left side of the upper screen surface; the wind field intensity in this area determines whether the grain mixture input through the shaking plate or the feeding device can smoothly enter area ②, avoiding the grain mixture from falling into area ① and affecting the grains from falling into the sieve holes, which is related to the cleaning efficiency.

[0090] Area ② is located above the middle upper screen surface, with its length accounting for 60% of the total length of the upper screen, and the left side of the area is adjacent to the right side of area ①. The wind field intensity in this area determines whether the materials blown from area ① can be successfully dispersed and complete the initial cleaning, which affects the cleaning efficiency and cleaning loss rate.

[0091] Area ③ is located above the upper screen surface of the tail section, its length accounts for 10% of the total length of the upper screen, and the right side of the area is aligned with the right side of the upper screen surface; the wind field intensity in this area determines whether the last part of impurities in the material mixture can be successfully cleaned, affecting the cleaning impurity rate.

[0092] Area ④ is located below the lower screen surface, with the same length as the lower screen surface, and the left side of the area is aligned with the left side of the lower screen surface; the wind field intensity in this area determines whether all impurities except seeds in the material mixture can be effectively separated and blown out of the cleaning chamber. It affects the cleaning effect, and the appropriate wind speed can greatly reduce the impurity rate of the cleaning.

[0093] Area ⑤ is located directly above area ①, with its left side aligned with the leftmost end of the cleaning chamber and its right side aligned with the right side of area ①. There is a possibility of eddy currents in this area. If the eddy currents are relatively complete and strong, impurities in the grain mixture will be affected by the eddy currents and confined to a certain area, which will adversely affect the cleaning efficiency.

[0094] Area ⑥ is located below area ④, with its left side aligned with the left side of area ② and its right side aligned with the right side of area ②. There is also the possibility of eddy currents in this area. If the eddy currents are relatively complete and strong, the impurities in the grain mixture will be affected by the eddy currents and will repeatedly roll over the top of the grain collection point without being cleared out, which will increase the impurity rate in the cleaning process.

[0095] As a preferred embodiment of the present invention, the rule reasoning of step S5 includes:

[0096] S5.1 Obtain the minimum floating velocity of grains according to the knowledge of suspension velocity parameters of rice and wheat components stored in the post-processing knowledge base;

[0097] S5.2 compare the maximum wind speed in the divided area with the minimum grain grain floating speed;

[0098] S5.3 Analyze the wind farm intensity and wind farm type based on the wind farm type judgment rule knowledge, wind farm intensity judgment rule knowledge, and unreasonable wind condition modification suggestion knowledge stored in the post-processing knowledge base.

[0099] S5.4 Determine whether the wind speed is reasonable. If it is unreasonable, analyze the reason. Determine whether there is an eddy current. If there is an eddy current, analyze the strength of the eddy current. Determine whether the eddy current has an adverse effect. If so, analyze the adverse effect to obtain the grain cleaning reasoning result.

[0100] As a specific embodiment of the present invention, the present invention uses rule-based reasoning to analyze the reasonable wind speed and wind field within the region, and its reasoning process is as follows: Figure 6 Shown:

[0101] 1. Reasoning process of airflow analysis above the screen surface in the cleaning chamber

[0102] During the operation of the cleaning screen, the distribution of the grain mixture along the length of the upper screen surface shows a pattern of less in the front, more in the middle, and less at the end. In the front section of the screen, the grains and impurities are easily piled up and mixed together, and only a small number of grains close to the screen surface and close to the screen holes are successfully cleaned; as the grain mixture moves along the screen surface, the material is gradually shaken and blown away, and the probability of grains falling into the screen holes increases; at the end of the screen, the volume of the grain mixture will be sharply reduced, the residual impurities will be blown out, and the remaining grains will pass through the cleaning screen.

[0103] 2. Reasoning process of airflow analysis under the screen surface under the cleaning chamber

[0104] Compared with the upper screen surface, the airflow on the lower screen surface is considered qualified as long as it is evenly distributed and the wind speed is generally lower than the grain suspension speed.

[0105] 3. Reasoning process at the feeding port of the cleaning chamber and the grain auger

[0106] The two places correspond to areas ⑤ and ⑥. The inferred wind field is qualified if there is no eddy current or there is eddy current and the effective average wind speed is less than the minimum suspension speed of the harvested crops.

[0107] 4. Post-processing process conclusion reasoning process

[0108] Integrate the reasoning process of 1-3 above and design the conclusion template as follows:

[0109] The wind field data set near the z=X mm section is analyzed. It can be seen from the result file data named csv-name that above the upper screen surface, the assessed wind speed before the screen is V①m / s, wind speed EVA①; the assessed wind speed in the screen is V②m / s, wind speed EVA②; the assessed wind speed at the screen tail is V③m / s, wind speed EVA③; among which the wind speed before the screen COM1 is the wind speed in the screen, the wind speed in the screen COM2 is the wind speed after the screen, and the wind speed before the screen COM3 is 1.5 times the wind speed in the screen; the assessed wind speed of the airflow below the lower screen surface is V④m / s, and the wind speed COM4 is the place where crop particles fall into the grain collection area.

[0110] Therefore, the airflow distribution on the upper screen surface is EVA, wherein the wind speed before the upper screen is EVA①, CON1, the wind speed in the middle of the upper screen is EVA②, CON2, and the wind speed at the tail of the upper screen is EVA③, CON3.

[0111] Located near the feeding port of the cleaning chamber, JUD1 vortex, the regional assessment wind speed is about V⑤m / s, JUD2 material transmission, JUD3 material is drawn into the vortex and causes accumulation, which reduces the hidden danger of cleaning efficiency, CON4.

[0112] Located in the area near the grain collection in the cleaning room, JUD4 vortex, the regional assessment wind speed is around V⑥m / s, and JUD5 grain collection work.

[0113] The meaning of key variables: X is the section position coordinate, the position information of the section cut; V① is the evaluation wind speed of the front section above the upper screen surface (area ①); V② is the evaluation wind speed of the middle section above the upper screen surface (area ②); V③ is the evaluation wind speed of the tail section above the upper screen surface (area ③); V④ is the evaluation wind speed of the entire section below the lower screen surface (area ④); V⑤ is the evaluation wind speed near the feed inlet of the cleaning chamber (area ⑤); V⑥ is the evaluation wind speed near the grain collection point of the cleaning chamber (area ⑥); EVA is the evaluation of the rationality of the airflow distribution on the upper screen surface; EVA① is the evaluation of the evaluation wind speed of area ①; EVA② is the evaluation of the evaluation wind speed of area ②; EVA③ is the evaluation of the evaluation wind speed of area ③; COM1 is the wind speed of area ① and area ②. COM2 is the comparison result of wind speed between area ② and area ③; COM3 is the comparison result of wind speed between area ① and 1.5 times of area ②; COM4 is the influence of wind speed in area ④ on particle falling; CON1 is the influence of wind speed in area ①; CON2 is the influence of wind speed in area ②; CON3 is the influence of wind speed in area ③; CON4 is the influence of eddy current in area ⑤; JUD1 is the result of whether eddy current exists in area ⑤; JUD2 is the influence of wind speed in area ⑤ on material conveying; JUD3 is the result of whether eddy current exists in area ⑤; JUD4 is the result of whether eddy current exists in area ⑥; JUD5 is the influence of wind speed in area ⑥ on material conveying.

[0114] Combined with the experience knowledge of screen surface airflow analysis in the knowledge base, the complete process of post-processing process conclusion reasoning is summarized, such as Figure 7 As shown;

[0115] As a specific embodiment of the present invention, the present invention selects the rice of model Zhendao No. 18 as the object of the reasoning case, stores the minimum and maximum values ​​of the suspension speed of the rice of model Zhendao No. 18 through the knowledge base, and the post-processing process conclusion reasoning example is as follows:

[0116] For the velocity field data set near the z=-0.05mm section, the wind speed above the upper screen surface was evaluated to be 4.8580m / s before the screen, which was low; the wind speed in the screen was evaluated to be 2.3251m / s, which was low; the wind speed at the tail of the screen was evaluated to be 4.0239m / s, which was high; the wind speed before the screen was greater than the wind speed in the screen, the wind speed in the screen was less than the wind speed after the screen, and the wind speed before the screen was less than 1.5 times the wind speed in the screen; the wind speed of the airflow below the lower screen surface was evaluated to be 6.0693m / s, and the wind speed did not affect the crop particles falling into the grain collection area.

[0117] Therefore, the airflow distribution on the upper screen surface is reasonable, in which the wind speed before the upper screen is too small, which may easily cause low cleaning efficiency, the wind speed in the upper screen is too low, which may easily cause high impurity content, and the wind speed at the tail of the upper screen is too high, which may easily cause high loss rate.

[0118] There are eddies in the area near the feed port of the cleaning chamber. The assessed wind speed in the area is around 0.95m / s, which is not conducive to material transportation. There is a risk that materials will be drawn into the eddies and accumulate, thus reducing the cleaning efficiency. However, considering that the eddies are relatively small in intensity, this risk can be ignored.

[0119] There are no eddies in the area near the grain collection point in the cleaning room, and the regional assessed wind speed is around 1.1m / s, which does not affect the grain collection work.

Claims

1. A harvester cleaning flow field analysis method based on target detection and rule reasoning, characterized in that: The method comprises: S1 obtains the harvester cleaning flow field simulation result data; S2 draws a scatter plot of the key cross-section of the flow field in the cleaning chamber based on the Python data analysis and visualization tool library and simulation result data to construct a training data set; S3 selects YOLOv8 as the recognition model for key points of the scatter plot of the key cross-section of the flow field in the cleaning chamber, and uses the training data set to train and optimize the recognition model to obtain the key point recognition model; S4 combines the key points of the key cross-section scatter diagram of the cleaning chamber flow field identified by the key point recognition model, and divides the harvester cleaning chamber into regions through division rules; S5 analyzes the wind field intensity and wind field type in the divided area based on rule reasoning to obtain the post-processing conclusion reasoning result.

2. The method for analyzing the flow field of harvester cleaning based on target detection and rule reasoning according to claim 1 is characterized in that: The step S1 selects various models of harvester cleaning chamber models and adjusts the model parameters, including: screen type, model size, and distance between upper and lower screens; uses SOLIDWORKS to establish a three-dimensional model of the harvester cleaning chamber, and uses fluid simulation software ANSYS FLUENT to simulate and analyze the flow field of the cleaning chamber to obtain simulation result data.

3. The method for analyzing the flow field of harvester cleaning based on target detection and rule reasoning according to claim 1 is characterized in that: The step S2 comprises: S2.1 uses the Python data analysis and visualization tool library to filter out key cross-section data from the simulation result data, and processes the filtered key cross-section data into a key cross-section scatter plot to clearly show the upper and lower screen positions and lengths; S2.2 By setting the thickness of the key section and controlling the sparsity of the scatter plot, the key section scatter plots under different slice thicknesses are obtained. At the same time, the position coordinates of the key points can be determined according to the sparsity of the points and the existence of the points; S2.3 performs data enhancement on the key cross-section scatter plot, including random scaling, random offset, horizontal mirroring, vertical mirroring, and random rotation of the scatter plot to obtain a training data set.

4. The method for analyzing the flow field of harvester cleaning based on target detection and rule reasoning according to claim 1 is characterized in that: The step S3 includes: S3.1 Label the training data set. Drag the mouse in labelme to get a key point labeling box. Then add classification labels to the boxed area, and then mark the area where the key points are located in the box. S3.2 randomly divide the labeled training data set into a training set and a validation set according to proportion, wherein the training set is used to train the target detection model, and the validation set is used to verify and test the key point recognition model; S3.3 Configure the experimental environment, train the key point recognition model with the training set, and verify the trained key point recognition model with the verification set; S3.4 uses intersection-over-union ratio, recall rate, precision rate, average precision and mean of average precision as evaluation indicators of key point recognition model, evaluates key point recognition performance through evaluation indicators, and analyzes the convergence of key point recognition model through the convergence curve of evaluation indicators and the loss function curve of the model; S3.5 uses the gradient descent optimizer to adjust relevant parameters to optimize the key point recognition model and obtain a key point recognition model with better performance.

5. The method for analyzing the flow field of harvester cleaning based on target detection and rule reasoning according to claim 1 is characterized in that: The cleaning room area division in step S4 includes: S4.1 predict the key point coordinates using the key point recognition model, and determine the region division rules based on the prediction results; S4.2 Use the regional division rules to divide the harvester cleaning room into areas, including: Area ① is located above the upper screen surface of the front section, with a length of 30% of the total length of the upper screen, and the left side of the area is aligned with the left side of the upper screen surface; Area ② is located above the middle upper screen surface, and its length accounts for 60% of the total length of the upper screen. The left side of the area is adjacent to the right side of area ①. Area ③ is located above the upper screen surface of the tail section, and its length accounts for 10% of the total length of the upper screen, and the right side of the area is aligned with the right side of the upper screen surface; Area ④ is located below the lower screen surface, with the same length as the lower screen surface, and the left side of the area is aligned with the left side of the lower screen surface; Area ⑤ is located directly above Area ①, with its left side aligned with the leftmost end of the cleaning room, and its right side aligned with the right side of Area ①; Area ⑥ is located below area ④, with its left side aligned with the left side of area ② and its right side aligned with the right side of area ②.

6. The method for analyzing the flow field of harvester cleaning based on target detection and rule reasoning according to claim 1 is characterized in that: The rule reasoning in step S5 includes: S5.1 Obtain the minimum floating velocity of grains according to the knowledge of suspension velocity parameters of rice and wheat components stored in the post-processing knowledge base; S5.2 compare the maximum wind speed in the divided area with the minimum grain grain floating speed; S5.3 Analyze the wind farm intensity and wind farm type according to the wind farm type judgment rule knowledge, wind farm intensity judgment rule knowledge, and unreasonable wind condition modification suggestion knowledge stored in the post-processing knowledge base; S5.4 Determine whether the wind speed is reasonable. If it is unreasonable, analyze the reason. Determine whether there is an eddy current. If there is an eddy current, analyze the strength of the eddy current. Determine whether the eddy current has an adverse effect. If so, analyze the adverse effect to obtain the grain cleaning reasoning result.

7. The method for analyzing the flow field of harvester cleaning based on target detection and rule reasoning according to claim 1 is characterized in that: The S4.1 includes: S4.1.1 Load the key section cross-section scatter plot and pre-process the input data; The backbone network in the S4.1.2YOLOv8 model extracts key cross-section scatter plot features and locates and classifies key points through the detection head; S4.1.3 The YOLOv8 model performs non-maximum suppression on prediction boxes with a high degree of overlap, removes duplicate prediction boxes, and retains the best prediction results; S4.1.4 There are four prediction results of the key point coordinates in the key section scatter diagram, which are located at the front end of the upper screen surface, the rear end of the upper screen surface, the front end of the lower screen surface, and the rear end of the lower screen surface in the cleaning chamber; the area division rules based on key points are determined by analyzing the fluid simulation result data in the cleaning chamber and the distribution of airflow in the cleaning chamber.

Images