Grain pile form dynamic monitoring method based on intelligent visual system

Abstract

The invention provides a grain pile form dynamic monitoring method based on an intelligent visual system, and the method comprises the steps: extracting a speed field and a direction field of grain flow motion from video data in a grain flow falling process through employing an optical flow field analysis method, and obtaining the initial information of a grain flow track; according to the track change information, the initial diffusion range of the grain flow at the bin bottom is analyzed, whether the diffusion mode is uniform or not is judged, and if diffusion is not uniform, the corresponding area is marked as an irregular diffusion area; a dynamic diffusion model is established based on the track change information, the dynamic diffusion model is adopted to process an irregular diffusion area, and the accumulation speed and the form change trend of the grain pile are predicted in combination with the speed field and the direction field of grain flow movement; and according to the predicted grain pile stacking speed and form change trend, operation parameters of the conveying equipment are adjusted, and the operation parameters comprise the speed and angle of a conveying belt.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for dynamically monitoring the shape of a grain pile based on an intelligent vision system. Background Art

[0002] During the process of grain silo filling operation, conveying equipment continuously imports grains into the silo. The grains freely fall from the end of the conveyor belt, forming a dynamic grain flow. The grain flow presents a specific trajectory during the falling process, and its trajectory is affected by factors such as the physical properties of grain particles, the height and angle of the conveying equipment, and the air flow in the silo. After the grain flow contacts the bottom of the silo, it begins to spread and gradually accumulate, forming a grain pile. The formation process of the grain pile has obvious dynamic characteristics, and its spreading pattern depends on the fluidity of the grains, the shape of the bottom of the silo, and the stacking height of the grain pile. As the grain pile continuously grows, the stacking shape of the grains gradually transitions from local diffusion to overall uniform distribution. However, local uneven stacking may occur during this process. For example, the grains may accumulate too fast or too slow in a certain area, resulting in an uneven or inclined surface of the grain pile. During the formation process of the grain pile, there is a complex dynamic relationship between the falling trajectory of the grain flow and the spreading pattern. The falling speed and angle of the grain flow directly affect the initial spreading range of the grains at the bottom of the silo, and the spreading range determines the stacking speed and shape of the grain pile. If the grain flow trajectory is unstable or the spreading pattern is uneven, it may lead to local stacking differences in the grain pile, thereby affecting the storage efficiency and safety of the grains in the silo. To monitor the uniformity of the grain pile formation process in real time, it is necessary to dynamically analyze the falling trajectory and spreading pattern of the grain flow. Through optical flow field analysis technology, the velocity field and direction field of the grain flow movement can be captured, thereby evaluating the uniformity during the grain pile stacking process. However, the complexity of the dynamic characteristics of the grain flow poses challenges to optical flow field analysis. For example, sudden changes in the grain flow trajectory, irregular changes in the spreading pattern, and interference from environmental factors in the silo will all affect the accuracy and real-time performance of optical flow field analysis. In addition, the dynamic changes in the shape of the grain pile have a direct impact on the control strategy of the filling operation. If the operating parameters of the conveying equipment cannot be adjusted in time, it may lead to uneven grain pile shape, resulting in an uneven or inclined surface of the grain pile, thereby affecting the subsequent outloading operation efficiency. Summary of the Invention

[0003] The present invention provides a method for dynamically monitoring the shape of a grain pile based on an intelligent vision system, mainly including:

[0004] Using the optical flow field analysis method to extract the velocity field and direction field of the grain flow movement from the video data during the falling process of the grain flow, and obtaining the initial information of the grain flow trajectory;

[0005] According to the velocity field and direction field of the grain flow movement, if the change rate in a local area of the velocity field or direction field exceeds a preset change rate threshold, it is determined that there is a sudden change in the grain flow trajectory, and the area with the sudden change is marked as the trajectory mutation area;

[0006] Adopt the local optical flow field analysis method to refine the velocity field and direction field in the trajectory mutation area, and obtain the trajectory change information, which includes the change amplitude, change direction, change duration and change area range;

[0007] According to the trajectory change information, analyze the initial diffusion range of the grain flow at the bottom of the silo, and judge whether the diffusion mode is uniform. If the diffusion is not uniform, mark the corresponding area as the irregular diffusion area;

[0008] Based on the trajectory change information, establish a dynamic diffusion model, use the dynamic diffusion model to process the irregular diffusion area, and combine the velocity field and direction field of the grain flow movement to predict the stacking speed and morphological change trend of the grain pile;

[0009] According to the predicted stacking speed and morphological change trend of the grain pile, adjust the operating parameters of the conveying equipment, and the operating parameters include the conveyor belt speed and angle;

[0010] Real-time monitor the grain flow trajectory and diffusion mode of the adjusted conveying equipment, and judge whether the standard deviation of the grain pile height, the uniformity of the coverage area and the distribution of the stacking density of the grain pile are improved. If so, record the final grain flow trajectory and diffusion mode data.

[0011] Further, the video data during the falling process of the grain flow is analyzed by the optical flow field analysis method to extract the velocity field and direction field of the grain flow movement, and the initial information of the grain flow trajectory is obtained, including: according to the change of the gray value between adjacent frames in the video frame sequence image obtained during the falling process of the grain flow, the region matching method is used to identify the edge contour region of the grain flow, the identified grain flow region is divided into multiple grid units, the corresponding relationship between the illumination value and the gray value is calculated for each grid unit to generate a grain flow density distribution map, a gray value gradient constraint equation and a brightness constancy constraint equation are established between two adjacent frames of images, and the constraint equations are solved according to the multi-scale iterative method. The constraint equations refer to the gray value gradient constraint equation and the brightness constancy constraint equation, and the distribution data of the grain flow optical flow vector field is obtained from the solution results of the constraint equations; the grain flow velocity vector in each grid unit is calculated according to the distribution data of the grain flow optical flow vector field, the least square method is used to fit the grid unit velocity vector field, and the overall movement trend data of the grain flow is obtained; if the included angle between the velocity vector of any grid unit and the velocity vector of the adjacent unit in the overall movement trend data of the grain flow is greater than the preset threshold, a tracking point is set at the edge of the grid unit, and the displacement of each tracking point is calculated by the pyramid Lucas-Kanade optical flow tracking algorithm; the displacement data of the tracking points is subjected to Gaussian filtering to obtain smoothed trajectory data, and the centroid coordinate sequence of the grain flow is calculated by the minimum circumscribed rectangle method; a quadratic polynomial fitting function is constructed according to the centroid coordinate sequence of the grain flow as the grain flow movement feature description function, and geometric parameters such as the area, edge length and diffusion radius of the grain flow region are calculated from the function coefficients; the Euclidean distance calculation method is used to compare the change of the geometric parameters between the front and back frames to obtain the grain flow movement parameters, and the initial information of the grain flow falling trajectory is obtained through the grain flow movement parameters.

[0012] Further, according to the velocity field and direction field of the grain flow, if the change rate of the velocity field or direction field exceeds a preset change rate threshold in a local area, it is determined that there is a mutation in the grain flow trajectory, and the area where the mutation exists is marked as the trajectory mutation area, including: evenly dividing the grain flow area according to the grain flow velocity field data, using the Sobel operator to calculate the first-order difference values of the velocity data point sets in the horizontal and vertical directions in each grid cell respectively, and calculating the velocity field change rate matrix from the difference values; for the direction vector sets in each grid cell in the grain flow direction field data, using the three-point fitting method to calculate the included angle change amount between adjacent direction vectors, and calculating the direction field change rate matrix through the included angle change amount; using the double-threshold comparison method to detect the velocity field change rate matrix and the direction field change rate matrix, if the velocity field change rate in any grid cell exceeds the preset velocity threshold or the direction field change rate exceeds the preset direction threshold, then mark this grid cell as a mutation candidate point; performing connectivity analysis on all mutation candidate points using the region growing algorithm, determining the set of adjacent mutation candidate points through the eight-neighborhood search method, extracting the region boundary point sequence using the boundary tracking algorithm, and constructing a piecewise cubic spline curve as the trajectory mutation area boundary curve through the boundary point sequence; according to the trajectory mutation area boundary curve, using the integral calculation method to obtain the regional area value, using the arc length calculation method to obtain the regional perimeter value, and obtaining the trajectory mutation area range parameter from the area and perimeter values.

[0013] Further, the local optical flow field analysis method is used to refine the velocity field and direction field in the trajectory mutation region to obtain trajectory change information. The trajectory change information includes the change amplitude, change direction, change duration, and change region range, and includes: performing local grid division with a fixed pixel size on the optical flow field data according to the trajectory mutation region range, and recalculating the local velocity field and direction field for each grid cell using the pyramid optical flow algorithm to obtain the velocity vector and direction vector sequences within the grid cell; for the velocity vector sequence within the grid cell, fitting the velocity change curve using the cubic polynomial least squares method, calculating the velocity difference between adjacent frames through the velocity change curve, and performing trapezoidal integration on the difference to obtain the velocity change amplitude parameter; constructing a curve of the direction angle changing with time according to the direction vector sequence within the grid cell, using the local extreme value detection method to extract the inflection point coordinate sequence in the direction change curve, and fitting the inflection point coordinate sequence with a quadratic curve to obtain the direction change angle parameter; for the velocity change amplitude parameter, using the continuous frame counting method to obtain the number of frames with changes, and calculating the change duration parameter through the number of frames and the sampling frequency; according to the velocity change amplitude parameter and the direction change angle parameter, calculating the motion change intensity value within each grid cell using the weighted average method, and normalizing the motion change intensity value to obtain the motion trend data; for the motion trend data, using the regional centroid calculation method to obtain the sequence of the central point coordinates of the mutation region, constructing the regional boundary contour using the minimum circumscribed rectangle method through the coordinate sequence, and calculating the change region range parameter from the boundary contour; constructing a trajectory change feature vector according to the change region range parameter, the change duration parameter, the velocity change amplitude parameter, and the direction change angle parameter, and obtaining the complete trajectory change information from the feature vector.

[0014] Further, based on the trajectory change information, analyze the initial diffusion range of the grain flow at the bottom of the silo, and determine whether the diffusion pattern is uniform. If the diffusion is non-uniform, mark the corresponding area as an irregular diffusion area, including: according to the change amplitude and change direction data in the trajectory change information, use the region growing algorithm based on eight-neighborhood to segment the area covered by the grain flow at the bottom of the silo, obtain the boundary point sequence of the initial diffusion range at the bottom of the silo through the gray similarity as the growth condition, and construct the diffusion range contour curve from the boundary point sequence; for the area divided by the diffusion range contour curve, divide the area into uniform grids of a fixed pixel size, use the pixel counting method to count the number of grain flow pixels in each grid unit, and calculate the volume density value of each grid unit from the pixel number; perform maximum-minimum normalization on the volume density values of each grid unit, use the linear mapping method to convert the normalized result into the diffusion intensity value, and construct the diffusion intensity distribution map from the diffusion intensity value; according to the diffusion intensity distribution map, use the Sobel operator to calculate the diffusion intensity gradients in the horizontal and vertical directions respectively, and obtain the diffusion intensity change rate data through the gradient value; for the diffusion intensity change rate data, use the weighted average method to calculate the overall diffusion uniformity value of the area, and calculate the diffusion irregularity parameter through the difference degree between the diffusion uniformity value and the ideal uniform distribution; if the diffusion irregularity parameter exceeds the preset uniformity threshold, use the connected component labeling algorithm to mark the area where the diffusion intensity gradient is greater than the gradient mean value, and obtain the irregular diffusion area from the marking result; for the irregular diffusion area, use the boundary tracking algorithm to extract the region contour point sequence, construct the irregular region boundary curve from the contour point sequence, and obtain the irregular region range data from the boundary curve.

[0015] Furthermore, a dynamic diffusion model is established based on the trajectory change information, and the dynamic diffusion model is used to process the diffusion irregular region. Combining the velocity field and direction field of the grain flow movement, the stacking velocity and the morphological change trend of the grain pile are predicted, including: according to the change amplitude and change direction data in the trajectory change information, a dynamic diffusion prediction model is constructed by using a double-layer feedforward neural network, and the diffusion rate and direction prediction parameters are obtained through training with historical data, and the grain flow diffusion movement trend in the irregular region is calculated from the prediction parameters; for the diffusion movement trend data, the irregular region is spatially divided by using a uniform grid with a fixed pixel size, and the change value of the stacking amount in the grid cell is calculated through the velocity field and direction field data, and the stacking density change rate per unit time is obtained from the change value of the stacking amount; according to the stacking density change rate, the density distribution of the grid cell is reconstructed by using the bilinear interpolation method, the grain pile height field function is constructed through the density distribution data, the three-dimensional morphological data of the grain pile are calculated from the height field function, the contour line sequence at different heights is obtained by using the contour line extraction algorithm, and the stacking angle of each layer is calculated from the contour line sequence, and the grain pile edge contour curve is constructed from the stacking angle data; according to the grain pile edge contour curve, morphological smoothing processing is performed on the curve by using opening operation and closing operation, and the horizontal expansion area of the grain pile is calculated through the processed contour curve; for the horizontal expansion area of the grain pile, spatial integration operation is performed in combination with the height field function, the change function of the grain pile volume over time is obtained from the integration result, and the predicted stacking velocity value is obtained by taking the derivative of the change function; according to the predicted stacking velocity value and the actual observation data, error calculation is performed, and the parameters of the dynamic diffusion prediction model are optimized by using the backpropagation algorithm, and the stacking morphological change trend is obtained from the optimized model.

[0016] Further, adjusting the operating parameters of the conveying equipment according to the predicted stacking speed and morphological change trend of the grain pile, where the operating parameters include the conveyor belt speed and angle, includes: establishing a conveyor belt parameter response function using an adaptive control algorithm based on the predicted stacking speed and morphological change trend data of the grain pile. The input layer includes the stacking speed and morphological characteristics, the hidden layer uses the hyperbolic tangent activation function, and the output layer generates the speed adjustment amount and the angle adjustment amount; for the speed adjustment amount and the angle adjustment amount, a regulation amount limiting function is constructed based on the preset equipment operation constraints, and the regulation amount is constrained within a range through the limiting function, and the boundary values of the conveyor belt operating parameters are obtained from the constraint results. A fuzzy rule inference engine is used to perform a linkage analysis on the speed and angle parameters. The input variables are divided into three fuzzy sets: low speed, medium speed, and high speed, and the output variables are divided into three fuzzy sets: small angle, medium angle, and large angle; for the fuzzy inference result, defuzzification is performed using the centroid method, and the speed adjustment step and the angle adjustment step are calculated through the defuzzification result, and a parameter adjustment path curve is constructed from the adjustment step data; according to the parameter adjustment path curve, the mean square error is used as the objective function, and the speed response curve and the angle response curve are calculated through gradient descent iteration, and the belt speed increment sequence and the angle increment sequence are obtained from the response curves. A piecewise linear mapping method is used to perform discretization processing within the set control interval, and the conveyor belt control instruction sequence is obtained through the discretization result; according to the conveyor belt control instruction sequence, a linear interpolation algorithm is used to smooth the instruction data, and the operating parameters of the conveying equipment are adjusted in real time through the processed instruction data, and the real-time operating parameters of the conveyor belt are obtained from the adjustment data.

[0017] Furthermore, the grain flow trajectory and diffusion pattern of the conveying equipment after real-time monitoring and adjustment are analyzed to determine whether the standard deviation of the grain pile height, the uniformity of the coverage area, and the distribution of the bulk density of the grain pile are improved. If improved, the final grain flow trajectory and diffusion pattern data are recorded, including: according to the operating state of the adjusted conveying equipment, an infrared height measurement sensor array is used to perform grid scanning on the surface of the grain pile, and a three-dimensional height distribution map of the grain pile is constructed through the scanning data. The height data of each measurement point is extracted from the height distribution map, the least squares method is used to calculate the average height, the deviations of each measurement point from the average value are summed, and the standard deviation index of the grain pile height is obtained by dividing the sum of the deviations by the number of measurement points; according to the three-dimensional height distribution map of the grain pile, the region growing algorithm based on the height gradient is used to segment the coverage area of the grain pile, the regions with the height difference between adjacent points less than the preset threshold are merged into the same category, the contour line of the coverage area is obtained through region merging, the region is divided by a regular hexagon grid, and the local uniformity value is obtained by calculating the ratio of the height change rate to the area within each grid unit, and the area uniformity parameter is calculated from the local uniformity value; according to the regular hexagon grid division result, a laser density detector array arranged vertically downward is used to perform density scanning on each grid unit, a bulk density distribution map is obtained through the scanning data, convolution operation is performed using a Gaussian kernel function, the density difference between adjacent grid units is calculated through the convolution result, and the density gradient distribution data is obtained from the density difference; if the standard deviation index of the height is lower than the preset standard deviation threshold and the area uniformity parameter is higher than the preset uniformity threshold, four-way cameras are used to collect the grain flow trajectory from multiple angles, a complete trajectory sequence is obtained through image stitching, the trajectory features are extracted using a convolutional neural network, the features are classified through a fully connected layer, and a diffusion pattern database is constructed from the classification results, and the content of the database is stored in a structured manner.

[0018] The technical solution provided by the embodiment of the present invention may include the following beneficial effects:

[0019] The present invention discloses a method for dynamically monitoring the shape of a grain pile based on an intelligent vision system. This method processes the video data of the grain flow during the falling process through the optical flow field analysis technology, extracts the velocity field and direction field of the grain flow movement, and identifies the trajectory mutation regions. Further, the local optical flow field analysis method is used to refine the trajectory change information, and the initial diffusion range and uniformity of the grain flow at the bottom of the bin are analyzed. Based on this, the present invention establishes a dynamic diffusion model to predict the stacking speed and morphological change trend of the grain pile, and adjusts the operating parameters of the conveying equipment accordingly. Through real-time monitoring and feedback adjustment, the present invention can effectively improve the height uniformity, coverage area uniformity, and bulk density distribution of the grain pile, realize the optimized control of grain storage in the bin, and improve the storage efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1It is a flowchart of a dynamic monitoring method for the shape of a grain pile based on an intelligent vision system according to the present invention.

[0021] Figure 2 It is another flowchart of a dynamic monitoring method for the shape of a grain pile based on an intelligent vision system according to the present invention. Specific embodiments

[0022] In order to enable those skilled in the art of this technology to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this specification.

[0023] Such as Figure 1-2 , a dynamic monitoring method for the shape of a grain pile based on an intelligent vision system in this embodiment may specifically include:

[0024] S101. During the falling process of the grain flow, use the intelligent vision system to collect video data and adopt the optical flow field analysis technology to extract the velocity field and direction field of the grain flow movement, generate the initial information of the grain flow trajectory, and at the same time calculate the geometric characteristics and motion parameters of the grain flow area through a multi-level analysis method to support subsequent diffusion evaluation.

[0025] In the embodiment of the present invention, when the grain flow falls, it is guided into the warehouse by the conveying equipment, and the intelligent vision system collects the video frame sequence in real time through the camera for analyzing the dynamic characteristics of the grain flow. The optical flow field analysis technology can capture the motion characteristics of the grain flow between frames and generate the velocity field and direction field data, providing a basis for trajectory monitoring.

[0026] S1011. By analyzing the gray value changes of adjacent frames in the video frame sequence, the region matching method is used to identify the edge contour region of the grain flow and divide the grid cells. Calculate the relationship between the illuminance and the gray value of each cell to generate the grain flow density distribution map. Subsequently, a constraint equation set is established based on the density distribution map and the optical flow vector field distribution data is solved. In the embodiment of the present invention, first, the collected video frame sequence is preprocessed to extract the change characteristics of the gray values between adjacent frames. The region matching method is used to locate the edge contour of the grain flow, and the grain flow region is divided into multiple grid cells, such as grid cells with a size of 32×32 pixels. For each grid cell, calculate the corresponding relationship between the illuminance and the gray value to generate a density distribution map reflecting the grain flow distribution law. Then, based on the gray value gradient and brightness constancy characteristics in the density distribution map, a gray value gradient constraint equation and a brightness constancy constraint equation are established respectively. The constraint equation set is solved by the multi-scale iteration method to obtain the optical flow vector field distribution data of the grain flow, which describes the motion state of the grain flow in each grid cell. Taking the grain unloading scenario of a certain granary as an example, the time interval between adjacent frames is 0.04 seconds, the vector distribution in the central region is uniform, and the edge region shows a velocity gradient change.

[0027] S1012. Calculate the grain flow velocity vector in each grid cell according to the optical flow vector field distribution data, use the least square method to fit the velocity vector field to generate the overall motion trend data of the grain flow, and calculate the grain flow region area and diffusion radius through the quadratic polynomial fitting function. At the same time, the tracking point and filtering technology are used to refine the motion parameters to obtain the initial information of the grain flow trajectory.

[0028] In the embodiment of the present invention, based on the optical flow vector field distribution data, the grain flow velocity vector of each grid cell is calculated, and the overall movement trend data of the grain flow is formed by least square fitting. If the included angle between the velocity vector of a certain grid cell and its adjacent cell exceeds a preset threshold, such as 25 degrees, tracking points are set at the edge of this area, and the displacement of the tracking points is calculated using the pyramid Lucas-Kanade optical flow tracking algorithm. Gaussian filtering is applied to the displacement data to generate smoothed trajectory data, and the centroid coordinate sequence of the grain flow is calculated using the minimum bounding rectangle method. A quadratic polynomial fitting function is constructed based on the centroid sequence, and its coefficients reflect the movement characteristics of the grain flow. Geometric parameters such as the area, edge length, and diffusion radius of the grain flow area are calculated through the function. Further, the Euclidean distance method is used to compare the changes in geometric parameters between the front and rear frames to quantify the movement parameters of the grain flow. For example, when the Euclidean distance exceeds 8, it indicates that a significant change has occurred in the grain flow trajectory. Taking the actual grain unloading process as an example, the vertical displacement of the centroid is about 15 pixels per frame, and the horizontal offset is controlled within 5 pixels, and the fitting function accurately describes the trajectory curvature. In the embodiment of the present invention, the grain flow shows a jet-like distribution during falling, and the gray value in the edge area varies between 110 and 180. Through the above analysis, the generated initial trajectory information includes the velocity field, direction field, and geometric parameters, laying a foundation for subsequent diffusion mode judgment. This step does not limit the specific algorithm details too much, and can be optimized by technicians according to the scenario. In practical applications, these parameters are derived from experimental statistics and can effectively characterize the movement law of the grain flow, ensuring the accuracy and real-time nature of monitoring.

[0029] S102. During the falling process of the grain flow, judge the trajectory mutation situation according to the grain flow movement velocity field and direction field data extracted by the intelligent vision system. If the change rate of the velocity field or direction field in a local area exceeds the preset threshold, it is determined as a trajectory mutation, and the corresponding area is marked as a trajectory mutation area. At the same time, the boundary and range parameters of the mutation area are determined through grid analysis and curve fitting techniques to support subsequent diffusion analysis.

[0030] In the embodiment of the present invention, the velocity field and direction field of the grain flow movement reflect its dynamic characteristics, and trajectory anomalies can be detected in a timely manner by monitoring the local changes in these two fields. Trajectory mutations are usually caused by grain flow blockage, deflection, or external interference, and need to be accurately identified to optimize the grain pile shape.

[0031] S1021. Divide the grain flow area into uniform grids according to the grain flow velocity field data. Use the Sobel operator to calculate the first-order difference values of the velocity data in the horizontal and vertical directions within each grid cell and generate a velocity field change rate matrix. At the same time, calculate the change amount of the included angle between adjacent vectors for the direction field data to construct a direction field change rate matrix, providing a quantitative basis for mutation detection.

[0032] In the embodiment of the present invention, the grain flow area is divided into a uniform grid of 16×16 pixels to ensure the analysis accuracy. The Sobel operator is used to perform differential calculation on the velocity data to obtain the change characteristics in the horizontal and vertical directions respectively. For example, in a certain grain bin discharging scenario, the normal grain flow velocity change rate remains below 0.15, while it can rise above 0.4 during blockage. The direction field analysis calculates the angle change between adjacent vectors through the three-point fitting method to generate the direction field change rate matrix. During normal falling, the angle change is usually less than 15 degrees, while it can reach more than 35 degrees during deflection. This quantization method clearly reflects the abnormal degree of the grain flow movement.

[0033] S1022. Perform double-threshold detection on the velocity field change rate matrix and the direction field change rate matrix. If the velocity field change rate of a certain grid cell exceeds the preset velocity threshold or the direction field change rate exceeds the preset direction threshold, it is marked as a mutation candidate point. Subsequently, the connectivity of the candidate points is analyzed through the region growing algorithm, and the boundary of the initial mutation region is extracted to further refine the mutation range.

[0034] In the embodiment of the present invention, the double-threshold detection sets the velocity field threshold to 0.35 and the direction field threshold to 30 degrees. When a grid cell exceeds any one of the thresholds, it is marked as a mutation candidate point. For example, in a certain discharging detection, 3 velocity anomaly points and 2 direction anomaly points are identified within the area of 350×400 pixels. Using the region growing algorithm, adjacent candidate points are clustered through eight-neighborhood search to form the initial mutation region. This method effectively integrates the scattered anomaly points and ensures the integrity of the mutation region.

[0035] Boundary extraction is a key step in mutation region recognition. For the initial mutation region, the boundary tracking algorithm is used to trace the edge point by point to generate a sequence of boundary points. Taking a certain detection as an example, the obtained sequence contains 86 points. Subsequently, a smooth boundary curve is constructed through piecewise cubic spline interpolation, and the curve length reaches 320 pixels. This interpolation technique can smooth the boundary noise and improve the accuracy of region description.

[0036] S1023. Calculate the area and perimeter parameters of the mutation region based on the boundary curve of the mutation region. The area value is obtained through the integration method, and the perimeter value is calculated using the arc length formula. The geometric shape characteristics of the mutation region are analyzed based on the ratio of the area to the perimeter, so as to provide data support for judging the cause of the abnormal grain flow trajectory.

[0037] In the embodiments of the present invention, the boundary curve analysis further quantifies the characteristics of the mutation region. Through integral calculation, the area of a certain mutation region is 4,800 square pixels, the perimeter is 360 pixels, and the area-perimeter ratio is 13.3. This long and narrow shape indicates that the grain flow may be deflected due to the influence of obstacles. In practical applications, such parameters can be associated with the operating state of the feeding equipment. For example, when the ratio is high, it indicates that there may be an angle deviation of the conveyor belt or interference from foreign objects in the bin. By monitoring these geometric features, it can provide a basis for subsequent adjustments.

[0038] The identification and parameterization process of the trajectory mutation region lays the foundation for the optimization of the grain pile shape. The dual analysis of the velocity field and the direction field ensures the comprehensiveness of mutation detection, while the mesh generation and curve fitting techniques improve the accuracy of boundary extraction. In an actual scenario, such as the feeding process of a certain grain transfer station, through the above methods, the problem area can be quickly located when the grain flow is abnormal, providing reliable data support for real-time regulation of conveying parameters. This step does not overly limit the threshold setting or algorithm details, which can be optimized and adjusted by technicians according to specific application scenarios.

[0039] S103. During the falling process of the grain flow, the local optical flow field analysis method is used for the trajectory mutation region to extract refined velocity field and direction field data, generate trajectory change information including the change amplitude, change direction, change duration, and regional range, and construct a feature vector through polynomial fitting and weighted calculation to support the prediction of the grain pile shape. The local optical flow field analysis method can capture the microscopic motion characteristics of the trajectory mutation region, providing accurate data support for subsequent diffusion mode analysis. The mutation region is usually caused by the obstruction or deflection of the grain flow, and the refined analysis can reveal its influence degree and change trend.

[0040] In the embodiments of the present invention, the trajectory mutation region is divided into local grids of 16×16 pixels, and each grid contains 256 pixel points to ensure the comprehensiveness of detail capture. The pyramid optical flow algorithm is used to recalculate the velocity field and direction field of each grid cell, generating a sequence of velocity vectors and direction vectors. For example, in a feeding scenario of a certain granary, the velocity vector of the normal grain flow changes smoothly, while the vectors in the mutation region fluctuate significantly, and the local velocity can increase from 2 pixels per frame to 8 pixels per frame. This algorithm improves the adaptability to complex motions through multi-scale decomposition.

[0041] S1031. For the sequence of velocity vectors within the grid cell, the least squares method of cubic polynomials is used to fit the velocity change curve and calculate the velocity difference between adjacent frames to obtain the velocity change amplitude. At the same time, the inflection point characteristics are analyzed through the sequence of direction vectors to generate the direction change angle, and the change duration is statistically analyzed in combination with the number of consecutive frames to quantify the mutation characteristics.

[0042] In the embodiment of the present invention, the fitting of the velocity vector sequence adopts a cubic polynomial, which can smooth the noise and reflect the change trend. After calculating the velocity difference between adjacent frames, the velocity change amplitude parameter is obtained by the trapezoidal integration method. During normal falling, the difference is mostly between 2 and 3 pixels per frame, while it can reach more than 8 pixels per frame during mutation. For the direction vector, a direction angle change curve is constructed and the inflection points are extracted by local extreme value detection. After quadratic curve fitting, the direction change angle parameter is generated. In actual detection, an inflection point angle exceeding 45 degrees often indicates the deflection of the grain flow. The change duration is obtained by counting consecutive frames. Calculated at a sampling rate of 50 frames per second, a short disturbance is about 0.2 seconds, and a serious anomaly can exceed 0.5 seconds. These parameters jointly describe the intensity and duration of the mutation.

[0043] S1032. Calculate the motion change intensity value by the weighted average method according to the velocity change amplitude and the direction change angle, and perform normalization processing to generate motion trend data. Subsequently, obtain the central point coordinate sequence of the mutation region through the calculation of the regional centroid and construct the boundary contour. Finally, generate the trajectory change feature vector by integrating various parameters to completely characterize the change information.

[0044] In the embodiment of the present invention, the motion change intensity value is calculated with a velocity-to-direction weight ratio of 3:2, highlighting the dominant role of velocity mutation. After normalization, motion trend data is generated, which is convenient for cross-regional comparison. The central point coordinate sequence is calculated by using the regional centroid method, and the boundary contour is constructed by the minimum circumscribed rectangle method to obtain the range of the change region. For example, in a certain detection, the area of the normal disturbance region is about 2000 square pixels, while the blocked region exceeds 5000 square pixels. The trajectory change feature vector integrates parameters such as velocity change amplitude, direction change angle, duration, and range, providing a basis for multi-dimensional anomaly assessment.

[0045] Taking a certain grain processing factory as an example, the analysis of the velocity and direction characteristics of the mutation region reveals the deflection caused by equipment failure. The boundary contour shows a long and narrow shape, indicating that there may be obstacles in the silo. The weighted calculation and coordinate analysis quantify the spatio-temporal characteristics of the anomaly, providing a reliable reference for adjusting the conveying strategy.

[0046] In the embodiment of the present invention, the generated trajectory change information realizes the complete conversion from the original data to the feature vector through multi-level analysis. In practical applications, a quick response can be made when the grain flow mutation first appears. For example, when the direction change angle suddenly increases to 45 degrees and lasts for 0.5 seconds, combined with the regional characteristics of an area exceeding 5000 square pixels, it can be judged as a serious blockage and subsequent regulation can be triggered in a timely manner. The obtained feature vector lays a solid foundation for the establishment of a dynamic diffusion model, enhancing the practicality and adaptability of the monitoring.

[0047] S104. After the grain flow falls to the bottom of the bin, analyze the initial diffusion range of the grain flow based on the trajectory change information and judge its uniformity. If an uneven diffusion pattern is found, mark it as an irregular diffusion area. At the same time, obtain the boundary curve of the irregular area through region segmentation and density distribution calculation to guide subsequent stacking adjustment. The diffusion behavior of the grain flow at the bottom of the bin directly affects the morphological stability of the grain pile. Uneven diffusion may lead to excessive or insufficient local stacking, affecting the storage efficiency. By refining the analysis of the diffusion range and intensity, it can provide a basis for the regulation of conveying equipment.

[0048] S1041. According to the change amplitude and direction data in the trajectory change information, use the eight-neighborhood region growing algorithm to segment the area covered by the grain flow at the bottom of the bin and extract the initial diffusion range boundary point sequence based on the gray-scale similarity condition. Subsequently, divide the grid based on the boundary point sequence and calculate the volume density value of each grid cell to construct the basic data of the diffusion intensity distribution. In the embodiment of the present invention, the region growing algorithm is based on eight-neighborhood search, and pixels with a gray-scale value difference less than 10 are classified into the same region. Taking a certain granary as an example, for the grain flow falling from a height of 10 meters, its boundary point sequence usually contains 200 to 300 points, reflecting the contour of the diffusion range. Then, divide the covered area into a 16×16 pixel grid, and each unit covers 256 square pixels. By counting the number of pixels of the grain flow through the pixel counting method, under normal circumstances, the number of pixels in each unit is between 180 and 220, and the volume density value distribution is concentrated, with a difference not exceeding 15%. This step lays a data foundation for the quantification of the diffusion intensity.

[0049] In the embodiment of the present invention, the evaluation of the diffusion intensity is the core of judging the uniformity. Perform maximum-minimum normalization on the volume density value, linearly map the result to the range of 0 to 1, and generate the diffusion intensity value. Taking actual observation as an example, during uniform diffusion, the diffusion intensity distribution map is circular or elliptical, gradually changing from the center to the edge. Use the Sobel operator to calculate the diffusion intensity gradients in the horizontal and vertical directions. Under normal conditions, the gradient ratio in the two directions is close to 1, indicating that the diffusion direction is balanced. If the gradient ratio deviates significantly, it indicates an abnormal diffusion pattern.

[0050] S1042. Calculate the diffusion intensity change rate and the overall uniformity value based on the diffusion intensity distribution map. If the diffusion irregularity parameter exceeds the preset uniformity threshold, identify the gradient abnormal area through the connected component labeling algorithm and extract the boundary curve of the diffusion irregular area, so as to quantify the irregular range to support the optimization decision. It can be understood that the diffusion intensity change rate is obtained by calculating the gradient value, and the diffusion uniformity value is obtained after weighted averaging. When the uniformity value is greater than 0.85, it indicates that the diffusion is better, and the irregularity parameter is calculated by comparing with the ideal value of 0.95. If the parameter exceeds the threshold, for example, it reaches 0.2 in a certain detection, the connected component labeling is started. Taking the gradient exceeding the average value by 20% as the standard, the irregular area is divided. In a certain detection, 3 abnormal areas are found, and the largest area is 1200 square pixels. The boundary tracking algorithm extracts a sequence of contour points, including 80 to 120 points, and the constructed boundary curve clearly describes the spatial characteristics of the irregular area.

[0051] In practical applications, a uniformly diffused grain flow can ensure balanced coverage of the bottom of the warehouse and avoid local overload. Tests in a certain granary show that the timely marking of irregular areas helps to adjust the conveying angle, improving the stacking uniformity by about 10%. This step fully characterizes the diffusion behavior through multi-level analysis, from boundary extraction to intensity distribution, and then to irregular marking, providing strong support for dynamic monitoring. The threshold or grid size can be adjusted according to the scene requirements to optimize the effect.

[0052] S105. During the grain flow stacking process, establish a dynamic diffusion model based on the trajectory change information, combine the velocity field and direction field of the grain flow to predict the stacking speed and the trend of morphological changes of the grain pile, and at the same time generate a height field function through neural network training and grid analysis to quantify the stacking characteristics and optimize the model parameters. The dynamic diffusion model uses the trajectory change information to predict the diffusion behavior of the grain flow and its impact on the grain pile morphology. The model is trained with historical data to ensure that the prediction results are close to the actual stacking process, providing a scientific basis for real-time adjustment.

[0053] S1051. Construct a two - layer feed - forward neural network model based on the change amplitude and direction data in the trajectory change information and train it with historical trajectory data to obtain the diffusion rate and direction parameters. Subsequently, uniformly divide the diffusion irregular region into a grid and calculate the change rate of the stacking density to support the morphology prediction. In the embodiment of the present invention, the two - layer feed - forward neural network contains 10 input nodes, corresponding to 5 feature dimensions of the change amplitude and direction respectively. After being processed by two hidden layers, the predicted values of the diffusion rate and direction are output. Taking a certain granary as an example, when the input falling speed is 2.5 meters per second, the model predicts that the diffusion rate is between 0.8 and 1.2 meters per second, and the direction angle is less than 15 degrees. The training data comes from historical trajectories to ensure that the model adapts to various scenarios. Then, divide the irregular region into a 16×16 pixel grid, and calculate the change in the stacking volume using the velocity field and direction field. Under normal circumstances, the difference in the stacking volume between adjacent grids is about 20%, and in the uneven region, it can reach 50%. The density growth rate in the central region is about twice that of the edge. This grid - based analysis provides fine - grained data for subsequent prediction.

[0054] Exemplarily, based on the change rate of the stacking density, use the bilinear interpolation method to reconstruct the density distribution of the grid cells and generate a height - field function. In a 400×400 pixel area, the interpolated height - field shows a bell - shaped distribution, with the central height reaching 1.8 meters and gradually decreasing towards the edge. Bilinear interpolation calculates through neighborhood weighting to ensure a natural density transition, and its smoothness is better than that of the simple linear method. Subsequently, generate contour lines at different heights through the contour - extraction algorithm. The stacking angle decreases from 35 degrees at the bottom to 25 degrees at the top, which is consistent with the characteristics of the angle of repose of the grain, adding a physical basis for the morphology description.

[0055] S1052. Perform a spatial integration calculation based on the height - field function to obtain the function of the change in the grain - pile volume over time and take the derivative to get the predicted value of the stacking speed. At the same time, smooth the contour curve through morphological processing and optimize the model in combination with actual data to improve the prediction accuracy.

[0056] In the embodiment of the present invention, the spatial integration converts the height - field into a volume - change function. Taking a certain observation as an example, when the height of the grain - pile reaches 1.5 meters, the volume growth speed is about 0.8 cubic meters per second. The predicted value of the stacking speed obtained by taking the derivative reflects the stacking dynamics. The contour curve is smoothed through opening and closing operations with a radius of 5 pixels, and the horizontal expansion area is calculated after removing the noise. The initial growth rate is about 0.5 square meters per second and later drops to 0.2 square meters per second. After comparing the predicted value with the measured value, the model is optimized through backpropagation. After 50 iterations, the error is reduced to within 8%, improving the prediction reliability.

[0057] In the actual scenario, the changing trends of the volume and area predicted by the model reveal the stacking asynchrony. For example, the center rapidly increases while the edge expands slowly. This characteristic indicates the necessity of adjusting the conveying parameters. Through fine grid division and neural network analysis, a complete process from trajectory data to morphology prediction is achieved, providing an efficient tool for the real-time monitoring and optimization of grain bin filling. The number of network layers or grid accuracy can be adjusted according to specific requirements to further improve the effect.

[0058] S106. During the grain pile stacking process, adjust the operating parameters of the conveying equipment according to the predicted stacking speed and morphological change trends, including the conveyor belt speed and angle, and generate a control instruction sequence through an adaptive control algorithm and fuzzy inference to ensure that the equipment responds smoothly to the stacking requirements. Among them, the adjustment of the conveying parameters is the execution link of dynamic monitoring, which directly affects the optimization effect of the grain pile morphology. Through the combination of prediction data and equipment characteristics, precise control is achieved.

[0059] S1061. According to the predicted grain pile stacking speed and morphological characteristic data, use an adaptive control algorithm to construct a conveyor belt parameter response function and generate a speed adjustment amount and an angle adjustment amount. At the same time, constrain the adjustment range through a limiting function to ensure the safe operation of the equipment.

[0060] In the embodiment of the present invention, the adaptive control algorithm adopts a three-layer neural network structure. The input layer receives the stacking speed and morphological characteristics. The hidden layer uses the hyperbolic tangent activation function to process the non-linear relationship. The output layer generates the adjustment amount. Taking a certain grain depot as an example, when the stacking speed is 0.8 meters per second, the algorithm outputs a speed adjustment amount of 0.15 meters per second and an angle adjustment amount of 2.5 degrees. To ensure safety, the conveyor belt speed is limited to 0.5 to 2.0 meters per second, and the angle is between 15 and 45 degrees. If the adjustment amount exceeds, such as the speed reaches 2.2 meters per second, the limiting function truncates it to 2.0 meters per second. This constraint mechanism avoids equipment overload and improves the operation stability.

[0061] S1062. For the speed and angle adjustment amounts, use a fuzzy rule inference engine for linkage analysis and divide the fuzzy sets. Subsequently, defuzzify through the centroid method to generate an adjustment step size and construct a parameter adjustment path curve to optimize the adjustment process.

[0062] In the embodiment of the present invention, the fuzzy inference engine divides the speed into three fuzzy sets: low speed of 0.5 to 0.8 meters per second, medium speed of 0.8 to 1.2 meters per second, and high speed of 1.2 to 2.0 meters per second, and divides the angle into three fuzzy sets: small angle of 15 to 25 degrees, medium angle of 25 to 35 degrees, and large angle of 35 to 45 degrees. According to the empirical rules, a small angle is preferred at high speed to slow down the stacking. After defuzzification by the centroid method, a speed adjustment step size such as 0.1 meters per second and an angle step size such as 2 degrees are generated. The adjustment path curve constructed through the step size data ensures that the parameter changes are smooth and controllable, avoiding sudden changes that affect the equipment.

[0063] In an embodiment of the present invention, based on the adjustment path curve, with the mean square error as the objective function, the velocity and angle response curves are optimized through gradient descent. The weight ratio of the velocity error to the angle error is 6:4. The velocity curve stabilizes after 5 iterations, and the angle curve requires 7 to 8 iterations. In the obtained increment sequence, the velocity change is controlled within 0.1 meters per second, and the angle does not exceed 3 degrees. Subsequently, piecewise linear mapping is adopted, and the increments are discretized with 5 characteristic points to ensure a natural transition. For example, when the velocity increases from 0.8 to 1.0 meters per second, the 5-segment mapping generates a smooth sequence to avoid shocks.

[0064] In an embodiment of the present invention, through the linear interpolation algorithm, the discrete instructions are interpolated at 0.1-second intervals to make the changes in the conveyor belt parameters continuous. Taking a certain adjustment as an example, when the velocity increases from 0.8 to 1.0 meters per second and the angle is adjusted from 25 to 28 degrees, the smoothed instruction sequence makes the equipment response stable and the stacking uniformity is improved by about 12%. This step realizes the seamless connection between prediction and control through multi-level optimization, from the response function to fuzzy inference and then to instruction generation, providing efficient support for the optimization of the grain pile shape. The fuzzy set or the number of iterations can be adjusted according to the equipment characteristics to further improve the accuracy.

[0065] S107. After the conveying equipment adjusts the operating parameters, the grain flow trajectory and diffusion pattern are monitored in real time, and whether the uniformity of the grain pile is improved is judged through height, area, and density analysis. If the improvement is significant, the optimized trajectory and diffusion pattern data are recorded to support subsequent optimization. Through multi-dimensional data analysis, it is ensured that the standard deviation of the grain pile height, the uniformity of the coverage area, and the density distribution reach the expected targets.

[0066] S1071. According to the adjusted state of the conveying equipment, an infrared height measurement sensor array is used to perform grid scanning on the surface of the grain pile to construct a three-dimensional height distribution map to extract height data. At the same time, the region growing algorithm is adopted to segment the coverage area and calculate the standard deviation and uniformity parameters of the height.

[0067] In an embodiment of the present invention, the infrared height sensor array is arranged at 20-cm intervals to perform a comprehensive scan of the surface of the grain pile. Taking a 6 m × 8 m granary as an example, the height distribution map generated by 1200 measurement points clearly reflects the surface morphology, with the central height being about 1.8 m and the edge dropping to 0.8 m. The least squares method is used to calculate the average height, and the standard deviation is obtained by summing the deviations and dividing by the number of points. The ideal value is within 0.2 m, and after actual optimization, it can be stabilized at 0.15 m. The region growing algorithm based on the height gradient uses a 0.1-m difference as the threshold to segment the coverage area and generate a boundary contour line to ensure the segmentation accuracy and the reliability of boundary recognition.

[0068] In the embodiment of the present invention, the analysis of density and uniformity further quantifies the stacking state. The covered area is divided into regular hexagonal grids with a side length of 40 cm, and each grid contains 4 height points, taking into account both data representativeness and calculation efficiency. The laser density detector array scans vertically at a frequency of 10 Hz to generate a density distribution map. Under normal circumstances, the density difference between adjacent grids is less than 5%. After smoothing the noise through convolution with a 3×3 Gaussian kernel, the density gradient distribution is calculated, and the local uniformity is obtained based on the height change rate and area ratio within the grid, and then the overall uniformity parameter is generated. The optimized uniformity can be increased from 0.75 to above 0.9, significantly improving the stacking distribution.

[0069] S1072. If the monitoring indicators show that the height standard deviation is lower than the preset threshold and the area uniformity is higher than the target value, the grain flow trajectory sequence is collected through four cameras and a diffusion pattern database is constructed using a convolutional neural network to record the optimization results.

[0070] In the embodiment of the present invention, when the height standard deviation is lower than 0.2 m and the uniformity exceeds 0.85, it is determined that the stacking is improved. The four cameras take pictures from multiple angles at 50 frames per second, and after splicing, a complete trajectory sequence is formed, covering the whole process from falling to stacking. The convolutional neural network adopts a three-layer structure. The first layer extracts morphological features, the second layer analyzes the movement trajectory, and the third layer fuses spatio-temporal information. After classification by the fully connected layer, a diffusion pattern database is generated. The database stores the optimized data in a structured manner, providing a reference basis for subsequent regulation.

[0071] In practical applications, after a certain grain depot is adjusted, the height standard deviation drops from 0.25 m to 0.15 m, and the density gradient smoothness increases by about 15%. The trajectory sequence shows that the grain flow diffuses more evenly. The pattern data recorded in the database supports the continuous iteration of parameter optimization, ensuring the stacking stability during long-term operation. This step, through the cooperation of sensors and algorithms, fully realizes the effect evaluation and experience accumulation from scanning to analysis and then to recording, and the threshold or grid density can be adjusted according to the scenario to optimize the monitoring accuracy.

[0072] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these changes and modifications.

Claims

1. A dynamic monitoring method for the shape of a grain pile based on an intelligent vision system, characterized in that, The method includes: Using the optical flow field analysis method to extract the velocity field and direction field of the grain flow movement from the video data during the falling process of the grain flow, and obtaining the initial information of the grain flow trajectory; According to the velocity field and direction field of the grain flow movement, if the change rate of the velocity field or direction field in a local area exceeds the preset change rate threshold, it is determined that there is a mutation in the grain flow trajectory, and the area where the mutation exists is marked as the trajectory mutation area; Using the local optical flow field analysis method to refine the velocity field and direction field of the trajectory mutation area, and obtaining the trajectory change information, where the trajectory change information includes the change amplitude, change direction, change duration, and change area range; According to the trajectory change information, analyze the initial diffusion range of the grain flow at the bottom of the bin, and judge whether the diffusion mode is uniform. If the diffusion is not uniform, mark the corresponding area as the diffusion irregular area; Based on the trajectory change information, establish a dynamic diffusion model, use the dynamic diffusion model to process the diffusion irregular area, and combine the velocity field and direction field of the grain flow movement to predict the stacking speed and morphological change trend of the grain pile; According to the predicted stacking speed and morphological change trend of the grain pile, adjust the operating parameters of the conveying equipment, where the operating parameters include the conveyor belt speed and angle; Real-time monitor the grain flow trajectory and diffusion mode of the adjusted conveying equipment, and judge whether the standard deviation of the grain pile height, the uniformity of the coverage area, and the distribution of the stacking density of the grain pile are improved. If they are improved, record the final grain flow trajectory and diffusion mode data.

2. The method according to claim 1, wherein The step of using the optical flow field analysis method to extract the velocity field and direction field of the grain flow movement from the video data during the falling process of the grain flow, and obtaining the initial information of the grain flow trajectory includes: Using the region matching method to identify the edge contour region of the grain flow, and obtaining the grain flow region grid cell division data and the grain flow density distribution map; Establish the gray value gradient constraint equation and the brightness constancy constraint equation according to the grain flow density distribution map, and solve the gray value gradient constraint equation and the brightness constancy constraint equation by the multi-scale iteration method to obtain the grain flow optical flow vector field distribution data; Calculate the grain flow velocity vector within the grid cell for the grain flow optical flow vector field distribution data, and use the least squares method to fit the velocity vector field to obtain the overall movement trend data of the grain flow; Construct a quadratic polynomial fitting function according to the overall movement trend data of the grain flow, and calculate the geometric parameters of the grain flow area and the diffusion radius from the function coefficients; Use the Euclidean distance calculation method to compare the change of geometric parameters between the front and back frames to obtain the grain flow movement parameters, and obtain the initial information of the grain flow falling trajectory.

3. The method according to claim 1, wherein The step of according to the velocity field and direction field of the grain flow movement, if the change rate of the velocity field or direction field in a local area exceeds the preset change rate threshold, it is determined that there is a mutation in the grain flow trajectory, and the area where the mutation exists is marked as the trajectory mutation area includes: Uniformly divide the grain flow area according to the grain flow velocity field data, use the Sobel operator to calculate the first-order difference values of the velocity data points in the horizontal and vertical directions within the grid cell, and obtain the velocity field change rate matrix from the first-order difference values; For the direction vector set within the grid cell, use the three-point fitting method to calculate the included angle change amount between adjacent direction vectors, and obtain the direction field change rate matrix from the included angle change amount; Perform double-threshold detection on the velocity field change rate matrix and the direction field change rate matrix. If the velocity field change rate exceeds the preset velocity threshold or the direction field change rate exceeds the preset direction threshold, mark the grid cell as a mutation candidate point; Use the region growing algorithm to perform connectivity analysis on the mutation candidate points, obtain the region boundary point sequence through the boundary tracking algorithm, and construct a piecewise cubic spline curve through the boundary point sequence to obtain the trajectory mutation region.

4. The method according to claim 1, wherein Adopt the local optical flow field analysis method to refine the velocity field and direction field of the trajectory mutation region to obtain trajectory change information, where the trajectory change information includes the change amplitude, change direction, change duration, and change region range, including: Perform local grid division on the range of the trajectory mutation region, and use the pyramid optical flow algorithm to obtain the velocity vector and direction vector sequences from the divided grids; Fit the velocity change curve using the least squares method of cubic polynomials according to the velocity vector sequence, and calculate the velocity difference between adjacent frames through the velocity change curve to obtain the velocity change amplitude parameter and the direction change angle parameter; Use the weighted average method to calculate the velocity change amplitude parameter and the direction change angle parameter, and obtain the motion change intensity value from the calculation results; According to the motion change intensity value, use the region centroid calculation method to obtain the coordinate sequence of the center point of the mutation region, and construct the region boundary contour through the coordinate sequence to obtain the trajectory change feature vector, which is used to characterize the trajectory change information.

5. The method according to claim 1, characterized in that, Based on the trajectory change information, analyze the initial diffusion range of the grain flow at the bottom of the bin, and judge whether the diffusion mode is uniform. If the diffusion is not uniform, mark the corresponding region as an irregular diffusion region, including: Use the region growing algorithm based on eight-neighborhoods to segment the area covered by the grain flow at the bottom of the bin, and obtain the boundary point sequence of the initial diffusion range at the bottom of the bin by using the gray similarity as the growth condition; Divide uniform grids of pixel size according to the boundary point sequence of the initial diffusion range at the bottom of the bin, and use the pixel counting method to obtain the number of grain flow pixels in the grid cell to get the volume density value; Perform maximum-minimum normalization processing on the volume density value of the grid cell, and obtain the diffusion intensity value through the linear mapping method to construct the diffusion intensity distribution map; If the diffusion irregularity parameter in the diffusion intensity distribution map exceeds the preset uniformity threshold, use the connected component labeling algorithm to label the region where the diffusion intensity gradient is greater than the gradient mean value to obtain the boundary of the irregular diffusion region.

6. The method according to claim 1, characterized in that, Establish a dynamic diffusion model based on the trajectory change information, use the dynamic diffusion model to process the irregular diffusion region, and combine the velocity field and direction field of the grain flow movement to predict the stacking speed and morphological change trend of the grain pile, including: Train a two-layer feedforward neural network model according to the historical trajectory data, and obtain the grain diffusion rate parameter and the diffusion direction parameter from the neural network model; For the diffusion rate parameter and the diffusion direction parameter, use uniform grid division for the irregular region, and calculate the stacking density change rate data from the grid cells; According to the stacking density change rate data, reconstruct the grid cell density distribution by the bilinear interpolation method, and construct the grain pile height field function from the density distribution; For the grain pile height field function, perform a spatial integration operation to obtain a function of the change in grain pile volume over time, and take the derivative of the change function to obtain a predicted value of the stacking speed.

7. The method according to claim 1, characterized in that, Adjust the operating parameters of the conveying equipment according to the predicted grain pile stacking speed and morphological change trend. The operating parameters include the conveyor belt speed and angle, and it includes: According to the stacking speed and morphological feature data, use an adaptive control algorithm to establish a conveyor belt parameter response function, and obtain the speed adjustment amount and angle adjustment amount from the response function; For the speed adjustment amount and angle adjustment amount, use a fuzzy rule inference engine to perform parameter linkage analysis. The fuzzy rule inference engine divides the speed variable into a low-speed fuzzy set, a medium-speed fuzzy set, and a high-speed fuzzy set, and divides the angle variable into a small-angle fuzzy set, a medium-angle fuzzy set, and a large-angle fuzzy set; According to the fuzzy inference result, use the centroid method for defuzzification, obtain the speed adjustment step size and angle adjustment step size from the defuzzification result, and construct a parameter adjustment path curve; For the parameter adjustment path curve, use a piecewise linear mapping method to discretize the belt speed increment sequence and the angle increment sequence, and obtain a conveyor belt control instruction sequence from the discretization result.

8. The method according to claim 1, wherein Real-time monitor the grain flow trajectory and diffusion pattern of the adjusted conveying equipment, and judge whether the standard deviation of the grain pile height, the uniformity of the coverage area, and the distribution of the stacking density of the grain pile are improved. If it is improved, record the final grain flow trajectory and diffusion pattern data, including: Use an infrared height measurement sensor array to perform a grid scan on the surface of the grain pile, construct a three-dimensional height distribution map of the grain pile through the scanned data, and obtain the height data of the measurement points from the three-dimensional height distribution map of the grain pile; According to the three-dimensional height distribution map of the grain pile, use a region growing algorithm based on height gradient to segment the coverage area of the grain pile, and obtain the contour line of the coverage area through region merging; For the contour line of the coverage area, divide the area using a regular hexagon grid and perform density scanning using a laser density detector array, and obtain a stacking density distribution map from the scanned data; If the height standard deviation index is lower than the preset standard deviation threshold and the area uniformity parameter is higher than the preset uniformity threshold, use a four-way camera to obtain a trajectory sequence, and construct a diffusion pattern database through the trajectory sequence using a convolutional neural network.

Images