Material pile boundary automatic detection system and method of unattended bucket-wheel stacker-reclaimer

Abstract

The invention relates to the technical field of material pile boundary detection, and particularly discloses an automatic material pile boundary detection system and method for an unattended bucket-wheel stacker-reclaimer. Radar equipment is used for scanning a target material pile to obtain point cloud data of the target material pile, and after coordinate system transformation and filtering processing are carried out on the point cloud data, the target material pile boundary is detected; carrying out local window-based geometric feature extraction on the preprocessed point cloud data by adopting a deep learning algorithm so as to mine local shape information of the target material pile, and carrying out core information-based aggregation analysis on geometric features of the target material pile under each local window; the global shape features and boundary information of the target material pile are captured, and automatic detection of the boundary of the target material pile is achieved through the semantic segmentation technology. By means of the mode, the defects of a traditional radar data processing mode can be overcome, accurate detection of material pile boundary information is achieved, and powerful support is provided for automatic operation of the bucket-wheel stacker reclaimer.

Description

Technical Field

[0001] This application relates to the technical field of stockpile boundary detection, and specifically relates to an automatic detection system and method for the stockpile boundary of an unattended bucket wheel stacker-reclaimer. Background Art

[0002] In modern industrial production, especially in the storage and handling fields of bulk materials such as coal and ore, the bucket wheel stacker-reclaimer is an important large-scale mechanical equipment, which is widely used in places such as ports, mines, and power plants. Traditional bucket wheel stacker-reclaimers mainly rely on manual operation. Operators need to constantly pay attention to the state of the stockpile and the operation of the equipment to ensure the smooth progress of the reclaiming and stacking operations. However, this manual operation method has many drawbacks. On the one hand, the working environment of operators is often relatively harsh, with serious pollution such as dust and noise, which poses a threat to the physical health of operators; on the other hand, the efficiency and accuracy of manual operation largely depend on the experience and skill level of operators, and it is easy to make operation mistakes, resulting in low production efficiency and difficult to meet the production requirements of large scale and high efficiency.

[0003] With the continuous development of industrial automation and intelligent technologies, the unattended bucket wheel stacker-reclaimer has emerged. The unattended bucket wheel stacker-reclaimer can realize automatic stacking and reclaiming operations, reduce manual intervention, and improve production efficiency and safety. During the operation of the unattended bucket wheel stacker-reclaimer, accurately detecting the stockpile boundary is one of the key technologies for realizing automatic operation. Accurate detection of the stockpile boundary can help the equipment accurately control the position and movement trajectory of the bucket wheel, avoid situations such as over-reclaiming or insufficient stacking, thereby improving the utilization rate of materials and reducing production costs.

[0004] Currently, there are various methods for stockpile boundary detection, such as detection methods based on visual images, detection methods based on laser scanning, and detection methods based on radar scanning, etc. Although the method based on laser scanning has high accuracy, the equipment cost is expensive, the scanning range is limited, and it is easily affected by environmental factors such as bad weather and dust; the method based on visual images has problems such as being sensitive to lighting conditions and the image being easily blocked, resulting in difficulties in ensuring the accuracy and reliability of the detection results; while the detection method based on radar scanning has a large scanning range and strong environmental adaptability, but the processing and analysis of radar data are relatively complex. Traditional radar data processing methods often rely on simple geometric rules or threshold settings, and it is difficult to accurately capture the complex and variable stockpile shape and fine boundary information, which may lead to insufficient accuracy of the stockpile boundary detection results.

[0005] Therefore, an optimized automatic detection system and method for the stockpile boundary of an unattended bucket wheel stacker-reclaimer are expected. Summary of the Invention

[0006] The embodiments of the present application aim to solve at least one of the technical problems existing in the prior art, and provide a system and method for automatically detecting the boundary of a stockpile of an unmanned bucket wheel stacker and reclaimer, which uses radar equipment to scan the target stockpile to obtain point cloud data of the target stockpile, and after coordinate system transformation and filtering of the point cloud data, uses a deep learning algorithm to extract geometric features based on local windows from the preprocessed point cloud data to mine the local shape information of the target stockpile, and then performs aggregation analysis based on core information on the geometric features of the target stockpile under each local window to capture the global shape features and boundary information of the target stockpile, thereby realizing automatic detection of the boundary of the target stockpile through semantic segmentation technology. In this way, the shortcomings of the traditional radar data processing method can be overcome, and accurate detection of the boundary information of the stockpile can be realized, providing strong support for the automated operation of the bucket wheel stacker and reclaimer.

[0007] On the one hand, an embodiment of the present application provides a method for automatically detecting a pile boundary of an unmanned bucket wheel stacker and reclaimer, comprising:

[0008] Use radar equipment to scan the target stockpile to obtain original point cloud data;

[0009] Performing data preprocessing on the original point cloud data to obtain preprocessed point cloud data;

[0010] Performing local window-based geometric feature extraction on the preprocessed point cloud data to obtain a set of window geometric feature semantic encoding feature matrices;

[0011] Performing a stockpile geometric feature aggregation analysis based on core information aggregation on the set of the window geometric feature semantic coding feature matrices to obtain a target stockpile geometric feature aggregation coding matrix;

[0012] The target stockpile geometric feature aggregation coding matrix is ​​semantically segmented to obtain a stockpile boundary recognition result.

[0013] Optionally, each data point in the original point cloud data is distance-angle information of the surface of the stockpile in a polar coordinate system.

[0014] Optionally, performing data preprocessing on the original point cloud data to obtain preprocessed point cloud data includes:

[0015] Converting the original point cloud data into a material field coordinate system to obtain coordinate system transformed point cloud data;

[0016] Mean filtering and / or Gaussian filtering are performed on the coordinate system transformed point cloud data to obtain the preprocessed point cloud data.

[0017] Optionally, perform geometric feature extraction based on a local window on the preprocessed point cloud data to obtain a set of window geometric feature semantic encoding feature matrices, including:

[0018] Divide the preprocessed point cloud data into a number of windows to obtain a set of window point cloud data;

[0019] Perform two-dimensional convolutional encoding on each window point cloud data in the set of window point cloud data to obtain the set of window geometric feature semantic encoding feature matrices.

[0020] Optionally, perform stockpile geometric feature aggregation analysis based on core information aggregation on the set of window geometric feature semantic encoding feature matrices to obtain a target stockpile geometric feature aggregation encoding matrix, including:

[0021] Input the set of window geometric feature semantic encoding feature matrices into an information kernel coarse-grained aggregation network to obtain a target stockpile geometric feature coarse-grained semantic aggregation encoding matrix;

[0022] Calculate the kernel aggregation compensation factor of each window geometric feature semantic encoding feature matrix in the set of window geometric feature semantic encoding feature matrices relative to the target stockpile geometric feature coarse-grained semantic aggregation encoding matrix to obtain a set of target stockpile geometric feature semantic kernel aggregation compensation factors;

[0023] Based on the set of target stockpile geometric feature semantic kernel aggregation compensation factors, perform semantic compensation aggregation encoding on the set of window geometric feature semantic encoding feature matrices and the target stockpile geometric feature coarse-grained semantic aggregation encoding matrix to obtain the target stockpile geometric feature aggregation encoding matrix.

[0024] Optionally, calculate the kernel aggregation compensation factor of each window geometric feature semantic encoding feature matrix in the set of window geometric feature semantic encoding feature matrices relative to the target stockpile geometric feature coarse-grained semantic aggregation encoding matrix to obtain a set of target stockpile geometric feature semantic kernel aggregation compensation factors, including:

[0025] Perform point convolutional encoding based on the Sigmoid activation function on the window geometric feature semantic encoding feature matrix and the target stockpile geometric feature coarse-grained semantic aggregation encoding matrix respectively to obtain a normalized window geometric feature semantic encoding feature matrix and a normalized target stockpile geometric feature coarse-grained semantic aggregation encoding matrix;

[0026] Calculate the position-wise difference matrix between the normalized window geometric feature semantic encoding feature matrix and the normalized target stockpile geometric feature coarse-grained semantic aggregation encoding matrix, and take the absolute value of the position-wise difference matrix to obtain a target stockpile geometric feature semantic kernel aggregation difference compensation encoding matrix;

[0027] Input the semantic kernel convergence difference compensation coding matrix of the target stockpile geometric features into a compensation feature importance scoring module based on a neural network layer to obtain the semantic kernel convergence compensation factor of the target stockpile geometric features.

[0028] Optionally, inputting the semantic kernel convergence difference compensation coding matrix of the target stockpile geometric features into a compensation feature importance scoring module based on a neural network layer to obtain the semantic kernel convergence compensation factor of the target stockpile geometric features includes:

[0029] Multiply the semantic kernel convergence difference compensation coding matrix of the target stockpile geometric features by a weight parameter vector, and add the multiplication result to a bias term to obtain a semantic kernel convergence compensation feature modulation vector of the target stockpile geometric features;

[0030] Multiply the semantic kernel convergence compensation feature modulation vector of the target stockpile geometric features by a semantic compensation feature importance scoring conversion vector of the target stockpile geometric features to obtain the semantic kernel convergence compensation factor of the target stockpile geometric features.

[0031] Optionally, multiplying the semantic kernel convergence difference compensation coding matrix of the target stockpile geometric features by a weight parameter vector, and adding the multiplication result to a bias term to obtain a semantic kernel convergence compensation feature modulation vector of the target stockpile geometric features includes:

[0032] Calculate the ratio between the Euclidean norm of the window geometric feature semantic coding feature matrix and the Euclidean norm of the target stockpile geometric feature coarse-grained semantic convergence coding matrix. If the ratio is less than 1, add 1 to the ratio and then calculate the logarithm to the base 2 as the bias term;

[0033] If the ratio is greater than or equal to 1, use the ratio as the bias term.

[0034] Optionally, based on the set of semantic kernel convergence compensation factors of the target stockpile geometric features, perform semantic compensation convergence coding on the set of window geometric feature semantic coding feature matrices and the target stockpile geometric feature coarse-grained semantic convergence coding matrix to obtain the target stockpile geometric feature aggregation coding matrix, including:

[0035] Perform compensation explicit modeling based on a gating function on the set of semantic kernel convergence compensation factors of the target stockpile geometric features to obtain a set of semantic kernel convergence compensation weight factors of the target stockpile geometric features;

[0036] Inputting the set of the target stockpile geometric feature semantic core convergence compensation weight factors, the target stockpile geometric feature coarse-grained semantic convergence coding matrix and the window geometric feature semantic coding feature matrix into a node fine-grained dynamic compensation convergence network to obtain the target stockpile geometric feature fine-grained semantic compensation convergence coding matrix;

[0037] The target stockpile geometric feature fine-grained semantic compensation aggregation coding matrix and the target stockpile geometric feature coarse-grained semantic aggregation coding matrix are input into a residual unit to obtain the target stockpile geometric feature aggregation coding matrix.

[0038] On the other hand, the present application provides an automatic detection system for a stockpile boundary of an unmanned bucket wheel stacker and reclaimer, comprising:

[0039] The original point cloud data acquisition module is used to scan the target stockpile using radar equipment to obtain original point cloud data;

[0040] A data preprocessing module, used for performing data preprocessing on the original point cloud data to obtain preprocessed point cloud data;

[0041] A geometric feature extraction module, used for performing geometric feature extraction based on a local window on the preprocessed point cloud data to obtain a set of window geometric feature semantic encoding feature matrices;

[0042] A feature aggregation analysis module, used for performing a stockpile geometric feature aggregation analysis based on core information aggregation on the set of the window geometric feature semantic coding feature matrices to obtain a target stockpile geometric feature aggregation coding matrix;

[0043] The semantic segmentation module is used to perform semantic segmentation on the target stockpile geometric feature aggregation coding matrix to obtain a stockpile boundary recognition result.

[0044] Compared with the prior art, the system and method for automatically detecting the boundary of a stockpile of an unmanned bucket wheel stacker and reclaimer provided in this application utilizes radar equipment to scan the target stockpile to obtain point cloud data of the target stockpile, and after coordinate system transformation and filtering of the point cloud data, a deep learning algorithm is used to extract geometric features based on local windows from the preprocessed point cloud data to mine the local shape information of the target stockpile, and then the global shape features and boundary information of the target stockpile are captured by performing aggregation analysis based on core information on the geometric features of the target stockpile under each local window, thereby realizing automatic detection of the boundary of the target stockpile through semantic segmentation technology. In this way, the shortcomings of the traditional radar data processing method can be overcome, and accurate detection of the boundary information of the stockpile can be realized, providing strong support for the automated operation of the bucket wheel stacker and reclaimer. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Flow chart of the method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application;

[0046] Figure 2 Schematic diagram of data flow of the method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application;

[0047] Figure 3 Flow chart of sub-step S2 of the method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application;

[0048] Figure 4 Flow chart of sub-step S3 of the method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application;

[0049] Figure 5 Flow chart of sub-step S4 of the method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application;

[0050] Figure 6 Block diagram of the system for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application. Detailed implementation manners

[0051] As shown in the present application and the claims, unless the context clearly indicates otherwise, words such as "a", "an", "one" and / or "the" are not specifically singular and may also include plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0052] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or the server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0053] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the operations before or below are not necessarily executed precisely in sequence. On the contrary, as needed, various steps can be executed in reverse order or simultaneously. At the same time, other operations can also be added to these processes, or one or several steps can be removed from these processes.

[0054] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0055] It is worth noting that in this application, all actions to obtain data are carried out in compliance with the relevant data protection laws and policies of the country where the data is located, and with the authorization given by the owner of the corresponding device.

[0056] In response to the technical problems described in the above background technology, this application proposes a method for automatically detecting the boundary of a stockpile of an unmanned bucket wheel stacker and reclaimer, which uses radar equipment to scan the target stockpile to obtain point cloud data of the target stockpile, and after coordinate system transformation and filtering of the point cloud data, a deep learning algorithm is used to extract geometric features based on local windows from the preprocessed point cloud data to mine the local shape information of the target stockpile, and then the global shape features and boundary information of the target stockpile are captured by performing aggregation analysis based on core information on the geometric features of the target stockpile under each local window, thereby realizing automatic detection of the boundary of the target stockpile through semantic segmentation technology. In this way, the shortcomings of traditional radar data processing methods can be overcome, accurate detection of stockpile boundary information can be achieved, and strong support can be provided for the automated operation of bucket wheel stackers and reclaimers.

[0057] Figure 1 The present invention is a flowchart of a method for automatically detecting a stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to an embodiment of the present application. Figure 1 and Figure 2 As shown, the method for automatically detecting the stockpile boundary of the unmanned bucket wheel stacker and reclaimer includes the following steps: S1, scanning the target stockpile using radar equipment to obtain original point cloud data. S2, performing data preprocessing on the original point cloud data to obtain preprocessed point cloud data. S3, performing local window-based geometric feature extraction on the preprocessed point cloud data to obtain a set of window geometric feature semantic coding feature matrices. S4, performing a stockpile geometric feature aggregation analysis based on core information convergence on the set of window geometric feature semantic coding feature matrices to obtain a target stockpile geometric feature aggregation coding matrix. S5, performing semantic segmentation on the target stockpile geometric feature aggregation coding matrix to obtain a stockpile boundary recognition result.

[0058] In the above method for automatically detecting the boundary of the stockpile of the unattended bucket wheel stacker-reclaimer, in step S1, a radar device is used to scan the target stockpile to obtain the original point cloud data. In a specific example of the present application, each data point in the original point cloud data is the distance-angle information of the stockpile surface in the polar coordinate system. It should be understood that the radar device emits electromagnetic waves and receives the electromagnetic waves reflected from the stockpile surface, and can calculate the distance between the radar and each point on the stockpile surface according to the propagation time and speed of the electromagnetic waves. At the same time, the radar rotates itself, and by combining the angle information obtained by the rotation angle sensor, the distance-angle information of the stockpile surface in the polar coordinate system can be obtained, generating the point cloud data reflecting the shape of the stockpile, thereby providing the basic data for the subsequent automatic detection of the stockpile boundary.

[0059] Specifically, the radar device can obtain the distance and angle information of the stockpile surface by emitting electromagnetic waves and receiving the reflected signals. These information are stored in the form of the polar coordinate system to form the original point cloud data. The data points in the polar coordinate system consist of two main parameters: distance and angle. The distance represents the straight-line distance between the radar device and a certain point on the stockpile surface, and the angle represents the azimuth angle of the point relative to the radar device. This data representation method can intuitively reflect the spatial distribution characteristics of the stockpile surface and provide the basis for subsequent data processing and analysis.

[0060] During the scanning process of the radar device, data is usually collected at a certain angular resolution and distance resolution. The angular resolution determines the scanning accuracy of the radar in the horizontal direction, while the distance resolution determines the measurement accuracy of the radar in the vertical direction. Higher resolution can capture more detailed information, but it will also increase the data volume and pose higher requirements for subsequent data processing. Therefore, in practical applications, it is necessary to reasonably set the scanning parameters of the radar device according to the requirements of the specific scenario to balance data accuracy and processing efficiency.

[0061] To ensure the stable operation of the radar device and obtain consistent and reliable original point cloud data, it is also necessary to pay attention to the influence of surrounding environmental factors. For example, dust, rain and snow may interfere with the propagation of radar signals and affect the quality of the final data. Therefore, taking appropriate protective measures, such as installing a dust cover or selecting a radar model with anti-interference ability, are all necessary steps. At the same time, considering that the change of light under different times and weather conditions may indirectly affect the working effect of the radar (although the radar mainly relies on electromagnetic waves rather than visible light), it is also necessary to calibrate and maintain the device regularly to ensure that it is always in the best working state.

[0062] In the above method for automatically detecting the boundary of the stockpile of the unattended bucket wheel stacker-reclaimer, in step S2, the original point cloud data is preprocessed to obtain the preprocessed point cloud data. Among them, Figure 3It is a flowchart of sub-step S2 of the automatic detection method for the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application. As Figure 3 shown, the step S2 includes steps: S21, transforming the original point cloud data into the stockyard coordinate system to obtain coordinate-transformed point cloud data. S22, performing mean filtering and / or Gaussian filtering on the coordinate-transformed point cloud data to obtain the preprocessed point cloud data.

[0063] Specifically, in the step S21, the original point cloud data is transformed into the stockyard coordinate system to obtain coordinate-transformed point cloud data. Specifically, the present application takes into account that the original point cloud data is obtained based on the polar coordinate system of the radar itself, and different radar devices may have different polar coordinate system settings. Moreover, there may be multiple radars or other devices that need to work together in the stockyard. The data in the polar coordinate system lacks generality under different radar devices or different scanning angles, which is not conducive to subsequent unified analysis and processing. Therefore, in order to enable all data to be processed and analyzed under a unified standard, the present application further transforms the original point cloud data into a common stockyard coordinate system to ensure the accuracy and consistency of the data, and obtains coordinate-transformed point cloud data. In the embodiment of the present application, the stockyard coordinate system is a Cartesian coordinate system. This conversion process is based on basic trigonometric relationships, that is, converting the distance and angle information of each data point into the corresponding X, Y, and Z coordinate values, so as to obtain the new coordinate position of the point in the stockyard coordinate system.

[0064] During the entire coordinate transformation process, it is crucial to ensure data accuracy. For this purpose, the issue of numerical stability also needs to be considered. For example, rounding errors may occur during floating-point operations, affecting the final coordinate transformation result. To avoid this situation, a higher-precision data type can be selected or a specific numerical algorithm can be adopted to reduce the impact of errors. In addition, for large-scale point cloud data sets, directly transforming all points one by one may consume a large amount of time and computing resources. In this case, a block processing method can be adopted, that is, the entire point cloud data is divided into several small blocks, and the coordinate transformation is performed separately and then the results are merged. This can not only improve the processing speed but also effectively manage the memory usage.

[0065] During the implementation of the coordinate transformation, verifying the accuracy of the transformation result cannot be ignored either. A commonly used method is to select some points with obvious features as reference points and compare whether the coordinate value changes of these points in the radar coordinate system and the stockyard coordinate system meet the expectations. If a large deviation is found, it is necessary to recheck the calculation process of the coordinate transformation matrix or evaluate the accuracy of the hardware device used. In addition, visualization tools can be used to intuitively display the point cloud data before and after the coordinate transformation, and observe the change trend of the overall shape and structure to judge the quality of the coordinate transformation.

[0066] Specifically, in step S22, mean filtering and / or Gaussian filtering are performed on the coordinate system-transformed point cloud data to obtain the preprocessed point cloud data. It should be understood that during the acquisition process of the point cloud data, due to the accuracy limitations of the radar device itself, environmental factors (such as electromagnetic interference, etc.), and errors during signal transmission, noise points will inevitably be introduced. These noise points may appear as isolated outliers, with obvious differences in spatial distribution from the surrounding normal point cloud data, and may interfere with subsequent geometric analysis and boundary detection, resulting in inaccurate results. Therefore, in order to improve the accuracy and reliability of the point cloud data, the present application further performs filtering processing on the point cloud data after coordinate system transformation. Specifically, mean filtering and / or Gaussian filtering methods are used to smooth each point in the point cloud data to remove or weaken the interference of noise points. Mean filtering calculates the average value of points within a certain range around each point and uses this average value as the filtering result of this point, thereby achieving smoothing processing. Gaussian filtering, according to the shape of the Gaussian function, performs weighted averaging on points within a certain range around each point, and the weights are determined by the Gaussian function. The points closer to the center have larger weights, thereby achieving smoothing processing. By mean filtering and / or Gaussian filtering, the noise points in the point cloud data can be effectively removed or weakened, improving the accuracy and reliability of the data, and providing more accurate basic data for subsequent geometric feature extraction and boundary detection.

[0067] In the above method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer, in step S3, geometric feature extraction based on a local window is performed on the preprocessed point cloud data to obtain a set of window geometric feature semantic encoding feature matrices. Among them, Figure 4 is a flowchart of sub-step S3 of the method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application. As Figure 4 shown, step S3 includes steps: S31, dividing the preprocessed point cloud data into several windows to obtain a set of window point cloud data. S32, performing two-dimensional convolutional encoding on each window point cloud data in the set of window point cloud data to obtain the set of window geometric feature semantic encoding feature matrices.

[0068] Specifically, in step S31, the preprocessed point cloud data is divided into several windows to obtain a set of window point cloud data. It should be understood that due to the complex and variable shape of the stockpile, the stockpiles at different positions may have different shapes. For example, the geometric features at the top and bottom of the stockpile may vary greatly. Directly extracting geometric features and detecting boundaries for the entire stockpile may face a large computational burden and accuracy loss. Therefore, to reduce the computational complexity and improve the detection accuracy, the preprocessed point cloud data in this application is divided into several local windows. Each window contains a part of the point cloud data, representing a certain local area of the stockpile. In this way, complex global problems can be transformed into multiple relatively simple local problems for processing. In the embodiment of this application, first, the size of the window is set according to factors such as the size of the stockpile, the density of the point cloud data, and the scanning range of the radar device, to ensure that each window can contain a sufficient number of point cloud data to reflect the local geometric features of the stockpile. Then, according to a preset division rule, such as grid division, overlapping division, etc., the preprocessed point cloud data is spatially divided into several independent or partially overlapping sub-regions to obtain a set of window point cloud data. In this way, fine analysis of different local areas of the stockpile can be realized, and the accuracy and reliability of boundary detection can be improved.

[0069] Specifically, the window size directly affects the accuracy of the final analysis result and the demand for computing resources. Smaller windows can capture finer details, but at the same time, they will increase the computational amount. Larger windows may lose some important local information. Therefore, determining an appropriate window size according to the characteristics of the stockpile in the actual application scenario and the required analysis accuracy becomes the primary task. For example, when dealing with large-scale stockpiles with relatively flat and slowly changing surfaces, larger-sized windows can be selected to reduce the computational burden. While in the face of areas with complex shapes and rich details, the window size needs to be reduced to better capture local features.

[0070] After determining the window size, the next step is to consider how to reasonably arrange these windows over the entire point cloud dataset. A common approach is to use a regular grid partitioning method, that is, to establish a uniformly distributed grid system within the entire point cloud space, with each grid cell corresponding to a window. The advantage of this method is that it is simple and easy to implement, facilitating automated processing, and can ensure seamless connection between windows, avoiding omission of any part of the data. However, this approach may also have limitations. Especially when dealing with non-uniformly distributed point cloud data, some regions may not be fully analyzed due to low point density, or conversely, excessive redundant calculations may occur in high-density regions. To overcome these problems, an adaptive window partitioning strategy can be adopted, that is, dynamically adjusting the window size or position according to the actual distribution of the point cloud data. By analyzing the density and distribution characteristics of each part of the point cloud data, smaller windows are used for dense regions to improve the ability to capture details, while larger windows are used for sparse regions to save computational resources.

[0071] In addition to the above-mentioned methods based on regular grids and adaptivity, a dedicated window partitioning scheme can also be designed in combination with the specific characteristics of the stockpile shape. For example, when dealing with a stockpile with an obvious hierarchical structure or a specific geometric shape, the window boundaries can be predefined according to the known shape characteristics, so that each window can specifically cover a certain specific area or structure. This not only improves the accuracy of feature extraction but also helps to simplify the subsequent data analysis process. In addition, considering the possible data loss or outliers during the radar scanning process, a reasonable window partitioning should also include a certain fault tolerance mechanism, such as reserving overlapping regions or cross-verifying part of the data of adjacent windows, so as to enhance the robustness of the overall analysis results.

[0072] During the implementation of window partitioning, attention should also be paid to maintaining the consistency and continuity of the data. This means that when dividing different windows, the data points within each window should be ensured to fully reflect the real situation in this area as much as possible, avoiding information fragmentation caused by artificial division. To this end, the point cloud data can be appropriately pre-organized before partitioning, such as sorting or clustering according to spatial positions, for subsequent window allocation. At the same time, for the convenience of management and retrieval, a unique identifier is usually assigned to each window, and its corresponding coordinate range or other relevant information is recorded. This approach is beneficial for quickly locating the data within a specific window in subsequent steps and also facilitates the integration and comparative analysis of the results of multiple windows.

[0073] To ensure the quality of window division, regular evaluation of the division effect is also an essential part. This can be carried out in various ways, such as checking whether the number of data points within each window after division meets the expectations, verifying whether there are obvious data breaks at the window boundaries, and using visualization tools to intuitively display the division results, etc. If any unreasonable division is found, the window parameters should be adjusted in a timely manner or the division strategy should be re-planned. It should be noted that during the adjustment process, the balance between computational efficiency and analysis accuracy should be taken into account to avoid excessive optimization resulting in too high computational costs or too long processing times.

[0074] Specifically, in step S32, two-dimensional convolutional encoding is performed on each piece of window point cloud data in the set of window point cloud data to obtain a set of window geometric feature semantic encoding feature matrices. Specifically, this application takes into account that discrete point cloud data only contains the geometric position information on the surface of the stockpile, but lacks an in-depth description of its geometric shape and structure and is difficult to be directly used for boundary detection. Therefore, in order to extract richer geometric feature descriptions from the point cloud data, this application uses the method of two-dimensional convolutional encoding to process each piece of window point cloud data respectively to extract its geometric features. Specifically, first, the window point cloud data is projected onto a two-dimensional plane to form a bird's eye view (BEV) to retain the spatial relationship of the original point cloud and at the same time convert it into an image form suitable for two-dimensional convolutional processing. In the bird's eye view, the x and y axes respectively represent the lateral and longitudinal positions of the stockpile on the horizontal plane, and the information of the z axis can be represented by different gray values to reflect the height information of the stockpile surface. Then, a two-dimensional convolutional neural network is used to extract features from the converted image. Those of ordinary skill in the art should know that the two-dimensional convolutional neural network performs sliding convolutional operations on the window point cloud data by using two-dimensional convolutional kernels to perceive the distribution pattern of points, and can effectively capture the local area geometric features of the stockpile, such as straight lines, curves, corner points, etc., so as to generate a set of window geometric feature semantic encoding feature matrices that accurately describe the geometric shape and structure of each local area of the target stockpile.

[0075] In the above method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer, in step S4, an aggregation analysis of the stockpile geometric features based on core information aggregation is performed on the set of window geometric feature semantic encoding feature matrices to obtain a target stockpile geometric feature aggregation encoding matrix. It should be understood that since the window geometric feature semantic encoding feature matrix only describes the geometric features of a certain local area of the stockpile, in order to obtain a comprehensive understanding of the shape and structure of the entire stockpile, the present application proposes a method for aggregating and analyzing the stockpile geometric features based on core information aggregation. By performing coarse-grained core information aggregation and fine-grained local feature compensation on the set of window geometric feature semantic encoding feature matrices, the local geometric features of each window are effectively integrated to achieve a comprehensive description of the overall geometric shape and structure of the stockpile. Among them, Figure 5 is a flowchart of sub-step S4 of the method for automatically detecting the stockpile boundary of an unattended bucket wheel stacker-reclaimer according to an embodiment of the present application. As Figure 5 shown, step S4 includes steps: S41, inputting the set of window geometric feature semantic encoding feature matrices into an information core coarse-grained aggregation network to obtain a target stockpile geometric feature coarse-grained semantic aggregation encoding matrix. S42, calculating the core aggregation compensation factor of each window geometric feature semantic encoding feature matrix in the set of window geometric feature semantic encoding feature matrices relative to the target stockpile geometric feature coarse-grained semantic aggregation encoding matrix to obtain a set of target stockpile geometric feature semantic core aggregation compensation factors. S43, based on the set of target stockpile geometric feature semantic core aggregation compensation factors, performing semantic compensation aggregation encoding on the set of window geometric feature semantic encoding feature matrices and the target stockpile geometric feature coarse-grained semantic aggregation encoding matrix to obtain the target stockpile geometric feature aggregation encoding matrix.

[0076] Specifically, step S41 is represented by the formula:

[0077]

[0078] where X represents the set of window geometric feature semantic encoding feature matrices, x1, x2, x i and x n respectively represent the 1st, 2nd, i-th, and n-th window geometric feature semantic encoding feature matrices in the set of window geometric feature semantic encoding feature matrices, n is the number of matrices in the set of window geometric feature semantic encoding feature matrices, max(·) and min(·) respectively represent taking the maximum value and the minimum value, e i represents the median of the feature distribution boundary of x i , softmax(·) represents the normalized exponential function, a i represents x iThe attention weight, Concat{·} represents feature concatenation, v h represents the coarse-grained semantic aggregation encoding matrix of the geometric features of the target stockpile.

[0079] That is, construct an information core coarse-grained aggregation network to process the set of the window geometric feature semantic encoding feature matrices, calculate their global aggregation weights based on the statistical features of each window geometric feature semantic encoding feature matrix, and through the way of feature weighted concatenation aggregation, highlight the key features that have a greater impact on the shape and structure of the entire stockpile, and realize the coarse-grained aggregation encoding of each window geometric feature semantic encoding feature matrix, so as to obtain the coarse-grained semantic aggregation encoding matrix of the geometric features of the target stockpile.

[0080] Specifically, step S42 includes: performing point convolution encoding based on the Sigmoid activation function on the window geometric feature semantic encoding feature matrix and the coarse-grained semantic aggregation encoding matrix of the geometric features of the target stockpile respectively to obtain a normalized window geometric feature semantic encoding feature matrix and a normalized coarse-grained semantic aggregation encoding matrix of the geometric features of the target stockpile. Calculate the position-wise difference matrix between the normalized window geometric feature semantic encoding feature matrix and the normalized coarse-grained semantic aggregation encoding matrix of the geometric features of the target stockpile, and take the absolute value of the position-wise difference matrix to obtain the target stockpile geometric feature semantic kernel aggregation difference compensation encoding matrix. Input the target stockpile geometric feature semantic kernel aggregation difference compensation encoding matrix into a compensation feature importance scoring module based on a neural network layer to obtain the target stockpile geometric feature semantic kernel aggregation compensation factor. More specifically, inputting the target stockpile geometric feature semantic kernel aggregation difference compensation encoding matrix into a compensation feature importance scoring module based on a neural network layer to obtain the target stockpile geometric feature semantic kernel aggregation compensation factor includes: multiplying the target stockpile geometric feature semantic kernel aggregation difference compensation encoding matrix by a weight parameter vector, and adding the multiplication result with a bias term to obtain a target stockpile geometric feature semantic kernel aggregation compensation feature modulation vector. Multiply the target stockpile geometric feature semantic kernel aggregation compensation feature modulation vector by a target stockpile geometric feature semantic compensation feature importance scoring conversion vector to obtain the target stockpile geometric feature semantic kernel aggregation compensation factor.

[0081] The above step S42 is expressed by the formula as:

[0082]

[0083] where g(·,·) represents a compensation factor calculation network, Sigmoid(·) represents the Sigmoid activation function, conv 1×1 (·) represents a 1×1 convolution operation, W 11 and W 21respectively represent the weight parameter matrix of the geometric features of the target stockpile and the weight parameter matrix of the coarse-grained aggregation of the geometric features of the target stockpile, x i ′ represents the semantic encoding feature matrix of the geometric features of the standardized window, v coarse ′ represents the coarse-grained semantic aggregation encoding matrix of the geometric features of the standardized target stockpile, represents subtraction by position point, d i represents the semantic kernel aggregation difference compensation encoding matrix of the geometric features of the target stockpile, W i is the weight parameter vector of the compensation feature importance scoring module, b i represents the bias term, r i t represents the semantic compensation feature importance scoring conversion vector of the geometric features of the target stockpile, m i represents x i corresponding semantic kernel aggregation compensation factor of the geometric features of the target stockpile.

[0084] Here, the present application takes into account that although the above global aggregation method can capture the overall trend of the geometric features of the target stockpile, the semantic compression aggregation process of the global geometric features may dilute or even lose some important local detail features. To make up for this deficiency, the present application further introduces a local feature compensation mechanism, which dynamically generates the kernel aggregation compensation factor of each window geometric feature semantic encoding feature matrix by measuring the feature deviation between each window geometric feature semantic encoding feature matrix and the coarse-grained semantic aggregation encoding matrix of the target stockpile geometric features, so as to describe the personalized semantic offset of each window geometric feature relative to the overall geometric shape of the target stockpile, so as to finely restore and enhance the local detail information that may be diluted or lost in the global aggregation encoding process in the subsequent feature compensation process.

[0085] Specifically, in a preferred example of the present application, multiplying the weight parameter vector by the semantic kernel aggregation difference compensation encoding matrix of the geometric features of the target stockpile, and adding the multiplication result to the bias term to obtain the semantic kernel aggregation compensation feature modulation vector of the geometric features of the target stockpile, including: calculating the ratio between the Euclidean norm of the window geometric feature semantic encoding feature matrix and the Euclidean norm of the coarse-grained semantic aggregation encoding matrix of the target stockpile geometric features, if the ratio is less than 1, then adding 1 to the ratio and calculating the logarithm to the base 2 as the bias term. If the ratio is greater than or equal to 1, then taking the ratio as the bias term, which is expressed by the formula:

[0086]

[0087] where, ‖·‖2 represents calculating the Euclidean norm of the matrix, and log2(·) represents the logarithmic function to the base 2.

[0088] Here, for the deviation compensation between the semantic encoding feature matrix of the window geometric feature and the coarse-grained semantic aggregation encoding matrix of the target stockpile geometric feature, it can be measured by quantifying the regret metric based on the information kernel compression hypothesis in the kernel aggregation decision process, that is, the game-theoretic counterfactual regret value, to measure the performance deviation of the kernel aggregation strategy as a scenario strategy. Specifically, first, a normalized decision point loss description based on the policy action is provided for the counterfactual regret value through the matrix norm representation, that is, the matrix norm representation of the semantic encoding feature matrix of the window geometric feature and the coarse-grained semantic aggregation encoding matrix of the target stockpile geometric feature. Then, for the possible matrix distribution action game scenario differences, the compensation rule correction of the personalized information of the local window geometric feature is carried out respectively with the information distribution degree of the regret value and the relative distribution amplitude of the regret value, so as to consider the personalized information of the local window geometric feature of the target stockpile as an un-taken action in the decision-making, and perform bias compensation based on the way of assuming its potential benefit based on the information kernel aggregation hypothesis.

[0089] Specifically, in a specific example of the present application, the step S43 includes: first, performing compensation explicit modeling based on a gating function on the set of semantic kernel aggregation compensation factors of the target stockpile geometric feature to obtain a set of semantic kernel aggregation compensation weight factors of the target stockpile geometric feature, which is expressed by the formula:

[0090]

[0091] where ω represents the gating threshold, e is the natural constant, Gate(·) represents the gating function, and τ i represents the semantic kernel aggregation compensation weight factor of the target stockpile geometric feature corresponding to x i

[0092] That is, after obtaining the semantic kernel aggregation compensation factors of the target stockpile geometric feature, the present application further regulates the action size of each semantic kernel aggregation compensation factor of the target stockpile geometric feature through compensation explicit modeling based on the gating function, dynamically selects information based on the non-linear constraint, so as to screen the action intensity of each semantic kernel aggregation compensation factor of the target stockpile geometric feature, and generate a set of semantic kernel aggregation compensation weight factors of the target stockpile geometric feature, so as to ensure that important personalized local detail features are effectively compensated in the global feature aggregation process, while irrelevant or redundant information is suppressed.

[0093] Then, the set of semantic kernel aggregation compensation weight factors of the target stockpile geometric feature, the set of the coarse-grained semantic aggregation encoding matrix of the target stockpile geometric feature, and the set of the semantic encoding feature matrix of the window geometric feature are input into the node fine-grained dynamic compensation aggregation network to obtain the fine-grained semantic compensation aggregation encoding matrix of the target stockpile geometric feature, which is expressed by the formula: ​

[0094] v com = Concat{τ i ·(x i - v coarse )}

[0095] where v com represents the fine-grained semantic compensation convergence encoding matrix of the geometric features of the target stockpile.

[0096] That is, based on the obtained semantic kernel convergence compensation weight factor of the geometric features of the target stockpile, feature compensation aggregation encoding is performed on the difference information between the window geometric feature semantic encoding feature matrix and the coarse-grained semantic convergence encoding matrix of the target stockpile geometric features (i.e., the local detail information lost during the global aggregation process) to generate the fine-grained semantic compensation convergence encoding matrix of the target stockpile geometric features, thereby realizing the fine description and comprehensive integration of the personalized semantic information of each window geometric feature.

[0097] Finally, the fine-grained semantic compensation convergence encoding matrix of the geometric features of the target stockpile and the coarse-grained semantic convergence encoding matrix of the geometric features of the target stockpile are input into the residual unit to obtain the aggregated encoding matrix of the geometric features of the target stockpile, which is expressed by the formula:

[0098] v fine = α·v coarse + β·v com

[0099] where α and β represent different weight parameters, and v fine represents the aggregated encoding matrix of the geometric features of the target stockpile.

[0100] That is, through the residual unit, linear weighted aggregation is performed on the fine-grained semantic compensation convergence encoding matrix of the geometric features of the target stockpile and the coarse-grained semantic convergence encoding matrix of the geometric features of the target stockpile to combine the global semantic feature generalization expression and personalized semantic information description of the geometric features of the target stockpile, generate the aggregated encoding matrix of the geometric features of the target stockpile, and thereby comprehensively and accurately reflect the geometric shape and structure of the target stockpile. In this way, the aggregated encoding matrix of the geometric features of the target stockpile not only contains the overall shape information of the stockpile but also incorporates the detailed features of each local area, providing a solid foundation for the subsequent boundary recognition of the stockpile.

[0101] In the above-mentioned method for automatically detecting the stockpile boundary of the unattended bucket wheel stacker and reclaimer, the step S5 performs semantic segmentation on the target stockpile geometric feature aggregation coding matrix to obtain the stockpile boundary recognition result. It should be understood that semantic segmentation is the process of dividing an image into multiple semantic regions, each region corresponding to a specific category or object. In the present application, a semantic segmentation model based on a neural network is used to train it through a large amount of labeled stockpile and background sample data, and the model parameters are adjusted to achieve a better classification effect. Then, the target stockpile geometric feature aggregation coding matrix is ​​input into the trained semantic segmentation model, and each feature point in the target stockpile geometric feature aggregation coding matrix is ​​pixel-by-pixel classified according to the parameters and classification rules obtained by the training to determine the category (such as stockpile or background) to which each pixel belongs, thereby identifying and distinguishing the stockpile from the background area, and dividing the stockpile boundary in the matrix. In practical applications, the bucket wheel stacker and reclaimer can automatically plan the stacking and reclaiming paths according to the detected stockpile boundary information, thereby improving work efficiency and accuracy.

[0102] In summary, the method for automatic detection of the stockpile boundary of the unmanned bucket wheel stacker and reclaimer based on the embodiment of the present application is explained, which uses radar equipment to scan the target stockpile to obtain the point cloud data of the target stockpile, and after the coordinate system transformation and filtering of the point cloud data, a deep learning algorithm is used to extract the geometric features of the pre-processed point cloud data based on the local window to mine the local shape information of the target stockpile, and then the global shape features and boundary information of the target stockpile are captured by performing aggregation analysis based on the core information on the geometric features of the target stockpile under each local window, so as to realize the automatic detection of the target stockpile boundary through semantic segmentation technology. In this way, the shortcomings of the traditional radar data processing method can be overcome, and the accurate detection of the stockpile boundary information can be realized, providing strong support for the automated operation of the bucket wheel stacker and reclaimer.

[0103] Furthermore, the present application also provides an automatic detection system for the stockpile boundary of an unmanned bucket wheel stacker and reclaimer.

[0104] Figure 6 FIG. 1 is a block diagram of an automatic detection system for a pile boundary of an unmanned bucket wheel stacker and reclaimer according to an embodiment of the present application. Figure 6As shown, the automatic material pile boundary detection system 100 of the unattended bucket wheel stacker-reclaimer according to the embodiment of the present application includes: a raw point cloud data acquisition module 110, which is used to scan a target material pile with a radar device to obtain raw point cloud data. A data preprocessing module 120, which is used to perform data preprocessing on the raw point cloud data to obtain preprocessed point cloud data. A geometric feature extraction module 130, which is used to perform geometric feature extraction based on a local window on the preprocessed point cloud data to obtain a set of window geometric feature semantic coding feature matrices. A feature aggregation and analysis module 140, which is used to perform material pile geometric feature aggregation analysis based on core information aggregation on the set of window geometric feature semantic coding feature matrices to obtain a target material pile geometric feature aggregation coding matrix. A semantic segmentation module 150, which is used to perform semantic segmentation on the target material pile geometric feature aggregation coding matrix to obtain a material pile boundary recognition result.

[0105] Here, those skilled in the art can understand that the specific operations of each module in the above automatic material pile boundary detection system of the unattended bucket wheel stacker-reclaimer have been introduced in detail above with reference to Figures 1 to 5 the description of the automatic material pile boundary detection method of the unattended bucket wheel stacker-reclaimer, and therefore, the repeated description thereof will be omitted.

[0106] The basic principles of the present invention have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, advantages, effects, etc. mentioned in the present invention are only examples and not limitations, and it cannot be considered that these advantages, advantages, effects, etc. are essential for each embodiment of the present invention. In addition, the specific details of the above embodiments are only for the purpose of illustration and easy understanding, rather than limitations. The above details do not limit the present invention to necessarily adopt the above specific details to implement.

[0107] In the above embodiments, the descriptions of each embodiment have their own emphases. For parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiments described above are only illustrative. For example, the unit division is only a logical function division, and there may be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0108] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be construed as limiting the claimed invention.

[0109] In addition, it is obvious that the term "including" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units stated in the system claims can also be implemented by one unit through software or hardware.

[0110] Finally, it should be noted that the above description has been given for purposes of illustration and description. In addition, the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer, characterized in that: include: Use radar equipment to scan the target stockpile to obtain original point cloud data; Performing data preprocessing on the original point cloud data to obtain preprocessed point cloud data; Performing local window-based geometric feature extraction on the preprocessed point cloud data to obtain a set of window geometric feature semantic encoding feature matrices; Performing a stockpile geometric feature aggregation analysis based on core information aggregation on the set of the window geometric feature semantic coding feature matrices to obtain a target stockpile geometric feature aggregation coding matrix; The target stockpile geometric feature aggregation coding matrix is ​​semantically segmented to obtain a stockpile boundary recognition result.

2. The method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to claim 1, characterized in that: Each data point in the original point cloud data is the distance-angle information of the surface of the stockpile in the polar coordinate system.

3. The method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to claim 2, characterized in that: The raw point cloud data is preprocessed to obtain preprocessed point cloud data, including: Converting the original point cloud data into a material field coordinate system to obtain coordinate system transformed point cloud data; Mean filtering and / or Gaussian filtering are performed on the coordinate system transformed point cloud data to obtain the preprocessed point cloud data.

4. The method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to claim 3, characterized in that: The preprocessed point cloud data is subjected to local window-based geometric feature extraction to obtain a set of window geometric feature semantic encoding feature matrices, including: Dividing the preprocessed point cloud data into a plurality of windows to obtain a set of window point cloud data; Two-dimensional convolution coding is performed on each window point cloud data in the set of window point cloud data to obtain a set of window geometric feature semantic coding feature matrices.

5. The method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to claim 4, characterized in that: The set of the window geometric feature semantic coding feature matrices is subjected to a stockpile geometric feature aggregation analysis based on core information aggregation to obtain a target stockpile geometric feature aggregation coding matrix, including: Inputting the set of the window geometric feature semantic coding feature matrices into the information core coarse-grained aggregation network to obtain the target stockpile geometric feature coarse-grained semantic aggregation coding matrix; Calculating the core convergence compensation factor of each window geometric feature semantic coding feature matrix in the set of the window geometric feature semantic coding feature matrix relative to the target stockpile geometric feature coarse-grained semantic convergence coding matrix to obtain a set of target stockpile geometric feature semantic core convergence compensation factors; Based on the set of semantic kernel convergence compensation factors of the target stockpile geometric feature, the set of window geometric feature semantic coding feature matrices and the target stockpile geometric feature coarse-grained semantic convergence coding matrix are semantically compensated and converged to obtain the target stockpile geometric feature aggregation coding matrix.

6. The method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to claim 5, characterized in that: Calculating the core convergence compensation factor of each window geometric feature semantic coding feature matrix in the set of the window geometric feature semantic coding feature matrix relative to the target stockpile geometric feature coarse-grained semantic convergence coding matrix to obtain a set of target stockpile geometric feature semantic core convergence compensation factors, including: Performing point convolution coding based on Sigmoid activation function on the window geometric feature semantic coding feature matrix and the target stockpile geometric feature coarse-grained semantic aggregation coding matrix respectively to obtain a standardized window geometric feature semantic coding feature matrix and a standardized target stockpile geometric feature coarse-grained semantic aggregation coding matrix; Calculating a position difference matrix between the standardized window geometric feature semantic coding feature matrix and the standardized target stockpile geometric feature coarse-grained semantic convergence coding matrix, and taking an absolute value of the position difference matrix to obtain a target stockpile geometric feature semantic core convergence difference compensation coding matrix; The target stockpile geometric feature semantic kernel convergence difference compensation encoding matrix is ​​input into a compensation feature importance scoring module based on a neural network layer to obtain the target stockpile geometric feature semantic kernel convergence compensation factor.

7. The method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to claim 6, characterized in that: Inputting the target stockpile geometric feature semantic kernel convergence difference compensation encoding matrix into a compensation feature importance scoring module based on a neural network layer to obtain the target stockpile geometric feature semantic kernel convergence compensation factor, including: Multiplying the target stockpile geometric feature semantic kernel convergence difference compensation coding matrix by the weight parameter vector, and adding a bias term to the multiplication result to obtain the target stockpile geometric feature semantic kernel convergence compensation feature modulation vector; The target stockpile geometric feature semantic core convergence compensation feature modulation vector is multiplied by the target stockpile geometric feature semantic compensation feature importance score conversion vector to obtain the target stockpile geometric feature semantic core convergence compensation factor.

8. The method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to claim 7, characterized in that: The weight parameter vector is multiplied by the target stockpile geometric feature semantic kernel convergence difference compensation coding matrix, and the multiplication result is added with a bias term to obtain the target stockpile geometric feature semantic kernel convergence compensation feature modulation vector, including: Calculating the ratio between the Euclidean norm of the window geometric feature semantic coding feature matrix and the Euclidean norm of the target stockpile geometric feature coarse-grained semantic convergence coding matrix, if the ratio is less than 1, adding one to the ratio and calculating the logarithm value with base 2 as the bias term; If the ratio is greater than or equal to 1, the ratio is used as the bias term.

9. The method for automatically detecting the stockpile boundary of an unmanned bucket wheel stacker and reclaimer according to claim 8, characterized in that: Based on the set of semantic kernel convergence compensation factors of the target stockpile geometric features, the set of the window geometric feature semantic coding feature matrices and the target stockpile geometric feature coarse-grained semantic convergence coding matrix are semantically compensated and converged to obtain the target stockpile geometric feature aggregation coding matrix, including: Performing compensation explicit modeling based on a gating function on the set of the target stockpile geometric feature semantic kernel convergence compensation factors to obtain a set of target stockpile geometric feature semantic kernel convergence compensation weight factors; Inputting the set of the target stockpile geometric feature semantic core convergence compensation weight factors, the target stockpile geometric feature coarse-grained semantic convergence coding matrix and the window geometric feature semantic coding feature matrix into a node fine-grained dynamic compensation convergence network to obtain the target stockpile geometric feature fine-grained semantic compensation convergence coding matrix; The target stockpile geometric feature fine-grained semantic compensation aggregation coding matrix and the target stockpile geometric feature coarse-grained semantic aggregation coding matrix are input into a residual unit to obtain the target stockpile geometric feature aggregation coding matrix.

10. An automatic detection system for the stockpile boundary of an unmanned bucket wheel stacker and reclaimer, characterized in that: include: The original point cloud data acquisition module is used to scan the target stockpile using radar equipment to obtain original point cloud data; A data preprocessing module, used for performing data preprocessing on the original point cloud data to obtain preprocessed point cloud data; A geometric feature extraction module, used for performing geometric feature extraction based on a local window on the preprocessed point cloud data to obtain a set of window geometric feature semantic encoding feature matrices; A feature aggregation analysis module, used for performing a stockpile geometric feature aggregation analysis based on core information aggregation on the set of the window geometric feature semantic coding feature matrices to obtain a target stockpile geometric feature aggregation coding matrix; The semantic segmentation module is used to perform semantic segmentation on the target stockpile geometric feature aggregation coding matrix to obtain a stockpile boundary recognition result.

Images