Train carriage bulk material leveling method and system based on material leveling machine and medium

Abstract

The invention provides a train carriage bulk material leveling method and system based on a material leveling machine and a medium, and relates to the technical field of bulk material loading, and the method comprises the following steps: S1, carrying out real-time synchronous scanning on a train carriage through a plurality of laser radar scanners, and obtaining real-time point cloud data of the carriage; s2, extracting key feature point cloud data through a 3D sparse convolutional neural network based on the real-time point cloud data of the carriage; and S3, processing the key feature point cloud data through point cloud voxelization, and projecting on a plane to generate a two-dimensional distribution diagram. According to the method, multi-laser radar real-time synchronous scanning is combined with the 3D sparse convolutional neural network, two-dimensional mapping is performed through voxelization processing, and the moving position required by the leveling machine in the next step is tracked and predicted according to the two-dimensional image, so that the point cloud data processing efficiency is remarkably improved; the problem of insufficient real-time performance caused by large data volume in a traditional method is effectively solved, and meanwhile the requirements for high real-time performance and high precision of material leveling operation are met.

Description

Technical Field

[0001] The present invention relates to the technical field of bulk material loading, and in particular to a train carriage bulk material leveling method, system and medium based on a leveling machine. Background Art

[0002] After bulk materials are loaded into a train car, the sides may be lower and the center higher across the width. Railway authorities require that the materials inside the cars be level. Currently, several solutions are commonly used for loading, including manual leveling inside the cars or the use of dedicated spiral flattening machines. With the rapid development of railway transportation, the automation and intelligence of train loading operations are increasing, and manual leveling is gradually being eliminated. Spiral flattening machines are a crucial piece of loading equipment in rail freight, primarily used for loading bulk materials such as coal and ore. Current spiral flattening machines load cargo using spiral guide rail conveyors and are equipped with basic control systems and monitoring devices, such as position sensors and weight sensors, to monitor essential parameters during the loading process. With the increase in rail freight volume, some railway transportation companies have begun to introduce digital control systems on spiral flattening machines to collect and control basic data during the loading process.

[0003] However, the existing spiral flat car control method still has many shortcomings: the material distribution during the loading process is uneven, which can easily cause the car to tilt and safety hazards; for actual point cloud calculations, it uses a lot of data, and processing a large amount of laser point cloud data will result in excessive computing power, making the real-time processing of the point cloud data scanned by the laser radar unable to meet the real-time response requirements of flat materials. Summary of the Invention

[0004] In view of this, the present invention proposes a method, system and medium for leveling bulk materials in train carriages based on a leveling machine. It combines real-time synchronous scanning of multiple laser radars with a 3D sparse convolutional neural network, and performs two-dimensional mapping through voxelization processing. The next required movement position of the leveling machine is predicted based on two-dimensional image tracking, which significantly improves the efficiency of point cloud data processing, effectively solves the problem of insufficient real-time performance caused by large data volume in traditional methods, and meets the high real-time and high-precision requirements of leveling operations.

[0005] The technical solution of the present invention is achieved as follows:

[0006] In one aspect, the present invention provides a method for leveling bulk materials in a train carriage based on a leveling machine, comprising the following steps:

[0007] S1 multiple lidar scanners scan train carriages synchronously in real time to obtain real-time point cloud data of the carriages;

[0008] S2 extracts key feature point cloud data based on the real-time point cloud data of the carriage through a 3D sparse convolutional neural network;

[0009] S3 processes key feature point cloud data by voxelizing the point cloud and projects it onto a plane to generate a two-dimensional distribution map;

[0010] S4 determines the target position on the material pile to which the leveler needs to move at the next moment based on the two-dimensional distribution map, and the leveler moves toward the target position;

[0011] S5 repeats steps S1-S4 until the flattening operation is completed.

[0012] Based on the above technical solution, preferably, step S1 includes the following sub-steps:

[0013] S11 obtains the original point cloud data of the corresponding area through each lidar scanner;

[0014] S12 applies the calibration matrix to the original point cloud data obtained by each lidar scanner and converts it into the global coordinate system to obtain the processed point cloud;

[0015] S13 fuses multiple processed point clouds and removes redundant points in overlapping areas to generate a complete car point cloud;

[0016] S14 removes outliers and background from the complete carriage point cloud to obtain real-time point cloud data of the carriage.

[0017] Based on the above technical solution, preferably, the 3D sparse convolutional neural network includes an input layer, a sparse convolutional encoder, an attention separation module and a feature decoder. The input layer is used to input the real-time point cloud data of the vehicle compartment, the sparse convolutional encoder is used to capture features of different granularities through hierarchical convolution, the feature separation module is used to suppress and separate low-value areas, and the feature decoder is used to output key feature point cloud data.

[0018] On the basis of the above technical solution, preferably, before step S3, voxel grid division is performed on the three-dimensional space, and step S3 includes the following sub-steps:

[0019] S31 assigns each point cloud in the key feature point cloud data to a voxel in the corresponding space, and only records the voxels containing the point cloud;

[0020] S32 calculates statistical features of points in each voxel and generates feature vectors of each voxel;

[0021] S33 projects all voxels on a horizontal plane to obtain a two-dimensional distribution map, wherein the two-dimensional distribution map represents the height of the stockpile in grayscale values.

[0022] Based on the above technical solution, preferably, step S4 includes the following sub-steps:

[0023] S41 locates candidate targets based on the two-dimensional distribution map;

[0024] S42 initializes the tracker for each candidate target and associates the target position in the historical frame;

[0025] S43 uses Kalman filtering to predict the changing trend of the target position, combines the current detection results to update the status, determines the target position at the next moment, and the flat material machine moves towards the target position.

[0026] On the basis of the above technical solution, preferably, step S6 is also included. After the leveling is completed, the leveling effect is verified by a second scan of the corresponding carriage. If it meets the standards, the current leveling operation is ended. If it does not meet the standards, the leveling operation is re-entered for the carriage and step S1 is started.

[0027] Further preferably, before step S6, a target value of the stockpile is preset, and step S6 includes the following sub-steps:

[0028] S61 obtains point cloud verification data after the carriage is leveled;

[0029] S62 obtains stockpile parameters from point cloud verification data;

[0030] S63 calculates statistical indicators of the material pile in the carriage according to the material pile parameters and the material pile target value, wherein the statistical indicators include the material pile height difference and the maximum deviation value;

[0031] S64 performs grid partitioning through the point cloud in the point cloud verification data and identifies local exceeding standards areas based on statistical indicators;

[0032] If all partitions meet the statistical indicators in S65, the material leveling is determined to be completed and the operation is ended. If there is an area that does not meet the standards, the process re-enters step S1 for that carriage.

[0033] On the basis of the above technical solution, preferably, before step S3, it also includes multi-scale feature fusion of the key feature point cloud data in step S2, constructing a multi-resolution feature pyramid, completing the point cloud of the missing area through the adversarial generative network, applying non-local mean filtering to eliminate point cloud noise interference, establishing a temporal association model, comparing continuous frame point cloud data to correct dynamic errors, and outputting the processed key feature point cloud data to step S3.

[0034] On the other hand, the present invention provides a train carriage bulk material leveling system based on a leveling machine, which is embedded with the above-mentioned train carriage bulk material leveling method.

[0035] On the other hand, the present invention provides a computer-readable storage medium, on which a computer can read a material leveling machine pushing program, and when the material leveling machine pushing program is executed by a processor, the above-mentioned train car bulk material leveling method is implemented.

[0036] The train carriage bulk material leveling method and system based on the leveling machine of the present invention has the following beneficial effects compared with the prior art:

[0037] (1) By combining real-time synchronous scanning of multiple laser radars with 3D sparse convolutional neural networks and performing two-dimensional mapping through voxel processing, the next required movement position of the flat material machine is predicted based on two-dimensional image tracking, significantly improving the efficiency of point cloud data processing and effectively solving the problem of insufficient real-time performance caused by large data volumes in traditional methods, while meeting the high real-time and high-precision requirements of flat material operations;

[0038] (2) Through the collaborative work of three stages: candidate target detection, multi-target tracking, and motion prediction: first, the high points or abnormal areas of the material pile are identified as candidate targets from the grayscale image, then the historical trajectory is matched through time series correlation, and finally a filtering algorithm is used to predict the target position change trend and correct the dynamic error. The entire process integrates image processing, target tracking, and state estimation technology, taking into account both real-time and robustness, and providing continuous and smooth movement instructions for the flat material machine;

[0039] (3) By performing voxelization after sparse convolution, the sparse and irregular problems of the original point cloud data are effectively solved. Secondly, the multi-feature fusion calculation method enhances the ability to describe the pile morphology. Finally, the projection and mapping algorithms set up ensure the accurate transmission of height information. These technical features together ensure that the final generated two-dimensional distribution map has sufficient accuracy and is easy to process by subsequent algorithms, showing good stability and reliability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 The present invention is a flow chart of a method for leveling bulk materials in a train carriage based on a leveling machine. DETAILED DESCRIPTION

[0042] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] It should be noted that the flattening machine pushes the material peaks on the open car pile toward both sides by setting a pushing screw. It only needs to move in the length direction and vertical direction relative to the open car, and does not need to move horizontally in the open car. This embodiment is based on this implementation.

[0044] In the prior art, for automated material leveling, the width of the carriage is usually determined first. Since the width is relatively fixed, after the width is determined, it is only necessary to obtain the length and height of the stockpile to obtain the volume of the stockpile. The volume of the stockpile is often calculated directly using three-dimensional point cloud data. After the volume of the stockpile is calculated, the standard height of the stockpile leveling can be determined, and then the leveling machine is directly moved to the standard height to perform the leveling operation. This method requires less data calculation and can meet real-time requirements. However, due to factors such as the conveying screw, the uniform motion of the carriage traction head, and the accumulation of material in front during the pushing process of the leveling machine, a single leveling operation often fails, resulting in the need for secondary leveling. Since the open car moves at a uniform speed under the action of the traction head, the leveling time for each carriage is limited. If continuous leveling fails, the time for secondary leveling cannot be met. For carriages with unqualified leveling and moved outside the leveling area, they can only be leveled manually, which increases labor costs.

[0045] Another point cloud processing method uses full 3D point cloud computing to calculate the operation path of the flat material machine. However, due to the large amount of calculation, the output path is usually 1-2S, which cannot be adaptively adjusted in time to cope with different flat material conditions. There will also be a large number of unqualified flat material conditions, especially unqualified stacking height in local areas.

[0046] In response to the above problems, Figure 1 As shown, the method for leveling bulk materials in a train carriage based on a leveling machine of the present invention includes steps S1-S6.

[0047] Step S1: Multiple lidar scanners perform real-time synchronous scanning of the train carriage to obtain real-time point cloud data of the carriage.

[0048] In actual operation, multiple lidar scanners are arranged in the flat material operation area, specifically by setting them directly above the train tracks through brackets. In some specific embodiments, the number of lidar scanners can be set to seven or eight.

[0049] This step involves deploying multiple LiDAR scanners to perform real-time, simultaneous scanning of train cars, sequentially collecting raw point cloud data and performing preprocessing on the raw point cloud data. This preprocessing includes global coordinate system conversion, multi-source data fusion, and data optimization. Specifically, each LiDAR acquires a localized raw point cloud, which is uniformly converted to a global coordinate system using a calibration matrix. The multi-viewpoint point clouds are then fused and redundant data in overlapping areas is removed to generate a complete car point cloud. Finally, outlier removal and background culling are used to filter out noise and irrelevant environmental information, resulting in high-precision, real-time point cloud data for the car, providing reliable input for subsequent analysis.

[0050] Specifically, this step includes steps S11-S14.

[0051] Step S11: Obtain original point cloud data of the corresponding area through each laser radar scanner.

[0052] Multiple LiDAR scanners scan the car's surface synchronously in a pre-set layout. Each scanner covers a specific area of the car (such as the top or side), independently collecting raw 3D point cloud data within its field of view. This raw point cloud contains the geometric information of the bulk material surface and any ambient noise (such as background noise from equipment supports, tracks, etc.).

[0053] Step S12: The original point cloud data acquired by each lidar scanner is converted to a global coordinate system using a calibration matrix to obtain a processed point cloud.

[0054] Based on the calibration parameters of the lidar (including position and attitude), the original point cloud obtained by each scanner is transformed into a coordinate system through the calibration matrix, and the point clouds in all local coordinate systems are mapped to a unified global coordinate system, eliminating the spatial misalignment problem caused by the differences in the posture of multiple sensors, and forming a spatially aligned multi-source point cloud dataset, that is, the processed point cloud.

[0055] Step S13: Fuse the multiple processed point clouds and remove redundant points in the overlapping areas to generate a complete car point cloud.

[0056] Multiple groups of point clouds in the global coordinate system are registered and fused, and the overlapping areas between different point clouds are detected using spatial indexing algorithms (such as KD-Tree). Duplicate points are removed through voxel downsampling or nearest neighbor distance threshold method, and unique point cloud data is retained to generate a complete and non-redundant three-dimensional point cloud model of the car surface.

[0057] Step S14: remove outliers and background from the complete car point cloud to obtain real-time point cloud data of the car.

[0058] For outlier removal: Statistical filtering or radius filtering algorithms can be used to remove isolated abnormal points caused by sensor noise or splashing materials based on point cloud density or neighborhood distance threshold.

[0059] For background removal: the car body geometry model (such as car body boundary parameters) can be combined to separate the bulk material point cloud inside the car body from the external environment points (such as tracks, robotic arms) through spatial masking or region growing methods, retaining only the valid data related to the material pile, and outputting the optimized car body real-time point cloud data.

[0060] At this point, the real-time point cloud data of the carriage can be obtained through step S1. The real-time point cloud data of the carriage refers to the point cloud data of the carriage and its material pile after preprocessing in the current state. It is real-time and is also the basis for the subsequent flat material machine to perform real-time adaptive adjustment actions.

[0061] Step S2: Based on the real-time point cloud data of the carriage, extract key feature point cloud data through a 3D sparse convolutional neural network.

[0062] In some specific embodiments, the 3D sparse convolutional neural network includes an input layer, a sparse convolutional encoder, an attention separation module and a feature decoder. The input layer is used to input real-time point cloud data of the vehicle compartment, the sparse convolutional encoder is used to capture features of different granularities through hierarchical convolution, the feature separation module is used to suppress and separate low-value areas, and the feature decoder is used to output key feature point cloud data.

[0063] The core task of step S2 is to extract key feature point clouds from the real-time point cloud data of the carriage obtained in step S1, providing high-quality input for subsequent material pile height analysis and path planning. This step uses a 3D sparse convolutional neural network as the basic architecture, and through multi-stage feature extraction and optimization, ensures that the point cloud data has high precision, low noise, and spatiotemporal consistency. First, the sparse convolutional encoder performs multi-scale feature extraction on the input point cloud to capture the geometric and semantic information of the material pile; then, the attention mechanism separates the key areas from the background noise to improve the feature quality; finally, through post-processing steps such as feature decoding, missing completion, denoising, and timing correction, the optimized key feature point cloud is output. The entire process takes into account both computational efficiency and feature expression capabilities, laying a solid foundation for subsequent processing.

[0064] The training process of this 3D sparse convolutional neural network starts with the collection of multi-operational point cloud data. LiDAR scanning is used to obtain train car data under different material pile shapes, environmental interference, and motion scenarios. The point cloud is then semantically segmented and key feature points are manually annotated. Enhancement strategies such as rotation and translation, noise injection, and dynamic occlusion are used to improve data diversity. A network architecture with multi-scale feature fusion is constructed, combined with a sparse convolutional encoder to extract hierarchical features. A dual-channel attention mechanism is used to suppress irrelevant areas. The model is optimized in stages through a progressive training strategy. The overall contour is learned at a coarse granularity, and then gradually refined to accurately locate key points. Dynamic loss weights are introduced to balance segmentation and regression tasks, supplemented by hard sample mining to enhance adaptability to complex cases. Finally, a lightweight model is deployed through channel pruning and quantization compression. Its robustness and real-time processing capabilities are verified under extreme working conditions such as lighting changes, high dust, and vibration interference. This ensures accurate identification of key areas of the material pile and guidance of the leveling machine in actual leveling operations.

[0065] In a specific embodiment, this step S2 includes steps S21 - S28 .

[0066] Step S21: Set a sparse convolutional encoder.

[0067] This step first inputs the real-time point cloud data of the carriage output in step S1 into the 3D sparse convolutional neural network. Due to the sparsity of the original point cloud data, directly using traditional dense convolution will result in a large amount of invalid calculations. Therefore, the point cloud is converted into a non-empty voxel grid through sparse voxelization, which significantly improves the computational efficiency. The sparse convolution encoder adopts a hierarchical structure. Shallow convolutions (such as 3×3×3 kernels) extract low-resolution global features (such as the boundaries of the carriage and the overall outline of the pile), while deep convolutions (such as 1×1×1 kernels) focus on high-resolution local details (such as the surface bumps and particle distribution of the pile). Each layer of convolution is followed by a LeakyReLU activation function to enhance the nonlinear modeling capability. The final output multi-scale feature map contains both geometric shape and semantic information, providing rich feature expression for subsequent processing.

[0068] Step S22: Setting the attention separation module.

[0069] Based on feature extraction, this step further optimizes the point cloud features, suppresses irrelevant noise and enhances key areas. The channel attention mechanism (SE Block) first analyzes the importance of different feature channels, generates channel weights through global average pooling (GAP) and fully connected layers (FC), and weakens low-value channels such as car wall reflections. Subsequently, the spatial attention module (CBAM) calculates spatial weights on the feature map to highlight the core areas of the pile (such as material peaks and depressions) while suppressing background interference. After combining the two, the feature map is dynamically enhanced through weighted fusion to ensure that subsequent processing focuses only on key point clouds related to the leveling task. This step significantly improves the signal-to-noise ratio of the feature and provides high-quality input for the decoding stage.

[0070] Step S23: Setting the feature decoder.

[0071] The goal of the decoding stage is to restore the optimized feature map to a sparse point cloud while retaining key information. Sparse transposed convolution is used to gradually upsample the feature map to restore the spatial resolution of the point cloud. To avoid loss of details, the shallow high-resolution features of the encoding stage are fused with the deep semantic features of the decoding stage through skip connections, taking into account both local accuracy and global consistency. Finally, 1×1 convolution and Softmax classifier are used to perform binary classification (key point / background point) on each point, retaining only the stockpile-related point cloud with high confidence. The output data is a sparse key feature point cloud after de-redundancy, which greatly reduces the complexity of subsequent calculations.

[0072] The above is the specific composition of the 3D sparse convolutional neural network. In addition to extracting key feature point cloud data through the 3D sparse convolutional neural network, the accuracy and robustness of subsequent processing can also be improved by enhancing the point cloud data.

[0073] The point cloud data can be enhanced through steps S24-S28.

[0074] Multi-scale feature fusion is performed on the key feature point cloud data to construct a multi-resolution feature pyramid. The point cloud of the missing area is completed and generated through the adversarial generative network. Non-local mean filtering is applied to eliminate point cloud noise interference. A temporal association model is established. The dynamic error is corrected by comparing the continuous frame point cloud data. The processed key feature point cloud data is output to step S3.

[0075] Step S24: Construct a multi-scale feature pyramid.

[0076] In order to enhance the model's perception of features of different scales, this step requires the construction of a multi-scale feature pyramid. First, the key point cloud is subjected to multi-level voxel downsampling (such as 5cm, 10cm, 20cm) to generate feature maps of different resolutions. Low-resolution features capture the overall shape of the pile (such as slope trends), while high-resolution features retain local details (such as particle accumulation). Feature maps at different levels are spliced and fused through lateral connections to achieve complementarity between details and semantics. This multi-scale fusion strategy effectively improves the model's robustness to changes in pile height, and is particularly suitable for complex stacking scenarios.

[0077] Step S25: Generate a complete image of the correct area through a generative adversarial network.

[0078] The original point cloud may contain missing data due to scanning occlusions or sensor blind spots. This step uses a generative adversarial network for completion: the generator (U-Net structure) predicts the point cloud distribution of the missing area based on contextual information, while the discriminator (PatchGAN) evaluates the authenticity of the generated point cloud, optimizing the completion effect through adversarial training. The loss function combines (measuring shape similarity) with the adversarial loss (ensuring distribution consistency) to seamlessly integrate the completed point cloud with the real data. The output is a complete stockpile point cloud, eliminating the interference of missing data on stockpile leveling decisions.

[0079] Step S26: performing fine denoising on the point cloud data through non-local mean filtering.

[0080] Although the previous steps have suppressed some noise, the point cloud may still contain outliers caused by sensor jitter or scattered particles. This step uses non-local mean filtering for refined denoising: for each point, a similarity weight is calculated (based on geometric and feature distances) with the global point cloud, and noisy points are smoothed using weighted averaging. Compared to traditional filtering methods, this algorithm can remove noise while preserving the stockpile's true geometric features (such as sharp edges), ensuring the accuracy of subsequent height calculations.

[0081] Step S27: Establish a temporal correlation model to compare the temporal and spatial consistency of consecutive frame point clouds, identify and correct dynamic errors.

[0082] During the flattening process, material flow or mechanical vibration within the carriage can cause dynamic deformation of the point cloud. This step establishes a temporal correlation model (such as LSTM or sliding window matching) to compare the spatiotemporal consistency of point clouds in consecutive frames, identifying and correcting dynamic errors. A Kalman filter is used to predict the motion trend of the point cloud and update the status based on the current detection results, eliminating transient occlusions or drift interference. This mechanism significantly improves the temporal stability of point cloud data and avoids misjudgments caused by dynamic changes.

[0083] Step S28: Integrate the point cloud data.

[0084] Finally, all optimized point cloud data is integrated: low-confidence points remaining from the completion and denoising processes are removed, and outliers are filtered using physical constraints (such as a maximum height threshold). The resulting data is a high-precision, complete, and spatially consistent key feature point cloud, which can be directly input into step S3 for voxelization. This step ensures the reliability and efficiency of subsequent flat material path planning.

[0085] S3 processes key feature point cloud data by voxelizing the point cloud and projects it onto a plane to generate a two-dimensional distribution map.

[0086] The main task of this step is to convert the key feature point cloud data extracted in step S2 into a two-dimensional height distribution map that can be used for the flat material machine path planning. This step converts the discrete three-dimensional point cloud data into a structured spatial representation through voxelization, and then generates an intuitive two-dimensional grayscale image through projection transformation. This conversion process not only retains the key height information of the material pile but also greatly reduces the complexity of subsequent processing, providing reliable spatial data support for the flat material machine's intelligent decision-making. The entire processing flow includes three key sub-steps: voxel grid division and point cloud allocation, voxel feature calculation and feature vector generation, and finally two-dimensional projection and grayscale distribution map generation.

[0087] In some specific embodiments, this step includes sub-steps S31-S33.

[0088] Step S31: assign each point cloud in the key feature point cloud data to a voxel in the corresponding space, and only record the voxels containing the point cloud.

[0089] This step first divides the three-dimensional space into regular sections and establishes a voxel grid coordinate system. According to the actual size of the carriage and the requirements for flat material accuracy, the space is divided into cubic units. A cube of 5cm×5cm×5cm can be selected. This size selection ensures sufficient spatial resolution while avoiding excessive computational burden. In specific implementation, a sparse storage structure is used to record the voxels containing the point cloud, which greatly improves storage efficiency. During the point cloud allocation process, each three-dimensional point is mapped to the corresponding voxel unit according to its spatial coordinates. At the same time, the system will dynamically adjust the voxel size in high-density areas to ensure finer sampling in dense material pile areas. The output of this step is voxelized point cloud data with spatial index information, which lays the foundation for subsequent feature calculations.

[0090] Step S32: Calculate statistical features for points in each voxel and generate feature vectors for each voxel.

[0091] After completing the voxel allocation of the point cloud, this step extracts and calculates features for each non-empty voxel. The main features extracted include: the maximum height value, average height value, height variance and point cloud density of all points in the voxel. After normalization, these features are combined into a multi-dimensional feature vector that characterizes the voxel. In order to improve the robustness of the features, the height data is also Gaussian filtered to effectively smooth the impact of measurement noise. In particular, in key areas such as the edge of the pile, additional geometric features such as the rate of change of the surface normal vector are calculated to more accurately describe the local morphology of the pile. These carefully designed features provide rich and reliable data support for subsequent height distribution analysis.

[0092] Step S33: Project all voxels on a horizontal plane to obtain a two-dimensional distribution map, wherein the two-dimensional distribution map represents the height of the pile in grayscale values.

[0093] This step converts the three-dimensional voxel data into a two-dimensional height distribution map. First, all voxels are projected vertically onto the horizontal plane, and the maximum height value of each projection position is taken as the grayscale reference of that position. The height value is then mapped to a grayscale range of 0-255 through a linear transformation, where higher grayscale values correspond to higher stockpile positions. In order to improve the image quality, a series of post-processing operations including missing value interpolation and edge enhancement can be performed. The final grayscale image not only intuitively shows the overall height distribution of the stockpile, but also retains sufficient detail information for subsequent analysis. This two-dimensional distribution map will become the direct basis for the path planning of the flat material machine, and its quality directly affects the effect of the entire flat material operation.

[0094] Step S4: Determine the target position on the material pile to which the leveler needs to move at the next moment based on the two-dimensional distribution map, and move the leveler toward the target position.

[0095] The goal of this step is to dynamically determine the target position to which the flattening machine needs to move next, based on the two-dimensional grayscale distribution map generated in step S3, to ensure efficient and accurate bulk material leveling. This step works in tandem through three stages: candidate target detection, multi-target tracking, and motion prediction. First, high points or abnormal areas in the material pile are identified as candidate targets from the grayscale map. Then, historical trajectories are matched through temporal correlation. Finally, a filtering algorithm is used to predict target position change trends and correct dynamic errors. The entire process integrates image processing, target tracking, and state estimation technologies, balancing real-time and robustness to provide continuous and smooth movement instructions for the flattening machine.

[0096] The step S4 includes sub-steps S41-S43.

[0097] Step S41: Locate candidate targets according to the two-dimensional distribution map.

[0098] This step extracts candidate target areas that need to be flattened (such as material peaks and local protrusions) from the two-dimensional grayscale distribution map. First, the grayscale image is subjected to multi-scale Gaussian filtering, and filters with different kernel sizes (such as 3×3, 5×5) are used to eliminate high-frequency noise (such as sensor noise or flying particle interference) while retaining the edge details of the material pile. Subsequently, based on the local extreme value detection algorithm, a sliding window (10×10 pixels) is used to traverse the image, and the maximum grayscale value and its position in each window are calculated. If the maximum value of a window is significantly higher than the neighborhood mean (the threshold is set to 2 times the standard deviation), it is marked as a candidate target. In order to further eliminate false detections, the morphological opening operation (first corrosion and then expansion) is combined to filter areas that are too small (such as less than 5×5 pixels) or irregular areas, and finally the coordinates, area and average height information of the candidate target are output as input for tracker initialization.

[0099] Step S42: Initialize the tracker for each candidate target and associate the target position in the historical frame.

[0100] For each candidate target, initialize an independent multi-target tracker (such as KCF kernel correlation filter tracker or SORT algorithm). The core task of the tracker is to associate the current frame target with the historical frame trajectory to ensure the continuity of the target ID.

[0101] The specific implementation includes: Feature description: extracting the grayscale histogram, local binary pattern (LBP) texture features and centroid coordinates of the candidate target area to construct a multi-dimensional feature vector. Historical frame matching: using the Hungarian Algorithm to perform target matching, calculate the feature similarity (cosine distance) and spatial distance (Euclidean distance) between the current candidate target and the historical target, and determine the optimal matching pair after comprehensive weighting. New target processing: for unmatched candidate targets, it is determined to be a new material pile area, assigned a new ID and added to the tracking queue; for historical targets that have not been matched continuously, if they are not detected for 3 consecutive frames, they are determined to have been leveled and removed from the queue. Through this mechanism, the system can distinguish between real material pile high points and instantaneous interference, avoiding repeated processing of the same area or missing new targets.

[0102] Step S43: Use Kalman filtering to predict the changing trend of the target position, update the state in combination with the current detection result, determine the target position at the next moment, and move the flat material machine towards the target position.

[0103] In order to cope with the dynamic changes in the target position caused by the movement of the flat feeder and the flow of materials, this step uses Kalman filtering to achieve the prediction and correction of the target trajectory, including state modeling: defining the target state vector as position, velocity and acceleration, and establishing a linear motion equation (such as a uniform velocity model or a uniform acceleration model). Prediction stage: Based on the state vector and covariance matrix of the previous moment, the target position and uncertainty range at the current moment are predicted through the state transfer equation. Measurement update: The actual position of the target detected in step S41 is used as the observation value, and the residual between the predicted value and the observed value is calculated. The Kalman gain is used to dynamically adjust the state estimation weight to reduce the impact of detection noise. Trajectory smoothing and output: Output the corrected target position coordinates, and perform polynomial fitting on the trajectory to eliminate instantaneous jitter and generate the movement instructions of the flat feeder (such as target point coordinates and moving speed). The introduction of Kalman filtering effectively solves the problem of trajectory jumps caused by short-term occlusion of the target or detection fluctuations, ensuring the smoothness and continuity of the flat feeder's moving path.

[0104] Step S5: Repeat steps S1-S4 until the flattening operation is completed.

[0105] In this step, the flattening machine can move according to the flattening machine movement instruction generated in step S43. While moving, the next required moving position is continuously determined by continuously looping steps S1-S4 until the next required moving position is outside the material pile area. At this time, the flattening operation can be terminated.

[0106] It's important to note that the output movement instructions must be based on the mechanical characteristics and motion constraints of the flat material machine (such as maximum speed and acceleration), using a cubic spline interpolation algorithm to generate a smooth motion trajectory to avoid mechanical shock caused by sudden stops and starts. The trajectory is also optimized, taking into account the spatial distribution characteristics of the material pile. The generated trajectory instructions must include position sequences, velocity curves, and acceleration parameters to provide a precise control benchmark for subsequent execution.

[0107] Step S6: After the leveling is completed, the leveling effect is verified by a second scan of the corresponding carriage. If it meets the standard, the current leveling operation is terminated. If it does not meet the standard, the leveling operation is restarted for the carriage and step S1 is started.

[0108] The purpose of this step is to verify the flattening process through a second scan, determine whether it meets the preset standards, and decide whether to re-execute the flattening process based on the verification results. This is done through a progressive process of data collection, parameter analysis, and zoning determination, combined with quantitative statistics and spatial distribution assessment, to ensure that the flatness of the stockpile within the carriage meets the required standards. If any localized areas exceed the standard, a re-flattening instruction is triggered, forming a closed-loop quality control mechanism, eliminating the cost of manual re-inspection and improving the reliability of automated operations.

[0109] In addition, before this step, the target value of the material pile needs to be preset. The target value of the material pile can be determined by collecting initial three-dimensional point cloud data of the material pile before the operation and performing calculations to determine parameters such as the volume of the material pile.

[0110] Step S61: Obtaining point cloud verification data after the carriage is leveled.

[0111] After the leveling machine completes its operation, the system activates the multi-lidar scanner in step S1 to perform a second full-range scan of the carriage. Compared with the initial scan, this scan uses a higher-precision mode (such as reducing the angular resolution of the lidar to 0.1°) to ensure that subtle height differences are captured. The scan data undergoes calibration matrix conversion, multi-source fusion, and redundancy elimination (the same process as steps S11-S14) to generate point cloud data of the carriage after leveling. In order to eliminate environmental interference, dynamic background modeling technology is additionally introduced to separate temporary background objects (such as falling debris) added during the movement of the carriage in real time to ensure the purity of the verification data.

[0112] Step S62: Obtain stockpile parameters from the point cloud verification data.

[0113] Key pile parameters are extracted from the secondary scan point cloud data. These include global parameters: the average height, maximum height, and estimated volume of the pile within the vehicle compartment; local parameters: the mean and standard deviation of the height of each grid cell calculated based on grid partitioning (consistent with the voxelization method in step S3); and morphological parameters: the slope distribution of the pile surface (through point cloud normal analysis) and the flatness index (calculated based on height variance). After parameter extraction, they are compared with preset target values (e.g., maximum allowable height difference ±5 cm, slope ≤15°) to form a preliminary assessment.

[0114] Step S63: Calculate statistical indicators of the material pile in the carriage according to the material pile parameters and the material pile target value, wherein the statistical indicators include the material pile height difference and the maximum deviation value.

[0115] Based on the extracted stockpile parameters, the following statistical indicators are calculated:

[0116] Height Difference Indicator: Global Height Difference = Maximum Height - Minimum Height; Local Height Difference = The difference between the highest and lowest points in each grid cell; Maximum Deviation: The maximum absolute deviation between the actual height and the target height; Flatness Compliance Rate: The percentage of grid cells that meet the height threshold (if ≥ 95% is considered fully met). The system automatically determines global compliance based on preset thresholds. If any indicator exceeds the limit (e.g., height difference > 8cm or compliance rate < 95%), the system is marked as "Not Met."

[0117] Step S64: gridding and partitioning the point cloud in the point cloud verification data, and identifying local exceeding-standard areas based on statistical indicators.

[0118] In order to accurately locate the non-compliant areas, the verification point cloud was divided into fine grids (e.g., 1m×1m grids). Grid attribute calculation: the mean height, maximum deviation, and mean slope were calculated within each grid. Exceeding standard judgment: if the height deviation within the grid is greater than a preset threshold (e.g., ±3cm) or the slope is greater than a preset angle, it is marked as “locally exceeding the standard”. Spatial clustering: the DBSCAN clustering algorithm was used to merge adjacent exceeding standard grids to form a continuous exceeding standard area (e.g., area greater than 0.5m 2 areas that require reprocessing).

[0119] The output is a list of areas exceeding the standard with spatial coordinates, clearly identifying the specific locations that require rework.

[0120] Step S65: If all partitions meet the statistical indicators, the material leveling is determined to be completed and the operation is terminated. If there is an area that does not meet the standards, the process re-enters step S1 for that car.

[0121] The train carriage bulk material leveling system based on the leveling machine of the present invention is embedded with the above-mentioned train carriage bulk material leveling method.

[0122] The computer-readable storage medium of the present invention can read the material leveling machine pushing program on the computer-readable storage medium, and the material leveling machine pushing program, when executed by the processor, realizes the above-mentioned train car bulk material leveling method.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for leveling bulk materials in train carriages based on a leveling machine, characterized in that: The following steps are involved: S1 multiple lidar scanners scan train carriages synchronously in real time to obtain real-time point cloud data of the carriages; S2 extracts key feature point cloud data based on the real-time point cloud data of the carriage through a 3D sparse convolutional neural network; S3 processes key feature point cloud data by voxelizing the point cloud and projects it onto a plane to generate a two-dimensional distribution map; S4 determines the target position on the material pile to which the leveler needs to move at the next moment based on the two-dimensional distribution map, and the leveler moves toward the target position; S5 repeats steps S1-S4 until the flattening operation is completed.

2. The method for leveling bulk materials in train carriages based on a leveling machine according to claim 1, characterized in that: The step S1 includes the following sub-steps: S11 obtains the original point cloud data of the corresponding area through each lidar scanner; S12 applies the calibration matrix to the original point cloud data obtained by each lidar scanner and converts it into the global coordinate system to obtain the processed point cloud; S13 fuses multiple processed point clouds and removes redundant points in overlapping areas to generate a complete car point cloud; S14 removes outliers and background from the complete carriage point cloud to obtain real-time point cloud data of the carriage.

3. The method for leveling bulk materials in train carriages based on a leveling machine according to claim 1, characterized in that: The 3D sparse convolutional neural network includes an input layer, a sparse convolutional encoder, an attention separation module and a feature decoder. The input layer is used to input real-time point cloud data of the vehicle compartment. The sparse convolutional encoder is used to capture features of different granularities through hierarchical convolution. The feature separation module is used to suppress and separate low-value areas. The feature decoder is used to output key feature point cloud data.

4. The method for leveling bulk materials in train carriages based on a leveling machine according to claim 1, characterized in that: Before step S3, the three-dimensional space is divided into voxel grids, and step S3 includes the following sub-steps: S31 assigns each point cloud in the key feature point cloud data to a voxel in the corresponding space, and only records the voxels containing the point cloud; S32 calculates statistical features of points in each voxel and generates feature vectors of each voxel; S33 projects all voxels on a horizontal plane to obtain a two-dimensional distribution map, wherein the two-dimensional distribution map represents the height of the stockpile in grayscale values.

5. The method for leveling bulk materials in train carriages based on a leveling machine according to claim 1, characterized in that: The step S4 includes the following sub-steps: S41 locates candidate targets based on the two-dimensional distribution map; S42 initializes the tracker for each candidate target and associates the target position in the historical frame; S43 uses Kalman filtering to predict the changing trend of the target position, combines the current detection results to update the status, determines the target position at the next moment, and the flat material machine moves towards the target position.

6. The method for leveling bulk materials in train carriages based on a leveling machine according to claim 1, characterized in that: It also includes step S6. After the leveling is completed, the leveling effect is verified by a second scan of the corresponding carriage. If it meets the standard, the current leveling operation is ended. If it does not meet the standard, the leveling operation is re-entered for the carriage and step S1 is started.

7. The method for leveling bulk materials in train carriages based on a leveling machine according to claim 6, characterized in that: Before step S6, a target value of the stockpile is preset. Step S6 includes the following sub-steps: S61 obtains point cloud verification data after the carriage is leveled; S62 obtains stockpile parameters from point cloud verification data; S63 calculates statistical indicators of the material pile in the carriage according to the material pile parameters and the material pile target value, wherein the statistical indicators include the material pile height difference and the maximum deviation value; S64 performs grid partitioning through the point cloud in the point cloud verification data and identifies local exceeding standards areas based on statistical indicators; If all partitions meet the statistical indicators in S65, the material leveling is determined to be completed and the operation is ended. If there is an area that does not meet the standards, the process re-enters step S1 for that carriage.

8. The method for leveling bulk materials in train carriages based on a leveling machine according to claim 1, characterized in that: Before step S3, the method also includes performing multi-scale feature fusion on the key feature point cloud data in step S2, constructing a multi-resolution feature pyramid, completing and generating the missing area point cloud through a generative adversarial network, applying non-local mean filtering to eliminate point cloud noise interference, establishing a temporal association model, comparing continuous frame point cloud data to correct dynamic errors, and outputting the processed key feature point cloud data to step S3.

9. A train carriage bulk material leveling system based on a leveling machine, characterized in that: The invention is embedded with the method for leveling bulk materials in a railway carriage according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer can read the material leveling machine pushing program on the storage medium, and when the material leveling machine pushing program is executed by the processor, it realizes the train car bulk material leveling method according to any one of claims 1 to 8.

Images