Coal yard inventory scheduling system based on real-time data

Abstract

The application discloses a coal yard inventory scheduling system based on real-time data, which comprises a perception layer, an edge computing layer, a digital twin layer, a decision layer and an execution layer. The perception layer is used for acquiring basic data. The edge computing layer can collect and clean the data of the perception layer, extract feature vectors of the environment and equipment state, and construct a three-dimensional density field. The digital twin layer receives the data obtained by the edge computing layer, and constructs a three-dimensional dynamic digital twin engine. The decision layer generates control instructions after receiving order information and calculating. The execution layer performs work according to the instructions issued by the decision layer. The application realizes accurate dynamic scheduling and management of the coal yard inventory through multi-source real-time data perception, rapid processing of the edge computing node and digital twin simulation. Through the simulation rehearsal and mechanical analysis of the digital twin layer, the collapse risk is early warned, and the accident rate is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal yard scheduling, and particularly relates to a coal yard inventory scheduling system based on real-time data. BACKGROUND

[0002] Traditional coal yard scheduling mainly relies on manual inspection and static data, which has significant problems: manual measurement by equipment is long, such as several hours for measuring the volume of coal piles; inventory cannot be updated in real time, leading to scheduling decisions relying on outdated information and easily causing imbalance between supply and demand; there is a lack of dynamic structure monitoring, and hidden dangers such as coal pile collapse and slope overrun cannot be found in time; vehicle path planning relies on experience, which easily causes empty loading of equipment, route conflicts, and difficulty in responding to sudden environmental changes such as extreme weather. Although existing technologies introduce some automatic equipment, the systems are isolated, and the data processing capability is weak, so full-link collaborative optimization cannot be achieved. SUMMARY

[0003] Therefore, the present application provides a coal yard inventory scheduling system based on real-time data to solve the problems in the prior art.

[0004] In order to achieve the above purpose, the present application provides the following technical scheme:

[0005] A coal yard inventory scheduling system based on real-time data, comprising a perception layer, an edge computing layer, a digital twin layer, a decision layer, and an execution layer,

[0006] The perception layer comprises a high-density laser radar matrix, an unmanned aerial vehicle inspection system equipped with an infrared thermal imager, a dynamic weighing of a weighbridge, an environment perception network, and a device state monitoring sensor.

[0007] The edge computing layer takes a plurality of sensors with computing and transmission capabilities provided in the perception layer as edge computing nodes; the edge computing nodes can collect data of the perception layer, clean the collected data, and extract feature vectors of the environment and device state; the edge computing nodes perform point cloud processing on the point cloud data of the perception layer and construct a three-dimensional density field.

[0008] The digital twin layer receives data obtained by the edge computing layer, constructs a three-dimensional dynamic digital twin engine, and fuses multispectral data to map a heat value distribution in the digital twin engine.

[0009] The decision layer, after receiving order information, runs a hybrid optimization algorithm based on the feature vectors preprocessed by the edge computing nodes, simulates a virtual coal yard with physical characteristics constructed by the digital twin layer, and then generates device control instructions and issues them to the execution layer.

[0010] The execution layer, after receiving the instructions issued by the decision layer, performs operations according to the instructions.

[0011] Further: the deployment interval of the laser radar is calculated according to an algorithm, and the deployment interval algorithm is d=2R tan (θ / 2n), wherein R is the radius of the coal pile, θ is the field of view angle of the radar, n is the overlap rate, n≥30%, and the laser radar is installed on the tower.

[0012] Further: the point cloud processing flow is as follows:

[0013] (1) Background model initialization: when there is no dynamic object interference, multiple frames of static scene point clouds are collected to construct an initial background model;

[0014] (2) Dynamic background update: periodically compare the newly collected point cloud with the background model to dynamically update the background and avoid false detection of static structural changes as dynamic objects;

[0015] (3) Original point cloud downsampling: voxel grid filtering or random sampling is performed on the high-density laser radar point cloud to reduce the data volume and improve the subsequent processing efficiency; key feature points are retained, including the coal pile surface curvature change area;

[0016] (4) Dynamic region detection: compare the current frame point cloud with the background model to extract the dynamic region through distance threshold method or deep learning segmentation model; label the dynamic point cloud cluster to distinguish between coal pile surface natural deformation and real dynamic objects;

[0017] (5) Dynamic object tracking: multiple target tracking is performed on the detected dynamic objects to predict the motion trajectory; the real-time position, speed and path of the dynamic object are output;

[0018] (6) Multi-view point cloud registration: align the point clouds collected by multiple devices at different times through ICP algorithm or feature matching to eliminate coordinate offset;

[0019] (7) Voxelization processing: convert the registered point cloud into a voxel grid to fill the three-dimensional spatial structure of the coal pile; optimize the voxel density distribution to reduce the void;

[0020] (8) Density field reconstruction: based on the voxelized data, calculate the local point cloud density field, and estimate the coal pile quality by combining the density field and volume data.

[0021] Further: the processing method of the edge computing node is as follows:

[0022] (1) Collect the real-time scanning of the deployed laser radar on the coal pile surface to generate high-precision three-dimensional point cloud data; synchronously access the environmental sensor data to correct the point cloud collection error;

[0023] (2) Real-time reception and cleaning of point cloud data, rejection of noise points, standardization of data format, and adaptation to subsequent processing modules;

[0024] (3) Temporarily store raw point cloud and preprocessing results, support high-efficiency throughput of sudden large-scale scanning data;

[0025] (4) Point cloud registration, align point clouds of multi-view scanning, eliminate coordinate offset caused by device movement or environmental jitter; adopt ICP algorithm for registration, build complete coal pile three-dimensional model; dynamically update registration results;

[0026] (5) Based on coal yard ground point cloud, fit reference plane as volume calculation reference surface, perform surface reconstruction on registered point cloud, generate closed coal pile surface, and calculate coal pile volume;

[0027] (6) Detect and identify the structural risk of the coal pile;

[0028] (7) Real-time volume change rate monitoring module, issue alarm for volume mutation;

[0029] (8) Deploy lightweight AI model on edge computing node to identify foreign matter or coal quality classification in coal pile;

[0030] (9) Push the volume calculation result to the coal yard management system; visualize the three-dimensional coal pile model, label key parameters, and synchronize the volume data to the cloud.

[0031] Further: the construction steps of the digital twin engine are as follows:

[0032] (1) Point cloud data preprocessing and enhancement:

[0033] Receive point cloud data processed by edge computing node; synchronously access coal yard environment parameters and operation log; segment and remove residual interference, retain coal pile main point cloud; based on density field information, enhance high-density area details and smooth low-density noise; use pre-trained model to distinguish semantic categories of coal pile, ground and loading and unloading equipment;

[0034] (2) Three-dimensional model reconstruction and structure optimization:

[0035] 1) Voxelization processing, convert point cloud to high-resolution voxel grid, retain density field data as voxel attribute; dynamically adjust voxel resolution, balance calculation efficiency and model accuracy;

[0036] 2) Surface reconstruction, use MarchingCubes algorithm to generate coal pile surface triangular mesh model from voxel grid;

[0037] 3) Hole repair, based on density field interpolation to fill in missing areas of the model caused by occlusion;

[0038] 4) Build layered model, coarse-grained model for global visualization, fine-grained model for local mechanical analysis;

[0039] (3) Physical property binding:

[0040] Density binding, mapping voxel density field to coal pile mass distribution property;

[0041] Calorific value texture mapping, assigning different regions with calorific value parameters according to coal classification, supporting combustion efficiency simulation;

[0042] Moisture parameter binding, correlating environmental sensor data, labeling coal pile surface humidity gradient.

[0043] Further: the digital twin engine supports multi-dimensional data superimposed display; and provides a section analysis tool to view the internal density and humidity distribution of the coal pile in real time.

[0044] Further: the decision-making process of the decision-making layer is implemented as follows:

[0045] (1) Environmental perception

[0046] 1) Real-time data acquisition, equipment state monitoring, real-time acquisition of the running parameters of the loader and transport vehicle equipment through sensors, including GPS position, oil pressure, engine speed, load state, and remaining oil quantity; synchronizing current job order data from the enterprise ERP system or dispatching platform, including coal type, demand, priority, delivery time, and customer location; integrating weather stations and on-site sensors to collect environmental data and assess the impact of environmental data on job efficiency;

[0047] 2) Multi-source data integration, integrating device status, order demand, and environmental parameters into edge computing nodes for timestamp alignment and data format standardization, and building a global perception data pool;

[0048] (2) State encoding

[0049] 1) Spatial feature extraction, converting three-dimensional point cloud data scanned by the coal yard laser radar into voxel grid, preserving the spatial features of the coal pile; identifying key areas including loading and unloading areas, roads, and coal storage areas based on semantic segmentation models, and encoding as a spatial topology graph;

[0050] 2) Time series feature modeling, extracting time series data of the device within the past 1 hour to construct a time series matrix; analyzing time series data through a long short-term memory network to capture device state change trends;

[0051] 3) Multi-modal feature fusion, combining spatial voxel features, time series encoding results, and environmental parameters into a unified high-dimensional feature vector as the input state representation of the mathematical optimization model;

[0052] (3) Hybrid optimization solution

[0053] 1) Reinforcement learning strategy generation, call pre-trained reinforcement learning model, generate candidate scheduling strategy set based on high-dimensional feature vector of current state;

[0054] 2) Mathematical programming solution, convert the scheduling problem into a mathematical optimization model, the objective function is to minimize the total cost, the constraint conditions include device capacity, delivery time limit and safety distance; use the solver to solve the mathematical optimization model, output the optimal scheduling scheme;

[0055] 3) Digital twin simulation verification, load the candidate strategy in the digital twin engine, simulate the whole process, evaluate the actual effect; adjust the strategy weight according to the simulation result, select the most robust feasible scheme;

[0056] (4) Instruction issuing

[0057] 1) The optimized scheduling scheme is decomposed into a device executable instruction sequence;

[0058] 2) The instruction is issued to the target device; and real-time receiving device execution state, dynamically triggering the exception handling mechanism.

[0059] Further: After the job is completed, the execution layer feeds back information to the enterprise ERP / MES system.

[0060] The present application has the following advantages: The present application realizes the precise dynamic scheduling and management of coal yard inventory through multi-source real-time data perception, edge computing node rapid processing and digital twin simulation. In addition, by setting up digital twin simulation rehearsal and mechanical analysis, the collapse risk is early warned, and the accident rate is reduced.

[0061] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or will be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more intuitively illustrate the prior art and the present application, the following exemplary drawings are given. It should be understood that the specific shape, structure shown in the drawings should not be regarded as a limitation condition for realizing the present application; for example, based on the technical concept disclosed in the present application and the exemplary drawings, those skilled in the art can easily make routine adjustments or further optimization to some units (components) such as increase / decrease / attribute division, specific shape, positional relationship, connection mode, size ratio relationship, etc.

[0063] Figure 1 A system block diagram of a coal yard inventory scheduling system based on real-time data is provided for an embodiment of the present application.

[0064] Figure 2A data interaction architecture schematic diagram of a coal yard inventory scheduling system based on real-time data is provided.

[0065] Figure 3 A flowchart of point cloud processing in the coal yard inventory scheduling system based on real-time data is provided.

[0066] Figure 4 A structure schematic diagram of an edge computing node in the coal yard inventory scheduling system based on real-time data is provided.

[0067] Figure 5 A flowchart of digital twin engine construction in the coal yard inventory scheduling system based on real-time data is provided.

[0068] Figure 6 A scheduling decision flowchart of the coal yard inventory scheduling system based on real-time data is provided. DETAILED DESCRIPTION

[0069] The embodiments of the present application will be described in detail by the specific embodiments, and the person skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the specification. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. It should be understood that these embodiments are only for further illustration of the present application, and cannot be understood as a limitation on the protection scope of the present application. The technical engineers in the field can make some non-essential improvements and adjustments to the present application according to the content of the above-mentioned application; based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor are within the protection scope of the present application.

[0070] Please refer to Figures 1-2 A coal yard inventory scheduling system based on real-time data, comprising a perception layer, an edge computing layer, a digital twin layer, a decision layer and an execution layer,

[0071] The perception layer is a multi-modal data perception layer, comprising a deployed high-density laser radar matrix (5G+LiDAR), a UAV inspection system equipped with an infrared thermal imager, a dynamic weighing based on RFID binding, an environmental perception network (temperature and humidity / dust / water seepage detection), a device state monitoring sensor, etc.

[0072] The laser radar matrix adopts a Velodyne VLS-128 laser radar (vertical angle of view 40°, 2.4 million points per second); the deployment interval of the laser radar is calculated according to an algorithm, the deployment interval algorithm is d = 2R tan (θ / 2n), wherein R is the radius of the coal pile, θ is the field of view angle of the radar, and n is an overlap rate ≥ 30%, and the laser radar is installed on the tower; the scanning strategy of the laser radar matrix is: automatically adjusting the scanning frequency (1-5 Hz) according to the height of the coal pile; and automatically filtering the interference of moving vehicles (point cloud preprocessing) based on the RANSAC algorithm. Then, the coal pile profile is calculated according to a spatial calibration algorithm, and rain and mist are compensated and dust is filtered to realize anti-interference, the rain and mist compensation adopts a 1550 nm wavelength laser penetration scheme, and the dust filtering is an adaptive threshold algorithm based on a particle swarm optimization.

[0073] The unmanned aerial vehicle inspection system first flies according to a preset path, ensures full coverage of a target area, and maintains a safe distance and a stable attitude. The ground station synchronously views a thermal imaging picture, and marks an abnormal area (such as a local high temperature point); synchronously stores infrared original data, a visible light image, a GPS coordinate and an environmental parameter.

[0074] For the data collected in the flight process, the processing flow is as follows:

[0075] a. Non-uniformity correction (NUC): eliminating the pixel response difference of the infrared detector itself;

[0076] b. Temperature calibration: converting the original signal into a temperature value based on a blackbody radiation reference value;

[0077] c. Humidity compensation: correcting the influence of humidity on thermal radiation according to the humidity data recorded during the inspection;

[0078] d. Atmospheric transmittance correction: adjusting the radiation value of a long-distance target in combination with atmospheric parameters (such as CO2 and water vapor content);

[0079] e. Radiation calibration: converting the corrected data into a standard radiation unit;

[0080] f. Pseudo-color mapping: converting the temperature distribution into an intuitive color image, and highlighting the abnormal area.

[0081] The equipment state monitoring sensor can acquire the running parameters of the loader and the transport vehicle equipment in real time, including a GPS position, an oil pressure, an engine speed, a load state, a remaining oil amount and the like.

[0082] The edge computing layer sets up a node sensor with an edge computing capability in the sensing layer, and the edge computing node can collect the data of the sensor, clean the collected data, and extract an environmental feature vector.

[0083] The edge computing layer can process the raw point cloud data collected by the perception layer and construct a three-dimensional density field.

[0084] Referring to Figure 3 , the point cloud processing flow is as follows:

[0085] (1) Background model initialization: When there is no dynamic object (such as vehicles, personnel) interference, multiple frames of static scene point clouds are collected to construct an initial background model (such as based on statistical mean or Gaussian mixture model).

[0086] (2) Dynamic background update: periodically compare the newly collected point cloud with the background model, and dynamically update the background (such as the shape change of the coal pile due to loading and unloading), to avoid false detection of static structural changes as dynamic objects; use a progressive update algorithm to balance environmental change sensitivity and computational overhead.

[0087] (3) Raw point cloud downsampling: voxel grid filtering or random sampling is performed on high-density lidar point clouds to reduce data volume (for example, from millions of point clouds to hundreds of thousands of point clouds) and improve subsequent processing efficiency; key feature points (such as areas with changes in coal pile surface curvature) are preserved to avoid excessive simplification affecting accuracy.

[0088] (4) Dynamic region detection: compare the current frame point cloud with the background model, and extract the dynamic region (such as moving loading and unloading vehicles, inspection robots) through distance threshold method or deep learning segmentation model; label the dynamic point cloud cluster to distinguish between natural deformation of the coal pile surface and real dynamic objects.

[0089] (5) Dynamic object tracking: multi-target tracking (such as based on Kalman filter or Hungarian algorithm) is performed on the detected dynamic objects to predict the motion trajectory; the real-time position, speed and path of the dynamic object (such as the driving route of the vehicle in the coal yard) are output to support obstacle avoidance and operation scheduling.

[0090] (6) Multi-view point cloud registration: align point clouds collected at different times or by different devices (such as fixed radar and vehicle-mounted radar data) through ICP algorithm or feature matching (FPFH descriptor) to eliminate coordinate offset; for the problem of sparse features on the coal pile surface, semantic auxiliary registration is introduced (such as using the outline of the loading and unloading equipment as a matching reference).

[0091] (7) Voxelization processing: convert the registered point cloud into a voxel grid to fill the three-dimensional spatial structure of the coal pile; optimize the voxel density distribution to reduce voids (such as point cloud missing due to dust shielding).

[0092] (8) Density field reconstruction: based on the voxelized data, calculate the local point cloud density field to reflect the compactness of different regions of the coal pile (such as the density difference between loose surface coal dust and internal compacted coal pile); combine the density field with the volume data to estimate the coal pile mass (requires presetting the coal density parameter).

[0093] Referring to Figure 4 The processing flow of the edge computing node is as follows:

[0094] Data access, collect sensor data, that is, collect real-time scanning of the surface of the coal pile by the deployed LiDAR (LiDAR), and generate high-precision three-dimensional point cloud data; synchronously access environmental sensor data (such as temperature and humidity, wind speed), and correct point cloud collection errors.

[0095] Stream processing engine, real-time reception and cleaning of point cloud data, rejection of noise points (such as flying birds, temporary equipment interference), standardization of data format (such as conversion to PCD or PLY format), and adaptation to subsequent processing modules.

[0096] Data caching, temporary storage of original point cloud and preprocessing results, supporting efficient throughput of sudden large-scale scanning data.

[0097] Point cloud registration, aligning point clouds of multi-view scanning (such as multi-circle scanning data of the unmanned aerial vehicle flying around the coal pile), eliminating coordinate offset caused by device movement or environmental jitter; using ICP algorithm or NDT (normal distribution transformation) registration, constructing a complete three-dimensional model of the coal pile; dynamically updating the registration result to adapt to real-time changes in the shape of the coal pile due to loading and unloading operations.

[0098] Volume calculation, based on the ground point cloud of the coal yard, fitting a reference plane as a volume calculation reference surface, surface reconstruction (such as triangulation or voxelization) of the registered point cloud, and generation of a closed coal pile surface; through spatial integration method or slice projection method, the volume of the coal pile (unit: cubic meter) is calculated, with an error control within ±2%; re-calculation is triggered every hour or on demand, supporting real-time inventory statistics.

[0099] Abnormality detection, detecting and identifying structural risks such as coal pile collapse and slope overrun (such as using curvature analysis or comparison with historical models).

[0100] Real-time calculation, the real-time calculation module monitors the volume change rate, and issues an alarm for sudden volume changes, such as triggering an alarm during abnormal loading and unloading (such as theft or dumping accidents).

[0101] Model inference, deploying a lightweight AI model (such as PointNet++) to identify foreign objects (such as mixed debris) in the coal pile or classify coal quality (such as distinguishing between lump coal and pulverized coal). The lightweight AI model is accelerated through TensorRT and achieves a second-level inference response on the edge.

[0102] Result output, pushing the volume calculation result to the coal yard management system to generate inventory daily / weekly reports (including volume, density estimation, and calorific value analysis); visualizing the three-dimensional coal pile model and labeling key parameters (maximum height, base area, volume change trend).

[0103] Closed-loop control: abnormal events (such as landslides) trigger audible and visual alarms, and automatically dispatch inspection drones for re-measurement. Volume data is synchronized to the cloud for long-term trend analysis and supply chain optimization.

[0104] Digital twin layer, receiving data from edge computing layer, constructs a three-dimensional dynamic digital twin engine (digital twin model), and fuses multi-spectral data to map out heat value distribution in the digital twin engine.

[0105] Referring to Figure 5 , the construction steps of the digital twin engine are as follows:

[0106] (1) Point cloud data preprocessing and enhancement

[0107] Laser point cloud input: receiving point cloud data processed by edge computing nodes (registration, dynamic object removal, and density field reconstruction have been completed); synchronously accessing coal yard environment parameters (temperature, humidity, wind speed) and operation logs (loading and unloading time, coal type).

[0108] Noise filtering and optimization: RANSAC algorithm: segment and remove residual interference (such as temporary equipment, flying birds), and retain the main point cloud of the coal pile.

[0109] Density field guided filtering: based on density field information, enhance the details of high-density areas, and smooth low-density noise (such as loose coal powder on the surface).

[0110] Semantic segmentation: use pre-trained models to distinguish semantic categories such as coal piles, ground, and loading and unloading equipment, and provide labels for subsequent physical property binding.

[0111] (2) Three-dimensional model reconstruction and structure optimization

[0112] Voxelization processing: convert point cloud to high-resolution voxel grid (such as 0.05m³), and retain density field data as voxel attributes; dynamically adjust voxel resolution to balance calculation efficiency and model accuracy (such as using fine voxels in frequently loaded and unloaded areas).

[0113] Surface reconstruction: use MarchingCubes algorithm to generate coal pile surface triangle mesh model from voxel grid, ensuring smooth transition and closed surface.

[0114] Hole repair: based on density field interpolation to fill in missing areas of the model caused by obstruction (such as dust, equipment).

[0115] Multi-scale modeling: construct hierarchical models, coarse-grained models for global visualization, and fine-grained models for local mechanical analysis.

[0116] (3) Physical property binding and simulation empowerment

[0117] Physical property mapping includes:

[0118] Density binding, mapping voxel density field to coal pile mass distribution attribute (unit: kg / m³);

[0119] Calorific value texture mapping, assigning different regions with calorific value parameters (unit: MJ / kg) according to coal type classification (e.g. lignite, anthracite), supporting combustion efficiency simulation;

[0120] Moisture parameter binding, correlating environmental sensor data, labeling coal pile surface layer moisture gradient;

[0121] Mechanical simulation and stability analysis includes:

[0122] Finite element analysis (FEA), simulating stress distribution under coal pile stress (e.g. rainfall infiltration, mechanical compaction), predicting collapse risk areas;

[0123] Slope safety warning, combining coal pile geometric model and friction coefficient, calculating critical slope and marking over-limit areas;

[0124] Dynamic update mechanism:

[0125] Real-time synchronization of edge node processing results (e.g. coal pile deformation caused by new loading and unloading), triggering local model reconstruction and attribute update.

[0126] Interactive three-dimensional twin interface is constructed, supporting multi-dimensional data superimposed display (e.g. calorific value distribution heat map, stress cloud map). Provides section analysis tools to view coal pile internal density and moisture distribution in real time.

[0127] Based on volume and density data, automatically calculate coal pile mass, generate inventory report; finite element analysis results trigger collapse warning, push to inspection personnel handheld terminal; combined with calorific value distribution and boiler parameters, optimize coal scheduling strategy.

[0128] Lightweight twin model is retained on the edge, supporting real-time alarm and local decision; high-precision model and historical data are synchronized to the cloud for long-term trend analysis (e.g. coal quality degradation prediction).

[0129] Referring to Figure 6 , decision layer, after receiving order information, according to the relevant feature vector preprocessed by edge computing node (including real-time collected equipment state data, environmental data, etc.), running hybrid optimization algorithm, simulating and rehearsing based on the virtual coal yard with physical characteristics constructed by digital twin layer, then generating equipment control instructions and issuing to execution layer.

[0130] The implementation steps of the decision-making process of the decision-making layer are as follows:

[0131] (1) Environmental perception

[0132] 1) Real-time data acquisition: Equipment status monitoring, real-time acquisition of loader, transport vehicle equipment operating parameters through sensors, including GPS position, oil pressure, engine speed, load status, remaining oil quantity, etc.; synchronize current job order data from enterprise ERP system or dispatch platform, including coal type, demand, priority, delivery time and customer location; integrate weather stations and field sensors to collect wind speed, temperature, humidity, visibility and other environmental data to assess their impact on job efficiency (such as strong wind weather limiting unmanned aerial vehicle inspection).

[0133] 2) Multi-source data integration: Integrate equipment status, order demand and environmental parameters into edge computing nodes, align timestamps and standardize data formats, and build a global perception data pool.

[0134] (2) State encoding

[0135] 1) Spatial feature extraction: Convert 3D point cloud data scanned by laser radar in the coal yard to a voxel grid with a resolution of 128x128x128, retaining spatial features such as coal pile volume and density distribution; based on semantic segmentation models, identify key areas such as loading and unloading areas, roads, and coal storage areas, and encode them as spatial topology graphs.

[0136] 2) Time series feature modeling: Extract time series data such as device running track, job frequency, and fault records within the past 1 hour, and build a time series matrix; analyze time series data through long short-term memory network (LSTM) to capture device state change trends (such as oil consumption attenuation law, load fluctuation period).

[0137] 3) Multi-modal feature fusion: Combine spatial voxel features, time series encoding results, and environmental parameters into a unified high-dimensional feature vector as the input state representation of the optimization model.

[0138] (3) Hybrid optimization solution

[0139] 1) Reinforcement learning strategy generation: Call pre-trained reinforcement learning models (such as PPO, DQN) to generate a set of candidate scheduling strategies based on current state features, such as vehicle path planning and loader task allocation priority.

[0140] 2) Mathematical programming solution: Convert the scheduling problem into a mathematical optimization model, with the objective function being to minimize total cost (fuel consumption, time delay, etc.), and constraints including device capacity, delivery deadline and safety distance; use solvers such as CPLEX, Gurobi to efficiently solve the model, output the optimal scheduling scheme (such as vehicle-coal pile matching relationship, loading and unloading time window).

[0141] 3) Digital twin simulation verification: load the candidate strategy in the digital twin engine, simulate the entire process (such as vehicle movement, coal pile loading and unloading), evaluate the actual effect (such as task completion rate, resource utilization rate); adjust the strategy weight according to the simulation result, and select the most robust feasible solution.

[0142] (4) Instruction issuing

[0143] 1) Control instruction generation: decompose the optimized scheduling scheme into device executable instruction sequence, such as loader operation path coordinates, transport vehicle navigation route, and loading and unloading robot action parameters.

[0144] 2) Real-time issuing and execution: issue instructions to target devices through time-sensitive network (TSN); and receive device execution status (such as position update, task progress) in real time, dynamically trigger abnormal processing mechanism (such as path re-planning, fault device replacement).

[0145] The execution layer receives the instructions issued by the decision layer and performs the work according to the instructions.

[0146] After the work is completed, the information is fed back to the enterprise ERP / MES system.

[0147] In this embodiment, the perception layer and the edge computing layer use a lightweight communication protocol; the digital twin layer and the decision layer interact through a state space interface; and the execution layer device supports plug and play.

[0148] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A coal yard inventory scheduling system based on real-time data, characterized by, The system comprises a perception layer, an edge computing layer, a digital twin layer, a decision layer, and an execution layer. The perception layer comprises a high-density laser radar matrix, an infrared thermal imager unmanned aerial vehicle inspection system, a dynamic weighing of a ground balance, an environment perception network, and a device state monitoring sensor. The edge computing layer uses several sensors with computing and transmission capabilities in the perception layer as edge computing nodes. The edge computing nodes can collect data from the perception layer, clean the collected data, and extract feature vectors of the environment and device state. The edge computing nodes perform point cloud processing on the point cloud data of the perception layer and construct a three-dimensional density field. The digital twin layer receives data from the edge computing layer, constructs a three-dimensional dynamic digital twin engine, and maps a heat value distribution in the digital twin engine by fusing multispectral data. The decision layer receives order information, runs a hybrid optimization algorithm based on the feature vectors preprocessed by the edge computing nodes, simulates a virtual coal yard with physical characteristics constructed by the digital twin layer, generates device control instructions, and sends the instructions to the execution layer. The execution layer receives the instructions from the decision layer and performs operations according to the instructions. The point cloud processing process is as follows: (1) Background model initialization: When there is no dynamic object interference, multiple frames of static scene point clouds are collected to construct an initial background model. (2) Dynamic background update: Periodically compare the newly collected point cloud with the background model to dynamically update the background and avoid false detection of static structural changes as dynamic objects. (3) Original point cloud downsampling: Perform voxel grid filtering or random sampling on high-density laser radar point clouds to reduce data volume and improve subsequent processing efficiency. (4) Dynamic region detection: Compare the current frame point cloud with the background model to extract dynamic regions through distance thresholding or deep learning segmentation models. (5) Dynamic object tracking: Track the detected dynamic objects and predict their motion trajectories. Output the real-time position, velocity, and path of the dynamic objects. (6) Multi-view point cloud registration: Align point clouds collected at different times or by different devices using ICP algorithm or feature matching to eliminate coordinate offsets. (7) Voxelization processing: Convert the registered point cloud into a voxel grid to fill the three-dimensional spatial structure of the coal pile. (8) Density field reconstruction: Calculate the local point cloud density field based on the voxelized data, combine the density field with the volume data, and estimate the coal pile quality. The decision-making process of the decision layer is as follows: a. Environment perception 1) Real-time data acquisition and equipment state monitoring. Real-time acquisition of the operating parameters of the loader and transport vehicle equipment through sensors, including GPS position, oil pressure, engine speed, load state, and remaining oil quantity. Synchronization of current job order data from the enterprise ERP system or dispatch platform, including coal type, demand quantity, priority, delivery time, and customer location. Integration of weather stations and on-site sensors to collect environmental data and assess the impact of environmental data on job efficiency. 2) Multi-source data integration. Integration of equipment status, order demand, and environmental parameters into the edge computing node for timestamp alignment and data format standardization to build a global perception data pool. b. State encoding 1) Spatial feature extraction. Conversion of three-dimensional point cloud data scanned by the coal yard laser radar into a voxel grid to preserve the spatial features of the coal pile. Identification of key areas including loading and unloading areas, roads, and coal storage areas based on a semantic segmentation model and encoding as a spatial topology graph. 2) Time series feature modeling. Extraction of time series data from the past 1 hour of equipment to construct a time series matrix. Analysis of time series data through a long short-term memory network to capture trends in equipment state changes. 3) Multi-modal feature fusion. Combining spatial voxel features, time series encoding results, and environmental parameters into a unified high-dimensional feature vector as the input state representation of the mathematical optimization model. c. Mixed optimization solution 1) Reinforcement learning strategy generation. Calling a pre-trained reinforcement learning model to generate a candidate set of scheduling strategies based on the high-dimensional feature vector of the current state. 2) Mathematical programming solution. Converting the scheduling problem into a mathematical optimization model with a target function of minimizing total cost and constraint conditions including equipment capacity, delivery deadline, and safety distance. Solving the mathematical optimization model using a solver to output the optimal scheduling scheme. 3) Digital twin simulation verification. Loading the candidate strategy in the digital twin engine to simulate the entire process and evaluate the actual effect. Adjusting the strategy weights based on the simulation results to select the most robust feasible solution. d. Instruction issuance 1) Decomposing the optimized scheduling scheme into a sequence of device executable instructions. 2) Issuing the instructions to the target device and receiving the device execution status in real time to dynamically trigger the exception handling mechanism. The processing method of the edge computing node is as follows: S1. Real-time scanning of the coal pile surface by the deployed laser radar to generate high-precision three-dimensional point cloud data. Synchronizing environmental sensor data to correct point cloud acquisition errors. S2. Real-time reception and cleaning of point cloud data, removing noise points, standardizing data formats, and adapting to subsequent processing modules. S3. Temporarily storing raw point cloud and preprocessing results to support efficient throughput of large-scale scanning data in emergencies. S4. Point cloud registration. Aligning point clouds from multiple perspectives to eliminate coordinate shifts caused by device movement or environmental jitter. Using the ICP algorithm for registration to build a complete three-dimensional model of the coal pile. Dynamically updating the registration results. S5. Fitting a reference plane based on the coal yard ground point cloud as a volume calculation reference surface. Reconstructing the surface of the registered point cloud to generate a closed coal pile surface and calculating the coal pile volume. S6. Detecting and identifying the structural risk of the coal pile. S7. The real-time calculation module monitors the volume change rate and issues an alarm for volume mutations; S8. A lightweight AI model is deployed based on the edge computing node to identify foreign objects or coal quality classification in the coal pile; S9. Push the volume calculation results to the coal yard management system; visualize the three-dimensional coal pile model, label key parameters, and synchronize the volume data to the cloud; The construction steps of the digital twin engine are as follows: I. Point cloud data preprocessing and enhancement: Receive the point cloud data processed by the edge computing node; synchronously access the coal yard environment parameters and operation logs; segment and remove residual interference, and retain the main point cloud of the coal pile; Based on the density field information, enhance the details of high-density areas and smooth low-density noise; use a pre-trained model to distinguish the semantic categories of coal piles, ground, and loading and unloading equipment to provide labels for subsequent physical property binding; II. Three-dimensional model reconstruction and structure optimization: A1. Voxelization processing converts point cloud to high-resolution voxel grid, retaining density field data as voxel attributes; dynamically adjust voxel resolution to balance calculation efficiency and model accuracy; A2. Surface reconstruction uses the Marching Cubes algorithm to generate a coal pile surface triangle mesh model from the voxel grid; A3. Hole repair based on density field interpolation fills in missing areas of the model caused by occlusion; A4. Build a layered model, coarse-grained model for global visualization, fine-grained model for local mechanical analysis; III. Physical property binding: Density binding maps voxel density field to coal pile mass distribution attribute; Heat value texture mapping assigns different regions with heat value parameters according to coal type classification, supporting combustion efficiency simulation; Humidity parameter binding correlates environmental sensor data to label coal pile surface layer humidity gradient.

2. The coal yard inventory scheduling system based on real-time data according to claim 1, characterized in that, The deployment interval of the laser radar is calculated according to the algorithm, and the deployment interval algorithm is: d=2R·tan(θ / 2n), where R is the radius of the coal pile, θ is the radar field of view angle, n is the overlap rate, n≥30%, and the laser radar is installed on the tower.

3. The coal yard inventory scheduling system based on real-time data according to claim 1, characterized in that, The digital twin engine supports multi-dimensional data superimposed display; and provides a section analysis tool to view the internal density and humidity distribution of the coal pile in real time.

4. The coal yard inventory scheduling system based on real-time data according to claim 1, characterized in that, After the operation is completed, the execution layer feeds back information to the enterprise ERP / MES system.

5. The coal stockyard inventory scheduling system based on real-time data according to claim 1, characterized in that, The infrared thermal imager unmanned aerial vehicle inspection system is first flown according to the preset path to ensure full coverage of the target area, maintain a safe distance and stable attitude; the ground station synchronously views the thermal imaging picture and marks the abnormal area; synchronously stores the infrared raw data, visible light image, GPS coordinates and environmental parameters.

Images