Coal pile weight calculation and dynamic weighing method based on intelligent sensor

Abstract

The invention discloses a coal pile weight calculation and dynamic weighing method based on an intelligent sensor. The method comprises the following steps: sparsely deploying intelligent sensing nodes at the bottom of a coal pile and acquiring initial data; constructing a double-loop model of an edge online sequence extreme learning machine and a cloud physical information neural network; the edge model processes the real-time data stream and outputs a fast weight estimation value and a virtual observation residual error; the cloud model reconstructs full-field stress distribution by using the data and calculates a virtual sensing value; adjusting an edge model learning target through virtual value feedback between double rings; carrying out integration on the reconstructed stress field to obtain accurate total mass and local mass, and carrying out visualization; and monitoring residual features to trigger early warning and adaptively adjust physical parameters of the model. According to the method, high-precision, continuous, dynamic and self-adaptive direct weighing of the mass of the coal pile is realized under the condition of low-cost sparse sensing deployment, and the contradiction among dynamics, precision and engineering feasibility of a traditional method is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things and intelligent sensor data processing technology, and in particular to a method for calculating and dynamically weighing coal piles based on intelligent sensors. Background Technology

[0002] In industrial sectors involving bulk material storage and transportation, such as coal, power, and ports, accurate and real-time acquisition of coal inventory is fundamental for inventory management, cost accounting, and production scheduling. Currently, the mainstream technical approaches in this field mainly revolve around two types of methods: static inventory and dynamic measurement.

[0003] Static inventory methods aim at periodic inventory checks. Manual measurement and volumetric density estimation are simple to operate but have significant errors; their accuracy is greatly affected by the surveyor's experience and the material's density value. Three-dimensional laser scanning technology uses point cloud data of the coal pile surface to construct a three-dimensional model to calculate volume, improving the automation and accuracy of volume measurement. However, the final mass still relies on empirical density conversion, and the system cost is high, making continuous monitoring impossible. Aerial photogrammetry is suitable for large-scale open-pit coal yards, but it also faces the density error problem in converting volume to mass.

[0004] In recent years, there has been some exploration of inventory monitoring methods based on deploying weighing sensors at the bottom of coal piles or on supporting structures. While this method theoretically allows for direct weighing, it faces significant challenges in large-scale industrial applications: accurately sensing non-uniformly distributed coal pile loads requires high-density sensor deployment beneath the load-bearing surface, leading to complex engineering and high costs; stress transmission within the coal pile exhibits high nonlinearity, making it difficult for sparsely deployed sensors to reconstruct the overall stress distribution through simple interpolation; environmental factors such as foundation settlement and temperature changes introduce significant interference, affecting long-term measurement stability. Existing technologies have clear limitations: high-precision static inventory techniques cannot meet the demands of dynamic real-time monitoring, while ideal continuous direct weighing technology is difficult to implement due to cost, engineering feasibility, and model accuracy issues.

[0005] Therefore, how to provide a method for calculating and dynamically weighing coal piles based on intelligent sensors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for calculating and dynamically weighing coal piles based on intelligent sensors. Addressing the challenge of achieving high-precision, continuous, and dynamic direct weighing of coal piles under engineering-feasible conditions using existing technologies, this invention proposes a dual-loop collaborative solution based on edge computing and a cloud-based physical model. This method deploys a sparse sensor array to acquire data, uses an online sequential extreme learning machine on the edge side for rapid estimation and residual generation, and employs a cloud-based physical information neural network for full-field stress reconstruction and virtual sensor value calculation conforming to mechanical laws. Online iterative optimization of the model is achieved through residual feedback between the two loops. This invention effectively integrates the speed advantages of temporal learning with the accuracy constraints of a physical model, significantly reducing system deployment and maintenance costs while achieving high-precision, continuous, and adaptive dynamic sensing of coal pile quality.

[0007] A method for calculating and dynamically weighing the weight of a coal pile based on a smart sensor, according to an embodiment of the present invention, includes the following steps: S1. Deploy intelligent sensing nodes in a two-dimensional sparse grid under the bearing layer at the bottom of the coal pile, record the physical coordinates of each node, collect sensor data under no-load and known standard load conditions, and obtain an initial training data set. S2. Construct an edge computing model based on an adaptive sliding window online sequence extreme learning machine, and construct a physical information neural network model embedded with multi-physics constraints; S3. The edge computing model collects real-time data streams from intelligent sensing nodes, updates the online sequence extreme learning machine through an adaptive sliding window mechanism, obtains a fast estimate of the total weight of the coal pile, and calculates the virtual observation residual vector. S4. The physical information neural network model receives the fast estimate and the virtual observation residual vector. It uses the sensor readings and fast estimate corrected by the virtual observation residual vector as input constraints, and calculates the full-field two-dimensional stress distribution with the coal pile base coordinate points as input. Based on this distribution, it calculates the virtual sensing value at the coordinates of the smart sensing node in reverse. S5. The physical information neural network model sends the virtual sensing value to the edge computing model. The edge computing model compares the virtual sensing value with the original real-time data stream to obtain the residual feedback signal, and uses the signal to adjust the learning objective function of the online sequence extreme learning machine. S6. Perform numerical integration on the two-dimensional stress distribution across the entire field to obtain accurate total mass values ​​of the coal pile and local mass values, and generate a visual state diagram containing stress cloud maps based on the mass values ​​and stress distribution. S7. Monitor the long-term statistical characteristics of the virtual observation residual vector. When the statistical characteristics meet the preset abnormal mode, trigger the system warning and adjust the preset equivalent physical parameters in the physical information neural network model according to the warning result.

[0008] Optionally, S1 specifically includes: S11. Determine the basic boundary of the bottom bearing layer of the coal pile, and plan a two-dimensional rectangular projection grid covering the expected accumulation range of the coal pile based on the basic boundary. S12. On the nodes of the two-dimensional rectangular projection grid, select target deployment points according to the preset sparse layout rules. The sparse layout rules require that the grid spacing between adjacent target deployment points be greater than the preset threshold. S13. Install intelligent sensing nodes at each selected target deployment point. The intelligent sensing nodes integrate pressure sensors, temperature sensors and triaxial tilt sensors. S14. Establish a global coordinate system, measure and record the three-dimensional physical coordinates of each installed smart sensor node in the global coordinate system. The three-dimensional physical coordinates include planar coordinates and elevation values. S15. Control the coal pile to be in a completely unloaded state, collect the pressure, temperature and tilt angle readings of the smart sensor node at this time, and record them as the initial unloaded dataset. S16. Apply a standard load of known total mass to the bearing layer at the bottom of the coal pile. After the load stabilizes, collect the pressure, temperature and tilt angle readings of the smart sensor node and record them as the standard load dataset. S17. Merge and label the initial empty dataset with the standard load dataset to form the initial training data set.

[0009] Optionally, S2 specifically includes: S21. Based on the initial training data set, determine the input dimension and output dimension of the edge computing model. The input dimension is determined by the number of smart sensing nodes and the number of sensor types. The output dimension is one-dimensional, corresponding to the total weight of the coal pile. S22. Construct the network structure of an online sequence extreme learning machine, set the number of nodes in the input layer, the number of nodes in the hidden layer, and the activation function, and randomly generate the connection weights from the input layer to the hidden layer and the bias of the hidden layer. S23. Set the initial length of the adaptive sliding window and define the sliding window length adjustment rules based on the data error within the window; S24. Define the input of the physical information neural network model as the two-dimensional coordinate points of the coal pile base region, and the output as the normal stress value at that coordinate point; S25. Add balance equation constraint terms and sensor data matching constraint terms to the loss function of the physical information neural network model. The balance equation constraint terms are composed of the stress distribution and gravity balance relationship. S26. Using the standard load dataset in the initial training dataset, initialize and train the network parameters of the physical information neural network model.

[0010] Optionally, S3 specifically includes: S31. The edge computing model collects real-time pressure, temperature and tilt angle readings from all smart sensing nodes at a fixed sampling frequency, and concatenates the readings of all sensors at each sampling moment into a real-time sensing vector according to the node order. S32. Store the real-time sensing vectors into the data buffer in chronological order. The length of the data buffer is greater than or equal to the current length of the adaptive sliding window. S33. Determine whether the number of real-time sensing vectors stored in the data buffer has reached the current length of the adaptive sliding window. If not, continue collecting data. S34. Take a sequence of real-time sensing vectors from the data buffer that is equal in number to the current length of the adaptive sliding window, and use it as training samples for the current window. S35. Input the training samples of the current window into the online sequence extreme learning machine, execute the online weight update algorithm, and update the hidden layer output weight matrix of the online sequence extreme learning machine. S36. Use the updated online sequence extreme learning machine to perform forward computation on the last real-time sensing vector in the current window training samples to obtain a fast estimate of the total weight of the coal pile. S37. Calculate the online sequence extreme learning machine output value corresponding to all real-time sensing vectors in the current window training samples, compare the output value with the reference weight value converted based on the physical relationship of the sensor reading at the corresponding time, obtain a set of difference sequences, perform statistical analysis on the difference sequences, and generate virtual observation residual vectors.

[0011] Optionally, S4 specifically includes: S41. The physical information neural network model receives fast estimates and virtual observation residual vectors from the edge computing model, and simultaneously receives raw sensor readings with timestamps corresponding to the fast estimates. S42. Add the virtual observation residual vector and the original sensor readings algebraically according to the corresponding sensor nodes to obtain the sensor readings after correction by the virtual observation residual vector. S43. The fast estimate is converted into an equivalent uniformly distributed load boundary condition, which together with the corrected sensor readings constitutes the input constraint set of the physical information neural network model. S44. Within the coal pile base area, generate a dense two-dimensional coordinate point grid according to a preset resolution, and input all coordinate points into the physical information neural network model. S45. The physical information neural network model uses the input constraint set as the training target, executes the backpropagation optimization algorithm, adjusts the network parameters, and continues until the loss function converges. The loss function includes balance equation constraint terms and sensor data matching constraint terms. S46. Perform forward calculations on the dense two-dimensional coordinate point grid using the converged physical information neural network model, and output the normal stress value corresponding to each coordinate point to form a full-field two-dimensional stress distribution. S47. Extract the normal stress value at the coordinate point corresponding to the physical coordinate position of the smart sensing node from the two-dimensional stress distribution of the whole field. Convert the normal stress value into an equivalent pressure reading through the sensor calibration relationship and use it as the virtual sensing value at the coordinate of the smart sensing node.

[0012] Optionally, S5 specifically includes: S51. The physical information neural network model sends the calculated virtual sensing values ​​to the edge computing model through a communication link. S52. The edge computing model extracts the original sensor pressure readings corresponding to the timestamps of the virtual sensing values ​​from the data buffer to form the original pressure reading vector. S53. Match the received virtual sensing values ​​with the original pressure reading vector according to the same smart sensing node index, calculate the difference node by node, and form the original difference vector. S54. Perform low-pass filtering on the original difference vector to remove high-frequency noise components and obtain a smoothed residual feedback signal vector. S55. Extract the mean and covariance matrix of the residual feedback signal vector as adjustment parameters for the learning objective function of the online sequence extreme learning machine; S56. Using the mean of the residual feedback signal vector as the bias correction term and the inverse of the covariance matrix of the residual feedback signal vector as the weight regularization matrix, a new learning objective function for the online sequence extreme learning machine is constructed. S57. In the next execution of the online sequence extreme learning machine weight update, replace the original learning objective function with a new learning objective function that includes a bias correction term and a weight regularization matrix.

[0013] Optionally, S6 specifically includes: S61. Obtain the full-field two-dimensional stress distribution data output by the physical information neural network model. This data includes the normal stress value corresponding to each coordinate point in the coal pile base area. S62. Perform area-weighted summation of the normal stress values ​​at all coordinate points in the full-field two-dimensional stress distribution data, and divide the summation result by the gravitational acceleration constant to obtain the total mass value of the coal pile. S63. Define a polygonal boundary within the coal pile base area. The polygonal boundary is determined by the coordinates of multiple vertices. S64. Extract the normal stress values ​​of the coordinate points located inside the polygon boundary from the full-field two-dimensional stress distribution data, perform area-weighted summation on these normal stress values, and divide the summation result by the gravitational acceleration constant to obtain the local region mass value of the region defined by the polygon boundary. S65. Based on the normal stress value and position coordinates of each coordinate point in the full-field two-dimensional stress distribution data, a two-dimensional stress cloud map is generated using a color mapping algorithm. S66. Mark the total mass value of the coal pile, the mass value of the local area, and the sequence of changes in the total mass of the coal pile over time in numerical and graphical form at the preset positions of the two-dimensional stress cloud map. S67. Combine the two-dimensional stress cloud map labeled with quality information with the independent quality change time series curve to generate a real-time status map of the coal pile digital twin.

[0014] Optionally, S7 specifically includes: S71. Collect virtual observation residual vectors and store them in chronological order to form a historical residual sequence; S72. Calculate the mean vector and covariance matrix of each virtual observation residual vector in the historical residual sequence to obtain the residual statistical characteristic sequence; S73. Match the residual statistical feature sequence with the preset anomaly pattern library, which includes sensor drift mode, foundation settlement mode and coal pile internal structure anomaly mode. S74. When the offset direction of the mean vector of the residual statistical feature sequence and the trend of the change of the eigenvalue of the covariance matrix meet the judgment conditions of any pattern in the preset abnormal pattern library, it is determined that the preset abnormal pattern has been detected. S75. Generate an early warning signal containing the anomaly type, location, and level based on the matched anomaly pattern type. S76. Based on the abnormal type and level of the warning signal, query the preset parameter adjustment mapping table to obtain the corresponding physical information neural network model equivalent physical parameter adjustment amount; S77. Apply the obtained equivalent physical parameter adjustment amount to the physical information neural network model to complete the parameter adjustment.

[0015] The beneficial effects of this invention are: (1) By deploying a sparse intelligent sensor array and constructing a dual-ring collaborative architecture of an edge fast response ring and a cloud precise reconstruction ring, this invention realizes online physical consistency reconstruction and real-time quality calculation of the stress distribution of the entire coal pile, effectively reducing the system hardware deployment cost and engineering complexity, breaking through the technical bottleneck between accuracy and feasibility of traditional dense sensor deployment or single data-driven models, and enhancing the reliability of intelligent metering and status perception in industrial bulk material stockpiles.

[0016] (2) By adopting the coupled iterative mechanism of online sequence extreme learning machine and physical information neural network, the present invention realizes the closed-loop feedback optimization between fast estimation value and high-precision physical field, which significantly improves the adaptive learning ability and long-term measurement stability in the dynamic weighing process, and shows better robustness and adaptability in complex industrial application scenarios where the coal pile shape changes continuously and the sensor has slow drift.

[0017] (3) In terms of realizing continuous dynamic direct weighing of coal piles, this invention effectively solves the problem of insufficient measurement accuracy caused by the difficulty of reflecting the mechanical state of the whole field due to sparse sensing data through a series of technical means such as virtual observation residual generation, physical constraint loss function construction and online adjustment of double-loop model parameters. It breaks through the dependence of traditional methods on empirical density conversion or high-cost dense deployment, and realizes the technological progress from indirect volume measurement to direct mechanical sensing and from static periodic inventory to dynamic continuous monitoring, effectively improving the digital and refined management level of coal storage and logistics. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the process for calculating and dynamically weighing the weight of a coal pile based on intelligent sensors, as proposed in this invention.

[0019] Figure 2 This is an architecture diagram of the dual-ring collaborative computing model proposed in this invention. Detailed Implementation

[0020] Combination Figures 1-2 The present invention will be described in further detail below. These accompanying drawings are simplified schematic diagrams, illustrating only the basic structure of the invention and showing the main components relevant to the invention. Figure 1 and Figure 2 The present invention provides a method for calculating and dynamically weighing the weight of a coal pile based on intelligent sensors, comprising the following steps: S1. Deploy intelligent sensing nodes in a two-dimensional sparse grid under the bearing layer at the bottom of the coal pile, record the physical coordinates of each node, collect sensor data under no-load and known standard load conditions, and obtain an initial training data set. S2. Construct an edge computing model based on an adaptive sliding window online sequence extreme learning machine, and construct a physical information neural network model embedded with multi-physics constraints; S3. The edge computing model collects real-time data streams from intelligent sensing nodes, updates the online sequence extreme learning machine through an adaptive sliding window mechanism, obtains a fast estimate of the total weight of the coal pile, and calculates the virtual observation residual vector. S4. The physical information neural network model receives the fast estimate and the virtual observation residual vector. It uses the sensor readings and fast estimate corrected by the virtual observation residual vector as input constraints, and calculates the full-field two-dimensional stress distribution with the coal pile base coordinate points as input. Based on this distribution, it calculates the virtual sensing value at the coordinates of the smart sensing node in reverse. S5. The physical information neural network model sends the virtual sensing value to the edge computing model. The edge computing model compares the virtual sensing value with the original real-time data stream to obtain the residual feedback signal, and uses the signal to adjust the learning objective function of the online sequence extreme learning machine. S6. Perform numerical integration on the two-dimensional stress distribution across the entire field to obtain accurate total mass values ​​of the coal pile and local mass values, and generate a visual state diagram containing stress cloud maps based on the mass values ​​and stress distribution. S7. Monitor the long-term statistical characteristics of the virtual observation residual vector. When the statistical characteristics meet the preset abnormal mode, trigger the system warning and adjust the preset equivalent physical parameters in the physical information neural network model according to the warning result.

[0021] In this embodiment, S1 specifically includes: S11. Determine the basic boundary of the bottom bearing layer of the coal pile, and plan a two-dimensional rectangular projection grid covering the expected accumulation range of the coal pile based on the basic boundary. S12. On the nodes of the two-dimensional rectangular projection grid, select target deployment points according to the preset sparse layout rules. The sparse layout rules require that the grid spacing between adjacent target deployment points be greater than the preset threshold. S13. Install intelligent sensing nodes at each selected target deployment point. The intelligent sensing nodes integrate pressure sensors, temperature sensors and triaxial tilt sensors. S14. Establish a global coordinate system, measure and record the three-dimensional physical coordinates of each installed smart sensor node in the global coordinate system. The three-dimensional physical coordinates include planar coordinates and elevation values. S15. Control the coal pile to be in a completely unloaded state, collect the pressure, temperature and tilt angle readings of the smart sensor node at this time, and record them as the initial unloaded dataset. S16. Apply a standard load of known total mass to the bearing layer at the bottom of the coal pile. After the load stabilizes, collect the pressure, temperature and tilt angle readings of the smart sensor node and record them as the standard load dataset. S17. Merge and label the initial empty dataset with the standard load dataset to form the initial training data set.

[0022] In this embodiment, S2 specifically includes: S21. Based on the initial training data set, determine the input dimension and output dimension of the edge computing model. The input dimension is determined by the number of smart sensing nodes and the number of sensor types. The output dimension is one-dimensional, corresponding to the total weight of the coal pile. S22. Construct the network structure of an online sequence extreme learning machine, set the number of nodes in the input layer, the number of nodes in the hidden layer, and the activation function, and randomly generate the connection weights from the input layer to the hidden layer and the bias of the hidden layer. S23. Set the initial length of the adaptive sliding window and define the sliding window length adjustment rules based on the data error within the window; S24. Define the input of the physical information neural network model as the two-dimensional coordinate points of the coal pile base region, and the output as the normal stress value at that coordinate point; S25. Add balance equation constraint terms and sensor data matching constraint terms to the loss function of the physical information neural network model. The balance equation constraint terms are composed of the stress distribution and gravity balance relationship. S26. Using the standard load dataset in the initial training dataset, initialize and train the network parameters of the physical information neural network model.

[0023] In this embodiment, S3 specifically includes: S31. The edge computing model collects real-time pressure, temperature and tilt angle readings from all smart sensing nodes at a fixed sampling frequency, and concatenates the readings of all sensors at each sampling moment into a real-time sensing vector according to the node order. S32. Store the real-time sensing vectors into the data buffer in chronological order. The length of the data buffer is greater than or equal to the current length of the adaptive sliding window. S33. Determine whether the number of real-time sensing vectors stored in the data buffer has reached the current length of the adaptive sliding window. If not, continue collecting data. S34. Take a sequence of real-time sensing vectors from the data buffer that is equal in number to the current length of the adaptive sliding window, and use it as training samples for the current window. S35. Input the training samples of the current window into the online sequence extreme learning machine, execute the online weight update algorithm, and update the hidden layer output weight matrix of the online sequence extreme learning machine. S36. Use the updated online sequence extreme learning machine to perform forward computation on the last real-time sensing vector in the current window training samples to obtain a fast estimate of the total weight of the coal pile. S37. Calculate the online sequence extreme learning machine output value corresponding to all real-time sensing vectors in the current window training samples, compare the output value with the reference weight value converted based on the physical relationship of the sensor reading at the corresponding time, obtain a set of difference sequences, perform statistical analysis on the difference sequences, and generate virtual observation residual vectors.

[0024] In this embodiment, S4 specifically includes: S41. The physical information neural network model receives fast estimates and virtual observation residual vectors from the edge computing model, and simultaneously receives raw sensor readings with timestamps corresponding to the fast estimates. S42. Add the virtual observation residual vector and the original sensor readings algebraically according to the corresponding sensor nodes to obtain the sensor readings after correction by the virtual observation residual vector. S43. The fast estimate is converted into an equivalent uniformly distributed load boundary condition, which together with the corrected sensor readings constitutes the input constraint set of the physical information neural network model. S44. Within the coal pile base area, generate a dense two-dimensional coordinate point grid according to a preset resolution, and input all coordinate points into the physical information neural network model. S45. The physical information neural network model uses the input constraint set as the training target, executes the backpropagation optimization algorithm, adjusts the network parameters, and continues until the loss function converges. The loss function includes balance equation constraint terms and sensor data matching constraint terms. S46. Perform forward calculations on the dense two-dimensional coordinate point grid using the converged physical information neural network model, and output the normal stress value corresponding to each coordinate point to form a full-field two-dimensional stress distribution. S47. Extract the normal stress value at the coordinate point corresponding to the physical coordinate position of the smart sensing node from the two-dimensional stress distribution of the whole field. Convert the normal stress value into an equivalent pressure reading through the sensor calibration relationship and use it as the virtual sensing value at the coordinate of the smart sensing node.

[0025] In this embodiment, S5 specifically includes: S51. The physical information neural network model sends the calculated virtual sensing values ​​to the edge computing model through a communication link. S52. The edge computing model extracts the original sensor pressure readings corresponding to the timestamps of the virtual sensing values ​​from the data buffer to form the original pressure reading vector. S53. Match the received virtual sensing values ​​with the original pressure reading vector according to the same smart sensing node index, calculate the difference node by node, and form the original difference vector. S54. Perform low-pass filtering on the original difference vector to remove high-frequency noise components and obtain a smoothed residual feedback signal vector. S55. Extract the mean and covariance matrix of the residual feedback signal vector as adjustment parameters for the learning objective function of the online sequence extreme learning machine; S56. Using the mean of the residual feedback signal vector as the bias correction term and the inverse of the covariance matrix of the residual feedback signal vector as the weight regularization matrix, a new learning objective function for the online sequence extreme learning machine is constructed. S57. In the next execution of the online sequence extreme learning machine weight update, replace the original learning objective function with a new learning objective function that includes a bias correction term and a weight regularization matrix.

[0026] In this embodiment, S6 specifically includes: S61. Obtain the full-field two-dimensional stress distribution data output by the physical information neural network model. This data includes the normal stress value corresponding to each coordinate point in the coal pile base area. S62. Perform area-weighted summation of the normal stress values ​​at all coordinate points in the full-field two-dimensional stress distribution data, and divide the summation result by the gravitational acceleration constant to obtain the total mass value of the coal pile. S63. Define a polygonal boundary within the coal pile base area. The polygonal boundary is determined by the coordinates of multiple vertices. S64. Extract the normal stress values ​​of the coordinate points located inside the polygon boundary from the full-field two-dimensional stress distribution data, perform area-weighted summation on these normal stress values, and divide the summation result by the gravitational acceleration constant to obtain the local region mass value of the region defined by the polygon boundary. S65. Based on the normal stress value and position coordinates of each coordinate point in the full-field two-dimensional stress distribution data, a two-dimensional stress cloud map is generated using a color mapping algorithm. S66. Mark the total mass value of the coal pile, the mass value of the local area, and the sequence of changes in the total mass of the coal pile over time in numerical and graphical form at the preset positions of the two-dimensional stress cloud map. S67. Combine the two-dimensional stress cloud map labeled with quality information with the independent quality change time series curve to generate a real-time status map of the coal pile digital twin.

[0027] In this embodiment, S7 specifically includes: S71. Collect virtual observation residual vectors and store them in chronological order to form a historical residual sequence; S72. Calculate the mean vector and covariance matrix of each virtual observation residual vector in the historical residual sequence to obtain the residual statistical characteristic sequence; S73. Match the residual statistical feature sequence with the preset anomaly pattern library, which includes sensor drift mode, foundation settlement mode and coal pile internal structure anomaly mode. S74. When the offset direction of the mean vector of the residual statistical feature sequence and the trend of the change of the eigenvalue of the covariance matrix meet the judgment conditions of any pattern in the preset abnormal pattern library, it is determined that the preset abnormal pattern has been detected. S75. Generate an early warning signal containing the anomaly type, location, and level based on the matched anomaly pattern type. S76. Based on the abnormal type and level of the warning signal, query the preset parameter adjustment mapping table to obtain the corresponding physical information neural network model equivalent physical parameter adjustment amount; S77. Apply the obtained equivalent physical parameter adjustment amount to the physical information neural network model to complete the parameter adjustment.

[0028] Example 1: To verify the feasibility of this invention in practice, a large bulk cargo yard at a coastal port was selected as the application scenario. This yard stores thermal coal, with irregular trapezoidal coal piles measuring approximately 120 meters × 80 meters at the base and a maximum height of about 15 meters. The designed maximum storage capacity is approximately 100,000 tons. On-site operations are frequent, with tens of thousands of tons of coal being stored and retrieved daily via stacker-reclaimers. Real-time and accurate inventory measurement is required to support production scheduling and trade settlement. Traditional methods face severe challenges in this scenario: 3D laser scanners are greatly affected by weather and have inventory cycles lasting several hours, failing to meet dynamic requirements; belt scales can only measure flow rate during transport and cannot directly obtain static inventory; densely deploying weighing sensors on such a large concrete support platform is extremely costly and difficult to construct, and it is difficult to cope with long-term drift problems caused by uneven foundation settlement.

[0029] This invention specifically addresses the technical challenge of achieving high-precision, continuous dynamic direct weighing under low-cost, sparse sensor deployment conditions. The specific implementation process is as follows: First, six intelligent sensor nodes are deployed under the concrete support layer at the bottom of the coal pile, with a grid spacing of 40 meters × 40 meters. Each node integrates a high-precision piezoresistive pressure sensor, a digital temperature sensor, and a triaxial MEMS tilt sensor, and is connected to an edge computing gateway via a wired network. After recording the precise three-dimensional coordinates of each node, the system collects pressure, temperature, and tilt angle readings from all sensors under two conditions: a completely unloaded coal yard and a known load of 500 tons applied using a standard weight cart. This forms an initial training data set encompassing both unloaded and loaded states.

[0030] Based on this initial dataset, a dual-ring collaborative computing model was constructed. The edge computing model employed an online sequential extreme learning machine with 18 input layer nodes (corresponding to 6 nodes × 3 types of sensors) and 50 hidden layer nodes, using the Sigmoid function as the activation function and an initial adaptive sliding window length of 100 sampling points. The cloud-based physical information neural network model adopted a fully connected neural network structure, containing 8 hidden layers with 128 neurons per layer, using the Tanh activation function, with the weight coefficient of the balance equation constraint term in the loss function set to 1.0 and the weight of the sensor data matching term set to 10.0. The physical information neural network was pre-trained using standard load data, with an initial training iteration of 5000 times, employing the Adam optimizer and an initial learning rate of 0.001.

[0031] The system enters the online operation phase. The edge computing gateway collects all sensor data at a frequency of 1 Hz, stitches them into a sensing vector in real time, and stores it in a buffer. Whenever the buffer data reaches the current sliding window length, the latest 100 sets of data are retrieved and input into the online sequence extreme learning machine for online weight updates, and then a fast estimate of the current total weight of the coal pile is output. Simultaneously, the difference sequence between the model prediction value and the reference value based on the physical conversion of sensor readings within this window is calculated, and statistical analysis is performed to generate a 6-dimensional virtual observation residual vector containing the mean and variance. This fast estimate and residual vector are uploaded to the cloud server in real time.

[0032] After receiving the data, the physical information neural network model begins operation. It first corrects the original pressure sensor readings using the received virtual observation residual vector to compensate for possible system drift. Then, the fast estimate is used as the overall load constraint, forming the input constraint set along with the corrected pressure readings at each node. Within a 120m × 80m rectangular area representing the bottom of the coal pile, 9600 coordinate points are generated at 1-meter intervals and input along with the input constraint set into the physical information neural network. The network aims to minimize the physical equation residuals and match the sensor readings, optimizing through a backpropagation algorithm. The loss function typically converges after 300 to 500 iterations. The converged network performs forward calculations on all 9600 coordinate points, outputting the full-field normal stress distribution. Based on this distribution, interpolation is used to accurately calculate the pressure values ​​that "should be present" at the six actual sensor coordinate points, i.e., the virtual sensing values.

[0033] The cloud sends the calculated virtual sensor values ​​back to the edge gateway. The edge computing model compares these virtual values ​​with the actual raw pressure readings collected at the corresponding time points, obtaining a new set of residuals. After low-pass filtering of these residuals, their mean and covariance information are extracted and used to dynamically adjust the learning objective function of the online sequence extreme learning machine for the next weight update. Specifically, an additional term is added to the original objective function, using the inverse of the covariance matrix as a regularization term and the mean as a bias correction term. This forms a complete closed loop of "fast edge estimation - cloud physical reconstruction - virtual value feedback - edge model adjustment".

[0034] Each time a cloud-based reconstruction is completed, the system performs numerical integration on the generated full-field stress distribution map. Integration is performed over the global region to obtain the total mass; it can also be performed within any polygonal region selected by the user in the visualization interface to obtain the local mass. All mass data, along with the stress cloud map and mass change curves, are integrated to form a real-time refreshed digital twin monitoring screen.

[0035] The system continuously runs and monitors the long-term statistical characteristics of the virtual observation residual vector. During the 30-day test in this embodiment, the system automatically detected a slow unidirectional drift in the residual mean of sensor node 2, with an increase in the covariance eigenvalue, matching the preset "sensor drift mode." The system then triggers a level-two warning and, based on the mapping table, lowers the equivalent elastic modulus parameter of the corresponding region in the physical information neural network model by 5%. After this adaptive adjustment, subsequent reconstruction results show that the overall accuracy of the system has recovered, requiring no manual intervention for calibration.

[0036] Through the above implementation, the technical advantages of this invention are concretely and quantitatively presented. To objectively evaluate performance, five different typical operational moments (covering stockpiling, retrieving, and static states) were selected during the testing period, and simultaneous measurements were performed using the method of this invention, the traditional three-dimensional laser scanning volume method (post-inventory counting), and the manual measurement method based on a total station for comparison. The manual measurement method, due to its complexity, was only used as an emergency reference benchmark. Specific comparison data are shown in the table below: Table 1: Comparison of coal pile weight measurement results using different methods (unit: tons)

[0037] A detailed analysis of the data in the table above reveals the following: First, at static or quasi-static times (T1, T2, T4, T5), the measured values ​​from the method of this invention are systematically slightly higher than those from the three-dimensional laser scanning method. This is primarily due to the difference in the fundamental principles of the two methods: the scanning method converts volume into empirical density, and its accuracy is greatly affected by the density value, as the density of coal fluctuates on-site; while the method of this invention directly senses weight based on mechanical principles, unaffected by empirical density. The error rates of both methods range from 0.84% ​​to 2.28%, which fully meets industrial management requirements (typically requiring an error within 3%), considering large-tonnage weighing scenarios. Second, at dynamic operation times (T3), the three-dimensional laser scanning method, requiring a stable environment and time-consuming scanning, cannot provide effective results; while the method of this invention achieves continuous output, demonstrating its core advantage of "dynamic" weighing. Finally, compared with the estimated values ​​from manual gauging, both the results of this invention and the three-dimensional scanning method show reasonable differences, but the results of this invention and the scanning method exhibit better consistency, confirming the reliability of the method of this invention.

Claims

1. A method for calculating and dynamically weighing the weight of a coal pile based on intelligent sensors, characterized in that, Includes the following steps: S1. Deploy intelligent sensing nodes in a two-dimensional sparse grid under the bearing layer at the bottom of the coal pile, record the physical coordinates of each node, collect sensor data under no-load and known standard load conditions, and obtain an initial training data set. S2. Construct an edge computing model based on an adaptive sliding window online sequence extreme learning machine, and construct a physical information neural network model embedded with multi-physics constraints; S3. The edge computing model collects real-time data streams from intelligent sensing nodes, updates the online sequence extreme learning machine through an adaptive sliding window mechanism, obtains a fast estimate of the total weight of the coal pile, and calculates the virtual observation residual vector. S4. The physical information neural network model receives the fast estimate and the virtual observation residual vector. It uses the sensor readings and fast estimate corrected by the virtual observation residual vector as input constraints, and calculates the full-field two-dimensional stress distribution with the coal pile base coordinate points as input. Based on this distribution, it calculates the virtual sensing value at the coordinates of the smart sensing node in reverse. S5. The physical information neural network model sends the virtual sensing value to the edge computing model. The edge computing model compares the virtual sensing value with the original real-time data stream to obtain the residual feedback signal, and uses the signal to adjust the learning objective function of the online sequence extreme learning machine. S6. Perform numerical integration on the two-dimensional stress distribution across the entire field to obtain accurate total mass values ​​of the coal pile and local mass values, and generate a visual state diagram containing stress cloud maps based on the mass values ​​and stress distribution. S7. Monitor the long-term statistical characteristics of the virtual observation residual vector. When the statistical characteristics meet the preset abnormal mode, trigger the system warning and adjust the preset equivalent physical parameters in the physical information neural network model according to the warning result.

2. The method for calculating and dynamically weighing coal pile weight based on intelligent sensors according to claim 1, characterized in that, S1 specifically includes: S11. Determine the basic boundary of the bottom bearing layer of the coal pile, and plan a two-dimensional rectangular projection grid covering the expected accumulation range of the coal pile based on the basic boundary. S12. On the nodes of the two-dimensional rectangular projection grid, select target deployment points according to the preset sparse layout rules. The sparse layout rules require that the grid spacing between adjacent target deployment points be greater than the preset threshold. S13. Install intelligent sensing nodes at each selected target deployment point. The intelligent sensing nodes integrate pressure sensors, temperature sensors and triaxial tilt sensors. S14. Establish a global coordinate system, measure and record the three-dimensional physical coordinates of each installed smart sensor node in the global coordinate system. The three-dimensional physical coordinates include planar coordinates and elevation values. S15. Control the coal pile to be in a completely unloaded state, collect the pressure, temperature and tilt angle readings of the smart sensor node at this time, and record them as the initial unloaded dataset. S16. Apply a standard load of known total mass to the bearing layer at the bottom of the coal pile. After the load stabilizes, collect the pressure, temperature and tilt angle readings of the smart sensor node and record them as the standard load dataset. S17. Merge and label the initial empty dataset with the standard load dataset to form the initial training data set.

3. The method for calculating and dynamically weighing coal piles based on intelligent sensors according to claim 1, characterized in that, S2 specifically includes: S21. Based on the initial training data set, determine the input dimension and output dimension of the edge computing model. The input dimension is determined by the number of smart sensing nodes and the number of sensor types. The output dimension is one-dimensional, corresponding to the total weight of the coal pile. S22. Construct the network structure of an online sequence extreme learning machine, set the number of nodes in the input layer, the number of nodes in the hidden layer, and the activation function, and randomly generate the connection weights from the input layer to the hidden layer and the bias of the hidden layer. S23. Set the initial length of the adaptive sliding window and define the sliding window length adjustment rules based on the data error within the window; S24. Define the input of the physical information neural network model as the two-dimensional coordinate points of the coal pile base region, and the output as the normal stress value at that coordinate point; S25. Add balance equation constraint terms and sensor data matching constraint terms to the loss function of the physical information neural network model. The balance equation constraint terms are composed of the stress distribution and gravity balance relationship. S26. Using the standard load dataset in the initial training dataset, initialize and train the network parameters of the physical information neural network model.

4. The method for calculating and dynamically weighing coal pile weight based on intelligent sensors according to claim 1, characterized in that, S3 specifically includes: S31. The edge computing model collects real-time pressure, temperature and tilt angle readings from all smart sensing nodes at a fixed sampling frequency, and concatenates the readings of all sensors at each sampling moment into a real-time sensing vector according to the node order. S32. Store the real-time sensing vectors into the data buffer in chronological order. The length of the data buffer is greater than or equal to the current length of the adaptive sliding window. S33. Determine whether the number of real-time sensing vectors stored in the data buffer has reached the current length of the adaptive sliding window. If not, continue collecting data. S34. Take a sequence of real-time sensing vectors from the data buffer that is equal in number to the current length of the adaptive sliding window, and use it as training samples for the current window. S35. Input the training samples of the current window into the online sequence extreme learning machine, execute the online weight update algorithm, and update the hidden layer output weight matrix of the online sequence extreme learning machine. S36. Use the updated online sequence extreme learning machine to perform forward computation on the last real-time sensing vector in the current window training samples to obtain a fast estimate of the total weight of the coal pile. S37. Calculate the online sequence extreme learning machine output value corresponding to all real-time sensing vectors in the current window training samples, compare the output value with the reference weight value converted based on the physical relationship of the sensor reading at the corresponding time, obtain a set of difference sequences, perform statistical analysis on the difference sequences, and generate virtual observation residual vectors.

5. The method for calculating and dynamically weighing coal piles based on intelligent sensors according to claim 1, characterized in that, S4 specifically includes: S41. The physical information neural network model receives fast estimates and virtual observation residual vectors from the edge computing model, and simultaneously receives raw sensor readings with timestamps corresponding to the fast estimates. S42. Add the virtual observation residual vector and the original sensor readings algebraically according to the corresponding sensor nodes to obtain the sensor readings after correction by the virtual observation residual vector. S43. The fast estimate is converted into an equivalent uniformly distributed load boundary condition, which together with the corrected sensor readings constitutes the input constraint set of the physical information neural network model. S44. Within the coal pile base area, generate a dense two-dimensional coordinate point grid according to a preset resolution, and input all coordinate points into the physical information neural network model. S45. The physical information neural network model uses the input constraint set as the training target, executes the backpropagation optimization algorithm, adjusts the network parameters, and continues until the loss function converges. The loss function includes balance equation constraint terms and sensor data matching constraint terms. S46. Perform forward calculations on the dense two-dimensional coordinate point grid using the converged physical information neural network model, and output the normal stress value corresponding to each coordinate point to form a full-field two-dimensional stress distribution. S47. Extract the normal stress value at the coordinate point corresponding to the physical coordinate position of the smart sensing node from the two-dimensional stress distribution of the whole field. Convert the normal stress value into an equivalent pressure reading through the sensor calibration relationship and use it as the virtual sensing value at the coordinate of the smart sensing node.

6. The method for calculating and dynamically weighing coal pile weight based on intelligent sensors according to claim 1, characterized in that, S5 specifically includes: S51. The physical information neural network model sends the calculated virtual sensing values ​​to the edge computing model through a communication link. S52. The edge computing model extracts the original sensor pressure readings corresponding to the timestamps of the virtual sensing values ​​from the data buffer to form the original pressure reading vector. S53. Match the received virtual sensing values ​​with the original pressure reading vector according to the same smart sensing node index, calculate the difference node by node, and form the original difference vector. S54. Perform low-pass filtering on the original difference vector to remove high-frequency noise components and obtain a smoothed residual feedback signal vector. S55. Extract the mean and covariance matrix of the residual feedback signal vector as adjustment parameters for the learning objective function of the online sequence extreme learning machine; S56. Using the mean of the residual feedback signal vector as the bias correction term and the inverse of the covariance matrix of the residual feedback signal vector as the weight regularization matrix, a new learning objective function for the online sequence extreme learning machine is constructed. S57. In the next execution of the online sequence extreme learning machine weight update, replace the original learning objective function with a new learning objective function that includes a bias correction term and a weight regularization matrix.

7. The method for calculating and dynamically weighing coal pile weight based on intelligent sensors according to claim 1, characterized in that, S6 specifically includes: S61. Obtain the full-field two-dimensional stress distribution data output by the physical information neural network model. This data includes the normal stress value corresponding to each coordinate point in the coal pile base area. S62. Perform area-weighted summation of the normal stress values ​​at all coordinate points in the full-field two-dimensional stress distribution data, and divide the summation result by the gravitational acceleration constant to obtain the total mass value of the coal pile. S63. Define a polygonal boundary within the coal pile base area. The polygonal boundary is determined by the coordinates of multiple vertices. S64. Extract the normal stress values ​​of the coordinate points located inside the polygon boundary from the full-field two-dimensional stress distribution data, perform area-weighted summation on these normal stress values, and divide the summation result by the gravitational acceleration constant to obtain the local region mass value of the region defined by the polygon boundary. S65. Based on the normal stress value and position coordinates of each coordinate point in the full-field two-dimensional stress distribution data, a two-dimensional stress cloud map is generated using a color mapping algorithm. S66. Mark the total mass value of the coal pile, the mass value of the local area, and the sequence of changes in the total mass of the coal pile over time in numerical and graphical form at the preset positions of the two-dimensional stress cloud map. S67. Combine the two-dimensional stress cloud map labeled with quality information with the independent quality change time series curve to generate a real-time status map of the coal pile digital twin.

8. The method for calculating and dynamically weighing coal pile weight based on intelligent sensors according to claim 1, characterized in that, Specifically, S7 includes: S71. Collect virtual observation residual vectors and store them in chronological order to form a historical residual sequence; S72. Calculate the mean vector and covariance matrix of each virtual observation residual vector in the historical residual sequence to obtain the residual statistical characteristic sequence; S73. Match the residual statistical feature sequence with the preset anomaly pattern library, which includes sensor drift mode, foundation settlement mode and coal pile internal structure anomaly mode. S74. When the offset direction of the mean vector of the residual statistical feature sequence and the trend of the change of the eigenvalue of the covariance matrix meet the judgment conditions of any pattern in the preset abnormal pattern library, it is determined that the preset abnormal pattern has been detected. S75. Generate an early warning signal containing the anomaly type, location, and level based on the matched anomaly pattern type. S76. Based on the abnormal type and level of the warning signal, query the preset parameter adjustment mapping table to obtain the corresponding physical information neural network model equivalent physical parameter adjustment amount; S77. Apply the obtained equivalent physical parameter adjustment amount to the physical information neural network model to complete the parameter adjustment.

Images