Intelligent flow stabilizing and blockage removing method for raw coal bin
By deploying heterogeneous sensors and pre-trained models in the raw coal bunker, accurate modeling and prediction of blockage type, degree, and spatial location are achieved, solving the problem of insufficient blockage identification and prediction capabilities in existing technologies, and improving operational reliability and material flow stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGXI DATANG INT XINYU NO 2 POWER GENERATION CO LTD
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-19
Smart Images
Figure CN122243178A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power plant flow stabilization and blockage clearing technology, and more specifically, to an intelligent flow stabilization and blockage clearing method for raw coal bunkers. Background Technology
[0002] As the raw coal bunker serves as a crucial hub in the fuel supply system of coal-fired power units, its operation in traditional technology relies on operators making judgments based on personal experience and using local commands to perform intermittent operations on the actuators in order to clear any potential blockages within the bunker. However, practice has shown that traditional technology lacks real-time perception and data collection of the material flow within the bunker. Furthermore, the reactive and intermittent operation mode adopted is not only inefficient but also prone to causing drastic fluctuations in coal feed due to improper timing and intensity of operations.
[0003] To overcome the aforementioned lack of data support, existing technologies utilize basic industrial computers and systems to construct an automated execution sequence triggered by discrete events. Through a simple program based on computer-aided design, it enables the vibration or purging device to be driven in a fixed sequence under the triggering of specific sensor signals. By establishing a simple signal-action mapping relationship, the unblocking process is transformed into a repeatable automated process, improving the consistency of the response.
[0004] However, in actual use, it still has some shortcomings. For example, the data processing unit cannot model and identify the type, degree and formation mechanism of blockage, which means that it can only execute pre-fixed and unadaptive operation instruction sequences. At the same time, it may trigger invalid operations due to false alarms from sensors, which will interfere with the stable material flow. It cannot predict blockages before they occur and generate flexible intervention strategies, which ultimately restricts the reliability of long-term operation. Summary of the Invention
[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent flow stabilization and blockage clearing method for raw coal bunkers, which solves the problems mentioned in the background art through the following scheme.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for intelligent flow stabilization and blockage clearing in raw coal bunkers includes: S1: By deploying multiple heterogeneous sensors in the raw coal bunker, time-series data characterizing the state of materials in the bunker are collected in parallel, and preprocessed to output standardized first risk characteristics. S2: Based on the first risk feature, the second risk feature is calculated using a feature extraction algorithm; S3: Input the second risk feature into the pre-trained congestion identification model and output the third risk feature of the current congestion situation. The third risk feature includes at least the congestion type, congestion degree index and spatial location information. S4: Using the third risk characteristic as the initial state, input it into the flow prediction model, which is used to deduce the changing trend of the congestion index; S5: Based on the aforementioned trend, perform multi-objective optimization calculations. The objective function of the multi-objective optimization calculations includes at least minimizing the disturbance of the coal flow at the outlet and minimizing the energy consumption of equipment operation, so as to generate the optimal instruction sequence for each unblocking actuator. S6: Execute the optimal instruction sequence and update the parameters of the congestion identification model and the flow prediction model.
[0007] Preferably, in S1, the first risk feature is a multi-dimensional real-time data stream that has undergone time alignment and spatial registration, which includes at least: The pressure field matrix time series data, which reflects the spatiotemporal distribution of pressure on the silo wall, is obtained by collecting and preprocessing an array of pressure sensors deployed on the inner wall of the cone section of the raw coal silo. Point cloud time series data reflecting the three-dimensional morphology and collapse rate of the material surface, obtained by a three-dimensional lidar deployed on the top of the raw coal bunker and preprocessed. Acoustic emission spectrum time series data reflecting the internal friction and extrusion state of the material is obtained by acoustic emission sensors deployed on the outer wall of the raw coal bunker cone and preprocessed.
[0008] Preferably, step S2, obtaining the second risk characteristic, specifically includes: Based on the time-series data in the first risk feature, feature sets in the time domain, frequency domain, and spatial domain are calculated in parallel. The time domain feature set includes the statistical moment features and rate of change features of each pressure sensing channel within the sliding time window. The frequency domain feature set includes the spectral energy distribution and entropy features obtained by fast Fourier transform and wavelet packet decomposition. The spatial domain feature set includes the spatial gradient distribution calculated based on the pressure array and the material surface morphology topology features based on three-dimensional point cloud computing. The time-domain feature set, frequency-domain feature set, and spatial-domain feature set are concatenated into a feature vector. Then, a principal component analysis algorithm based on the variance contribution rate and dynamically adjusted weights is used to reduce the dimensionality of the feature vector, generating a low-dimensional state feature vector as the initial second risk feature. The current state feature vector is concatenated with the historical core state feature vectors of consecutive moments in a predetermined time series, and then input into a recurrent neural network. The recurrent neural network captures the temporal dependencies of state evolution, and the output feature vector is the second risk feature.
[0009] Preferably, the congestion identification model in S3 is a fusion architecture, which specifically includes: The gradient boosting decision tree ensemble model, serving as the backbone model, is used to perform nonlinear fitting and classification regression calculations on the second risk feature, outputting a preliminary probability distribution of blockage type and a blockage degree index. A lightweight convolutional neural network, used as an auxiliary model, is employed for spatial feature extraction and pattern recognition of two-dimensional pressure distribution images formed from standardized array-type pressure sensor data. The gradient boosting decision tree ensemble model and the output of the lightweight convolutional neural network are integrated and calculated through a weighted fusion method to generate the final third risk feature.
[0010] Preferably, step S3, obtaining the third risk feature, specifically includes: For the input two-dimensional pressure grayscale image, the lightweight convolutional neural network and its classification results are used to calculate the class activation heatmap H. The intensity value of each pixel in H represents the degree of contribution of the target region to the recognition of the current blockage type by the lightweight convolutional neural network. On the activation heatmap H, find the coordinates of the pixel with the highest intensity and denote them as the focal pixel coordinates. Its coordinates correspond to the center of the region where the pressure anomaly is most concentrated on the two-dimensional pressure grayscale image plane; Acquire point cloud data of the material surface reconstructed by 3D LiDAR scanning at the same time, and determine the position of the center of the silo discharge port in the point cloud coordinate system. ; The image focal pixel coordinates are transformed using a pre-calibrated perspective transformation matrix M. Mapped to the coordinate system of the warehouse where the lidar point cloud is located; The estimated three-dimensional spatial coordinates (X, Y, Z) of the blockage point are calculated, specifically as follows: Where T represents the transpose.
[0011] Preferably, the calibration method for the perspective transformation matrix M is as follows: In an empty warehouse, the known three-dimensional coordinates within the warehouse are... Markers were placed at multiple feature points, and the pixel coordinates of the markers in the pressure distribution image were acquired simultaneously. The matrix M is obtained by solving the system of equations, specifically as follows: in, It is represented as the scale factor, and T represents the transpose.
[0012] Preferably, in S4, the flow prediction model is a sequence-to-sequence neural network model with embedded physical constraints; Using the third risk feature within a previous preset time window and the operation instructions already executed within the corresponding time window as joint inputs, the encoder and decoder structure and the built-in physical loss function constraints are used to extrapolate the changing trend of the congestion index and the corresponding prediction uncertainty range within a future decision cycle in a rolling manner.
[0013] Preferably, in S4, both the encoder and decoder are stacked with two GRU layers, each with a hidden state dimension of 128.
[0014] The technical effects and advantages of this invention are as follows: 1. This invention achieves accurate modeling and identification of blockage type, degree and spatial location by acquiring multi-dimensional time-series data and capturing time-series dependencies in parallel using heterogeneous sensors. This reduces the probability of false alarms from sensors, reduces the interference of invalid operations on material flow, and solves the defects of existing technologies that cannot model and identify blockage characteristics and are prone to false alarms. 2. This invention uses a sequence-to-sequence neural network model with embedded physical constraints to infer the trend of congestion changes, thereby achieving trend prediction before congestion occurs and generating flexible intervention strategies instead of fixed instructions. This improves operational reliability and alleviates the problems of existing technologies being unable to predict and lacking adaptability in instructions. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating the steps of an intelligent flow stabilization and blockage clearing method for raw coal bunkers according to an embodiment of this application. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.
[0018] Hereinafter, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first," "second," and "third" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.
[0019] As attached Figure 1 The method for intelligent flow stabilization and blockage clearing in raw coal bunkers, as shown, collects pre-processed material time-series data through heterogeneous sensors, extracts features, inputs the data into a model to identify the blockage situation, and generates the optimal blockage clearing command by combining flow prediction and multi-objective optimization. The command is then executed and the model is iteratively updated. Specifically, the method includes the following steps: S1: By deploying multiple heterogeneous sensors in the raw coal bunker, time-series data characterizing the state of materials in the bunker are collected in parallel, and preprocessed to output standardized first risk characteristics. S2: Based on the first risk feature, the second risk feature is calculated using a feature extraction algorithm; S3: Input the second risk feature into the pre-trained congestion identification model and output the third risk feature of the current congestion situation. The third risk feature includes at least the congestion type, congestion degree index and spatial location information. S4: Using the third risk characteristic as the initial state, input it into the flow prediction model, which is used to deduce the changing trend of the congestion index; S5: Based on the aforementioned trend, perform multi-objective optimization calculations. The objective function of the multi-objective optimization calculations includes at least minimizing the disturbance of the coal flow at the outlet and minimizing the energy consumption of equipment operation, so as to generate the optimal instruction sequence for each unblocking actuator. S6: Execute the optimal instruction sequence and update the parameters of the congestion identification model and the flow prediction model.
[0020] Specifically, in S1, the first risk characteristic is composed of a multi-dimensional real-time data stream that is time-aligned and spatially registered, including at least: time-series data of a pressure field matrix reflecting the spatiotemporal distribution of pressure on the coal bunker wall, obtained by preprocessing an array of pressure sensors deployed on the inner wall of the raw coal bunker cone section; time-series data of a point cloud reflecting the three-dimensional morphology and collapse rate of the material surface, obtained by preprocessing a three-dimensional lidar deployed on the top of the raw coal bunker; and time-series data of an acoustic emission spectrum reflecting the internal friction and compression state of the material, obtained by preprocessing an acoustic emission sensor deployed on the outer wall of the raw coal bunker cone section. In this process, the first risk characteristic is to ensure that all heterogeneous sensor data are collected under a unified time reference through hardware synchronization signals, and to generate a fused data stream with millisecond-level timestamp consistency and a unified numerical scale after performing standardized processing procedures including low-pass filtering, wavelet denoising, real-time outlier detection and removal, and maximum-minimum normalization. This data stream provides a complete, consistent and high-quality data foundation for subsequent steps to deeply perceive and model the dynamic operating status of materials in the warehouse, thereby directly supporting the system to adaptively and with low latency generate an optimized instruction sequence based on global state assessment.
[0021] In this embodiment, the first risk feature, in each processing cycle, has at least the following output fields: a unified timestamp, a warehouse wall pressure value, three-dimensional world coordinates, and energy values of key frequency bands extracted from the original acoustic emission spectrum.
[0022] It should be noted that, in order to collect the time-series data, three types of complementary sensors are deployed at key locations within the raw coal bunker to form a three-dimensional monitoring network, including: an array-type pressure sensor, a three-dimensional lidar, and an acoustic emission sensor. In this embodiment, the array-type pressure sensor uses an industrial-grade piezoelectric pressure sensor array of model PS-AT201, which is tightly mounted in a matrix on the inner wall lining of the conical section of the raw coal bunker. Its layout coordinates cover the entire conical surface area from the starting point of the conical transition section to 1 meter above the discharge port. Each sensor node continuously measures the normal pressure of the material on the bunker wall at a sampling frequency of 100Hz, generating a time-series signal reflecting the spatial distribution of pressure. The three-dimensional lidar uses a model HOKUYO... The UTM-30LX 2D laser scanner uses a pan-tilt head to rotate the tilt axis and construct a 3D scan. Installed inside a dedicated protective cover at the top inspection door of the raw coal bunker, its scanning center is perpendicular to the material surface. At a scanning frequency of 1Hz, it acquires 3D point cloud data of the material surface for calculating surface morphology, collapse rate, overall material level, and acoustic emission sensors. The acoustic emission sensor is a PAC-WDI-AST broadband acoustic emission sensor, installed at a specific position on the lower outer wall of the raw coal bunker cone. It is used to monitor stress wave signals generated by internal friction, compression, and rupture of the material. It acquires raw waveform data at a high sampling frequency of 500kHz and generates spectral data in real time through Fourier transform for analyzing the internal flow of the material and the microscopic activities that initiate blockages.
[0023] Furthermore, to ensure the temporal consistency of the collected time-series data and the accuracy of subsequent fusion analysis, a hardware-software combined synchronization strategy is adopted: Hardware synchronization signal: A high-precision crystal oscillator clock source is set as the master clock, which generates a 1PPS synchronization signal; Synchronization mechanism: All sensors receive the 1PPS hardware synchronization signal; Simultaneously, when the time-series data is collected, in addition to its own high-precision local timestamp, it is also tagged with the global frame number of the latest received 1PPS signal; When receiving the time-series data, all data streams are first aligned to a unified sampling clock source reference based on the 1PPS global frame number, and then microsecond-level interpolation alignment is performed on each data stream based on its local timestamp, ultimately forming a timestamp alignment of all sensor data within millisecond precision; To ensure that data from different sensors can be fused and analyzed in a unified spatial dimension, a precise mapping relationship between all data and the physical space of the warehouse needs to be established, specifically... The implementation is as follows: Taking the center point of the raw coal bunker's discharge port as the origin, with the vertically upward direction as the positive Z-axis, mutually perpendicular X and Y axes are defined in the horizontal plane to form a fixed bunker coordinate system. The installation position of the pressure sensor array is predetermined by mechanical drawings. The position of each pressure sensor node can be directly calculated and transformed to the world coordinate system based on its row, column, and installation spacing. The transformation relationship is stored in the system in the form of a lookup table. After the three-dimensional lidar is installed, its transformation matrix from the sensor coordinate system to the world coordinate system is determined through a hand-eye calibration process. Specifically, at least three distinct calibration target spheres with known world coordinates are placed within the bunker. The lidar scans to obtain the center coordinates of these target spheres in their sensor coordinate system. The optimal rotation-translation matrix is calculated by solving the spatial point set registration problem. The acoustic emission sensor is installed at a single point, and its three-dimensional coordinates in the world coordinate system are determined by measurement.
[0024] Furthermore, to eliminate noise and outliers and standardize data scales to form a high-quality, standardized data stream, at least the following steps are performed: filtering and denoising. For pressure signals, a fourth-order Butterworth low-pass filter is used with a cutoff frequency set to 10Hz to preserve the gradual pressure variation characteristics generated by material flow and filter out high-frequency noise caused by mechanical vibration. For acoustic emission spectrum data, a wavelet threshold-based denoising model is applied to effectively separate characteristic frequency band energy related to material flow and suppress environmental electromagnetic noise. For lidar point clouds, background removal is first performed to remove fixed structural points on the silo walls, followed by a statistical outlier removal filtering algorithm to remove discrete noise points caused by dust or splashing particles. Outliers... For each sensor data stream, the mean μ and standard deviation σ are calculated in real time within a 60-second sliding time window. Any data point whose value exceeds the range [μ-3σ, μ+3σ] is identified as an instantaneous outlier and removed. The outlier is then filled using linear interpolation of adjacent valid data points. If a sensor continuously marks more than a set number of data points as outliers, a sensor fault alarm is triggered, and a fault-tolerant estimation mode based on other sensor data is initiated. The normalization strategy performs max-min normalization on the preprocessed data streams, with the normalization range uniformly set to [0,1]. In this embodiment, the specific formula for normalization is: ,in, and For pressure data, the theoretical empty pressure and the design maximum static pressure at the sensor installation location are used; for material level data, the empty height and the full height are used; for acoustic emission energy, the long-term statistical background noise level and the energy peak under a typical severe blockage event are used.
[0025] It should be noted that the cutoff frequency of the low-pass filter is defined as 10Hz, which is determined based on long-term observation and analysis of the wall pressure signal spectrum during the operation of the raw coal bunker. Its purpose is to retain the low-frequency pressure change characteristics formed by material flow and accumulation, while effectively filtering out the high-frequency mechanical noise introduced by the vibration of equipment such as belt conveyors and coal feeders. The sliding time window is defined as 60 seconds, which is an empirical value after balancing the rate of change of operating conditions and statistical stability. It aims to provide sufficient data samples for the calculation of the mean and standard deviation, while also reflecting the recent trend of material state changes in a timely manner.
[0026] In this embodiment, compared with the isolated, binary signals that usually rely on only a single or a few level switches or pressure switches in the prior art, the first risk feature generated by S1 has achieved a fundamental improvement in data dimension, quality and integration. Its data can depict a multi-dimensional panorama of material pressure distribution gradient, material surface dynamics, internal friction activity and other conditions in the silo, so that subsequent steps can accurately distinguish the blockage type, quantify the blockage degree, locate the blockage origin, and provide a reliable initial state input for the prediction model.
[0027] Specifically, in S2, based on the time-series data in the first risk feature, feature sets in the time domain, frequency domain, and spatial domain are calculated in parallel; wherein, the time domain feature set includes the statistical moment features and rate of change features of each pressure sensing channel within the sliding time window, the frequency domain feature set includes the spectral energy distribution and entropy features obtained by fast Fourier transform and wavelet packet decomposition, and the spatial domain feature set includes the spatial gradient distribution calculated based on the pressure array and the material surface morphology topology features based on three-dimensional point cloud computing.
[0028] It should be noted that the time-domain feature set includes at least statistical features, rate of change features, and event count features; the frequency-domain feature set includes at least the spectrum, frequency band energy, and wavelet packet energy entropy; and the spatial-domain feature set includes at least pressure field gradient features, material surface morphology features, and asymmetry features.
[0029] In a preferred embodiment, obtaining the time-domain feature set includes: for each pressure sensor channel, taking a sliding time window of 30 seconds, and calculating the statistical characteristics of the signal within the window, including but not limited to mean, standard deviation, skewness, and kurtosis; wherein, the mean reflects the static load, the standard deviation quantifies the fluctuation intensity, and skewness and kurtosis jointly describe the symmetry and peak degree of the pressure distribution, which can effectively distinguish between steady flow, pulsating flow and impact load; the rate of change feature can be obtained by calculating the collapse rate of the material surface height, i.e., the amount of decrease in the average material surface height per unit time, or by calculating the gradient of the change in the global pressure mean, i.e., the amount of change in the mean of the entire pressure array per unit time; the event counting feature may include: for the acoustic emission signal, setting an energy threshold based on 3 times the standard deviation of the background noise level, and counting the number of pulses whose acoustic emission signal envelope exceeds the threshold within a sliding time window of 30 seconds, reflecting the frequency of internal micro-friction and fracturing activities.
[0030] In a preferred embodiment, obtaining the frequency domain feature set includes: analyzing the spectrum: performing Fast Fourier Transform (FFT) on the mean-removed pressure fluctuation signal and acoustic emission signal respectively, with a transform length of 1024 points, extracting the dominant frequency and amplitude of the pressure fluctuation signal in the 1-5Hz frequency band, and extracting the centroid frequency of the acoustic emission signal in the 100-400kHz frequency band; the frequency band energy is obtained by dividing the power spectral density of the acoustic emission signal into two key frequency bands: the low-frequency band LFB (50-150kHz) mainly corresponds to frictional acoustic emission, and the high-frequency band HFB (200-350kHz) mainly corresponds to fracture acoustic emission, and calculating the energy integral of the two frequency bands. and And calculate their ratio. , This is a sensitive indicator for distinguishing blockage types; blockage formation is accompanied by more particle breakage. Significantly increased; while friction intensifies when adhering to the wall. Decrease or remain unchanged; the wavelet packet energy entropy is obtained by performing a 3-level wavelet packet decomposition on the pressure fluctuation signal using the db4 wavelet basis, obtaining the 8 node coefficients of the 3rd level, calculating the energy of each node, and normalizing it to obtain the probability distribution, thereby calculating the wavelet packet energy entropy. An increase in the wavelet packet energy entropy value indicates that the energy is more evenly and chaotically distributed in the frequency band, which often occurs in the early stage of flow instability or blockage.
[0031] In a preferred embodiment, obtaining the spatial feature set includes: meshing the spatial coordinates of the array-type pressure sensors; calculating the lateral and longitudinal pressure gradients of each sensor point within its local 3×3 neighborhood; statistically analyzing the mean and variance of all sensor points, where the mean directly reflects the steepness of the pressure change in the vertical direction, and the variance characterizes the spatial non-uniformity of the pressure gradient; performing moving least squares surface fitting on the three-dimensional point cloud to obtain a smooth surface, and calculating the average curvature and surface roughness of the surface at the sampling points, wherein the surface roughness is defined as the ratio of the point cloud to the surface roughness. The standard deviation of the fitted surface distance is used to determine the material surface during normal flow. It is approximately a flat or regular concave surface with low curvature and roughness. When "well flow" or "segregation" occurs, the curvature distribution is abnormal and the roughness increases. Taking the central axis of the silo as the reference, the pressure measurement area of the cone wall is divided into four quadrants: 0°, 90°, 180°, and 270°. The average value of the pressure readings of all sensors in each quadrant is calculated. The standard deviation of the average value of the four quadrants is further calculated and defined as the pressure asymmetry index to effectively detect unilateral wall adhesion or segregation caused by uneven material composition or silo structure problems.
[0032] Furthermore, the time-domain feature set, frequency-domain feature set, and spatial-domain feature set are concatenated into a feature vector in a fixed order, and a principal component analysis algorithm based on the variance contribution rate to dynamically adjust the weights is used to reduce the dimensionality of the feature vector, generating a low-dimensional state feature vector as the preliminary second risk feature.
[0033] It should be noted that, during the initialization phase, the principal component analysis algorithm collects data from the target raw coal bunker for no less than 7 days, covering a training set of features from various typical working conditions, including empty bunkers, normal full-load operation, and artificially simulated varying degrees of arching and wall adhesion. The standard principal component analysis algorithm is then applied to the training set to calculate the variance contribution rate of each principal component. A cumulative variance contribution rate threshold of 95% is set, and the first k=15 principal components whose cumulative contribution rate exceeds the threshold for the first time are selected to ensure that most of the effective information is retained after dimensionality reduction. During online operation, the real-time feature vector is projected onto the k principal component directions to obtain the state feature vector. The dynamic adjustment of weights based on the variance contribution rate is reflected in the following: the projection essentially uses the variance contribution rate of each principal component direction as weight to linearly reconstruct the original features; the principal component direction with a high variance contribution rate has a dominant weight when reconstructing the original data information.
[0034] Furthermore, the current state feature vector is concatenated with the historical core state feature vectors of consecutive moments within a predetermined time series, and then input into a recurrent neural network. The recurrent neural network captures the temporal dependencies of state evolution, and the output feature vector is the second risk feature.
[0035] It should be noted that the recurrent neural network adopts a single-layer gated recurrent unit network with 32 hidden layer neurons. The state feature vectors of the current time and the past N-1 time steps are arranged in chronological order to form an input sequence of length N. The sequence length N is an adjustable parameter, usually set to correspond to a time length of 1 to 3 minutes, that is, N is set to 90 to 180. The hidden state of the last time step of the network contains the memory information of the state evolution within the entire historical window from t-N+1 to t, and it is output as the second risk feature.
[0036] In this embodiment, the second risk feature can simultaneously express a series of related phenomena such as "the pressure gradient is slowly increasing", "the proportion of high-frequency acoustic emission energy is rising", and "local depressions appear on the material surface". These are the typical comprehensive features of "blockage" formation, enabling the system to obtain the data basis for fine-scale modeling and identification of blockage type, degree and formation mechanism for the first time.
[0037] Specifically, in S3, the congestion identification model is a fusion architecture, which includes: a gradient boosting decision tree ensemble model as the backbone model, used to perform nonlinear fitting and classification regression calculation on the second risk feature, and output a preliminary congestion type probability distribution and congestion degree index; a lightweight convolutional neural network as an auxiliary model, used to extract spatial features and recognize patterns from the two-dimensional pressure distribution image formed by standardized array pressure sensor data; the outputs of the gradient boosting decision tree ensemble model and the lightweight convolutional neural network are integrated and calculated through a weighted fusion method to generate the final third risk feature.
[0038] In this embodiment, the two-dimensional grayscale image is directly generated by mapping data from an array-type pressure sensor. The number of rows and columns in the image correspond to the layout rows and columns of the sensor array, respectively. The pixel grayscale value at coordinates (i,j) in the image... , This represents the pressure reading of the sensor at position (i,j) at time t. and These represent the lower and upper limits of the sensor's measurement range.
[0039] It should be noted that the training of the blockage identification model is divided into two stages: The first stage involves establishing a three-dimensional geometric model of the raw coal bunker. Using the discrete element method (DEM), a large-scale simulation dataset is generated, simulating various working conditions such as normal flow, arching, wall adhesion, and tubular flow under different coal types, humidity levels, and bunker locations. This generates a dataset with corresponding real blockage labels. The gradient boosting decision tree ensemble model and the lightweight convolutional neural network are then independently pre-trained. The gradient boosting decision tree ensemble model uses multi-class logarithmic loss and mean squared error as loss functions for classification and regression tasks, respectively. The lightweight convolutional neural network uses cross-entropy loss as the classification loss function. The second stage involves determining the backbone parameters of the two models and determining their output data, which includes at least... , , , The data is stitched together using small-scale field data containing real labels, and the fusion weights are adjusted accordingly. Fine-tuning the network The parameters and the final decision-making layer are fine-tuned end-to-end, with scaling factors... The value used to adjust the correction strength of spatial features to the final CDI value is usually set to a small positive number. In this embodiment, it is defined as between 0.05 and 0.15. It is determined by observing the change in CDI prediction error after fine-tuning on a small-scale validation set. The aim is to introduce spatial information to improve accuracy while avoiding over-correction that leads to output instability.
[0040] In this embodiment, the gradient boosting decision tree ensemble model is implemented using the LightGBM framework, with the following key training hyperparameters defined: 500 trees, a maximum tree depth of 8, a learning rate of 0.05, and a minimum data size of 20 for leaf nodes to prevent overfitting. The lightweight convolutional neural network contains three convolutional blocks, each consisting of a convolutional layer, a batch normalization layer, and a ReLU activation function, with a kernel size of 3x3, a stride of 1, and 32, 64, and 128 channels, respectively. The convolutional layers are followed by a global average pooling layer and two fully connected layers, ultimately outputting the classification probability.
[0041] In one possible implementation, generating the third risk feature includes: the gradient boosting decision tree ensemble model taking the dimensionality-reduced second risk feature output by S2 as input; simultaneously, the lightweight convolutional neural network taking a two-dimensional grayscale image reconstructed from normalized array pressure sensor data as input; and simultaneously performing parallel forward computation, the gradient boosting decision tree ensemble model outputs two intermediate results: a multidimensional probability vector regarding the congestion type. The lightweight convolutional neural network outputs a congestion type probability vector based on image classification, along with a scalar value as an initial congestion level index. The final determination of the congestion type is made by and The final probability vector is obtained by weighted averaging. The type corresponding to the highest probability value is taken as the final recognition result.
[0042] It should be noted that the class activation mapping technique in the lightweight convolutional neural network is used to perform visualization analysis on the two-dimensional grayscale image, generating a heat map that highlights the image area that contributes the most to the network decision, namely the pressure abnormal area. The focal pixel coordinates of the heat map, combined with the material surface geometry collected in real time by the three-dimensional lidar, are mapped onto the real three-dimensional spatial coordinates of the raw coal bunker through a pre-calibrated perspective transformation matrix, thereby determining the spatial location information of the blockage.
[0043] Furthermore, the final determination of the congestion type is a decision-making process based on probability fusion, specifically involving the following steps: The type probability vector output by the gradient boosting decision tree ensemble model is... The type probability vector output by the lightweight convolutional neural network Perform a weighted average to calculate the final comprehensive probability vector. Specifically, it is expressed as: in, Represented as the preset fusion weight coefficient, 0≤ ≤1, During the model training phase, a grid search on the validation set is used to determine the type recognition accuracy, with the goal of maximizing the overall accuracy of type recognition; in one specific implementation, a setting is made... =0.7 indicates that the gradient boosting decision tree ensemble model based on the second risk feature is given a higher confidence level in its discrimination results; traversing the comprehensive probability vector Select the element with the highest probability value. The type represented by the index corresponding to its maximum value is determined as the current final blockage type; in a specific implementation, index 0 corresponds to "normal", 1 corresponds to "arched blockage", and 2 corresponds to "wall-attached blockage".
[0044] Furthermore, the congestion index is a continuous quantification value, which is obtained through a process of preliminary estimation and fine-tuning based on spatial characteristics: the scalar value output by the regression head of the gradient boosting decision tree ensemble model is directly adopted. As a preliminary estimate of the degree of congestion, It is calculated based on the second risk feature, and its value range is usually normalized to [0,1], where 0 represents completely unobstructed and 1 represents completely blocked; the global average pooling output of the last convolutional layer of the lightweight convolutional neural network is extracted as a vector representing the global spatial anomaly pattern of the pressure distribution. Fine-tuning the network using a pre-trained fully connected layer right This process generates a fine-tuning factor. The fine-tuning factor Used to correct the preliminary index based on the spatial heterogeneity of the pressure field, resulting in the final blockage index. Calculate using the following formula: in, Represented as a small scaling factor, it controls the magnitude of fine-tuning of spatial features. The min and max functions ensure... The value is restricted to the range [0,1].
[0045] It should be noted that the fine-tuning network... It is a simple two-layer fully connected network whose input is a vector extracted by a lightweight convolutional neural network. The hidden layer has a dimension of 64, uses the ReLU activation function, and has a single output neuron that outputs a fine-tuning factor. In the second training phase, the network, together with the fusion layer, is trained using live data with mean squared error as the loss function.
[0046] Furthermore, the determination of the spatial location information relies on visual interpretability techniques and geometric mapping: For the input two-dimensional pressure grayscale image, the lightweight convolutional neural network and its classification results are used to calculate a class activation heatmap H through gradient-weighted class activation mapping technology. The intensity value of each pixel in H represents the degree of contribution of the target region to the lightweight convolutional neural network in identifying the current blockage type. High-intensity regions correspond to areas with significant pressure anomalies. On the class activation heatmap H, the coordinates of the pixel with the highest intensity are found and recorded as the focal pixel coordinates. Its coordinates correspond to the center of the region where the pressure anomaly is most concentrated on the two-dimensional pressure grayscale image plane; acquire the material surface point cloud data reconstructed by three-dimensional lidar scanning at the same time, and determine the position of the center of the silo discharge port in the point cloud coordinate system. The image focal pixel coordinates are transformed using a pre-calibrated perspective transformation matrix M. Mapped to the coordinate system of the warehouse where the lidar point cloud is located; the calibration method of the perspective transformation matrix M is as follows: in the empty warehouse state, with known three-dimensional coordinates within the warehouse. Markers were placed at multiple feature points, and the pixel coordinates of these markers in the pressure distribution image were acquired simultaneously. The matrix M is obtained by solving the system of equations, specifically as follows: in, The scale factor is represented by T, which represents the transpose; the estimated three-dimensional spatial coordinates (X, Y, Z) of the blockage point are calculated, specifically as follows: In this embodiment, based on actual physical constraints, namely that the point should be inside the warehouse wall, the mapped point is projected onto the inner surface of the warehouse wall to finally obtain the spatial location information.
[0047] In this embodiment, a breakthrough improvement was achieved in the comparative test of historical operating data and simulation data containing multiple complex working conditions: First, the recognition accuracy and quantification capability have achieved a leap forward. The average recognition accuracy of key blockage types such as "arching" and "wall adhesion" has been increased to 94.5%, and a continuous blockage degree index can be output simultaneously. The correlation coefficient between the index and the expert evaluation results is above 0.92, while the spatial positioning error is less than 0.3 meters. This has enabled comprehensive and accurate modeling of the blockage formation mechanism, development degree, and specific location. Second, relying on multi-sensor information fusion, the false triggering rate caused by single-point signal anomalies has been successfully suppressed to below 4%, eliminating most of the interference of invalid operations on the coal flow from the source.
[0048] Specifically, in S4, the flow prediction model is a sequence-to-sequence neural network model with embedded physical constraints. The third risk feature within a previous preset time window and the sequence of operation instructions executed within the corresponding time window are used as joint inputs. Through its encoder-decoder structure and the built-in physical loss function constraints, the model infers the changing trend of the congestion index and the corresponding prediction uncertainty range within a future decision cycle in an online and rolling manner.
[0049] It should be noted that both the encoder and decoder of the sequence-to-sequence model are constructed using stacked gated recurrent units. In this embodiment, both the encoder and decoder are stacked with two layers of GRU, each with a hidden state dimension of 128. The encoder progressively encodes a 30-length input sequence into a context vector. The decoder then uses this context vector as the initial state to autoregressively generate the CDI prediction sequence for the next 50 time steps. Simultaneously, a fixed time interval of 60 seconds is used as a decision cycle. At the beginning of each cycle, the latest sensor data is collected, and the process from S1 to S5 is executed in a rolling manner, thereby achieving continuous and online updates to the material flow state and unblocking strategy.
[0050] In this embodiment, the training data for the flow prediction model is generated by high-fidelity discrete element method simulation of the raw coal bunker. The simulation simulates 10,000 different working conditions, including different coal types, moisture content, initial material levels, and blockage patterns. For each working condition, a time series data segment with a length of 300 time steps is generated, with a sampling interval of 10 seconds. Each data sample includes: the input sequence of the third risk feature and operation instruction encoding for the first 30 time steps, and the target sequence of the true CDI values for the subsequent 50 time steps. The Adam optimizer is used for training, with an initial learning rate of 0.001 and a batch size of 32. The training is carried out for a total of 500 rounds, and an early stopping strategy is adopted to prevent overfitting.
[0051] It should be noted that, to overcome the shortcomings of purely data-driven models in generalizing under unknown conditions and purely mechanistic models in computational complexity, a regularization term derived from physical equations is added to the standard data fitting loss during model training. This includes at least: mass conservation constraints. in, Expressed as the equivalent density of the material. Represented as the macroscopic flow velocity field of the material, it penalizes instances of mass non-conservation in the predicted results, ensuring that the predicted trend is physically reasonable; momentum change constraint: in, Expressed as the matter derivative, Represented as stress tensor, Represented as gravitational acceleration, this correlates pressure, gravity, and material acceleration, enabling the model to learn evolutionary patterns consistent with mechanical principles; the total loss function of the physical information neural network is: in, This is expressed as the mean square error or other data fitting loss between the predicted CDI sequence and the true value. These are represented as hyperparameters, used to balance the accuracy of data fitting with the degree of adherence to physical laws.
[0052] It should be noted that the equivalent density A macroscopic average quantity calculated using empirical values of material level height, silo geometric volume, and coal powder bulk density obtained from 3D lidar; the macroscopic flow velocity field Its vertical component can be directly estimated from the material surface collapse rate obtained by continuous scanning of a three-dimensional lidar, while its horizontal component or shear information can be indirectly inferred by analyzing the pressure wave transmission mode or spatial correlation measured by an array of pressure sensors; the stress tensor The normal component is directly related to the pressure distribution of the silo wall measured directly by the array pressure sensor, and the shear stress component can be estimated based on the silo wall friction model and the normal pressure. At the same time, the parameters involved in the regularization term derived from the physical equation have all been normalized and transformed into dimensionless quantities.
[0053] In this embodiment, the hyperparameters Optimized by grid search, set =0.1, =0.05. Furthermore, based on the power plant's requirements for stability and response speed, the flow prediction model, during deployment and operation, encodes the third risk feature output by S3 and the historical sequence of operation commands from the most recent period into a high-dimensional vector. The input is used as the initial input to the flow prediction model. For each historical time t, the numerical variables in the third risk feature are standardized, and the congestion type variables are one-hot encoded. Simultaneously, the operation command executed at that time is converted into a multi-dimensional binary vector, where each bit represents the activation state of a specific actuator. All the encoded values are concatenated to form the input vector for that time. In this embodiment, 30 consecutive input vectors constitute the encoder's input sequence. For the same input... Perform TN forward propagations, randomly discarding some neurons each time. The final prediction output is the average of the multiple predictions. The uncertainty is calculated as follows: in, This is expressed as the average of multiple predictions. This is represented as the forward computation of the network. Represented as the network weights in the t-th sample; prediction uncertainty The magnitude of the risk level is directly mapped to the risk level; in this embodiment, a threshold is set. :when The time is low risk. The risk level is currently medium. It is a high-risk situation.
[0054] In this embodiment, the number of samples TN is defined as 50 to achieve a balance between uncertainty estimation accuracy and computational overhead; the threshold The uncertainty of all predicted samples on the validation set is respectively. The 30th and 70th percentiles of the distribution.
[0055] It should be noted that the flow prediction model is based on the neural network model. Through its forward propagation calculation, it can quickly deduce the continuous change trend curve of the congestion degree index in the future prediction time domain with a delay of seconds, and simultaneously output a comprehensive risk level.
[0056] In this embodiment, a raw coal bunker underwent a three-month comparative operation test. Key comparisons are as follows: Regarding the modeling and identification capabilities for blockages, existing technologies can only report a binary state of "empty" or "blocked," failing to distinguish between types, resulting in a false alarm rate of 18%. This method, through S2-S3, reduces the blockage false alarm rate to below 3% and can accurately distinguish between arching and wall adhesion, with an identification accuracy exceeding 95%. Regarding the adaptability of operating strategies, existing technologies trigger approximately 420 clearing actions per month, of which, after post-analysis, actions deemed "unnecessary" or "excessive" account for as high as 35%, significantly interfering with coal feeding stability. This method, through S4-S5, optimizes the total number of actions per month to approximately 150, with over 95% being preventative flexible actions based on risk prediction. The proportion of unnecessary actions is reduced to less than 5%, and the standard deviation of outlet coal flow fluctuation is reduced by 60%.
[0057] Specifically, in S5, the generation of congestion clearing strategies is abstracted as a problem of seeking optimal solutions for multiple objectives under multiple constraints. Through online computation, the decision variables are defined as vectors A=[a1,a2,...,a...]. n ], where each element a i This represents the sequence of actions of a traffic clearing execution agency within a future decision-making cycle. The sequence is further decomposed into a series of binary action identifiers, action triggering time points, and action intensity parameters. Its dimensions are determined by the number of actuators and the discretization granularity of the decision-making cycle.
[0058] Furthermore, we construct an objective function F(A) containing at least three optimization objectives: minimizing the outlet coal flow disturbance. ,in, This is represented by the predicted change curve of the outlet coal feeder signal, simulated based on the prediction model after applying strategy A. This is expressed as the desired stable flow rate reference value, used to minimize the cumulative negative impact of all actions on coal output stability; and to minimize equipment operating energy consumption. .in, Let be the intensity of the ai-th action. This is expressed as the energy conversion coefficient of the actuator, used to encourage the use of the fewest and gentlest actions to solve problems; maximizing the effectiveness of preventative unblocking. .in, It represents the decrease in the congestion index at the end of the prediction period after applying strategy A, compared to the predicted decrease under the no-intervention scenario; it is aggregated into a single scalar objective function using a weighted summation method: The weighting coefficients [w1, w2, w3] can be configured according to running preferences.
[0059] In this embodiment, the weighting coefficients [w1,w2,w3] are set to [0.3,0.2,0.5] when the predicted risk level is consistently higher than the high threshold, prioritizing clearing blockages; if the flow deviation increases significantly, they are adjusted to [0.6,0.2,0.2], prioritizing flow stabilization; under stable operating conditions, they are set to [0.25,0.6,0.15], emphasizing energy saving; the high threshold is defined as a risk level of 3, out of a total of 4 levels; consistently higher than the high threshold means a risk level ≥ 3 for 3 consecutive prediction periods; a significantly increased flow deviation means that the absolute deviation between the current outlet flow and the expected stable flow reference value exceeds 15% of the expected stable flow reference value, and the duration exceeds 10 seconds.
[0060] It should be noted that the optimization calculations are performed under strict constraints: Risk constraint: Strategy A must ensure that at any time during the execution cycle, the risk level derived by the prediction model does not exceed the safety threshold; Equipment physical constraint: The action frequency, interval time, and maximum action intensity of each actuator are limited by its physical characteristics; Coupling constraint: The mutual influence between actuator actions is considered; In this embodiment, this includes, but is not limited to, for any two air cannons whose three-dimensional Euclidean distance between their installation positions is less than 2 meters, the interval between their action triggering times must be greater than 5 seconds; For the same vibrator, the interval between two consecutive actions must not be less than the minimum cooling time specified by its manufacturer.
[0061] In this embodiment, an improved genetic algorithm is used for online solution. The algorithm parameters are set as follows: population size 100, maximum number of iterations 50, crossover probability 0.8, and mutation probability 0.1. The solution process is as follows: 100 sets of unblocking instruction sequences that meet the basic physical constraints are randomly generated to form the initial population. For each individual in the population, i.e., a strategy A, it is input along with the current material state into the flow prediction model constructed by S4. The model acts as a fast digital simulator to deduce the future working conditions after executing strategy A and outputs the predicted outlet coal flow signal, the exponential change in the degree of blockage, and the real-time risk level. Each individual uses a mixture of real and integer values. For each actuator aᵢ, the encoding consists of three parts: a real number between [0,1], representing the threshold for whether to take action within the decision cycle; an integer, representing the discrete time step number of the action trigger time relative to the start of the cycle; and a real number, representing the action intensity parameter, whose range is normalized to the [0,1] interval, and the actual intensity is linearly mapped to the allowable range of the actuator. Subsequently, the fitness value of the individual is calculated according to the objective function F(A). Based on the fitness, selection, crossover, and mutation are performed to generate a new generation of population. This process is repeated iteratively until the maximum number of iterations is reached. The individual with the best fitness is decoded to obtain the optimal instruction sequence.
[0062] Specifically, in S6, the optimal instruction sequence generated in S5 is securely sent to the underlying actuators via a standard industrial communication protocol. A parallel feedback monitoring thread is immediately initiated, and after instruction execution begins, the multi-source sensor network in S1 is reactivated to continuously collect all material status data covering the instruction execution process and for 5 minutes after its completion. All generated data is packaged into a digital case according to a predefined structured format and stored in the case library. Each case includes at least: the first risk characteristic that triggered this decision, the second risk characteristic, the third risk characteristic, the predicted CDI trend curve, the risk level, the optimal instruction sequence, and operating condition labels such as timestamp, coal type, and initial material level in the storage bin.
[0063] Furthermore, after the instructions are executed, the core model is periodically and automatically iteratively optimized: the triggering mechanism is triggered by two conditions, and it will be executed automatically if either condition is met: case accumulation trigger: when the number of newly accumulated valid cases in the case library reaches a preset threshold of 100; performance degradation trigger: when the average prediction deviation of the most recent M=20 decisions is continuously exceeded by 150% of the historical baseline value.
[0064] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for intelligent flow stabilization and blockage clearing in raw coal bunkers, characterized in that, include: S1: By deploying multiple heterogeneous sensors in the raw coal bunker, time-series data characterizing the state of materials in the bunker are collected in parallel, and preprocessed to output standardized first risk characteristics. S2: Based on the first risk feature, the second risk feature is calculated using a feature extraction algorithm; S3: Input the second risk feature into the pre-trained congestion identification model and output the third risk feature of the current congestion situation. The third risk feature includes at least the congestion type, congestion degree index and spatial location information. S4: Using the third risk characteristic as the initial state, input it into the flow prediction model, which is used to deduce the changing trend of the congestion index; S5: Based on the aforementioned trend, perform multi-objective optimization calculations. The objective function of the multi-objective optimization calculations includes at least minimizing the disturbance of the coal flow at the outlet and minimizing the energy consumption of equipment operation, so as to generate the optimal instruction sequence for each unblocking actuator. S6: Execute the optimal instruction sequence and update the parameters of the congestion identification model and the flow prediction model.
2. The intelligent flow stabilization and blockage clearing method for raw coal bunkers according to claim 1, characterized in that: The first risk characteristic of S1 is a multi-dimensional real-time data stream that has undergone time alignment and spatial registration, which includes at least: The pressure field matrix time series data, which reflects the spatiotemporal distribution of pressure on the silo wall, is obtained by collecting and preprocessing an array of pressure sensors deployed on the inner wall of the cone section of the raw coal silo. Point cloud time series data reflecting the three-dimensional morphology and collapse rate of the material surface, obtained by a three-dimensional lidar deployed on the top of the raw coal bunker and preprocessed. Acoustic emission spectrum time series data reflecting the internal friction and extrusion state of the material is obtained by acoustic emission sensors deployed on the outer wall of the raw coal bunker cone and preprocessed.
3. The intelligent flow stabilization and blockage clearing method for raw coal bunkers according to claim 1, characterized in that: S2, obtaining the second risk characteristic, specifically includes: Based on the time-series data in the first risk feature, feature sets in the time domain, frequency domain, and spatial domain are calculated in parallel. The time domain feature set includes the statistical moment features and rate of change features of each pressure sensing channel within the sliding time window. The frequency domain feature set includes the spectral energy distribution and entropy features obtained by fast Fourier transform and wavelet packet decomposition. The spatial domain feature set includes the spatial gradient distribution calculated based on the pressure array and the material surface morphology topology features based on three-dimensional point cloud computing. The time-domain feature set, frequency-domain feature set, and spatial-domain feature set are concatenated into a feature vector. Then, a principal component analysis algorithm based on the variance contribution rate and dynamically adjusted weights is used to reduce the dimensionality of the feature vector, generating a low-dimensional state feature vector as the initial second risk feature. The current state feature vector is concatenated with the historical core state feature vectors of consecutive moments in a predetermined time series, and then input into a recurrent neural network. The recurrent neural network captures the temporal dependencies of state evolution, and the output feature vector is the second risk feature.
4. The intelligent flow stabilization and blockage clearing method for raw coal bunkers according to claim 1, characterized in that: The S3, the congestion identification model, is a fusion architecture, which specifically includes: The gradient boosting decision tree ensemble model, serving as the backbone model, is used to perform nonlinear fitting and classification regression calculations on the second risk feature, outputting a preliminary probability distribution of blockage type and a blockage degree index. A lightweight convolutional neural network, used as an auxiliary model, is employed for spatial feature extraction and pattern recognition of two-dimensional pressure distribution images formed from standardized array pressure sensor data. The gradient boosting decision tree ensemble model and the output of the lightweight convolutional neural network are integrated and calculated through a weighted fusion method to generate the final third risk feature.
5. The intelligent flow stabilization and blockage clearing method for raw coal bunkers according to claim 4, characterized in that: S3, obtaining the third risk characteristic, specifically includes: For the input two-dimensional pressure grayscale image, the lightweight convolutional neural network and its classification results are used to calculate the class activation heatmap H. The intensity value of each pixel in H represents the degree of contribution of the target region to the recognition of the current blockage type by the lightweight convolutional neural network. On the activation heatmap H, find the coordinates of the pixel with the highest intensity and denote them as the focal pixel coordinates. Its coordinates correspond to the center of the region where the pressure anomaly is most concentrated on the two-dimensional pressure grayscale image plane; Acquire point cloud data of the material surface reconstructed by 3D LiDAR scanning at the same time, and determine the position of the center of the silo discharge port in the point cloud coordinate system. ; The image focal pixel coordinates are transformed using a pre-calibrated perspective transformation matrix M. Mapped to the coordinate system of the warehouse where the lidar point cloud is located; The estimated three-dimensional spatial coordinates (X, Y, Z) of the blockage point are calculated, specifically as follows: Where T represents the transpose.
6. The intelligent flow stabilization and blockage clearing method for raw coal bunkers according to claim 5, characterized in that: The calibration method for the perspective transformation matrix M is as follows: In an empty warehouse, the known three-dimensional coordinates within the warehouse are... Markers were placed at multiple feature points, and the pixel coordinates of the markers in the pressure distribution image were acquired simultaneously. The matrix M is obtained by solving the system of equations, specifically as follows: in, It is represented as the scale factor, and T represents the transpose.
7. The intelligent flow stabilization and blockage clearing method for raw coal bunkers according to claim 1, characterized in that: The flow prediction model in S4 is a sequence-to-sequence neural network model with embedded physical constraints. Using the third risk feature within a previous preset time window and the operation instructions already executed within the corresponding time window as joint inputs, the encoder and decoder structure and the built-in physical loss function constraints are used to extrapolate the changing trend of the congestion index and the corresponding prediction uncertainty range within a future decision cycle in a rolling manner.
8. The intelligent flow stabilization and blockage clearing method for raw coal bunkers according to claim 7, characterized in that: In S4, both the encoder and decoder are stacked with two GRU layers, each with a hidden state dimension of 128.