Millimeter wave radar material level monitoring method, device and equipment and storage medium

Abstract

The invention relates to the technical field of millimeter-wave radars, and discloses a millimeter-wave radar material level monitoring method, device and equipment and a storage medium. The method comprises the steps that a plurality of millimeter-wave radar sensors are installed at the top of a coal feeder bin, a reflection point cloud data set is collected, Euclidean clustering is carried out, and an effective point cloud clustering set is obtained; performing multi-stage information screening to obtain an effective material level cluster set of each millimeter wave radar sensor; carrying out signal detection alternating-to-parallel ratio and reliability analysis to generate fused material level data; inputting the fused material level data into a multi-model interaction and unscented Kalman filter model for material level dynamic tracking to obtain a time sequence material level curve and a prediction confidence interval; and iterative second-order cone programming calculation is executed based on the time sequence material level curve and the prediction confidence interval, and a coal feeder control parameter sequence is obtained. According to the method, the problem of radar point cloud disorder caused by a complex environment in the coal feeder is effectively solved, and efficient control parameter optimization under different load conditions is realized.

Description

Technical Field

[0001] The present invention relates to the field of millimeter wave radar technology, and in particular to a millimeter wave radar material level monitoring method, device, equipment and storage medium. Background Art

[0002] Traditional material level monitoring relies primarily on technologies such as weight-type, ultrasonic, or single millimeter-wave radar sensors. However, these methods often face numerous challenges in the complex industrial environment of coal feeders. In particular, single millimeter-wave radar technology, while offering advantages such as non-contact and dust resistance, still suffers from limited detection range, low accuracy, and poor stability in practical applications.

[0003] The limitations of single millimeter-wave radar in coal feeder material level monitoring are mainly reflected in the following aspects: First, due to the complex spatial structure of the coal feeder silo, single-point monitoring cannot fully cover the entire material level space, resulting in blind spots in monitoring; second, the complex internal environment of the silo makes the radar reflection signal chaotic, and various media such as coal powder, metal parts, and moving structures are mixed, causing serious signal interference and reducing the accuracy of material level identification; third, the working state of the coal feeder changes under different load conditions, and factors such as coal powder adhesion and vibration lead to poor stability of monitoring results; fourth, existing algorithms have difficulty in processing multi-source heterogeneous data and model uncertainty, and cannot adapt to dynamically changing working environments. Summary of the Invention

[0004] The main purpose of the present invention is to provide a millimeter-wave radar level monitoring method, device, equipment and storage medium. The present invention effectively deals with the problem of cluttered radar point clouds caused by the complex environment inside the coal feeder, and realizes efficient control parameter optimization under different load conditions.

[0005] To achieve the above object, the present invention provides a millimeter wave radar material level monitoring method, comprising the following steps: Multiple millimeter-wave radar sensors are installed on the top of the coal feeder silo to collect a reflection point cloud dataset, and Euclidean clustering is performed on the reflection point cloud dataset to obtain a valid point cloud cluster set; Performing multi-level information screening on the effective point cloud cluster set to obtain an effective material level cluster set of each millimeter wave radar sensor; Performing signal detection intersection-over-union ratio and reliability analysis on the effective material level cluster sets of each millimeter-wave radar sensor to generate fused material level data; Inputting the fused material level data into the multi-model interaction and unscented Kalman filter model to perform dynamic material level tracking, and obtaining a time series material level curve and a prediction confidence interval; An iterative second-order cone programming calculation is performed based on the time-series material level curve and the prediction confidence interval to obtain a coal feeder control parameter sequence.

[0006] The present invention also provides a millimeter wave radar material level monitoring device, comprising: an acquisition unit, configured to install multiple millimeter-wave radar sensors on the top of the coal feeder silo, collect a reflection point cloud dataset, and perform Euclidean clustering on the reflection point cloud dataset to obtain a valid point cloud cluster set; A screening unit, configured to perform multi-level information screening on the effective point cloud cluster set to obtain an effective material level cluster set of each millimeter wave radar sensor; An analysis unit, configured to perform signal detection intersection-over-union ratio and reliability analysis on the effective material level cluster sets of each millimeter-wave radar sensor to generate fused material level data; A dynamic tracking unit is used to input the fused material level data into a multi-model interaction and unscented Kalman filter model to perform dynamic material level tracking, thereby obtaining a time series material level curve and a prediction confidence interval; The calculation unit is used to perform iterative second-order cone programming calculation based on the time-series material level curve and the prediction confidence interval to obtain a coal feeder control parameter sequence.

[0007] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0008] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0009] In summary, the technical solution provided by the present invention achieves full coverage monitoring of the silo space through the triangular or polygonal distribution installation of multiple radar sensors, eliminates monitoring blind spots, and provides all-round material level sensing capabilities. Through the adaptive clustering radius mechanism, dual-threshold seed point selection and hierarchical clustering architecture, the radar point cloud clutter problem caused by the complex environment inside the coal feeder is effectively handled, and the accuracy and robustness of point cloud target extraction are enhanced. Through the three-level screening mechanism (spatial constraints, reflection characteristic constraints and time consistency constraints) and the working condition adaptive mechanism, the problem of complex reflection source interference is successfully solved, and the accuracy of material level signal recognition is significantly improved. Through the unified material level representation model and weighted fusion method, the problem of multi-sensor data conflict and complementarity is solved, and the coverage and measurement accuracy of material level detection are enhanced. By constructing three dynamic models and introducing an interactive multi-model mechanism, the problem of the traditional method's sudden drop in accuracy when the state changes is effectively solved, and the stability and prediction accuracy of material level tracking are significantly improved. The multi-stage solution strategy and iterative SOCP algorithm were adopted to successfully overcome the high computational complexity of traditional methods, achieve efficient control parameter optimization under different load conditions, improve the stability of coal feeder operation and reduce energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a schematic diagram of the steps of a millimeter wave radar material level monitoring method according to one embodiment of the present invention; Figure 2 This is a structural block diagram of a millimeter wave radar material level monitoring device according to an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0011] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0013] Reference Figure 1 This embodiment provides a millimeter wave radar material level monitoring method, comprising the following steps: S1, multiple millimeter-wave radar sensors are installed on the top of the coal feeder silo to collect a reflection point cloud dataset, and Euclidean clustering is performed on the reflection point cloud dataset to obtain a valid point cloud cluster set; A sensor layout was designed on top of the coal feeder silo to ensure comprehensive millimeter-wave radar coverage of the entire silo space. Three or more 77GHz millimeter-wave radar sensors were selected and arranged in a triangular or polygonal configuration, with each radar installed at an angle within ±5° to ensure sufficient overlap and complementarity within the space, minimizing blind spots. After installation, each millimeter-wave radar sensor was configured, setting key parameters such as operating frequency, bandwidth, sampling frequency, and beam direction to form a data sampling parameter set. Each millimeter-wave radar sensor was connected to a separate data acquisition unit. Each unit integrates a signal conditioning module, an analog-to-digital conversion module, and a data preprocessing module. The signal conditioning module amplifies and filters the raw radar echo signal. The amplification stage uses a gain-adjustable amplifier to enhance weak signal recognition, while the filtering stage employs a 5MHz Butterworth bandpass filter to effectively suppress industrial interference and background noise, ensuring signal purity. The filtered analog echo signal is input into the analog-to-digital conversion module for high-precision digitization. A high-performance ADC chip with a sampling rate of at least 20 MHz and a resolution of 16 bits is used to convert the analog signal into a digital signal. The digital signal is then input into the data preprocessing module for range-Doppler processing, converting the time-domain echo signal into initial point cloud data containing target range, relative velocity, and reflection intensity. This point cloud data represents the core spatial reflection information detected by the millimeter-wave radar, recording key parameters such as range, reflection intensity, and timestamp in the form of data pairs. After forming the initial point cloud, all data from different radars are aggregated to a central processing unit via a high-speed industrial bus. This central unit features a time synchronization mechanism supported by the IEEE 1588 protocol, precisely aligning data frames from different sensors to form a complete, temporally consistent point cloud dataset, known as the reflection point cloud dataset. Euclidean clustering is then performed on the reflection point cloud dataset. This clustering process, based on Euclidean distance, groups densely packed point clouds by setting appropriate cluster radii, thereby identifying cluster units representing physical object boundaries or surfaces, ultimately outputting a set of valid point cloud clusters.

[0014] Outlier points are filtered from the reflective point cloud dataset using statistical methods. The average spatial distance between each point and its nearest neighbors is calculated, and a global statistical distribution analysis is performed. When the local average distance of a point is significantly higher than the global mean, it is considered an outlier and removed. A coordinate transformation is performed on the point cloud data after outliers are filtered. By measuring the obtained sensor spatial pose and position matrix, the position transformation of each point in three-dimensional space is achieved with sub-centimeter accuracy. This establishes preprocessed point cloud data in a unified coordinate system. Local point cloud density features are extracted from the preprocessed point cloud in the unified coordinate system. By counting the number of neighboring points or the density of their distribution within a certain neighborhood, the spatial clustering radius used in subsequent clustering is dynamically adjusted, allowing the clustering strategy to adaptively adapt to the distribution changes of the point cloud in different regions. A clustering parameter set is generated based on the local density, providing a basis for subsequent clustering operations. A dual-threshold seed point selection strategy is implemented within the point cloud data. For each point, two metrics are calculated: the density of its surrounding neighborhood, which measures whether the point is located within a structure; and the echo signal strength of the point, which reflects the nature of the reflecting surface. The system selects only those points whose density and signal strength values both exceed preset thresholds as valid seed points, thus eliminating noise points or unrepresentative edge points. Hierarchical clustering is then performed on valid seed points, using a larger radius for coarse clustering to divide the point cloud into a number of coarse regions. Within each region, the clustering is then refined using a pre-set adaptive clustering radius to ensure internal consistency and external independence. Cluster validation and boundary optimization are then performed on the initial set of clusters. Invalid small clusters are eliminated, and the remaining clusters undergo boundary shape optimization, such as using a convex hull boundary construction algorithm, to ensure that each cluster has a clear and compact boundary outline, thus forming a valid set of point cloud clusters.

[0015] S2, perform multi-level information screening on the effective point cloud cluster set to obtain the effective material level cluster set of each millimeter wave radar sensor; Specifically, a coal feeder spatial reference model is established, which includes the geometric boundary information of the silo, the spatial extent of the internal mechanical structure, and the active area at the normal material level. The internal structure of the coal feeder is modeled using 3D laser scanning or structured light measurement technology, forming a 3D spatial reference framework that includes the silo inner wall boundary, the mechanical structure boundary, and the active area at the material level. Each cluster in the valid point cloud cluster set is spatially matched to the coal feeder spatial reference model. By calculating the minimum distance between the cluster center and the boundary of each spatial area, the spatial distance feature of each cluster relative to the silo structure is extracted. These features constitute a spatial distance feature set, which contains the distance relationship between the cluster's location in 3D space and the silo inner wall, mechanical components, and active area. Based on the spatial distance feature set, a first-level spatial constraint filtering operation is performed to determine whether the cluster is within the set active area at the normal material level. Clusters whose distance to the active area is less than a specified spatial threshold are retained, while clusters that are significantly close to the mechanical structure or the silo boundary are eliminated, resulting in a spatial filtering result set. The spatially filtered result set is then subjected to reflection characteristic constraint screening. This screening mechanism constructs a characteristic function based on the differences in millimeter-wave reflection responses of different object materials. This function comprehensively considers the cluster's average reflection intensity, the amplitude of reflection intensity variation, and the spatial distribution of reflection intensity. Each cluster is assigned a comprehensive score to determine whether it exhibits typical coal reflectivity characteristics. Only clusters with scores above a threshold are retained, resulting in a reflection characteristic screening result set. To mitigate the effects of occasional reflections or random misjudgments on the recognition results, the reflection characteristic screening result set of the current frame is compared with the data of the previous time frame. The temporal persistence and positional variation trends of the same cluster are analyzed. Only clusters that remain stable across multiple consecutive time frames are considered valid coal locations, ultimately outputting a temporal consistency screening result set. An operating condition recognition mechanism is introduced to determine the current operating condition based on indicators such as sensor status, motor speed, and coal feed rate. Based on this information, the spatial distance threshold, reflection intensity threshold, and temporal consistency threshold used in each of the aforementioned screening stages are dynamically adjusted, allowing the screening process to adapt to the changing coal level characteristics under different operating conditions. This mechanism ultimately yields a valid coal level cluster set for each millimeter-wave radar sensor.

[0016] S3, performing signal detection intersection-combination ratio and reliability analysis on the effective material level clustering set of each millimeter-wave radar sensor to generate fused material level data; It should be noted that the effective material level clusters of each millimeter-wave radar sensor are projected onto a unified grid model. This grid model covers the entire silo monitoring area and is divided into equally spaced two-dimensional grid cells. Each grid cell is associated with a height value and a confidence parameter, forming an initial material level surface set. The cluster information observed by each radar sensor is mapped to the corresponding grid location, and its observation height and reliability information are recorded on the grid. Any two surface data provided by different sensors in the initial material level surface set are compared pairwise. By calculating the degree of spatial overlap between them in the grid dimension, the signal detection intersection-in-union ratio (IoU) is obtained. The ratio of the area jointly covered by the two sensors in the material level grid to their total coverage area is calculated to construct an IoU matrix, which reflects the degree of consistency in the monitoring data space between the individual sensors. To ensure the accuracy and rationality of the fusion results, an adaptive consistency threshold is set based on this IoU matrix. This threshold is dynamically adjusted according to the number of effective sensors participating in the fusion in the current system, thereby adapting to real-world conditions such as changes in the number of sensors or fluctuations in coverage overlap. When the IoU value of any pair of sensors in the IoU matrix falls below the consistency threshold, the system automatically triggers a conflict resolution mechanism. This involves identifying which sensor in the pair exhibits the greatest deviation and analyzing the source of the error based on factors such as historical stability, signal strength, and spatial location. Consequently, observations with significant deviations are adjusted or eliminated, generating a conflict resolution result. Each millimeter-wave radar sensor is assigned a reliability weight. This weight is dynamically calculated based on a comprehensive consideration of multiple factors, including the sensor's accuracy in historical measurements, the susceptibility of its installation location in the silo to interference, the stability of the current observation signal, and the continuity of previous and subsequent frames. This creates a sensor weight set. Based on this weight set and the conflict resolution results, each spatial location in the grid model is fused. Specifically, when multiple sensor observations are present at the same grid location, a weighted fusion approach is used to calculate the different height values. This weighted fusion takes the height values provided by each sensor, their corresponding weights, and their confidence levels, and generates a fused height value. The fused confidence level is then updated accordingly to reflect the trustworthiness of the integrated multi-source information. Spatial smoothing is performed on the fused height values and fusion confidence to remove mutations and jagged effects caused by discrete measurements between grids. Outliers are removed based on the height change trends of adjacent grids to ensure the spatial continuity and physical rationality of the fused material level surface. At the same time, to address data missing in certain areas due to occlusion or radar signal attenuation, a blind spot compensation mechanism based on spatial continuity constraints and historical time series data is introduced. This mechanism uses known reliable data from adjacent time frames or spatial neighborhoods to reasonably estimate and fill in missing areas, improving the integrity and stability of the material level data and ultimately obtaining fused material level data.

[0017] S4, inputting the fused material level data into the multi-model interaction and unscented Kalman filter model to perform dynamic material level tracking, and obtaining the time series material level curve and prediction confidence interval; Specifically, representative material level feature point data is extracted from the fused grid material level data. These feature points are located at material level surface boundaries, slope corners, or areas sensitive to height changes, and possess strong geometric characteristics and high representativeness. Curvature analysis, gradient identification, and confidence screening of the grid height field are performed to select a set of highly reliable spatial feature points. Their three-dimensional positions and corresponding confidence parameters at the current moment are then constructed as the input for dynamic tracking. The extracted feature point data require a stable correspondence between different time frames. Therefore, a joint probabilistic data association algorithm is introduced, and a time threshold parameter is set to control the maximum allowable range of change between the previous and next frames. This time threshold parameter is dynamically adjusted based on the current material level height change rate, with its double value serving as the maximum interval for temporal association to adapt to the actual dynamics of material level changes. An association probability matrix is constructed based on spatial matching and feature consistency analysis between the previous and next time frames. Each element in the matrix represents the probability of a feature point matching between the current frame and the previous frame. The calculation process comprehensively considers multiple factors, including point distance, normal vector direction consistency, and material level surface continuity. Based on the obtained association probability matrix, three dynamic models are constructed based on typical material level state variation patterns: a uniform velocity model, applicable to stable feeder operation, describing slow material level fluctuations with highly linear variations; a uniform acceleration model, applicable to startup / shutdown or sudden transition phases, describing material level variations with predictable acceleration; and a random acceleration model, applicable to conditions of severe fluctuations or abnormal disturbances, describing processes with uncertain material level height and large fluctuations. These three dynamic models together constitute a dynamic model set that covers the material level variation characteristics of most coal feeding conditions. To dynamically switch and combine models based on the operating state during actual tracking, an interactive multi-model mechanism is introduced within the dynamic model set. The transition probabilities between different models are estimated online and in real time, and a model transition probability matrix is constructed. Each element in this matrix represents the probability of the system transitioning from one model state to another. By weighted fusion of the outputs of multiple models and dynamically adjusting the participation ratio of each model, a model combination result containing the current optimal model combination is generated in real time. This result integrates the predictive capabilities and matching degree of each model, achieving smooth transitions between multiple motion states. The model combination results are input into an unscented Kalman filter to perform nonlinear state propagation and state estimation. The unscented Kalman filter propagates the state and covariance by selecting a set of sigma points distributed within the state space, avoiding the limitations of the traditional extended Kalman filter in derivation of the system model, thereby improving adaptability to nonlinear material level systems and estimation accuracy. The state estimation output by the filter includes information such as the current material level height, rate of change, and acceleration. Based on the filtering results of the continuous time series, the system forms a time-series material level curve, reflecting the dynamic evolution of the material level over time.The filter's process noise covariance matrix and measurement noise covariance matrix are dynamically adjusted based on the current operating conditions. The process noise reflects the uncertainty of the material level variation model, while the measurement noise accounts for the impact of millimeter-wave radar measurement errors. By continuously updating these two types of noise parameters, the system more effectively models uncertainty and calculates prediction confidence intervals for a future period based on the covariance results of the filter estimate, ultimately generating a time-series material level curve and prediction confidence intervals.

[0018] S5, based on the time series material level curve and the prediction confidence interval, iterative second-order cone programming calculation is performed to obtain the coal feeder control parameter sequence.

[0019] Among these, a mathematical model was established to describe the dynamic changes in material level. The core of this model is the material level change rate model. This model analyzes the input and output processes of materials in the silo and expresses the rate of change of material level over time as the difference between the input and output functions. These two functions are nonlinear functions of the key control variables of the coal feeder—the feed rate, rotational speed, and gate opening. By physically modeling the material conveying process of the coal feeding system and fitting historical data, these functions are expressed as continuously differentiable nonlinear structures, thereby characterizing the response mechanism of material level height to the control variables. Upper and lower limits and dynamic change rate limits are set for each control variable to form a set of control parameter constraints. This set of constraints includes the maximum and minimum values of the feed rate, rotational speed, and gate opening within the operating range, as well as the maximum allowable variation of these three within each time step, to prevent mechanical shock or system instability caused by sudden parameter changes. Based on the aforementioned constraints and the time-series material level curve and its predicted confidence interval obtained by the unscented Kalman filter in the previous stage, an optimization objective function is constructed. This objective function comprehensively considers multiple deviation factors, including the deviation between the current material level and the target material level, the deviation between the coal feed rate and its ideal reference value, the deviation in rotational speed, and the deviation in gate opening. The goal is to minimize the weighted sum of squares of all deviations within the predicted time domain, thereby meeting the material level control requirements while maintaining the economy and smoothness of the control operation. Because the optimization objective function contains nonlinear constraints and a non-convex structure, the problem is transformed into a quadratic constrained quadratic programming problem with a quadratic objective function and quadratic constraints, and is solved using an iterative second-order cone programming algorithm. This algorithm linearizes or convexifies the original non-convex problem into a subproblem with a second-order cone structure in each iteration, enabling rapid approach to the local optimal solution while ensuring the constraints are feasible. A convergence tolerance threshold and a maximum number of iterations are set as termination criteria. When the parameter update change is less than the set threshold or the number of iterations reaches the upper limit, the initial control parameter solution is output as the first-stage solution. The actual operating constraints of the coal feeder are introduced into the initial control parameter solution. The initial solution is re-optimized by combining the dynamic programming algorithm. The impact of the control sequence on the material level changes in the next few steps is analyzed through a rolling window method. The control amount is adjusted to take into account both real-time performance and prediction accuracy, and the optimized control parameter sequence is obtained. In order to enhance the system's adaptability to external disturbances, such as changes in coal quality, equipment fluctuations, or sudden changes in load, a disturbance detection mechanism is set in the control link to evaluate in real time whether the current prediction error is abnormally expanding. If a disturbance is detected, the compensation processing mechanism is activated. This mechanism corrects parameters based on historical disturbance response experience and quickly switches to the preset control parameter template under different load levels, thereby ensuring the robustness and response speed of the control system. Combining the optimized parameter sequence with the actual operating condition judgment, the system outputs the coal feeding rate, speed, and gate opening control parameters applicable at the current moment to form the coal feeder control parameter sequence.

[0020] In one example, multiple millimeter-wave radar sensors were installed on the top of a coal feeder silo to collect a reflection point cloud dataset. Euclidean clustering was then performed on the reflection point cloud dataset to obtain a valid point cloud cluster set, including: Multiple millimeter-wave radar sensors are installed on top of the coal feeder silo to fully cover the silo space. The millimeter-wave radar sensors are then parameterized to obtain data sampling parameters. Connect each millimeter-wave radar sensor to an independent data acquisition unit, which includes a signal conditioning module, an analog-to-digital conversion module, and a data preprocessing module; The radar echo signal is collected according to the data sampling parameters, and the radar echo signal is amplified and filtered through the signal conditioning module to obtain an amplified echo signal, and the amplified echo signal is converted into a digital signal through the analog-to-digital conversion module; The digital signal is input into the data preprocessing module to perform range-Doppler processing to obtain initial point cloud data, and the initial point cloud data is transmitted to the central processing unit for time synchronization to generate a reflection point cloud data set; Euclidean clustering is performed on the reflection point cloud dataset to obtain a valid point cloud cluster set.

[0021] In this example, multiple millimeter-wave radar sensors are installed on top of the coal feeder silo. Three or more millimeter-wave radar sensors are selected and distributed symmetrically in a triangular or polygonal structure. The installation angle of the sensors is controlled within a vertical deviation range of no more than ±5° to ensure that the beam of each sensor not only covers the area directly below it, but also overlaps with adjacent sensors, thereby effectively eliminating blind spots and avoiding data loss caused by uneven material stacking heights or obstructions caused by internal structures. At the same time, after the physical installation is completed, all millimeter-wave radar sensors are configured with parameters, including key parameters such as operating frequency band, transmission bandwidth, pulse repetition frequency, beam angle, signal gain, as well as core performance indicators such as sampling frequency and distance resolution. After the parameter configuration is completed, each millimeter-wave radar sensor is independently connected to the corresponding data acquisition unit. This unit serves as the core device for front-end signal acquisition and preliminary processing. It integrates a signal conditioning module, an analog-to-digital conversion module, and a data preprocessing module to form a signal processing link. The signal conditioning module performs gain control and filtering on the raw radar echo signal, amplifying weak signals through a low-noise amplifier. A bandpass filter then applies bandpass filtering to suppress frequency components, preserving the radar's effective frequency band while suppressing interference signals in irrelevant frequency bands. This results in an amplified and filtered echo signal. This analog signal then enters the analog-to-digital conversion module for digitization. Using a high-sampling rate (e.g., 20 MHz) and high-precision (e.g., 16-bit) analog-to-digital converter, it maps the analog waveform into a digital signal suitable for further processing by the digital system, maintaining waveform integrity in both time and amplitude dimensions to ensure accurate range measurement. The digital signal is then input into the data preprocessing module within the data acquisition unit for combined range-Doppler processing. This process applies window function weighting and a fast Fourier transform to the radar signal, extracting target information in both the range and velocity domains. This generates raw point cloud data at each time point. This data includes the spatial distance of the reflection point, as well as multiple feature dimensions such as reflection intensity, relative velocity, and timestamp. The initial point cloud collected by each millimeter-wave radar is transmitted in real time to the central processing unit via a high-speed industrial communication bus (such as a CAN-FD bus or industrial Ethernet). Within the central processing unit, all point cloud data is time-synchronized to avoid time deviations caused by sampling delays or internal clock drift among the radar devices. This time synchronization mechanism utilizes the IEEE 1588 Precision Time Protocol, which limits the time error of multi-source data to less than 1 millisecond, ensuring temporal consistency and physical rationality for subsequent fusion analysis. After time-synchronized integration, a reflection point cloud dataset containing the acquisition results from multiple sensors is obtained, featuring high temporal resolution, high spatial accuracy, and multi-source redundancy. Euclidean clustering is performed on this reflection point cloud dataset to extract spatially structured and physically meaningful reflection targets from the large-scale, complex point cloud data.The algorithm analyzes the spatial locations of the point cloud and constructs an adjacency graph based on the Euclidean distance between each point. Cluster search is then performed within a set initial cluster radius, grouping points that are sufficiently close together. To improve the adaptability and accuracy of clustering, a density-adaptive mechanism is introduced in the cluster radius setting. This dynamically adjusts the radius based on the density of each point within its local area, enhancing the algorithm's ability to handle both sparse and dense areas. To prevent false clustering starts, a dual-threshold mechanism is introduced in seed point selection. This combines the neighborhood density and reflection intensity of a point. Only when both meet the preset conditions is the point considered a valid seed point for cluster expansion, thereby improving the quality of the initial clustering. After performing Euclidean clustering on the entire point cloud, the initial clustering results are validated, eliminating noisy clusters with too few points. The boundaries of each cluster are optimized, for example, by extracting the outer contour using a convex hull algorithm, to ensure geometric clarity and physical continuity of the cluster structure. This results in a set of valid point cloud clusters with high confidence and spatial consistency.

[0022] In one example, Euclidean clustering is performed on a reflection point cloud dataset to obtain a valid point cloud cluster set, including: Filter outliers on the reflected point cloud data set to obtain point cloud data without outliers, and perform coordinate transformation on the point cloud data without outliers to obtain a preprocessed point cloud in a unified coordinate system. Extract local point cloud density features from the preprocessed point cloud in a unified coordinate system, dynamically adjust the clustering radius based on the local point cloud density features, establish an adaptive clustering radius mechanism, and obtain a clustering parameter set; Based on the clustering parameter set, a dual-threshold seed point selection strategy is implemented to calculate the neighborhood density value and signal strength value of each candidate point respectively, and only points whose neighborhood density value and signal strength value meet the threshold conditions at the same time are selected as valid seed points; The effective seed points are hierarchically clustered to obtain an initial cluster set, and cluster validity verification and boundary optimization are performed on the initial cluster set to obtain a valid point cloud cluster set.

[0023] In this example, outlier points are identified and removed from a reflection point cloud dataset collected by multiple millimeter-wave radar sensors. In actual coal feeder operation, due to factors such as equipment vibration, electromagnetic interference, reflection from metal attachments, and dynamic coal dust lifting, the radar system will capture some point cloud data with extreme spatial locations or abnormal reflection intensities. These points are spatially isolated and lack a continuous neighborhood structure. A statistically-based outlier detection algorithm is used to analyze the spatial neighborhood of each point. The average distance between the point and its nearest several (e.g., eight) neighbors is calculated and compared with the average distance and standard deviation of the global point cloud data. If the local average distance of a point is significantly higher than twice the overall standard deviation, it is marked as an outlier and removed. This mechanism effectively removes isolated stray points and anomalous reflection points, resulting in cleaned point cloud data with a tighter structure and more reliable signals. The cleaned point cloud is mapped from the local coordinate system of each radar sensor to a common reference coordinate system. Laser ranging, 3D calibration, and relative attitude measurement during the installation phase are used to obtain the spatial rotation matrix and translation vector of each sensor relative to the silo structure. The 3D coordinates of each point are then mapped to a unified coordinate system using a rigid transformation. Coordinate normalization eliminates viewpoint deviations caused by differences in radar installation positions, resulting in a preprocessed point cloud set in the same spatial reference system. Density features are then extracted from the point cloud within this unified coordinate system. Point cloud density is a key metric for describing the complexity of the local region of a point in space. It is expressed as the number of neighboring points within a fixed radius or the average distance between points. Higher density values indicate more compact structures and the presence of physical boundaries or surfaces, while lower density values indicate sparse background or distant reflections. Based on this characteristic, the local density distribution of each point is estimated by searching for the number of neighbors within a certain neighborhood, with each point as the center. This information is used to establish an adaptive clustering radius mechanism. The core of this mechanism lies in dynamically adjusting the search radius used in the clustering process based on the density of the local point cloud. Dense areas use a smaller clustering radius to improve boundary resolution, while sparse areas use a larger radius to avoid missegmentation. In this way, the clustering algorithm maintains high adaptability and robustness in complex and changing point cloud structures, forming a set of clustering parameters optimized for the current data structure. A dual-threshold seed point selection strategy is implemented based on the clustering parameter set, and candidate seed points are selected based on two key metrics: neighborhood density and radar reflection signal strength. The former ensures that the selected seed point is located within a dense area, thereby ensuring stable expansion capabilities. The latter determines whether the point originates from a physical material surface or a physical object boundary, as the actual reflection signal at the coal-material interface is significantly stronger than the background or mechanical structure. A point is selected as a valid seed point only if both metrics exceed the set threshold.The effective seed points are hierarchically clustered to enhance the algorithm's ability to recognize multi-scale structures. A larger initial clustering radius is used to perform the first round of coarse clustering on the global point cloud to quickly lock the approximate contour range of each major material level area and reflection area; on this basis, guided by the aforementioned adaptive clustering radius, a second round of refined clustering is performed within each coarse clustering area to redivide the boundaries and internal structures, thereby achieving a combination of global positioning and local refinement. This hierarchical clustering process enables the algorithm to maintain a complete perception of complex structures while adapting to boundary details and local undulations, effectively preventing problems such as insufficient aggregation or over-segmentation caused by a fixed clustering radius. After clustering is completed, the initial cluster set is validated and its boundaries optimized. The validation is performed by counting the number of points, average density, and shape structure indicators of each cluster, eliminating small clusters composed of isolated noise and with a number of points below the set threshold, in order to eliminate invalid results caused by the algorithm's mis-aggregation. Boundary optimization, based on geometric analysis methods, involves constructing a convex hull boundary for each cluster point set to describe its spatial contours, removing burrs or breakpoints along the boundary, and forming a well-enclosed, well-defined 3D cluster structure. The result is a set of valid point cloud clusters with a clear spatial distribution, good density continuity, and clearly defined boundaries.

[0024] In one example, a multi-level information screening is performed on the valid point cloud cluster set to obtain a valid material level cluster set of each millimeter wave radar sensor, including: Establish a coal feeder spatial reference model, which includes the silo inner wall boundary, mechanical structure boundary and normal material level activity area; Perform spatial matching between each cluster in the valid point cloud cluster set and the coal feeder spatial reference model to obtain a spatial distance feature set; Perform spatial constraint screening based on the spatial distance feature set, retain the clusters located in the normal material level activity area, and eliminate the reflection points that are obviously deviated from the normal material level activity area to obtain the spatial screening result set; Perform reflection characteristic constraint filtering on the spatial filtering result set to obtain a reflection characteristic filtering result set, compare the reflection characteristic filtering result set with the historical frame data, perform time consistency constraint filtering, and obtain a time consistency filtering result set; The screening threshold is dynamically adjusted according to the current operating status of the coal feeder, and the working condition adaptation processing is performed on the time consistency screening result set to obtain the effective material level clustering set of each millimeter wave radar sensor.

[0025] In this example, a spatial reference model of the coal feeder is constructed. This model serves as the benchmark for all subsequent screening and judgment. Its structure consists of three core components: the silo's inner wall boundary, the mechanical structure boundary, and the spatial extent of the normal material level active area. High-precision 3D laser scanning or structured light measurement technology is used to capture the spatial topography of the feeder's silo interior. Based on design drawings and on-site measurements, 3D contour point sets are extracted for each boundary type. A standardized spatial structural model is generated using CAD modeling tools. The silo's inner wall boundary refers to the edge of the physical container holding the coal, constraining the spatial limits of coal distribution. The mechanical structure boundary includes the areas occupied by fixed or moving components within the feeder, such as the motor, reducer, and support structures. These areas generate high-intensity reflections in the point cloud but do not represent material level. The normal material level active area is the effective area of coal, defined based on long-term operating experience. It lies between the bottom outlet and the top safety distance of the silo. Its spatial extent must cover the actual material level fluctuation range while excluding the influence of interfering boundaries. It serves as a key spatial reference for point cloud screening. When the spatial model is established, the effective point cloud cluster sets obtained by the clustering algorithm in the previous stage are spatially matched with the model one by one, that is, the center point, boundary point or centroid coordinates of each cluster are geometrically compared with the boundaries of different areas in the model, and the minimum Euclidean distance of each cluster point set to each reference boundary is calculated to form a spatial distance feature set, recording the spatial relationship between the cluster and the inner wall of the silo, the mechanical structure and the material level activity area. Based on these distance features, a spatial constraint screening process is performed. By judging whether each cluster is within the set material level activity area space, cluster points that are obviously close to the silo wall, the edge of the equipment, or in non-coal distribution areas such as above and below the equipment are screened out. For example, when the distance from the cluster center to the boundary of the normal material level activity area is greater than a certain threshold, the cluster is identified as background information caused by non-coal reflection and is eliminated to form a spatial screening result set. Spatially constrained filtering is performed based on a spatial distance feature set. A reflection characteristic scoring mechanism is constructed based on historical sample training or empirical reflection models. This mechanism considers three dimensions: the average reflection intensity of the cluster, the variance of the reflection intensity within the cluster, and the reflection pattern characteristics of the cluster points in their spatial distribution, such as whether they exhibit continuous surface characteristics or blurred boundaries. These features are weighted and synthesized into a scoring function, and a reasonable threshold is set for judgment. Only clusters with scores above the set value are retained as the reflection characteristic screening result set. The reflection characteristic screening result set of the current frame is then compared with the historical data of the previous frame or even multiple frames for temporal consistency.Based on the spatial position, reflection intensity change rate, and geometric stability of the clusters, the persistence of the same cluster between adjacent frames is compared. Only clusters that maintain relatively stable spatial position, continuous structural morphology, and consistent reflection characteristics across several consecutive frames are considered valid material level reflection objects. Clusters that appear only in a single frame or a short period of time and quickly disappear are considered sporadic interference points and are eliminated. This approach constructs a temporal consistency screening result set. To adapt to the various operational changes faced by coal feeders during actual operation, such as startup, shutdown, rapid load increases or decreases, and changes in coal type, the current operating state is incorporated into the screening strategy, implementing an adaptive operating condition adjustment mechanism. This mechanism determines the current system operating condition by real-time analysis of indicators such as coal feed rate, motor speed, control commands, and material level change rate. It then dynamically adjusts the aforementioned spatial constraint threshold, reflection characteristic score lower limit, and temporal consistency evaluation parameters based on pre-set rules. This makes the entire screening process flexible and adaptable to diverse environmental changes, ensuring that true signals are not filtered during startup, excessive misjudgments are avoided during stable periods, and fault tolerance is maintained during periods of fluctuation. Based on the quadruple joint screening of spatial model constraints, reflection characteristic constraints, temporal stability constraints and working condition adaptive mechanism, a set of highly reliable, structurally continuous and temporally stable effective material level clusters are extracted from multi-source radar data, corresponding to the spatial perception results of each millimeter-wave radar sensor.

[0026] In one example, signal detection intersection-over-union and reliability analysis are performed on the effective material level clusters of each millimeter-wave radar sensor to generate fused material level data, including: The effective material level clustering set of each millimeter-wave radar sensor is projected onto the unified grid model to obtain the initial material level surface set; Calculate the signal detection intersection-and-union ratio for any two surfaces in the initial material level surface set to form an intersection-and-union ratio matrix; An adaptive consistency threshold is set based on the intersection-in-union ratio matrix. When the intersection-in-union ratio of any pair of sensors is lower than the threshold, the conflict resolution mechanism is triggered to obtain the conflict resolution result. Assigning reliability weights to each millimeter-wave radar sensor to obtain a sensor weight set; Based on the sensor weight set and conflict processing results, a weighted fusion calculation is performed on the different height value detection results of multiple sensors at the same grid point to obtain the fused height value and fusion confidence; The fused height value and fusion confidence are post-processed by spatial smoothing and outlier removal, and the detection blind spots are filled by spatial continuity constraints and historical data to obtain the fused material level data.

[0027] In this example, the effective material level clusters from each millimeter-wave radar sensor are projected onto a unified grid model. This grid model covers the entire projected area of the silo and is divided into sections at a fixed 2D resolution. Each grid cell represents a fixed physical area, and its data structure includes three basic attributes: spatial coordinates, height, and confidence. Through this process, the 3D cluster data from radar sensors in different orientations are uniformly projected onto the same 2D position index system, forming a set of initial material level surfaces corresponding to multiple independent sensors. Each surface represents the sensor's height estimate of the silo surface and records the coordinates of the grid points it covers, as well as the corresponding height and confidence level. The signal detection intersection-over-union ratio (IoU) is calculated for any two surfaces in the initial material level surface set. For each pair of initial surfaces, the ratio of their overlapping area to their joint coverage area at the grid level is calculated. The intersection is the number of grid cells where both sensors provide data, and the union is the total number of grid cells where either sensor provides data. The ratio of these two grid cells is the signal detection intersection-over-union ratio (IoU) for that pair of sensors. In this way, a complete intersection-in-union (IoU) matrix is calculated. Each element of this matrix represents the degree of similarity between a pair of sensors' spatially correlated material level detection results. Values closer to 1 indicate high detection consistency, while lower values indicate significant discrepancies or potential errors in the material level estimates between the two sensors. To determine whether this discrepancy is acceptable, an adaptive consistency threshold is constructed based on this IoU matrix. This threshold is determined based on the number of sensors and the number of active sensors in the system. The greater the number of sensors, the lower the tolerance for single-pair inconsistency. Therefore, a dynamic threshold function is constructed that varies with the number of active sensors. For example, when the number of sensors is large, the threshold is higher, reflecting the system's strict consistency requirements; when the number of sensors is small, the threshold is appropriately lowered to enhance the system's redundancy and fault tolerance. If the IoU between any pair of sensors falls below the current adaptive threshold, the system immediately triggers a conflict resolution mechanism. This mechanism determines the source of the conflict based on factors such as the stability of each sensor's historical detections, the reflection intensity distribution of the current frame, the degree of spatial matching, and the grid coverage density. It then decides whether to exclude any anomalous surfaces or downgrade the detection values within the conflicting area, thereby preventing erroneous data from influencing the final fusion results. After conflict resolution is complete, each millimeter-wave radar sensor is assigned a reliability weight to form a sensor weight set. This weight is dynamically calculated based on each sensor's multi-dimensional performance indicators, including historical stability (such as the average error and anomaly rate over the past few frames), installation location reliability (such as the ratio of viewing angle coverage to blind spot ratio), the mean and variance of the signal strength in the current frame, the continuity of the spatial distribution within the detection area, and the sensor hardware status. By weightedly combining these indicators, each sensor is assigned a different reliability rating to form a sensor weight set. Based on the sensor weight set and the conflict resolution results, a weighted fusion calculation is performed on the different height detection results of multiple sensors at the same grid point, outputting a fused height value and fusion confidence.The fused height values and fused confidence levels are then post-processed with spatial smoothing and outlier removal. Spatial smoothing suppresses jumps within the local area by applying a weighted median filter or Gaussian filter to the height values of the neighborhood surrounding each grid point, thereby avoiding sudden changes in material level caused by single-point anomalies. Outlier removal identifies isolated, spatially distributed, anomalous grid points with significant height differences from surrounding points. The outlier removal process determines whether these points represent physical anomalies or system errors based on the changing trends of the current and historical frames. If these points are considered anomalous, they are replaced with interpolation or directly removed. To compensate for local data loss caused by radar wave obstruction, reflection blind spots, and coal dust adsorption, spatial continuity constraints and historical data compensation mechanisms are introduced. In the spatial dimension, the edges of missing areas are expanded and filled based on the relationship between the height gradients and curvatures of adjacent grid points. In the temporal dimension, blind spots are estimated based on the stable trends of the previous frame or multiple frames of historical data, restoring the continuity of the material level surface in both space and time, ultimately yielding fused material level data.

[0028] In one example, the fused material level data is input into the multi-model interaction and unscented Kalman filter model to perform dynamic material level tracking, and a time-series material level curve and prediction confidence interval are obtained, including: Extract material level feature point data from fused material level data; A time threshold parameter is set for the material level feature point data. The time threshold parameter is twice the current material level change rate. The previous and next frame data are matched using a joint probability data association algorithm to obtain an association probability matrix. Based on the correlation probability matrix, a uniform speed model for stable coal feeding state, a uniform acceleration model for start-stop transition state, and a random acceleration model for fluctuating state are constructed, and the uniform speed model, uniform acceleration model, and random acceleration model are taken as a dynamic model set. Introducing an interactive multi-model mechanism into the dynamic model set and estimating the model transition probability matrix online to determine the model combination result containing the current optimal model combination; The model combination result is input into the unscented Kalman filter to perform material level state estimation and nonlinear system propagation to obtain the filtering result; A time series material level curve is formed according to the filtering results, and the process noise covariance matrix and the measurement noise covariance matrix are dynamically adjusted according to the changes in working conditions. The prediction confidence interval is calculated to obtain the time series material level curve and the prediction confidence interval.

[0029] In this example, key feature points are extracted from the fused, high-confidence material level grid data as input elements for the dynamic state. These feature points are located where the material level surface structure experiences dramatic structural changes, such as at the highest points of the pile, local corners, boundary transitions, or locations with the most pronounced vibration trends. The extraction process comprehensively considers factors such as grid curvature, gradient direction, spatial continuity, and confidence strength to select a representative, spatially evenly distributed, and physically meaningful set of feature points. These points constitute a discrete sample of the material level state space for subsequent state propagation and matching. After extracting the feature points, due to the continuous nature of material level changes and a certain time lag, point-level matching relationships are established between previous and subsequent frames to ensure tracking path consistency and state estimation accuracy. A time threshold parameter is set for feature point matching, which is dynamically adjusted based on the current material level change rate. A time tolerance of twice the current level change rate is selected to ensure that acceptable matching point pairs are found within a reasonable range, even in scenarios with dramatic fluctuations. Within this time threshold, a joint probabilistic data association algorithm is used to fully match the feature points of the current frame with those of the previous frame or frames. This algorithm considers spatial distance and integrates multi-dimensional factors such as the point's direction vector, surface normal, curvature similarity, and previous and subsequent height trend changes. For each current feature point, a joint probabilistic matching relationship is established with historical feature points. Based on this, a complete association probability matrix is constructed. Each element in the matrix reflects the probability that a current point and a historical point belong to the same physical point, thus establishing a probabilistic basis for cross-frame tracking paths. Based on the temporal evolution trajectory provided by the association probability matrix, the coal feeding state is modeled, and three representative motion models are constructed to form a dynamic model set. The uniform velocity model describes the linear rise or fall of material level under conditions of continuous coal feeding and stable system load. Its state transitions assume a linear change in height over time while maintaining a constant velocity. The uniform acceleration model is used during startup and shutdown processes or load regulation, where the rate of rise or fall of material level varies. Therefore, an acceleration term is introduced into the state vector to capture the nonlinear acceleration trend. The random acceleration model describes unpredictable and highly fluctuating coal feeding conditions, such as sudden changes in coal quality, equipment interference, or temporary blockages. Its state transitions incorporate a strong process noise component to simulate state uncertainty under complex disturbances. These three models, each with their own unique descriptive capabilities, together form a switchable multi-model system that adapts to the dynamic characteristics of material level under different operating conditions. To enable intelligent model switching during actual operation, an interactive multi-model mechanism is introduced within this dynamic model set. This interactive multi-model framework allows multiple models to participate in state estimation at each moment, and their participation weights are dynamically adjusted via a model transition probability matrix. The transition probability matrix represents the probability of any model being transformed into another model at the current moment. The matrix is obtained from historical operation data statistics, or estimated and updated online in real time to ensure that the model switching is smooth and has a realistic basis.Within the interacting multi-model framework, state predictions from three models are run in parallel for each feature point, yielding corresponding state estimates and covariances. The predictions from each model are then weighted and fused based on the model weights and transition matrix at the previous moment, along with the matching degree of the current observation. This output represents the estimated result of the optimal combined model. This combined prediction is then fed into an unscented Kalman filter, which performs nonlinear propagation and correction on the state vector of each feature point. Compared to the traditional extended Kalman filter, the unscented Kalman filter achieves high-order state estimation by constructing a set of Sigma points distributed in the state space and propagating their responses in the nonlinear system. This eliminates the need to differentiate the system function and avoids the error amplification associated with Jacobian matrix linearization. The unscented Kalman filter can handle highly nonlinear systems with uncertain noise covariances, making it suitable for material level tracking scenarios under complex coal feeding conditions. Its state vector includes the current height, as well as velocity, acceleration, and system dynamic response indicators. Together with the confidence level in the fused data, it constrains state updates, enhancing the stability and reliability of the estimate. An unscented Kalman filter continuously estimates the state of each frame, reconstructing a complete time-series material level curve on the time axis. This curve reflects the dynamic evolution of the material level, including the current real-time altitude and inferring the short-term future altitude change trend. To enhance predictive capability and noise immunity, the process noise covariance matrix and measurement noise covariance matrix within the filter are adjusted based on the currently detected operating state. If the system is stable, the process noise weight is appropriately reduced to enhance the filter's response speed. If the system is fluctuating or uncertain, the process noise level is increased to expand the estimation interval and prevent misjudgment caused by rapid convergence. The measurement noise covariance is dynamically adjusted based on the fusion confidence level. A higher confidence level indicates higher measurement reliability, while a lower confidence level requires a larger observation error model. Based on the covariance results of the filtered output state, a time-series material level curve and the upper and lower confidence interval boundaries are constructed.

[0030] In one example, an iterative second-order cone programming calculation is performed based on the time-series material level curve and the prediction confidence interval to obtain a coal feeder control parameter sequence, including: The feeder's input and output functions are expressed as nonlinear functions of the feed rate, feeder speed, and gate opening, and the material level change rate model is obtained. A control parameter constraint set is set for the coal feeder control parameters, which includes the upper and lower limits of the value range of the coal feeding rate, the coal feeder speed and the gate opening, as well as the maximum change rate limit; Based on the time series material level curve and prediction confidence interval, an optimization objective function including material level deviation, coal feeding rate deviation, speed deviation and gate opening deviation is constructed; The optimization objective function is transformed into a quadratic constrained quadratic programming problem. The non-convex optimization problem is decomposed into multiple convex sub-problems through an iterative second-order cone programming algorithm. The convergence condition and iteration upper limit are set to obtain the initial control parameter solution. The actual operation constraints of the coal feeder are introduced into the initial control parameter solution, and the parameters are optimized and adjusted through dynamic programming to obtain the optimized control parameter sequence. Based on the optimized control parameter sequence and the current load condition, disturbance detection and compensation processing are performed, and online strategy switching is performed in combination with the preset control parameter template to generate the coal feeder control parameter sequence.

[0031] In this example, the feeder's inlet and outlet functions are expressed as nonlinear functions of the feed rate, feeder speed, and gate opening. The feed function is determined by the feed rate, speed, and gate opening. In actual operation, the material flow velocity depends not only on the belt speed or spiral speed but also on the dimensional variations of the material cross-section, which are directly influenced by the gate opening. The outlet function, on the other hand, is related to the combustion rate corresponding to the power generation load. While this is a passive reflection, it is also influenced by controllable factors such as the ash return ratio and airflow disturbances. After modeling these two functions as multivariate nonlinear functions, the difference between them constitutes the material level change rate function—the trend of the material level increase or decrease per unit time. Based on this model, bounds and rate of change constraints are imposed on the control variables to construct a set of control parameter constraints. This set of constraints defines upper and lower limits for the feed rate, speed, and gate opening within the physical equipment capabilities. For example, the feed rate should not exceed the maximum conveying capacity of the belt, the speed should be controlled within the rated motor speed range, and the gate opening is limited by the minimum closed and maximum open values of the structural design. At the same time, in order to prevent the control instructions from undergoing drastic mutations between consecutive time steps, causing mechanical shock or system oscillation, the constraint set sets a maximum rate of change limit for each control variable to ensure the continuity and smoothness of the control quantity. An optimization objective function is constructed based on the time-series material level curve and prediction confidence interval generated by the unscented Kalman filter. The design of this function takes into account the deviation between the current material level and the target material level, and incorporates the degree of deviation between each control parameter and its expected reference value into the optimization objective. That is, the objective function contains four core contents, namely the material level deviation term, the coal feed rate deviation term, the speed deviation term and the gate opening deviation term. Among them, the material level deviation term reflects the control strength of the system on the material level stability, while the deviation term of the control variable reflects the balance weight of energy consumption optimization, system life maintenance and control stability. These deviation terms are weighted in the form of square penalties to construct the objective function, and form a multi-objective optimization problem with nonlinear variable relationships and constraint structures in the overall structure. Because both the input and output functions are nonlinear expressions and the optimization objective is non-convex in function space, the system transforms the problem into a solvable form. Specifically, it constructs an approximately solvable quadratically constrained quadratic programming problem framework and solves it using an iterative second-order cone programming algorithm. During the second-order cone programming process, the original non-convex optimization problem is linearly relaxed and convexified. At each iteration, a set of convex subproblems that satisfy the current variable boundaries and constraints are constructed. These subproblems are solved using weighted interior point methods or projected gradient methods, continuously approaching the global optimal solution.Each iteration selects the optimal direction within the current step size within the control variable space and updates the control parameters. The system also evaluates the improvement in the objective function and the satisfaction of constraints. Convergence criteria are set, such as the objective function change being less than a threshold, the parameter update amplitude being below a set standard, or the maximum number of iterations reaching an upper limit, to control the algorithm's operational scale. Ultimately, an initial control parameter solution is obtained upon termination of the iteration. The actual operating constraints of the coal feeder are incorporated into the initial control parameter solution to account for more dynamic constraints and execution path optimization in actual operation. Based on the concept of a rolling horizon, a state transition prediction path is constructed for multiple future time steps. The cumulative impact of the control sequence on the dynamic changes in the material level over the future time period is evaluated, and the original control sequence is optimized to enhance foresight and adaptability in dynamic operation. For example, if a rapid decline in the material level is detected during the prediction process, the dynamic programming module preemptively increases the coal feed rate or adjusts the gate opening to mitigate hysteresis effects and maintain the material level within the confidence interval. The resulting optimized control parameter sequence satisfies both the optimal solution conditions and the dynamic response. Disturbance detection and compensation are then performed based on the optimized control parameter sequence and the current load conditions. The disturbance detection mechanism monitors the rate of change of material level, the deviation between the control output and the system response in real time, identifies sudden disturbances by setting a disturbance discrimination threshold, and performs parameter compensation processing after confirming the existence of a disturbance. For example, it restores its response path based on the history of similar disturbances, or adjusts the control instructions through a preset compensation matrix to enhance the system's dynamic response to emergencies. At the same time, to meet the needs of switching between various operating modes, a control parameter template library is established, pre-defining the optimal control instruction structure for different power generation loads, coal quality grades, start-up and shutdown phases, and maintenance modes. After real-time judgment of the current system operating conditions, it switches to the corresponding template. By superimposing the templates, it corrects the errors caused by the disturbance, forming a stable and efficient online control strategy, and ultimately generating a set of coal feeder control parameter sequences for the current actual operating conditions.

[0032] Reference Figure 2 This embodiment provides a millimeter wave radar material level monitoring device, including: Acquisition unit 1 is used to install multiple millimeter wave radar sensors on the top of the coal feeder silo, collect a reflection point cloud data set, and perform Euclidean clustering on the reflection point cloud data set to obtain a valid point cloud cluster set; Screening unit 2, used to perform multi-level information screening on the valid point cloud cluster set to obtain the valid material level cluster set of each millimeter wave radar sensor; Analysis unit 3, used to perform signal detection intersection-over-union ratio and reliability analysis on the effective material level cluster set of each millimeter wave radar sensor to generate fused material level data; Dynamic tracking unit 4 is used to input the fused material level data into the multi-model interaction and unscented Kalman filter model to perform dynamic material level tracking, and obtain a time series material level curve and a prediction confidence interval; The calculation unit 5 is used to perform iterative second-order cone programming calculation based on the time-series material level curve and the prediction confidence interval to obtain the coal feeder control parameter sequence.

[0033] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0034] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, memory, display screen, input device, network interface and database connected via a system bus. The processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above method is implemented.

[0035] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0036] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which implements the above-described method when executed by a processor. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0037] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware using a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media provided herein and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM.

[0038] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0039] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A millimeter wave radar material level monitoring method, characterized in that: The following steps are involved: Multiple millimeter-wave radar sensors are installed on the top of the coal feeder silo to collect a reflection point cloud dataset, and Euclidean clustering is performed on the reflection point cloud dataset to obtain a valid point cloud cluster set; Performing multi-level information screening on the effective point cloud cluster set to obtain an effective material level cluster set of each millimeter wave radar sensor; Performing signal detection intersection-over-union ratio and reliability analysis on the effective material level cluster sets of each millimeter-wave radar sensor to generate fused material level data; Inputting the fused material level data into the multi-model interaction and unscented Kalman filter model to perform dynamic material level tracking, and obtaining a time series material level curve and a prediction confidence interval; An iterative second-order cone programming calculation is performed based on the time-series material level curve and the prediction confidence interval to obtain a coal feeder control parameter sequence.

2. The millimeter wave radar material level monitoring method according to claim 1, characterized in that: The method includes installing multiple millimeter-wave radar sensors on the top of the coal feeder silo to collect a reflection point cloud dataset, and performing Euclidean clustering on the reflection point cloud dataset to obtain a valid point cloud cluster set, including: Multiple millimeter-wave radar sensors are installed on the top of the coal feeder silo to form full coverage of the silo space, and the millimeter-wave radar sensors are parameterized to obtain data sampling parameters; Connecting each millimeter-wave radar sensor to an independent data acquisition unit, wherein the data acquisition unit includes a signal conditioning module, an analog-to-digital conversion module, and a data preprocessing module; Acquire a radar echo signal according to the data sampling parameters, amplify and filter the radar echo signal through the signal conditioning module to obtain an amplified echo signal, and convert the amplified echo signal into a digital signal through the analog-to-digital conversion module; Inputting the digital signal into the data preprocessing module to perform range-Doppler processing to obtain initial point cloud data, and transmitting the initial point cloud data to the central processing unit for time synchronization to generate a reflection point cloud data set; Euclidean clustering is performed on the reflection point cloud dataset to obtain a valid point cloud cluster set.

3. The millimeter wave radar material level monitoring method according to claim 2, characterized in that: The performing Euclidean clustering on the reflection point cloud dataset to obtain a valid point cloud cluster set includes: Filtering outliers on the reflected point cloud data set to obtain point cloud data from which the outliers have been filtered out, and performing coordinate transformation on the point cloud data from which the outliers have been filtered out to obtain a preprocessed point cloud in a unified coordinate system; Extracting local point cloud density features from the preprocessed point cloud in the unified coordinate system, dynamically adjusting the clustering radius according to the local point cloud density features and establishing an adaptive clustering radius mechanism to obtain a clustering parameter set; Executing a dual-threshold seed point selection strategy based on the clustering parameter set, respectively calculating a neighborhood density value and a signal strength value of each candidate point, and selecting only points whose neighborhood density value and signal strength value simultaneously meet the threshold conditions as valid seed points; Hierarchical clustering is performed on the effective seed points to obtain an initial cluster set, and cluster validity verification and boundary optimization are performed on the initial cluster set to obtain a valid point cloud cluster set.

4. The millimeter wave radar material level monitoring method according to claim 1, characterized in that: The multi-level information screening of the effective point cloud cluster set to obtain the effective material level cluster set of each millimeter wave radar sensor includes: Establishing a coal feeder spatial reference model, wherein the coal feeder spatial reference model includes a silo inner wall boundary, a mechanical structure boundary, and a normal material level activity area; Performing spatial matching on each cluster in the valid point cloud cluster set and the coal feeder spatial reference model to obtain a spatial distance feature set; Performing spatial constraint screening based on the spatial distance feature set, retaining clusters located in the normal material level activity area, and removing reflection points that are obviously deviated from the normal material level activity area, to obtain a spatial screening result set; Performing reflection characteristic constraint screening on the spatial screening result set to obtain a reflection characteristic screening result set, and comparing the reflection characteristic screening result set with historical frame data, performing time consistency constraint screening to obtain a time consistency screening result set; The screening threshold is dynamically adjusted according to the current operating state of the coal feeder, and the working condition adaptation processing is performed on the time consistency screening result set to obtain the effective material level cluster set of each millimeter wave radar sensor.

5. The millimeter wave radar material level monitoring method according to claim 1, characterized in that: The performing signal detection intersection-and-union ratio and reliability analysis on the effective material level cluster sets of each millimeter wave radar sensor to generate fused material level data includes: The effective material level clustering set of each millimeter-wave radar sensor is projected onto the unified grid model to obtain the initial material level surface set; Calculating signal detection intersection-and-union ratios for any two surfaces in the initial material level surface set to form an intersection-and-union ratio matrix; An adaptive consistency threshold is set based on the intersection-in-union ratio matrix, and a conflict resolution mechanism is triggered when the intersection-in-union ratio of any pair of sensors is lower than the threshold to obtain a conflict resolution result; Assigning a reliability weight to each of the millimeter-wave radar sensors to obtain a sensor weight set; Based on the sensor weight set and the conflict processing result, performing a weighted fusion calculation on the different height value detection results of multiple sensors at the same grid point to obtain a fused height value and a fusion confidence; The fused height value and the fused confidence are subjected to spatial smoothing and outlier elimination post-processing, and detection blind spots are filled by spatial continuity constraints and historical data to obtain fused material level data.

6. The millimeter wave radar material level monitoring method according to claim 1, characterized in that: The fused material level data is input into the multi-model interaction and unscented Kalman filter model to perform dynamic material level tracking to obtain a time series material level curve and a prediction confidence interval, including: Extracting material level feature point data from the fused material level data; A time threshold parameter is set for the material level feature point data, where the time threshold parameter is twice the current material level change rate, and a joint probability data association algorithm is used to match the previous and next frame data to obtain an association probability matrix; Based on the association probability matrix, a uniform speed model of a stable coal feeding state, a uniform acceleration model of a start-stop transition state, and a random acceleration model of a fluctuating state are constructed, and the uniform speed model, the uniform acceleration model, and the random acceleration model are used as a dynamic model set; Introducing an interactive multi-model mechanism into the dynamic model set and estimating a model transition probability matrix online to determine a model combination result including a current optimal model combination; Inputting the model combination result into an unscented Kalman filter, performing material level state estimation and nonlinear system propagation, and obtaining a filtering result; A time series material level curve is formed according to the filtering result, and the process noise covariance matrix and the measurement noise covariance matrix are dynamically adjusted according to the change of working conditions, and the prediction confidence interval is calculated to obtain the time series material level curve and the prediction confidence interval.

7. The millimeter wave radar material level monitoring method according to claim 1, characterized in that: The iterative second-order cone programming calculation is performed based on the time series material level curve and the prediction confidence interval to obtain the coal feeder control parameter sequence, including: The feeder's input and output functions are expressed as nonlinear functions of the feed rate, feeder speed, and gate opening, and the material level change rate model is obtained. Setting a control parameter constraint set for the coal feeder control parameters, wherein the control parameter constraint set includes upper and lower limits of the value range of the coal feeding rate, the coal feeder speed and the gate opening, and a maximum change rate limit; Constructing an optimization objective function including a material level deviation term, a coal feeding rate deviation term, a rotation speed deviation term, and a gate opening deviation term based on the time series material level curve and the prediction confidence interval; The optimization objective function is converted into a quadratic constrained quadratic programming problem, the non-convex optimization problem is decomposed into multiple convex sub-problems by an iterative second-order cone programming algorithm, the convergence condition and the iteration upper limit are set, and the initial control parameter solution is obtained; The actual operation constraints of the coal feeder are introduced into the initial control parameter solution, and the parameters are optimized and adjusted through dynamic programming to obtain an optimized control parameter sequence; Based on the optimized control parameter sequence and the current load condition, disturbance detection and compensation processing is performed, and online strategy switching is performed in combination with a preset control parameter template to generate a coal feeder control parameter sequence.

8. A millimeter wave radar material level monitoring device, characterized in that: For implementing the steps of the millimeter wave radar material level monitoring method according to any one of claims 1 to 7, the millimeter wave radar material level monitoring device comprises: an acquisition unit, configured to install multiple millimeter-wave radar sensors on the top of the coal feeder silo, collect a reflection point cloud dataset, and perform Euclidean clustering on the reflection point cloud dataset to obtain a valid point cloud cluster set; A screening unit, configured to perform multi-level information screening on the effective point cloud cluster set to obtain an effective material level cluster set of each millimeter wave radar sensor; An analysis unit, configured to perform signal detection intersection-over-union ratio and reliability analysis on the effective material level cluster sets of each millimeter-wave radar sensor to generate fused material level data; A dynamic tracking unit is used to input the fused material level data into a multi-model interaction and unscented Kalman filter model to perform dynamic material level tracking, thereby obtaining a time series material level curve and a prediction confidence interval; The calculation unit is used to perform iterative second-order cone programming calculation based on the time-series material level curve and the prediction confidence interval to obtain a coal feeder control parameter sequence.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the millimeter wave radar material level monitoring method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the millimeter wave radar material level monitoring method according to any one of claims 1 to 7 are implemented.

Images