Slag quantity precision detection system and method based on binocular vision

Abstract

The application provides a kind of accurate detection system and method of slag quantity based on binocular vision, system includes dry slag machine automatic speed measurement module, calculates the real-time running speed of dry slag machine conveying chain plate;Slag quantity detection module, the time sequence alignment of slag layer three-dimensional point cloud reconstructed by binocular stereo vision is carried out, and the volume of slag quantity is calculated;Interference automatic identification module is used to distinguish dry slag machine conveying chain plate structural member and coal slag, and provides higher quality slag layer three-dimensional point cloud data for slag quantity detection module.The application synchronously collects image by binocular stereo vision, reconstructs high-density three-dimensional point cloud of slag layer by combining improved semi-global matching stereo matching algorithm;Further, the volume of accumulated coal slag is accurately calculated by surface gridding based on Delaunay triangulation and scanning line integration method;And by associating the density parameters in coal database, the volume value is converted into mass data, which realizes accurate and quantitative detection of slag quantity from three-dimensional form to quality index.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, specifically to a binocular vision-based system and method for accurate slag quantity detection, and more particularly to a binocular vision-based system and method for accurate slag quantity detection with automatic speed and interference sensing. Background Technology

[0002] In thermal power plants, the ash removal machine is a critical piece of equipment at the bottom of the boiler, responsible for conveying the solid residue (coal ash) that falls after coal combustion. The amount of coal ash accumulated directly affects the safe operation of the ash removal machine: excessive accumulation can cause chain jamming, equipment overload, or even shutdown accidents, while fluctuations in ash volume may also reflect coking or abnormal combustion problems inside the boiler. Traditional monitoring methods relying on manual inspections or simple sensors are insufficient for real-time and accurate quantification. Operationally, there is an urgent need to dynamically grasp ash volume data through automated detection methods to support equipment early warning, optimize ash removal strategies, and correlate with boiler combustion status analysis, thus fulfilling the core requirement of intelligent operation and maintenance of power plants.

[0003] The amount of slag is a key indicator for evaluating boiler combustion efficiency: excessive slag indicates incomplete fuel combustion (increased carbon loss), while changes in slag quality (such as melting and agglomeration) may reflect uneven temperature distribution or unreasonable air distribution within the furnace. By accurately detecting the amount and morphology of slag, power plants can optimize combustion control parameters (such as primary air volume and pulverized coal fineness) in real time to improve combustion efficiency. Simultaneously, abnormal increases in slag can provide early warning of the risk of coking on boiler heating surfaces, preventing a decrease in heat transfer efficiency due to scaling. This provides data support for developing soot blowing strategies and preventative maintenance, ensuring economical unit operation from the outset.

[0004] Existing detection technologies have shortcomings in terms of automation and accuracy. Manual inspection relies on experience and visual inspection, which is highly subjective, produces discrete data, and cannot provide continuous monitoring. Mechanical sensors (such as weighing instruments and level gauges) are easily affected by vibrations from dry slag machines, high temperatures, and dust, resulting in significant accuracy drift, short lifespan, and high maintenance costs. Monocular vision solutions struggle to distinguish between the slag layer and background under complex operating conditions (such as glare from conveyor belts and steam obstruction), and the lack of depth information leads to excessive deviations in volume calculations. These shortcomings hinder the realization of combustion optimization closed-loop control.

[0005] This invention provides a binocular vision-based system and method for precise slag quantity detection, enabling automatic perception of speed and interference. The binocular vision technology simultaneously acquires images using two cameras and reconstructs a three-dimensional point cloud of the slag layer using a stereo matching algorithm, achieving non-contact, precise detection. This solution provides power plants with a low-maintenance, high-information-density core data source for combustion optimization. Summary of the Invention

[0006] In view of the deficiencies in the prior art, the purpose of this invention is to provide a system and method for accurate detection of slag quantity based on binocular vision.

[0007] A binocular vision-based precision slag quantity detection system according to the present invention includes: The automatic speed measurement module for the dry slag machine is used to calculate the real-time running speed of the conveyor chain plate of the dry slag machine; The slag volume detection module performs time-series alignment of the slag layer three-dimensional point cloud reconstructed by binocular stereo vision based on the real-time running speed of the conveyor chain plate of the dry slag machine, and calculates the slag volume. An automatic interference identification module is used to distinguish between the conveyor chain plate structure of the dry slag machine and coal slag, and to provide higher quality three-dimensional point cloud data of the slag layer for the slag quantity detection module.

[0008] Preferably, the automatic speed measurement module of the dry slag machine is based on a high-speed industrial binocular camera and AI algorithm. It calculates the actual running speed of the conveyor chain plate of the dry slag machine by analyzing the pixel changes of target features in a continuous frame image sequence.

[0009] Preferably, the slag quantity detection module accurately calculates the volume of coal slag within a specified sensing range of the dry slag machine based on binocular vision, and obtains the mass through density conversion; it reconstructs the three-dimensional point cloud of the slag layer using binocular stereo vision, and combines point cloud filtering and grid integration to achieve volume measurement.

[0010] Preferably, the automatic interference identification module includes: The image preprocessing submodule is used to enhance the quality of the acquired raw images; The feature extraction submodule uses various image processing algorithms to extract key features of the chain plate partition, including geometric features, texture features, and motion features; The deep learning inference submodule uses an optimized neural network model for real-time inference computation.

[0011] Preferably, the automatic speed measurement module of the dry slag machine applies histogram equalization and adaptive threshold segmentation to improve the contrast of feature points, as expressed by the formula:

[0012] in, The original image is represented by (x, y), where x represents the horizontal coordinate and y represents the vertical coordinate. CLAHE is contrast-limited adaptive histogram equalization, α is the weighting coefficient, and median is the median filter to suppress dust noise. Feature point initialization includes pre-marking points with a certain contrast on the chain plate surface, or automatically extracting natural features; the Shi-Tomasi corner detection algorithm is used, with the following formula:

[0013] Where λ1 and λ2 are the eigenvalues ​​of the Harris matrix, and τ is the threshold.

[0014] Preferably, the slag quantity detection module uses a binocular camera for image acquisition, and the camera calibration is completed using the Zhang Zhengyou calibration method to calculate the intrinsic parameter matrix K and distortion coefficients.

[0015] Where u represents the x-coordinate of the pixel in the image coordinate system, v represents the y-coordinate of the pixel in the image coordinate system, and together they represent the pixel position of the 3D point projected onto the 2D image plane; Xw represents the X-axis coordinate of the 3D point in the world coordinate system, Yw represents the Y-axis coordinate of the 3D point in the world coordinate system, Zw represents the Z-axis coordinate of the 3D point in the world coordinate system, and together they describe the position of the object in the real 3D space; fx and fy are the focal lengths, cx and cy are the principal point coordinates, and R and t are the rotation and translation matrices; calibration uses a checkerboard template, and the reprojection error is <0.1 pixels; The Bouguet algorithm is applied for stereo correction, aligning the epipolar lines of the image, simplifying the matching process, and outputting the corrected image. and .

[0016] Preferably, the process of stereo matching and point cloud generation performed by the slag quantity detection module includes: Using an improved SGM algorithm, semi-global matching is employed, reducing the cost volume. The calculation is as follows:

[0017] Where p is the number of pixels, d is the disparity, and R... p Let λ be the neighborhood window, λ be the gradient weight, and q be the gradient weight. d For matching points; Point cloud reconstruction, parallax map converted into point cloud :

[0018] Where B is the baseline distance and f is the focal length; Point cloud filtering, applying statistical outlier filtering to remove noise, if points The following conditions must be met:

[0019] Then remove it, where, Let k be the nearest neighbor set, and k be the size of the nearest neighbor set. Let be a point in the point set, and μ and σ be the mean and standard deviation of the distance.

[0020] Preferably, the process of slag layer segmentation and volume integration performed by the slag quantity detection module includes: Dynamic background removal, pre-stored clean point cloud state Real-time point cloud registration using ICP algorithm :

[0021] Differential extraction of slag layer point cloud ; Slag layer meshing: Surface mesh S is constructed using Delaunay triangulation.

[0022] Volume calculation, based on grid volume integration using the scan line method:

[0023] Where M is the number of grids, A k Let the projected area of ​​the triangular facet be . The average depth of the vertices.

[0024] Preferably, multi-frame image fusion technology is used to accumulate useful information by analyzing the temporal relationship of consecutive multi-frame images, while suppressing the influence of random noise.

[0025] A method for accurate slag quantity detection based on binocular vision according to the present invention includes: Step S1: Calculate the real-time running speed of the conveyor chain plate of the dry slag machine; Step S2: Based on the real-time running speed of the conveyor chain plate of the dry slag machine, perform time-series alignment on the 3D point cloud of the slag layer reconstructed by binocular stereo vision, and calculate the slag volume. Step S3: Distinguish between the conveyor chain plate structure of the dry slag machine and the coal slag, and provide higher quality 3D point cloud data of the slag layer for step S2.

[0026] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention acquires images synchronously using binocular stereo vision, and reconstructs a high-density three-dimensional point cloud of slag layer by combining an improved semi-global matching (SGM) stereo matching algorithm; then, it uses surface meshing based on Delaunay triangulation and scan line integration method to accurately calculate the volume of accumulated coal slag; and by associating density parameters in the coal type database, it converts the volume value into mass data, realizing accurate and quantitative detection of slag quantity from three-dimensional morphology to quality indicators.

[0027] 2. This system effectively suppresses typical interferences by integrating image processing and multi-sensor data. At the perception level, the hardware design, featuring a dust cover and active cooling, is adapted to high-temperature environments. At the algorithm level, image enhancement is achieved through adaptive histogram equalization, wavelet denoising, and improved Wiener filtering. Statistical outlier filtering (SORF) is used to remove point cloud noise, and algorithms such as RANSAC and Kalman filtering are employed to handle occlusion and anomalies in feature point tracking. This ensures the stable operation and data reliability of the detection system under complex conditions such as dust, steam, metal reflection, and chain vibration.

[0028] 3. The system of this invention possesses high-speed processing capabilities. The processing delay of the dry slag machine speed detection module is less than 10 milliseconds (ms), and the slag volume calculation update frequency can reach 10 Hz. By calculating the chain plate speed in real time and performing time-series alignment, continuous and dynamic measurement of the slag volume on the moving chain plate is achieved. The detection results are output in real time in a time-series format, providing minute-level or even second-level dynamic slag volume trend data for boiler combustion optimization, supporting real-time early warning and closed-loop control. Attached Figure Description

[0029] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a structural diagram of the present invention.

[0030] Figure 2 This is a schematic diagram of the automatic speed measurement module for the dry slag machine of the present invention.

[0031] Figure 3 This is a schematic diagram of the slag quantity detection module of the present invention.

[0032] Figure 4 This is a schematic diagram of the interference automatic identification module of the present invention.

[0033] Figure 5 This is an external view of the detection device according to an embodiment of the present invention.

[0034] Figure 6 This is a flowchart of the method of the present invention. Detailed Implementation

[0035] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0036] like Figure 1As shown, a binocular vision-based precision slag quantity detection system includes: The automatic speed measurement module for the dry slag machine is used to calculate the real-time running speed of the conveyor chain plate of the dry slag machine; The slag volume detection module performs time-series alignment of the slag layer three-dimensional point cloud reconstructed by binocular stereo vision based on the real-time running speed of the conveyor chain plate of the dry slag machine, and calculates the slag volume. An automatic interference identification module is used to distinguish between the conveyor chain plate structure of the dry slag machine and coal slag, and to provide higher quality three-dimensional point cloud data of the slag layer for the slag quantity detection module.

[0037] The specific contents of each module are as follows: I. Automatic Speed ​​Measurement Module for Dry Slag Machine like Figure 2 As shown, the core objective of this module is to achieve accurate and automatic detection of the conveyor chain speed of the dry slag machine. The technical solution is based on a high-speed industrial binocular camera and AI algorithms. By analyzing pixel changes in target features (such as chain nodes, scraper edges, or slag block displacement) in a continuous frame image sequence, the actual operating speed of the chain is calculated. The software algorithm is based on the PyTorch framework and includes feature point tracking, motion blur correction, and Kalman filter fusion to ensure high accuracy under dust interference and chain vibration. The implementation steps of this module are divided into four stages: image acquisition, feature extraction, displacement calculation, and speed fusion output. Each stage includes detailed algorithms and error control mechanisms.

[0038] (1) Image acquisition and preprocessing Image acquisition is fundamental to speed detection, requiring high-speed, high-contrast image sequences to capture the details of the chain conveyor's movement. In hardware deployment, the camera is mounted axially perpendicular to the chain conveyor's direction of movement, 0.5-1.0 meters from the chain surface, to cover a 2-3 meter wide sensing area. The camera is equipped with a dust cover and an active cooling system to withstand the high-temperature environment of the dry slag machine. The image acquisition frequency is set to match the chain conveyor's maximum speed to avoid motion blur. Preprocessing steps include: Image enhancement: Histogram equalization and adaptive thresholding are applied to improve feature point contrast. The formula is expressed as:

[0039] in, For the original image, For the enhanced image, (x, y) represents the coordinate position of a pixel in the image, where x represents the horizontal (column) coordinate and y represents the vertical (row) coordinate. CLAHE is contrast-limited adaptive histogram equalization, α is the weighting coefficient, and median is median filtering to suppress dust noise.

[0040] Feature point initialization: High-contrast markers are pre-made on the chain plate surface, or natural features (such as chain node corners) are automatically extracted. The Shi-Tomasi corner detection algorithm is used, with the following formula:

[0041] Where λ1 and λ2 are the eigenvalues ​​of the Harris matrix, and τ is the threshold (empirical value 0.01) to ensure stable feature point extraction in low-texture environments.

[0042] (2) Feature point tracking and displacement calculation This step uses AI algorithms to analyze the pixel displacement of feature points between consecutive frames and calculate the chain plate motion vector. The core employs an improved optical flow method combined with RANSAC robustness optimization to handle dust occlusion and changes in illumination.

[0043] Optical flow implementation: The Pyramidal Lucas-Kanade algorithm is used to calculate the displacement vectors of feature points between adjacent frames. The formula is:

[0044] in, and For image gradient, The time gradient is represented by the summation region, which is a 15×15 pixel neighborhood window of the feature points. The algorithm runs at multiple scales within the pyramid, iterating from low-resolution layers to improve accuracy in large displacement scenes.

[0045] Motion fuzzy correction: Vibration displacement based on IMU output Image blur is corrected using deconvolution. The point spread function (PSF) is modeled as follows:

[0046] in, This represents the amplitude of the vibration displacement.

[0047] The corrected image is obtained iteratively using the Richardson-Lucy algorithm:

[0048] Reduce displacement calculation errors, among which Indicates the number of iterations. Represents a blurred image. This indicates the corrected image.

[0049] Outlier removal: The RANSAC algorithm is used to fit a motion consistency model, removing outlier vectors (such as incorrect matches caused by dust occlusion). The RANSAC iterative process is as follows: randomly sample a subset of feature points, fit an affine transformation model, calculate the inlier rate, and retain the set of largest inliers. The output is the set of effective displacement vectors. average displacement ,in For effective displacement vector set Size.

[0050] (3) Speed ​​calculation and fusion output Velocity calculation integrates visual displacement and IMU data, and achieves dynamic compensation through a state-space model.

[0051] Preliminary speed calculation: based on frame rate and average pixel displacement Calculate the chain speed ,in This refers to the physical size of the pixel (calibrated value, such as 0.1mm / pixel).

[0052] Multi-sensor fusion: Establishing a Kalman filter state-space model, state variables (Velocity, acceleration), observed values ,in To calculate the chain speed, This represents the velocity calculated from IMU (Inertial Measurement Unit) data. The state equation and observation equation are:

[0053] Where A is the state transition matrix (empirical value [1, Δt; 0, 1]), H is the observation matrix [1, 0], and w k and v k For process noise and observation noise (covariance matrix calibrated based on field data). Kalman prediction-update loop output fusion speed. .

[0054] Real-time output and interface: Speed ​​data is output to the system bus in timestamp format, supporting timing alignment of the sludge module. The software is implemented in Python, with a processing latency of <10ms.

[0055] II. Slag Quantity Detection Module like Figure 3As shown, the core objective of this module is to accurately calculate the volume of coal slag within a specified sensing range of the dry slag machine based on binocular vision, and obtain the mass through density conversion to support boiler combustion efficiency analysis. The technical solution utilizes binocular stereo vision to reconstruct a 3D point cloud of the slag layer, combining point cloud filtering and mesh integration to achieve volume measurement. The hardware system includes a parallel baseline binocular camera. The software algorithm covers stereo matching, point cloud reconstruction, and volume integration, with density values ​​dynamically retrieved from a coal type database.

[0056] The implementation process is divided into four stages: image acquisition and calibration, stereo matching and point cloud generation, slag layer segmentation and volume integration, and mass conversion. The processing flow needs to run on an edge server (such as an Intel Xeon) to output the slag quantity curve in real time. The technical details are described step by step below.

[0057] (1) Binocular image acquisition and system calibration In the hardware deployment, a binocular camera is mounted above the slag dryer, covering a 1m × 2m sensing area. The cameras are triggered synchronously to avoid motion distortion. Calibration is fundamental to volumetric accuracy. Camera calibration: Zhang Zhengyou's calibration method is used to calculate the intrinsic parameter matrix K and distortion coefficients. The formula is:

[0058] Where u represents the x-coordinate (horizontal position) of a pixel in the image coordinate system, and v represents the y-coordinate (vertical position) of a pixel in the image coordinate system. Together, they represent the pixel position after the 3D point is projected onto the 2D image plane. Xw represents the X-axis coordinate of the 3D point in the world coordinate system, Yw represents the Y-axis coordinate of the 3D point in the world coordinate system, and Zw represents the Z-axis coordinate of the 3D point in the world coordinate system. Together, they describe the position of the object in the real 3D space. fx and fy are the focal lengths, cx and cy are the principal point coordinates, and R and t are the rotation and translation matrices. Calibration uses a checkerboard template, and the reprojection error is <0.1 pixels.

[0059] Stereo correction: The Bouguet algorithm is applied to align the epipolar lines of the image, simplifying the matching process. Output corrected image. and .

[0060] (2) Stereo matching and point cloud generation This step calculates the disparity map through stereo matching and converts it into a 3D point cloud to handle the irregular shape and texture loss of the slag layer.

[0061] Improved SGM algorithm: Employs semi-global matching, cost volume The calculation is as follows:

[0062] Where p is the number of pixels, d is the disparity, and R...p Let q be the neighborhood window, λ be the gradient weight (0.8), and q be the gradient weight. d For the matching point, The intensity value of the left image. The intensity value of the corresponding point in the right image. The gradient of the left image. The gradient is the corresponding point in the right image. Multi-path aggregation optimizes disparity smoothness.

[0063] Point cloud reconstruction: Conversion of disparity map into point cloud :

[0064] Where B is the baseline distance and f is the focal length. The point cloud density is approximately 10^5 points / m².

[0065] Point cloud filtering: Applying statistical outlier filtering (SORF) to remove noise, if the point cloud... The following conditions must be met:

[0066] Then remove it, where, For the nearest neighbor set, Size of the point set Let μ be a point in the point set, and σ be the mean and standard deviation of the distance. The signal-to-noise ratio of the point cloud is improved by >20dB after filtering.

[0067] (3) Slag layer segmentation and volume integration The key to segmenting the slag layer point cloud and calculating its volume lies in distinguishing between the slag and the chain plate background.

[0068] Dynamic background removal: Pre-stored clean point cloud state The point cloud captured at time t is registered using the ICP algorithm. :

[0069] Differential extraction of slag layer point cloud .

[0070] Slag layer meshing: Surface mesh S is constructed using Delaunay triangulation.

[0071] Volume calculation: Integrating the mesh volume based on the scan line method:

[0072] Where M is the number of grids, A k Let the projected area of ​​the triangular facet be . The average depth of the vertices is given. The volume update frequency is 10Hz, which meets the requirements for dynamic monitoring.

[0073] (4) Mass conversion and output Mass calculation combines volume and density: Density retrieval: ρ is obtained from the coal type database (e.g., bituminous coal ρ=1.4g / cm³), and updated in real time based on combustion parameters.

[0074] Quality output: The data is pushed to the DCS system in time-series format.

[0075] Error control: Volumetric error is adaptively optimized through mesh density; density error is compensated through laboratory calibration.

[0076] III. Automatic Interference Identification Module like Figure 4 As shown, this module mainly includes image preprocessing, feature extraction, and deep learning inference. The image preprocessing step is responsible for enhancing the quality of the acquired raw images, including noise reduction, contrast adjustment, and geometric correction, providing high-quality input data for subsequent processing. The feature extraction step uses various image processing algorithms to extract key features of the chain plate partition, including geometric features, texture features, and motion features. The deep learning inference step is the core of the system, employing an optimized neural network model for real-time inference calculations.

[0077] (1) Image preprocessing techniques (a) Image quality enhancement In the complex industrial environment of thermal power plants, image quality is often affected by a variety of factors, including dust pollution, changes in lighting, and thermal radiation interference. To address these issues, the system employs a comprehensive set of image quality enhancement technologies.

[0078] Adaptive histogram equalization improves image contrast, making the edge features of the chain plate partition clearer.

[0079] A wavelet transform-based denoising algorithm is used to remove high-frequency noise from the image while preserving the integrity of edge information. For image blurring caused by dust, an improved Wiener filtering algorithm is employed for image restoration.

[0080] To address image distortion caused by thermal radiation, a physical model-based correction algorithm was developed, which performs geometric correction on the image by analyzing the temperature distribution.

[0081] To cope with changes in lighting conditions, the system also integrates automatic exposure control and white balance adjustment functions to ensure stable image quality under different times and weather conditions.

[0082] The combined application of these preprocessing techniques significantly improves the quality of the original image, laying a solid foundation for subsequent feature extraction and recognition analysis.

[0083] (b) Geometric correction and registration Due to mechanical vibrations during the operation of the slag dryer and potential slight displacements in the camera's mounting position, the original images often suffer from geometric distortion. The system employs a geometric correction technique based on feature point matching to address this issue.

[0084] Identify stable feature points in the image, such as fixed marks on chain plates or edge features of the equipment, as reference benchmarks for correction.

[0085] The camera's motion parameters are estimated by calculating the displacement of feature points between consecutive frames.

[0086] Based on these parameters, the system uses a perspective transformation model to perform geometric correction on the image, eliminating image distortion caused by camera motion.

[0087] To improve correction accuracy, the system also integrates a sub-pixel-level feature point detection algorithm, which can detect minute geometric changes.

[0088] Considering the thermal expansion effect of the chain plate itself, a temperature compensation model was established for the system, and the geometric correction parameters were dynamically adjusted according to the ambient temperature.

[0089] The application of these geometric correction techniques ensures spatial consistency between consecutive frames, providing a reliable foundation for subsequent motion analysis and feature matching.

[0090] (c) Multi-frame image fusion To improve the image signal-to-noise ratio and enhance the visibility of key features, the system employs multi-frame image fusion technology. This technology accumulates useful information by analyzing the temporal relationship of consecutive frames, while suppressing the influence of random noise.

[0091] The fusion algorithm first performs motion estimation on consecutive frames to determine the correspondence between images.

[0092] Then, a weighted average method is used to fuse the corresponding pixels, with the weights allocated based on the pixel's signal-to-noise ratio and motion consistency. For moving regions, the system employs an adaptive fusion strategy to reduce the impact of motion blur. For static regions, fusion of more frames is used to achieve better noise suppression.

[0093] In addition, the system takes into account the differences in importance of different areas and uses more refined fusion parameters for key areas containing chain plate partitions.

[0094] Multi-frame fusion technology not only significantly improves image quality but also enhances the stability and detectability of chain plate partition features, providing more reliable input data for subsequent deep learning analysis. This utilization of temporal information also provides the system with richer feature dimensions, contributing to improved final recognition accuracy.

[0095] (2) Feature extraction and analysis (a) Edge contour feature extraction Edge contour features are one of the important visual characteristics that distinguish chain conveyor belt partitions from coal slag. Chain conveyor belt partitions typically have regular geometric shapes and clear boundaries, while coal slag exhibits irregular contour features. The system employs multi-scale edge detection technology to extract these features.

[0096] An improved Canny edge detection algorithm is used to detect all edge information in an image. This algorithm can maintain stable detection performance under different contrast conditions through adaptive threshold selection.

[0097] The geometric characteristics of the edges were analyzed using a gradient orientation histogram method, and the edge distribution in different directions was statistically analyzed. Due to its regular geometric structure, the chain plate partition exhibits obvious edge response peaks in specific directions, while the edge distribution of coal slag is relatively uniform.

[0098] The system also analyzes the continuity and smoothness characteristics of the edges, and quantifies the regularity of the contour by calculating parameters such as the length of the edge line segments and the curvature change.

[0099] To improve the robustness of the features, the system employs a multi-scale analysis method to extract edge features at different spatial resolutions, and then performs feature fusion.

[0100] This multi-level edge analysis method can effectively capture the geometric features of the chain plate partition, providing an important basis for subsequent classification and recognition.

[0101] (b) Texture feature analysis Texture features are another important identification criterion. The surface of the chain plate usually has a relatively uniform metallic texture, while coal slag exhibits an irregular granular texture. The system uses a texture analysis method based on gray-level co-occurrence matrix to extract these features.

[0102] By calculating the gray-level co-occurrence matrix at different directions and distances, the system can obtain statistical characteristics of the texture, such as contrast, uniformity, correlation, and entropy. These characteristics can effectively quantify the differences in texture properties across different regions.

[0103] The system also employs a wavelet transform-based texture analysis method, which decomposes the image into different frequency sub-bands and analyzes the energy distribution characteristics of each sub-band. Wavelet texture features are particularly effective in capturing the directionality and scale characteristics of textures.

[0104] To further enhance the ability to discriminate texture features, the system also integrates a local binary pattern algorithm, which can effectively describe the microstructural features of local textures.

[0105] By integrating multiple texture analysis methods, a high-dimensional texture feature vector was constructed, which can comprehensively reflect the texture characteristics of different material surfaces. These texture features, combined with edge features, provide rich input information for the deep learning model, significantly improving the accuracy and stability of recognition.

[0106] (c) Motion pattern recognition Motion characteristics are another important dimension for distinguishing chain conveyor belts from coal slag, as the two differ significantly in their motion patterns. As a component of conveying equipment, chain conveyor belts exhibit regular motion trajectories and consistent speeds, while coal slag may display random motion or a relatively stationary state.

[0107] The system employs optical flow analysis to extract motion features, estimating the motion vector field by calculating pixel displacements between consecutive frames. For the chain plate partition region, the optical flow vector exhibits directional consistency and amplitude stability, while the optical flow vector in the slag region is relatively chaotic. The system quantifies the motion features by statistically analyzing the directional distribution and amplitude variations of the optical flow vector.

[0108] The system also employs a trajectory analysis-based method to track the motion trajectories of feature points over time, analyzing the smoothness and periodicity of the trajectories. The feature point trajectories of the chain plate partition typically exhibit good linearity, while the trajectories of coal slag may show irregular changes.

[0109] To address the occlusion problem in motion analysis, the system employs a robust feature point tracking algorithm that can maintain tracking continuity even under partial occlusion.

[0110] The extraction of motion features provides the system with temporal dimension information, which significantly enhances the discrimination ability of the recognition algorithm.

[0111] (3) Deep learning model design (a) Network architecture optimization To address the specific needs of chain plate partition recognition, the system employs a specially optimized deep convolutional neural network architecture. The network design fully considers the characteristics of industrial images and the requirements of real-time processing.

[0112] The backbone network adopts a residual connection structure, which can effectively alleviate the gradient vanishing problem in deep networks while maintaining high computational efficiency.

[0113] The network's input layer is designed with a multi-channel structure, including not only traditional RGB image channels but also dedicated feature channels such as edge feature maps, texture feature maps, and motion feature maps. This multimodal input design allows the network to utilize multiple types of feature information simultaneously.

[0114] In the middle layer of the network, an attention mechanism module is employed, which can automatically learn the importance weights of different features, highlighting the feature dimensions most helpful for the recognition task. The output layer of the network adopts a multi-branch design, which not only outputs the classification results, but also provides confidence assessment and region localization information.

[0115] To meet the real-time processing requirements in industrial environments, the network structure has been optimized for lightweight design. Computational complexity has been reduced through model compression and quantization techniques while maintaining recognition accuracy.

[0116] This customized network design ensures the system's efficiency and accuracy in practical applications.

[0117] (b) Training data construction High-quality training data is a key factor for the success of deep learning models. The system's training dataset construction process fully considers the actual operating conditions of thermal power plants.

[0118] Data acquisition covers various operating states, including normal operation, heavy load operation, start-up and shutdown processes, and other operating conditions. It also considers different environmental conditions, such as varying light intensities, dust concentrations, and temperatures.

[0119] To ensure the representativeness of the data, the data collection process covered different time periods, including daytime, nighttime, and different seasonal conditions.

[0120] To enhance the model's generalization ability, the system also employs data augmentation techniques, expanding the training dataset through methods such as rotation, scaling, brightness adjustment, and noise addition. Furthermore, a dedicated hard sample dataset was constructed, containing challenging recognition scenarios such as abnormal lighting, severe occlusion, and motion blur.

[0121] This comprehensive data construction strategy ensures that the model maintains stable performance under various complex conditions.

[0122] (c) Model training optimization The model training process employs a multi-stage optimization strategy to achieve optimal recognition performance.

[0123] In the initial training phase, a large learning rate is used for rapid convergence, followed by a gradual reduction of the learning rate for fine-tuning. The loss function design comprehensively considers multiple factors such as classification accuracy, boundary precision, and class balance, employing a combination of weighted cross-entropy loss and focus loss to effectively address the class imbalance problem.

[0124] Several regularization techniques were employed during training, including weight decay, batch normalization, and random deactivation, to prevent overfitting. To improve the model's robustness, adversarial training was also used, adding small adversarial perturbations to enhance its resistance to noise.

[0125] The training process employs a rigorous validation strategy, using an independent validation set to monitor model performance and an early stopping mechanism to prevent overfitting.

[0126] It also employs model ensemble technology, training multiple models with different initialization parameters and network structures, and obtaining the final recognition result through a voting mechanism.

[0127] like Figure 5 As shown, this ensemble approach significantly improves the system's accuracy and stability. The entire training process is accelerated through a distributed computing platform, ensuring both training efficiency and model quality.

[0128] like Figure 6 As shown, a method for accurate slag quantity detection based on binocular vision includes: Step S1: Calculate the real-time running speed of the conveyor chain plate of the dry slag machine; Step S2: Based on the real-time running speed of the conveyor chain plate of the dry slag machine, perform time-series alignment on the 3D point cloud of the slag layer reconstructed by binocular stereo vision, and calculate the slag volume. Step S3: Distinguish between the conveyor chain plate structure of the dry slag machine and the coal slag, and provide higher quality 3D point cloud data of the slag layer for step S2.

[0129] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0130] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A binocular vision-based system for precise slag quantity detection, characterized in that, include: The automatic speed measurement module for the dry slag machine is used to calculate the real-time running speed of the conveyor chain plate of the dry slag machine; The slag volume detection module performs time-series alignment of the slag layer three-dimensional point cloud reconstructed by binocular stereo vision based on the real-time running speed of the conveyor chain plate of the dry slag machine, and calculates the slag volume. An automatic interference identification module is used to distinguish between the conveyor chain plate structure of the dry slag machine and coal slag, and to provide higher quality three-dimensional point cloud data of the slag layer for the slag quantity detection module.

2. The binocular vision-based precision slag quantity detection system according to claim 1, characterized in that, The automatic speed measurement module of the dry slag machine is based on a high-speed industrial binocular camera and AI algorithm. It calculates the actual running speed of the conveyor chain plate of the dry slag machine by analyzing the pixel changes of target features in a continuous frame image sequence.

3. The binocular vision-based precision slag quantity detection system according to claim 1, characterized in that, The slag quantity detection module accurately calculates the volume of coal slag within the specified sensing range of the dry slag machine based on binocular vision, and obtains the mass through density conversion; it reconstructs the three-dimensional point cloud of the slag layer using binocular stereo vision, and combines point cloud filtering and grid integration to achieve volume measurement.

4. The binocular vision-based precision slag quantity detection system according to claim 1, characterized in that, The automatic interference identification module includes: The image preprocessing submodule is used to enhance the quality of the acquired raw images; The feature extraction submodule uses various image processing algorithms to extract key features of the chain plate partition, including geometric features, texture features, and motion features; The deep learning inference submodule uses an optimized neural network model for real-time inference computation.

5. The binocular vision-based precision slag quantity detection system according to claim 2, characterized in that, The automatic speed measurement module for the dry slag machine applies histogram equalization and adaptive threshold segmentation to improve the contrast of feature points, as expressed by the formula: in, The original image is represented by (x, y), where x represents the horizontal coordinate and y represents the vertical coordinate. CLAHE is contrast-limited adaptive histogram equalization, α is the weighting coefficient, and median is the median filter to suppress dust noise. Feature point initialization includes pre-marking points with a certain contrast on the chain plate surface, or automatically extracting natural features; the Shi-Tomasi corner detection algorithm is used, with the following formula: Where λ1 and λ2 are the eigenvalues ​​of the Harris matrix, and τ is the threshold.

6. The binocular vision-based precision slag quantity detection system according to claim 1, characterized in that, The slag quantity detection module uses a binocular camera for image acquisition, and the camera calibration is completed using the Zhang Zhengyou calibration method. The intrinsic parameter matrix K and distortion coefficients are calculated. Where u represents the x-coordinate of the pixel in the image coordinate system, v represents the y-coordinate of the pixel in the image coordinate system, and together they represent the pixel position of the 3D point projected onto the 2D image plane; Xw represents the X-axis coordinate of the 3D point in the world coordinate system, Yw represents the Y-axis coordinate of the 3D point in the world coordinate system, Zw represents the Z-axis coordinate of the 3D point in the world coordinate system, and together they describe the position of the object in the real 3D space; fx and fy are the focal lengths, cx and cy are the principal point coordinates, and R and t are the rotation and translation matrices; calibration uses a checkerboard template, and the reprojection error is <0.1 pixels; The Bouguet algorithm is applied for stereo correction, aligning the epipolar lines of the image, simplifying the matching process, and outputting the corrected image. and .

7. The binocular vision-based precision slag quantity detection system according to claim 6, characterized in that, The process of stereo matching and point cloud generation by the slag quantity detection module includes: Using an improved SGM algorithm, semi-global matching is employed, reducing the cost volume. The calculation is as follows: Where p is the number of pixels, d is the disparity, and R... p Let λ be the neighborhood window, λ be the gradient weight, and q be the gradient weight. d For matching points; Point cloud reconstruction, parallax map converted into point cloud : Where B is the baseline distance and f is the focal length; Point cloud filtering, applying statistical outlier filtering to remove noise, if points The following conditions must be met: Then remove it, where, Let k be the nearest neighbor set, and k be the size of the nearest neighbor set. Let be a point in the point set, where μ and σ are the mean and standard deviation of the distance.

8. The binocular vision-based precision slag quantity detection system according to claim 7, characterized in that, The process of slag layer segmentation and volume integration performed by the slag quantity detection module includes: Dynamic background removal, pre-stored clean point cloud state Real-time point cloud registration using ICP algorithm : Differential extraction of slag layer point cloud ; Slag layer meshing: Surface mesh S is constructed using Delaunay triangulation. Volume calculation, based on grid volume integration using the scan line method: Where M is the number of grids, A k Let the projected area of ​​the triangular facet be . The average depth of the vertices.

9. The binocular vision-based precision slag quantity detection system according to claim 1, characterized in that, By employing multi-frame image fusion technology, useful information is accumulated by analyzing the temporal relationship of consecutive multi-frame images, while suppressing the influence of random noise.

10. A method for accurate detection of slag quantity based on binocular vision, characterized in that, include: Step S1: Calculate the real-time running speed of the conveyor chain plate of the dry slag machine; Step S2: Based on the real-time running speed of the conveyor chain plate of the dry slag machine, perform time-series alignment on the 3D point cloud of the slag layer reconstructed by binocular stereo vision, and calculate the slag volume. Step S3: Distinguish between the conveyor chain plate structure of the dry slag machine and the coal slag, and provide higher quality 3D point cloud data of the slag layer for step S2.

Images