Aggregate motion tracking method and system during concrete mixing process based on monocular X-ray imaging

Abstract

The present invention discloses a method and system for tracking aggregate motion during concrete mixing based on monocular X-ray imaging. The system comprises: a single X-ray source that generates a cone beam of rays; a flat-panel detector located opposite the X-ray source; a tracer aggregate containing special markings; an image acquisition and processing unit; and a three-dimensional position and attitude calculation unit. The present invention reduces system costs: compared to binocular X-ray systems, the equipment required is halved, significantly reducing construction and maintenance costs. It also simplifies the system structure: a single X-ray source-detector combination reduces space requirements and installation complexity. It also achieves high-precision three-dimensional tracking: XY plane positioning accuracy can reach ±1mm, Z-axis positioning accuracy can reach ±2mm, and attitude calculation accuracy is controlled within ±5°.

Description

Technical Field

[0001] The present invention relates to the fields of concrete material science and engineering, process control and intelligent sensing technology, and in particular to a method and system for tracking aggregate motion during concrete mixing based on monocular X-ray imaging. Background Art

[0002] In concrete, the uniformity of the three-dimensional spatial distribution of aggregates is the core factor that determines its performance. Uneven aggregate distribution can have serious impacts on multiple levels: in terms of mechanical properties, the uneven distribution of aggregates can lead to stress concentration, weak interface transition zones, and significantly reduced compressive and flexural strengths; in terms of durability, differences in aggregate packing density can form permeation channels, exacerbating chloride ion erosion and freeze-thaw damage, and reducing the durability of concrete; in terms of construction performance, uneven aggregate distribution can change rheological properties and increase engineering risks such as pumping blockage and pouring segregation. Especially in modern concrete such as high-strength concrete and self-compacting concrete, which have extremely high requirements for aggregate particle size distribution and volume fraction, precise control of aggregate spatial distribution has become the key to improving project quality. However, traditional methods that rely on apparent indicators such as slump cannot effectively reflect the true spatial distribution of aggregates, resulting in large discreteness in concrete performance and difficulty in accurately controlling project quality.

[0003] One of the primary goals of concrete mixing is to achieve uniform distribution of aggregates, cement paste, and other components. Advanced concrete technologies, such as ultra-high-performance concrete (UHPC), self-compacting concrete, and emerging 3D-printed concrete, place even stricter demands on the spatial distribution of aggregates, which are directly linked to the uniformity of mixing.

[0004] The root cause of these problems lies in the opacity of concrete. Because its internal state cannot be directly observed, the concrete mixing process has long been considered a "black box." Mixing quality relies primarily on empirical judgment and post-process testing, lacking scientific, real-time process monitoring. This not only limits our understanding of the mechanisms that shape concrete properties, but also hinders the real-time and precise nature of quality control.

[0005] Existing monitoring technologies face many bottlenecks:

[0006] 1. Opacity challenge: Traditional optical methods cannot penetrate concrete and cannot directly observe the internal aggregate distribution.

[0007] 2. High cost and complexity: Although X-ray binocular or multi-camera CT systems can reconstruct aggregate distribution in three dimensions, the equipment is expensive, the system is complex, and the data processing volume is huge, making it difficult to widely deploy them in industrial sites for routine uniformity monitoring.

[0008] 3. Difficulty in obtaining three-dimensional information: Conventional monocular X-ray systems have low costs and simple structures, but it is difficult to directly obtain precise three-dimensional position and six-degree-of-freedom posture information, making it difficult to accurately evaluate the uniformity of aggregate distribution in three-dimensional space.

[0009] 4. Tracer Limitations: Existing methods that rely on specialized tracer aggregates have limitations in sample preparation, tracer recovery, environmental impact, and the ability to represent the behavior of real aggregates. There is an increasing need for “label-free” tracing of natural aggregates to assess their true distribution.

[0010] 5. Insufficient algorithm robustness: Under complex working conditions such as high aggregate concentration (for example, volume fraction exceeding 40%), rapid movement (for example, aggregate speed can reach 0.1-1 m / s), and frequent occlusion, the accuracy and robustness of existing tracking algorithms are difficult to meet the needs of effectively tracking large amounts of aggregate for statistical distribution analysis.

[0011] 6. Lack of real-time feedback and intelligent control: Existing monitoring systems are mostly open-loop systems, which cannot dynamically optimize mixing parameters according to the real-time aggregate distribution uniformity status and achieve closed-loop intelligent control.

[0012] 7. Limitations of uniformity assessment: Traditional two-dimensional or indirect uniformity assessment methods are difficult to fully reflect the true three-dimensional distribution state and local unevenness, and the evaluation of mixing quality is not comprehensive and accurate enough.

[0013] In response to the above technical problems, the present invention provides a method and system for tracking aggregate motion during concrete mixing based on monocular X-ray imaging. Summary of the Invention

[0014] The present invention provides a method and system for tracking aggregate motion during concrete mixing based on monocular X-ray imaging, which solves the problems in the background technology.

[0015] In order to achieve the above object, the present invention adopts the following technical solutions:

[0016] Aggregate motion tracking system during concrete mixing process based on monocular X-ray imaging, the system includes the following:

[0017] A single X-ray source produces a cone beam of rays;

[0018] Flat panel detector, located on the opposite side of the X-ray source;

[0019] Tracer aggregate containing special marking points;

[0020] Image acquisition and processing unit;

[0021] 3D position and attitude calculation unit;

[0022] The special marking points in the tracer aggregate produce high-contrast projections under X-ray irradiation. The image acquisition and processing unit captures these projection signals through a flat-panel detector to form a two-dimensional projection image containing the aggregate position and posture information. The image acquisition and processing unit pre-processes the projection image, extracts the aggregate characteristic parameters, measures the projection diameter, edge sharpness, and grayscale contrast information of the marking points, and passes the processed structured data to the three-dimensional position and posture solution unit. The three-dimensional position and posture solution unit receives the image processing results and calculates the three-dimensional position of the aggregate in space and its own rotation angle based on the principles of projection geometry.

[0023] Furthermore, the special marking points include:

[0024] The main marking point of the lead ball with a diameter of 2-3 mm is located inside the aggregate;

[0025] There are 2-3 lead auxiliary marking points with a diameter of 0.5-1 mm located near the surface of the aggregate, and the lead auxiliary marking points are distributed non-collinearly.

[0026] Furthermore, the voltage of the X-ray source is 380-400 kV, the current is 8-12 mA, the focal spot size is less than 0.5 mm, and the beam angle is 30°-40°.

[0027] Furthermore, it also includes a system online self-calibration module, a closed-loop control unit, a digital twin module, and a multi-source data fusion unit;

[0028] The system's online self-calibration module provides an accurate geometric parameter basis for the aggregate motion tracking system, ensuring the accuracy of three-dimensional position and attitude solutions, and improving the long-term stability of the system through parameter compensation; the multi-source data fusion unit receives aggregate motion and distribution data from the X-ray system, integrates multi-dimensional sensor information such as temperature, acoustics, and vibration, and provides comprehensive status information for the digital twin module and closed-loop control; the digital twin module drives the virtual model status update based on the fused data of the multi-source data fusion unit, performs parameter calibration and "what-if analysis", and provides predictive decision support for the closed-loop control unit; the closed-loop control unit generates the optimal control strategy based on the fused data and the twin model prediction results of the digital twin module, realizing adaptive optimization of the mixing process.

[0029] Furthermore, the three-dimensional position and attitude solving unit adopts one of the following solutions:

[0030] Algorithms for direct identification and tracking of natural aggregates or aggregates without pre-defined markers;

[0031] An end-to-end model based on deep learning that infers 3D position and pose directly from X-ray projection images;

[0032] Physical information neural networks that integrate physical law constraints are used to optimize three-dimensional information solutions.

[0033] Another object of the present invention is to provide a method for tracking aggregate motion during concrete mixing based on monocular X-ray imaging, comprising the following steps:

[0034] S1, aggregate image recognition and tracking;

[0035] S2, three-dimensional position derivation;

[0036] S3, solve the three-dimensional pose of the aggregate based on the PnP algorithm;

[0037] S4. Based on the three-dimensional position information of the aggregate obtained in step S2, evaluating the true three-dimensional distribution uniformity of the aggregate in the measured material;

[0038] S5. Inputting the calculated three-dimensional information and / or the assessed distribution uniformity as real-time monitoring data;

[0039] S7. Outputting the control instruction to the concrete mixing equipment to dynamically adjust at least one mixing parameter to achieve closed-loop control of the mixing process.

[0040] Furthermore, in step S2, the three-dimensional position derivation method includes:

[0041] S2.1, Z-axis position derivation method:

[0042] Using the principles of projective geometry, the relationship between the distance z from the object to the radiation source and the projection size is established: d = D × (SDD-z) / z, where D is the actual diameter of the lead ball, d is the projection diameter, and SDD is the distance from the radiation source to the detector.

[0043] By measuring the diameter d of the lead ball in the projection image and combining it with the pre-calibrated parameters, the above equation is solved to obtain the z value;

[0044] By continuously monitoring the changes in the aggregate projection size, the position changes of the aggregate along the X-ray direction can be deduced in real time, and stable operation can be achieved even when the aggregate rotates and the projection shape changes;

[0045] In the Z-axis position derivation method in step S2.1, an image processing algorithm with enhanced projection edge features is used to measure the projection size of the shot put, and a multi-feature fusion depth estimation model is used to reduce errors, while optimizing temporal consistency.

[0046] The steps of the image processing algorithm for projected edge feature enhancement are:

[0047] Applying non-isotropic diffusion filtering to reduce noise while preserving edges;

[0048] Use the Canny edge detection operator to extract the precise edge of the shot put projection;

[0049] Apply sub-pixel ellipse fitting algorithm to improve diameter measurement accuracy to 0.1 pixel;

[0050] The multi-feature fusion depth estimation model:

[0051] Construct depth estimation function: z = f(d,s,c,p);

[0052] Where d is the projection diameter, s is the edge sharpness, c is the contrast, and p is the relative position feature of the marker point;

[0053] The function is trained using a random forest regression model;

[0054] Reduce the error caused by a single feature by fusing multiple features;

[0055] The timing consistency optimization method:

[0056] Design the Kalman filter state vector, which contains position and velocity information;

[0057] Set process noise and observation noise parameters;

[0058] Apply a prediction-update loop to filter out random errors;

[0059] Consider the physical constraints of aggregate movement to enhance prediction accuracy.

[0060] S2.2, XY plane positioning algorithm:

[0061] Based on the geometric principle of X-ray projection, the mapping relationship between the aggregate center projection coordinates (u, v) and the actual coordinates (x, y) is established: x = (u-u0) ×z / f; y = (v-v0)×z / f, where (u0, v0) is the optical center projection coordinate, f is the equivalent focal length, and z is the distance from the object to the ray source;

[0062] Furthermore, in step S4, based on the three-dimensional position information of the aggregate, the true three-dimensional distribution uniformity of the aggregate in the measured material is evaluated by one of the following methods:

[0063] Methods based on spatial statistics, including nearest neighbor distance distribution, Voronoi diagram analysis, and radial distribution function;

[0064] Based on the grid method, the three-dimensional space of the mixing container is divided into several small cubic units, and the coefficient of variation of the aggregate quantity or volume proportion in each unit is calculated.

[0065] Furthermore, in step S3, solving the three-dimensional posture of the aggregate based on the PnP algorithm includes:

[0066] S31. Establish a perspective-n-point problem model;

[0067] S32. Apply the EPnP or P3P algorithm to solve the rotation matrix and translation vector;

[0068] S33. Convert the rotation matrix into Euler angle or quaternion representation;

[0069] S34. Apply the RANSAC algorithm and time series smoothing to optimize the posture solution results.

[0070] Furthermore, the steps of aggregate image recognition and tracking in step S1 are as follows:

[0071] S11, extracting characteristic parameters of aggregate in X-ray projection;

[0072] S12, establishing an aggregate feature fingerprint library;

[0073] S13, aggregate identification based on deep learning algorithm;

[0074] S14, applying a multi-target tracking algorithm to simultaneously track the motion states of multiple aggregates;

[0075] S15. Record and analyze the three-dimensional motion trajectory of aggregate in the mixing space.

[0076] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0077] 1. Reduce system costs: Compared with binocular X-ray systems, the equipment requirements are reduced by half, significantly reducing construction and maintenance costs;

[0078] 2. Simplified system structure: Single X-ray source-detector combination reduces space requirements and installation complexity;

[0079] 3. Achieve high-precision three-dimensional tracking: XY plane positioning accuracy can reach ±1mm, Z axis positioning accuracy can reach ±2mm, and attitude solution accuracy is controlled within ±5°;

[0080] 4. Improve temporal resolution: Through algorithm optimization and hardware acceleration, the system throughput can reach over 100fps, meeting the requirements of high-dynamic scenes;

[0081] 5. Enhanced applicability: Applicable to various concrete mixes and mixing environments, and can be expanded to other opaque suspension research;

[0082] 6. Simultaneous tracking of multiple aggregates: Even in grayscale images, the motion states and trajectories of multiple aggregates can be tracked simultaneously;

[0083] 7. Provide quantitative indicators of uniformity: through uniformity algorithm, the mixing quality is evaluated in real time, providing an objective basis for optimizing the mixing process;

[0084] 8. Solve the "black box" problem of concrete: achieve fully transparent characterization, accurate evaluation and intelligent control of the concrete mixing process;

[0085] 9. Significantly improved monitoring accuracy and robustness: Advanced algorithms and optimized imaging technology significantly improve 3D tracking accuracy and robustness under more complex working conditions;

[0086] 10. Achieving “label-free” tracking and expanding applicability: Breakthroughs in label-free tracking technology will greatly simplify sample preparation, reduce costs, and avoid the potential impact of tracers on concrete properties;

[0087] 11. Intelligent closed-loop control and optimization of the mixing process: A closed-loop control system based on real-time monitoring of aggregate distribution uniformity and digital twin technology enables adaptive optimization of the mixing process;

[0088] 12. Provide more comprehensive and in-depth insights into material behavior: True 3D uniformity assessment, local inhomogeneity identification, and correlation with rheological parameters provide unprecedented means for a deeper understanding of the internal structural evolution and performance formation mechanisms of concrete materials;

[0089] 13. Improve concrete quality control and project safety: Through real-time, precise process monitoring and intelligent optimization based on aggregate distribution uniformity, concrete performance dispersion and project hazards caused by uneven mixing can be significantly reduced;

[0090] 14. Environmental friendliness and sustainability: The development of environmentally friendly tracer materials and the implementation of "label-free" tracking can help reduce potential environmental pollution, and intelligent optimization of the mixing process can save energy and reduce consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 This is a schematic diagram of the overall structure of the monocular X-ray aggregate tracking system of the present invention, which shows the layout relationship between the X-ray source, the mixing container, and the flat panel detector;

[0092] Figure 2 It is a flow chart of the “markerless” aggregate tracking algorithm;

[0093] Figure 3 This is a framework diagram of the aggregate pose estimation algorithm based on physical information neural network (PINN);

[0094] Figure 4 This is a schematic diagram of the digital twin system architecture for the concrete mixing process;

[0095] Figure 5 This is the flow chart of the closed-loop stirring control system. DETAILED DESCRIPTION

[0096] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0097] Aggregate motion tracking system and method during concrete mixing process based on monocular X-ray imaging.

[0098] Example 1: Implementation of a monocular X-ray imaging system

[0099] This embodiment provides a detailed implementation scheme of a monocular X-ray imaging system:

[0100] 1. System structure design: such as Figure 1 As shown in the figure, the system uses a single X-ray source and a flat-panel detector in an opposing arrangement. The X-ray source is located on one side of the mixing vessel, and the flat-panel detector is located on the other side. The X-rays emitted by the source pass through the concrete mixing vessel and are received by the flat-panel detector, forming a projection image.

[0101] 2. X-ray source selection and parameter setting:

[0102] Use an industrial-grade X-ray source with a voltage of 380-400kV;

[0103] The current is set to 8-12mA to provide sufficient penetration;

[0104] The focal size is controlled below 0.5mm to ensure projection clarity;

[0105] A cone beam is used with a beam angle of 30°-40°, so that the X-rays are distributed divergently in space.

[0106] 3. Selection and configuration of flat panel detectors:

[0107] Effective imaging area: 500mm×500mm;

[0108] Pixel size: 150 μm;

[0109] Resolution: 3072×3072;

[0110] Maximum frame rate: 200fps;

[0111] Dynamic range: 18bit;

[0112] 4. System calibration and correction method:

[0113] Using a calibration body with known geometric dimensions, projection images are acquired at different depths;

[0114] Measure the changes in the projection size of the calibration body at different depths;

[0115] Establishing a magnification scale-distance mapping relationship curve;

[0116] Use the polynomial fitting method to establish the functional relationship: d = f(z);

[0117] Or use a lookup table method to establish a numerical mapping.

[0118] 5. Adaptive ray parameter module:

[0119] The system integrates a feedback module that dynamically adjusts the X-ray source's tube voltage (adjustment range 100-450 kV) and tube current (adjustment range 1-15 mA) through a control algorithm based on the average grayscale value received by the detector (the target average grayscale value is set at 50-70% of the detector dynamic range), image contrast (target contrast-to-noise ratio > 5), or assessment of concrete penetration depth.

[0120] In the initial stage of mixing, the concrete is relatively sparse, and the parameters can be appropriately lowered to reduce scattering and radiation; as mixing progresses and admixtures act, if the density of the mixture increases, the parameters will be automatically increased to ensure penetration.

[0121] 6. System online self-calibration module:

[0122] Reference markers with known geometric features (such as 5 mm diameter tungsten balls) are introduced at fixed positions or periodically in the stirred vessel.

[0123] The image processing unit periodically (every 5-10 minutes) detects the projection position and size of these reference markers.

[0124] If the detection results deviate from the calibration values ​​(position deviation > 0.5 pixels, size deviation > 0.2%), the calibration module will fine-tune the projection geometry model parameters (such as equivalent focal length, optical center position, detector inclination, etc.), or compensate for the subsequent position and attitude solution results.

[0125] Example 2: Three-dimensional position derivation method

[0126] 1.XY plane positioning algorithm:

[0127] Based on the geometric principle of X-ray projection, the mapping relationship between the aggregate center projection coordinates (u, v) and the actual coordinates (x, y) is established:

[0128] x=(u- u0)×z / f;

[0129] y=(v- v0)×z / f;

[0130] Where (u0, v0) is the projection coordinate of the optical center, f is the equivalent focal length, and z is the distance from the object to the ray source.

[0131] 2. Z-axis position derivation method:

[0132] Using the principles of projection geometry, the relationship between the distance z from the object to the ray source and the projection size is established: d = D × (SDD-z) / z;

[0133] Where D is the actual diameter of the lead ball, d is the projected diameter, and SDD is the distance from the ray source to the detector;

[0134] By measuring the diameter d of the lead ball in the projection image and combining it with the pre-calibrated parameters, the above equation is solved to obtain the z value.

[0135] By continuously monitoring the changes in the aggregate projection size, the position changes of the aggregate along the X-ray direction can be deduced in real time, and it can work stably even when the projection shape changes due to the rotation of the aggregate.

[0136] 3. The image processing algorithm for projected edge feature enhancement is:

[0137] Applying non-isotropic diffusion filtering to reduce noise while preserving edges;

[0138] Use the Canny edge detection operator to extract the precise edge of the shot put projection;

[0139] A sub-pixel ellipse fitting algorithm is applied to improve the diameter measurement accuracy to 0.1 pixel.

[0140] 4. Multi-feature fusion depth estimation model:

[0141] Construct depth estimation function: z=f(d,s,c,p);

[0142] Where d is the projection diameter, s is the edge sharpness, c is the contrast, and p is the relative position feature of the marker point;

[0143] The function is trained using a random forest regression model;

[0144] Reduce the error caused by a single feature by fusing multiple features.

[0145] 5. Timing consistency optimization method:

[0146] Design the Kalman filter state vector, which contains position and velocity information;

[0147] Set process noise and observation noise parameters;

[0148] Apply a prediction-update loop to filter out random errors;

[0149] Consider the physical constraints of aggregate movement to enhance prediction accuracy.

[0150] Example 3: Posture calculation from a monocular perspective

[0151] 1. PnP problem modeling method:

[0152] Model the aggregate pose solution problem as a perspective-n-point problem;

[0153] Mathematically expressed as: [u, v] = K[R|t][X, Y, Z, 1]ᵀ;

[0154] Where K is the camera intrinsic parameter matrix, [R|t] is the posture parameter, [X, Y, Z] is the 3D coordinate of the marker point, and [u, v] is the projection coordinate.

[0155] 2.EPnP algorithm implementation details:

[0156] Represent n 3D points as a linear combination of 4 control points;

[0157] Establish the projection equations of the control points;

[0158] Solve linear systems via singular value decomposition (SVD);

[0159] For the case of 3 markers, the P3P algorithm is used to solve the problem directly.

[0160] 3. Posture representation and conversion method:

[0161] Convert the solved rotation matrix R into Euler angle representation (rll, pitch, yaw)

[0162] Or convert to quaternion representation to reduce singularity problems;

[0163] Establish the local coordinate system of the aggregate so that the principal axes are aligned with the principal inertia axes of the aggregate.

[0164] 4. Specific technologies for improving posture accuracy:

[0165] Optimize the layout of marking points so that they are distributed in different directions on the aggregate surface as much as possible;

[0166] Apply the RANSAC algorithm to filter out possible outliers;

[0167] The ellipsoid of the lead ball projection is used as an auxiliary constraint;

[0168] Apply low-pass filtering to smooth the time series and remove high-frequency noise.

[0169] 5. Posture verification and correction method:

[0170] Verification experiments were conducted using aggregates with known postures;

[0171] Analyze the accuracy distribution of posture calculation at different angles and depths;

[0172] Establish an error compensation model and perform specific optimization for high error areas;

[0173] Under typical working conditions, the control rotation angle error is within ±5°.

[0174] 6. End-to-end deep learning model construction:

[0175] Network architecture: A convolutional neural network is designed with ResNet-50 or EfficientNet-B3 as the backbone network to extract image features, followed by fully connected layer branches to regress the 2D center, depth, and pose parameters of the aggregate projection.

[0176] For sequential images, LSTM (1-2 layers, 128-256 hidden units) or Transformer layers (encoder-decoder structure, 4-8 attention heads) are introduced to capture temporal dependencies.

[0177] Generate a large number of (10 5 -10 6 ) training images with precise 3D position and pose labels.

[0178] The loss function includes 2D position loss, depth loss, and posture loss, and the total loss is the weighted sum of each component.

[0179] 7. Integration of Physical Information Neural Networks (PINNs):

[0180] In the loss function of the deep learning model, an additional physical constraint loss term is introduced to penalize prediction results that do not conform to known physical laws.

[0181] Physical laws include: kinematic constraints (speed and acceleration should be within a reasonable range), dynamic constraints (considering gravity, fluid resistance, and collisions between aggregates), and geometric constraints (aggregates cannot penetrate the mixer wall).

[0182] The automatic differentiation technique is used to calculate the derivative of the network output with respect to the input, and the physical laws are directly embedded in the loss function.

[0183] Methods such as Monte Carlo dropout (randomly inactivating neurons with a probability of 0.1-0.3 in the prediction stage and performing 50-100 forward propagations) are used to quantify the uncertainty of the prediction.

[0184] Example 4: Aggregate Image Recognition and Tracking

[0185] 1. Aggregate feature extraction:

[0186] Extract aggregate contours in the projection image based on grayscale distribution features;

[0187] Calculate the geometric parameters of each aggregate, including area, perimeter, major-minor axis ratio, shape factor, etc.;

[0188] Extract the relative position relationship and size characteristics of the marked points in the aggregate;

[0189] Generate unique feature vectors of aggregates as the basis for identification and tracking.

[0190] 2. Aggregate identification algorithm:

[0191] Establish an aggregate feature fingerprint library to store the feature vector of each target aggregate;

[0192] Apply deep learning models such as convolutional neural networks (CNN) to train aggregate recognition systems;

[0193] Even if the aggregates look similar in grayscale projection, different aggregates can be distinguished by small feature differences;

[0194] The recognition algorithm has an accuracy of over 95% and can still work effectively even in partial occlusion.

[0195] 3.Simultaneous tracking of multiple aggregates:

[0196] Based on the matching of aggregate features and the continuity of motion speed, the inter-frame aggregate association is realized;

[0197] Apply a multi-target tracking algorithm, such as SORT (Simple Online and Realtime Tracking) or DeepSORT.

[0198] Establish an aggregate movement trajectory database to record the complete spatiotemporal information of each aggregate.

[0199] It can maintain stable tracking performance at up to 30% aggregate density.

[0200] 4. Occlusion processing strategy:

[0201] An occlusion prediction model is developed to record the motion of aggregates before they are occluded.

[0202] A trajectory prediction algorithm is applied to maintain tracking continuity during short-term occlusion of aggregates.

[0203] When the aggregate reappears, its identity is restored through feature matching.

[0204] The system can handle temporary occlusion of aggregate for up to 2 seconds.

[0205] 5. Three-dimensional motion analysis:

[0206] Combining the three-dimensional position derivation and attitude solution results, the complete six-degree-of-freedom motion of the aggregate is reconstructed.

[0207] Analyze the movement pattern of aggregate in the mixing space, including parameters such as translational velocity, angular velocity, and acceleration.

[0208] Statistically analyze the collective motion characteristics of multiple aggregates to evaluate mixing uniformity and efficiency.

[0209] Generate a 3D flow field analysis of the mixing space to intuitively display the rheological properties of concrete.

[0210] 6. Direct tracking of “unmarked” natural aggregates:

[0211] Image preprocessing: Advanced noise reduction algorithms (such as BM3D) and contrast enhancement techniques (such as CLAHE, with a window size of 8x8) are used in combination with a deep learning-based image enhancement network to highlight the edges and textures of natural aggregates.

[0212] Aggregate segmentation: Use the improved watershed algorithm, region growing algorithm, or train a specialized semantic segmentation network (such as U-Net and its variants) to accurately segment the 2D projection of a single natural aggregate from the complex background and contacting aggregates.

[0213] Feature extraction: For each segmented natural aggregate projection, extract its shape descriptor (such as contour Fourier descriptor, Zernike moment), texture features, size, and possible internal features.

[0214] Depth estimation and pose inference: through model-based fitting (if a 3D model library of aggregates can be obtained in advance) or learning-based methods (training deep neural networks to directly regress the 3D position and pose of natural aggregate projection images from single or multiple frames).

[0215] Tracking: Use a tracking algorithm based on appearance features and motion prediction (such as a tracker that combines a Siamese network for feature matching and a Kalman filter for motion modeling).

[0216] 7. Graph-based tracking methods:

[0217] Graph construction: In each frame, the detected aggregates are used as nodes, and edges are constructed based on the appearance similarity of the aggregates, the proximity between the motion prediction position and the actual detection position, and the possible topological relationship between the aggregates. The edge weight represents the possibility of association.

[0218] Tracking is path search / matching: This method transforms the multi-target tracking problem into finding the optimal path or performing graph matching on a spatiotemporal graph. Graph neural networks (such as a 2-3 layer Graph Attention Network) or the classic Hungarian algorithm, minimum cost flow, and other methods are used for data association.

[0219] Long-term occlusion processing: When an aggregate reappears after being occluded for a long time, more reliable identity re-identification is performed by combining its motion trajectory before occlusion and its relative relationship pattern with other unoccluded aggregates. A dedicated re-identification network can be trained.

[0220] Example 5: Design and application of new tracer aggregate

[0221] 1. Coding Aggregate Design:

[0222] Select the actual aggregate density (2.5-2.7 g / cm 3 ), matrix materials of similar size (such as epoxy resin or polymers compatible with concrete mortar);

[0223] The internal part is embedded with a specific coding pattern (such as a 3x3 dot matrix that can encode 2) composed of materials with different X-ray absorption coefficients (such as tungsten powder, molybdenum wire, diameter 0.1-0.5mm) through precision machining or 3D printing technology. 9 information).

[0224] The pattern can be identified under X-ray projection and decoded into the unique ID of the aggregate.

[0225] 2. “Smart” aggregate:

[0226] Explore integrating microstructures or materials inside or on the surface of aggregates that are sensitive to specific physical fields (e.g., micro-deformations caused by temperature changes >5°C or stress changes >1MPa);

[0227] These structures or materials can produce identifiable tiny characteristic changes under X-ray irradiation (such as grayscale changes in a specific area >5%, diffraction spot displacement >0.1 pixel);

[0228] Extremely high-resolution imaging and sophisticated signal interpretation algorithms are required.

[0229] 3. Environmentally friendly tracer materials:

[0230] Select metal oxides with high atomic number but low biotoxicity (such as bismuth oxide Bi2O3, with a density of about 8.9 g / cm 3 ; Zirconium oxide ZrO2, density about 5.68 g / cm 3 ) or high-density ceramics (such as silicon nitride Si3N4, with a density of about 3.2 g / cm 3; Barium sulfate BaSO4 filled polymer composite material) as a marker material to replace lead.

[0231] The contrast ratio within the concrete X-ray energy range (100-400keV) and the long-term chemical stability in alkaline environment were verified experimentally.

[0232] 4. Multimodal tracer aggregate:

[0233] Aggregates are designed to be inspected simultaneously by X-rays and other modalities (e.g. magnetic induction frequencies 1-10kHz, ultrasonic frequencies 0.5-5MHz).

[0234] Used for data cross-validation or supplementary information to improve the robustness and accuracy of tracking.

[0235] Example 6: Aggregate Distribution Uniformity Assessment

[0236] This embodiment provides a method for quickly evaluating concrete mixing uniformity. This method eliminates the need to track the identity of individual aggregates and enables real-time evaluation of mixing quality by analyzing the overall distribution characteristics of aggregates in the projection image.

[0237] 1. Aggregate segmentation and identification:

[0238] Using the grayscale difference of X-ray projection images, each independently moving aggregate area is segmented;

[0239] The connected domain analysis algorithm is applied to identify the interconnected grayscale regions as a single aggregate;

[0240] Through morphological processing, the aggregate boundaries are optimized and noise and small connections are removed;

[0241] For each identified aggregate region, its projected area and centroid position are calculated.

[0242] 2. Basic algorithm flow for uniformity calculation:

[0243] Get the number n of all aggregates in the current projection image and the projection area a[i] of each aggregate (i=1,2,…,n);

[0244] Calculate the total projected area of ​​aggregate A=Σa[i];

[0245] Get the total area S of the mixing chamber on the projection plane;

[0246] Calculate the area expansion coefficient e=A / S;

[0247] Taking the centroid of each aggregate as the center, expand each aggregate by a factor of e (overlapping is allowed);

[0248] Calculate the total area A' covered by all aggregates after expansion (the overlapping part is counted only once);

[0249] Calculate the uniformity index U = A' / S, which ranges from 0 to 1. The closer the value is to 1, the more uniform the distribution.

[0250] 3. Image processing implementation details:

[0251] Create a binary mask image of the same size as the projected image, with all initial values ​​set to 0;

[0252] For each aggregate, draw a circle or ellipse with an area of ​​e×a[i] centered at its centroid;

[0253] Mark the drawn area as 1 in the mask image;

[0254] Count the number of pixels with a value of 1 in the mask image and calculate its ratio to the total pixels, which is the uniformity U.

[0255] 4. Weighted uniformity model:

[0256] Considering the different contributions of aggregates at different positions to uniformity, a position weight factor is introduced;

[0257] Set a higher weight for the center area of ​​the mixing tank and a lower weight for the edge area;

[0258] Calculate the position weight w[i] based on the distance from the center of mass of the aggregate to the center of the mixing bin;

[0259] The corrected uniformity calculation formula is U_w = Σ(w[i]×a'[i]) / Σ(w[i]×s[i]);

[0260] Where a'[i] is the effective coverage area after the expansion of the i-th aggregate, and s[i] is the ideal coverage area at that location.

[0261] 5. Uniformity timing analysis:

[0262] Record the change curve U(t) of uniformity U over time t during the entire stirring process;

[0263] Analyze the convergence characteristics of U(t) and determine the time point t* when the stirring tends to be stable;

[0264] The uniformity change rate dU / dt is calculated. When it is continuously lower than the threshold ε, it is determined that the stirring has reached the optimal state.

[0265] A library of uniformity characteristic curves for different concrete mix ratios is established to guide actual production.

[0266] 6.Uniformity assessment application:

[0267] The current uniformity value and its changing trend are displayed in real time on the control interface.

[0268] Set an alarm mechanism based on the uniformity threshold and issue a reminder when the uniformity falls below the preset value.

[0269] Based on historical data, a relationship model between mixing time and uniformity is established to optimize mixing process parameters.

[0270] Combined with other quality indicators, a comprehensive evaluation system is constructed to provide a scientific decision-making basis for concrete production.

[0271] 7. In addition, the development of true three-dimensional uniformity index, based on the three-dimensional coordinates of aggregate, develops indicators and algorithms that can directly evaluate the distribution uniformity in three-dimensional space. The calculation steps of true three-dimensional uniformity index are as follows:

[0272] Step 1: Get the three-dimensional centroid coordinates (x_i, y_i, z_i) of all tracked aggregates at a certain moment;

[0273] Step 2: Based on spatial statistical methods:

[0274] Nearest neighbor distance distribution: Calculate the three-dimensional Euclidean distance from each aggregate to its nearest neighbor aggregate, and analyze the mean, variance, and distribution of these distances. A more concentrated distribution (small variance) and a smaller mean (but not zero to avoid overlap) generally indicate a more uniform distribution.

[0275] Voronoi diagram analysis: Construct a three-dimensional Voronoi diagram of the aggregate centroid and analyze the volume distribution of the Voronoi units. The more uniform the volume distribution (coefficient of variation < 0.1), the more uniform the aggregate distribution.

[0276] Radial distribution function / pair correlation function g(r): Statistically calculates the probability density of other aggregates appearing in a spherical shell with a distance r to r+dr centered on any aggregate. The ideal uniform distribution has a specific g(r)=1 (for large r).

[0277] Step 3: Grid-based method: Divide the three-dimensional space of the mixing container into several small cubic units (with a side length of 1-3 times the average aggregate particle size), count the number, volume percentage, or mass of aggregate in each unit, and calculate the coefficient of variation, information entropy, or standard deviation of these values ​​across all units as an indicator of unevenness.

[0278] 8. Local uniformity and mixed dead zone identification:

[0279] Using a grid-based approach, identify areas where the aggregate quantity or volume fraction is much lower than average (e.g., less than 30-50% of the average) or where no aggregate passes through for extended periods of time (e.g., for more than 10% of the mix cycle). These areas may be mixing dead zones.

[0280] Visualize these areas and analyze their causes (e.g. impeller design, speed, material viscosity, etc.).

[0281] 9. Predictive uniformity analysis:

[0282] Record the uniformity index U(t) at a series of moments during the stirring process.

[0283] Use time series models (such as ARIMA, LSTM) or physics-based mixed models to fit the evolution law of U(t).

[0284] Based on the data of the current mixing stage, the remaining time required to reach the target uniformity is predicted, or the steady-state uniformity that can be achieved under the current mixing parameters is predicted.

[0285] Example 7: Intelligent system integration and closed-loop control

[0286] 1. Software and hardware integration solution:

[0287] Develop integrated control software with modular design;

[0288] Design workflows that coordinate X-ray source control, image acquisition, processing, and storage;

[0289] The software adopts a multi-threaded architecture to separate image acquisition and processing modules;

[0290] Implement parameter setting interface to facilitate adjustment of operating parameters.

[0291] 2. Real-time performance optimization technology:

[0292] Use CUDA to implement GPU accelerated image processing and pose solving.

[0293] It adopts pipeline processing architecture to process multiple frames of images in parallel.

[0294] Key algorithm optimization, such as using lookup tables instead of complex calculations.

[0295] Image region selective processing to analyze only the region of interest.

[0296] 3. Fusion with multi-source data:

[0297] Real-time correlation with stirring motor parameters (speed, torque).

[0298] Establish the relationship between aggregate movement index and mixing energy.

[0299] Fusion with rheological test data to correlate flow field characteristics with material properties.

[0300] Develop a data visualization interface to intuitively display multidimensional data relationships.

[0301] 4. Application scenario expansion case:

[0302] Used for studying the rheological properties of various opaque suspensions.

[0303] Used for optimization and monitoring of industrial mixing processes.

[0304] Used for analysis of the dynamic behavior of granular materials.

[0305] Used to monitor material flow status during pumping process.

[0306] 5. Edge computing deployment:

[0307] X-ray image acquisition cards are directly connected to edge computing devices with GPUs (such as NVIDIA AGX Orin);

[0308] Complete image preprocessing (frame rate > 100fps), aggregate detection, preliminary tracking, and some pose calculation tasks on this device;

[0309] The processed structured data (such as aggregate ID, 2D / 3D coordinates, timestamp, data packet size <1KB / aggregate / frame) is sent to the central server through the network.

[0310] 6. Closed-loop stirring control system implementation:

[0311] Input: The central server receives in real time the aggregate distribution uniformity index U(t) (updated 1-5 times per second), local unevenness information, average kinetic energy of aggregate, etc. from the X-ray monitoring system.

[0312] Control logic / decision making unit:

[0313] Rule-based control: For example, if dU / dt approaches zero but U(t) is still below the target value, increase the impeller speed or change the stirring direction; if a persistent mixing dead zone is detected, try changing the stirring mode.

[0314] Model-based predictive control: Utilizes a dynamic model of the mixing process to predict the impact of different control inputs on uniformity and optimizes the control sequence online to achieve the target uniformity in the fastest and most energy-efficient manner.

[0315] Reinforcement Learning Controller: Train a reinforcement learning agent to learn the optimal stirring control policy by interacting with the stirring environment.

[0316] Output: Control instructions are sent to the PLC or motor driver of the mixing equipment to adjust the mixing parameters.

[0317] 7. Construction of digital twin of mixing process:

[0318] Geometric model: Accurate 3D CAD model of the mixing vessel and mixing blades.

[0319] Physical Model:

[0320] Concrete rheological models (such as Bingham model = +η p , Herschel-Bulkley model = +K n , parameters can be adjusted online according to real-time monitoring data);

[0321] Aggregate motion model (considering fluid drag, collisions between aggregates, and interactions with blades and vessel walls);

[0322] Data interface: Receives real-time data on aggregate position, speed, uniformity, etc. from the X-ray monitoring system and other sensors to drive the status update and parameter calibration of the twin model;

[0323] Visualization and interactive interface: Provides real-time 3D visualization of the internal state of the mixing process, allowing users to perform "what-if" analysis and simulate the impact of different process parameters.

[0324] 8. Multi-source data fusion:

[0325] Install temperature sensors on the mixing equipment to monitor concrete temperature changes (related to hydration reaction and frictional heat generation, with an expected temperature rise of 5-20°C);

[0326] Install acoustic sensors or acceleration sensors to monitor the load on the agitator blades, the sound of aggregate collision, or the vibration of the equipment;

[0327] These data are aligned in time with the aggregate motion and distribution data obtained by the X-ray system (synchronization accuracy <10ms) for fusion analysis;

[0328] Leverage machine learning models (e.g., multimodal fusion networks, e.g., attention-based fusion layers) to comprehensively analyze all sensor data to more accurately assess the mix state, predict early concrete properties, or detect abnormal conditions.

[0329] Compared with the existing technology, the main innovation of the present invention is that it realizes the three-dimensional tracking of aggregates inside concrete through a single X-ray system. The key technologies include special tracer aggregate design, multi-feature fusion depth estimation and PnP-based posture solution. In particular, the method of deducing the Z-axis position through the change of projection size, and the image recognition algorithm for distinguishing and tracking multiple aggregates in the grayscale projection image, solve the technical difficulties of monocular X-ray systems in three-dimensional space monitoring. In addition, the original method of assessing the uniformity of aggregate distribution of the present invention provides a simple and efficient means of real-time monitoring of mixing quality, which can objectively evaluate the mixing uniformity without the need for complex aggregate identity tracking. The present invention greatly reduces the cost and complexity of equipment, making the technology widely applicable to the study of concrete mixing mechanism and quality control.

[0330] The leap from "label reliance" to "label-free priority" in this invention has greatly expanded the universality and convenience of the technology; the algorithm upgrade from "simple model" to "AI empowerment + physical constraints" has significantly improved the analysis capabilities in complex scenarios; the system evolution from "open-loop monitoring" to "closed-loop intelligent control + digital twin" has enabled the mixing process from "observable" to "predictable, optimizable, and adaptive"; the monitoring expansion from "single aggregate" to "multi-component, multi-scale" adapts to the needs of the development of modern high-performance concrete; the uniformity characterization from "two-dimensional / indirect" to "true three-dimensional, local refinement" provides a more scientific basis for quality evaluation.

[0331] This invention is not only applicable to concrete mixing research but can also be extended to other fields such as the rheology of opaque suspensions and the dynamic analysis of granular materials. The system's modular design allows for adjustments based on specific application scenarios to meet varying accuracy and real-time requirements. This invention not only provides an unprecedentedly powerful tool for in-depth research on concrete mixing mechanisms and refined quality control of production processes, but also offers valuable technical solutions and broad application prospects for other industrial fields involving opaque multiphase fluid mixing and particle dynamics analysis, such as the chemical, pharmaceutical, food, and mining industries.

[0332] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. Aggregate motion tracking system during concrete mixing process based on monocular X-ray imaging, characterized by: The system includes the following: A single X-ray source produces a cone beam of rays; Flat panel detector, located on the opposite side of the X-ray source; Tracer aggregate containing special marking points; Image acquisition and processing unit; 3D position and attitude calculation unit; Special marker points in the tracer aggregate produce high-contrast projections under X-ray irradiation. The image acquisition and processing unit captures these projection signals through a flat-panel detector, forming a two-dimensional projection image containing the aggregate's position and posture information. The image acquisition and processing unit preprocesses the projection image, extracts the aggregate's characteristic parameters, measures the projection diameter, edge sharpness, and grayscale contrast information of the marker points, and transmits the processed structured data to the three-dimensional position and posture solution unit. The three-dimensional position and posture solution unit receives the image processing results and calculates the three-dimensional position of the aggregate in space and its own rotation angle based on the principles of projective geometry. It also includes a system online self-calibration module, a closed-loop control unit, a digital twin module, and a multi-source data fusion unit; The system's online self-calibration module provides an accurate geometric parameter basis for the aggregate motion tracking system, ensuring the accuracy of three-dimensional position and attitude calculations and improving the system's long-term stability through parameter compensation. The multi-source data fusion unit receives aggregate motion and distribution data from the X-ray system, integrating temperature, acoustic, and vibration multi-dimensional sensor information to provide comprehensive status information for the digital twin module and closed-loop control. The digital twin module drives the virtual model status update based on the fused data of the multi-source data fusion unit, performs parameter calibration and hypothesis analysis, and provides predictive decision support for the closed-loop control unit; the closed-loop control unit generates the optimal control strategy based on the fused data and the twin model prediction results of the digital twin module to achieve adaptive optimization of the mixing process.

2. The aggregate motion tracking system during concrete mixing based on monocular X-ray imaging according to claim 1, characterized in that: The special marking points include: The main marking point of the lead ball with a diameter of 2-3 mm is located inside the aggregate; There are 2-3 lead auxiliary marking points with a diameter of 0.5-1 mm located near the surface of the aggregate, and the lead auxiliary marking points are distributed non-collinearly.

3. The aggregate motion tracking system during concrete mixing based on monocular X-ray imaging according to claim 1, characterized in that: The voltage of the X-ray source is 380-400 kV, the current is 8-12 mA, the focal spot size is less than 0.5 mm, and the beam angle is 30°-40°.

4. The aggregate motion tracking system during concrete mixing based on monocular X-ray imaging according to claim 1, characterized in that: The three-dimensional position and attitude solving unit adopts one of the following solutions: Algorithms for direct identification and tracking of natural aggregates or aggregates without pre-defined markers; An end-to-end model based on deep learning that infers 3D position and pose directly from X-ray projection images; Physical information neural networks that integrate physical law constraints are used to optimize three-dimensional information solutions.

5. A method for tracking aggregate motion during concrete mixing based on monocular X-ray imaging, characterized by: The method for tracking aggregate motion during concrete mixing based on monocular X-ray imaging according to claim 1 comprises the following steps: S1, aggregate image recognition and tracking; S2, three-dimensional position derivation; S3, solve the three-dimensional pose of the aggregate based on the PnP algorithm; S4. Based on the three-dimensional position information of the aggregate obtained in step S2, evaluating the true three-dimensional distribution uniformity of the aggregate in the measured material; S5. Inputting the calculated three-dimensional information and / or the assessed distribution uniformity as real-time monitoring data; S7. Outputting a control instruction to the concrete mixing equipment to dynamically adjust at least one mixing parameter to achieve closed-loop control of the mixing process.

6. The method for tracking aggregate motion during concrete mixing based on monocular X-ray imaging according to claim 5, characterized in that: In step S2, the three-dimensional position derivation method includes the following steps: S2.1, Z-axis position derivation method: Using the principles of projective geometry, the relationship between the distance z from the object to the radiation source and the projection size is established: d = D × (SDD-z) / z, where D is the actual diameter of the lead ball, d is the projection diameter, and SDD is the distance from the radiation source to the detector; By measuring the diameter d of the lead ball in the projection image and combining it with the pre-calibrated parameters, the above equation is solved to obtain the z value; By continuously monitoring the changes in the aggregate projection size, the position changes of the aggregate along the X-ray direction can be deduced in real time, and stable operation can be achieved even when the aggregate rotates and the projection shape changes; In the Z-axis position derivation method in step S2.1, an image processing algorithm with enhanced projection edge features is used to measure the projection size of the shot put, and a multi-feature fusion depth estimation model is used to reduce errors, while optimizing temporal consistency. The steps of the image processing algorithm for projected edge feature enhancement are: Applying non-isotropic diffusion filtering to reduce noise while preserving edges; Use the Canny edge detection operator to extract the precise edge of the shot put projection; Apply sub-pixel ellipse fitting algorithm to improve diameter measurement accuracy to 0.1 pixel; The multi-feature fusion depth estimation model: Construct depth estimation function: z = f(d, s, c, p); Where d is the projection diameter, s is the edge sharpness, c is the contrast, and p is the relative position feature of the marker point; The function is trained using a random forest regression model; Reduce the error caused by a single feature by fusing multiple features; The timing consistency optimization method: Design the Kalman filter state vector, which contains position and velocity information; Set process noise and observation noise parameters; Apply a prediction-update loop to filter out random errors; Considering the physical constraints of aggregate movement to enhance prediction accuracy; S2.2, XY plane positioning algorithm: Based on the geometric principle of X-ray projection, the mapping relationship between the aggregate center projection coordinates (u, v) and the actual coordinates (x, y) is established: x = (u-u0) × z / f; y = (v-v0) × z / f, where (u0, v0) is the optical center projection coordinate, f is the equivalent focal length, and z is the distance from the object to the ray source.

7. The method for tracking aggregate motion during concrete mixing based on monocular X-ray imaging according to claim 5, characterized in that: In step S4, based on the three-dimensional position information of the aggregate, the true three-dimensional distribution uniformity of the aggregate in the measured material is evaluated by one of the following methods: Methods based on spatial statistics, including nearest neighbor distance distribution, Voronoi diagram analysis, and radial distribution function; Based on the grid method, the three-dimensional space of the mixing container is divided into several small cubic units, and the coefficient of variation of the aggregate quantity or volume proportion in each unit is calculated.

8. The method for tracking aggregate motion during concrete mixing based on monocular X-ray imaging according to claim 5, characterized in that: In step S3, calculating the three-dimensional posture of the aggregate based on the PnP algorithm includes: S31. Establish a perspective-n-point problem model; S32. Apply the EPnP or P3P algorithm to solve the rotation matrix and translation vector; S33. Convert the rotation matrix into Euler angle or quaternion representation; S34. Apply the RANSAC algorithm and time series smoothing to optimize the posture solution results.

9. The method for tracking aggregate motion during concrete mixing based on monocular X-ray imaging according to claim 5, characterized in that: The steps of aggregate image recognition and tracking in step S1 are as follows: S11, extracting characteristic parameters of aggregate in X-ray projection; S12, establishing an aggregate feature fingerprint library; S13, aggregate identification based on deep learning algorithm; S14, applying a multi-target tracking algorithm to simultaneously track the motion states of multiple aggregates; S15. Record and analyze the three-dimensional motion trajectory of aggregate in the mixing space.

Images