A novel real-time measurement system for rheological parameters of concrete materials

Abstract

This invention relates to the field of building materials technology and discloses a real-time measurement system for the rheological parameters of a new concrete material. The system includes an image acquisition module for acquiring dynamic image sequences of the new concrete material in motion; an image processing module for image registration of the extracted effective analysis region sequence to obtain a preprocessed effective dynamic image sequence; a feature extraction module for inputting the preprocessed effective dynamic image sequence into a fusion model of an attention-enhanced graph convolutional neural network and a three-dimensional convolutional neural network to extract a spatiotemporal fusion feature vector characterizing the rheological behavior of concrete; and a real-time inversion module for inputting the spatiotemporal fusion feature vector into a physically constrained rheological parameter inversion model to output the yield stress and plastic viscosity of the concrete material in real time. This invention achieves real-time and accurate measurement.

Description

Technical Field

[0001] This invention relates to the field of building materials technology, and more specifically to a new real-time measurement system for the rheological parameters of concrete materials. Background Technology

[0002] Concrete rheological parameters directly determine its key construction properties such as pouring fluidity and pumping resistance. Accurate measurement is crucial to ensuring project quality. Existing measurement methods are mainly contact-based, such as rotational rheometers, which require sampling from the construction site and then laboratory testing. This not only fails to reflect the rheological characteristics of concrete in actual motion in real time, but also has the problems of damaging the original state of the material during sampling and the test results lagging behind the construction progress. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a new real-time measurement system for the rheological parameters of concrete materials.

[0004] The first aspect of this invention provides a real-time measurement system for the rheological parameters of a new concrete material, the system comprising:

[0005] The image acquisition module is used to acquire dynamic image sequences of new concrete material in motion in real time.

[0006] The image processing module is used to eliminate the interference of uneven ambient lighting and shadows on the texture features of concrete surface by using homomorphic filtering. It uses optical flow method combined with background subtraction to automatically identify and extract continuous areas of concrete flow in the image to obtain effective analysis areas. The extracted effective analysis area sequence is image registered to obtain a preprocessed effective dynamic image sequence.

[0007] The feature extraction module is used to input the preprocessed effective dynamic image sequence into the attention-enhanced graph convolutional neural network and the three-dimensional convolutional neural network fusion model to extract the spatiotemporal fusion feature vector characterizing the rheological behavior of concrete.

[0008] The real-time inversion module is used to input the spatiotemporal fusion feature vector into the rheological parameter inversion model constrained by physical information, and output the yield stress and plastic viscosity of concrete material in real time.

[0009] Optionally, in a first implementation of the first aspect of the present invention, the image processing module includes:

[0010] The conversion submodule is used to convert the acquired dynamic image sequence into a grayscale image, perform a logarithmic transformation on the grayscale image, and perform a two-dimensional Fourier transform on the logarithmically transformed image to obtain a frequency domain image.

[0011] The filtering submodule is used to suppress low-frequency components and retain high-frequency components in the frequency domain image through a Gaussian high-pass filter, and to perform a two-dimensional inverse Fourier transform on the filtered frequency domain image.

[0012] The recovery submodule is used to perform an exponential transformation on the inversely transformed image to restore it to the original grayscale value range, resulting in a grayscale image sequence that eliminates uneven lighting and shadow interference while preserving the texture features of the concrete surface.

[0013] The calculation submodule is used to calculate the optical flow field of each frame of the image using the Lucas-Kanade optical flow method. It selects 30 static images of concrete when it is not flowing, calculates the mean and standard deviation of the gray value of each pixel, and generates the initial background frame.

[0014] The first determining submodule is used to perform a difference operation between each frame image and the initial background frame to obtain a foreground difference image, and to perform binarization processing on the foreground difference image to determine the region with a gray level difference greater than a first threshold as a candidate foreground region.

[0015] The intersection operation submodule is used to perform intersection operations on regions in the optical flow field whose motion vector magnitude is greater than the second threshold and candidate foreground regions to obtain continuous regions of concrete flow.

[0016] The cropping submodule is used to crop the pixels within the minimum bounding rectangle coordinate range of each frame image based on the minimum bounding rectangle coordinates of the continuous region of concrete flow, so as to obtain an effective analysis region sequence containing only the concrete flow region.

[0017] The spatial alignment submodule is used to select the first frame of the effective analysis region sequence as the reference frame. For each frame to be registered in the effective analysis region sequence, the SIFT feature extraction algorithm is used to realize the spatial alignment between the frame to be registered and the reference frame. Registration is completed for all frames in the effective analysis region sequence to obtain the preprocessed effective dynamic image sequence.

[0018] Optionally, in a second implementation of the first aspect of the present invention, the computing submodule includes:

[0019] In the current frame of grayscale image, corners with high contrast are extracted as feature points using the Shi-Tomasi corner detection algorithm;

[0020] Valid feature points are filtered by thresholding corner response values. For feature points in two adjacent frames, a pixel grayscale change equation is established. The horizontal and vertical motion vectors of the feature points are solved by the least squares method to obtain the optical flow field of each frame.

[0021] Optionally, in a third implementation of the first aspect of the present invention, the spatial alignment submodule includes:

[0022] Within the effective regions of the frame to be registered and the reference frame, a scale space is constructed, and extreme points at each scale are detected.

[0023] The extreme points are located and filtered, the principal direction of each feature point is calculated, and feature descriptors are generated.

[0024] The similarity between the feature descriptors of the frame to be registered and the reference frame is calculated by Euclidean distance, and feature point pairs with similarity less than the third threshold are selected.

[0025] Eight pairs of feature points are randomly selected, the affine transformation matrix is ​​calculated, and the reprojection error of other feature point pairs in the affine transformation matrix is ​​verified. Feature point pairs with errors less than the fourth threshold are retained.

[0026] The optimal affine transformation matrix is ​​determined iteratively. Based on the optimal affine transformation matrix, the effective area of ​​the frame to be registered is remapped pixel by pixel. The coordinates of each pixel in the effective area of ​​the frame to be registered are converted into the corresponding coordinates in the coordinate system of the reference frame. The pixel grayscale value of the converted coordinates is calculated by bilinear interpolation to achieve spatial alignment between the frame to be registered and the reference frame.

[0027] Optionally, in a fourth implementation of the first aspect of the present invention, the feature extraction module includes:

[0028] The first convolutional submodule is used to stack the preprocessed effective dynamic image sequence along the time dimension. The input is a three-dimensional convolutional neural network containing three convolutional layers and two pooling layers. The convolution and pooling processes are repeated to obtain the apparent motion feature vector.

[0029] The second convolutional submodule is used to construct a graph structure based on the preprocessed effective dynamic image sequence and input it into a graph convolutional neural network containing two graph convolutional layers to obtain a spatial structure relationship feature vector.

[0030] The fusion submodule is used to transform the apparent motion feature vector and the spatial structure relationship feature vector into feature vectors of the same dimension through two fully connected layers, fuse them through a cross-modal attention mechanism, and generate a spatiotemporal fusion feature vector after passing through the ReLU function.

[0031] Optionally, in a fifth implementation of the first aspect of the present invention, the first convolutional submodule includes:

[0032] Using a 3×3×3 three-dimensional convolution kernel, the kernel slides across the spatial and temporal dimensions of the effective dynamic image sequence to perform convolution operations on the spatial pixel differences and temporal pixel changes of each local region, generating an initial spatiotemporal feature map.

[0033] By using a 2×2×2 max pooling layer, the maximum value within each pooling window is selected, and the above convolution and pooling process is repeated. The feature map is then transformed into an apparent motion feature vector through a global average pooling layer. The apparent motion feature vector is used to characterize the flow velocity and surface morphology changes of concrete.

[0034] Optionally, in a sixth implementation of the first aspect of the present invention, the second convolutional submodule includes:

[0035] The first layer of graph convolution kernels updates node features through weighted summation, capturing local spatial relationships. The second layer of graph convolution kernels strengthens the features of key regions through an attention mechanism. The fully connected layer integrates node features into a spatial structure relationship feature vector, which is used to characterize the distribution and contact relationship between cement paste and aggregate.

[0036] Optionally, in a seventh implementation of the first aspect of the present invention, the real-time inversion module includes:

[0037] The standardization processing submodule is used to perform Z-score standardization on the spatiotemporal fusion feature vector. The standardized feature vector is then input into the rheological parameter inversion model containing three fully connected layers, and the preliminary rheological parameter prediction values ​​are output.

[0038] The second determination submodule is used to calculate the deviation between the preliminary rheological parameter predictions and the physical constraints determined based on the Bingham fluid model, and to obtain the physical rule loss term.

[0039] The adjustment submodule is used to adjust the preliminary rheological parameter predictions based on the current physical rule loss terms, and transmits the adjusted yield stress and plastic viscosity values ​​in real time.

[0040] Optionally, in an eighth implementation of the first aspect of the present invention, the standardization processing submodule includes:

[0041] The first fully connected layer uses the ReLU activation function to perform a non-linear transformation on the standardized feature vector; the second fully connected layer uses the ReLU activation function to extract key features; and the third fully connected layer outputs preliminary rheological parameter predictions.

[0042] Optionally, in a ninth implementation of the first aspect of the present invention, a method for implementing a real-time measurement system for rheological parameters of a new concrete material is provided, the method comprising the following steps:

[0043] Real-time acquisition of dynamic image sequences of new concrete material in motion;

[0044] Homomorphic filtering is used to eliminate the interference of uneven ambient lighting and shadows on the texture features of concrete surface. Optical flow method combined with background subtraction is used to automatically identify and extract continuous areas of concrete flow in the image to obtain effective analysis areas. Image registration is performed on the extracted effective analysis area sequence to obtain a preprocessed effective dynamic image sequence.

[0045] The preprocessed effective dynamic image sequence is input into the attention-enhanced graph convolutional neural network and 3D convolutional neural network fusion model to extract the spatiotemporal fusion feature vector characterizing the rheological behavior of concrete.

[0046] The spatiotemporal fusion feature vector is input into the rheological parameter inversion model constrained by physical information, and the yield stress and plastic viscosity of concrete material are inverted and output in real time.

[0047] The technical solution provided by this invention acquires a dynamic image sequence of new concrete material in motion in real time;

[0048] Homomorphic filtering is employed to eliminate the interference of uneven ambient lighting and shadows on the texture features of concrete surfaces. Optical flow combined with background subtraction is used to automatically identify and extract continuous regions of concrete flow in the image, obtaining the effective analysis region. Image registration is performed on the extracted effective analysis region sequence to obtain a preprocessed effective dynamic image sequence. This preprocessed effective dynamic image sequence is input into a fusion model of an attention-enhanced graph convolutional neural network and a 3D convolutional neural network to extract a spatiotemporal fusion feature vector characterizing the rheological behavior of concrete. This spatiotemporal fusion feature vector is then input into a physically constrained rheological parameter inversion model to invert and output the yield stress and plasticity of the concrete material in real time. Viscosity; This invention employs non-contact image acquisition, combined with homomorphic filtering to effectively eliminate uneven lighting and shadow interference. It also utilizes optical flow and background subtraction to accurately extract the effective flow area of ​​concrete. Image registration enhances data consistency, solving the characteristic distortion problem of traditional non-contact measurements. Through an attention-enhanced graph convolution and 3D convolution fusion model, it efficiently extracts the spatiotemporal features of rheological behavior. A physically constrained inversion model ensures that the output yield stress and plastic viscosity conform to fluid mechanics laws, achieving real-time, accurate measurement without on-site sampling. This avoids the damage to material state caused by contact measurements and dynamically reflects the rheological characteristics of concrete, providing reliable data support for engineering construction quality control. Attached Figure Description

[0049] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0050] Figure 1A schematic diagram of the structure of a real-time measurement system for rheological parameters of a new concrete material provided in an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the feature extraction module provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the structure of the real-time inversion module provided in an embodiment of the present invention. Detailed Implementation

[0053] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0054] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A flowchart of a real-time measurement system for rheological parameters of a new concrete material provided in this embodiment of the invention is shown. The system includes:

[0055] The image acquisition module is used to acquire dynamic image sequences of new concrete material in motion in real time.

[0056] The image processing module is used to eliminate the interference of uneven ambient lighting and shadows on the texture features of concrete surface by using homomorphic filtering. It uses optical flow method combined with background subtraction to automatically identify and extract continuous areas of concrete flow in the image to obtain effective analysis areas. The extracted effective analysis area sequence is image registered to obtain a preprocessed effective dynamic image sequence.

[0057] The feature extraction module is used to input the preprocessed effective dynamic image sequence into the attention-enhanced graph convolutional neural network and the three-dimensional convolutional neural network fusion model to extract the spatiotemporal fusion feature vector characterizing the rheological behavior of concrete.

[0058] The real-time inversion module is used to input the spatiotemporal fusion feature vector into the rheological parameter inversion model constrained by physical information, and output the yield stress and plastic viscosity of concrete material in real time.

[0059] In this embodiment, high-speed industrial cameras are selected as image acquisition devices. Two to four cameras are evenly arranged around the concrete movement device to ensure that all camera lenses cover the complete area of ​​the concrete movement. Next, each camera is calibrated using a standard calibration plate. By capturing images of the calibration plate at different angles, the camera's intrinsic and extrinsic parameters are calculated to eliminate lens distortion and viewing angle deviation, ensuring consistent spatial accuracy of subsequently acquired images. Camera parameters are set according to the expected movement speed of the concrete: a frame rate of 50-200fps to capture continuous motion details, a resolution of 1920×1080 to preserve material composition details, and an exposure time of 10-50μs to avoid motion blur. Simultaneously, the camera triggering method is set to external synchronization, synchronized with the control of the concrete movement device. The system establishes a signal connection to ensure that the camera startup is synchronized with the start of concrete movement, preventing time misalignment in the image sequence. It then activates the concrete movement device, such as driving the rheometer cylinder to rotate, causing flow within the concrete and simultaneously triggering all cameras to begin acquisition. The cameras continuously capture images at a set frame rate, converting each instantaneous state of the concrete flow process into images, forming a continuous dynamic image sequence. The acquisition duration is set to 10-30 seconds based on measurement requirements, ensuring coverage of the complete stage from the start of concrete flow to stable flow. The image sequences acquired by each camera are stored on a local server according to naming rules, using TIFF format to preserve original pixel information. Simultaneously, a trigger signal is sent to the feature extraction module, transmitting the storage path and basic parameters of the image sequence.

[0060] In this embodiment, the image processing module includes:

[0061] The conversion submodule is used to convert the acquired dynamic image sequence into a grayscale image, perform a logarithmic transformation on the grayscale image, and perform a two-dimensional Fourier transform on the logarithmically transformed image to obtain a frequency domain image.

[0062] The filtering submodule is used to suppress low-frequency components and retain high-frequency components in the frequency domain image through a Gaussian high-pass filter, and to perform a two-dimensional inverse Fourier transform on the filtered frequency domain image.

[0063] The recovery submodule is used to perform an exponential transformation on the inversely transformed image to restore it to the original grayscale value range, resulting in a grayscale image sequence that eliminates uneven lighting and shadow interference while preserving the texture features of the concrete surface.

[0064] The calculation submodule is used to calculate the optical flow field of each frame of the image using the Lucas-Kanade optical flow method. It selects 30 static images of concrete when it is not flowing, calculates the mean and standard deviation of the gray value of each pixel, and generates the initial background frame.

[0065] The first determining submodule is used to perform a difference operation between each frame image and the initial background frame to obtain a foreground difference image, and to perform binarization processing on the foreground difference image to determine the region with a gray level difference greater than a first threshold as a candidate foreground region.

[0066] The intersection operation submodule is used to perform intersection operations on regions in the optical flow field whose motion vector magnitude is greater than the second threshold and candidate foreground regions to obtain continuous regions of concrete flow.

[0067] The cropping submodule is used to crop the pixels within the minimum bounding rectangle coordinate range of each frame image based on the minimum bounding rectangle coordinates of the continuous region of concrete flow, so as to obtain an effective analysis region sequence containing only the concrete flow region.

[0068] The spatial alignment submodule is used to select the first frame of the effective analysis region sequence as the reference frame. For each frame to be registered in the effective analysis region sequence, the SIFT feature extraction algorithm is used to realize the spatial alignment between the frame to be registered and the reference frame. Registration is completed for all frames in the effective analysis region sequence to obtain the preprocessed effective dynamic image sequence.

[0069] In this embodiment, the calculation submodule includes: extracting high-contrast corner points as feature points in the current frame grayscale image using the Shi-Tomasi corner detection algorithm; filtering effective feature points by using a corner response value threshold; establishing pixel grayscale change equations for feature points in two adjacent frames; and solving the horizontal and vertical motion vectors of the feature points using the least squares method to obtain the optical flow field of each frame image.

[0070] In this embodiment, when performing Shi-Tomasi corner detection in the current frame's grayscale image, the image is Gaussian smoothed using a 3×3 Gaussian kernel. Each pixel in the image is traversed, and the grayscale value of that pixel and its eight surrounding pixels is weighted and averaged using Gaussian weights, replacing the original pixel's grayscale value to reduce image noise interference with subsequent corner detection. Next, the Sobel operator is used to calculate the gradient maps of the image in the x and y directions. For each pixel, the Sobel_x operator is used to calculate its horizontal grayscale change rate, and the Sobel_y operator is used to calculate its vertical grayscale change rate. Subsequently, for each pixel... A 2×2 autocorrelation matrix is ​​constructed at the location, which consists of the sum of squares of the gradients in the x-direction, the sum of the products of the gradients in the x and y directions, and the sum of squares of the gradients in the y-direction. The matrix elements are obtained by accumulating the gradient values ​​within a 3×3 local window. The two eigenvalues ​​of the autocorrelation matrix are solved, and the smaller eigenvalue is taken as the corner response value of the pixel. Finally, a threshold for the corner response value is set, and pixels with response values ​​greater than the threshold are selected as candidate feature points. At the same time, non-maximum suppression is applied to the candidate feature points, and only the pixel with the largest response value is retained within a 5×5 local window to avoid dense overlap of feature points. Finally, a set of effective feature points with high contrast is obtained.

[0071] For each effective feature point in two adjacent frames, a 7×7 local window is defined. Based on the assumption of local gray-level invariance of the optical flow field, that is, the pixel gray-level values ​​in the local area around the same feature point do not change with time between adjacent frames, a pixel gray-level change equation is established for each pixel in the local window to form an overdetermined linear equation system. The overdetermined equation system is solved by the least squares method to obtain the horizontal and vertical motion vectors of the feature points. The motion vectors of all effective feature points are matched one-to-one with their coordinate positions in the image to form a matrix containing the motion trend of each pixel. This matrix is ​​the optical flow field of each frame of the image.

[0072] In this embodiment, the spatial alignment submodule includes:

[0073] Within the effective regions of the frame to be registered and the reference frame, a scale space is constructed, and extreme points at each scale are detected.

[0074] The extreme points are located and filtered, the principal direction of each feature point is calculated, and feature descriptors are generated.

[0075] The similarity between the feature descriptors of the frame to be registered and the reference frame is calculated by Euclidean distance, and feature point pairs with similarity less than the third threshold are selected.

[0076] Eight pairs of feature points are randomly selected, the affine transformation matrix is ​​calculated, and the reprojection error of other feature point pairs in the affine transformation matrix is ​​verified. Feature point pairs with errors less than the fourth threshold are retained.

[0077] The optimal affine transformation matrix is ​​determined iteratively. Based on the optimal affine transformation matrix, the effective area of ​​the frame to be registered is remapped pixel by pixel. The coordinates of each pixel in the effective area of ​​the frame to be registered are converted into the corresponding coordinates in the coordinate system of the reference frame. The pixel grayscale value of the converted coordinates is calculated by bilinear interpolation to achieve spatial alignment between the frame to be registered and the reference frame.

[0078] In this embodiment, please refer to Figure 2 The feature extraction module includes:

[0079] The first convolutional submodule is used to stack the preprocessed effective dynamic image sequence along the time dimension. The input is a three-dimensional convolutional neural network containing three convolutional layers and two pooling layers. The convolution and pooling processes are repeated to obtain the apparent motion feature vector.

[0080] The second convolutional submodule is used to construct a graph structure based on the preprocessed effective dynamic image sequence and input it into a graph convolutional neural network containing two graph convolutional layers to obtain a spatial structure relationship feature vector.

[0081] The fusion submodule is used to transform the apparent motion feature vector and the spatial structure relationship feature vector into feature vectors of the same dimension through two fully connected layers, fuse them through a cross-modal attention mechanism, and generate a spatiotemporal fusion feature vector after passing through the ReLU function.

[0082] In this embodiment, the first convolutional submodule includes:

[0083] Using a 3×3×3 three-dimensional convolution kernel, the kernel slides across the spatial and temporal dimensions of the effective dynamic image sequence to perform convolution operations on the spatial pixel differences and temporal pixel changes of each local region, generating an initial spatiotemporal feature map.

[0084] By using a 2×2×2 max pooling layer, the maximum value within each pooling window is selected, and the above convolution and pooling process is repeated. The feature map is then transformed into an apparent motion feature vector through a global average pooling layer. The apparent motion feature vector is used to characterize the flow velocity and surface morphology changes of concrete.

[0085] In this embodiment, the second convolution submodule includes:

[0086] The first layer of graph convolution kernels updates node features through weighted summation, capturing local spatial relationships. The second layer of graph convolution kernels strengthens the features of key regions through an attention mechanism. The fully connected layer integrates node features into a spatial structure relationship feature vector, which is used to characterize the distribution and contact relationship between cement paste and aggregate.

[0087] In this embodiment, cross-modal attention fusion generates a spatiotemporal fusion feature vector. A cross-modal attention module is constructed, and the apparent motion feature vector and the spatial structure relation feature vector are respectively input into two fully connected layers, transforming them into feature vectors of the same dimension. The correlation matrix between the apparent motion feature vector and the spatial structure relation feature vector is calculated. Softmax normalization is performed on each row of the correlation matrix to obtain the attention weight of the spatial structure relation feature vector on the apparent motion feature vector, and similarly, the attention weight of the apparent motion feature vector on the spatial structure relation feature vector is obtained. The apparent motion feature vector is multiplied by its corresponding attention weight, and the spatial structure relation feature vector is multiplied by its corresponding attention weight. The sum of the two is then passed through an activation function to generate the spatiotemporal fusion feature vector.

[0088] In this embodiment, please refer to Figure 3 The real-time inversion module includes:

[0089] The standardization processing submodule is used to perform Z-score standardization on the spatiotemporal fusion feature vector. The standardized feature vector is then input into the rheological parameter inversion model containing three fully connected layers, and the preliminary rheological parameter prediction values ​​are output.

[0090] The second determination submodule is used to calculate the deviation between the preliminary rheological parameter predictions and the physical constraints determined based on the Bingham fluid model, and to obtain the physical rule loss term.

[0091] The adjustment submodule is used to adjust the preliminary rheological parameter predictions based on the current physical rule loss terms, and transmits the adjusted yield stress and plastic viscosity values ​​in real time.

[0092] In this embodiment, the standardization processing submodule includes:

[0093] The first fully connected layer uses the ReLU activation function to perform a non-linear transformation on the standardized feature vector; the second fully connected layer uses the ReLU activation function to extract key features; and the third fully connected layer outputs preliminary rheological parameter predictions.

[0094] In this embodiment, physical rules are determined based on the Bingham fluid model: yield stress ≥ 0, plastic viscosity ≥ 0, and shear stress = yield stress + plastic viscosity × shear rate. The deviation between the preliminary predicted value and the physical rules is calculated. If the yield stress < 0, the deviation is |yield stress|; if the plastic viscosity < 0, the deviation is |plastic viscosity|. The square of the difference between the actual shear stress and the predicted shear stress is calculated. The predicted shear stress is the predicted yield stress + predicted plastic viscosity × shear rate, which is used as the force balance deviation. The above deviations are summed to obtain the physical rule loss term. In the real-time application stage, the pre-trained inversion model weights are called, and the preliminary predicted value is fine-tuned according to the current physical rule loss term: if the yield stress < 0, it is adjusted to 0; if the plastic viscosity < 0, it is adjusted to 0; if the force balance deviation > the threshold, the predicted value is corrected according to the deviation ratio, such as the predicted yield stress = predicted yield stress + deviation × 0.1. The fine-tuned yield stress and plastic viscosity values ​​are transmitted to the monitoring interface for display in real time and stored in the database according to the timestamp, completing a real-time inversion of rheological parameters.

[0095] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A new real-time measurement system for rheological parameters of concrete materials, characterized in that, The system includes: The image acquisition module is used to acquire dynamic image sequences of new concrete material in motion in real time. The image processing module is used to eliminate the interference of uneven ambient lighting and shadows on the texture features of concrete surface by using homomorphic filtering. It uses optical flow method combined with background subtraction to automatically identify and extract continuous areas of concrete flow in the image to obtain effective analysis areas. The extracted effective analysis area sequence is image registered to obtain a preprocessed effective dynamic image sequence. The feature extraction module is used to input the preprocessed effective dynamic image sequence into the attention-enhanced graph convolutional neural network and the three-dimensional convolutional neural network fusion model to extract the spatiotemporal fusion feature vector characterizing the rheological behavior of concrete. The real-time inversion module is used to input the spatiotemporal fusion feature vector into the rheological parameter inversion model constrained by physical information, and output the yield stress and plastic viscosity of concrete material in real time. The image processing module includes: The conversion submodule is used to convert the acquired dynamic image sequence into a grayscale image, perform a logarithmic transformation on the grayscale image, and perform a two-dimensional Fourier transform on the logarithmically transformed image to obtain a frequency domain image. The filtering submodule is used to suppress low-frequency components and retain high-frequency components in the frequency domain image through a Gaussian high-pass filter, and to perform a two-dimensional inverse Fourier transform on the filtered frequency domain image. The recovery submodule is used to perform an exponential transformation on the inversely transformed image to restore it to the original grayscale value range, resulting in a grayscale image sequence that eliminates uneven lighting and shadow interference while preserving the texture features of the concrete surface. The calculation submodule is used to calculate the optical flow field of each frame of the image using the Lucas-Kanade optical flow method. It selects 30 static images of concrete when it is not flowing, calculates the mean and standard deviation of the gray value of each pixel, and generates the initial background frame. The first determining submodule is used to perform a difference operation between each frame image and the initial background frame to obtain a foreground difference image, and to perform binarization processing on the foreground difference image to determine the region with a gray level difference greater than a first threshold as a candidate foreground region. The intersection operation submodule is used to perform intersection operations on regions in the optical flow field whose motion vector magnitude is greater than the second threshold and candidate foreground regions to obtain continuous regions of concrete flow. The cropping submodule is used to crop the pixels within the minimum bounding rectangle coordinate range of each frame image based on the minimum bounding rectangle coordinates of the continuous region of concrete flow, so as to obtain an effective analysis region sequence containing only the concrete flow region. The spatial alignment submodule is used to select the first frame of the effective analysis region sequence as the reference frame. For each frame to be registered in the effective analysis region sequence, the SIFT feature extraction algorithm is used to realize the spatial alignment between the frame to be registered and the reference frame. Registration is completed for all frames in the effective analysis region sequence to obtain the preprocessed effective dynamic image sequence.

2. The real-time measurement system for rheological parameters of a new concrete material as described in claim 1, characterized in that, The computation submodule includes: In the current frame of grayscale image, corners with high contrast are extracted as feature points using the Shi-Tomasi corner detection algorithm; Valid feature points are filtered by thresholding corner response values. For feature points in two adjacent frames, a pixel grayscale change equation is established. The horizontal and vertical motion vectors of the feature points are solved by the least squares method to obtain the optical flow field of each frame.

3. The real-time measurement system for rheological parameters of a new concrete material as described in claim 1, characterized in that, The spatial alignment submodule includes: Within the effective regions of the frame to be registered and the reference frame, a scale space is constructed, and extreme points at each scale are detected. The extreme points are located and filtered, the principal direction of each feature point is calculated, and feature descriptors are generated. The similarity between the feature descriptors of the frame to be registered and the reference frame is calculated by Euclidean distance, and feature point pairs with similarity less than the third threshold are selected. Eight pairs of feature points are randomly selected, the affine transformation matrix is ​​calculated, and the reprojection error of other feature point pairs in the affine transformation matrix is ​​verified. Feature point pairs with errors less than the fourth threshold are retained. The optimal affine transformation matrix is ​​determined iteratively. Based on the optimal affine transformation matrix, the effective area of ​​the frame to be registered is remapped pixel by pixel. The coordinates of each pixel in the effective area of ​​the frame to be registered are converted into the corresponding coordinates in the coordinate system of the reference frame. The pixel grayscale value of the converted coordinates is calculated by bilinear interpolation to achieve spatial alignment between the frame to be registered and the reference frame.

4. The real-time measurement system for rheological parameters of a new concrete material as described in claim 1, characterized in that, The feature extraction module includes: The first convolutional submodule is used to stack the preprocessed effective dynamic image sequence along the time dimension. The input is a three-dimensional convolutional neural network containing three convolutional layers and two pooling layers. The convolution and pooling processes are repeated to obtain the apparent motion feature vector. The second convolutional submodule is used to construct a graph structure based on the preprocessed effective dynamic image sequence and input it into a graph convolutional neural network containing two graph convolutional layers to obtain a spatial structure relationship feature vector. The fusion submodule is used to transform the apparent motion feature vector and the spatial structure relationship feature vector into feature vectors of the same dimension through two fully connected layers, fuse them through a cross-modal attention mechanism, and generate a spatiotemporal fusion feature vector after passing through the ReLU function.

5. The real-time measurement system for rheological parameters of a new concrete material as described in claim 4, characterized in that, The first convolutional submodule includes: Using a 3×3×3 three-dimensional convolution kernel, the kernel slides across the spatial and temporal dimensions of the effective dynamic image sequence to perform convolution operations on the spatial pixel differences and temporal pixel changes of each local region, generating an initial spatiotemporal feature map. By using a 2×2×2 max pooling layer, the maximum value within each pooling window is selected, and the above convolution and pooling process is repeated. The feature map is then transformed into an apparent motion feature vector through a global average pooling layer. The apparent motion feature vector is used to characterize the flow velocity and surface morphology changes of concrete.

6. The real-time measurement system for rheological parameters of a new concrete material as described in claim 4, characterized in that, The second convolutional submodule includes: The first layer of graph convolution kernels updates node features through weighted summation, capturing local spatial relationships. The second layer of graph convolution kernels strengthens the features of key regions through an attention mechanism. The fully connected layer integrates node features into a spatial structure relationship feature vector, which is used to characterize the distribution and contact relationship between cement paste and aggregate.

7. The real-time measurement system for rheological parameters of a new concrete material as described in claim 1, characterized in that, The real-time inversion module includes: The standardization processing submodule is used to perform Z-score standardization on the spatiotemporal fusion feature vector. The standardized feature vector is then input into the rheological parameter inversion model containing three fully connected layers, and the preliminary rheological parameter prediction values ​​are output. The second determination submodule is used to calculate the deviation between the preliminary rheological parameter predictions and the physical constraints determined based on the Bingham fluid model, and to obtain the physical rule loss term. The adjustment submodule is used to adjust the preliminary rheological parameter predictions based on the current physical rule loss terms, and transmits the adjusted yield stress and plastic viscosity values ​​in real time.

8. The real-time measurement system for rheological parameters of a new concrete material as described in claim 7, characterized in that, The standardization processing submodule includes: The first fully connected layer uses the ReLU activation function to perform a non-linear transformation on the standardized feature vector; the second fully connected layer uses the ReLU activation function to extract key features; and the third fully connected layer outputs preliminary rheological parameter predictions.

9. A method for implementing a real-time measurement system for the rheological parameters of a new concrete material as described in claim 1, characterized in that, The method includes the following steps: Real-time acquisition of dynamic image sequences of new concrete material in motion; Homomorphic filtering is used to eliminate the interference of uneven ambient lighting and shadows on the texture features of concrete surface. Optical flow method combined with background subtraction is used to automatically identify and extract continuous areas of concrete flow in the image to obtain effective analysis areas. Image registration is performed on the extracted effective analysis area sequence to obtain a preprocessed effective dynamic image sequence. The preprocessed effective dynamic image sequence is input into the attention-enhanced graph convolutional neural network and 3D convolutional neural network fusion model to extract the spatiotemporal fusion feature vector characterizing the rheological behavior of concrete. The spatiotemporal fusion feature vector is input into the rheological parameter inversion model constrained by physical information, and the yield stress and plastic viscosity of concrete material are inverted and output in real time.

Images