A computer vision-based method and system for detecting the workability of concrete

Abstract

This invention discloses a computer vision-based method and system for detecting the workability of concrete. It employs a non-contact visual sensing approach to simultaneously acquire video sequences of freshly mixed concrete in a free-flowing state at both the macroscopic slump flow scale and the microscopic particle migration scale. An improved spatiotemporal feature pyramid network is used to extract multi-scale morphological features from the video sequences, obtaining a multi-scale visual feature set including the flow profile evolution rate, surface ripple spectrum, aggregate suspension trajectory, and paste coating state. A graph neural network model integrating prior knowledge of fluid mechanics is constructed to map discrete visual features to continuous rheological parameters, outputting workability indicators such as slump, spread, yield stress, and plastic viscosity in real time. This invention significantly improves the level of intelligent quality control in concrete production.

Description

Technical Field

[0001] This invention relates to the field of concrete quality testing technology, and in particular to a method and system for testing the workability of concrete based on computer vision. Background Technology

[0002] Concrete workability refers to the technological properties of fresh concrete during mixing, transportation, pouring, and vibration, and is a key indicator for concrete quality control. Good or bad workability directly affects construction efficiency, cost control, and ultimately, structural performance and durability.

[0003] Currently, the testing of concrete workability mainly relies on traditional methods such as slump tests and spread tests. These methods have the following problems: First, the testing process requires manual operation, which is time-consuming and labor-intensive, and it is difficult to achieve continuous online monitoring; second, the test results depend on the experience and skills of the operators, which is highly subjective and has poor repeatability; third, traditional methods can only obtain single indicators such as slump and spread, and cannot comprehensively characterize the rheological properties of concrete, such as yield stress and plastic viscosity; in addition, manual testing has the problem of detection lag, which makes it impossible to achieve real-time quality control of the concrete production process.

[0004] Therefore, there is an urgent need for a method and system that can detect the workability of concrete in a non-contact, automated, and real-time manner to improve detection efficiency and accuracy and achieve intelligent quality control in concrete production. Summary of the Invention

[0005] This invention provides a computer vision-based method and system for detecting the workability of concrete, which solves the problems of existing concrete workability detection methods such as reliance on manual labor, low efficiency, single indicators, and inability to monitor in real time, and realizes non-contact, automated, and real-time detection of the workability indicators of fresh concrete.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for detecting the workability of concrete based on computer vision, the method comprising the following steps: Step 1: Using a non-contact visual sensing method, video sequences of freshly mixed concrete in a free-flowing state are simultaneously acquired from both the macroscopic slump flow scale and the microscopic particle migration scale.

[0007] The first video sequence of macroscopically acquired data on the overall flow morphology changes of freshly mixed concrete. The second video sequence was acquired at the microscale to obtain the particle motion state on the surface of freshly mixed concrete. .

[0008] Step 2: An improved spatiotemporal feature pyramid network is used to extract multi-scale morphological features from the first video sequence and the second video sequence to obtain a multi-scale visual feature set.

[0009] The multi-scale visual feature set ,in This is the eigenvector representing the evolution rate of the flow profile. The surface ripple spectrum feature vector, This represents the feature vector of the aggregate suspension trajectory. This is the feature vector of the slurry coating state.

[0010] Step 3: Construct a graph neural network model that integrates prior knowledge of fluid mechanics, and incorporate the multi-scale visual feature set. The discrete visual features are mapped to continuous rheological parameters, and the workability index vector of fresh concrete is calculated and output in real time based on the rheological parameters.

[0011] The workability index vector of concrete ,in Slump For scalability, For yield stress, It is a plastic viscosity.

[0012] Furthermore, the macroscopic data acquisition is performed using a first industrial camera, taking images from a top-down angle directly above the concrete flow plane, with a frame rate of [missing information]. Sampling time interval Obtain the moment when freshly mixed concrete begins to slump. Until the flow stops The first video sequence of the entire process: .

[0013] The microscale data acquisition was performed using a second industrial camera, taking images from a lateral oblique angle at the concrete flow front edge, with a frame rate of [missing information]. Sampling time interval The second video sequence of particle movement on the concrete surface was obtained. ;in, To ensure the temporal resolution of the movement of micro-particles; the first industrial camera and the second industrial camera are synchronized frame by time synchronization signal.

[0014] Furthermore, the improved spatiotemporal feature pyramid network includes a spatial feature pyramid module, a temporal feature pyramid module, a feature fusion module, and a feature decoupling module.

[0015] The spatial feature pyramid module uses multi-scale convolutional kernels to extract feature maps at different spatial resolutions. The set of scale factors for the spatial feature pyramid is as follows: , among which, the Each scaling factor: ;for Input image at time , No. The layer space features are represented as follows: ;in Indicates the downsampling factor as Pooling operations, Indicates the first Layered convolution operations; multi-scale spatial feature fusion is performed as follows: The temporal feature pyramid module uses three-dimensional convolution and temporal attention mechanisms to extract motion features between video frames; for example, the temporal feature pyramid module uses three-dimensional convolution and temporal attention mechanisms to extract motion features between video frames. Centered on time, with a time window radius of A sequence of consecutive frames of images: The time characteristics are represented as follows: ;in The radius of the time window. This represents the sampling time interval for the corresponding video sequence.

[0016] The feature fusion module adaptively fuses spatial and temporal features. The fusion feature at any given time is represented as follows: ;in and For learnable fusion weights, Indicates feature concatenation operation; The feature decoupling module will fuse features. The features are mapped to four categories of visual features through four parallel feature decoupling branches:

[0017] in , , , These are the feature decoupling mapping functions for flow profile, surface ripples, aggregate trajectory, and slurry coating, respectively.

[0018] Furthermore, the flow profile evolution rate feature vector Specific extraction methods include: Semantic segmentation is performed on each frame of the first video sequence to obtain a binary mask of the concrete flow region; the boundary contour of the binary mask is extracted, and the area enclosed by the contour and the diameter of the equivalent circle are calculated; Calculate the flow profile evolution rate at time step 1, calculate the flow profile evolution acceleration, and construct the flow profile evolution rate eigenvector:

[0019] in For the maximum evolution rate, The average evolution rate, For the average evolution acceleration, The initial diameter, This is the final stable diameter.

[0020] Furthermore, the surface ripple spectrum feature vector Specific extraction methods include: exist Multiple sampling windows are selected within the concrete flow area at any given time. A two-dimensional Fourier transform is performed on the grayscale image within each window to obtain the spatial spectrum. The isotropic power spectral density is calculated. Extraction is performed from the power spectral density. The surface ripple spectrum feature vector is constructed from the ripple feature parameters at time step: ;in , , These are the time averages of the main frequency, spectral energy, and spectral entropy, respectively. , These are the time standard deviations of spectral energy and spectral entropy, respectively.

[0021] Furthermore, the aggregate suspension trajectory feature vector Specific extraction methods include: The second video sequence is processed using an object detection network. The coarse aggregate particles in the sample are identified and located to obtain... The set of aggregate particle positions in each frame of the image at any given time is used. A multi-target tracking algorithm is employed to correlate the aggregate particles across frames, constructing a set of aggregate particle motion trajectories. The instantaneous velocity, instantaneous motion direction, and coefficient of variation of each individual aggregate particle are calculated. A feature vector of the aggregate suspension trajectory is then constructed. ;in For the maximum average speed, The average direction of motion, This represents the standard deviation of the direction of motion.

[0022] Furthermore, the slurry coating state feature vector The specific extraction methods include: converting the color space of the concrete surface image from RGB to Lab color space, extracting the luminance channel L and chromaticity channels a and b; using an adaptive threshold segmentation method to distinguish between slurry areas and exposed aggregate areas, calculating slurry coverage and slurry coating uniformity indices, extracting slurry gloss features and slurry color consistency features, and constructing a slurry coating state feature vector. ;in, These are the slurry coverage, slurry coating uniformity index, slurry gloss characteristics, and slurry color consistency characteristics.

[0023] Furthermore, constructing a graph neural network model that integrates prior knowledge of fluid mechanics includes: A feature map is constructed from a set of multi-scale visual features, where the node set These correspond to the flow profile feature node, ripple spectrum feature node, aggregate trajectory feature node, and slurry coating feature node, respectively. Node feature matrix Edge set Based on the physical relationships between features, physical constraints between rheological parameters are introduced: Where τ is the shear stress, For yield stress, Plastic viscosity, The shear rate and the constraint relationships are encoded as edge weight matrices of a graph neural network. A physically constrained graph attention mechanism is used for feature aggregation, mapping the final layer output features of the graph neural network into a rheological parameter vector through a fully connected layer: .

[0024] Furthermore, a sliding time window is used to process continuous video frames in real time, and the output performance index vector is updated at a frequency of no less than 1Hz.

[0025] In a second aspect, the present invention provides a computer vision-based concrete workability detection system for performing the method of the first aspect. The system includes: a dual-scale video acquisition module, a video preprocessing module, a spatiotemporal feature pyramid network module, a graph neural network mapping module, and a workability index output module.

[0026] The dual-scale video acquisition module includes a first industrial camera for macroscopic collapse flow acquisition and a second industrial camera for microscopic particle migration acquisition, as well as a time synchronization controller; the first industrial camera is positioned directly above the concrete flow plane, and its frame rate is [missing information]. The second industrial camera is positioned laterally at the leading edge of the flow, and its frame rate is [missing information]. ,and The two cameras achieve frame synchronization acquisition through the time synchronization controller.

[0027] The video preprocessing module is used to process the acquired video sequence. and Denoising, distortion correction, and spatiotemporal alignment are performed.

[0028] The spatiotemporal feature pyramid network module is used to extract multi-scale visual feature sets from the preprocessed video sequence. It includes a flow profile evolution rate feature extraction unit, a surface ripple spectrum feature extraction unit, an aggregate suspension trajectory feature extraction unit, and a slurry coating state feature extraction unit.

[0029] The graph neural network mapping module is used to fuse Bingham's prior knowledge of fluid dynamics and to integrate the multi-scale visual feature set. Mapped to rheological parameter vector It includes graph construction units, physical constraint injection units, message passing units, and rheological parameter regression units.

[0030] The performance index output module is used to output the rheological parameter vector. Calculate and output slump in real time Scalability Yield stress and plastic viscosity The output update frequency is no less than 1Hz, and the results are displayed visually through the display interface.

[0031] Compared with the prior art, the beneficial effects of the present invention are: This invention employs a dual-scale visual sensing approach, simultaneously acquiring concrete flow information at both the macroscopic flow morphology and microscopic particle motion scales, thus achieving a comprehensive characterization of concrete workability. Multi-scale morphological features are extracted through a spatiotemporal feature pyramid network, including key features such as flow profile evolution rate, surface ripple spectrum, aggregate suspension trajectory, and slurry coating state. A graph neural network model incorporating prior knowledge of fluid mechanics is introduced to establish a mapping relationship between visual features and rheological parameters, enabling real-time, non-contact detection of multiple workability indicators such as slump, spread, yield stress, and plastic viscosity, significantly improving the efficiency and accuracy of concrete workability testing. Attached Figure Description

[0032] Figure 1 This is a flowchart of a computer vision-based concrete workability testing method according to the present invention. Figure 2 This is a schematic diagram of the components of a computer vision-based concrete workability testing system according to the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0034] Example 1 like Figure 1 As shown, this invention provides a computer vision-based method for detecting the workability of concrete, characterized in that the method includes the following steps: Step 1: Using a non-contact visual sensing method, video sequences of freshly mixed concrete in a free-flowing state are simultaneously acquired from both the macroscopic slump flow scale and the microscopic particle migration scale.

[0035] The first video sequence of macroscopically acquired data on the overall flow morphology changes of freshly mixed concrete. The second video sequence was acquired at the microscale to obtain the particle motion state on the surface of freshly mixed concrete. .

[0036] The first industrial camera uses an industrial-grade CMOS camera with a resolution of 1920×1080 pixels and a shooting frame rate of... Set to 30fps, with a 60° field of view, and installed 1.2m above the concrete flow plane to ensure complete capture of the entire concrete flow process within a diameter of 800mm.

[0037] The second industrial camera is a high-speed industrial camera with a resolution of 2048×1536 pixels, and the shooting frame rate is... Set to 60fps, with a 50mm focal length, mounted 0.5m from the leading edge of the flow path, and a 45° pitch angle, the system is designed to clearly capture the movement of aggregate particles larger than 5mm on the concrete surface. A hardware-triggered time synchronization controller ensures synchronization accuracy better than 1ms, guaranteeing consistent time references between the two video streams. The shooting environment utilizes uniform illumination from LED surface light sources, with an illuminance of at least 1000 lux, to eliminate the impact of shadows and reflections on image quality.

[0038] The macroscopic data acquisition was performed using a first industrial camera, taking images from a top-down view directly above the concrete flow plane, with a frame rate of [missing information]. Sampling time interval Obtain the moment when freshly mixed concrete begins to collapse. Until the flow stops The first video sequence of the entire process .

[0039] The microscale data acquisition was performed using a second industrial camera, taking images from a lateral oblique angle at the concrete flow front edge, with a frame rate of [missing information]. Sampling time interval The second video sequence of particle movement on the concrete surface was obtained. ;in, To ensure the temporal resolution of the movement of micro-particles; the first industrial camera and the second industrial camera are synchronized frame by time synchronization signal.

[0040] Step 2: An improved spatiotemporal feature pyramid network is used to extract multi-scale morphological features from the first video sequence and the second video sequence to obtain a multi-scale visual feature set.

[0041] The multi-scale visual feature set ,in This is the eigenvector representing the evolution rate of the flow profile. The surface ripple spectrum feature vector, This represents the feature vector of the aggregate suspension trajectory. This is the feature vector of the slurry coating state.

[0042] The improved spatiotemporal feature pyramid network includes a spatial feature pyramid module, a temporal feature pyramid module, a feature fusion module, and a feature decoupling module.

[0043] The spatial feature pyramid module uses multi-scale convolutional kernels to extract feature maps at different spatial resolutions. The set of scale factors for the spatial feature pyramid is as follows: , among which, the Each scaling factor: ;for Input image at time , No. The layer space features are represented as follows: ;in Indicates the downsampling factor as Pooling operations, Indicates the first Layered convolution operations; multi-scale spatial feature fusion is performed as follows: .

[0044] In this embodiment, the number of layers of the spatial feature pyramid Setting it to 4 corresponds to the following set of scale factors: These correspond to feature maps at the original resolution, 1 / 2 resolution, 1 / 4 resolution, and 1 / 8 resolution, respectively.

[0045] Each convolutional operation uses a 3×3 kernel with 64, 128, 256, and 512 channels respectively, and ReLU is used as the activation function. Max pooling is used for pooling, with the stride being the same as the downsampling factor. For multi-scale feature fusion, the low-resolution feature map is first upsampled to its original resolution using bilinear interpolation, and then concatenated along the channel dimension to obtain a 960-dimensional spatial feature vector. .

[0046] The temporal feature pyramid module uses three-dimensional convolution and temporal attention mechanisms to extract motion features between video frames; for example, the temporal feature pyramid module uses three-dimensional convolution and temporal attention mechanisms to extract motion features between video frames; Centered on time, with a time window radius of A sequence of consecutive frames of images: The time characteristics are represented as follows: ;in The radius of the time window. This represents the sampling time interval for the corresponding video sequence.

[0047] Time window radius The threshold is set to 5, meaning that 11 consecutive frames (5 frames before and 5 frames after the current frame) are taken as input. The 3D convolutional kernel size is 3×3×3, with a temporal stride of 1 and a spatial stride of 2. The temporal attention mechanism uses a multi-head self-attention structure with 8 attention heads and a hidden layer dimension of 256.

[0048] The feature fusion module adaptively fuses spatial and temporal features. The fusion feature at any given time is represented as follows: ,in and For learnable fusion weights, This indicates a feature concatenation operation. Fusion weights. and It is obtained through adaptive learning via an attention mechanism, and the specific calculation method is as follows: First, spatial features and time characteristics The weights are compressed to 64 dimensions through fully connected layers, then concatenated and passed through two fully connected layers and Softmax normalization to output a two-dimensional weight vector. ,satisfy .

[0049] In the training process of this embodiment, the initial stage and All values ​​are initialized to 0.5. After training and convergence, the time feature weights exhibit dynamic changes for different flow stages: in the early stages of flow, the time feature weights... The spatial feature weights are relatively large (approximately 0.6-0.7) to capture rapidly changing motion information; as the flow tends to stabilize in the later stages, the spatial feature weights are... The feature size is relatively large (approximately 0.6-0.7) to accurately describe the final form; the fused feature dimensions are 1472 (960+512).

[0050] The feature decoupling module will fuse features. The features are mapped to four categories of visual features through four parallel feature decoupling branches: ;in , , , These are the feature decoupling mapping functions for flow profile, surface ripples, aggregate trajectory, and slurry coating, respectively.

[0051] The flow profile evolution rate feature vector Specific extraction methods include: For the first video sequence Each frame of the image Perform semantic segmentation to obtain a binary mask for the concrete flow region. Extract the boundary contour of the binary mask and calculate the area enclosed by the contour. And the equivalent circle diameter: .

[0052] The semantic segmentation network adopts the U-Net architecture, with the encoder using a pre-trained ResNet-50 as the backbone network and the decoder employing a layer-wise upsampling and skip connection structure. The training dataset contains 5000 labeled images of concrete flow, with label categories of concrete region and background region.

[0053] The segmentation accuracy achieved an average intersection-union ratio (mIoU) of 95.6% on the test set, and a binary mask was obtained. Subsequently, morphological opening operations (5×5 structuring element) were used to remove noisy regions, and closing operations (7×7 structuring element) were used to fill internal holes. Boundary contour extraction employed the Canny edge detection algorithm, with a gradient threshold of 50 at the lower limit and 150 at the upper limit. For irregular contours, the equivalent circle diameter was calculated using an area equivalence method, whereby the area enclosed by the actual contour was equivalent to the diameter of a circle with the same area.

[0054] calculate rate of evolution of the flow profile at any moment : .

[0055] Calculate the acceleration of the flow profile evolution : .

[0056] Construct the feature vector of the flow profile evolution rate: ;in For the maximum evolution rate, The average evolution rate, For the average evolution acceleration, The initial diameter, This is the final stable diameter.

[0057] The surface ripple spectrum feature vector Specific extraction methods include: exist Multiple sampling windows are selected within the concrete flow area at any given time. A two-dimensional Fourier transform is performed on the grayscale image within each window to obtain the spatial spectrum. ; Calculate the isotropic power spectral density : ;in For frequency radius, For radius The number of sampling points on the circumference.

[0058] The sampling window size was set to 128×128 pixels. A uniform grid method was used to select sampling windows in the concrete flow area, with an overlap rate of 50% between windows. An average of 16-36 effective sampling windows were extracted per frame (the number varied with the area of ​​the flow area). The Hanning window function was applied to the grayscale image of each window to reduce spectral leakage, and then a two-dimensional fast Fourier transform (2D-FFT) was performed.

[0059] Extract from power spectral density Ripple characteristic parameters at time: dominant frequency Spectral energy and spectral entropy : ; ; ; Constructing the surface ripple spectrum feature vector: ;in , , These are the time averages of the main frequency, spectral energy, and spectral entropy, respectively. , These are the time standard deviations of spectral energy and spectral entropy, respectively.

[0060] The aggregate suspension trajectory feature vector Specific extraction methods include: The second video sequence is processed using an object detection network. The coarse aggregate particles in the sample are identified and located to obtain... The set of aggregate particle positions in each frame of the image at any given time: ;in for The number of aggregate particles detected at any given time.

[0061] A multi-target tracking algorithm is used to perform inter-frame correlation of aggregate particles, constructing a set of aggregate particle motion trajectories: ;in Indicates the first The complete trajectory of each aggregate, M is the total number of aggregates tracked.

[0062] Calculate the instantaneous velocity of a single aggregate particle: ; Calculate the instantaneous direction of motion: And the coefficients of variation: ;in The average velocity of all aggregate particles, This represents the standard deviation of the velocity.

[0063] Construct the feature vector of aggregate suspension trajectory: ;in For the maximum average speed, The average direction of motion, This represents the standard deviation of the direction of motion.

[0064] The specific method for extracting the feature vector Fp of the slurry coating state includes: The color space of the concrete surface image was converted from RGB to Lab color space, and the luminance channel L and chrominance channels a and b were extracted. The color space conversion was performed using the `cvtColor` function from the OpenCV library. Adaptive thresholding segmentation automatically determined the optimal threshold using the Otsu method, based on the combined characteristics of the L channel (luminance) and b channel (yellow-blue tint) in the Lab space. The slurry area typically exhibits higher luminance (L>120) and a yellowish tint (b>10), while the exposed aggregate area exhibits lower luminance and a grayish tint. The analysis region was divided into 64 sub-regions using an 8×8 grid.

[0065] An adaptive threshold segmentation method is used to distinguish slurry regions. and exposed aggregate areas Calculate the slurry coverage: ;in This represents the total area of ​​the analysis region; The analysis region is divided into Calculate the uniformity index of slurry coating in a sub-region of equal area. : ;in Let be the slurry coverage of the i-th sub-region. This represents the average coverage rate.

[0066] Extracting the gloss characteristics of the slurry: .

[0067] Extracting the consistency of slurry color: .

[0068] Constructing the feature vector of slurry coating state: .

[0069] Step 3: Construct a graph neural network model that integrates prior knowledge of fluid mechanics, and map the discrete visual features in the multi-scale visual feature set to continuous rheological parameters. Based on the rheological parameters, calculate and output the workability index vector of fresh concrete in real time.

[0070] The workability index vector of concrete ,in Slump For scalability, For yield stress, It is a plastic viscosity.

[0071] Constructing a graph neural network model that integrates prior knowledge of fluid mechanics includes: The multi-scale visual feature set F={Fc,Fw,Fa,Fp} is constructed as a feature map. , where the node set These correspond to the flow profile feature nodes, ripple spectrum feature nodes, aggregate trajectory feature nodes, and slurry coating feature nodes, respectively; node feature matrix edge set Based on the physical relationships between features, the following are constructed: flow-corrugated edge (ecw), flow-aggregate edge (eca), corrugated-slurry edge (ewp), and aggregate-slurry edge (eap).

[0072] In this embodiment, the graph neural network contains three graph attention convolutional layers, with hidden feature dimensions of 128, 64, and 32 for each layer, respectively. The construction of the edge set E is based on rheological principles: the flow profile feature and the ripple spectrum feature are connected by the edge ecw, reflecting the correlation between macroscopic flow and surface state; the flow profile feature and the aggregate trajectory feature are connected by the edge eca, reflecting the consistency between overall flow and particle motion; the ripple spectrum feature and the slurry coating feature are connected by the edge ewp, reflecting the relationship between surface texture and slurry state; and the aggregate trajectory feature and the slurry coating feature are connected by the edge eap, reflecting the interaction between aggregate suspension and slurry properties.

[0073] Introducing physical constraints between rheological parameters: Where τ is the shear stress, For yield stress, Plastic viscosity, The constraint relationship, represented by the shear rate, is encoded as an edge weight matrix in a graph neural network.

[0074] Feature aggregation is performed using a physically constrained graph attention mechanism. The node update formula for the layer is: ;in Attention coefficient For nodes The set of neighboring nodes, For activation function, Corresponding node The initial eigenvectors.

[0075] The final layer output features of the graph neural network Mapped to a rheological parameter vector through a fully connected layer: ;in To output the weight matrix, This is the bias vector.

[0076] The calculation formula for workability indicators is combined with the output of a graph neural network, specifically including: the calculation of slump S uses the equivalent diameter when the flow profile is stable. Relationship with initial height: ;in The initial height of the collapse cylinder. The moment when the flow stops The final equivalent diameter (derived from the eigenvectors) supply), This is the geometric correction factor; The calculation of the spread SF combines the time integral of the final equivalent diameter and the flow profile evolution rate: ;in The rate of evolution of the flow profile (derived from the eigenvectors) In approximate).

[0077] Yield stress The calculations combine the morphological characteristics and slurry state at the point where the flow stops: ;in For concrete density, It is the acceleration due to gravity. S represents the final height. The slurry coverage (derived from the feature vector) supply), is an empirical coefficient, and n is the slurry influence index, which is determined through experimental calibration.

[0078] Plastic viscosity The calculation combines flow rate and aggregate suspension characteristics: ;in The maximum flow profile evolution rate (provided by the eigenvector Fc) is the maximum flow profile evolution rate. For reference shear rate, The velocity variation coefficient (derived from the eigenvector) supply), and These are the model coefficients.

[0079] A sliding time window is used to process consecutive video frames in real time, and a performance index vector with an update frequency of no less than 1Hz is output. .

[0080] To verify the effectiveness of the method in this embodiment, a three-month on-site test was conducted at a commercial concrete mixing plant. During the test, a total of 2,860 batches of concrete samples were tested, covering strength grades from C25 to C60.

[0081] Comparison with traditional manual slump tests shows that the correlation coefficient of the slump test results obtained by this method is 0.96, with an average absolute error of 7.2 mm; the correlation coefficient for spread test is 0.94, with an average absolute error of 11.5 mm. A sample comparison with laboratory rheometer tests (180 groups in total) shows that the correlation coefficient for yield stress test is 0.91, and the correlation coefficient for plastic viscosity test is 0.88. The time required for a single test is reduced from 5-8 minutes using traditional methods to less than 30 seconds, improving testing efficiency by more than 10 times. Furthermore, this method successfully identified 23 batches of concrete with abnormal workability (slump deviation exceeding 30 mm or segregation), allowing for timely feedback to the production system for mix proportion adjustments and preventing quality accidents.

[0082] Example 2 like Figure 2 The diagram shows the components of a computer vision-based concrete workability testing system according to the present invention. The system includes: a dual-scale video acquisition module, a video preprocessing module, a spatiotemporal feature pyramid network module, a graph neural network mapping module, and a workability index output module.

[0083] The dual-scale video acquisition module includes a first industrial camera for macroscopic collapse flow acquisition and a second industrial camera for microscopic particle migration acquisition, as well as a time synchronization controller; the first industrial camera is positioned directly above the concrete flow plane, and its frame rate is [missing information]. The second industrial camera is positioned laterally at the leading edge of the flow, and its frame rate is [missing information]. ,and The two cameras achieve frame synchronization acquisition through the time synchronization controller.

[0084] The video preprocessing module is used to process the acquired video sequence. and Denoising, distortion correction, and spatiotemporal alignment are performed.

[0085] The spatiotemporal feature pyramid network module is used to extract multi-scale visual feature sets from the preprocessed video sequence. It includes a flow profile evolution rate feature extraction unit, a surface ripple spectrum feature extraction unit, an aggregate suspension trajectory feature extraction unit, and a slurry coating state feature extraction unit.

[0086] The graph neural network mapping module is used to fuse Bingham's prior knowledge of fluid dynamics and to integrate the multi-scale visual feature set. Mapped to rheological parameter vector It includes graph construction units, physical constraint injection units, message passing units, and rheological parameter regression units.

[0087] The performance index output module is used to output the rheological parameter vector. Calculate and output slump in real time Scalability Yield stress and plastic viscosity The output update frequency is no less than 1Hz, and the results are displayed visually through the display interface.

[0088] The hardware platform of the detection system in this embodiment uses an industrial-grade embedded computer, configured with: an Intel Core i7-11700 processor, 32GB DDR4 memory, a 512GB NVMe solid-state drive, and an NVIDIA RTX 3060 graphics card (for deep learning inference acceleration). The operating system is Ubuntu 20.04 LTS, and the deep learning frameworks used are PyTorch 1.10 and TensorRT 8.0. The overall system power consumption is approximately 350W, the operating temperature range is 0-45℃, and the protection rating is IP54, making it suitable for industrial environments such as mixing plants.

[0089] The system supports one-click start and automatic initialization, completing camera connection, model loading, and interface startup within 60 seconds of power-on, entering a ready state. The system provides a remote maintenance interface, supporting online updates of model parameters and remote viewing of system logs. In actual deployment, it is recommended to set up the testing station near the mixer truck unloading port or the laboratory slump testing platform to ensure the representativeness of concrete samples. The system has been deployed and applied in 5 commercial concrete companies, accumulating over 10,000 hours of operation with a system availability rate of 99.2%, significantly improving the quality control level and intelligence of concrete production.

[0090] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A computer vision-based method for detecting the workability of concrete, characterized in that, The method includes the following steps: Step 1: Using a non-contact visual sensing method, video sequences of freshly mixed concrete in a free-flowing state are simultaneously acquired from both the macroscopic slump flow scale and the microscopic particle migration scale. The first video sequence of macroscopically acquired data on the overall flow morphology changes of freshly mixed concrete. The second video sequence was acquired at the microscale to obtain the particle motion state on the surface of freshly mixed concrete. ; Step 2: An improved spatiotemporal feature pyramid network is used to extract multi-scale morphological features from the first video sequence and the second video sequence respectively, to obtain a multi-scale visual feature set. The multi-scale visual feature set ,in This is the eigenvector representing the evolution rate of the flow profile. The surface ripple spectrum feature vector, This represents the feature vector of the aggregate suspension trajectory. The feature vector of the slurry coating state; Step 3: Construct a graph neural network model that integrates prior knowledge of fluid mechanics, map the discrete visual features in the multi-scale visual feature set to continuous rheological parameters, and calculate and output the workability index vector of fresh concrete in real time based on the rheological parameters. The workability index vector of concrete ,in Slump For scalability, For yield stress, It is a plastic viscosity.

2. The method according to claim 1, characterized in that, The macroscopic data acquisition was performed using a first industrial camera, taking images from a top-down view directly above the concrete flow plane, with a frame rate of [missing information]. Sampling time interval ; Obtain the first video sequence of the entire process of freshly mixed concrete from the moment of slump to the moment of cessation of flow: The microscale data acquisition was performed using a second industrial camera, taking images from a lateral oblique angle at the concrete flow front edge, with a frame rate of [missing information]. Sampling time interval The second video sequence of particle movement on the concrete surface was obtained. in, To ensure the temporal resolution of the movement of micro-particles; the first industrial camera and the second industrial camera are synchronized frame by time synchronization signal.

3. The method according to claim 2, characterized in that, The improved spatiotemporal feature pyramid network includes a spatial feature pyramid module, a temporal feature pyramid module, a feature fusion module, and a feature decoupling module; The spatial feature pyramid module uses multi-scale convolutional kernels to extract feature maps at different spatial resolutions. The set of scale factors for the spatial feature pyramid is as follows: , among which, the Each scaling factor: ;for Input image at time , No. The layer space features are represented as follows: ;in Indicates the downsampling factor as Pooling operations, Indicates the first Layered convolution operations; multi-scale spatial feature fusion is performed as follows: ； The temporal feature pyramid module uses three-dimensional convolution and temporal attention mechanisms to extract motion features between video frames; for example, the temporal feature pyramid module uses three-dimensional convolution and temporal attention mechanisms to extract motion features between video frames; Centered on time, with a time window radius of A sequence of consecutive frames of images: The time characteristics are represented as follows: ;in The radius of the time window. This represents the sampling time interval for the corresponding video sequence; The feature fusion module adaptively fuses spatial and temporal features. The fusion feature at any given time is represented as follows: ； in and For learnable fusion weights, Indicates feature concatenation operation; The feature decoupling module will fuse features. The features are mapped to four categories of visual features through four parallel feature decoupling branches: ； in , , , These are the feature decoupling mapping functions for flow profile, surface ripples, aggregate trajectory, and slurry coating, respectively.

4. The method according to claim 3, characterized in that, The flow profile evolution rate feature vector The specific extraction method includes: performing semantic segmentation on each frame of the first video sequence to obtain a binary mask of the concrete flow region; extracting the boundary contour of the binary mask, calculating the area enclosed by the contour and the diameter of the equivalent circle; calculating... The flow profile evolution rate at time t is calculated, the flow profile evolution acceleration is calculated, and the flow profile evolution rate eigenvector is constructed.

5. The method according to claim 4, characterized in that, The surface ripple spectrum feature vector Specific extraction methods include: Multiple sampling windows are selected within the concrete flow area at any given time. A two-dimensional Fourier transform is performed on the grayscale image within each window to obtain the spatial spectrum. The isotropic power spectral density is calculated. Extraction is performed from the power spectral density. The surface ripple spectrum feature vector is constructed by taking the ripple characteristic parameters at time t.

6. The method according to claim 5, characterized in that, The aggregate suspension trajectory feature vector The specific extraction methods include: using an object detection network to analyze the second video sequence. The coarse aggregate particles in the sample are identified and located to obtain... The set of aggregate particle positions in each frame of the image at any given time is used. A multi-target tracking algorithm is used to correlate the aggregate particles between frames to construct a set of aggregate particle motion trajectories. The instantaneous velocity, instantaneous motion direction, and coefficient of variation of a single aggregate particle are calculated. The feature vector of the aggregate suspension trajectory is then constructed.

7. The method according to claim 6, characterized in that, The feature vector of the slurry coating state The specific extraction methods include: converting the color space of the concrete surface image from RGB space to Lab color space, extracting the luminance channel L and chromaticity channels a and b; using an adaptive threshold segmentation method to distinguish between the slurry area and the exposed aggregate area, calculating the slurry coverage rate and slurry coating uniformity index, extracting the slurry gloss characteristics and slurry color consistency characteristics, and constructing a slurry coating state feature vector.

8. The method according to claim 7, characterized in that, Constructing a graph neural network model that integrates prior knowledge of fluid mechanics includes: A feature map is constructed from a set of multi-scale visual features, where the node set These correspond to the flow profile feature node, ripple spectrum feature node, aggregate trajectory feature node, and slurry coating feature node, respectively. Node feature matrix edge set Based on the physical relationships between features, physical constraints between rheological parameters are introduced: ;in For shear stress, For yield stress, Plastic viscosity, The shear rate and the constraint relationships are encoded as edge weight matrices of a graph neural network. A physically constrained graph attention mechanism is used for feature aggregation, mapping the final layer output features of the graph neural network into a rheological parameter vector through a fully connected layer: .

9. The method according to claim 8, characterized in that, A sliding time window is used to process continuous video frames in real time, and the output performance index vector is updated at a frequency of no less than 1Hz.

10. A computer vision-based concrete workability testing system, used to perform the method according to any one of claims 1 to 9, characterized in that, The system includes: a dual-scale video acquisition module, a video preprocessing module, a spatiotemporal feature pyramid network module, a graph neural network mapping module, and a performance index output module. The dual-scale video acquisition module includes a first industrial camera for macroscopic collapse flow acquisition and a second industrial camera for microscopic particle migration acquisition, as well as a time synchronization controller; the first industrial camera is positioned directly above the concrete flow plane, and its frame rate is [missing information]. The second industrial camera is positioned laterally at the leading edge of the flow, and its frame rate is [missing information]. ,and The two cameras achieve frame synchronization acquisition through the time synchronization controller. The video preprocessing module is used to process the acquired video sequence. and Denoising, distortion correction, and spatiotemporal alignment are performed. The spatiotemporal feature pyramid network module is used to extract multi-scale visual feature sets from the preprocessed video sequence. It includes a flow profile evolution rate feature extraction unit, a surface ripple spectrum feature extraction unit, an aggregate suspension trajectory feature extraction unit, and a slurry coating state feature extraction unit; The graph neural network mapping module is used to fuse Bingham's prior knowledge of fluid dynamics and to integrate the multi-scale visual feature set. Mapped to rheological parameter vector It includes graph construction units, physical constraint injection units, message passing units, and rheological parameter regression units; The performance index output module is used to output the rheological parameter vector. Calculate and output slump in real time Scalability Yield stress and plastic viscosity The output update frequency is no less than 1Hz, and the results are displayed visually through the display interface.

Images