Visual intelligent detection method and system for rockfill material grading

Abstract

The invention belongs to the technical field of water conservancy and hydropower engineering construction, and relates to a visual intelligent detection method and system for rockfill material gradation, and the detection method comprises the steps: 1) arranging an information collection system on a rockfill material transport line; 2) acquiring a license plate of a rockfill material transport vehicle and an image of rockfill materials carried by the rockfill material transport vehicle based on the information acquisition system in the step 1); 3) performing rockfill material grading detection on the image of the rockfill material carried by the rockfill material transport vehicle obtained in the step 2) to obtain a grading detection result; and 4) taking the grading detection result obtained in the step 3) and the license plate of the rockfill material transport vehicle obtained in the step 2) corresponding to the detection result as visual intelligent detection results and issuing the visual intelligent detection results to the outside. The visual intelligent detection method and system for rockfill material grading are high in detection efficiency and accurate in detection result.

Description

Technical Field

[0001] The present invention belongs to the technical field of water conservancy and hydropower engineering construction, and relates to a visual intelligent detection method and system for the gradation of rockfill materials, and particularly relates to a visual intelligent detection method and system for the gradation of rockfill materials suitable for rapid detection of the particle gradation of large-scale rockfill materials in rockfill dams or similar projects. Background Art

[0002] Rockfill dams have been widely used in hydropower projects due to their strong adaptability and high economy. Among them, the gradation of the filler is a key control index determining the filling quality of the dam body. However, the existing gradation detection methods mainly rely on digging pits and sampling for screening after compaction. These methods are not only complex in operation and have large construction interference, but also have insufficient sample representativeness and belong to post-verification, making it difficult to meet the real-time control requirements before filling.

[0003] With the development of information technology, detection methods based on morphological perception provide a new technical path for the efficient analysis of the gradation of rockfill materials. These methods are mainly divided into two categories: point cloud scanning and visual perception. Point cloud scanning quantifies the particle morphology through three-dimensional laser or structured light technology, but its point cloud segmentation efficiency is low, which limits its popularization and application in large-scale projects. In contrast, two-dimensional visual perception technology extracts particle contours and calculates gradations using image segmentation, which has the advantages of high efficiency and convenience. Especially, image segmentation technology based on deep learning has made breakthroughs in terms of robustness and generalization, significantly enhancing the potential for large-scale application at the construction site. However, on the one hand, the existing image segmentation technology has insufficient segmentation performance for small particles, especially rockfill materials with a particle size less than 5 mm; on the other hand, for rockfill materials in a natural stacked state, there are inter-particle occlusions on the surface layer and inter-layer occlusions in the deep layer, resulting in the problem of missing visual information. The above factors lead to difficult gradation analysis and become the main technical bottleneck in visual gradation detection. Summary of the Invention

[0004] In order to solve the above technical problems existing in the background art, the present invention provides a visual intelligent detection method and system for the gradation of rockfill materials with high detection efficiency and accurate detection results.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A visual intelligent detection method for the gradation of rockfill materials, characterized in that: the visual intelligent detection method for the gradation of rockfill materials includes the following steps:

[0007] 1) Set up an information collection system on the rockfill material transportation route;

[0008] 2) Based on the information collection system in step 1), obtain the license plate of the rockfill material transportation vehicle and the image of the rockfill materials carried by the rockfill material transportation vehicle;

[0009] 3) Perform gradation detection on the heap stone material images carried by the heap stone material transport vehicles obtained in step 2) to obtain the gradation detection results;

[0010] 4) Publish the gradation detection results obtained in step 3) and the license plate of the heap stone material transport vehicle obtained in step 2) corresponding to the detection results as visual intelligent detection results.

[0011] Preferably, the information acquisition system in step 1) at least includes a heap stone material image acquisition device and a heap stone material transport vehicle information acquisition device; the heap stone material transport vehicle information acquisition device includes an RFID reader, an RFID tag installed on the transport vehicle, and a network transmission unit; the RFID tag is at least bound with the license plate of the heap stone material transport vehicle.

[0012] Preferably, the specific implementation manner of step 3) is:

[0013] 3.1) Use an image classification and recognition model to determine whether the content of P5 in the heap stone material images carried by the heap stone material transport vehicles exceeds the standard; if the judgment result is that it exceeds the standard, obtain the detection result of unqualified gradation; if the judgment result is that it does not exceed the standard, proceed to step 3.2);

[0014] 3.2) Obtain the gradation curve of the heap stone material images carried by the heap stone material transport vehicles, and obtain the detection result that the gradation of the heap stone material carried by the heap stone material transport vehicle is qualified material or the detection result that the gradation of the heap stone material is unqualified material according to the gradation curve of the heap stone material images carried by the heap stone material transport vehicles.

[0015] Preferably, the image classification and recognition model in step 3.1) is the result of using the YOLOv8-cls model and training the YOLOv8-cls model; the heap stone material P5 is the heap stone material particles with a particle size less than 5 mm.

[0016] Preferably, the specific construction method of the image classification and recognition model is:

[0017] a.1) Collect heap stone material image samples for training the model under various different imaging conditions, and the heap stone material image samples for training the model at least cover the situations where the content of P5 in the heap stone material is qualified and unqualified;

[0018] a.2) Label the heap stone material image samples for training the model in step a.1) in a binary classification manner to obtain initial samples;

[0019] a.3) Perform data augmentation on the initial samples to obtain a sample expansion set; the data augmentation method is translation, rotation, brightness transformation, and / or flipping;

[0020] a.4) Divide the sample augmentation set obtained in step a.3) into a training set and a validation set according to a ratio of 8:2;

[0021] a.5) Train the YOLOv8-cls model with the training set obtained in step a.4) to obtain a trained model; at the same time, verify the trained model using the validation set obtained in step a.4), and finally obtain an image classification and recognition model.

[0022] Preferably, the specific implementation method of step 3.2) is as follows:

[0023] 3.2.1) Segment and extract the contour of the rockfill image carried by the rockfill transport vehicle;

[0024] 3.2.2) Estimate the three-dimensional size of the rockfill particles based on the results of step 3.2.1), and calculate the visual gradation of the rockfill according to the three-dimensional size of the rockfill;

[0025] 3.2.3) Construct a mapping model between the visual gradation and the actual gradation of the rockfill, and infer the actual gradation of the rockfill according to the mapping model;

[0026] 3.2.4) Obtain the gradation curve of the actual gradation of the rockfill in step 3.2.3), and compare the gradation curve of the actual gradation of the rockfill with the envelope line of the design gradation curve. When the gradation curve of the actual gradation of the rockfill is within the envelope line of the design gradation curve, the test result that the gradation of the rockfill is qualified is obtained; when the gradation curve of the actual gradation of the rockfill is not within the envelope line of the design gradation curve, the test result that the gradation of the rockfill is unqualified is obtained.

[0027] Preferably, for the rockfill image segmentation model used in step 3.2.1), the rockfill image segmentation model is the result of using the YOLOv8-seg model and training the YOLOv8-seg model;

[0028] The specific implementation method of the rockfill image segmentation model is as follows:

[0029] b.1) Collect mixed-level rockfill image samples under various different imaging conditions;

[0030] b.2) Label the mixed-level rockfill image samples obtained in step b.1) to obtain initial samples;

[0031] b.3) Perform data augmentation on the initial samples to obtain a sample augmentation set; the data augmentation methods are translation, rotation, brightness transformation, and / or flipping;

[0032] b.4) Divide the sample augmentation set obtained in step b.3) into a training set and a validation set according to a ratio of 8:2;

[0033] b.5) Use the training set obtained in step b.4) to train the YOLOv8-seg model to obtain a trained model; at the same time, use the validation set obtained in step b.4) to validate the trained model, and finally obtain a rockfill image segmentation model;

[0034] The specific implementation method of the contour extraction is as follows:

[0035] c.1) Use the convex hull algorithm to fit the incomplete contour to obtain the key points of the contour;

[0036] c.2) Use the least squares method to fit an equivalent ellipse based on the contour key points.

[0037] Preferably, the specific implementation method of step 3.2.2) is as follows:

[0038] 3.2.2.1) Obtain the equivalent ellipsoid of the rockfill particles and the projection of the ellipsoid on the plane, and the projection of the ellipsoid on the plane is an ellipse; the expression of the ellipsoid is:

[0039]

[0040] The expression of the ellipse is:

[0041] a 11 X 2 +2a 12 XY+a 22 Y 2 +a 33 =0 (2)

[0042] Where:

[0043]

[0044] Where:

[0045] θ and are respectively the azimuth angle and the pitch angle between the projection plane and the main plane of the ellipsoid;

[0046] For the equation of the projection relationship between the major axis A, minor axis B of the ellipse with formula (2) and the three semi-major axes a, b, c of the ellipsoid is:

[0047]

[0048] Where:

[0049] I=a 11 +a 22 ;

[0050]

[0051] 3.2.2.2) Solve the equation of the projection relationship obtained in step 3.2.2.1) to obtain the estimated values of the three-dimensional dimensions of the rockfill particles. The estimated values of the three-dimensional dimensions of the rockfill particles are the three semi-major axes a, b, and c of the ellipse;

[0052] The solution method is as follows:

[0053] Introduce the particle shape characteristic index aspect ratio λ 1 and thickness-to-width ratio λ 2 as well as the probability distributions of the stacking azimuth angle θ and the pitch angle :

[0054]

[0055] Use the Monte Carlo method to solve the set of feasible particle three-dimensional dimensions, and use the expectation of the solution set as the representative solution. The conditions for the feasible solutions are:

[0056]

[0057] Where:

[0058] n is the particle number of the rockfill;

[0059] ε A and ε B are the error thresholds of the major and minor semi-axes respectively;

[0060] A * is the size of the major semi-axis of the equivalent ellipse;

[0061] B * is the size of the minor semi-axis of the equivalent ellipse;

[0062] According to the feasible solution space, obtain the estimated values of the three semi-axis lengths a, b, and c of the ellipse. The expressions for the estimated values of the three semi-axis lengths of the ellipse are respectively:

[0063]

[0064] Where N is the number of feasible solutions;

[0065] is the estimated value of the semi-axis length a;

[0066] is the estimated value of the semi-axis length b;

[0067] is the estimated value of the semi-axis length c;

[0068] 3.2.2.3) Calculate the visual gradation of the rockfill materials based on the three-dimensional particle sizes of the rockfill materials estimated in step 3.2.2.2). The specific implementation method of the visual gradation of the rockfill materials is as follows:

[0069] The density of the rockfill materials is ρ. Calculate the mass M and gradation GIR of each particle of the rockfill materials * :

[0070]

[0071] where:

[0072] n represents the total number of rockfill material particles within the particle size range;

[0073] N' represents the total number of rockfill material particles in (1).

[0074] Preferably, the specific implementation method of step 3.2.3) is as follows:

[0075] d.1) Collect rockfill material samples from multiple trips, identify the visual gradation of the samples vehicle by vehicle, and obtain the actual gradation of the rockfill materials through screening tests by sampling the rockfill materials in the vehicle to obtain a model training sample set;

[0076] d.2) Divide the training sample set into a training set and a test set according to the ratio of 0.85:0.15;

[0077] d.3) In the training set obtained in step d.3), use the visual gradation as the input sample and the screening gradation as the output sample to train a support vector machine regression model; at the same time, verify the trained support vector machine regression model through the test set obtained in step d.3) to obtain a mapping model between the visual gradation and the actual gradation of the rockfill materials;

[0078] d.4) Infer the actual gradation of the rockfill materials according to the mapping model between the visual gradation and the actual gradation of the rockfill materials obtained in step d.3).

[0079] A detection system for implementing the visual intelligent detection method of the rockfill material gradation as described above, characterized in that: the detection system includes an information collection system, a rockfill material gradation detection module, an information association module, and a detection result external release module;

[0080] The information collection system includes a rockfill material image collection device and a rockfill material transport vehicle information collection device; the rockfill material transport vehicle information collection device includes an RFID reader, an RFID tag installed on the transport vehicle, and a network transmission unit; the RFID tag is at least bound with the license plate of the rockfill material transport vehicle;

[0081] The information collection system is used to collect the license plate of the rockfill material transport vehicle and the image of the rockfill materials carried by the rockfill material transport vehicle;

[0082] The rockfill gradation detection module is used to detect the gradation of the rockfill images carried by rockfill transport vehicles and obtain the gradation detection results;

[0083] The information association module is used to associate the gradation detection results with the license plates of the rockfill transport vehicles to form the results for external release;

[0084] The detection result external release module is used to send the results for external release to the construction party or the management party.

[0085] The advantages of the present invention are as follows:

[0086] The present invention discloses a visual intelligent detection method and system for the gradation of rockfill. The method includes: 1) setting up an information collection system on the rockfill transport route; 2) obtaining the license plates of the rockfill transport vehicles and the rockfill images carried by the rockfill transport vehicles based on the information collection system in step 1); 3) detecting the gradation of the rockfill images carried by the rockfill transport vehicles obtained in step 2) to obtain the gradation detection results; 4) releasing the gradation detection results obtained in step 3) and the license plates of the rockfill transport vehicles obtained in step 2) corresponding to the detection results as the visual intelligent detection results to the outside. The present invention collects images and realizes gradation detection during the rockfill transportation stage, aiming to solve the problems of low efficiency, high cost and insufficient representativeness of traditional manual gradation screening. Aiming at the limitations of the segmentation performance of small particle images of on-vehicle rockfill and the visual information loss caused by inter-particle occlusion and inter-layer occlusion in the natural stacking state, a gradation detection method and system based on visual intelligence are proposed. The invention has the advantages of non-contact, high efficiency and accuracy, can quickly detect the gradation of large-scale rockfill before filling, and provides reliable technical support for the construction quality control of rockfill dams. Description of the Drawings

[0087] Figure 1 is a flowchart of the visual intelligent detection method for the gradation of rockfill provided by the present invention. Detailed Embodiments

[0088] Next, in combination with the drawings and specific embodiments, the invention will be further described:

[0089] See Figure 1 , the present invention provides a visual intelligent detection method for the gradation of rockfill, which specifically includes the following steps:

[0090] S1. According to the transportation route of the rockfill materials at the construction site, select a suitable location to deploy the image and license plate information collection devices for rockfill materials, considering factors such as electricity and network. Fix the L-shaped vertical pole on the concrete foundation through embedded anchor bolts to ensure the stability of the vertical pole. Install the image acquisition camera, RFID reader, and network transmission unit.

[0091] Preferably, in step S1, it also includes: sticking an RFID tag on the vehicle transporting the rockfill materials, and this tag is bound with the license plate number information. Install it at a prominent and easily readable position on the vehicle head to ensure that the tag information can be smoothly read by the RFID reader during the vehicle's driving process.

[0092] S2. When the transportation vehicle drives into the gradation detection station, the RFID reader automatically reads the RFID tag information on the vehicle, and at the same time, the image acquisition device takes pictures and samples of the rockfill materials in the hopper according to the preset parameters and angles.

[0093] S3. Transmit the image data and RFID tag information collected in S2 to the gradation cloud analysis module in real time through the network transmission unit, and conduct rockfill material gradation detection based on the rockfill material images.

[0094] Preferably, the cloud gradation analysis system in S3 conducts gradation detection based on the images collected in step S2, including two stages:

[0095] The first stage: Use the image classification and recognition model to judge whether the P5 content (particles with a particle size less than 5 mm) of the rockfill materials exceeds the standard. If it is determined to exceed the standard, output that the gradation is unqualified; if not, enter the detection of the second stage.

[0096] Preferably, the image classification and recognition model for the P5 content of the rockfill materials adopts the YOLOv8-cls model, and the model training process includes:

[0097] Collect mixed gradation rockfill material images under various imaging conditions (such as lighting, weather, etc.), covering the cases where the P5 content is qualified and unqualified. Further, label the image samples in a binary classification manner, for example, mark "qualified" as 1 and "unqualified" as 0 to obtain the initial samples. Then, perform data augmentation on the initial samples to expand the sample set, and use operations such as translation, rotation, brightness transformation, and flipping to increase the diversity of the data. Finally, divide the expanded sample set into a training set and a validation set according to a ratio of 8:2, so as to use the trained image classification and recognition model for the P5 content of the rockfill materials to realize the discrimination of the qualification of the P5 content of the rockfill material images collected in S2.

[0098] The second stage, if the P5 content is not determined to exceed the standard in the first-stage detection, enter this stage, including the following analysis modules:

[0099] (1) Image segmentation and contour extraction of rockfill materials.

[0100] Preferably, the rockfill material image segmentation model adopts the YOLOv8-seg model, and the model training process includes: collecting mixed-graded rockfill material images under various imaging conditions (lighting, weather, etc.). Further, the contours of the rockfill materials in the images are labeled to obtain initial samples. Then, data augmentation is performed on the initial samples to expand the sample set, and operations such as translation, rotation, brightness transformation, and flipping are used to increase the diversity of the data. Finally, the expanded sample set is divided into a training set and a validation set according to a ratio of 8:2, so as to use the trained rockfill material image segmentation model to segment the rockfill material images collected in S2 to obtain the stone particle contours.

[0101] Preferably, the rockfill material contour extraction includes: in the stacked state, the mutual occlusion between surface particles results in incomplete stone particle contours obtained by segmentation. First, the convex hull algorithm is used to fit the incomplete contours to obtain the key points of the contours; since the overall shape of the particle contours is approximately elliptical, the least squares method is used to fit an equivalent ellipse based on the contour key points to maximize the retention of the original shape features of the particle contours. According to the shooting distance of the rockfill material image acquisition, the conversion relationship between the pixel size and the actual size of the image is determined, and then the sizes A * and B * .

[0102] (2) Three-dimensional estimation and visual gradation calculation.

[0103] Preferably, the three-dimensional sizes of the rockfill material particles are estimated based on the major and minor axis sizes of the equivalent ellipse obtained in (1), specifically including:

[0104] 1) The particles are approximated as ellipsoids, and the equivalent ellipse obtained in step (1) is the projection of the particle approximate ellipsoid on the plane. The ellipsoid equation is as follows:

[0105]

[0106] The projected ellipse equation of the ellipsoid on the plane can be obtained as:

[0107] a 11 X 2 +2a 12 XY+a 22 Y 2 +a 33 =0 (2)

[0108] Among them,

[0109]

[0110] Among them, θ and They are respectively the azimuth angle and the pitch angle between the projection plane and the main plane of the ellipsoid. For the relationship between the major and minor semi-axes A and B of the ellipse with the equation of (2) and the three semi-major axes a, b, and c of the ellipsoid, it can be expressed as:

[0111]

[0112] where I = a 11 + a 22 ,

[0113] 2) The three-dimensional size and angle of the particles have a total of five unknowns, namely a, b, c, θ, and . Given that there are only two parameters, A and B, the projection relationship equation is an indeterminate equation. Further, in order to solve the aforementioned indeterminate equation, the aspect ratio λ 1 of the particle shape characteristic index, the thickness-to-width ratio λ 2 , and the probability distributions of the stacking azimuth angle θ and the pitch angle are introduced:

[0114]

[0115] The Monte Carlo method is used to solve the feasible set of particle three-dimensional sizes, and the expectation of the solution set is taken as the representative solution. The conditions for the feasible solution are:

[0116]

[0117] where n is the particle number, and ε A and ε B are respectively the error thresholds of the major and minor semi-axes.

[0118] According to the feasible solution space, the three axis lengths a, b, and c are obtained, which are the estimated values of the particle three-dimensional sizes:

[0119]

[0120] where N is the number of feasible solutions.

[0121] Preferably, based on the estimated three-dimensional sizes of the rockfill particles, the visual gradation is calculated, specifically including:

[0122] The density of the rockfill is ρ, and the mass M and gradation GIR of each particle can be calculated * :

[0123]

[0124] where n represents the total number of rockfill particles within the particle size range, and N' represents the total number of rockfill particles in (1).

[0125] (3) Actual gradation inference. Based on the measured sieve gradation data, a mapping model between the visual gradation and the actual gradation of the surface layer rockfill is constructed using a support vector machine regression model.

[0126] Preferably, the process of constructing the mapping model includes: collecting rockfill samples of multiple trips, identifying the visual gradation of the samples vehicle by vehicle, and sampling the rockfill in the vehicle (2 - 3m 3 ), obtaining the actual gradation of the stone material through a screening test, and obtaining the model training sample set. Further, using the visual gradation as the input sample and the sieve gradation as the output sample, training the support vector machine regression model to realize the inference of the actual gradation based on the visual gradation obtained in (2).

[0127] Exemplarily, the training process of the support vector machine: using the visual gradation as the input and the sieve gradation as the output, dividing the samples into a training set and a test set according to the ratio of 0.85:0.15. In order to balance the differences in the order of magnitude of the mass percentages of different particle sizes, the data is normalized. The SVR algorithm uses a radial basis kernel function, and the penalty factor C and the parameter σ in the kernel function 2 are optimized using a particle swarm algorithm to obtain the optimal parameter combination of the two parameters, thus completing the model training. After training, the support vector regression model is considered as an existing tool. When the visual gradation is detected, it can be input into the support vector regression model to obtain the corresponding actual gradation. At the same time, the actual gradation obtained from the screening experiment is equivalent to the actual gradation of the rockfill: the actual gradation is unknown, and the gradation obtained through the screening experiment represents the actual gradation.

[0128] (4) Gradation curve determination: Comparing the gradation curve obtained in (3) with the envelope line of the design gradation curve. When the gradation detection result is within the range of the designed gradation envelope line, the rockfill of this vehicle is considered qualified; otherwise, it is unqualified.

[0129] S4. Associate the gradation detection result obtained in S3 with the license plate information collected in S2, and publish the result through the system push module.

Claims

1. A visual intelligent detection method for rockfill material gradation, characterized by: The visual intelligent detection method for rockfill material gradation comprises the following steps: 1) Set up an information collection system on the rockfill transportation route; 2) Based on step 1), the information collection system acquires the license plate of the rockfill material transport vehicle and the image of the rockfill material carried by the rockfill material transport vehicle; 3) performing a rockfill material gradation test on the rockfill material image carried by the rockfill material transport vehicle obtained in step 2) to obtain a gradation test result; 4) The gradation detection result obtained in step 3) and the license plate of the rockfill material transport vehicle obtained in step 2) corresponding to the detection result are published as visual intelligent detection results.

2. The visual intelligent detection method for rockfill material gradation according to claim 1, characterized in that: The information collection system in step 1) at least includes a rockfill image collection device and a rockfill transport vehicle information collection device; the rockfill transport vehicle information collection device includes an RFID card reader, an RFID tag installed on the transport vehicle, and a network transmission unit; the RFID tag is at least bound to the license plate of the rockfill transport vehicle.

3. The visual intelligent detection method for rockfill material gradation according to claim 2 is characterized by: The specific implementation method of step 3) is: 3.1) Using the image classification recognition model, determine whether the content of the rockfill material P5 in the image of the rockfill material carried by the rockfill material transport vehicle exceeds the standard; if the judgment result is that it exceeds the standard, obtain a test result of unqualified gradation; if the judgment result is that it does not exceed the standard, proceed to step 3.2); 3.2) Obtaining a gradation curve of the image of the rockfill material carried by the rockfill material transport vehicle, and obtaining a test result of whether the gradation of the rockfill material carried by the rockfill material transport vehicle is qualified material or unqualified material according to the gradation curve of the image of the rockfill material carried by the rockfill material transport vehicle.

4. The visual intelligent detection method for rockfill material gradation according to claim 3 is characterized by: The image classification recognition model in step 3.1) is the result of using a YOLOv8-cls model and training the YOLOv8-cls model; the rockfill material P5 is rockfill material particles with a particle size of less than 5 mm.

5. The visual intelligent detection method for rockfill material gradation according to claim 4 is characterized by: The specific construction method of the image classification and recognition model is: a.1) collecting rockfill material image samples for training the model under a variety of different imaging conditions, wherein the rockfill material image samples for training the model at least cover situations where the rockfill material P5 content is qualified and unqualified; a.2) labeling the rockfill image samples used for training the model in step a.1) in a binary classification manner to obtain initial samples; a.3) performing data enhancement on the initial samples to obtain an expanded sample set; the data enhancement method is translation, rotation, brightness transformation and / or flipping; a.4) Divide the sample expansion set obtained in step a.3) into a training set and a validation set in a ratio of 8:2; a.5) The training set obtained in step a.4) is used to train the YOLOv8-cls model to obtain a training model; meanwhile, the training model is verified using the verification set obtained in step a.4) to finally obtain an image classification and recognition model.

6. The visual intelligent detection method for rockfill material gradation according to claim 5, characterized in that: The specific implementation method of step 3.2) is: 3.2.1) Segmenting and contour extraction of images of rockfill materials carried by rockfill material transport vehicles; 3.2.2) estimating the three-dimensional size of the rockfill material particles according to the result of step 3.2.1), and calculating the visual gradation of the rockfill material according to the three-dimensional size of the rockfill material; 3.2.3) Construct a mapping model between the visual gradation and the actual gradation of the rockfill material, and infer the actual gradation of the rockfill material based on the mapping model; 3.2.4) Obtaining a gradation curve of the actual gradation of the rockfill material in step 3.2.3), and comparing the gradation curve of the actual gradation of the rockfill material with the envelope of the design gradation curve; when the gradation curve of the actual gradation of the rockfill material is within the envelope of the design gradation curve, a test result indicating that the gradation of the rockfill material is qualified material is obtained; when the gradation curve of the actual gradation of the rockfill material is not within the envelope of the design gradation curve, a test result indicating that the gradation of the rockfill material is unqualified material is obtained.

7. The visual intelligent detection method for rockfill material gradation according to claim 6, characterized in that: The rockfill image segmentation model used in the segmentation in step 3.2.1) is a result of using a YOLOv8-seg model and training the YOLOv8-seg model; The specific implementation of the rockfill image segmentation model is: b.1) Collect mixed-grade rockfill image samples under various imaging conditions; b.2) labeling the mixed-grade rockfill material image samples obtained in step b.1) to obtain initial samples; b.3) performing data enhancement on the initial samples to obtain an expanded sample set; the data enhancement method is translation, rotation, brightness transformation and / or flipping; b.4) Divide the sample expansion set obtained in step b.3) into a training set and a validation set in a ratio of 8:2; b.5) ​​training the YOLOv8-seg model with the training set obtained in step b.4) to obtain a training model; and verifying the training model with the verification set obtained in step b.4) to finally obtain a rockfill image segmentation model; The specific implementation method of the contour extraction is: c.1) Use the convex hull algorithm to fit the incomplete contour and obtain the key points of the contour; c.2) The equivalent ellipse is obtained by fitting the least square method based on the contour key points.

8. The visual intelligent detection method for rockfill material gradation according to claim 7, characterized in that: The specific implementation method of step 3.2.2) is: 3.2.2.1) Obtain the equivalent of the rockfill material particles to an ellipsoid and the projection of the ellipsoid on the plane, wherein the projection of the ellipsoid on the plane is an ellipse; the expression of the ellipsoid is: The expression of the ellipse is: a 11 X 2 +2a 12 XY+a 22 Y 2 +a 33 =0 (2) in: in: θ and are the azimuth and elevation angles between the projection plane and the principal plane of the ellipsoid, respectively; The equation for the projection relationship between the ellipse length A, the minor semi-axis B and the three semi-major axes a, b, c of the ellipsoid with formula (2) is: in: I=a 11 +a 22 ; 3.2.2.2) Solving the equation of the projection relationship obtained in step 3.2.2.1) to obtain an estimated value of the three-dimensional size of the rockfill material particle, wherein the estimated value of the three-dimensional size of the rockfill material particle is the three semi-major axes a, b, and c of the ellipse; The solution is: The particle shape characteristic indicators width-to-length ratio λ1, thickness-to-width ratio λ2, stacking azimuth angle θ and pitch angle are introduced The probability distribution of is: The Monte Carlo method is used to solve the feasible three-dimensional particle size set, and the expectation of the solution set is used as the representative solution. The conditions for the feasible solution are: in: n is the particle number of the rockfill material; ε A and ε B are the error thresholds of the major and minor semi-axes respectively; A * is the dimension of the major semi-axis of the equivalent ellipse; B * is the dimension of the minor semiaxis of the equivalent ellipse; According to the feasible solution space, the estimated values ​​of the three semi-axis lengths a, b, and c of the ellipse are obtained. The expressions of the estimated values ​​of the three semi-axis lengths of the ellipse are respectively: Where N is the number of feasible solutions; is the estimated value of the semi-axis length a; is the estimated value of the semi-axis length b; is the estimated value of the semi-axis length c; 3.2.2.3) Calculate the visual gradation of the rockfill material according to the three-dimensional size of the rockfill material particles estimated in step 3.2.2.2). The specific implementation method of the visual gradation of the rockfill material is: The density of rockfill is ρ, and the mass M and gradation GIR of each particle of rockfill are calculated. * : in: n represents the total number of rockfill particles within the particle size range; N′ represents the total number of rockfill particles in (1).

9. The visual intelligent detection method for rockfill material gradation according to claim 8, characterized in that: The specific implementation method of step 3.2.3) is: d.1) Collect rockfill samples from multiple vehicles, identify the visual gradation of samples one by one, and obtain the actual gradation of the rockfill by sampling the rockfill in the vehicle through screening tests to obtain the model training sample set; d.2) Divide the training sample set into a training set and a test set in a ratio of 0.85:0.15; d.3) In the training set obtained in step d.3), the visual gradation is used as an input sample and the screening gradation is used as an output sample to train a support vector machine regression model; at the same time, the trained support vector machine regression model is verified by the test set obtained in step d.3), and a mapping model between the visual gradation of the rockfill material and the actual gradation is obtained; d.4) The actual gradation of the rockfill material is obtained by inferring the mapping model between the visual gradation of the rockfill material obtained in step d.3) and the actual gradation.

10. A detection system for implementing the visual intelligent detection method for rockfill material gradation according to any one of claims 1 to 9, characterized in that: The detection system includes an information collection system, a rockfill material gradation detection module, an information association module, and a detection result external release module; The information collection system includes a rockfill material image collection device and a rockfill material transport vehicle information collection device; the rockfill material transport vehicle information collection device includes an RFID card reader, an RFID tag installed on the transport vehicle, and a network transmission unit; the RFID tag is at least bound to the license plate of the rockfill material transport vehicle; The information collection system is used to collect the license plate of the rockfill material transport vehicle and the image of the rockfill material carried by the rockfill material transport vehicle; The rockfill material gradation detection module is used to perform gradation detection on the rockfill material image carried by the rockfill material transport vehicle and obtain the gradation detection result; The information association module is used to associate the gradation detection result with the license plate of the rockfill material transportation vehicle to form a result for external release; The detection result external publishing module is used to send the external publishing results to the construction party or the management party.

Images