A method for verifying coal pile volume based on a single two-dimensional image of a coal yard

Abstract

The present invention discloses a method for verifying the volume of a coal pile based on a single two-dimensional image of a coal yard, which relates to the technical field of coal yard management and solves the technical problem of low coal counting efficiency based on computer vision measurement technology. The key point of its technical solution is to train multiple coal pile photos to achieve the effect of directly verifying the volume of the coal pile through only a single image without depth of field in the subsequent algorithm implementation process, thereby ensuring real-time performance. In this process, the image only requires a monocular camera to be taken, and there is no need to purchase multiple large-scale equipment such as binoculars, depth of field cameras or laser coal counting meters, which ensures that the algorithm achieves the economic goal. Finally, the purpose of verifying the volume of the coal pile more accurately and quickly is achieved by designing an algorithm framework. The present invention realizes the verification of the volume of the coal pile in the coal yard only through a single two-dimensional image, thereby reliably counting the coal, with low cost, convenience and speed, and can realize real-time monitoring of the coal yard.

Description

Technical Field

[0001] The present application relates to the technical field of coal yard management, and in particular to a method for verifying the volume of a coal pile based on a single two-dimensional image of a coal yard. Background Art

[0002] While various renewable energy projects are booming in the current power industry, traditional thermal power generation remains the primary source of power generation, accounting for over 60% of all power generation. However, coal mining accounts for the largest share of power generation costs for traditional thermal power plants, accounting for over 75% of total production costs. This significantly impacts the cash flow of power plants during daily operations, making coal resource management a crucial priority for the power industry.

[0003] Power plants make production scheduling, procurement, and inventory decisions based on production demand and existing coal reserves. This not only effectively optimizes coal inventory, avoids capital backlogs, and improves capital turnover, but also fundamentally addresses fuel management flaws such as lost tons and potential losses, achieving a balanced budget. Furthermore, with coal prices rising and power generation companies' profits declining, large coal-fired power plants are generally adopting a blended combustion strategy to reduce fuel costs and improve economic efficiency. Timely inventory of various coal types in coal yards also provides a technical means for accurate coal blending and provides a crucial basis for scientifically and rationally organizing production, economic scheduling, cost analysis, and economic efficiency evaluation at coal-fired power plants.

[0004] Since the 1990s, many universities, power plants and some industrial enterprises have carried out research on the precise measurement of coal storage in coal yards. According to different measurement principles, various research institutions have developed different coal counting systems, including automatic measurement systems composed of pulse ultrasonic ranging sensors, data acquisition instruments and microcomputers; coal counting systems based on prismless total stations; photoelectric detection systems composed of lasers and linear array CCD cameras; laser portable coal counting systems, etc., which have been widely used in the inventory of coal piles in resource-based enterprises such as electricity, steel, and metallurgy. To a certain extent, they have solved the problem of automated inventory of coal storage in power plants. However, there are also common problems such as large workload of coal counting, large measurement errors, inventory work being greatly affected by weather and environment, and only periodic inventory, which cannot reflect the status of coal storage and different types of coal in coal yards in real time and dynamically.

[0005] There are various technical implementation paths for existing coal mine inventory, among which laser scanning measurement technology and computer vision-based measurement technology are the main ones, and others include ultrasonic-based measurement technology and drone-based oblique photogrammetry technology.

[0006] In laser scanning measurement technology, due to the high price of three-dimensional laser scanners, domestic laser coal counting products all use two-dimensional laser scanners. According to the different platforms on which the two-dimensional laser scanners are mounted, laser coal counting systems can be divided into portable laser coal counting systems, fixed laser coal counting systems, and laser coal counting systems based on walking robots. Figure 1 As shown, when each system is working, a two-dimensional laser scanner is used to perform cross-sectional scanning on the surface of the coal yard to obtain the two-dimensional coordinate data of the cross-sectional contour of the coal yard, and finally match it with the position information obtained by the displacement sensor or angle sensor. Through a series of software processing, the surface morphology and three-dimensional coordinate data of the measured target are obtained, and then the coal yard reserves are calculated, the three-dimensional model is established and displayed, and finally the coal yard morphology data is obtained to achieve coal management.

[0007] Laser scanning coal counting technology is mature and widely used, but it still faces various limitations. More accurate coal counting requires large-scale equipment. For example, fixed laser coal counting systems utilize bucket wheel excavators for scanning. Compared to portable coal counting instruments, these systems often suffer from heavy workloads, large measurement errors, and inventory management that is significantly affected by weather and environmental factors. Inventory management can only be performed periodically, and the system cannot dynamically reflect the status of coal in the coal yard or different types of coal in real time. While these systems improve accuracy and convergence of results, they significantly increase the time required for coal counting and the cost of equipment operation and maintenance.

[0008] Ultrasonic coal measurement relies on calculating distance by multiplying the time difference between an ultrasonic wave's return from an obstacle and its propagation speed. However, due to the inability to accurately control the timing during ultrasonic measurement, measurement errors can occur. This is particularly true when precise measurement of points over a large area is required. Furthermore, the propagation speed of ultrasound varies depending on the transmission medium, leading to even greater uncertainty. The presence of dust in coal yards can also significantly affect measurement.

[0009] Whether it's coal panning using laser scanning or oblique photography, portable, fixed, or drone-based methods, each has its own advantages and disadvantages and is suitable for different operating scenarios. However, existing coal panning technology, product time, and maintenance costs cannot compare to current computer vision methods.

[0010] The measurement technology based on computer vision uses a laser rangefinder to measure the distance, and a digital camera to shoot around the entire coal yard. The images are stitched together to form a complete coal yard, thus completing the image acquisition work. The collected scattered coordinate data of the coal yard surface are then interpolated, and then interpolated by image fitting to complete the 3D reconstruction and volume calculation of the coal yard. The existing computer vision measurement technology is mainly based on 3D reconstruction work based on visual SLAM. Its technical implementation path is as follows: Figure 2shown.

[0011] For Panmei, the 3D reconstruction work is mainly done through Structure from Motion (SFM), which is a process of calculating the camera parameters from image sequences acquired from different angles to restore the 3D structure of the scene. The process usually uses the epipolar geometric relationship between feature matching points between two or more views to restore the camera position and posture information and the 3D structure of the scene. The camera position and posture information is the external parameters of the camera, including the camera rotation angle and camera translation during the shooting process. The 3D structure of the scene is described by a 3D point cloud. By mapping a certain number of 2D image feature points to a 3D space coordinate system, a 3D point cloud is formed to restore the 3D structure of the scene. The mapping relationship from the 2D plane to the 3D space is represented by a projection matrix composed of the camera external parameters and the camera internal parameters. The camera internal parameters include the focal length of the camera, the coordinates of the principal point, and the pixel size. The internal parameters of the camera are obtained through camera calibration. The process of the traditional structure from motion algorithm is as follows: Figure 3 As shown, a series of coordinate transformations are designed to map the feature points of the two-dimensional image to the three-dimensional space coordinate system, but it is ultimately a transformation from the pixel coordinate system to the world coordinate system. According to the above-mentioned camera calibration, feature point extraction and matching, and epipolar geometry principles, the camera internal parameter matrix, matching point pixel coordinates, camera rotation matrix, and camera translation vector are obtained respectively. These parameters can be used to restore the camera projection matrix, find the mapping relationship between the two-dimensional pixel coordinate points and the three-dimensional space coordinate points, solve the three-dimensional space coordinates corresponding to the image pixel points, generate a three-dimensional point cloud, and restore the three-dimensional structure of the scene.

[0012] While computer vision measurement reduces equipment construction and maintenance costs compared to laser scanning and ultrasonic-based measurement techniques, traditional methods, such as structure-from-motion (SMO) for 3D reconstruction of coal piles, still require extensive preparation. These issues include insufficient robustness of global SMO methods. The L1 norm of the rotation matrix is ​​relatively robust to outliers, while the relative translation relationship when solving for camera positions is sensitive to matching outliers. There are also issues with scene integrity, filtering extremal geometric edges, and the potential loss of partial images. Furthermore, calibration of system external parameters is required to determine the spatial information of each scene, as well as sensitivity to the selection of initial image pairs and the order in which cameras are added. Scene drift can lead to cumulative errors during large-scale reconstruction. Furthermore, efficiency is limited, as repeated bundle adjustments require significant computational time. These issues hinder the ability of existing computer vision-based techniques to monitor coal piles in real time, increasing the time cost of monitoring and reducing the efficiency of coal monitoring in power plants.

[0013] In summary, computer vision-based measurement technology is difficult to implement due to the large number of images from different viewpoints required during image sampling, resulting in a cumbersome reconstruction process and complex target feature point extraction. However, due to its low equipment cost and high updateability, it still has significant room for development with the rapid advancement of contemporary information technology. Therefore, addressing the existing issues in computer vision measurement technology to improve coal sorting efficiency is the goal of this application. Summary of the Invention

[0014] The present application provides a method for verifying the volume of a coal pile based on a single two-dimensional image of a coal yard. The technical purpose is to reduce the number and time of image capture required for traditional three-dimensional reconstruction, and to improve the accuracy of computer coal counting while reducing computer processing time, thereby improving the efficiency of coal counting in the coal yard.

[0015] The above technical objectives of this application are achieved through the following technical solutions:

[0016] A method for verifying the volume of a coal pile based on a single two-dimensional image of a coal yard comprises:

[0017] S1: Acquire multiple batches of monocular photos of the coal pile, and acquire volume data of the coal pile corresponding to the monocular photos;

[0018] S2: constructing a data set based on the monocular photo and the coal pile volume data, and preprocessing the data set to divide it into a training set and a test set;

[0019] S3: setting parameters of the coal pile volume verification model, and then training the coal pile volume verification model using the training set to obtain a trained coal pile volume verification model;

[0020] S4: Input the test set into the trained coal pile volume verification model for testing. When the predicted volume output by the trained coal pile volume verification model reaches a preset accuracy, a final coal pile volume verification model is obtained.

[0021] S5: Calibrate the coal pile volume using the final coal pile volume calibration model.

[0022] The beneficial effects of this application are:

[0023] (1) Volume measurement can be achieved by collecting only a single image and volume data, which significantly reduces the time cost required for traditional coal inventory in power plants, achieves the effect of real-time inventory of coal piles, reduces management costs and operation and maintenance costs, and improves accuracy.

[0024] (2) The algorithm does not require much upfront training cost. It only requires multiple batches of coal pile images from the relevant coal yard and the corresponding volume data of the coal piles for algorithm training.

[0025] (3) There is no need to purchase additional special equipment, such as binocular cameras, depth-of-field cameras, etc., and the equipment purchase cost and maintenance cost are low.

[0026] (4) It can be used in different types of coal yards and has a wide range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of a cross-sectional scan of a coal yard surface by a two-dimensional laser scanner in the prior art;

[0028] Figure 2 It is a flow chart of computer vision SLAM technology in the prior art;

[0029] Figure 3 This is a schematic diagram of the implementation of computer vision SLAM technology in the existing technology;

[0030] Figure 4 A flowchart of the method described in this application;

[0031] Figure 5 This is a two-dimensional picture diagram of a coal pile;

[0032] Figure 6 Schematic diagram of the residual block structure;

[0033] Figure 7 Flowchart for training a regression residual network;

[0034] Figure 8 The coal yard data preparation map provided in the embodiment of this application;

[0035] Figure 9 This is a schematic diagram of the model training process provided in the embodiments of this application, i.e., the final loss calculation. DETAILED DESCRIPTION

[0036] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0037] like Figure 4 As shown, the method for verifying the volume of a coal pile based on a single two-dimensional image of a coal yard described in this application includes:

[0038] S1: Acquire multiple batches of monocular photos of the coal pile, and acquire volume data of the coal pile corresponding to the monocular photos.

[0039] Specifically, after obtaining the monocular photo and the corresponding coal pile volume data, they are traversed and read in a specific manner, and the monocular photo and the corresponding coal pile volume data are marked and saved.

[0040] When performing volume verification on coal piles in a coal yard, the requirements for coal pile data are as follows:

[0041] Two-dimensional image (i.e. monocular image): specifically refers to a monocular image with sufficient height and covering the entire coal yard shape information when the internal and external parameters of the camera are consistent. The pixel and internal parameter accuracy of the image should be sufficient to record the original image of the coal pile shape in the coal yard. After saving the coal pile image and point cloud data together, the algorithm will traverse the folder to which they belong, read multiple images in it, and fill them according to the pre-set pixel specifications (such as 230*180). The two-dimensional image of the coal pile is as follows: Figure 5 shown.

[0042] Coal pile volume data: Record the volume of the target coal pile in the above monocular image, and save the corresponding file name of the coal pile 2D image in a CSV file for easy reading by the algorithm. The format is that there are two columns in the CSV file. The first column stores the file name of the coal pile 2D image, and the second column is the volume of the coal pile in the image corresponding to the file name (no volume unit is required).

[0043] S2: constructing a data set based on the monocular photograph and the coal pile volume data, and dividing the data set into a training set and a test set after preprocessing.

[0044] As a specific embodiment, the ratio of the training set to the test set is 9:1.

[0045] Specifically, preprocessing the data set includes: performing pixel filling on the monocular image and normalizing the coal pile volume data.

[0046] When filling pixels in a monocular image, this step requires establishing the parameters for feature extraction from the acquired sample data. The monocular image is read into a tensor according to the RGB matrix channels as follows:

[0047]

[0048] Among them, P i represents the matrix of the i-th monocular image; p HW Represents the value of each pixel in the monocular image, and p HW ∈[0,255]; H represents the total horizontal resolution of the monocular image, and W represents the total vertical resolution of the monocular image, that is, the image has a resolution of H*W. i Does not include gradient information, save format for reading input 2D image information.

[0049] P i As input information, it is necessary to pre-expand. Due to the undulation of coal pile boundaries and the uncertainty of coal yard size, the image data in the dataset needs to be padded to ensure the same data samples and provide balanced data for subsequent training. The specific filling method is: fill 0 to the matrix P iTo form a three-channel 230*180 matrix, the filled matrix is ​​expressed as:

[0050]

[0051] Normalize the coal pile volume data. This process will help with analysis. This process will traverse the volume column of the CSV file and address the maximum and minimum values ​​for storage, and strictly follow the maximum and minimum normalization (Min-Max Normalization) for normalization. This method has the following advantages: (1) Normalization speeds up the gradient descent to find the optimal solution, that is, speeds up the convergence of the training network; (2) Normalization improves the accuracy of the target calculation. The purpose of normalization in this process is to limit the preprocessed data to a certain range (such as [0,1] or [-1,1]), thereby eliminating the adverse effects caused by singular sample data.

[0052] S3: Setting the parameters of the coal pile volume verification model, and then training the coal pile volume verification model using a training set to obtain a trained coal pile volume verification model.

[0053] Specifically, the coal pile volume verification model is an improved regression residual network, which includes 4 residual blocks connected in sequence and a 1*1 fully connected layer. Residual networks were originally used in visual classification tasks, but this application uses them in actual physical engineering problems. By adding the original input image information and volume as the algorithm training set, multiple training and corrections are performed when the reconstruction does not meet the requirements. This solution can improve the accuracy of the algorithm through repeated training. In this process, in order to improve the algorithm running speed and algorithm simplicity, and to solve the problem of overfitting caused by selecting a shallow residual network due to a small number of samples, a reference residual network was finally selected. Each residual block includes 1 convolutional layer and 1 fully connected layer connected in sequence. The fully connected layer includes linear processing, nonlinear processing and linear processing in sequence, such as Figure 6 The output of the residual block is the sum of the output of its fully connected layer and the input of its convolutional layer. The input of the convolutional layer of residual block 1 is the preprocessed training set data, and the output of residual block 1 is the sum of the training set data and the output of its fully connected layer. The input of residual block 2 is the output of residual block 1, and so on, until the output of residual block 4 is output through a 1*1 fully connected layer.

[0054] like Figure 6As shown in the figure, an information flow is branched from the input stream Input to the output stream Output. This operation will prevent the gradient from disappearing during the gradient calculation during back propagation, thereby increasing the efficiency of back propagation and the frequency of model parameter correction. The model is built with reference to the above model structure and the network is debugged. The input has been debugged in the front part. As for the output, the traditional Resnet will use the Softmax layer for output in the classification task. This application improves the last fully connected layer and adjusts the final output to a fully connected layer with a dimension of 1. The algorithm flow is as follows Figure 7 shown.

[0055] The coal pile volume verification model is trained using the training set, including: inputting the preprocessed training set into the residual block, and finally outputting the predicted volume through a 1*1 fully connected layer; determining the error loss based on the coal pile volume data and the predicted volume in the input training set, and backpropagating the parameters using the loss function until the coal pile volume verification model reaches the preset accuracy, completing the training and outputting the trained coal pile volume verification model.

[0056] like Figure 4 and Figure 7 As shown, when the coal pile volume verification model is trained using a training set, the training set is first input into the coal pile volume verification model, and feature extraction is performed on the monocular photos in the training set to obtain feature images. Three-dimensional spatial information features are obtained through the feature images and the coal pile volume data. Then, the predicted volume is obtained based on the three-dimensional spatial information features, and a judgment is made as to whether the predicted volume is qualified. If qualified, the predicted volume is output. If unqualified, the parameters of the coal pile volume verification model are optimized until the training of the coal pile volume verification model is completed, and the trained coal pile volume verification model is obtained.

[0057] As a specific embodiment, common hyperparameter settings of the regression residual network are shown in Table 1:

[0058] Batch_size (number of samples used for training in a single pass) Shuffle Learnrate 10 False 1e-4

[0059] Table 1 Other hyperparameter settings for network design

[0060] like Figure 7 As shown, after the regression residual network outputs the diffusion result, that is, the predicted volume, the error loss is determined according to the volume parameters initially input, and the error function performs backpropagation of the parameters. The loss calculation in this stage mainly includes the loss calculation of the predicted volume and the pre-stored coal pile volume data. In this application, the loss function selects the L1loss and MSEloss functions that come with Pytorch to calculate together. The final loss is the sum of the two functions, and the loss output of the current training is recorded during the training process.

[0061] The final loss is expressed as:

[0062]

[0063] in, Represents the target prediction value received by the function; y represents the target true value received by the function, that is, the absolute value loss. The backpropagation efficiency of the dual loss calculation is much higher than that of the common loss function calculation. The choice of the dual loss function is due to the fact that the single loss function has always been ineffective during model testing. The algorithm can only significantly reduce the loss after repeated training for 500 rounds. The dual loss function calculation effect is better. When the remaining settings refer to the design of this application, the training can converge in 200 or even 150 rounds, completing the adjustment of the model parameters.

[0064] After the loss calculation, the learning rate of the back propagation process is adjusted to select the Adam algorithm in combination with the MultiStepLR algorithm. The settings are shown in Table 2:

[0065] Optimizer Milestones Gamma optimizer [50,200] 0.5

[0066] Table 2 MultiStepLR algorithm hyperparameter settings

[0067] In Table 2, Optimizer represents the training recorder, Milestones represents the learning rate adjustment interval, and Gamma represents the adjustment ratio.

[0068] After completing the training of the coal pile volume verification model, save the .pth training file to the running path, and then save the test set test results to a CSV file for result display.

[0069] S4: The test set is input into the trained coal pile volume verification model for testing. When the predicted volume output by the trained coal pile volume verification model reaches a preset accuracy, the final coal pile volume verification model is obtained.

[0070] Specifically, when the test set is input into the trained coal pile volume verification model for testing, the test set is first input into the trained coal pile volume verification model, and feature extraction is performed on the monocular photos in the test set to obtain feature pictures. The three-dimensional spatial information features are obtained through the feature pictures and the coal pile volume data. The predicted volume is then obtained based on the three-dimensional spatial information features, and a judgment is made on whether the predicted volume is qualified. If qualified, the predicted volume is output. If unqualified, the parameters of the trained coal pile volume verification model are continued to be optimized until the output predicted volume reaches the preset accuracy, thereby obtaining the final coal pile volume verification model.

[0071] S5: Calibrate the coal pile volume using the final coal pile volume calibration model.

[0072] The following will describe the technical solution of this application in detail with a specific operation example. The data source is the image generated by the coal pan of a power plant, such as Figure 8 As shown, the coal yard is strip-shaped, with most sizes measuring 50m*170m. It is required to collect images with as few blank areas as possible, and then input the volume data obtained from the coal panning. The specific workflow is as follows:

[0073] (1) Collect image data and mark each image with a name. Here, the image is collected using the spherical coordinates [0 50 100] with the center of the coal yard as the sphere center.

[0074] (2) Collect the volume data of the coal pile and save it into a CSV file according to the corresponding image name mark to facilitate algorithm reading. The reading tool uses pdf.to_csv to read the volume data.

[0075] (3) There are 224 documents in total. The last 10 documents are selected as the test set, and the remaining 214 documents are used as the training set.

[0076] (4) Read the first 214 files and save the name, image information RBG matrix, and final volume of each file into the same tensor.

[0077] (5) Run the program and wait for the network training to complete. After the training is completed, the algorithm loss function graph and the test set results will be generated, such as Figure 9 shown.

[0078] The final output of the algorithm is shown in Table 3:

[0079] Name True volume Predict volume error 1A_20210531_093941.png 10125 9713.922 4.06％ 1A_20210630_092853.png 19609 20407.951 4.07％ 1A_20210731_102908.png 22857 23178.143 1.41％ 1A_20210831_112055.png 30654 31798.453 3.73％ 1A_20210930_093850.png 23266 23917.963 2.80％ 1A_20211029_144900.png 14082 13500.999 4.13％ 1A_20211129_091622.png 18883 19154.275 1.44％ 1A_20211231_102226.png 19506 19562.453 0.29％ 1A_20220130_085010.png 22926 22796.656 0.56％ 1A_20220228_085408.png 10583 9749.756 7.87％

[0080] Table 3 Final prediction results of the coal pile volume verification model

[0081] Among them, Name is the name of the coal field, True volume is the volume data obtained after the coal is sorted by large equipment, Predict volume is the volume predicted by the model, and Error is the error between the two. The maximum error measured in the test set is +7.87%, and the average error is +3.04%. Subsequently, 224 coal field reselection test sets were used and the experiment was repeated to obtain the results in Table 4 and

[0082] Table 5.

[0083]

[0084]

[0085] Table 4. Test results of the model reselection test set.

[0086] Name Truevolume Predict volume error 1B_20210531_093941.png 8756 8824.127 0.78％ 1B_20210630_092853.png 28270 29257.168 3.49％ 1B_20210731_102908.png 10905 11234.913 3.03％ 1B_20210831_112055.png 17943 18306.777 2.03％ 1B_20210930_093850.png 25447 25758.992 1.23％ 1B_20211029_144900.png 22130 23043.742 4.13％ 1B_20211129_091622.png 14522 15016.995 3.41％ 1B_20211231_102226.png 25173 26479.85 5.19％ 1B_20220130_085010.png 16879 17485.895 3.60％ 1B_20220228_085408.png 8117.0005 7591.444 6.47％

[0087] Table 5 Model repeated experimental data display

[0088] It can be seen from Tables 4 and 5 that the comprehensive average error after multiple tests is 2.93%. The coal pile volume verification model performs well and can meet the needs of practical application.

[0089] The above are exemplary embodiments of the present application, and the protection scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for verifying the volume of a coal pile based on a single two-dimensional image of a coal yard, characterized in that: include: S1: Acquire multiple batches of monocular photos of the coal pile, and acquire volume data of the coal pile corresponding to the monocular photos; S2: constructing a data set based on the monocular photo and the coal pile volume data, and preprocessing the data set to divide it into a training set and a test set; S3: setting parameters of the coal pile volume verification model, and then training the coal pile volume verification model using the training set to obtain a trained coal pile volume verification model; S4: Input the test set into the trained coal pile volume verification model for testing. When the predicted volume output by the trained coal pile volume verification model reaches a preset accuracy, a final coal pile volume verification model is obtained. S5: calibrating the volume of the coal pile using the final coal pile volume calibration model; The preprocessing of the data set includes pixel filling of the monocular photo and normalization of the coal pile volume data; Perform pixel padding on the monocular photo and expand the RGB matrix of the monocular image into a three-channel 230*180 matrix, expressed as: ； in, Represents the matrix of the i-th monocular image after pixel filling, represents the value of each pixel in the monocular image, and , Indicates the total horizontal resolution of the monocular image before expansion, Indicates the vertical total of the monocular image resolution before expansion; The coal pile volume verification model is an improved regression residual network, which includes four residual blocks connected in sequence and a 1*1 fully connected layer. The coal pile volume verification model is trained using a training set, including: The preprocessed training set is input into the residual block, and finally the predicted volume is output through a 1*1 fully connected layer; The error loss is determined based on the coal pile volume data and predicted volume in the input training set, and the parameters are back-propagated by the loss function until the coal pile volume verification model reaches the preset accuracy. The training is completed and the trained coal pile volume verification model is output.

2. The method according to claim 1, wherein The ratio of the training set to the test set is 9:

1.

3. The method according to claim 1, wherein The loss function includes L1loss function and MSEloss function, which are expressed as: ； in, Represents the target prediction value received by the function; Represents the target true value accepted by the function, that is, the absolute value loss.

4. The method according to claim 3, wherein Each residual block includes a convolutional layer and a fully connected layer connected in sequence, and the fully connected layer includes linear processing, nonlinear processing and linear processing in sequence; the output of the residual block is the sum of the input of its convolutional layer and the output of its fully connected layer.

5. The method according to claim 3, wherein In the back propagation process, the learning rate is adjusted by using the Adam algorithm in combination with the MultiStepLR algorithm.

Images