An image-based intelligent prediction system and method for strength of slag-based filling body

Abstract

The present application provides an image-based intelligent prediction system and method for strength of slag-based filling body, and relates to the technical field of filling mining in metal mines. The present application mainly establishes the connection between the color and strength of the filling body through image intelligent prediction, and then predicts the in-situ mechanical parameters of the filling body, so as to achieve the purpose of quickly and accurately completing the in-situ mechanical parameter test. The present application only needs to establish the linear relationship and regression model between the filling body color and uniaxial compressive strength for the first time, and can then perform rapid strength detection. The present application organically combines image acquisition, in-situ strength test and deep learning algorithm, further performs overall evaluation on the safety of the filling body in the mine, and promotes the technical development of safety evaluation and risk identification in the mine.

Description

Technical Field

[0001] This invention relates to the field of metal mine backfilling mining technology, and in particular to an image-based intelligent prediction system and method for the strength of slag-based backfill bodies. Background Technology

[0002] Mine backfilling is a crucial guarantee for the effective recovery of mineral resources and the sustainable development of the mining industry. In backfill mining, to ensure safe production operations, the backfill body must possess high stability, and the support system based on the backfill body must effectively control rock strata movement and deformation. The stability of the backfill body affects the safe production of deep mines, and its strength is a key indicator for quantifying this stability; strength testing is the foundation for evaluating backfill stability. Given the overall trend of mining in my country continuously advancing to deeper levels, the increase in ground stress makes determining the stability of the backfill body particularly important. In recent years, to improve backfill quality and reduce costs, some large-scale modern mining enterprises in China have begun to use slag-based binders to replace cement. Slag-based binders have advantages such as low CO2 emissions, high strength, good corrosion resistance, low hydration heat release, and low cost, offering significant economic and technical benefits. Given the widespread application of slag in backfilling, how to conduct on-site evaluation of the strength of slag-based backfill bodies is of great significance for underground mining safety assessment.

[0003] How to quickly and accurately test the strength of infill bodies is a key issue in stability evaluation. Currently, conventional infill body strength testing mainly uses two methods: in-situ core sampling and laboratory strength testing. In-situ core sampling involves drilling holes on-site with drilling equipment to extract infill body samples, which are then processed into standard dimensions and transported to a laboratory for strength testing to obtain the physical and mechanical parameters of the samples. However, since most mines do not have high-precision presses and other equipment, and are limited by the cumbersome procedures of drilling, core sampling, and sample preparation, various factors such as vibration during core sampling and bumps and knocks during transportation and processing can damage the integrity of the samples. In addition, the test requires high dimensional flatness, which can easily affect the accuracy of the data. Furthermore, the daily production activities of the mine are also affected to some extent by the limitations of underground space and ventilation conditions. Laboratory strength testing involves preparing standard-sized infill bodies in the laboratory according to the filling process flow and using the same mix proportions. The samples are cured under conditions close to the on-site temperature and humidity. At different curing ages, the mechanical parameters of the infill body under the selected filling process are evaluated according to standard mechanical test methods. While indoor strength testing can accurately obtain the required mechanical parameters, the deposition, consolidation, and curing processes of infill slurry in the mining area differ from those in surface laboratories. The constant temperature and humidity environment of laboratory curing cannot replicate the complex temperature, humidity, and overlying strata pressure experienced in actual conditions. Furthermore, considering the size effect and non-uniform characteristics of the infill material leading to uneven strength distribution, the mechanical parameters obtained from infill material curing using only standard indoor specimens will deviate significantly from actual values, resulting in limited practicality. Therefore, exploring a method for quickly and easily predicting infill material strength in the field is imperative. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an image-based intelligent prediction system and method for the strength of slag-based backfill, which addresses the shortcomings of the prior art. The system mainly establishes a relationship between the color and strength of the in-situ backfill through intelligent image processing, thereby predicting the strength parameters of the slag-based backfill and achieving the purpose of rapid and accurate in-situ mechanical parameter testing.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] On the one hand, the present invention provides an image-based intelligent prediction system for the strength of slag-based backfill, including a strength and image information acquisition module, a data processing and correlation module, an intelligent model construction module, and a measured data analysis module.

[0007] The strength and image information acquisition module is used to conduct similar experiments in the laboratory, collect sample strength data and capture original images.

[0008] The data processing and association module is used to obtain color features based on the RGB values ​​obtained from the original image processing, and to establish a color feature-intensity relationship database with the intensity data.

[0009] The intelligent model building module is used to build an analysis model based on a feedforward neural network. The color feature-intensity relationship database is input into the model, and the algorithm is optimized to obtain a perfected model.

[0010] The measured data analysis module is used to input actual data into the optimized model for processing to obtain intensity prediction results.

[0011] Furthermore, the strength and image information acquisition module is equipped with common laboratory experimental equipment, capable of preparing a large number of filling body samples under different variables, and the high-precision mechanical testing machine can accurately obtain the sample strength parameters; the main part of the module is a CCD industrial camera and a portable camera box, used to capture the original image of the sample; the portable camera box has a built-in light source; providing a calibration background environment for shooting.

[0012] Furthermore, the data processing and association module preprocesses the images to be acquired using color image filtering to reduce noise. Based on the actual situation, it selects any algorithm from gamma transform, logarithmic transform, and histogram equalization to clearly distinguish between the high-resolution RGB image background and the filling samples, improving the contrast between the target and the background and avoiding image recognition confusion. A binary image is obtained by thresholding the original digital image. The resulting binary image uses image processing software to automatically select algorithms to complete two processes: edge detection and image segmentation. Since there are many algorithms available for these processes and related research is mature, they will not be elaborated further. After segmentation, the filling image obtains the pixel mean value as a color feature by assigning a color histogram to each color channel. A color feature-intensity relationship database is then established by correlating the color features with the corresponding filling intensity.

[0013] The color channels include 10 channels, namely R, G, B channels, and G / R, G / B, B / R, (GR) / (R+G), (GB) / (G+B), (RB) / (B+R), and G / (R+G+B) color channels.

[0014] Furthermore, the intelligent model construction module employs a manual sampling method combining Bayesian optimization and oversampling / undersampling to establish a feedforward neural network model for predicting the color feature-intensity relationship of the infill body, specifically:

[0015] Based on a color feature-intensity relationship database, the data is divided into training and testing sets. Color features in the RGB color space are selected as input parameters, and infill intensity is selected as the output parameter to establish a feedforward neural network model to predict intensity. After determining the input and output layers, the number of hidden layers and the number of neurons in each hidden layer are set according to the complexity of the actual situation. The average value of the cross-entropy loss function after 5 cross-validations is used as the objective function, and a Bayesian optimization algorithm is used for network optimization. At the same time, a manual sampling method combining oversampling and undersampling is used to process the training set to eliminate data imbalance. A permutation feature importance algorithm is introduced to analyze the relative importance of input variables, and finally the optimal feature combination is selected from the color features. Finally, the optimized feedforward neural network model is obtained.

[0016] Furthermore, the measured data analysis module uses the same shooting equipment and conditions as the strength and image information acquisition module to complete the acquisition of the original image of the infill body to be tested in the field, and performs the same steps as the data processing and association module to perform noise reduction and segmentation on the original image and extract the corresponding color feature data; the color feature data is then input into the optimized feedforward neural network model for processing, thereby obtaining the prediction result of the infill strength.

[0017] On the other hand, the present invention also provides an image-based intelligent prediction method for the strength of slag-based backfill bodies, implemented based on the above-mentioned image-based intelligent prediction system for the strength of slag-based backfill bodies, comprising the following steps:

[0018] Step 1: Conduct similar experiments in the laboratory with varying group ratio parameters, and collect sample strength data and corresponding original images;

[0019] Step 2: Obtain color feature parameters in the RGB color space based on the original image processing described in Step 1, and establish a color feature-intensity relationship database with the corresponding intensity data;

[0020] Step 3: Construct an analysis model based on a feedforward neural network, input the color feature-intensity relationship database into the model, complete algorithm optimization, and construct an image-based intensity prediction model;

[0021] Step 4: Input the actual data into the optimized model for processing to obtain the intensity prediction results.

[0022] Furthermore, the specific method of step 1 is as follows:

[0023] Similar experiments were conducted based on variations in the composition parameters of the actual backfilling mix design. These parameters included particle size distribution, binder content, mass concentration, component ratio, curing time, and temperature. Tailings and binder were obtained from the mining company. Four factors—mass concentration, binder content, component ratio, and curing time—were selected as independent variables. Response surface methodology was used to conduct surface laboratory tests. The strength performance indicators of the backfill were obtained using a high-precision mechanical testing machine. High-resolution RGB images were acquired using a CCD industrial camera and a portable camera.

[0024] Furthermore, the specific method for step 2 is as follows:

[0025] The obtained image is preprocessed using a color image filtering method. Based on the organic relationship between the R, G, and B components of the color image, the multidimensional color vector of the pixels is directly processed to eliminate noise in the image and avoid the influence of noise on image enhancement. Gamma transformation is used to improve the contrast of the image for accurate target identification in the subsequent process. In the RGB color space, based on the G channel image, a 2G-RB filtering algorithm that emphasizes the green component and suppresses the other two components is used to segment the infill body from the background image. After segmentation, the infill body image is processed by obtaining the pixel mean value as a color feature by mapping a color histogram for each color channel.

[0026] The color channels include 10 channels: R, G, B channels, and 7 new channels G / R, G / B, B / R, (GR) / (R+G), (GB) / (G+B), (RB) / (B+R), and G / (R+G+B) derived from the R, G, and B channels. The average pixel value is obtained from these 10 channel images, and 10 color features are extracted.

[0027] Furthermore, the specific method of step 3 is as follows:

[0028] The data in the color feature-intensity relationship database was randomly divided into training and test sets. Ten color features in the RGB color space were selected as input parameters for the neural network, and the infill strength was selected as the output parameter. A feedforward neural network model was established to predict the intensity. After the input and output layers were determined, the number of hidden layers and the number of neurons in each hidden layer were set according to the complexity of the actual situation. The average value of the cross-entropy loss function after five cross-validations was used as the objective function, and the network was optimized using a Bayesian optimization algorithm. The training set was divided into five equal parts of the same size. Four parts were selected for training, and the remaining parts were used for validation. The process of dividing the training set and randomly selecting training and validation parts was repeated five times, and the average loss of the five validations was taken. At the same time, a manual sampling method combining oversampling and undersampling was used to process the training set to eliminate the influence of the imbalance of intensity parameters due to the scarcity of light colors in the training set. The permutation feature importance algorithm was used to analyze the relative importance of the input variables, and the optimal feature combination was finally selected from the ten color features. Finally, the optimized feedforward neural network model was obtained and used to predict the intensity of infill bodies in actual mines.

[0029] Furthermore, the specific method of step 4 is as follows:

[0030] The original images of the filling body to be tested were captured on-site at the mine. The process was as follows: a filling body sample with uniform color and area similar to that of the laboratory experiment was selected, placed in a portable camera box, and the original image of the filling body was obtained according to the shooting conditions in step 1; according to the image processing process in step 2, the data of the corresponding color features of the filling body were extracted and input into the optimized model in step 3 for processing to obtain the prediction result of the strength of the filling body.

[0031] The beneficial effects of adopting the above technical solution are as follows: The present invention provides an image-based intelligent prediction system and method for the strength of slag-based backfill bodies. It mainly establishes a relationship between the color and strength of the backfill body through intelligent image prediction, thereby predicting the in-situ mechanical parameters of the backfill body. It features a strength and image information acquisition module, anti-interference calibration color sampling, simple operation, and the ability to quickly create a calibration photography environment, avoiding the influence of various factors such as light source dust, ambient temperature, and humidity in complex underground environments. It acquires high-resolution RGB images, ensuring the authenticity of the in-situ color sampling of the backfill body. It also features a data processing and correlation module, which can intelligently match the calibration environmental conditions to perform denoising and segmentation steps on the obtained images, quantifying the color distribution of each part of the target to improve the accuracy of color feature recognition. Furthermore, it features an intelligent model construction module, which constructs a color feature-intensity model of the backfill body image based on the big data color features of the backfill body combined with computational programming technology. Deep learning is used to establish a color feature-intensity relationship feature library, and then the color model is coupled to analyze and select the optimal results in the feature library to predict the in-situ mechanical parameters of the backfill body. It has a measured data analysis module, which has no requirements on the size, shape, or integrity of the sample. It is easy to operate; you only need to take a picture to obtain the RGB value and input it into the measured data analysis module. It is low in cost and highly accurate.

[0032] This invention leverages the theoretical advantages of neural networks in learning and mining sample features to provide a rapid method for predicting the strength of backfill materials. This solves the problem of mines relying on traditional borehole inspection, while simultaneously improving identification efficiency and reducing operational risks. This invention only requires establishing a linear relationship and regression model between backfill material color and uniaxial compressive strength initially to perform rapid strength detection. It organically combines image acquisition, in-situ strength testing, and deep learning algorithms to further conduct a comprehensive assessment of the safety of mine backfill materials, thus promoting the technological development of mine safety assessment and risk identification. Attached Figure Description

[0033] Figure 1 This is a diagram showing the strength difference of the filling material under different colors, provided in an embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the method flow provided in an embodiment of the present invention. Detailed Implementation

[0036] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0037] A notable characteristic of slag-based backfill is its bluish-green color, resulting from the formation of ferrous ions during slag hydration. The intensity of this bluish-green color depends on factors such as the amount of binder, curing time, and curing temperature; a deeper color generally indicates higher strength. Figure 1 As shown. In view of this, considering the outdated testing methods for the strength parameters of slag-based backfill bodies, and the problems of traditional coring being cumbersome, labor-intensive, unable to conduct timely strength testing on-site, and time-consuming and labor-intensive sample transportation, this embodiment provides an image-based intelligent prediction system for the strength of slag-based backfill bodies, such as... Figure 2 As shown, it includes an intensity and image information acquisition module, a data processing and correlation module, an intelligent model construction module, and a measured data analysis module, which are connected in sequence.

[0038] The strength and image information acquisition module is used for similar experiments in the laboratory to collect sample strength data and capture raw images. This module is equipped with common laboratory equipment, enabling the mass production of infill samples under various variables. A high-precision mechanical testing machine accurately obtains specimen strength parameters. The main body of the module consists of a CCD industrial camera and a portable camera case for capturing raw images of the samples. The portable camera case has a built-in light source to provide a calibrated background environment for shooting.

[0039] The data processing and association module is used to obtain color features based on the RGB values ​​obtained from the original image processing and to establish a color feature-intensity relationship database with the intensity data. This module preprocesses the acquired image using color image filtering to reduce noise. Depending on the actual situation, it selects any one of the following algorithms: gamma transform, logarithmic transform, or histogram equalization, to clearly distinguish between the background and the infill sample in the high-resolution RGB image, improving the contrast between the target and the background and avoiding image recognition confusion. A binary image is obtained by thresholding the original digital image. The obtained binary image uses image processing software such as Matlab and OpenCV to perform edge detection and image segmentation using selected algorithms. This process offers numerous algorithm options and is well-researched. After segmentation, the infill image obtains the pixel mean value as a color feature by mapping each color channel to a color histogram. The color features are then correlated with the corresponding infill intensity to build a color feature-intensity relationship database. The color channels include 10 channels, namely R, G, B channels, and G / R, G / B, B / R, (GR) / (R+G), (GB) / (G+B), (RB) / (B+R), and G / (R+G+B) color channels.

[0040] The intelligent model building module is used to construct an analysis model based on a feedforward neural network. The color feature-intensity relationship database is input into the model to optimize the algorithm and improve the model. This module employs a manual sampling method combining Bayesian optimization and oversampling / undersampling to establish a feedforward neural network model for predicting the color feature-intensity relationship of infill bodies. Specifically:

[0041] Based on a color feature-intensity relationship database, the data is divided into training and testing sets. Color features in the RGB color space are selected as input parameters, and infill intensity is selected as the output parameter to establish a feedforward neural network model to predict intensity. After determining the input and output layers, the number of hidden layers and the number of neurons in each hidden layer are set according to the complexity of the actual situation. The average value of the cross-entropy loss function after 5 cross-validations is used as the objective function, and a Bayesian optimization algorithm is used for network optimization. At the same time, a manual sampling method combining oversampling and undersampling is used to process the training set to eliminate data imbalance. A permutation feature importance algorithm is introduced to analyze the relative importance of input variables, and finally the optimal feature combination is selected from the color features. Finally, the optimized feedforward neural network model is obtained.

[0042] The measured data analysis module is used to input actual data into the optimized model for processing, thereby obtaining the intensity prediction result. This module uses the same imaging equipment and conditions as the intensity and image information acquisition module to acquire original images of the infill body to be tested in the field. Following the same steps as the data processing and correlation module, the original images undergo noise reduction and segmentation, and the corresponding color feature data is extracted. The color feature data is then input into the optimized feedforward neural network model for processing, thereby obtaining the predicted intensity of the infill body.

[0043] The method for intelligently predicting the strength of slag-based backfill using the aforementioned image-based intelligent prediction system for slag-based backfill is as follows: Figure 3 As shown, it includes the following steps:

[0044] Step 1: Conduct similar experiments in the laboratory with varying group ratio parameters, collect sample strength data and corresponding original images. The specific method is as follows:

[0045] Similar experiments were conducted based on variations in the composition parameters (including particle size distribution, binder content, mass concentration, component ratio, curing time, and temperature) of the actual on-site backfill mix design. Tailings and binder were obtained from the mining company. Four factors—mass concentration, binder content, component ratio, and curing time—were selected as independent variables. Response surface methodology was used to conduct surface laboratory tests. The strength performance indicators of the backfill body were obtained using a high-precision mechanical testing machine. High-resolution RGB images were acquired using a CCD industrial camera.

[0046] The CCD industrial camera is compatible with Windows systems and supports multiple drivers including WDM and OpenCV, enabling a controllable electronic shutter without the need for an additional mechanical shutter. It uses a USB 3.0 output interface for rapid, continuous acquisition of high-resolution RGB images. The main components used are the Microvision MV-VDM120SM / SC industrial camera and the SANOTO B270 portable camera case. The selected equipment is mature and reliable, capable of handling complex external environments and maintaining stable operation.

[0047] For ease of use and setup, the CCD industrial camera can be directly fixed to the mounting bracket provided with the camera box, ensuring the camera remains horizontal. When the filling sample is placed in the center of the camera box, the sample should be approximately 20cm away from the camera. The distance can be quickly adjusted according to the actual situation to obtain a clear image.

[0048] To prevent image capture from being affected by lighting conditions, the camera flash is not used during image capture. The camera box has six built-in light sources, sufficient for most photographic requirements. After the initial light source brightness settings are completed, the adjustment data is automatically saved to ensure the calibrated light source effect. The camera box isolates external light sources and dust, and effectively reduces stray reflections from the rough surface of the filling material. The camera's built-in white balance software is used to process the white balance of the photos. During the shooting process, a white board is used as the background, and the camera resolution is set to 1280×960. The resulting RGB images have high clarity, sufficient for subsequent image segmentation, feature extraction, and recognition algorithms.

[0049] Step 2: Obtain color feature parameters in the RGB color space based on the original image processing, and establish a color feature-intensity relationship database with the corresponding intensity data. The specific method is as follows:

[0050] The obtained image is preprocessed using a color image filtering method. This method directly processes the multidimensional color vectors of pixels based on the organic relationship between the R, G, and B components of the color image to eliminate noise and avoid its impact on image enhancement. Gamma transform is used to improve image contrast for accurate target identification. Preliminary experiments show that the infill bodies are mostly green. In the RGB color space, a 2G-RB filtering algorithm, emphasizing the green component and suppressing the other two components, can be used to segment the infill bodies from the background image based on the G channel image. After segmentation, the infill body image is analyzed by generating a color histogram for each color channel, and the pixel mean value is obtained as a color feature. Based on the R, G, and B channels, seven new channels are obtained (G / R, G / B, B / R, (GR) / (R+G), (GB) / (G+B), (RB) / (B+R), G / (R+G+B)), where the image division is a dot-matrix division of each element of the image matrix. Similar to the R, G, and B channels, the average pixel value is derived from these 10 channel images. This allows us to extract 10 color features from the 10 RGB channels.

[0051] Step 3: Construct an analysis model based on a feedforward neural network. Input the color feature-intensity relationship database into the model, optimize the algorithm, and construct an image-based intensity prediction model. The specific method is as follows:

[0052] The data in the existing color feature-intensity database was randomly divided into a training set (80%) and a test set (20%). Ten color features from the RGB color space were selected as input parameters for the neural network, and infill strength was used as the output parameter to build a feedforward neural network (FNN) model to predict the intensity. After determining the input and output layers, the number of hidden layers and the number of neurons in each hidden layer were set according to the complexity of the actual situation. The number of hidden layers, the number of neurons in each hidden layer, and the maximum number of iterations have the greatest impact on the performance of the neural network. These hyperparameters were optimized using a Bayesian optimization algorithm to improve the performance of the neural network. Before optimizing the neural network, an objective function should be defined. The cross-entropy loss function is widely used in neural networks. The smaller the value of the cross-entropy loss function, the higher the model accuracy. To improve generalization ability and avoid overfitting, the average value of the cross-entropy loss function of 5 cross-validations (CV) was used as the objective function. The training set was divided into five equal parts of the same size. Four parts were randomly selected for training, and the remaining parts were used for validation. The process of dividing the training set and randomly selecting training and validation parts was repeated 5 times, and the average loss of the 5 validations was taken. Simultaneously, a manual sampling method combining oversampling and undersampling (SMOTETomek) was employed to process the training set, eliminating the influence of imbalanced intensity parameters due to fewer light-colored features. To explain the feedforward neural network, a permutation feature importance algorithm was used to analyze the relative importance of input variables, ultimately selecting the optimal feature combination from 10 color features. Finally, the optimized feedforward neural network model was obtained, and its practicality was verified by predicting the strength of mine backfill in the field.

[0053] Step 4: Input the actual data into the optimized model from Step 3 for processing to obtain the intensity prediction results. The specific method is as follows:

[0054] The process involves capturing original images of the infill material to be tested at the mine site. The procedure is as follows: A uniformly colored infill sample with an area similar to that of the laboratory sample is selected and placed in a portable camera case. Following the same shooting conditions as in step 1, the original image of the infill material is acquired. Then, following the same process as image noise reduction and enhancement in step 2, the corresponding color feature data of the infill material is extracted and input into the optimized analysis model in step 3 for processing, thereby obtaining the predicted result of the infill material strength.

[0055] This embodiment only requires establishing a linear relationship and regression model between the color of the filling body and the uniaxial compressive strength the first time to perform rapid strength detection. It organically combines image acquisition, in-situ strength testing and deep learning algorithms to further conduct an overall assessment of the safety of mine filling bodies, and promote mine safety assessment and risk identification.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. An image-based intelligent prediction system for the strength of slag-based backfill, characterized in that: The system includes an intensity and image information acquisition module, a data processing and correlation module, an intelligent model construction module, and a measured data analysis module; The strength and image information acquisition module is used to conduct similar experiments in the laboratory, collect sample strength data, and capture original images. The module includes common laboratory equipment for preparing filling samples under different variables. It also includes a high-precision mechanical testing machine for accurately acquiring sample strength parameters, a CCD industrial camera, and a portable camera case for capturing original images of the samples. The portable camera case has a built-in light source to provide a calibrated background environment for shooting. The data processing and association module is used to obtain color features based on the RGB values ​​obtained from the original image processing, and to establish a color feature-intensity relationship database with the intensity data. The data processing and association module preprocesses the image to be acquired using color image filtering to reduce noise in the image. Depending on the actual situation, it selects any one of the following algorithms: gamma transform, logarithmic transform, or histogram equalization, to create a clear distinction between the background and the filled sample in the high-resolution RGB image, improving the contrast between the target and the background and avoiding image recognition confusion. Binary values ​​are obtained by thresholding the original digital image. The resulting binary image is processed by image processing software using an algorithm to perform edge detection and image segmentation. After segmentation, the infill image is processed by obtaining the mean pixel value as a color feature by assigning a color histogram to each color channel. The color features are correlated with the corresponding infill intensity to build a color feature-intensity relationship database. The color channels include 10 channels, namely R, G, B channels and G / R, G / B, B / R, (GR) / (R+G), (GB) / (G+B), (RB) / (B+R), and G / (R+G+B) color channels. The intelligent model building module is used to construct an analysis model based on a feedforward neural network. It inputs the color feature-intensity relationship database into the model, performs algorithm optimization, and obtains a refined model. Specifically, the intelligent model building module employs a Bayesian optimization and a manual sampling method combining oversampling and undersampling to establish a feedforward neural network model for predicting the color feature-intensity relationship of infill bodies. Specifically, based on the color feature-intensity relationship database, the data is divided into training and testing sets. Color features in the RGB color space are selected as input parameters, and infill body intensity is selected as the output parameter to establish a feedforward neural network model to predict intensity. After determining the input and output layers, the number of hidden layers and the number of neurons in each hidden layer are set according to the complexity of the actual situation. The average value of the cross-entropy loss function after 5 cross-validations is used as the objective function, and a Bayesian optimization algorithm is used for network optimization. Simultaneously, a manual sampling method combining oversampling and undersampling is used to process the training set to eliminate data imbalance. A permutation feature importance algorithm is introduced to analyze the relative importance of input variables, and finally, the optimal feature combination is selected from the color features. Finally, the optimized feedforward neural network model is obtained. The measured data analysis module is used to input actual data into the optimized model for processing to obtain intensity prediction results.

2. The image-based intelligent prediction system for the strength of slag-based backfill bodies according to claim 1, characterized in that: The measured data analysis module uses the same shooting equipment and conditions as the strength and image information acquisition module to complete the acquisition of the original image of the infill body to be tested in the field. The original image is then subjected to the same steps as the data processing and association module to perform noise reduction and segmentation, and the corresponding color feature data is extracted. The color feature data is then input into the optimized feedforward neural network model for processing, thereby obtaining the prediction result of the infill strength.

3. An image-based intelligent prediction method for the strength of slag-based backfill, implemented based on the image-based intelligent prediction system for the strength of slag-based backfill as described in claim 1, characterized in that: The method includes the following steps: Step 1: Conduct similar experiments in the laboratory with varying group ratio parameters, and collect sample strength data and corresponding original images; Step 2: Obtain color feature parameters in the RGB color space based on the original image processing described in Step 1, and establish a color feature-intensity relationship database with the corresponding intensity data; the specific method is as follows: The obtained image is preprocessed using a color image filtering method. Based on the organic relationship between the R, G, and B components of the color image, the multidimensional color vector of the pixels is directly processed to eliminate noise in the image and avoid the influence of noise on image enhancement. Gamma transformation is used to improve the contrast of the image for accurate target identification in the subsequent process. In the RGB color space, based on the G channel image, a 2G-RB filtering algorithm that emphasizes the green component and suppresses the other two components is used to segment the infill body from the background image. After segmentation, the infill body image is processed by obtaining the pixel mean value as a color feature by mapping a color histogram for each color channel. The color channels include 10 channels: R, G, B channels, and 7 new channels G / R, G / B, B / R, (GR) / (R+G), (GB) / (G+B), (RB) / (B+R), and G / (R+G+B) derived from the R, G, and B channels. The average pixel value is obtained from these 10 channel images, and 10 color features are extracted. Step 3: Construct an analysis model based on a feedforward neural network. Input the color feature-intensity relationship database into the model, optimize the algorithm, and construct an image-based intensity prediction model. The specific method is as follows: The data in the color feature-intensity relationship database was randomly divided into training and test sets. Ten color features from the RGB color space were selected as input parameters for the neural network, and the infill strength was selected as the output parameter. A feedforward neural network model was established to predict the intensity. After determining the input and output layers, the number of hidden layers and the number of neurons in each hidden layer were set according to the complexity of the actual situation. The average value of the cross-entropy loss function after five cross-validations was used as the objective function, and the network was optimized using a Bayesian optimization algorithm. The training set was divided into five equal parts of the same size. Four parts were selected for training, and the remaining parts were used for validation. The process of dividing the training set and randomly selecting training and validation parts was repeated five times, and the average loss of the five validations was taken. At the same time, a manual sampling method combining oversampling and undersampling was used to process the training set to eliminate the influence of the imbalance of intensity parameters due to fewer light colors in the training set. The relative importance of the input variables was analyzed using a permutation feature importance algorithm, and the optimal feature combination was finally selected from the ten color features. Finally, the optimized feedforward neural network model was obtained and used to predict the intensity of infill bodies in actual mines. Step 4: Input the actual data into the optimized model for processing to obtain the intensity prediction results.

4. The image-based intelligent prediction method for the strength of slag-based backfill bodies according to claim 3, characterized in that: The specific method for step 1 is as follows: Similar experiments were conducted based on variations in the composition parameters of the actual backfilling mix design. These parameters included particle size distribution, binder content, mass concentration, component ratio, curing time, and temperature. Tailings and binder were obtained from the mining company. Four factors—mass concentration, binder content, component ratio, and curing time—were selected as independent variables. Response surface methodology was used to conduct surface laboratory tests. The strength performance indicators of the backfill were obtained using a high-precision mechanical testing machine. High-resolution RGB images were acquired using a CCD industrial camera and a portable camera.

5. The image-based intelligent prediction method for the strength of slag-based backfill bodies according to claim 4, characterized in that: The specific method for step 4 is as follows: The original images of the filling body to be tested were captured on-site at the mine. The process was as follows: a filling body sample with uniform color and area similar to that of the laboratory experiment was selected, placed in a portable camera box, and the original image of the filling body was obtained according to the shooting conditions in step 1; according to the image processing process in step 2, the data of the corresponding color features of the filling body were extracted and input into the optimized model in step 3 for processing to obtain the prediction result of the strength of the filling body.

Images