A rock mass stability intelligent evaluation method, system and device

Abstract

This invention relates to the field of mining technology, specifically to an intelligent evaluation method, system, and equipment for rock mass stability. The method includes: designing a low-light enhancement algorithm to enhance images; performing data augmentation and polygon annotation on the enhanced images to form a dataset; constructing and training a rock mass structural surface recognition model using a U-Net network improved with ResNet 50; inputting the dataset into the model to obtain structural surface feature parameters; measuring and correcting the point load strength of rock samples from the field; calculating the rock compressive strength based on the corrected point load strength; calculating and correcting the basic quality indicators of the rock mass based on the structural surface feature parameters and the rock compressive strength; and classifying and evaluating the stability of the rock mass based on the corrected indicators. This invention improves the automation and accuracy of rock mass stability evaluation, and features strong field applicability and high evaluation efficiency.

Description

Technical Field

[0001] This invention relates to the field of mining technology, specifically to an intelligent evaluation method, system, and equipment for rock mass stability. Background Technology

[0002] Rock mass stability is a critical issue in mining and engineering construction, influenced by various factors such as the development characteristics of rock mass structural planes and rock strength. Current rock mass stability assessment methods primarily rely on manual field surveys, laboratory strength analysis, or manual experience-based judgment, which suffers from low efficiency, large errors, significant susceptibility to human factors, and poor timeliness. Existing visual recognition methods only utilize machine learning techniques for qualitative analysis, lacking quantitative statistical analysis of rock mass structure and consideration of the impact of rock strength and geological environmental factors. With increasing mining depth and more complex rock mass conditions, traditional methods are no longer sufficient to meet the demands for efficient and accurate rock mass quality assessment.

[0003] Currently, methods for intelligent evaluation of rock mass stability mainly include mechanical unloading analysis and rapid visual recognition. Patent CN202310603831.7 uses a single data dimension, still relying on stress unloading parameters for evaluation, and lacks quantitative analysis of rock mass structural features (such as joint and fracture distribution). Patent CN202411061833.9 uses machine learning algorithms to extract rock mass fracture features, but it cannot distinguish between rock mass structural types such as fissures, interbedded rock, and faults, and therefore cannot further quantify the rock mass quality. Furthermore, this patent, which relies on on-site camera deployment, data network transmission, control terminal data processing, and result generation, has poor applicability in underground mining. In CN202411010597.8, the data fusion capability of this patent is insufficient. It only performs visual image analysis based on the YOLOv5 model and lacks quantitative fusion of rock mechanical parameters (such as compressive strength) and geological environmental factors (such as groundwater and geostress). It does not introduce rock mass quality grading standards (such as the BQ method), and its YOLOv5 model is difficult to dynamically correct the identification results to adapt to changes in complex geological conditions. Summary of the Invention

[0004] In recent years, with the development of image processing technology and deep learning algorithms, it has become possible to automatically identify rock mass structural features and evaluate rock mass stability using image processing and deep learning models. Developing miniaturized rock mass stability evaluation equipment facilitates on-site use in underground mines and provides timely data support for related blasting and support designs. Addressing the aforementioned technical problems of single data dimensions, insufficient identification accuracy, and an incomplete evaluation system, this invention provides an intelligent rock mass stability evaluation method, system, and equipment. This invention improves U-Net to achieve automatic segmentation of structural surfaces, cracks, and rock inclusions, and combines multi-source data such as point load tests to construct a comprehensive evaluation system, enabling rapid and intelligent evaluation of rock mass quality. A complete rock mass stability evaluation system has been established for easy on-site use; multi-source data collaborative evaluation is achieved through improvements to U-Net and the BQ grading system. The system includes a data acquisition and processing module, a rock mass structural surface identification module, a rock strength feature analysis module, and a rock mass quality grading and stability evaluation module. The device integrates an industrial camera, LED ring light, UWB positioning tag, display device, and rock strength testing equipment. It can complete the entire process of image acquisition, structural surface recognition, rock strength analysis, and stability evaluation in underground or low-light environments, achieving centimeter-level positioning of the acquisition location and associated data storage. The method, system, and device described in this invention are applicable to the following low-light scenarios: underground roadways, tunnels, underground chambers, and above-ground open-pit mines for surrounding rock structure acquisition and stability evaluation under nighttime or dark weather conditions.

[0005] The technical means employed in this invention are as follows:

[0006] Firstly, a smart evaluation method for rock mass stability includes the following steps: Images of surrounding rock in underground mine roadways are collected, a low-light enhancement algorithm is designed to enhance the images, the enhanced images are augmented, and polygon annotations are performed on the augmented images to form a dataset. The low-light enhancement algorithm is implemented based on an illumination extraction module and a feature fusion module. A rock mass structure surface recognition model based on an improved U-Net network was constructed and trained. The dataset was input into the trained rock mass structure surface recognition model to obtain the structure surface feature parameters, which include the number of fractures per unit area, dip angle, rock inclusions, and trace length and width. The U-Net network was improved using ResNet 50. The point load strength of the rock sample in the field is measured, the point load strength is corrected, and the compressive strength of the rock is calculated based on the corrected load strength. Based on the structural surface characteristic parameters and rock compressive strength, the basic quality index of the rock mass is calculated. The basic quality index of the rock mass is then corrected by combining groundwater, geostress and weak structural surface occurrence factors to obtain the corrected rock mass quality index. The stability of the rock mass is then classified and evaluated based on the corrected rock mass quality index.

[0007] Furthermore, the method also includes the step of interfacing with the mine's existing UWB positioning system: integrating a UWB positioning tag into the acquisition device, and obtaining the device's location information in the roadway in real time through the UWB positioning tag; and storing the coordinate information in association with the acquired images, structural surface recognition results, and rock mass stability evaluation results.

[0008] Furthermore, the lighting extraction module includes a first channel splicing layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and a first linear combination layer connected in sequence. The feature fusion module includes, in sequence, a self-feature enhancement layer, a first Transformer module, a first downsampling module, a second Transformer module, a second downsampling module, a third Transformer module, a third downsampling module, a fourth Transformer module, a first Prompt module, a first upsampling module, a second channel concatenation layer, a fourth convolutional layer, a fifth Transformer module, a second Prompt module, a second upsampling module, a third channel concatenation layer, a fifth convolutional layer, a sixth Transformer module, a third Prompt module, a third upsampling module, a fourth channel concatenation layer, a seventh Transformer module, an output layer, and an element-wise addition layer. The output of the first Transformer module is connected to the input of the fourth channel concatenation layer, the output of the second Transformer module is connected to the input of the third channel concatenation layer, and the output of the third Transformer module is connected to the input of the second channel concatenation layer.

[0009] Furthermore, the computational process of the low-light enhancement algorithm includes: The input image is extracted from the dataset. The two input images are passed through the first channel stitching layer to obtain the first stitching feature map. The first stitching feature map is passed through the first convolutional layer and the second convolutional layer in sequence to obtain the intermediate lighting feature. The intermediate lighting feature is passed through the third convolutional layer to obtain the convolutional feature map. The convolutional feature map and the input image are input into the first linear combination layer to obtain the lighting mapping image. The illumination mapping image is passed sequentially through a self-feature enhancement layer and a first Transformer module to obtain a first Transformer feature. The first Transformer feature is then passed sequentially through a first downsampling module and a second Transformer module to obtain a second Transformer feature. The second Transformer feature is then passed sequentially through a second downsampling module and a third Transformer module to obtain a third Transformer feature. The third Transformer feature is sequentially passed through a third downsampling module, a fourth Transformer module, a first Prompt module, and a first upsampling module to obtain a first upsampling output. The first upsampling output and the third Transformer feature are then passed through a second channel concatenation layer to obtain a second concatenated feature map. The second concatenated feature map is sequentially passed through a fourth convolutional layer, a fifth Transformer module, a second Prompt module, and a second upsampling module to obtain a second upsampling output. The second upsampling output and the second Transformer feature are then passed through a third channel concatenation layer to obtain a third concatenated feature map. The third concatenated feature map is sequentially passed through a fifth convolutional layer, a sixth Transformer module, a third Prompt module, and a third upsampling module to obtain a third upsampling output. The third upsampling output and the first Transformer feature are then passed through a fourth channel concatenation layer to obtain a fourth concatenated feature map. The fourth concatenated feature map is then passed through a seventh Transformer module to obtain a seventh feature map. After the seventh feature map is output, it is passed to the input image through an element-wise addition layer to output the enhanced image.

[0010] Furthermore, the low-light enhancement algorithm achieves the supplementary lighting effect through a dual mechanism of illuminance reconstruction and feature compensation, and works in conjunction with a hardware supplementary lighting device, specifically including: The hardware lighting device uses an LED ring light, which is placed around the camera to provide active lighting in low-light environments, eliminate shadows, and enhance details in dark areas. The low-light enhancement algorithm, as a software correction module, performs secondary enhancement on the image acquired after hardware illumination. The illumination extraction module processes the input low-light image. I in The channel mean is calculated to obtain the initial illuminance estimation map. The initial illumination estimation map is concatenated with the original image through channels and then fed into a convolutional layer to generate an illumination mapping image. I illu The expression for the illumination mapping image is:

[0011] in, This indicates a channel splicing operation. This is a combined mapping of depthwise separable convolution and 1x1 convolution; The feature fusion module inputs the illumination map image into the Transformer encoder, extracts illumination features at different scales through the downsampling module, and introduces a Prompt module at each stage to activate and enhance dark area features. The Prompt module utilizes intermediate illumination features obtained from the illumination extraction module. F illu The enhanced features are obtained by weighted reconstruction of the feature map through an attention mechanism. The expression of the enhanced features is as follows:

[0012] in, F in Input features for the current layer, F enh For enhanced features; Finally, the enhanced feature map is fused with the original input image through an element-wise addition layer to generate the output image after illumination. I out The expression for the output image is:

[0013] in, F out The enhanced feature map output by the feature fusion module. These are learnable fusion weight coefficients.

[0014] Furthermore, the network architecture of the first Prompt module, the second Prompt module, and the third Prompt module is the same, and the workflow of the first Prompt module includes: The input feature map is fed into the prediction layer, which includes a pooling layer, a linear layer, and a softmax layer, to obtain the predicted feature map. The predicted feature map is linearly combined with the parameters to obtain the combined feature map. The combined feature map and the input feature map are then subjected to interpolation to obtain the interpolated feature map. The feature map after the difference processing is passed through the sixth convolutional layer, the fourth Prompt layer and the first normalization layer in sequence to obtain the first normalized feature map. The intermediate illumination feature is passed through the difference layer and the second normalization layer in sequence to obtain the second normalized feature map. The first normalized feature map and the second normalized feature map are passed through the first attention layer to obtain the output of the attention layer. The output of the attention layer is sequentially passed through the second linear combination layer, the third normalization layer, the feedforward neural network layer, the third linear combination layer, the fourth normalization layer, the separable two-dimensional convolutional layer, and the second attention layer to obtain the output of the first Prompt module.

[0015] Furthermore, the improvement of the rock mass structure surface recognition model lies in replacing the ordinary convolutional blocks of the U-Net network with ResNet 50 residual blocks; The workflow of the rock mass structure surface identification model includes: The input image is processed by ResNet 50 residual and max pooling four times to obtain the first feature map; The first feature map is processed by ResNet 50 residual and upconvolution four times to obtain the second feature map; The second feature map is subjected to ResNet 50 residual processing three times to obtain the structural surface feature parameters.

[0016] Furthermore, the number of cracks per unit area is calculated by converting pixels of the markers to the actual area. The formula for calculating the actual area is:

[0017] in, This is the actual area. The number of pixels in the image. The number of pixels of the marker. The actual area of ​​the marker. The number of cracks per unit area is calculated based on the actual area. The formula for calculating the number of cracks per unit area is as follows:

[0018] in, The number of cracks per unit area. The number of cracks detected. By identifying the angle between the weak structural surface and the hole axis, the weak structural surface is fitted into a straight line using a fitting method. The projection lengths of the straight line on the y-axis and x-axis are calculated, and then the inclination angle is calculated. The formula for calculating the inclination angle is as follows:

[0019] in, It is the angle of inclination. Let be the length of the line projected onto the x-axis. Let be the length of the line projected onto the y-axis. The trace length is calculated using the least squares fitting method, and the formula for calculating the trace length is as follows:

[0020] in, For the length of the trace, ( Let be the coordinates of the first endpoint of the line. Let x be the x-coordinate of the first endpoint. Let y be the y-coordinate of the first endpoint. Let be the coordinates of the second endpoint of the line. Let x be the x-coordinate of the second endpoint. Let y be the y-coordinate of the second endpoint. The actual length of the marker. The formula for calculating the trace width is:

[0021] in, 'For the width of the gap, The area of ​​the crack pixel. The calculation formula for the basic quality indicators of the rock mass is as follows:

[0022] in, These are the basic quality indicators for rock masses. For rock compressive strength, The rock mass integrity coefficient. The formula for calculating the corrected rock mass quality index is as follows: [ BQ ]= BQ 100 ( K 1+ K 2+ K 3) Wherein, [BQ] is the corrected rock mass quality index, K1 is the correction coefficient for the influence of groundwater, K2 is the correction coefficient for the influence of the attitude of weak structural planes, and K3 is the correction coefficient for the influence of in-situ stress. The calculation process for the compressive strength of the rock includes: The rock specimen was placed between two spherical conical pressure plates and continuously loaded before the specimen failed. The morphological characteristics, failure mode and bearing strength of the specimen were recorded. The dimensional parameters on the failure surface of the rock specimen were measured, including the spacing between loading points and the width of the fracture surface. Calculate the equivalent core diameter based on the stated size parameters:

[0023] in, The equivalent core diameter is... The width of the fracture surface. The spacing between loading points; Calculate the uncorrected point load strength based on the bearing capacity and the equivalent core diameter:

[0024] in, For uncorrected point load strength, For load-bearing strength; Introducing a correction factor, and based on the correction factor and the uncorrected point load strength, calculating the corrected point load strength:

[0025] in, To correct the point load strength, The correction factor is calculated using the following formula:

[0026] in, m To correct the index; Calculate the rock compressive strength based on the corrected point load strength:

[0027] in, It represents the compressive strength of the rock.

[0028] Secondly, a rock mass stability intelligent evaluation system is provided to implement the above-mentioned on-site rock mass stability intelligent evaluation method, comprising: The data acquisition and processing module is used to acquire images of the surrounding rock in underground mine roadways, enhance the images using a low-light enhancement algorithm, and perform data augmentation and polygon annotation on the enhanced images to form a rock mass structure surface recognition dataset. The rock mass structure surface recognition module is used to construct and train a rock mass structure surface recognition model using a ResNet 50 improved U-Net network. The dataset is input into the trained rock mass structure surface recognition model, and the structure surface feature parameters are automatically extracted. The structure surface feature parameters include the number of fractures per unit area, the orientation of the structure surface, the interbedded rock, and the length and width of the trace. The rock strength characteristic analysis module is used to correct the point load strength and calculate the rock compressive strength based on the corrected point load strength. The rock mass quality classification and stability evaluation module is used to calculate the basic quality index of the rock mass based on the structural surface characteristic parameters and rock compressive strength, and to modify the basic quality index of the rock mass by combining groundwater, geostress and weak structural surface occurrence factors to obtain modified rock mass quality index, and to classify and evaluate the stability of the rock mass based on the modified rock mass quality index. The intelligent evaluation device integrates a data acquisition and processing module, a rock mass structure surface identification module, a rock strength characteristic analysis module, and an embedded algorithm to achieve rapid intelligent on-site evaluation. The intelligent evaluation device includes an explosion-proof housing, an integrated control board, a touch screen display, a power module, a tripod, an industrial camera, an LED ring light, and a UWB positioning tag. The touch screen display is used to display the acquired images, recognition results, and stability evaluation results in real time.

[0029] Thirdly, a smart rock mass stability evaluation device includes: An industrial camera is used to collect raw image data of the surrounding rock in underground mine tunnels. The industrial camera acquires images through a single-frame capture mode, and generates a static image each time it is triggered to be sent to the recognition system. LED ring lights are placed around the industrial camera to provide active hardware lighting in low-light environments. UWB positioning tags are used to communicate with the existing UWB positioning system in the mine, providing real-time centimeter-level positioning and obtaining the location information of the equipment in the roadway. The display device is used to display the acquired images, rock mass structure surface identification results, rock compressive strength and rock mass stability evaluation results in real time. Rock strength testing equipment is used to measure the point load strength of rock samples; A processor is used to execute the computer program in the memory to implement the above-mentioned intelligent evaluation system for on-site rock mass stability.

[0030] Compared with the prior art, the present invention has the following advantages: 1. This invention achieves automatic identification and segmentation of rock mass structural surfaces by improving the U-Net deep learning model. Combined with parameters such as rock compressive strength obtained from field point load tests, the structural surface features and rock strength indices are quantitatively input into the BQ grading model, forming a full-process quantitative evaluation system. Compared to traditional methods of manually statistically analyzing structural surfaces and judging stability based on experience, the evaluation results are more objective and accurate.

[0031] 2. This invention utilizes an improved U-Net model within the PyTorch framework to achieve automatic identification of structural surfaces such as joints, fissures, and rock inclusions. This replaces the traditional manual annotation and detection methods, significantly reducing manual workload while avoiding errors in subjective human judgment, and substantially improving the efficiency and accuracy of rock mass structural surface detection.

[0032] 3. The technical system of this invention is both portable and easy to use. From high-definition industrial camera acquisition and PromptHDR enhancement processing, to portable point load testing equipment for strength measurement, and then to on-site input parameters to complete BQ index calculation and stability evaluation, the entire process can be completed quickly on the mine site, providing timely and reliable data support for subsequent blasting parameter design and support scheme formulation.

[0033] 4. This invention is applicable to various mining site environments and can adapt to the rock mass stability evaluation needs of different geological conditions and different types of mines. It is also applicable to the rock mass stability evaluation of tunnels, slopes, and other engineering projects.

[0034] Based on the above reasons, this invention can be widely promoted in fields such as mining. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of an intelligent evaluation method for rock mass stability according to the present invention.

[0037] Figure 2 This is a network architecture diagram of the low-light enhancement algorithm of the present invention.

[0038] Figure 3 This is a schematic diagram of the improved U-Net network structure of the present invention.

[0039] Figure 4 This is a schematic diagram of a rock strength field testing device according to an embodiment of the present invention.

[0040] Figure 5 This is a schematic diagram of the hardware composition of the intelligent rock mass quality evaluation device in an embodiment of the present invention.

[0041] Figure 6 This is a schematic diagram of the window of the intelligent evaluation system for rock mass stability in an embodiment of the present invention.

[0042] Figure 7(a) shows the original low-light image without hardware illumination or software correction.

[0043] Figure 7(b) shows the image after hardware-only supplemental lighting (LED ring light).

[0044] Figure 7(c) shows the final image after correction by hardware lighting and software low-light enhancement algorithm.

[0045] Figure 8 This is a schematic diagram of the overall structure and field scenario of the intelligent rock mass stability evaluation system and equipment in an embodiment of the present invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0047] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0048] like Figure 1 As shown, this invention provides an intelligent evaluation method for rock mass stability, the steps of which are as follows: A. Collect images of the surrounding rock in underground mine roadways, design a low-light enhancement algorithm to enhance the images, perform data augmentation on the enhanced images, and perform polygon annotation on the augmented images to form a dataset. The low-light enhancement algorithm is implemented based on the illumination extraction module and the feature fusion module.

[0049] Specifically, step A includes: A1. High-definition industrial cameras are used to collect high-definition image data of the surrounding rock in underground mine tunnels. Representative and universally applicable two-dimensional graphic data of the rock walls, i.e., images of the surrounding rock in underground mine tunnels, need to be collected. The industrial camera acquires images through a single-frame capture mode, generating a static image each time a shot is triggered and sending it to the recognition system.

[0050] A11. Determine the acquisition location and target: Select exposed tunnels in underground mines as image acquisition locations.

[0051] A12. Prepare a high-definition industrial camera: Select a high-resolution industrial camera to ensure that it can capture the details of the rock mass structure.

[0052] A13. Image Acquisition: Collect images of rock mass structural features extensively within underground mine tunnels. Ensure that the acquired images include a variety of scenes and lighting conditions. Pay attention to capturing structural features of different shapes, sizes, orientations, and occlusions during shooting.

[0053] A14. Adjust Image Size: Based on the model's processing capabilities and computing resources, appropriately scale or crop the acquired high-resolution images. Ensure the adjusted image size fits the model's input requirements while preserving sufficient detail.

[0054] A15. Recording and Organizing Images: Number and record the acquired images to ensure that each image has a clear source and location information. Organize and store the acquired images for subsequent analysis and processing.

[0055] A2. The PromptHDR algorithm, a low-light enhancement algorithm based on cue learning, is used to enhance the recovery of image details and improve the quality of images acquired in low-light environments. This algorithm works in conjunction with a hardware LED ring light to form a hardware-software combined lighting solution.

[0056] like Figure 2 As shown, the low-light enhancement algorithm is implemented based on an illumination extraction module and a feature fusion module. By designing an illumination extraction module to explicitly model the illumination distribution, color distortion caused by non-uniform illumination is effectively suppressed. A cue module that integrates the Mamba mechanism and illumination features is introduced to enhance the model's contextual understanding of the alleyway scene and its ability to preserve details of the alleyway's structural surfaces. Furthermore, a sampling module containing DySample and HWD components is constructed to achieve high-quality image reconstruction under multi-scale feature fusion. Images preprocessed by the low-light enhancement algorithm are more applicable and have higher segmentation accuracy in subsequent image segmentation tasks.

[0057] The first part is the Light Extraction Module (LEM) (Ⅰ). It calculates the illumination prior of the input image and combines it with the original RGB image. Using 1x1 convolutional layers and 5x5 depthwise separable convolutional layers, it extracts spatial features to generate an illumination map (illu_map). The intermediate illumination features (illu_fea) are then used to optimize the features of the Prompt block. An RGB image is received as input. The first step calculates the illumination prior by averaging the channels of the input image. Next, the original RGB image and the illumination prior are concatenated. The concatenated input is then processed by a first 1x1 convolutional layer to reconstruct the inter-channel features and generate a feature map. This feature map is then passed through a 5x5 depthwise separable convolutional layer, which allows the network to efficiently extract spatial features and capture local illumination patterns and variations. Finally, the feature map is passed through a second 1x1 convolutional layer to map the extracted features back to the original image space, generating the final illumination map. The final result includes the output illumination map (illu_map), used for feature extraction in subsequent networks; and the generated illumination features (illu_fea), used by the prompt module to optimize the features. The illumination features (illu_fea) include: brightness information (especially in dark areas), signal-to-noise ratio information, local contrast information, color accuracy information, edge sharpness information, and detail recovery rate. The implementation of this module can effectively suppress color distortion caused by non-uniform downhole illumination, improve the uniformity of the overall brightness distribution, and provide structured prior lighting information for subsequent stages.

[0058] Illumination Extraction Module. Traditional low-light image processing methods often fail to effectively address color distortion and insufficient illumination. Therefore, an illumination extraction module is designed in the first stage to specifically process the illumination feature information of low-light images. This module extracts the illumination information from the low-light image and maps it back to the original image space, generating an illumination-mapped image which is then sent to the next stage. The illumination features generated in this process will also be fused in the next stage. This stage effectively alleviates color deviation and brightness loss caused by insufficient illumination, preparing the image for the next stage.

[0059] The second part is a feature fusion module (FFM) based on a "U"-shaped network structure (Ⅱ). It extracts global and local features of the image through a Transformer, and combines downsampling, prompting, upsampling and feature concatenation. Finally, it concatenates the results of deep feature extraction with the initial input image to output the enhanced image.

[0060] Feature Fusion Module. Traditional Transformer architectures suffer from redundancy in long-distance information transmission, low processing efficiency, and loss of local details when processing low-light images. A U-shaped Transformer architecture is designed to fully utilize contextual information and improve efficiency in long-distance information transmission. A cueing mechanism is introduced to enhance the model's capabilities during feature extraction, ensuring the algorithm extracts more lighting details to improve overall image quality. A novel sampling module is used, which extracts multi-scale feature information through downsampling, preserving the global structure of the image and reducing noise, while simultaneously using upsampling to recover local details. Finally, the illumination map image generated in the illumination extraction stage is fused with the error component to output a high-quality enhanced image.

[0061] The Transformer Block structure is shown on the right side of the figure; the Prompt consists of a Global Context Hint (GCP) submodule and a Hierarchical Feature Refinement (HFR) submodule, used to optimize the extracted features, as shown in the specific structure below. Figure 2 Below. The global contextual cues submodule is responsible for extracting global structural information and generating guiding cues to maintain the overall rationality of the image's illumination distribution and color consistency. The hierarchical feature refinement module fuses the illumination features from the illumination extraction module (LEM) with the input features, refining local features through an attention mechanism to enhance edge sharpness, restore texture details, and suppress noise propagation. This design enables the model to dynamically understand the global semantic information of rock mass images in low-light conditions while preserving subtle structural features such as tunnels.

[0062] This invention does not modify the Transformer; it is essentially a combination of different modules that produces the desired low-light enhancement effect.

[0063] Finally, the illumination map output by the illumination extraction module is fused with the error components processed by the feature fusion module to generate an enhanced image, achieving efficient fusion of multi-scale features and high-quality reconstruction.

[0064] The calculation and implementation mechanism of supplementary lighting in the low-light enhancement algorithm is as follows: The supplementary lighting effect is achieved based on a dual mechanism of illumination reconstruction and feature compensation. Firstly, in the illumination extraction module, the input low-light image is processed... I in The initial illuminance estimation map is obtained by calculating the channel mean. The initial illumination estimation map is concatenated with the original image through channels and then fed into a convolutional layer to generate an illumination mapping image. I illu This process can be expressed by the formula:

[0065] in, This indicates a channel splicing operation. This is a combined mapping of depthwise separable convolution and 1x1 convolution; Secondly, in the feature fusion module, supplementary lighting is achieved through multi-scale feature enhancement. This involves mapping the illumination image... I illu The input is a Transformer encoder, which extracts illumination features at different scales through a downsampling module. At each stage, a Prompt module is introduced to activate and enhance dark area features. The Prompt module utilizes intermediate illumination features obtained from the illumination extraction module. F illu The feature map is reconstructed using a weighted attention mechanism, and its core calculation formula is:

[0066] in, F in Input features for the current layer, F enh This is an enhanced feature. The mechanism enables targeted brightness enhancement and detail restoration in low-light areas.

[0067] Finally, the enhanced feature map is fused with the original input image through an element-wise addition layer to generate the output image after illumination. I out :

[0068] in, F out The enhanced feature map output by the feature fusion module. These are learnable fusion weight coefficients used to balance the original and enhanced information, avoiding overexposure or distortion.

[0069] When this algorithm works in conjunction with a hardware LED ring light, the hardware lighting first eliminates shadows in major dark areas and improves overall brightness, while the software correction further restores edge details, suppresses noise, and balances brightness distribution. This combined hardware and software lighting solution optimizes for the uneven lighting, weakly textured surfaces, and equipment noise in underground tunnel environments, demonstrating significant improvements in brightness distribution, color fidelity, local contrast, signal-to-noise ratio, and detail retention. This helps improve the accuracy and reliability of subsequent intelligent detection algorithms in low-light environments.

[0070] The illumination extraction module includes a first channel splicing layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and a first linear combination layer connected in sequence.

[0071] The feature fusion module includes, in sequence, an SFE layer, a first Transformer module, a first downsampling module, a second Transformer module, a second downsampling module, a third Transformer module, a third downsampling module, a fourth Transformer module, a first Prompt module, a first upsampling module, a second channel stitching layer, a fourth convolutional layer, a fifth Transformer module, a second Prompt module, a second upsampling module, a third channel stitching layer, a fifth convolutional layer, a sixth Transformer module, a third Prompt module, a third upsampling module, a fourth channel stitching layer, a seventh Transformer module, an output layer, and an element-wise addition layer. The output of the first Transformer module is connected to the input of the fourth channel stitching layer, the output of the second Transformer module is connected to the input of the third channel stitching layer, and the output of the third Transformer module is connected to the input of the second channel stitching layer.

[0072] The computational process of the low-light enhancement algorithm includes: The first step is to extract the input images from the dataset, pass the two input images through the first channel stitching layer to obtain the first stitched feature map, pass the first stitched feature map through the first convolutional layer and the second convolutional layer in sequence to obtain the intermediate illumination features, pass the intermediate illumination features through the third convolutional layer to obtain the convolutional feature map, and input the convolutional feature map and the input image into the first linear combination layer to obtain the illumination mapping image.

[0073] The first step specifically processes the illumination feature information of low-light images. Illumination information is extracted from the low-light image and mapped back to the original image space to generate an illumination-mapped image, which is then sent to the next stage. The illumination features generated in this process will also be fused in the next stage. This process effectively alleviates the problems of color deviation and brightness loss caused by insufficient lighting, preparing the image for the next stage.

[0074] The second step involves passing the illumination-mapped image sequentially through a Self Feature Enhancement (SFE) layer and a first Transformer module to obtain the first Transformer feature. The first Transformer feature is then passed sequentially through a first downsampling module and a second Transformer module to obtain the second Transformer feature. Finally, the second Transformer feature is passed sequentially through a second downsampling module and a third Transformer module to obtain the third Transformer feature.

[0075] The SFE layer essentially extracts the lighting elements from the original image, thus enhancing it.

[0076] The third step involves passing the third Transformer feature sequentially through the third downsampling module, the fourth Transformer module, the first Prompt module, and the first upsampling module to obtain the first upsampling output. The first upsampling output and the third Transformer feature are then passed through a second-channel concatenation layer to obtain a second concatenated feature map. This second concatenated feature map is then passed sequentially through a fourth convolutional layer, a fifth Transformer module, a second Prompt module, and a second upsampling module to obtain a second upsampling output. The second upsampling output and the second Transformer feature are then passed through a third-channel concatenation layer to obtain a third concatenated feature map. This third concatenated feature map is then passed sequentially through a fifth convolutional layer, a sixth Transformer module, a third Prompt module, and a third upsampling module to obtain a third upsampling output. The third upsampling output and the first Transformer feature are then passed through a fourth-channel concatenation layer to obtain a fourth concatenated feature map. This fourth concatenated feature map is then passed through a seventh Transformer module to obtain a seventh feature map. After outputting the seventh feature map, it is combined with the input image through an element-wise addition layer to output the enhanced image.

[0077] Steps two and three address the problems of redundant information transmission over long distances, low information processing efficiency, and loss of local details inherent in traditional Transformer architectures when processing low-light images. A "U-shaped" Transformer architecture is designed to fully utilize contextual information and improve efficiency during long-distance information transmission. A cueing mechanism is introduced to enhance the model's capabilities in feature extraction, ensuring the algorithm can extract more lighting details from the image and improve overall image quality. A sampling module is designed to extract multi-scale feature information through downsampling, preserving the global structure of the image and reducing noise, while upsampling restores local details, ensuring accurate restoration of image details.

[0078] In the low-light enhancement algorithm, the upsampling module is Dysample and the downsampling module is HWD. The sampling structure is used to improve the quality of the final generated image.

[0079] Specifically, upsampling achieves precise detail recovery and enhances detail representation in blurred or dark areas by dynamically adjusting the sampling position. The upsampling module is the DySample module, whose core idea is to achieve flexible spatial transformation by learning the offset of input features. The core operation is to redesign the upsampling process from the perspective of point sampling, and to achieve flexible feature upsampling by learning the dynamic sampling point offset of the input features. First, the upsampling module assumes that continuous features will be obtained. The input feature map (size C×H×W) is expanded into a continuous feature map by bilinear interpolation, then dynamic sampling points are generated. By learning content-aware offsets, the sampling point positions are adjusted, and upsampling features are extracted from the continuous feature map. Finally, the sampling is completed using the grid_sample function.

[0080] The downsampling module extracts multi-scale information through Haar wavelet transform, effectively preserving global structure and detailed features, thus compensating for the shortcomings of the Transformer architecture in handling local details. The downsampling module is an HWD module, whose core idea is to reduce the spatial resolution of the feature map while retaining all information, and encoding some spatial information into the channel dimension. Subsequent convolutional layers are then used to extract discriminative features and filter redundant information. Downsampling after the Transformer extracts information allows for further extraction and preservation of multi-scale local features while retaining the global semantic information captured by the Transformer module. The downsampling module uses a Haar wavelet transform-based downsampling method, maximizing the preservation of important information while reducing the spatial resolution of the feature map.

[0081] Through the above targeted optimizations, the illumination mapping image and error components are finally fused to output a high-quality enhanced image.

[0082] In a preferred embodiment of the present invention, the network architecture of the first Prompt module, the second Prompt module, and the third Prompt module is the same, and the workflow of the first Prompt module includes: The first step is to input the input feature map into the prediction layer, which includes a pooling layer, a linear layer, and a softmax layer to obtain the predicted feature map.

[0083] The second step is to linearly combine the predicted feature map with the parameter (P_parm) to obtain the combined feature map. Then, the combined feature map and the input feature map are subjected to interpolation to obtain the interpolated feature map.

[0084] The parameter (P_parm) is used for feature transformation.

[0085] The third step involves passing the differenced feature map through the sixth convolutional layer, the fourth Prompt layer (Prompt1), and the first normalization layer to obtain the first normalized feature map. The intermediate illumination feature map is then passed through the difference layer and the second normalization layer to obtain the second normalized feature map. Finally, the first and second normalized feature maps are passed through the first attention layer to obtain the output of the attention layer.

[0086] The fourth step involves passing the output of the attention layer sequentially through the second linear combination layer, the third normalization layer, the feedforward neural network layer, the third linear combination layer, the fourth normalization layer, the separable two-dimensional convolutional layer, and the second attention layer to obtain the output of the first Prompt module.

[0087] The Prompt module consists of two sub-modules: Global Context Prompting (GCP) and Hierarchical Feature Refinement (HFR). The first part captures global information and generates prompts, enhancing the model's extraction of overall image features through guided cues. The second part fuses the cues with illumination features, striking a balance between local details and global information. The resulting features are used in subsequent feature processing. Its core principle is to utilize structured prior information obtained from the illumination extraction module to dynamically guide the enhancement process through "cues," enabling the model to adaptively focus on the structural context of the entire tunnel and subtle features of individual structural surfaces (such as faults and cracks), thus achieving more perceptually intelligent image enhancement.

[0088] A3. Perform data augmentation on the enhanced image. Use the "Create Polygons" tool to annotate the edges of structural surfaces with polygons, marking structural features (such as joints, fissures, and inclusions). Classify and name each annotated target to ensure accurate category information. Save the annotation results to form initial training data.

[0089] As a preferred embodiment of the present invention, data augmentation includes image rotation, scaling, cropping, flipping, color transformation, etc., to increase the diversity and quantity of training samples.

[0090] B. In the PyTorch framework, construct and train a rock mass structure surface recognition model based on an improved U-Net network. Input the dataset into the trained rock mass structure surface recognition model to obtain the structure surface feature parameters, including the number of fractures per unit area, dip angle, rock inclusions, and trace length and width. Improve the U-Net network using ResNet 50.

[0091] Using the PyTorch deep learning framework, a deep neural network for rock mass structure feature recognition is constructed using the Python language. Deep features of rock mass images are extracted through operations such as convolution and pooling.

[0092] B1, such as Figure 3 As shown, a deep neural network for rock mass structure feature recognition is constructed using Python within the PyTorch deep learning framework. A rock mass structure surface recognition model based on an improved U-Net network is built, extracting deep-level features from rock mass images through operations such as convolution and pooling. The ResNet50 network is combined with the U-Net network, and the vanishing gradient problem in deep neural networks is addressed by introducing "residual blocks."

[0093] The core advantage of incorporating ResNet50 into U-Net as the backbone network lies in its deep integration of ResNet50's powerful deep feature extraction capabilities (effectively capturing multi-scale semantic information through residual blocks) and U-Net's refined spatial recovery mechanism (fusing features from different levels using skip connections). This not only significantly improves segmentation accuracy and balances global semantics with local details, but also leverages ResNet50's pre-trained weights on ImageNet to achieve efficient data convergence and model robustness, thus achieving more accurate and stable segmentation results in complex scenarios such as image segmentation. Utilizing ResNet50's feature extraction capabilities in conjunction with U-Net to achieve pixel-level rock mass structure image segmentation allows for quantitative statistics of information such as structural surface dip angle, trace width, and length, as well as quantitative evaluation of rock mass stability.

[0094] U-Net's backbone feature extraction network is VGG; the improvement is to replace the backbone feature extraction network with the more effective ResNet 50. In short... Figure 3 ResNet is a feature extraction method, and the right-hand side shows the array results.

[0095] Specifically, the improvement of the U-Net network in this invention lies in replacing the ordinary convolutional blocks of the U-Net network with ResNet 50 residual blocks. The workflow of the rock mass structure surface recognition model includes: The input image is processed by ResNet 50 residual and max pooling four times to obtain the first feature map; the first feature map is processed by ResNet 50 residual and up convolution four times to obtain the second feature map; the second feature map is processed by ResNet 50 residual three times to obtain the structural surface feature parameters.

[0096] U-Net uses ResNet50 as its backbone feature extraction network, replacing the regular convolutional blocks in U-Net with ResNet50 residual blocks. In the original U-Net, each layer of the encoder (left half of the U) and decoder (right half of the U) consists of two consecutive 3x3 convolutions (followed by ReLU activations); now, each layer consists of residual blocks. The difference is that the encoder can be considered a complete ResNet50, while the decoder only uses ResNet residual blocks. Therefore, the ResNet50 on the left is the backbone feature extraction network, and the improved U-Net on the right is the enhanced feature extraction network.

[0097] B2. Train a deep neural network, optimize network parameters, and select the optimal model by comparing the detection accuracy and recognition speed of structural features such as rock joints, fissures, and rock inclusions to ensure that the model can accurately identify rock structure features.

[0098] B3. Use the trained rock mass structure recognition and segmentation model to identify the number of fractures per unit area, the orientation of structural planes, interbedded rock, and the length and width of traces in the rock mass images at the mine site.

[0099] B31. Use a trained deep neural network to identify structural features such as joints, fissures, and rock inclusions in rock mass images collected from the mine site. Develop a computer vision-based algorithm for determining object distances and an automatic statistical calculation program for rock mass parameters to statistically analyze structural feature parameters such as the number of fissure groups and rock mass integrity.

[0100] B32. Feature Statistics: The number of cracks per unit area is calculated by converting pixels of the marker to the actual area. The formula for calculating the actual area is:

[0101] in, This is the actual area. The number of pixels in the image. The number of pixels of the marker. In this embodiment, S is taken as 0.0625m², representing the actual area of ​​the marker. 2 .

[0102] The formula for converting actual length is:

[0103] in, To capture the actual length, The length of the image in pixels. The length of the marker in pixels. The actual length of the marker is 0.25m in this embodiment.

[0104] The number of cracks per unit area is calculated based on the actual area. The formula for calculating the number of cracks per unit area is:

[0105] in, The number of cracks per unit area. This represents the number of cracks detected.

[0106] By identifying the angle between the weak structural plane and the hole axis, the weak structural plane is fitted into a straight line using a fitting method. The projection lengths of the straight line on the y-axis and x-axis are calculated, and then the dip angle is calculated. The formula for calculating the dip angle is:

[0107] in, It is the angle of inclination. Let be the length of the line projected onto the x-axis. The length of the line projected onto the y-axis.

[0108] Identify the fracture structure planes and fit them to a straight line using the least squares fitting method. The length of this straight line is the trace length, which is calculated using the following formula:

[0109] in, For the length of the trace, ( Let be the coordinates of the first endpoint of the line. Let x be the x-coordinate of the first endpoint. Let y be the y-coordinate of the first endpoint. Let be the coordinates of the second endpoint of the line. Let x be the x-coordinate of the second endpoint. Let y be the y-coordinate of the second endpoint. This refers to the actual length of the marker.

[0110] Based on the calculated trace length, the crack is considered rectangular. The formula for calculating the trace width is:

[0111] in, For the gap width, The area of ​​the crack pixel.

[0112] C. Measure the point load strength of the rock sample in the field, correct the point load strength, and calculate the rock compressive strength based on the corrected load strength.

[0113] C1. Collect typical rock samples on-site and measure point loads using on-site rock strength testing equipment. On-site rock strength testing equipment includes, for example: Figure 4 As shown.

[0114] Specifically, the rock specimen is placed between two spherical conical pressure plates and continuously loaded before the specimen fails. The morphological characteristics, failure mode, and bearing strength of the specimen are recorded. The dimensional parameters on the failure surface of the rock specimen are measured, including the distance between loading points and the width of the fracture surface.

[0115] C2. Calculate the tensile strength and compressive strength of the rock based on the test data.

[0116] C21. Calculate the equivalent core diameter based on the dimensional parameters:

[0117] in, The equivalent core diameter is in mm. The width of the fracture surface. This refers to the spacing between the loading points.

[0118] C22. Calculate the uncorrected point load strength based on the bearing capacity and equivalent core diameter:

[0119] in, The uncorrected point load strength, in MPa. The bearing capacity is expressed in MPa.

[0120] C23. Introduce a correction factor, and calculate the corrected point load strength based on the correction factor and the uncorrected point load strength:

[0121] in, To correct for the point load strength, in MPa. The correction factor is calculated using the following formula:

[0122] in, m To correct the index, a value of 0.40 to 0.45 can be used.

[0123] C24. The tensile strength Rt and compressive strength Rc of rock can be obtained from the results of on-site rock point load strength tests. According to the empirical formulas provided by the International Society for Rock Mechanics (ISRM), the tensile strength and compressive strength of rock can be obtained from the point load strength, and the relationship is as follows:

[0124]

[0125] in, For the tensile strength of rock, It represents the compressive strength of the rock.

[0126] D. Based on the structural surface characteristic parameters and rock compressive strength, calculate the basic quality index of the rock mass. Combine the factors of groundwater, geostress and weak structural surface occurrence to correct the basic quality index of the rock mass, and obtain the corrected rock mass quality index. Based on the corrected rock mass quality index, classify and evaluate the stability of the rock mass.

[0127] D1. Combining the structural surface features and rock strength data identified by the improved U-Net model, the basic quality index (BQ) of the rock mass is calculated using the BQ rock mass quality grading evaluation method.

[0128] D11. Calculate the basic quality index BQ of the rock mass:

[0129] in, These are the basic quality indicators for rock masses. For rock compressive strength, This is the rock mass integrity coefficient.

[0130] The rock mass integrity coefficient was calculated with reference to Table 1.

[0131] Table 1. Integrity coefficient (K) and joint fracture statistics ( (Comparison)

[0132] D12. Determine the relationship between the integrity coefficient K and the rock compressive strength.

[0133] D13. When calculating BQ, the following conditions must be met: 1) When σ cw When >90K+30, with σ cw Substitute 90K + 30 into the equation to find the value of BQ. 2) When K > 0.04σ cw When +0.4, with K=0.04σ cw Substitute +0.4 to calculate the value of BQ.

[0134] D2. The BQ value is corrected by taking into account factors such as groundwater conditions, initial stress field, and the orientation of weak structural surfaces.

[0135] Calculate the corrected basic quality index [BQ] value of the rock mass: [ BQ ]= BQ 100 ( K 1+ K 2+ K 3) Wherein, [BQ] is the corrected rock mass quality index, K1 is the groundwater influence correction coefficient, K2 is the weak structural plane occurrence factor influence correction coefficient, and K3 is the geostress influence correction coefficient.

[0136] The values ​​of correction coefficients K1, K2, and K3 are determined using Tables 2, 3, and 4, respectively.

[0137] Table 2 Groundwater Influence Correction Factor K1

[0138] Note: 1. P is the fissure water pressure in the surrounding rock of the underground project (MPa); 2. Q is the water outflow rate per 10m tunnel length [L / (min·10m)] Table 3 Correction coefficients K2 for the attitude of major weak structural surfaces

[0139] Table 4 Correction factor K3 for the influence of natural stress

[0140] D3. Based on the corrected [BQ] value, classify the rock mass for quality and evaluate its stability. The evaluation shall be carried out according to the relevant indicators in Table 5.

[0141] Table 5 Rock Mass Quality Grading Evaluation Standards

[0142] This invention also provides an intelligent rock mass stability evaluation system for implementing the above-mentioned intelligent on-site rock mass stability evaluation method, comprising: The data acquisition and processing module is used to acquire images of the surrounding rock in underground mine tunnels, enhance the images using a low-light enhancement algorithm, and perform data augmentation and polygon annotation on the enhanced images to form a rock mass structure surface recognition dataset.

[0143] The rock mass structure surface recognition module is used to build and train a rock mass structure surface recognition model using a ResNet 50 improved U-Net network. The dataset is input into the trained rock mass structure surface recognition model, and the structural surface feature parameters are automatically extracted. The structural surface feature parameters include the number of fractures per unit area, the orientation of the structural surface, the interbedded rock, and the length and width of the traces.

[0144] The rock strength characteristic analysis module is used to correct the point load strength and calculate the rock compressive strength based on the corrected point load strength.

[0145] The rock mass quality classification and stability evaluation module is used to calculate the basic quality indicators of the rock mass based on the structural surface characteristic parameters and rock compressive strength. It then corrects the basic quality indicators of the rock mass by combining groundwater, geostress and weak structural surface occurrence factors to obtain the corrected rock mass quality indicators. Based on the corrected rock mass quality indicators, the stability of the rock mass is classified and evaluated.

[0146] The intelligent evaluation equipment integrates a data acquisition and processing module, a rock mass structure surface recognition module, a rock strength characteristic analysis module, and embedded algorithms to achieve rapid intelligent on-site evaluation. The intelligent evaluation equipment includes an explosion-proof housing, an integrated control board, a touch screen display, a power supply module, an industrial camera, an LED ring light, and UWB positioning tags. The positioning accuracy of the UWB positioning tags is at the centimeter level. The touch screen display is used to display the acquired images, recognition results, and stability evaluation results in real time.

[0147] like Figure 5 The image shows the hardware of the intelligent evaluation device, which consists of a control board, a touch screen, a power supply, and an industrial camera, serving as the carrier for image capture and program algorithm execution. like Figure 6 The image shown is of the intelligent detection program software, which consists of modules for parameter input, image display, operation control, and result output. It plays a role in hardware control and algorithm detection operation. This invention also provides an intelligent evaluation device for rock mass stability, comprising: Industrial cameras are used to collect raw image data of the surrounding rock in underground mine tunnels. Industrial cameras acquire images through a single-frame capture mode, and each trigger capture generates a static image which is then sent to the recognition system. LED ring lights are placed around industrial cameras to provide active hardware lighting in low-light environments. UWB positioning tags are used to communicate with the existing UWB positioning system in the mine, providing real-time centimeter-level positioning and obtaining the location information of the equipment in the roadway. The display device is used to display the acquired images, rock mass structure surface identification results, rock compressive strength and rock mass stability evaluation results in real time. Rock strength testing equipment is used to measure the point load strength of rock samples.

[0148] UWB positioning tags work in conjunction with existing UWB base stations in the mine to achieve centimeter-level positioning.

[0149] The equipment is suitable for the following scenarios with insufficient light: underground roadways, tunnels, underground chambers, and above-ground open-pit mines for the collection and stability evaluation of surrounding rock structures at night or in dark weather conditions.

[0150] This invention establishes a complete rock mass stability evaluation system based on Raspberry Pi, which is convenient for field use.

[0151] Example 1: On-site operation procedure This embodiment uses an underground iron ore tunnel as an application scenario to describe in detail the complete operation steps for rock mass stability evaluation using the system and equipment of this invention.

[0152] Step 1: Data Acquisition and Model Training: In a laboratory environment, a large number of images of surrounding rock in underground mine tunnels were acquired, covering different lighting conditions and structural surface types. Polygon annotation tools were used to mark the edges of structural surfaces such as joints, fissures, and rock inclusions. Based on the PyTorch framework, an improved U-Net network with ResNet50 as the backbone was used to train the recognition model, iterating more than 150 times until the model converged. After the trained model was solidified, it was deployed to the processor of the intelligent evaluation device.

[0153] Step Two: Hardware Equipment List: The industrial camera uses a resolution of 5 megapixels or higher. It generates RAW images by controlling sensor exposure via trigger signals, which are then processed by an ISP and output as JPEG or PNG still images before being sent to the processor. An LED ring light is installed around the camera to actively illuminate and eliminate shadows in low-light environments. All electronic components are integrated into an explosion-proof housing, meeting underground explosion-proof requirements. An integrated control board runs a Linux system, deploying deep learning models and evaluation algorithms. A 7-inch or larger touchscreen displays acquired images, recognition results, and stability evaluations. A UWB positioning tag communicates with the mine's existing positioning system. A portable point load cell is used to measure the point load strength of rocks on-site. A rechargeable lithium battery ensures continuous operation for more than 8 hours.

[0154] On-site setup procedure: Select the cross-section of the tunnel to be measured and remove surface loose rocks. Set up the tripod and aim the industrial camera at the surrounding rock area. After powering on and waiting for the system to start, click the positioning button on the touchscreen to confirm that the UWB tag is connected to the mine positioning system. Adjust the brightness of the LED ring light and preview the image to ensure that the structural surface is clearly visible. Finally, click the photo button to capture the current image.

[0155] Step 3: On-site identification and evaluation Image processing and structural surface recognition: The image after hardware-illuminated illumination is fed into a low-light enhancement algorithm for software correction, including illumination reconstruction and feature compensation. The enhanced image is then input into an improved U-Net model to automatically identify structural surfaces such as cracks, joints, and inclusions, outputting parameters such as the number of cracks per unit area, dip angle, trace length, and trace width. The system performs pixel-to-actual-size conversion based on the markers.

[0156] Rock strength data acquisition: Rock samples are collected on-site, and point load strength is measured using a portable point load tester. Test results are manually entered via touchscreen or automatically transmitted via Bluetooth, and the system automatically calculates the rock compressive strength.

[0157] Stability level determination: The system substitutes the structural surface characteristic parameters and rock compressive strength into the BQ classification model, and combines the correction coefficients of factors such as groundwater, geostress, and the occurrence of weak structural surfaces to calculate the corrected rock mass quality index [BQ] value, and automatically determines the stability level of rock mass from level I to level V.

[0158] Step 4: Results Display and Data Management The touchscreen simultaneously displays the original image, the software-enhanced image, the image with structural surface identification and annotation (fractures are marked with red lines), a table of structural surface characteristic parameters, rock compressive strength values, corrected rock mass quality index [BQ], and stability evaluation results. All results are automatically saved to local storage, with each data point associated with centimeter-level location information. Data reports can be exported via USB flash drive or downhole wireless network.

[0159] Step 5: Summary of Hardware, Model, Algorithm, Results, and Value Creation In this embodiment, the hardware modules and the model algorithm work together, as detailed below: The industrial camera uses a single-frame capture mode to acquire static images of the surrounding rock, which can be used on-site without relying on a network and outputs the original image, providing basic data for the entire recognition process.

[0160] As a hardware active lighting module, the LED ring light provides active lighting in low-light environments, eliminating shadows, enhancing details in dark areas, and outputting images after hardware lighting, significantly improving image quality under complex lighting conditions in mines.

[0161] The low-light enhancement algorithm, as a secondary enhancement module in software, works in conjunction with hardware illumination. Through the illumination extraction module and the feature fusion module, it realizes illumination reconstruction and feature compensation, further enhancing the image after hardware illumination and outputting the enhanced image, providing high-quality input for subsequent structural surface recognition.

[0162] The U-Net model is improved as the core algorithm for structural surface recognition. ResNet50 is used as the backbone feature extraction network to replace the ordinary convolutional blocks of the U-Net network. The enhanced image is segmented and recognized at the pixel level, and structural surface feature parameters such as cracks, joints, and rock inclusions are output. This realizes the full automation of structural surface recognition, replacing the traditional manual statistical method and greatly improving efficiency and accuracy.

[0163] UWB positioning tags communicate with the mine's existing UWB positioning system to obtain real-time location information of equipment in the roadway, achieving high-precision traceability and repeatable measurement, and providing a location benchmark for long-term surrounding rock deformation monitoring.

[0164] As a rock strength testing device, the point load tester collects rock samples on-site for point load tests, calculates the rock compressive strength using empirical formulas, and outputs the rock compressive strength value, effectively supplementing the rock mechanical parameters that cannot be obtained by pure image recognition.

[0165] The BQ grading and evaluation module is based on the modified BQ formula. It integrates structural surface characteristic parameters with rock compressive strength, and combines correction coefficients such as groundwater, geostress, and the occurrence of weak structural surfaces to calculate the modified rock mass quality index [BQ] and determine the stability level. It outputs the final evaluation conclusion, providing direct guidance for mine blasting parameter design and support scheme formulation.

[0166] Figure 7(a) shows the original low-light image without hardware illumination or software correction. Figure 7(b) shows the image after hardware illumination (LED ring light) only. Figure 7(c) shows the final image after hardware illumination plus software low-light enhancement algorithm correction. Figure 8 This is a schematic diagram of the overall structure and field scenario of the intelligent rock mass stability evaluation system and equipment in an embodiment of the present invention.

[0167] In this embodiment, the entire process, from equipment setup to outputting evaluation results, can be completed within 10 minutes, achieving the goal of taking photos and generating data on-site. The UWB centimeter-level positioning function ensures that the positional deviation of multiple measurements at the same measuring point is less than 10cm, providing a reliable positional benchmark for long-term surrounding rock deformation monitoring.

[0168] 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; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rock mass stability intelligent evaluation method, characterized in that, Includes the following steps: Images of surrounding rock in underground mine roadways are collected, a low-light enhancement algorithm is designed to enhance the images, the enhanced images are augmented, and polygon annotations are performed on the augmented images to form a dataset. The low-light enhancement algorithm is implemented based on an illumination extraction module and a feature fusion module. A rock mass structure surface recognition model based on an improved U-Net network was constructed and trained. The dataset was input into the trained rock mass structure surface recognition model to obtain the structure surface feature parameters, which include the number of fractures per unit area, dip angle, rock inclusions, and trace length and width. The U-Net network was improved using ResNet 50. The point load strength of the rock sample in the field is measured, the point load strength is corrected, and the compressive strength of the rock is calculated based on the corrected load strength. Based on the structural surface characteristic parameters and rock compressive strength, the basic quality index of the rock mass is calculated. The basic quality index of the rock mass is then corrected by combining groundwater, geostress and weak structural surface occurrence factors to obtain the corrected rock mass quality index. The stability of the rock mass is then classified and evaluated based on the corrected rock mass quality index.

2. The intelligent evaluation method for rock mass stability according to claim 1, characterized in that, The method also includes the step of interfacing with the mine's existing UWB positioning system: integrating a UWB positioning tag into the acquisition device, and using the UWB positioning tag to obtain the device's location information in the roadway in real time; and storing the coordinate information in association with the acquired images, structural surface recognition results, and rock mass stability evaluation results.

3. The intelligent evaluation method for on-site rock mass stability according to claim 1, characterized in that, The illumination extraction module includes a first channel splicing layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and a first linear combination layer connected in sequence. The feature fusion module includes, in sequence, a self-feature enhancement layer, a first Transformer module, a first downsampling module, a second Transformer module, a second downsampling module, a third Transformer module, a third downsampling module, a fourth Transformer module, a first Prompt module, a first upsampling module, a second channel concatenation layer, a fourth convolutional layer, a fifth Transformer module, a second Prompt module, a second upsampling module, a third channel concatenation layer, a fifth convolutional layer, a sixth Transformer module, a third Prompt module, a third upsampling module, a fourth channel concatenation layer, a seventh Transformer module, an output layer, and an element-wise addition layer. The output of the first Transformer module is connected to the input of the fourth channel concatenation layer, the output of the second Transformer module is connected to the input of the third channel concatenation layer, and the output of the third Transformer module is connected to the input of the second channel concatenation layer.

4. The method according to claim 1, characterized in that, The calculation process of the low-light enhancement algorithm includes: The input image is extracted from the dataset. The two input images are passed through the first channel stitching layer to obtain the first stitching feature map. The first stitching feature map is passed through the first convolutional layer and the second convolutional layer in sequence to obtain the intermediate lighting feature. The intermediate lighting feature is passed through the third convolutional layer to obtain the convolutional feature map. The convolutional feature map and the input image are input into the first linear combination layer to obtain the lighting mapping image. The illumination mapping image is passed sequentially through a self-feature enhancement layer and a first Transformer module to obtain a first Transformer feature. The first Transformer feature is then passed sequentially through a first downsampling module and a second Transformer module to obtain a second Transformer feature. The second Transformer feature is then passed sequentially through a second downsampling module and a third Transformer module to obtain a third Transformer feature. The third Transformer feature is sequentially passed through a third downsampling module, a fourth Transformer module, a first Prompt module, and a first upsampling module to obtain a first upsampling output. The first upsampling output and the third Transformer feature are then passed through a second channel concatenation layer to obtain a second concatenated feature map. The second concatenated feature map is sequentially passed through a fourth convolutional layer, a fifth Transformer module, a second Prompt module, and a second upsampling module to obtain a second upsampling output. The second upsampling output and the second Transformer feature are then passed through a third channel concatenation layer to obtain a third concatenated feature map. The third concatenated feature map is sequentially passed through a fifth convolutional layer, a sixth Transformer module, a third Prompt module, and a third upsampling module to obtain a third upsampling output. The third upsampling output and the first Transformer feature are then passed through a fourth channel concatenation layer to obtain a fourth concatenated feature map. The fourth concatenated feature map is then passed through a seventh Transformer module to obtain a seventh feature map. After the seventh feature map is output, it is passed to the input image through an element-wise addition layer to output the enhanced image.

5. The intelligent rock mass stability evaluation method according to claim 1, characterized in that, The low-light enhancement algorithm achieves the supplementary lighting effect through a dual mechanism of illuminance reconstruction and feature compensation, and works in conjunction with a hardware supplementary lighting device, specifically including: The hardware lighting device uses an LED ring light, which is placed around the camera to provide active lighting in low-light environments, eliminate shadows, and enhance details in dark areas. The low-light enhancement algorithm, as a software correction module, performs secondary enhancement on the image acquired after hardware illumination. The illumination extraction module processes the input low-light image. I in The channel mean is calculated to obtain the initial illuminance estimation map. The initial illumination estimation map is concatenated with the original image through channels and then fed into a convolutional layer to generate an illumination mapping image. I illu The expression for the illumination mapping image is: in, This indicates a channel splicing operation. This is a combined mapping of depthwise separable convolution and 1x1 convolution; The feature fusion module inputs the illumination map image into the Transformer encoder, extracts illumination features at different scales through the downsampling module, and introduces a Prompt module at each stage to activate and enhance dark area features. The Prompt module utilizes intermediate illumination features obtained from the illumination extraction module. F illu The feature map is reconstructed using an attention mechanism with weights to obtain enhanced features. The expression for the enhanced features is as follows: in, F in Input features for the current layer, F enh For enhanced features; Finally, the enhanced feature map and the original input image are fused through an element-by-element addition layer to generate an output image after light compensation I out The expression of the output image is: wherein, F out enhanced feature maps output by the feature fusion module, are learnable fusion weight coefficients.

6. The intelligent rock mass stability evaluation method according to claim 3, characterized in that, The first, second, and third Prompt modules have the same network architecture. The workflow of the first Prompt module includes: The input feature map is fed into the prediction layer, which includes a pooling layer, a linear layer, and a softmax layer, to obtain the predicted feature map. The predicted feature map is linearly combined with the parameters to obtain the combined feature map. The combined feature map and the input feature map are then subjected to interpolation to obtain the interpolated feature map. The feature map after the difference processing is passed through the sixth convolutional layer, the fourth Prompt layer and the first normalization layer in sequence to obtain the first normalized feature map. The intermediate illumination feature is passed through the difference layer and the second normalization layer in sequence to obtain the second normalized feature map. The first normalized feature map and the second normalized feature map are passed through the first attention layer to obtain the output of the attention layer. The output of the attention layer is sequentially passed through the second linear combination layer, the third normalization layer, the feedforward neural network layer, the third linear combination layer, the fourth normalization layer, the separable two-dimensional convolutional layer, and the second attention layer to obtain the output of the first Prompt module.

7. The intelligent rock mass stability evaluation method according to claim 1, characterized in that, The improvement of the rock mass structure surface recognition model lies in replacing the ordinary convolutional blocks of the U-Net network with ResNet 50 residual blocks; The workflow of the rock mass structure surface identification model includes: The input image is processed by ResNet 50 residual and max pooling four times to obtain the first feature map; The first feature map is processed by ResNet 50 residual and upconvolution four times to obtain the second feature map; The second feature map is subjected to ResNet 50 residual processing three times to obtain the structural surface feature parameters.

8. The intelligent rock mass stability evaluation method according to claim 1, characterized in that, The number of cracks per unit area is calculated by converting pixels to actual area using markers. The formula for calculating the actual area is: in, This is the actual area. The number of pixels in the image. The number of pixels of the marker. The actual area of ​​the marker. The number of cracks per unit area is calculated based on the actual area. The formula for calculating the number of cracks per unit area is as follows: in, The number of cracks per unit area. The number of cracks detected. By identifying the angle between the weak structural surface and the hole axis, the weak structural surface is fitted into a straight line using a fitting method. The projection lengths of the straight line on the y-axis and x-axis are calculated, and then the inclination angle is calculated. The formula for calculating the inclination angle is as follows: wherein is the tilt angle, is the length of the straight line in the x-axis projection, is the length of the straight line in the y-axis projection, The trace length is calculated using the least squares fitting method, and the formula for calculating the trace length is as follows: in, For the length of the trace, ( Let be the coordinates of the first endpoint of the line. Let x be the x-coordinate of the first endpoint. Let y be the y-coordinate of the first endpoint. Let be the coordinates of the second endpoint of the line. Let x be the x-coordinate of the second endpoint. Let be the y-coordinate of the second endpoint. The actual length of the marker. The formula for calculating the trace width is: in, 'For the width of the gap, The area of ​​the crack pixel. The calculation formula for the basic quality indicators of the rock mass is as follows: wherein, is the basic quality index of rock mass, is the compressive strength of rock, is the integrity coefficient of rock mass, The formula for calculating the corrected rock mass quality index is as follows: [ BQ ]= BQ 100( K 1+ K 2+ K 3) Wherein, [BQ] is the corrected rock mass quality index, K1 is the correction coefficient for the influence of groundwater, K2 is the correction coefficient for the influence of the attitude of weak structural planes, and K3 is the correction coefficient for the influence of in-situ stress. The calculation process for the compressive strength of the rock includes: The rock specimen was placed between two spherical conical pressure plates and continuously loaded before the specimen failed. The morphological characteristics, failure mode and bearing strength of the specimen were recorded. The dimensional parameters on the failure surface of the rock specimen were measured, including the spacing between loading points and the width of the fracture surface. Calculate the equivalent core diameter based on the stated size parameters: wherein, is the equivalent core diameter, is the fracture face width, is the loading point spacing; Calculate the uncorrected point load strength based on the bearing capacity and the equivalent core diameter: wherein, is the uncorrected point load strength, is the bearing strength; Introducing a correction factor, and based on the correction factor and the uncorrected point load strength, calculating the corrected point load strength: wherein, is the corrected point load strength, is a correction factor, the calculation formula of which is: wherein m is a correction index; Calculate the rock compressive strength based on the corrected point load strength: in, It represents the compressive strength of the rock.

9. A rock mass stability intelligent evaluation system, used to implement the on-site rock mass stability intelligent evaluation method according to any one of claims 1-8, characterized in that, include: The data acquisition and processing module is used to acquire images of the surrounding rock in underground mine roadways, enhance the images using a low-light enhancement algorithm, and perform data augmentation and polygon annotation on the enhanced images to form a rock mass structure surface recognition dataset. The rock mass structure surface recognition module is used to construct and train a rock mass structure surface recognition model using a ResNet 50 improved U-Net network. The dataset is input into the trained rock mass structure surface recognition model, and the structure surface feature parameters are automatically extracted. The structure surface feature parameters include the number of fractures per unit area, the orientation of the structure surface, the interbedded rock, and the length and width of the trace. The rock strength characteristic analysis module is used to correct the point load strength and calculate the rock compressive strength based on the corrected point load strength. The rock mass quality classification and stability evaluation module is used to calculate the basic quality index of the rock mass based on the structural surface characteristic parameters and rock compressive strength, and to modify the basic quality index of the rock mass by combining groundwater, geostress and weak structural surface occurrence factors to obtain modified rock mass quality index, and to classify and evaluate the stability of the rock mass based on the modified rock mass quality index. The intelligent evaluation equipment integrates a data acquisition and processing module, a rock mass structure surface identification module, a rock strength characteristic analysis module, and an embedded algorithm to achieve rapid intelligent on-site evaluation. The intelligent evaluation equipment includes an explosion-proof housing, an integrated control board, a touch screen display, a power module, a tripod, an industrial camera, an LED ring light, and a UWB positioning tag. The touch screen display is used to display the acquired images, recognition results, and stability evaluation results in real time.

10. A rock mass stability intelligent evaluation device, characterized in that, include: An industrial camera is used to collect raw image data of the surrounding rock in underground mine tunnels. The industrial camera acquires images through a single-frame capture mode, and generates a static image each time it is triggered to be sent to the recognition system. LED ring lights are placed around the industrial camera to provide active hardware lighting in low-light environments. UWB positioning tags are used to communicate with the existing UWB positioning system in the mine, providing real-time centimeter-level positioning and obtaining the location information of the equipment in the roadway. The display device is used to display the acquired images, rock mass structure surface identification results, rock compressive strength and rock mass stability evaluation results in real time. Rock strength testing equipment is used to measure the point load strength of rock samples; A processor is configured to execute the computer program in the memory to implement the intelligent evaluation system for on-site rock mass stability as described in claim 9.

Images