Surrounding rock quality evaluation method and system based on TBM slag charge image intelligent identification

Abstract

The invention relates to a surrounding rock quality evaluation method, and discloses a surrounding rock quality evaluation method and system based on TBM slag charge image intelligent identification, and the method comprises the steps: collecting a slag charge image in real time; inputting the slag charge image into a trained recognition and segmentation model to obtain a pixel-level segmentation result about slag charge particles in the slag charge image; and performing quality evaluation on the rock mass which is being tunneled according to the pixel-level segmentation result and a set rock mass quality evaluation rule. According to the method, the surrounding rock quality can be further inverted according to the extracted slag particle morphological characteristics and grading curve, and the TBM tunneling performance can be analyzed in an auxiliary manner, so that the construction efficiency is improved, and the safety and reliability of tunneling operation are guaranteed.

Description

Technical Field

[0001] This invention relates to a method for assessing the quality of surrounding rock, specifically to a method and system for assessing the quality of surrounding rock based on intelligent image recognition of TBM slag. Background Technology

[0002] TBMs (Tunnel Boring Machines) have become the main equipment for modern tunnel construction. During TBM construction, their operating status and advancement efficiency are highly sensitive to the surrounding rock geological conditions. Traditional methods of sensing rock conditions mainly rely on visual assessment by operators, which is inefficient and highly subjective. Furthermore, due to the obstruction caused by the TBM equipment, it is often difficult to observe the rock conditions at the tunnel face, hindering operators from obtaining complete geological information in a timely manner and adjusting equipment parameters.

[0003] During the TBM construction process, the cutterhead continuously breaks and cuts the rock mass. The broken rock mass cut by the cutterhead is transported and discharged as excavated material via belt conveyor. The morphology, particle size distribution, and accumulation characteristics of the excavated material generated by the cutterhead are important indicators reflecting geological conditions and tunneling efficiency. They are crucial for assessing surrounding rock quality, optimizing tunneling parameters, and reflecting the operating status of the TBM equipment. Traditional excavated material analysis typically relies on manual sampling and screening experiments, which is extremely time-consuming and labor-intensive. The results obtained are often highly lagging and have poor sample representativeness, failing to meet the needs of TBM's high-efficiency construction schedule and real-time decision-making in intelligent construction.

[0004] In comparison, acquiring images of slag is much simpler. Industrial cameras enable non-contact, continuous acquisition, providing a data foundation for online slag analysis. Theoretically, image processing techniques can quickly extract parameters such as particle size, size distribution, and shape of the slag. However, TBM construction tunnels often suffer from poor lighting conditions, and the slag is typically covered in mud, contains a large number of fine particles, and is often piled up during belt conveyor transport. Rock blocks and fine particles often obscure or adhere to each other, resulting in low accuracy of traditional image processing methods in edge detection, target segmentation, and particle size identification, making it difficult to meet the high-precision analysis requirements under complex working conditions.

[0005] Deep learning technology possesses powerful learning and feature extraction capabilities in image recognition, overcoming the limitations of traditional image processing. Therefore, a real-time image acquisition system was used to rapidly capture slag images, and a deep learning model incorporating multiple algorithms was employed to achieve high-precision pixel-level segmentation. The segmentation results of the slag images were then quantified to obtain the morphological parameters of the slag particles.

[0006] When the TBM (Tunnel Boring Machine) tunnels into high-quality surrounding rock, the cutterhead advance speed slows down and the cycle tunneling time is prolonged due to the hardness and high uniaxial compressive strength of the rock. This necessitates increasing the main propulsion force and cutterhead rotation speed to maintain efficiency. At this stage, the excavated material is mostly flaky and powdery, with small and uniform particle size. Conversely, in lower-quality surrounding rock, due to lower rock strength and well-developed joints, the tunneling speed is faster, resulting in more large-diameter rock fragments with uneven particle size distribution, often featuring smooth joint surfaces. Furthermore, the gradation curves of surrounding rock with different integrity levels differ significantly, often exhibiting "L-shaped," "trapezoidal," and plateau-like characteristics. Therefore, by extracting excavated material morphology parameters through image recognition, rock mass quality evaluation indicators can be constructed, providing a valid basis for TBM tunneling parameter optimization and construction safety. Summary of the Invention

[0007] To address the problems of manual screening being time-consuming, labor-intensive, and ineffective, and traditional image processing techniques having low segmentation accuracy, poor robustness, and inability to objectively evaluate rock mass quality based on the obtained slag morphology parameters, this invention provides a surrounding rock quality assessment method based on intelligent image recognition of TBM slag. This method utilizes a deep learning model that integrates multiple algorithms, combined with a real-time image acquisition system on-site, to perform rapid, real-time, and high-precision morphological quantification and analysis of slag generated during TBM tunneling.

[0008] This invention is achieved through the following technical solution: A method for assessing surrounding rock quality based on intelligent image recognition of TBM slag includes: Real-time acquisition of slag images; The slag image is input into a trained recognition and segmentation model to obtain pixel-level segmentation results of slag particles in the slag image. The quality of the rock mass being excavated is assessed based on the pixel-level segmentation results and the established rock mass quality evaluation rules.

[0009] As an optimization, the recognition and segmentation model includes a Faster R-CNN object detection sub-model and a U-Net image segmentation sub-model. The Faster R-CNN object detection sub-model is used to extract the bounding box position parameters of the slag particles in the slag image, and the U-Net image segmentation sub-model is used to perform image segmentation on the slag image based on the bounding box position parameters to obtain the pixel-level segmentation result of each slag particle in the slag image.

[0010] As an optimization, the Faster R-CNN object detection sub-model includes a feature extraction network module, an FPN module, an RPN module, and a RoI Pooling module. The feature extraction network module includes several convolutional layers with scales gradually decreasing from input to output, used to extract features from the slag image to obtain initial feature maps of different scales. The FPN module is used to obtain multi-level semantic features from each of the initial feature maps, then progressively upsamples from high-level semantic features to lower levels, and during the upsampling process, performs a 1×1 convolution with the initial feature map of the same level and adds them together to fuse high-level semantic information and low-level spatial information. Finally, a 3×3 convolutional layer is used to obtain the final multi-scale fused feature map. The RPN module is used to obtain candidate regions of the slag image based on the multi-scale fused feature map. The RoI Pooling module is used to perform region-of-interest pooling on the multi-scale fused feature map using the candidate regions to finally obtain the bounding box position parameters of the slag particles.

[0011] As an optimization, the RPN module includes a 3×3 convolutional layer, two 1×1 convolutional layers, and anchor boxes. The multi-scale fused feature map is passed through the 3×3 convolutional layer and then through two 1×1 convolutional layers to obtain the target detection confidence and bounding box regression values. The target detection confidence and bounding box regression values ​​are combined with the anchor boxes to generate candidate regions.

[0012] As an optimization, the U-Net image segmentation sub-model includes several downsampling units and several upsampling units, with the number of downsampling units and upsampling units being the same. Along the data transmission direction, the dimension of the downsampling unit is the same as the dimension of the corresponding convolutional layer in the feature extraction network module. Simultaneously, the first downsampling unit inputs an intermediate image after the invalid region of the slag image is masked by the bounding box position parameters. The several downsampling units and several upsampling units are connected in a U-Net skip connection, directly passing the first feature map of the several downsampling units to the corresponding upsampling unit, and finally obtaining the pixel-level segmentation result of each slag particle.

[0013] As an optimization, the rock mass quality evaluation rule is specifically as follows: The pixel-level segmentation results of the slag particles are sampled and quantized into equivalent ellipsoids to calculate the volume V and mass M of the slag particles. , a and b represent the major and minor axes of the slag particles, respectively. This indicates the density of the slag particles; The volume V and mass M of the slag particles are used to obtain particle morphology characteristics, including average roundness. Gradient curvature coefficient and non-uniformity coefficient ; The quality of the surrounding rock is evaluated using the aforementioned particle morphology characteristics and a rock mass quality evaluation formula. The rock mass quality evaluation formula... ; N represents the mass of the rock mass and the number of slag particles. Indicates the fine particle content coefficient. Indicates the median particle size. It represents the ideal non-uniformity coefficient.

[0014] As an optimization, the average roundness of the slag particles is obtained by measuring their volume V and mass M. Gradient curvature coefficient and non-uniformity coefficient The process is as follows: The average roundness of each slag particle is calculated to obtain the average roundness in the corresponding slag image. ; Particle size distribution curves were plotted using the volume / mass ratios of different slag particles, thereby obtaining the gradation curvature coefficient. ; The non-uniformity coefficient is calculated based on the two characteristic particle sizes on the particle size distribution curve. , , To limit the particle size, The effective particle size.

[0015] This invention also discloses a surrounding rock quality assessment system based on intelligent image recognition of TBM slag, comprising: The acquisition module is used to acquire images of slag in real time. The recognition and segmentation module is used to input the slag image into the trained recognition and segmentation model to obtain pixel-level segmentation results of slag particles in the slag image. The evaluation module is used to evaluate the quality of the rock mass being excavated based on the pixel-level segmentation results and the set rock mass quality evaluation rules.

[0016] As an optimization, the acquisition module includes a line array camera, an encoder, a light source, and a controller mounted on the slag conveyor belt. The light emitted by the light source is directed towards the conveyor belt. The line array camera is located below the light source and is used to capture images of slag particles on the conveyor belt. The encoder is used to monitor the operating status of the conveyor belt in real time and determines the timing of the line array camera's capture by cooperating with the line array camera. The controller is used to control the capture by the line array camera and transmit the image data captured by the line array camera to the terminal through a communication module.

[0017] As an optimization, the terminal includes one or more of a cloud server, a local computer, and a mobile phone.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention effectively improves the efficiency of slag morphology recognition and analysis, overcomes the technical problem of low segmentation accuracy in the case of random slag stacking by traditional image recognition methods, and can further invert the surrounding rock quality based on the extracted slag particle morphology characteristics and gradation curves to assist in the analysis of TBM tunneling performance, thereby improving construction efficiency and ensuring the safety and reliability of tunneling operations. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart below shows a method for assessing the quality of surrounding rock based on intelligent image recognition of TBM slag, as described in this invention. Figure 2 This is a schematic diagram of the structure of the recognition and segmentation model in this invention; Figure 3 This is a curve of the rock mass fracture index. Figure 4 This is a schematic diagram of the hardware connection of a surrounding rock quality assessment system based on intelligent image recognition of TBM slag, as described in this invention. Figure 5 A schematic diagram comparing the original image of slag material, manually annotated images, and images annotated using a recognition and segmentation model; Figure 6 This is a graph showing the quantification results of the morphological parameters of each slag particle in the slag image; Figure 7 (a) is a schematic diagram of identification, segmentation, and morphological quantization; Figure 7 (b) is a schematic diagram of the gradation curve; Figure 8 This is a schematic diagram of the rock mass quality evaluation calculation results in a specific embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0021] This embodiment 1 provides a method for assessing the quality of surrounding rock based on intelligent image recognition of TBM slag, such as... Figure 1 As shown, it includes: S1. Real-time acquisition of slag images.

[0022] S2. Input the slag image into the trained recognition and segmentation model to obtain pixel-level segmentation results of slag particles in the slag image.

[0023] Due to the complex packing morphology and imaging conditions of slag, target particles (also known as slag particles) are often mixed with interfering substances (such as water and fine slag powder). Therefore, to accurately identify target particles, it is first necessary to detect and identify the correct target particles, and secondly, to accurately segment the target particles. This method employs a fusion deep learning model that combines the Faster R-CNN object detection algorithm and the U-Net image segmentation algorithm to achieve automated slag particle identification. The two sub-networks (sub-models) correspond to target particle detection and high-precision segmentation, respectively. Figure 2 .

[0024] In some embodiments, the recognition and segmentation model includes a Faster R-CNN object detection sub-model and a U-Net image segmentation sub-model. The Faster R-CNN object detection sub-model is used to extract bounding box position parameters of slag particles in the slag image, and the U-Net image segmentation sub-model is used to perform image segmentation on the slag image based on the bounding box position parameters to obtain pixel-level segmentation results for each slag particle in the slag image.

[0025] In some embodiments, the Faster R-CNN object detection sub-model includes a feature extraction network module, an FPN module, an RPN module, and an RoI Pooling module. The feature extraction network module includes several convolutional layers whose scale gradually decreases from input to output, used to extract features from the slag image to obtain initial feature maps of different scales.

[0026] like Figure 2 As shown, this embodiment has five convolutional layers, namely D1-D5, and the scale of D1-D5 gradually decreases.

[0027] The FPN module is used to obtain multi-level semantic features from each of the initial feature maps, and then gradually upsample from the high-level semantic features to the low-level features. During the upsampling process, it performs a 1×1 convolution with the initial feature map of the same level and adds them together to fuse the high-level semantic information and the low-level spatial information. Finally, it obtains the final multi-scale fused feature map through a 3×3 convolutional layer.

[0028] Figure 2 In the process, the FPN module extracts semantic features from D1-D5 respectively. - (From D to M, it is through 1×1 Conv ( Figure 2 (The arrow from D to M is used to implement this). - The semantic features are arranged sequentially from the lowest level to the highest level. Then, upsampling begins from M5, and during the upsampling process, the features are convolved with the initial feature map of the same layer using a 1×1 method and then added together. M5 and... Same. During the process from M5 to M4, M4 is... + The result after upsampling. Other procedures are the same.

[0029] The RPN module is used to obtain candidate regions of the slag image based on the multi-scale fused feature map.

[0030] In some embodiments, the RPN module includes a 3×3 convolutional layer, two 1×1 convolutional layers, and anchor boxes. The multi-scale fused feature map is passed through the 3×3 convolutional layer and then through the two 1×1 convolutional layers to obtain the target detection confidence and bounding box regression values. The target detection confidence and bounding box regression values ​​are combined with the anchor boxes to generate candidate regions.

[0031] The RoI Pooling module is used to perform region of interest pooling on the multi-scale fused feature map using the candidate regions, and finally obtains the bounding box position parameters of the slag particles.

[0032] This sub-model performs the object detection stage. First, the input image (slag image) is fed into the object detection sub-network (i.e., the Faster R-CNN object detection sub-model). Features are extracted through the feature extraction convolutional layers (D1-D5) in the feature extraction network module. For detection tasks, target slag particles in an image vary in size and shape. Larger feature maps have lower abstraction levels and retain more detailed information, making them suitable for detecting small targets, while smaller feature maps are better suited for detecting larger targets. If only a single-scale feature map is used for detection, information is easily lost. Therefore, theoretically, performing object detection on feature maps of different scales simultaneously should yield better results. Thus, the convolutional layers (D2-D5) are used to generate multi-scale feature representations (P2-P5) through a Feature Pyramid Network (FPN), achieving effective extraction and fusion of multi-scale features, thus enhancing the detection and segmentation accuracy for particles of different sizes. Faster R-CNN's Faster Propagation Network (FPN) first extracts multi-level features (M2-M5) from the feature extraction convolutional layers. Then, it progressively upsamples from top to bottom, starting with high-level semantic features (M5 to M2) (2×Up) to propagate semantic information to lower layers. During upsampling, it performs 1×1 convolutions with features from the same layer and adds them together, thus fusing high-level semantic information and low-level spatial information. Finally, a 3×3 convolution is used to obtain the final multi-scale fused feature map (P2-P5). The resulting feature map (P2-P5) is fed into Faster R-CNN's RPN module. After a 3×3 convolution, two 1×1 convolutions are performed to obtain the object detection confidence and bounding box regression values. These are combined with anchors to generate candidate regions (proposals), which are then used to perform ROI pooling on the multi-scale feature map (P2-P5). Finally, the bounding box positions of the target particles are output.

[0033] In some embodiments, the U-Net image segmentation sub-model includes a plurality of downsampling units and a plurality of upsampling units, wherein the number of downsampling units and the number of upsampling units are the same. Specifically, the number of downsampling units is the number of convolutional layers minus 1, and each downsampling unit corresponds to one of the N-1 convolutional layers with the largest scale, where N is the number of convolutional layers. Figure 2 As shown, N is 5, so the convolutional layers are D1-D5, and the downsampling units are S1-S4. S1-S4 correspond to D1-D4 respectively.

[0034] Along the data transmission direction, the dimension of the downsampling unit is the same as the dimension of the corresponding convolutional layer in the feature extraction network module. For example... Figure 2As shown, D1 is 64×256×256, so S1 is also 64×256×256; D2 is 256×128×128, so S2 is also 256×128×128; S3 has the same dimensions as D3, and S4 has the same dimensions as D4.

[0035] The first downsampling unit inputs an intermediate image after the invalid region of the slag image has been masked using the bounding box position parameters. That is, the intermediate image input to downsampling unit S1 is an image in which the positions of each target particle in the slag image have been located using the bounding box position parameters, and the pixel values ​​outside the bounding box have been cleared to zero.

[0036] A U-Net-style skip connection is used to connect several downsampling units and several upsampling units, directly passing the first feature map of several downsampling units to the corresponding upsampling units, and finally obtaining the pixel-level segmentation result of each slag particle.

[0037] like Figure 2 As shown, the intermediate image input to the U-Net image segmentation sub-model is skipped to the upsampling unit S8, the downsampling unit S1 is skipped to the upsampling unit S7, the downsampling unit S2 is skipped to the upsampling unit S6, and the downsampling unit S3 is skipped to the upsampling unit S5.

[0038] U-Net-style skip connections, also known as encoder-decoder skip connections, are characterized by directly connecting feature maps from different levels in the encoder (downsampling stage, as shown in the input S1-S4) to the corresponding scale levels in the decoder (upsampling stage, as shown in the S5-S8). This allows for the fusion of low-level detail features (such as edges and textures) with high-level semantic features (such as target category information) through concatenation or addition.

[0039] This sub-model performs the segmentation stage, feeding the bounding box parameters output by the detection sub-network (Faster R-CNN object detection sub-model) along with the original image (slag image) into the segmentation sub-network. The bounding boxes locate the positions of each target particle in the image. The model clears the pixel values ​​outside the bounding boxes to zero, thereby masking invalid regions, allowing the segmentation sub-network to focus only on the target particle regions. The segmentation sub-network first extracts input image features through four downsampling stages (S1-S4), consistent with D1-D4 in the detection sub-network, achieving low-level feature sharing to improve the consistency and feature representation capability of the overall model. Subsequently, the image enters four upsampling stages (S5-S8) to gradually restore the spatial resolution of the feature maps. During this process, the U-Net skip connection structure is introduced, directly passing the feature maps from the downsampling stages to the corresponding upsampling stages, avoiding the loss of spatial information during upsampling. Finally, the segmentation sub-network outputs pixel-level segmentation results for each target particle, achieving high-precision particle boundary extraction.

[0040] S3. Based on the pixel-level segmentation results and the established rock mass quality evaluation rules, the quality of the rock mass being excavated is assessed.

[0041] In some embodiments, the rock mass quality evaluation rules are specifically as follows: A1. Quantize the pixel-level segmentation results of the slag particles using an equivalent ellipsoid, and calculate the volume V and mass M of the slag particles. , a and b represent the major and minor axes of the slag particles, respectively. This indicates the density of the slag particles; A2. The volume V and mass M of the slag particles are used to obtain particle morphology characteristics, including average roundness. Gradient curvature coefficient and non-uniformity coefficient ; A3. The quality of the surrounding rock is evaluated using the aforementioned particle morphology characteristics and a rock mass quality evaluation formula. The rock mass quality evaluation formula... ; N represents the mass of the rock mass and the number of slag particles. Indicates the fine particle content coefficient. Indicates the median particle size. It represents the ideal non-uniformity coefficient.

[0042] The relationship between volume and mass and the Rock Quality Evaluation (RQI) formula is mainly reflected in the provision of basic parameters through "equivalent ellipsoid quantification," which supports the calculation of particle morphology and gradation characteristics in the RQI. The specific logical chain is as follows: The volume (V) and mass (M) calculated using the equivalent ellipsoid formula are the basis for subsequent derivation of morphological parameters such as "roundness" and "elongation". For example, roundness can be calculated using the ratio of the major axis (a) and the minor axis (b) (or a derived formula).

[0043] The core logic of RQI is to comprehensively evaluate rock mass quality by combining fragmentation characteristics (number of rock blocks, proportion of fine grains) and integrity characteristics (size, roundness, and gradation uniformity of rock blocks), with volume and mass playing a crucial role throughout the process. Crushing characteristics: Content coefficient of fine particles (<5mm) ( The number of rock blocks N depends on the identification of the volume / mass of individual particles.

[0044] Complete characteristics: Median particle size Average roundness All of these require first quantifying particle morphology through volume / mass, and then statistically deducing the results.

[0045] In some embodiments, the volume V and mass M of the slag particles are used to obtain the average roundness. Gradient curvature coefficient and coefficient of uniformity The process is as follows: The average roundness of each slag particle is calculated to obtain the average roundness in the corresponding slag image. ; Particle size distribution curves were plotted using the volume / mass ratios of different slag particles, thereby obtaining the gradation curvature coefficient. ; The non-uniformity coefficient is calculated based on the two characteristic particle sizes on the particle size distribution curve. , , To limit the particle size, The effective particle size is represented by i and j, which are positive numbers.

[0046] For example, This indicates that particles smaller than this size account for 60% of the total mass, also known as the "limited particle size". It reflects the particle size at the 60% position in the particle gradation, and in the gradation curve, it corresponds to the x-axis (particle size value) when the cumulative percentage content is 60%.

[0047] The particle size smaller than this value accounts for 10% of the total mass, also known as the "effective particle size". The x-axis (particle size value) of the corresponding gradation curve is the position of the particle size when the cumulative percentage content is 10%.

[0048] Plotting the particle size distribution curve: Through sieving tests (or image recognition, etc.), statistically analyze the particle mass (or volume, which must be consistent) of different particle size ranges, calculate the percentage of particles in each particle size range relative to the total mass, and then calculate the cumulative percentage content (accumulating from the smallest particle size). Plot the particle size distribution curve with "particle size" as the x-axis (usually logarithmic scale) and "cumulative percentage content" as the y-axis.

[0049] Sure and On the gradation curve, find the particle size corresponding to a cumulative percentage content of 60%, which is the [particle size value]. Find the particle size corresponding to a cumulative percentage content of 10%, which is... Through the formula The coefficient of non-uniformity is obtained.

[0050] In summary, the characterization of TBM slag falls under the category of morphology. To quantify the information carried by the slag, the segmented slag particles are quantified using equivalent ellipsoids. The principal axis (a) and secondary axis (b) correspond to the major and minor axes of the particles, respectively, and the volume and mass are calculated. Based on the quantified particle morphology parameters from the identification results, their roundness, elongation, and other indices are calculated, and gradation curves are plotted.

[0051] Rock mass quality evaluation formula based on particle morphology characteristics: Evaluate the quality of the surrounding rock. Among these, N This represents the number of rock fragments identified in the slag. The fine particle content coefficient is taken when the particle size is less than 5 mm. Indicates the degree to which the gradation curvature coefficient deviates from the ideal value; Represents the median particle size; Represents average roundness; The coefficient of non-uniformity; The coefficient of inhomogeneity is the ideal inhomogeneity. The numerator term... The larger the value, the more rock blocks, the higher the content of fine grains, and the poorer the gradation curve shape, indicating a high degree of rock mass fragmentation; denominator term The larger the value, the larger the rock blocks produced by cutting, and the higher their roundness and uniform gradation, indicating a high degree of rock mass integrity. Therefore, the formula as a whole combines the characteristics of surrounding rock fragmentation and integrity for comprehensive judgment. RQI The closer the value is to 0, the better the quality of the surrounding rock; the closer it is to 1, the worse the quality of the surrounding rock. For example... Figure 3 .

[0052] Example 2 discloses a surrounding rock quality assessment system based on intelligent image recognition of TBM slag, used to execute the surrounding rock quality assessment method based on intelligent image recognition of TBM slag in Example 1, including: The acquisition module is used to acquire images of slag in real time. The recognition and segmentation module is used to input the slag image into the trained recognition and segmentation model to obtain pixel-level segmentation results of slag particles in the slag image. The evaluation module is used to evaluate the quality of the rock mass being excavated based on the pixel-level segmentation results and the set rock mass quality evaluation rules.

[0053] In some embodiments, the acquisition module includes a line-scan camera, an encoder, a light source, and a controller mounted on the slag conveyor belt. The light emitted by the light source is directed towards the conveyor belt. The line-scan camera is located below the light source and is used to capture images of slag particles on the conveyor belt. The encoder is used to monitor the operating status of the conveyor belt in real time and, in conjunction with the line-scan camera, determines the timing for the camera to capture images. The controller is used to control the image capture by the line-scan camera and transmit the image data captured by the line-scan camera to a terminal via a communication module. The terminal includes one or more of a cloud server, a local computer, and a mobile phone.

[0054] 1. During TBM tunneling, an image acquisition system mounted on the on-site muck conveyor belt is used to acquire real-time images of the muck material. The core components of this system include a linear array camera, encoder, light source, controller, and lens, such as... Figure 4 .

[0055] 2. The acquired images are transmitted to the computer in the TBM control room via network interface card 1, and uploaded to the internet cloud server via network interface 2. Operators can view the system's operating status both locally and remotely. Figure 4 .

[0056] 3. After receiving the image of the slag material, the computer in the control room inputs the image into a deep learning network for training and prediction, and finally outputs the recognition and segmentation results of the slag material particles in the image.

[0057] The specific implementation of this method is illustrated below with concrete examples.

[0058] (1) Install an image acquisition system above the slag discharge belt conveyor of the TBM and deploy computer equipment in the TBM control room to realize the network connection between the data acquisition end and the processing end; (2) Acquire images of slag material in real time through an image acquisition system and transmit the image data to the computer in the control room; (3) The collected images are labeled and organized to construct a training dataset, which is then input into the deep learning model for training to obtain a pre-trained model; (3) During the tunneling process, the system will input the collected images into the pre-trained model in real time and output the segmentation results of the slag particles and their morphological parameters. (4) Automatically draw particle size distribution curves based on particle parameters to provide a basis for construction status assessment and parameter optimization.

[0059] (5) According to Figure 5 The morphological features and parameters obtained from the segmentation are evaluated according to the rock mass quality evaluation formula: Evaluate the quality of the surrounding rock.

[0060] in, Figure 6 This is a graph showing the quantification results of the morphological parameters of each slag particle in the slag image. Figure 7 (a) is a schematic diagram of identification, segmentation, and morphological quantification; (b) is a schematic diagram of the gradation curve. Figure 8 This is a schematic diagram of the rock mass quality evaluation calculation results in this embodiment.

[0061] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing the quality of surrounding rock based on intelligent image recognition of TBM slag, characterized in that, include: Real-time acquisition of slag images; The slag image is input into a trained recognition and segmentation model to obtain pixel-level segmentation results of slag particles in the slag image. The quality of the rock mass being excavated is assessed based on the pixel-level segmentation results and the established rock mass quality evaluation rules.

2. The method for assessing surrounding rock quality based on intelligent image recognition of TBM slag as described in claim 1, characterized in that, The recognition and segmentation model includes a Faster R-CNN object detection sub-model and a U-Net image segmentation sub-model. The Faster R-CNN object detection sub-model is used to extract the bounding box position parameters of the slag particles in the slag image, and the U-Net image segmentation sub-model is used to perform image segmentation on the slag image based on the bounding box position parameters to obtain the pixel-level segmentation result of each slag particle in the slag image.

3. The method for assessing surrounding rock quality based on intelligent image recognition of TBM slag as described in claim 2, characterized in that, The Faster R-CNN object detection sub-model includes a feature extraction network module, an FPN module, an RPN module, and a RoI Pooling module. The feature extraction network module includes several convolutional layers with scales gradually decreasing from input to output, used to extract features from the slag image to obtain initial feature maps of different scales. The FPN module is used to obtain multi-level semantic features from each of the initial feature maps, then progressively upsamples from high-level semantic features to lower levels, and during the upsampling process, performs a 1×1 convolution with the initial feature map of the same level and adds them together to fuse high-level semantic information and low-level spatial information. Finally, a 3×3 convolutional layer is used to obtain the final multi-scale fused feature map. The RPN module is used to obtain candidate regions of the slag image based on the multi-scale fused feature map. The RoIPooling module is used to perform region-of-interest pooling on the multi-scale fused feature map using the candidate regions to finally obtain the bounding box position parameters of the slag particles.

4. The method for assessing surrounding rock quality based on intelligent image recognition of TBM slag as described in claim 3, characterized in that, The RPN module includes a 3×3 convolutional layer, two 1×1 convolutional layers, and anchor boxes. The multi-scale fused feature map is passed through the 3×3 convolutional layer and then through two 1×1 convolutional layers to obtain the target detection confidence and bounding box regression values. The target detection confidence and bounding box regression values ​​are combined with the anchor boxes to generate candidate regions.

5. The method for assessing surrounding rock quality based on intelligent image recognition of TBM slag as described in claim 3, characterized in that, The U-Net image segmentation sub-model comprises several downsampling units and several upsampling units, with the number of downsampling units and upsampling units being the same. Along the data transmission direction, the dimension of each downsampling unit is the same as the dimension of the corresponding convolutional layer in the feature extraction network module. The first downsampling unit receives an intermediate image after invalid regions of the slag image have been masked using the bounding box position parameters. The downsampling units and upsampling units are connected in a U-Net-style skip connection, directly transmitting the first feature maps of the downsampling units to the corresponding upsampling units, ultimately obtaining pixel-level segmentation results for each slag particle.

6. The method for assessing surrounding rock quality based on intelligent image recognition of TBM slag as described in claim 1, characterized in that, The specific rules for evaluating rock mass quality are as follows: The pixel-level segmentation results of the slag particles are sampled and quantized into equivalent ellipsoids to calculate the volume V and mass M of the slag particles. , a and b represent the major and minor axes of the slag particles, respectively. This indicates the density of the slag particles; The volume V and mass M of the slag particles are used to obtain particle morphology characteristics, including average roundness. Gradient curvature coefficient and non-uniformity coefficient ; The quality of the surrounding rock is evaluated using the aforementioned particle morphology characteristics and a rock mass quality evaluation formula. The rock mass quality evaluation formula... ; The mass of the rock mass is represented by N, and the number of slag particles is represented by N. Indicates the fine particle content coefficient. Indicates the median particle size. This represents the ideal non-uniformity coefficient.

7. The method for assessing surrounding rock quality based on intelligent image recognition of TBM slag as described in claim 6, characterized in that, The average roundness is obtained from the volume V and mass M of the slag particles. Gradient curvature coefficient and non-uniformity coefficient The process is as follows: The average roundness of each slag particle is calculated to obtain the average roundness in the corresponding slag image. ; Particle size distribution curves were plotted using the volume / mass ratios of different slag particles, thereby obtaining the gradation curvature coefficient. ; The non-uniformity coefficient is calculated based on the two characteristic particle sizes on the particle size distribution curve. , , To limit the particle size, The effective particle size.

8. A surrounding rock quality assessment system based on intelligent image recognition of TBM slag, characterized in that, include: The acquisition module is used to acquire images of slag in real time. The recognition and segmentation module is used to input the slag image into the trained recognition and segmentation model to obtain pixel-level segmentation results of slag particles in the slag image. The evaluation module is used to evaluate the quality of the rock mass being excavated based on the pixel-level segmentation results and the set rock mass quality evaluation rules.

9. A surrounding rock quality assessment system based on intelligent image recognition of TBM slag as described in claim 8, characterized in that, The acquisition module includes a line array camera, an encoder, a light source, and a controller mounted on the slag conveyor belt. The light emitted by the light source is directed towards the conveyor belt. The line array camera is located below the light source and is used to capture images of slag particles on the conveyor belt. The encoder is used to monitor the operating status of the conveyor belt in real time and, in conjunction with the line array camera, determines the timing for the camera to capture images. The controller is used to control the image capture by the line array camera and transmits the image data captured by the line array camera to the terminal via a communication module.

10. A surrounding rock quality assessment system based on intelligent image recognition of TBM slag as described in claim 8, characterized in that, The terminal includes one or more of the following: cloud server, local computer, and mobile phone.

Images