Intelligent detection and identification method for flow plasticity type of shield muck

Abstract

The invention relates to the technical field of intelligent construction of tunnel engineering, in particular to an intelligent detection and identification method for flow plasticity types of shield muck. The method comprises the steps that firstly, an industrial camera with environmental adaptability is used for collecting an image of slag at a slag outlet of the shield tunneling machine; then, preprocessing including Retinex illumination enhancement, geometric correction and denoising is carried out on the image; inputting the processed image into a trained YOLOv8-CBAM model for analysis and reasoning, wherein the model enhances the ability of extracting muck textures and key features by embedding a CBAM attention mechanism; the model outputs three types of muck states of dry soil, suitable soil and rare earth and confidence degrees thereof; and finally, performing real-time linkage adjustment on construction parameters such as the injection amount and the propelling speed of the modifier according to an identification result to form intelligent closed-loop control. According to the invention, the accuracy and real-time performance of muck state identification are improved, the dependence on artificial experience is reduced, and the safety, high efficiency and continuity of shield construction are effectively guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction technology for tunnel engineering, and more specifically to an intelligent detection and identification method for the fluid plasticity type of shield tunnel excavated soil. Background Technology

[0002] Against the backdrop of increasingly rapid infrastructure construction, urban land resources are becoming increasingly scarce, and underground space development has become an important means to alleviate urban land use conflicts and optimize urban spatial structure. The large-scale implementation of projects such as subways, underground roads, and integrated utility tunnels has driven the rapid development of underground construction technology. In the process of rapid advancement in urban underground space development and subway construction, the shield tunneling method has become the mainstream technology in the field of underground engineering due to its advantages such as short construction cycle, minimal surface disturbance, and wide adaptability to geological conditions. Especially in urban subway tunnel construction, the Earth Pressure Balance (EPB) shield tunneling technology has become one of the mainstream construction methods in the tunnel engineering field due to its significant advantages such as high mechanization and minimal impact on the surrounding environment. This technology maintains a certain pressure in the soil chamber in front of the cutterhead to achieve ground stability and smooth discharge of excavated soil, placing higher demands on the precise control of the excavated soil state. During the tunneling process, in order to achieve stable "in and out" of excavated soil and maintain pressure within the shield tunneling chamber, it is necessary to regulate the plasticity of the excavated soil through excavated soil improvement methods to make it exhibit an ideal plastic flow state. In existing technologies, the detection and improvement status assessment of the fluidity and plasticity of construction waste mainly rely on the following methods: First, we will conduct indoor tests on the physical properties of the improved slag soil, such as analyzing the relationship between the improvement parameters and the fluidity and plasticity of the slag soil through slump tests and mixing tests. Secondly, the morphology of the slag and soil at the slag opening is observed by manual observation, and the improvement status of the slag and soil is comprehensively evaluated in combination with the tunneling parameters of the tunnel boring machine (such as cutterhead torque, total thrust of propulsion cylinders, etc.). Third, some studies have attempted to apply machine learning techniques (such as the traditional YOLOv5 object detection algorithm) to the identification of slag mound morphology in order to improve the automation of the assessment.

[0003] However, the aforementioned existing technologies still face many technical bottlenecks in practical applications, which severely restrict the intelligent, safe, and efficient advancement of tunnel boring machine (TBM) construction. Specific technical problems are as follows: Firstly, there is a lack of clarity in the rules for soil improvement. Currently, the industry lacks a unified and scientific standard system for evaluating the effectiveness of soil improvement and controlling parameters. Most existing indoor testing studies only use the properties of the soil itself as a single reference indicator, failing to fully consider the comprehensive impact of multiple factors such as changes in geological conditions and fluctuations in equipment operating parameters during tunnel boring machine (TBM) excavation. In actual construction, the state of soil improvement is not only affected by the amount of improver injected, but also closely related to tunneling parameters such as cutterhead rotation speed and propulsion speed. However, existing technologies have not yet established a comprehensive improvement rule system covering multiple factors, resulting in a lack of clear guidance standards for on-site engineers when controlling improvement parameters. They often have to rely on personal experience for "trial and error" subjective judgment, which greatly reduces construction efficiency and creates potential safety hazards.

[0004] Secondly, there is the subjectivity and lack of objectivity in condition assessment. In current engineering practice, the assessment of the improved state of excavated soil is highly dependent on the operator's technical level and subjective work state. Operators with different experience may make different judgments on the same excavated soil form, especially in the assessment of intermediate states such as "suitable soil," where the characteristics are relatively indistinct. The difference in judgment is even more obvious. The "suitable soil" refers to excavated soil that, after improvement, has good plastic deformation capacity, is fluid to soft plastic, has low internal friction, and low permeability. In addition, the tunnel boring machine (TBM) construction environment is usually quite harsh. Taking a subway project as an example, the average sound pressure level in the TBM operator's cab is as high as 92-103 dB, far exceeding the national daytime noise emission limit of 70 dB. Working in such a high-noise environment for a long time can easily lead to operator fatigue, which in turn can cause judgment errors. At the same time, manual assessment requires continuous monitoring by dedicated personnel, which not only consumes a lot of manpower but also makes it difficult to ensure the real-time and consistency of the assessment. If the improved state of the excavated soil is not fed back in time, it can directly lead to safety and quality problems such as excavation face instability and surface settlement during TBM construction.

[0005] Third, the lag in indoor testing methods. While indoor testing methods such as slump tests, flowability tests, and consistency tests can reflect the plasticity of excavated soil to a certain extent, they are limited by changes in geological conditions and testing efficiency, making it difficult to keep up with the dynamic changes in the strata ahead in real time. For example, when the tunnel boring machine passes through different geological layers, the particle size distribution, moisture content, and other characteristics of the excavated soil will change significantly. However, indoor tests are often based on outdated geological samples, resulting in the inability to synchronize the adjustment of improvement parameters with changes in the geological conditions. Moreover, the testing process takes a certain amount of time, and the cycle from sample collection to test result output is relatively long, making it difficult to meet the real-time control requirements of tunnel boring construction. This leaves the adjustment of improvement parameters in a passive state, unable to effectively prevent adverse phenomena such as "mud cake formation" and "gushing".

[0006] Fourth, the negative impact of the construction environment on the reliability of manual assessment. In addition to high noise levels, tunnel boring machine (TBM) operations may also face adverse conditions such as high temperatures, humidity, and dust. These factors can negatively impact the physiological and psychological state of operators. During prolonged operations, operators are prone to inattention and misjudgments, especially at night or during continuous construction periods. Furthermore, observation conditions at the muck outlet may be limited by insufficient lighting and debris splashing. For example, natural light is scarce inside the tunnel, requiring manual observation to rely on auxiliary lighting. Changes in the angle and intensity of the lighting can lead to errors in identifying the shape of the debris, further reducing the accuracy of manual assessment results.

[0007] Fifth, the level of intelligence is insufficient. The existing intelligent level of slag flow plasticity detection technology is difficult to meet the automation development needs of shield tunneling construction. Although some studies have attempted to apply machine learning technology to slag morphology recognition, most of them use traditional target detection algorithms, and the detection accuracy and robustness under complex working conditions need to be improved. For example, the detection model based on YOLOv5 has a low accuracy rate in identifying "suitable soil" when dealing with images with uneven lighting and blurred slag morphology. The average mAP of the three types of slag is only 0.781, which cannot meet the actual engineering needs for high-precision detection. Moreover, such models lack targeted optimization for shield tunneling construction scenarios and are difficult to effectively cope with the complex background of dynamic changes at the slag outlet and the multi-scale morphological characteristics of slag.

[0008] Sixth, insufficient data processing capabilities and a lack of high-quality key state data. Improved state assessment methods cannot meet the needs of multi-parameter fusion analysis in shield tunneling. The improved state of the excavated soil is affected by various factors such as cutterhead torque, total thrust of the tunneling cylinders, soil chamber pressure, auger torque, and excavated soil morphology. These parameters change in real time during tunneling, generating massive amounts of data. Existing assessment methods struggle to obtain objective, real-time data on the fluid plasticity of the excavated soil. Even if multi-source tunneling parameters such as cutterhead torque and soil chamber pressure are collected, the lack of this crucial dimension prevents effective fusion analysis and hinders the uncovering of deep correlations between data.

[0009] In summary, current technologies for detecting and assessing the fluidity and plasticity of construction waste have significant shortcomings in terms of improvement rule systems, assessment objectivity, real-time performance, environmental adaptability, intelligence level, and data processing capabilities, which seriously restrict the safe and efficient advancement of tunnel boring machine (TBM) construction. Summary of the Invention

[0010] In view of this, the present invention provides an intelligent detection and identification method for the fluid plasticity type of tunnel boring machine (TBM) excavated soil. This method aims to solve problems in existing technologies such as unclear rules for excavated soil improvement, strong subjectivity in state assessment, significant lag in indoor testing, the impact of complex construction environments on the reliability of manual judgment, insufficient accuracy of intelligent identification, and a lack of multi-source data processing capabilities. This improves the accuracy, real-time performance, and objectivity of excavated soil fluid plasticity state identification, reduces reliance on manual experience, achieves intelligent closed-loop control of the excavated soil improvement process, and ensures the safe, efficient, and continuous progress of TBM construction.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent detection and identification of the flow plasticity type of tunnel boring machine excavation includes the following steps: Images of the excavated soil at the tunnel boring machine's discharge port are captured using industrial cameras; The acquired images of slag and soil are preprocessed, and the preprocessing includes at least image enhancement processing based on Retinex theory; The preprocessed image is input into the trained YOLOv8-CBAM model for analysis and reasoning; the YOLOv8-CBAM model is constructed by embedding the CBAM attention module into the YOLOv8 network structure; the CBAM attention module, through the collaborative processing of the channel attention submodule and the spatial attention submodule, enables the model to focus on the key feature regions of texture and morphology of the slag image; The YOLOv8-CBAM model identifies and outputs the fluidity and plasticity type of the slag soil; Based on the identified type of plasticity of the excavated soil, the construction parameters of the tunnel boring machine are adjusted accordingly.

[0012] In one specific implementation, the CBAM attention module is embedded in the backbone and neck network of the YOLOv8 model.

[0013] In one specific implementation scheme, the fluid plasticity type includes three categories: dry soil, suitable soil, and rare earth; the model output includes category labels and corresponding recognition confidence scores.

[0014] In a specific feasible implementation plan, the specific method for adjusting the construction parameters in linkage is as follows: when the state of the excavated soil deviates from "suitable soil", the system automatically adjusts one or more parameters among the amount of amendment injected, the propulsion speed and the speed of the screw conveyor according to preset rules: if it is dry soil, the amount of amendment injected is increased; if it is rare earth soil, the amount of amendment injected is decreased.

[0015] In one specific implementation, the preprocessing further includes geometric correction and denoising of the image.

[0016] In one specific implementation, the YOLOv8-CBAM model is deployed on an edge computing device and the model inference time is reduced through TensorRT inference acceleration technology to meet the real-time requirements of tunnel boring machine construction.

[0017] In one specific implementation scheme, the YOLOv8-CBAM model achieves multi-task recognition through a multi-label learning architecture; The multi-label learning architecture uses a multi-label loss function and sets up multiple detection heads in parallel to synchronously output the fluidity and plasticity type and other state information of the slag.

[0018] In one specific implementation, the images of the excavated soil are acquired by an industrial camera deployed at the muck outlet of the tunnel boring machine. The industrial camera is equipped with explosion-proof, moisture-proof, or shock-proof functions to adapt to the harsh environment of tunnel boring construction.

[0019] In one specific implementation scheme, the identification results of the YOLOv8-CBAM model are transmitted to the tunnel boring machine main control system via the MODBUS or OPC-UA industrial protocol to trigger the linkage adjustment of the construction parameters.

[0020] Compared with existing technologies, the real-time classification and fluid plasticity assessment of excavated soil morphology at the muck outlet in earth pressure balance shield tunneling, by introducing the CBAM attention mechanism and deeply integrating it with the YOLOv8 target detection model, combined with image preprocessing enhancement technology and edge computing deployment scheme, achieves accurate and real-time identification of three types of excavated soil states: dry soil, suitable soil, and rare earth soil. Based on the identification results, construction parameters such as the amount of amendment injected and the propulsion speed are adjusted in a coordinated manner to form an intelligent closed-loop control. This effectively improves the accuracy, real-time performance, and automation level of excavated soil state identification, reduces reliance on human experience, and enhances the system's adaptability in complex construction environments. The intelligent detection and identification method for the fluid plasticity type of shield tunneling excavated soil described in this invention has the following beneficial effects: 1. Improve the accuracy and standardization of the identification of the fluidity and plasticity of tunnel boring machine excavation.

[0021] By integrating the YOLOv8 object detection model with the CBAM attention mechanism, a robust and powerful feature extraction model for construction waste image recognition was constructed, improving the accuracy of identifying suitable soil in intermediate states, which are difficult to judge using traditional methods. Leveraging the synergistic effect of channel and spatial attention in the CBAM attention module, the model can actively focus on key areas and salient features in construction waste images, effectively weakening background interference and improving its response to subtle texture features. This system can perform refined classification of three types of construction waste states, standardizing the previously experience-based evaluation process and providing more accurate data support for adjusting construction parameters. Specifically: By deeply integrating the YOLOv8 object detection model with the CBAM attention mechanism, the model can actively focus on key areas and subtle texture features (such as loose particles in dry soil and uniform texture in suitable soil) in the soil image, effectively weakening the interference of complex backgrounds. In particular, it significantly improves the recognition accuracy for the intermediate state of "suitable soil," which is difficult to judge by traditional methods. At the same time, the recognition results are output in the form of probability vectors, quantifying the uncertainty of the state and avoiding the subjective bias of human experience judgment. This standardizes the evaluation process that originally relied on personal experience, provides accurate and reliable data support for adjusting construction parameters, and reduces the construction risks caused by misjudgment.

[0022] 2. To achieve real-time and closed-loop control of the condition assessment of excavated soil during shield tunneling.

[0023] By utilizing industrial cameras to acquire real-time images of the slag outlet and combining this with a lightweight YOLOv8-CBAM inference model deployed on edge devices (such as Jetson Xavier / TX2), the system achieves second-level identification and continuous monitoring of the slag condition of the tunnel boring machine (TBM), overcoming the bottleneck of slow response times in traditional indoor testing methods. During construction, the system can automatically adjust key parameters such as the amount of amendment injected, the speed of the screw conveyor, and the propulsion rate based on the identification results, constructing a closed-loop control logic of identification, feedback, and regulation. This enhances the TBM's responsiveness to complex geological changes and effectively reduces the probability of problems such as mud cake formation and blowouts. Specifically: Leveraging the real-time image acquisition capabilities of industrial cameras and the second-level inference capabilities of edge computing devices (the inference time for a single image can meet the needs of real-time monitoring), this system breaks through the bottleneck of traditional indoor tests that rely on lagging samples and have long response cycles, enabling continuous and dynamic monitoring of the soil condition. Based on the recognition results, it can automatically adjust key parameters such as the amount of amendment injected, the speed of the screw conveyor, and the propulsion rate, constructing a complete closed loop of "recognition-feedback-control". This enhances the shield tunneling construction's responsiveness to complex geological changes, effectively reducing the probability of adverse phenomena and safety and quality problems such as "mud cake formation", "gushing", "excavation face instability", and "surface settlement", ensuring construction continuity and operational safety. It also supports remote log recording and uploading of inference results, providing traceable and quantifiable decision-making basis for the dispatch center and expert system.

[0024] 3. Reduce reliance on manual labor and limitations of environmental adaptability, and improve the level of intelligent construction.

[0025] By constructing a deep learning-based image recognition system, the automation, intelligence, and standardization of construction waste condition assessment have been achieved. The system exhibits good consistency in assessment results, is unaffected by differences in personnel skill levels, and maintains stable operation under long-term, high-intensity working conditions. The system supports stable operation in complex construction environments such as high noise, high humidity, and low illumination. Equipped with anti-fog and explosion-proof industrial cameras and image enhancement modules (such as the Retinex supplementary lighting model), it effectively solves the problem of decreased accuracy in traditional manual identification under low visibility and severe lighting interference conditions, saving construction labor costs and reducing personnel operational risks. Specifically: By constructing a fully automated image recognition system based on deep learning, the traditional evaluation mode of manual on-site observation is completely eliminated. The evaluation results are not affected by the differences in the operator's technical level or working status, and have high consistency and stability. It can operate continuously under long-term, high-intensity working conditions. The system is equipped with industrial cameras designed to be fog-proof, explosion-proof, and moisture-proof. Combined with Retinex image enhancement modules such as illumination compensation and noise filtering, it can effectively cope with complex environments such as high noise, high humidity, low light, and debris splashing in shield tunneling construction, solving the problem of decreased accuracy of traditional manual recognition under harsh working conditions. At the same time, it supports 24 / 7 unattended operation, saving construction labor costs, reducing the operational risks of personnel in hazardous working environments, promoting the fundamental transformation of shield tunneling construction from "experience-driven" to "data-driven", and providing strong technical support for the construction of smart construction sites.

[0026] 4. Supports multi-task expansion and multi-source system integration, enhancing the breadth of system applications.

[0027] The constructed waste soil detection system possesses excellent scalability and open interface capabilities, supporting the expansion of more recognition tasks based on a multi-label learning architecture, such as parallel detection of foreign objects, material stratification, and water seepage in the waste soil. The system reserves multiple communication interfaces, enabling seamless integration with the tunnel boring machine's main control system, amendment injection system, data visualization platform, and ground monitoring and early warning system to construct a comprehensive intelligent tunnel boring machine control platform. By integrating tunneling parameters, ground data, and waste soil conditions, it achieves cross-source information joint analysis, providing more comprehensive decision support for the entire lifecycle safety management and construction optimization of the project. Specifically: The YOLOv8-CBAM backbone architecture constructed by this method supports multi-label learning. By simply adding category labels, it can expand the detection of foreign objects (such as boulders, plastics, and metal parts), water seepage status, and unmixed conditioner status in tunnel boring machine (TBM) excavation soil, enabling multi-dimensional information extraction from complex excavation soil morphology. The system reserves multiple industrial communication interfaces such as MODBUS and OPC-UA, which can seamlessly connect with TBM tunneling parameter acquisition platforms, conditioner automatic control systems, data visualization platforms, and ground monitoring and early warning systems. It integrates multi-source information such as tunneling parameters, ground data, and surface settlement monitoring to build an intelligent TBM integrated control platform. In addition, the model inference engine has a built-in dual-channel hot standby and fault switching mechanism, and the equipment is equipped with intelligent temperature control and self-testing modules to ensure continuous and stable operation under emergencies such as network disconnection and power failure, meeting the requirements of high-intensity underground operations. Its modular and open design also facilitates subsequent function upgrades and adaptation to different brands and models of TBMs, and has broad engineering promotion prospects and industrial application value.

[0028] 5. To lay a high-quality data foundation for multi-source data fusion and intelligent decision-making in shield tunneling construction.

[0029] By identifying the fluid plastic state of construction waste in high precision and real-time, core parameters of the waste waste state are provided in an objective, quantifiable, and continuous time series manner. This key state data can be spatiotemporally aligned and fused with multi-dimensional tunneling parameters such as cutterhead torque, soil chamber pressure, and propulsion speed collected in real time by the tunnel boring machine, thereby constructing a more comprehensive construction state assessment model. This solves the bottleneck of traditional methods that cannot effectively fuse data due to the lack of reliable waste waste state data, and provides data support for subsequent implementation of advanced intelligent decision-making based on data-driven construction risk prediction and parameter adaptive optimization. Attached Figure Description

[0030] 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0031] Figure 1 A flowchart of YOLOv8-CBAM for identifying construction waste is provided in an embodiment of the present invention; Figure 2 This is a flowchart of an image preprocessing and data augmentation method provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the field equipment deployment and data interaction process of a YOLOv8-CBAM model system provided in an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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 are within the scope of protection of the present invention.

[0033] Taking Earth Pressure Balance (EPB) tunnel boring machine (TBM) construction technology as an example, the state of the excavated soil at the TBM construction site is highly uncertain. Affected by factors such as geological changes, moisture content fluctuations, and the proportion of injected soil conditioner, its fluid-plastic characteristics often undergo abrupt changes, thus impacting earth pressure balance, excavation efficiency, and the stability of cutterhead torque. Traditional excavated soil condition identification mainly relies on experienced operators' judgment of the soil's color, shape, and moisture content at the excavation port, combined with adjustments based on some laboratory test results. In complex and ever-changing construction environments, this method exhibits significant lag and subjectivity, making it difficult to achieve high-frequency, high-precision continuous monitoring and feedback control, severely restricting the automation and intelligence level of TBM construction.

[0034] like Figure 1 , Figure 2 , Figure 3 As shown, the present invention provides an intelligent detection and identification method for the fluid plasticity type of shield tunneling excavation soil, which is applied to the real-time classification and fluid plasticity assessment of the morphology of excavation soil at the slag outlet in earth pressure balance shield tunneling. The present invention will be described in detail below from three dimensions: core process, key technology implementation, and system physical deployment.

[0035] By gradually introducing AI-based image recognition technology into the field of tunnel boring machine (TBM) construction, industrial cameras are used to continuously acquire images of the slag outlet, which are then analyzed and judged in real time using deep learning models. This has become a key means to improve the efficiency and standardization of slag condition assessment. Among numerous deep learning algorithms, the YOLO (YouOnlyLookOnce) series of object detection models has become a popular tool for industrial intelligent recognition due to its excellent real-time performance and detection accuracy. YOLOv8, as the latest generation model in this series, has undergone in-depth optimization in network structure, inference mechanism, and training strategy while maintaining its original high efficiency, making it particularly suitable for scenarios with high real-time requirements, such as TBM construction.

[0036] The YOLOv8 backbone network employs a C2f (CrossStage Partialand ConcatFusion) structure, which enhances the extraction capability of deep semantic information through feature cross-layer propagation and multi-scale fusion. Simultaneously, it adopts an anchor-free mechanism, abandoning the traditional dense anchor box design, reducing the parameter adjustment work caused by prior box setting, effectively simplifying the detection process, and improving the robustness of target recognition in complex images such as deformed targets and blurred boundaries. However, YOLOv8 still faces difficulties in recognizing the "suitable soil" category in shield tunneling slag images. This is because "suitable soil" typically exhibits blurred boundaries, dense fine-grained structure, and uniform brightness distribution. Traditional convolutional layers tend to overlook these key microscopic changes when extracting features, leading to missed detections or misjudgments.

[0037] To further enhance the model's discrimination ability in complex soil images, this invention introduces the CBAM (Convolutional Block Attention Module) attention mechanism and embeds it into the YOLOv8 model structure, forming the YOLOv8-CBAM model. CBAM is a lightweight convolutional attention module that guides the neural network to focus on task-related feature regions through a dual mechanism of "channel attention + spatial attention." Its core idea is to allow the network to autonomously learn "what to look at" (channel attention) and "where to look" (spatial attention), thereby efficiently capturing key target information in the image and weakening invalid background interference. The CBAM channel attention module extracts global contextual information through max pooling and average pooling, inputting it into a shared fully connected network to generate channel weight tensors, thus dynamically adjusting the contribution of each channel to the output. This module helps enhance channel features related to semantic discrimination and improves the ability to capture subtle changes in soil texture and particle details. It is used to calculate the channel attention feature map. M c ( F ), aiming to analyze the input feature map F Different channels are assigned different weights to highlight the characteristics of important channels:

[0038] Where F represents the input feature map, AvgPool(F) performs average pooling on the feature map, calculating the average value of all elements within each region to capture the global average information of the feature map. MaxPool(F) performs max pooling on the feature map, taking the maximum value within each region to capture the global maximum response information of the feature map. MLP stands for Multilayer Perceptron, a neural network structure containing multiple fully connected layers used to perform nonlinear transformations on the pooled results. f represents the computational process of the Multilayer Perceptron, i.e., performing fully connected operations, activation operations, etc., on the input.

[0039] Next, the spatial attention module performs max pooling and average pooling operations on the aforementioned feature maps in the spatial dimension, and generates a two-dimensional spatial attention map through 7×7 convolution, representing the attention level of each pixel position in the image. The channel-weighted feature map is then multiplied pixel-by-pixel with the spatial weight map to further enhance the response of key pixel regions and improve the model's perception accuracy of complex image details. Channel-aggregated features generate a spatial distribution map. M s (F)：

[0040] Conv It is a convolution operation used to extract features and adjust dimensions of the concatenated pooling results to generate a spatial attention map.

[0041] Finally, by combining channel attention and spatial attention, the final feature map is generated. F ':

[0042] in F Input feature map M c ( F By calculating the channel attention map, M s For operators, on M c ( F )· F Perform spatial attention weighting operations.

[0043] When constructing YOLOv8-CBAM, during the network's forward propagation, after the feature maps pass through the second, third, and fourth C2f modules of the backbone network, they are immediately fed into three independent CBAM attention modules for processing, respectively used to enhance feature representations at different scales. Simultaneously, before feature fusion begins in the neck network, the features from each path originating from the backbone network and about to participate in the splicing are also preprocessed using dedicated CBAM attention modules. This series of embedding operations constitutes a complete feature optimization process, ensuring that multi-scale features receive targeted enhancement from the attention mechanism at the critical stages of extraction and fusion, thereby effectively improving the model's detection accuracy for construction waste targets under similar computational complexity.

[0044] The improvement of this invention in the YOLOv8 model is specifically reflected in the embedding position of the CBAM attention module and the key layer selection strategy. In the backbone network, the CBAM attention module is precisely embedded after the second, third, and fourth C2f modules. This design is based on the complementarity of information carried by feature layers of different depths: the 4x downsampled feature map output by the second C2f module contains rich spatial details, which is beneficial for capturing small-scale targets such as the edges and textures of scattered gravel; the 8x downsampled feature map output by the third C2f module balances spatial and semantic information, which is suitable for recognizing medium-sized muck piles and vehicle parts; the 16x downsampled feature map output by the fourth C2f module has highly abstract semantic information, which is used to perceive the overall scene such as large muck trucks or stockpiles. By embedding the CBAM attention module in these three locations, key channels related to muck features can be selected hierarchically and selectively through channel attention, and spatial attention can be used to focus on the target area and suppress background interference.

[0045] Furthermore, in the neck network structure of the model, this invention applies a CBAM attention module to the feature branches involved in the fusion before each feature fusion in the PAN-FPN. This aims to perform targeted calibration and enhancement of features at each scale before feature splicing, thereby improving the quality and discriminative power of the fused feature map. This multi-layered embedding strategy constitutes a complete attention enhancement system: shallow attention ensures sensitivity to small targets, mid-layer attention optimizes the feature quality of typical targets, deep attention ensures the discrimination accuracy of large targets and the scene, and the pre-fusion attention of the neck network ensures the effectiveness of multi-scale information fusion, ultimately improving the model's multi-scale target detection performance in complex slag and wasteland scenarios.

[0046] The YOLOv8-CBAM slag and soil identification process constructed in this invention is as follows: Figure 1 As shown, at the tunnel boring machine (TBM) construction site, the precise deployment of industrial cameras is prioritized to achieve real-time data acquisition. The industrial cameras are installed at locations directly above the muck outlet with unobstructed views and complete coverage of the dynamic muck discharge. Tools such as spirit levels and angle gauges are used to ensure stable camera installation angles, with vertical deviation controlled within ±1°. Based on the lighting characteristics of the TBM construction environment, suitable supplementary lighting equipment, such as LED supplementary lights with automatic intensity adjustment, is used to automatically compensate for light in low-light conditions at night and in tunnels, ensuring uniform brightness in the acquired images.

[0047] The industrial camera is set to 1920×1080 resolution and 30fps acquisition frame rate. It is directly connected to the edge computing device through an industrial-grade network cable to build a stable image transmission link. The image acquisition stability is checked every 5 minutes to check for image frame loss and blurring caused by vibration and electromagnetic interference, so as to ensure that the real-time acquired slag images are clear and continuous, providing a high-quality data source for subsequent processing.

[0048] The acquired images first enter the image preprocessing module 12, where image processing libraries such as OpenCV are used to perform basic operations such as denoising and grayscale correction. The key focus is on Retinex illumination compensation 13, employing the MSRCR (Multi-Scale Retinex Color Recovery) algorithm to decompose the image's illumination and reflection layers. Addressing common issues in tunnel boring machine (TBM) construction such as uneven brightness and backlighting / shadowing, the illumination layer is adaptively adjusted to enhance the details of the reflection layer. For example, when the image is overexposed due to strong local light at the slag outlet, the brightness of the corresponding area in the illumination layer is reduced to restore the slag texture; when the image is darkened due to weak light deep in the tunnel, the overall illumination layer gain is increased to clearly present the color and particles of the slag. The MSRCR algorithm is used for enhancing slag images to solve common problems in TBM construction environments such as uneven illumination, backlighting / shadowing, and image darkening, thereby improving the visibility of slag texture and details, serving as the input image for the subsequent YOLOv8-CBAM model. The MSRCR algorithm requires setting several core parameters in application. The number of scales is set to three, each corresponding to a different range of illumination compensation capabilities. The Gaussian surround standard deviations for each scale are set to 15, 80, and 250, respectively. Parameter settings and key steps in the algorithm can be implemented based on existing mature MSRCR technology.

[0049] After completing illumination compensation, data augmentation processing was carried out. Based on the characteristics of tunnel boring machine (TBM) excavation images, a specific augmentation strategy was designed: simulating excavation gushing scenarios, local stretching and distortion transformations were applied to the images; color jitter was implemented to address color differences in excavation from different geological layers, randomly adjusting the image hue and saturation values ​​with an amplitude controlled within ±15%; Gaussian blur and salt-and-pepper noise addition were introduced to simulate image acquisition interference under harsh construction conditions, expanding the diversity of the dataset.

[0050] The enhanced images are batch-processed using Python scripts and divided into training and validation sets in an 8:2 ratio, providing rich and realistic data support for model training. For example, the training dataset is collected from multiple actual shield tunneling projects with different geological conditions, aiming to cover the common geological variability and complexities encountered in shield tunneling. Data sources must cover at least four typical strata: clay, sand, gravel, and composite strata. For each stratum, images of slag in three fluid plastic states—dry soil, suitable soil, and rare earth—are systematically collected to ensure the model can learn key visual features under different conditions. Industrial cameras deployed in actual applications are selected as image acquisition devices to ensure consistency in specifications; for example, explosion-proof and moisture-proof industrial cameras are used, with a uniform image resolution of 1920×1080. To reflect real-world application scenarios, the acquisition environment must cover complex conditions such as low light, high dust, direct light, and shadows within the tunnel. In the data annotation stage after image acquisition, the annotation standard is clearly defined: only the slag and waste soil areas in the image are annotated, excluding interference from irrelevant areas such as the slag outlet mechanical structure and background pipelines. Mainstream annotation tools such as LabelImg can be used, employing a "bounding box + category label" method for annotation. Each slag and waste soil area corresponds to a bounding box, and the selected area must completely cover the slag and waste soil to avoid cropping key features. The annotated data needs to undergo consistency checks, for example, by randomly selecting 10% of the annotated data for consistency testing. Ultimately, the constructed dataset consists of 1,000 to 2,000 rigorously annotated slag and waste soil images. The method of constructing this dataset ensures that the data meets the requirements for model training in terms of diversity, representativeness, and quality.

[0051] The preprocessed image is input into the YOLOv8-CBAM model 15. The model is deployed on a Jetson Xavier edge computing device, and TensorRT is used for inference acceleration to ensure that the inference time for a single image is ≤200ms. During model inference, Backbone Attention 161 focuses on the texture features of the slag soil through channel and spatial attention mechanisms, such as the loose particles of dry soil and the uniform texture of suitable soil, to enhance the extraction of texture features 171; Neck Attention 162 performs weighted fusion of multi-scale feature maps to highlight the slag soil accumulation morphology, boundaries, etc., to focus on key areas 172. After multi-scale feature fusion 18, the model identifies the state of the slag soil 19 and outputs the identification results of three states: dry soil, suitable soil, and rare earth, while also providing the identification confidence level (accuracy ≥95%).

[0052] The system determines the operation based on whether the identified state deviates from 120. If the state deviates from "suitable soil" (e.g., dry soil identification confidence > 60%), it immediately links the construction parameter adjustment module to adjust construction parameters 1211: according to preset rules, under dry soil conditions, the amount of amendment injected is increased by 10%-20%, the propulsion speed is reduced by 5%, and the screw conveyor speed is increased by 3%; if the state is normal, it maintains the current parameters 1212. The adjustment command is transmitted to the shield tunneling main control system via the MODBUS-RTU protocol through the feedback construction system 122 operation, realizing a "identification-control-feedback" closed loop. A soil condition identification report is generated every 10 minutes, recording the changes in soil condition and construction efficiency data before and after parameter adjustment for subsequent process optimization.

[0053] Considering the complex and variable image acquisition conditions in tunnel boring machine (TBM) construction environments, images often suffer from blurriness, noise, and uneven brightness, severely impacting model inference quality. Therefore, this invention also designs a complete image preprocessing and enhancement workflow, including image denoising, histogram equalization, color normalization, and high dynamic range compression. Specifically, addressing common issues in underground construction such as uneven lighting and moisture shading, the Retinex image enhancement model is introduced to perform lighting compensation, resulting in more uniform brightness distribution and more prominent texture details. In the image enhancement stage, the system performs data enhancement operations such as rotation, blurring, color difference perturbation, and simulated jetting to construct various typical on-site image scenarios, improving the model's generalization ability and anti-interference capability.

[0054] The image preprocessing and data augmentation method constructed in this invention is as follows: Figure 2 As shown, before the tunnel boring machine (TBM) construction begins, the image acquisition equipment and preprocessing system are debugged. An industrial camera is installed 0.5m above the slag outlet, covering the slag area at a 45° angle, and equipped with a protective casing to resist splashing slag and high humidity erosion. The industrial camera is set to a resolution of 1280×720 and a frame rate of 25fps. A wireless transmission link is established through a 5G industrial router to ensure the stable uploading of the raw images acquired by the original image acquisition unit to the preprocessing server.

[0055] The image enters the parallel preprocessing stage. The geometric correction module 221 uses OpenCV's warpPerspective function to correct image geometric deformations (such as stretching and twisting of the slag shape) caused by installation offset and shield vibration, based on the industrial camera calibration parameters. First, standard checkerboard images are acquired, the intrinsic and extrinsic parameters of the industrial camera are calibrated, a distortion correction model is established, and each acquired image is corrected in real time to ensure that the image alignment error is ≤2 pixels.

[0056] The image enhancement module 222 operates synchronously, improving image contrast through histogram equalization and expanding the grayscale dynamic range to address the issue of concentrated grayscale in shield tunnel excavation images. It also uses gamma correction to optimize image brightness based on the characteristics of excavation images from different geological layers (e.g., clay layer images are darker, so γ=1.2 is set; sand layer images are brighter, so γ=0.8 is set). Combined with a brightness adjustment algorithm, it automatically identifies bright and dark areas of the image, increasing the brightness of dark areas and decreasing the brightness of bright areas, making the details of the excavation clearer.

[0057] The noise filtering module 223 uses median filtering to process salt-and-pepper noise (window size 3×3) and bilateral filtering to process Gaussian noise (sigmaColor=50, sigmaSpace=10). The preprocessed image 23 is then used as the input of the YOLOv8-CBAM model 24.

[0058] The model is loaded onto a GPU server and utilizes multi-threaded inference technology, supporting the processing of ≥10 frames per second. It automatically identifies the state of the construction waste (25) and outputs information such as state category, recognition confidence, and feature heatmap. The recognition results trigger the linkage adjustment of parameters (26). The system has a built-in fuzzy control algorithm that dynamically adjusts parameters such as the frequency of the amendment injection pump, the pressure of the propulsion cylinder, and the speed of the screw conveyor based on the degree of deviation between the state of the construction waste and the "suitable soil".

[0059] Adjustment commands are transmitted to field equipment via the OPC-UA protocol, with a response time of ≤1s. The intelligent decision output 27 includes multi-level content: generating a real-time soil condition report and pushing it to the construction monitoring screen; establishing a historical database to store preprocessing parameters, recognition results, and control commands for each frame of image, for retrospective analysis of construction conditions; when 5 consecutive frames of recognition results are abnormal (such as persistent dry soil condition), an audible and visual alarm is triggered, and the on-site inspection robot is linked for verification, forming a complete "collection-processing-decision-feedback" process to ensure accurate and efficient soil improvement during shield tunneling construction.

[0060] Based on the above methods, the YOLOv8-CBAM model can effectively identify three typical states of tunnel boring machine (TBM) excavated soil: dry soil, suitable soil, and rare earth soil. Among these, "suitable soil," as the most ideal excavated soil state, has direct guiding significance for setting construction parameters. The system can not only output category labels but also provide identification confidence distributions in the form of probability vectors, enabling fault-tolerant judgments of state transition zones (such as the dry-rare soil boundary). During construction, if the identification results show that the excavated soil state deviates from the "suitable soil" range, the system can use a feedback mechanism to adjust key parameters such as the amount of amendment injected, the propulsion speed, and the screw conveyor speed, achieving real-time closed-loop control.

[0061] In terms of system deployment, this invention supports the deployment of the YOLOv8-CBAM model on edge computing devices such as JetsonXavier and TX2. It features low power consumption, high computing power, and high integration, making it suitable for long-term stable operation in the confined space and high humidity / temperature environments of tunnel boring machines (TBMs). Industrial cameras are installed at the TBM's slag outlet to capture images and transmit them in real-time to the inference module. The system outputs the recognition results to the TBM's main control system via industrial protocols such as MODBUS and OPC-UA, enabling data linkage. Furthermore, the system has functions such as log recording, remote debugging, and local caching, providing detailed records and rapid responses to abnormal situations.

[0062] In terms of interface integration technology, the system adopts a "middleware adaptation layer + protocol conversion gateway" architecture to achieve seamless multi-platform integration. The middleware layer builds a data hub based on Kafka, standardizing slag identification data (including timestamps, target categories, confidence levels, etc.) through JSON Schema, while deploying an OPCUA server (such as KEPServerEX) to achieve variable mapping with the additive automatic control system. At the hardware level, an RS485 / Profinet module is integrated to read tunnel boring machine parameters via Modbus RTU; at the software level, an EtherCAT master station module (such as Beckhoff CX5130) is built-in, supporting 50ms-level real-time communication with the PLC, and providing RESTful API and MQTT protocol interfaces (including X.509 certificate authentication) to meet the remote access requirements of the dispatch center.

[0063] To support real-time control and intelligent platform expansion, the system is designed with a two-way interactive link and an open interface architecture. In the real-time control phase, 10Hz updated slag particle size data is pushed to the amendment system via Linux shared memory or TCPSocket (port 9000), and adjustment commands (protobuf format) from the screw conveyor are received. Remote monitoring uploads identification results to the dispatch center via an encrypted tunnel (SSL / TLS), supporting quality analysis based on time-series databases (such as InfluxDB). Standardized APIs are reserved at the expansion level. A data fusion engine (based on Apache Flink) enables spatiotemporal correlation analysis of multi-source data, including grounding layer information (GIS format), surface subsidence monitoring (BeiDou RTK data), and advanced geological early warning systems, providing underlying technical support for the intelligent shield tunneling integrated control platform.

[0064] The field device deployment process of the YOLOv8-CBAM model system constructed in this invention is as follows: 31 Figure 3As shown, the on-site equipment deployment includes: an image acquisition system 321, which uses a tripod to fix an industrial camera 1.2m above the tunnel boring machine's slag outlet, and an LED supplementary lighting module 331. The supplementary lighting and the industrial camera are powered and controlled synchronously. Through the DMX512 protocol, the supplementary lighting brightness is automatically adjusted according to the ambient light intensity (light intensity range 500-5000 lux) to solve the problems of insufficient lighting and uneven brightness in the tunnel.

[0065] The industrial camera is equipped with a 1920×1080 resolution sensor 341 and a 30fps acquisition frequency 351. It is directly connected to the industrial control computer of the edge computing subsystem 322 via the image input inference module 36 through the gigabit network port to build a low-latency image transmission channel. The image transmission packet loss rate is tested every hour to ensure that it is ≤1%.

[0066] The edge computing subsystem 322 uses the TX2 module 332 and is deployed in an industrial control chassis 342 with a constant temperature design. The chassis has a built-in temperature sensor and cooling fan to control the operating temperature between 20-35℃.

[0067] A YOLOv8-CBAM acceleration engine was built on the TX237, and TensorRT was used to quantize and prune the model, improving inference speed by 3 times, with single-frame image inference time ≤200ms. Simultaneously, a local storage module38 was configured, using a 1TB SSD, to store image data and recognition results according to the naming rule of "date + construction section". Historical data upload management is automatically performed every 24 hours39, synchronizing historical data to the management platform and ensuring data security through encrypted transmission via FTP protocol.

[0068] The data transmission subsystem 323, relying on the industrial-grade communication link 333, constructs a three-tiered network of "edge computing - tunnel boring machine main control - cloud management". It uses the PROFINET protocol to achieve real-time data interaction between the edge computing and traffic data feedback system 352, with a transmission cycle of ≤200ms, synchronizing the results of soil identification and equipment status parameters (such as industrial camera temperature and supplementary lighting brightness) in real time. Through a 5G VPN channel, it uploads aggregated data to the cloud management platform at 15-minute intervals for multi-site construction data comparison and analysis, and process optimization decisions, constructing a data interaction system of "on-site acquisition - edge processing - cloud decision-making" to support intelligent management and control of tunnel boring machine construction.

[0069] The specific technical solution for implementing multi-label learning on the YOLOv8-CBAM backbone architecture is as follows: An improved loss function design is adopted, replacing the original single-label classification loss with a multi-label binary cross-entropy loss. By assigning an independent label vector to each target (e.g., [slag, boulders, plastic, high water content, uneven amendment]), the model can simultaneously output the predicted probabilities of multiple non-mutually exclusive categories. In terms of network structure, multiple task heads are connected in parallel after the feature map output by the backbone: the foreign object detection head identifies targets such as boulders and plastics by adding a 1×1 convolutional layer (the number of output channels corresponds to the number of foreign object categories); the state analysis head uses a spatial attention pooling layer to extract global statistics from the feature map, used to determine the water seepage state and the uniformity of amendment mixing. During the training phase, a multi-task joint optimization strategy is adopted, and the gradient contribution of different learning tasks is balanced by dynamic weight coefficients (dynamically adjusted based on the validation loss of each task). During the inference phase, CUDA streaming parallel computing is used to compute the output of each task head, ensuring that only the output layer parameters need to be fine-tuned when adding label categories (without modifying the backbone network), realizing lightweight deployment of task expansion. Ultimately, the model can maintain real-time performance (inference speed decrease ≤10%) while simultaneously outputting the main body of the slag and related state information, providing multi-dimensional data support for construction safety decisions.

[0070] In terms of the technical implementation of model inference layer, the system security adopts a dual-channel hot standby architecture: the primary and backup inference servers are kept synchronized through a heartbeat detection mechanism (3 health message exchanges per second). The primary server backs up the inference results and running status to the backup server in real time. When the network disconnection of the primary server is detected (3 consecutive message losses) or power abnormality (voltage fluctuations exceeding ±15%), a seamless switch is triggered within 0.5 seconds. The backup server immediately takes over the inference task. At the same time, an independent UPS power supply ensures at least 30 minutes of power supply after a power outage, ensuring that the detection service is not interrupted in case of emergencies.

[0071] In terms of equipment safety, the camera adopts an explosion-proof housing design (compliant with GB3836.1-2010 standard), which can withstand internal explosion pressure and prevent flame leakage. The lens surface is covered with a fluororubber sealing ring to achieve IP68-level moisture resistance. The body has passed a 3-axis vibration test (10-2000Hz frequency range, 10g acceleration) to meet the requirements of underground bumpy environments. The system has a built-in intelligent temperature control module, which monitors the operating temperature of the inference unit and camera in real time through a temperature sensor (sampling accuracy ±0.5℃). When the temperature exceeds 60℃, the cooling fan is automatically activated (air speed adjustable up to 3000rpm), and an audible and visual alarm is triggered. The self-test module performs hardware diagnostics every 5 minutes (including 20 indicators such as camera focal length, network bandwidth, and storage capacity). Abnormal data is pushed to the maintenance terminal through an encrypted protocol, enabling early detection and timely intervention of faults and ensuring long-term stable operation of the system.

[0072] This invention, based on YOLOv8-CBAM, integrates multiple key technologies including image enhancement, attention mechanisms, edge computing, and industrial control to construct an integrated system for the identification and feedback control of the fluid plasticity of tunnel boring machine (TBM) excavated soil. This system not only improves the real-time performance and accuracy of excavated soil condition identification but also provides a solid technical foundation for intelligent control of TBM construction, demonstrating significant engineering application value and promising prospects for wider adoption. In the future, with further enhancements to model capabilities and the integration of multi-source data, this system is expected to play a role in a broader range of underground engineering fields, propelling urban underground space development towards a higher level of intelligence and digitalization.

[0073] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent detection and identification of the flow plasticity type of tunnel boring machine excavated soil, characterized in that, Includes the following steps: Images of the excavated soil at the tunnel boring machine's discharge port are captured using industrial cameras; The acquired images of slag and soil are preprocessed, and the preprocessing includes at least image enhancement processing based on Retinex theory; The preprocessed image is input into the trained YOLOv8-CBAM model for analysis and reasoning; the YOLOv8-CBAM model is constructed by embedding the CBAM attention module into the YOLOv8 network structure; the CBAM attention module, through the collaborative processing of the channel attention submodule and the spatial attention submodule, enables the model to focus on the key feature regions of texture and morphology of the slag image; The YOLOv8-CBAM model identifies and outputs the fluidity and plasticity type of the slag soil; Based on the identified type of plasticity of the excavated soil, the construction parameters of the tunnel boring machine are adjusted accordingly.

2. The intelligent detection and identification method for the plasticity type of shield tunnel excavation soil according to claim 1, characterized in that, The CBAM attention module is embedded in the backbone and neck network of the YOLOv8 model.

3. The intelligent detection and identification method for the flow plasticity type of shield tunnel excavated soil according to claim 1, characterized in that, The fluid plasticity types include three categories: dry soil, suitable soil, and rare earth; the model output includes category labels and corresponding recognition confidence scores.

4. The intelligent detection and identification method for the flow plasticity type of shield tunnel excavated soil according to claim 3, characterized in that, The specific method for adjusting construction parameters in linkage is as follows: when the state of the slag soil deviates from "suitable soil", the system automatically adjusts one or more parameters of the amount of amendment injected, the propulsion speed and the speed of the screw conveyor according to preset rules: if it is dry soil, the amount of amendment injected is increased; if it is rare earth soil, the amount of amendment injected is decreased.

5. The intelligent detection and identification method for the plasticity type of shield tunnel excavation soil according to claim 1, characterized in that, The preprocessing also includes geometric correction and noise reduction of the image.

6. The intelligent detection and identification method for the flow plasticity type of shield tunnel excavated soil according to claim 1, characterized in that, The YOLOv8-CBAM model is deployed on an edge computing device and uses TensorRT inference acceleration technology to reduce model inference time in order to meet the real-time requirements of tunnel boring machine construction.

7. The intelligent detection and identification method for the plasticity type of shield tunnel excavation soil according to claim 1, characterized in that, The YOLOv8-CBAM model achieves multi-task recognition through a multi-label learning architecture; The multi-label learning architecture uses a multi-label loss function and sets up multiple detection heads in parallel to synchronously output the fluidity and plasticity type and other state information of the slag.

8. The intelligent detection and identification method for the flow plasticity type of shield tunnel excavated soil according to claim 1, characterized in that, The images of the excavated soil were captured by an industrial camera deployed at the muck outlet of the tunnel boring machine. The industrial camera is equipped with explosion-proof, moisture-proof, or shockproof functions to adapt to the harsh environment of tunnel boring construction.

9. The intelligent detection and identification method for the plasticity type of shield tunnel excavation soil according to claim 1, characterized in that, The recognition results of the YOLOv8-CBAM model are transmitted to the tunnel boring machine main control system via the MODBUS or OPC-UA industrial protocol to trigger the linkage adjustment of the construction parameters.

Images