A TBM (Traffic Machine) Prediction Method Based on Intelligent Hierarchical Evaluation of Multi-Source Data
By constructing a hierarchical and decoupled intelligent assessment network, multi-source data is transformed into explicit engineering physical indices, solving the problems of single data source and lack of foresight in TBM malfunction risk warning, and realizing advanced identification and interpretable warning of malfunction risks.
Patent Information
- Application Number
- CN202611205987.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing TBM (Turbine Machine Tool) risk warning methods suffer from problems such as a single data source, lack of forward-looking warnings, and lack of interpretability of outputs. They cannot effectively correlate with the surrounding rock condition, rock fragmentation products, and abnormal mechanical response, making it difficult to guide targeted on-site handling.
A hierarchical and decoupled intelligent evaluation network is constructed. Through a multi-source data intelligent hierarchical evaluation method, the tunnel face image, rock debris image and tunneling parameters are transformed into stability index, deterioration index and anomaly index. A multi-head self-attention fusion network is used for comprehensive evaluation to generate a clear probability of jamming and remaining distance.
It enables proactive identification of card-related risks, generates complete early warning information including primary and secondary causes, and possesses engineering verifiability and operational guidance, significantly improving prediction accuracy and on-site decision-making efficiency.
Smart Images

Figure CN123117769A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of tunnel boring machine (TBM) construction safety early warning technology, specifically to a method for intelligent evaluation driven by physical mechanisms of tunnel face images, rock debris images, and tunneling parameters, and then for TBM risk classification prediction based on the fusion of the three results using an attention mechanism. Background Technology
[0002] Full-face tunnel boring machines (TBMs) have become core equipment for deep-buried long tunnel construction due to their advantages such as high efficiency, safety, and environmental friendliness. However, when traversing complex geological conditions such as fault fracture zones, soft rock with large deformation, and high ground stress, TBMs frequently experience cutterhead or shield jamming accidents, leading to project delays, increased costs, and even safety incidents. Therefore, accurately and proactively predicting jamming risks is of great significance for ensuring construction safety and progress.
[0003] Currently, TBM jamming risk early warning methods are mainly divided into three categories. The first category is numerical simulation methods, which establish rock-machine interaction models to analyze mechanical processes such as surrounding rock deformation and shield friction, thereby predicting the likelihood of jamming. This method is computationally complex, difficult to obtain parameters, and struggles to respond in real-time to dynamic changes during construction. The second category is machine learning methods based on monitoring data. This method utilizes parameters such as thrust, torque, speed, and penetration depth collected in real-time during TBM tunneling, combined with classification or regression models to determine jamming risk. This method relies on only a single tunneling parameter, resulting in incomplete information sources and low early warning accuracy. The third category is qualitative assessment methods based on expert experience and historical cases. This method summarizes past jamming cases to form empirical rules to guide on-site operations. The knowledge expressed in this type of method is often unstructured, highly subjective, and difficult to adapt to complex and changing geological conditions.
[0004] Furthermore, existing methods generally suffer from three common problems: a single data source, failing to effectively link the causal chain between "the surrounding rock condition, rock fragmentation products, and abnormal mechanical response"; a lack of proactive warning capabilities, with most methods only triggering alarms when parameters exceed limits or obvious image anomalies appear; and a lack of interpretability in the output, failing to explain whether the warning is primarily driven by surrounding rock instability or abnormal tunneling parameters, making it difficult to guide targeted on-site measures. While some existing technologies integrate image and parameter data, their integration methods are mostly early direct stitching or later fractional fusion, lacking physical anchors in the intermediate process, and the output requires further verification.
[0005] Therefore, there is an urgent need to develop an intelligent prediction method for TBM card machine risks that can integrate multi-source data, is interpretable, and enables real-time hierarchical early warning, in order to make up for the shortcomings of existing technologies. Summary of the Invention
[0006] To address the aforementioned issues, this application provides a TBM (Tunnel Boring Machine) machine prediction method based on intelligent hierarchical evaluation of multi-source data. By constructing a hierarchical and decoupled intelligent evaluation network, multi-source heterogeneous data such as tunnel face images, rock debris images, and tunneling parameters are transformed into a unified risk index. Then, a multi-head self-attention fusion network is used for comprehensive evaluation, significantly improving the accuracy, foresight, and interpretability of TBM machine prediction.
[0007] The first aspect of this invention provides a TBM (Traffic Machine Module) prediction method based on intelligent hierarchical evaluation of multi-source data, comprising: The process involves acquiring and preprocessing images of the tunnel face, rock debris, and tunneling parameter sequences to obtain preprocessed images of the tunnel face, rock debris, and tunneling parameter sequences. The preprocessed images of the tunnel face are then fed into an intelligent assessment network for surrounding rock structures to obtain a tunnel face stability index and a tunnel face information fusion feature vector. Similarly, the preprocessed rock debris images are fed into an intelligent assessment network for rock debris features to obtain a rock debris deterioration index and a rock debris information fusion feature vector. The preprocessed tunneling parameter sequences are then fed into an intelligent assessment network for tunneling status to obtain a tunneling anomaly index and a tunneling status information fusion feature vector. Finally, the tunnel face stability index and the tunnel face information fusion feature vector are concatenated and projected onto a specified dimension to obtain a first projection vector. The rock debris deterioration index... The first projection vector is obtained by concatenating the feature vector of the data with the rock cutting information and projecting it onto a specified dimension. The second projection vector is obtained by concatenating the feature vector of the tunneling anomaly index with the feature vector of the tunneling status information and projecting it onto a specified dimension. The first projection vector, the second projection vector, and the third projection vector are fed into an encoder module based on a multi-head self-attention mechanism to generate a fusion discrimination feature vector. The fusion discrimination feature vector is input into a multi-task prediction output layer to obtain the predicted jam probability and the remaining distance of the jam. Based on the jam probability and combined with the face stability index, the rock cutting deterioration index, and the tunneling anomaly index, a jam probability graded early warning is performed, and early warning information is obtained and the contribution of the three indices to the jam probability is displayed. The actual causes of the jam are analyzed through the above data.
[0008] Furthermore, the intelligent assessment network for surrounding rock structures includes: Semantic segmentation sub-network: performs pixel-level segmentation of the joints and fragmented areas of the face image to obtain face feature vectors and face image segmentation mask; Physical index calculation module: calculates face physical indexes based on the face image segmentation mask; Multilayer perceptron: fuses the face feature vectors and face physical indexes to output face stability index.
[0009] Furthermore, the physical parameters of the tunnel face include joint density, bunker fragmentation index, and integrity coefficient.
[0010] Furthermore, the intelligent assessment network for rock debris characteristics includes: A multi-task convolutional neural network extracts features from the rock debris image to obtain a rock debris image feature image; a particle size regression task head acquires the particle size distribution features of the rock debris image feature image to obtain rock debris particle size information; a shape classification task head classifies the rock debris image feature image to obtain rock debris shape information; a texture feature task head acquires the texture features of the rock debris image feature image to obtain rock debris texture feature information; and a multilayer perceptron fuses the feature information obtained from the three task heads to output a rock debris degradation index.
[0011] Furthermore, the rock fragment particle size information is the effective particle size d. 10 Median particle size d 30 , limited particle size d 60 Inhomogeneity coefficient and curvature coefficient The shape information is block-shaped, sheet-shaped, or needle-shaped.
[0012] Furthermore, the intelligent assessment network for tunneling status includes: A temporal prediction model using long short-term memory networks and attention mechanisms: Predicting tunneling parameters using the tunneling parameter sequence; a feature extraction module extracting features from the tunneling parameter sequence to obtain corresponding composite index features; and a multilayer perceptron fusing the predicted tunneling parameters and the composite index features to output a tunneling anomaly index.
[0013] Furthermore, the composite index is characterized by torque fluctuation index, penetration coefficient of variation, and specific energy change rate.
[0014] Furthermore, the predicted card machine probability loss function is Focal Loss, the remaining card machine distance loss function is Huber Loss, and the total loss is the weighted sum of the two. The Focal Loss formula is:
[0015] in The Focal Loss value. This represents the model's predicted probability for the true class. For focusing parameters; The Huber Loss formula is:
[0016] in The Huber Loss value. This represents the actual remaining distance of the card reader. The predicted remaining distance for the SIM card. This is the acceptable error threshold.
[0017] The total loss function is:
[0018]
[0019] in and These are the coefficients of the two loss functions.
[0020] Furthermore, in the step of feeding the first projection vector, the second projection vector, and the third projection vector into the encoder module based on a multi-head self-attention mechanism to generate a fused discriminative feature vector, The encoder module of the multi-head self-attention mechanism is used to concatenate the first projection vector, the second projection vector and the third projection vector into three head features and corresponding feature vectors. Each head feature and its corresponding feature vector are scaled and dot-producted, then normalized by Softmax, and used as weights for weighted summation to obtain the weighted vector corresponding to each feature head. The weighted vector corresponding to each feature head is used to generate a fused discriminative feature vector through multi-head concatenation and linear projection.
[0021] The Normalization is also configured to derive the percentage contribution of the face stability index, the rock debris deterioration index, and the tunneling anomaly index to the jamming probability, using the following formula:
[0022] in The percentage contribution of the three indices to mode m. Let m be the attention weight of mode m in the k-th attention head. The stability index of the tunnel face. The rock debris deterioration index. This is the tunneling anomaly index.
[0023] Furthermore, when a jamming event occurs, it is also configured to provide an online incremental learning process, which extracts the face image, rock debris image, and tunneling parameter sequence data at the time of the jamming event, and uses a hierarchical fine-tuning strategy to fine-tune the encoder module and the parameters within the multi-task prediction output layer of the multi-head self-attention mechanism. The hierarchical fine-tuning strategy involves freezing the parameters of the surrounding rock structure intelligent evaluation network, the rock debris feature intelligent evaluation network, and the tunneling status intelligent evaluation network; performing small-batch gradient updates on the encoder module of the multi-head self-attention mechanism and the multi-task prediction output layer; and mixing new event data and historical sampled data in a specified ratio for training to prevent catastrophic forgetting.
[0024] This application proposes a TBM (Traffic Machine Controller) prediction method based on intelligent hierarchical evaluation of multi-source data. Specifically, its beneficial effects are as follows: 1. By constructing a hierarchical and decoupled intelligent assessment network, three types of multi-source heterogeneous data—face images, rock debris images, and tunneling parameters—are transformed into intermediate indices with clear engineering physical meaning: the face stability index, the rock debris deterioration index, and the tunneling anomaly index. Each of these three indices corresponds to an independent link in the TBM jamming accident evolution chain. They can be independently output, independently verified, and independently guide engineering practice before fusion prediction. This fundamentally overcomes the shortcomings of existing end-to-end black box models, which lack traceability and interpretability of intermediate processes, making the early warning results engineering verifiable and operationally instructive.
[0025] 2. The system can quantify the contribution of the three indices to the final probability of system malfunction as a percentage, generating complete early warning information including both primary and secondary causes. Furthermore, even when a single index is abnormal but the probability of system malfunction is still low, the system still outputs an independent early warning, ensuring complete interpretability and precise guidance for handling the situation, significantly improving on-site decision-making efficiency and resource utilization.
[0026] 3. It boasts strong early warning capabilities, combining the forward-looking nature of images with the trend prediction ability of parameters to achieve proactive identification of jamming risks, rather than simply issuing alarms when a jam occurs, thus allowing ample time for on-site response. It ensures high prediction accuracy and has strong generalization ability. Employing multimodal fusion based on multi-head self-attention, it can automatically learn the optimal weights of various data sources under different geological conditions. Compared to single-data-source methods, its fusion prediction accuracy is significantly improved, and it is more adaptable to different geological conditions. Attached Figure Description
[0027] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0028] Figure 1 This is a flowchart of a TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data proposed in an embodiment of this application; Figure 2 This is a flowchart of an intelligent evaluation network for surrounding rock structures proposed in one embodiment of this application; Figure 3 This is a flowchart of an intelligent assessment network for rock debris characteristics proposed in one embodiment of this application; Figure 4 This is a flowchart of an intelligent assessment network for tunneling status proposed in one embodiment of this application; Figure 5 This is a flowchart of the encoder module and task prediction output layer of the multi-head self-attention mechanism proposed in an embodiment of this application. Detailed Implementation
[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0030] Existing methods generally suffer from three common problems: first, a single data source, failing to effectively link the causal chain between "the surrounding rock condition, rock fragmentation products, and abnormal mechanical response"; second, a lack of proactive warning capabilities, with most methods triggering alarms only when parameters exceed limits or obvious image anomalies appear; and third, a lack of interpretability in the output, failing to explain whether the warning is primarily driven by surrounding rock instability or abnormal tunneling parameters, making it difficult to guide targeted on-site measures. While some existing technologies integrate image and parameter data, their integration methods are mostly early direct stitching or later fractional fusion, lacking physical anchors in the intermediate process, and the output requires further verification.
[0031] In view of this, this application provides a TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data to alleviate the problems of single data source, lack of strong early warning and lack of interpretability of output in the prior art.
[0032] Please refer to Figure 1 , Figure 1 This is a flowchart of a TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data, proposed in an embodiment of this application. Figure 1 As shown, a TBM (Toyota Machine Machine) prediction method based on intelligent hierarchical evaluation of multi-source data includes: S1. Acquire the tunnel face image, rock debris image, and tunneling parameter sequence data and perform data preprocessing to obtain the preprocessed tunnel face image, preprocessed rock debris image, and preprocessed tunneling parameter sequence data; S2. The preprocessed tunnel face image is sent to the intelligent evaluation network for surrounding rock structure to obtain the tunnel face stability index and the tunnel face information fusion feature vector; the preprocessed rock debris image is sent to the intelligent evaluation network for rock debris features to obtain the rock debris deterioration index and the rock debris information fusion feature vector; the preprocessed tunneling parameter sequence data is sent to the intelligent evaluation network for tunneling status to obtain the tunneling anomaly index and the tunneling status information fusion feature vector. S3. The face stability index is concatenated with the face information fusion feature vector and projected onto a specified dimension to obtain a first projection vector; the rock debris deterioration index is concatenated with the rock debris information fusion feature vector and projected onto a specified dimension to obtain a second projection vector; the tunneling anomaly index is concatenated with the tunneling state information fusion feature vector and projected onto a specified dimension to obtain a third projection vector; the first projection vector, the second projection vector, and the third projection vector are fed into an encoder module based on a multi-head self-attention mechanism to generate a fusion discrimination feature vector. S4: Input the fused discriminative feature vector into the multi-task prediction output layer to obtain the predicted machine malfunction probability and the remaining distance to the machine malfunction; S5: Based on the jamming probability and combined with the working face stability index, the rock debris deterioration index and the tunneling anomaly index, perform a graded early warning of the jamming probability, obtain early warning information and display the contribution of the three indices to the jamming probability P, and analyze the actual causes of the jamming through the above data.
[0033] This invention presents a TBM jam prediction method based on intelligent hierarchical evaluation of multi-source data. By introducing face images, rock debris images, and tunneling parameter sequence data, it fully considers the lack of multi-source data analysis of TBM jams in existing technologies and transforms it into intermediate indices with clear engineering physical meaning. This fundamentally overcomes the defects of existing end-to-end black box models, such as the untraceability and uninterpretability of intermediate processes, and makes the early warning results engineering verifiable and operationally instructive. It can quantify the contribution of the three indices to the final jam probability in percentage form and generate complete early warning information including "main cause + auxiliary cause".
[0034] As one feasible approach, the face image acquisition method involves installing an explosion-proof wide-angle industrial camera behind the TBM shield to automatically capture images during each tunneling cycle or low-dust periods. The images are then corrected based on laser target points, and a dark channel prior algorithm is used to remove dust and water mist. The rock debris image acquisition method involves installing a linear array industrial camera and an LED strip light source above the conveyor belt's slag outlet to continuously acquire rock debris images. Histogram equalization is used to enhance contrast, and the rock debris area is segmented. The tunneling parameter sequence data acquisition method involves high-frequency, real-time reading of tunneling parameters such as thrust, cutterhead torque, tunneling speed, cutterhead rotation speed, and penetration sequence from the TBM's PLC system via industrial Ethernet. The parameters are then subjected to 5-point median filtering to remove burrs, and linear interpolation is used for short-term missing values to establish data quality indicators and eliminate sensor faulty segments.
[0035] It should be noted that the method for acquiring the face image, rock debris image, and tunneling parameter sequence data in this embodiment is only one feasible method. Those skilled in the art can choose the acquisition method for the face image, rock debris image, and tunneling parameter sequence data according to the actual situation.
[0036] Please refer to Figure 2 , Figure 2 This is a flowchart of an intelligent assessment network for surrounding rock structures proposed in one embodiment of this application. Figure 2 As shown, the intelligent assessment network for surrounding rock structures includes: Semantic segmentation sub-network: performs pixel-level segmentation of the joints and fragmented areas of the face image to obtain face feature vectors and face image segmentation mask; Physical index calculation module: calculates face physical indexes based on the face image segmentation mask; Multilayer perceptron: fuses the face feature vectors and face physical indexes to output face stability index.
[0037] Furthermore, the physical parameters of the tunnel face include joint density, bunker fragmentation index, and integrity coefficient.
[0038] As an implementable approach, the semantic segmentation sub-network uses the DeepLabV3+ architecture. The pixel-level segmentation includes joint traces, broken and loose areas, intact rock blocks, and seepage areas. Using these segmentation elements, a 1024-dimensional face feature vector and a face image segmentation mask are obtained. Using the face image segmentation mask, three face physical indices are obtained: joint density, shelter fragmentation index, and integrity coefficient. The meanings of these three physical indices are as follows: Joint density The number of joints per unit volume (joints / m³) reflects the degree of cutting of the rock mass; Rock fragmentation index (RBI): The ratio of the area of the fractured zone to the total area of the tunnel face, reflecting the overall degree of fragmentation of the tunnel face; Integrity coefficient The value is estimated in accordance with relevant engineering geological specifications and serves as a comprehensive evaluation index for rock mass integrity.
[0039] By concatenating the above three physical indicators with the 1024-dimensional face feature vector, and then passing them through two multilayer perceptrons, the final face stability index is obtained. .
[0040] in ∈[0,1], the denser the joints learned by the intelligent evaluation network of the surrounding rock structure, the larger the fractured zone. The smaller the mapping relationship, the more stable the surrounding rock becomes, and the higher the risk of the cutterhead jamming due to collapse.
[0041] Please refer to Figure 3 , Figure 3 This is a flowchart of an intelligent assessment network for rock debris characteristics proposed in one embodiment of this application. Figure 3 As shown, the intelligent assessment network for rock debris characteristics includes: Multi-task convolutional neural network: extracts features from the rock debris image to obtain a feature image of the rock debris; particle size regression task head: obtains the particle size distribution features of the feature image of the rock debris image to obtain rock debris particle size information; shape classification task head: classifies the feature image of the rock debris image to obtain rock debris shape information; texture feature task head: obtains the texture features of the feature image of the rock debris image to obtain rock debris texture feature information; multilayer perceptron: fuses the feature information obtained from the three task heads to output the rock debris degradation index.
[0042] Furthermore, the rock fragment particle size information is the effective particle size d. 10 Median particle size d 30 , limited particle size d 60 Inhomogeneity coefficient and curvature coefficient The shape information is block-shaped, sheet-shaped, or needle-shaped.
[0043] In the coefficient of non-uniformity and curvature coefficient The formulas for both are:
[0044]
[0045] As an feasible approach, the multi-task convolutional neural network adopts a CNN network architecture. After obtaining the feature image of the rock fragments, it is fed into three output heads for feature extraction. The particle size regression task head returns rock fragment particle size information. A flat particle size curve with a wide distribution range is a typical manifestation of rock fragmentation. The shape classification task returns rock fragment shape information. A high proportion of flaky and needle-shaped rock fragments is a significant indicator of weak and fractured strata. The texture feature task head returns 128-dimensional rock fragment texture information to capture subtle differences such as surface roughness and fracture freshness of the rock fragments.
[0046] The cinder particle size information, cinder shape information, and 128-dimensional cinder texture information are stitched together, and the cinder degradation index is finally obtained by using a multilayer perceptron on both sides. .
[0047] in ∈[0,1], the intelligent evaluation network for rock fragment features learns that the more fragmented and flaky the rock fragments are, the higher their degree of fragmentation. The larger the mapping relationship, the more likely the index is to indicate the self-stabilizing ability of the surrounding rock. The larger the index, the more fragmented and flaky the rock fragments are, and the higher the risk of the collapsing cutterhead jamming.
[0048] Please refer to Figure 4 , Figure 4 This is a flowchart of a network for intelligent assessment of tunneling status proposed in one embodiment of this application. Figure 4 As shown, the intelligent assessment network for tunneling status includes: A temporal prediction model using long short-term memory networks and attention mechanisms: Predicting tunneling parameters using the tunneling parameter sequence; a feature extraction module extracting features from the tunneling parameter sequence to obtain corresponding composite index features; and a multilayer perceptron fusing the predicted tunneling parameters and the composite index features to output a tunneling anomaly index.
[0049] Furthermore, the composite index is characterized by torque fluctuation index, penetration coefficient of variation, and specific energy change rate, and its formula is:
[0050]
[0051]
[0052] in, The torque fluctuation index. The standard deviation of the cutter head torque. This represents the average torque of the cutter head. The coefficient of variation for penetration. The standard deviation of penetration. The mean penetration value. The rate of change of specific energy For thrust, For penetration sequence, The area of the cutter head is denoted as ...
[0053] As an feasible approach, the temporal prediction model of the Long Short-Term Memory network and attention mechanism uses an LSTM-Attention model architecture. It takes the thrust, cutterhead torque, propulsion speed, cutterhead rotation speed, and penetration depth sequence from the past 50 minutes as input and predicts the parameter values for the next 5 minutes. The predicted future parameters are compared with the equipment's rated values or safety thresholds, and a composite index is calculated using the feature extraction module. The 128-dimensional hidden state features output by LSTM-Attention are concatenated with the 32-dimensional composite index features, and the tunneling anomaly index is obtained through a two-layer multilayer perceptron. .
[0054] Among them, the tunneling anomaly index ∈[0,1], when thrust and torque remain high while velocity or penetration decreases, or when shield pressure increases sharply and remains high, A significant increase indicates an abnormal tunneling condition, suggesting an impending risk of the machine stalling.
[0055] As an feasible approach, thresholds are set for the working face stability index, rock debris deterioration index, and tunneling anomaly index: the working face stability index threshold is 0.3; the rock debris deterioration index threshold is 0.7; and the tunneling anomaly index threshold is 0.7.
[0056] The three generated indices each correspond to an independent link in the TBM card machine accident evolution chain, have clear engineering and physical meanings, and there is an inherent logical connection and synergistic relationship among the three.
[0057] The physical significance of the tunnel face stability index: The tunnel face is the excavation face directly affected by the TBM cutterhead, and the structural integrity of its surrounding rock directly determines the possibility of rockfall and collapse after excavation. When the joint density of the tunnel face increases and the area of the fractured zone expands, the self-stabilizing ability of the surrounding rock decreases, and local or large-scale collapses are very likely to occur after excavation. The collapsed rock mass accumulates in front of and around the cutterhead, causing the rotational resistance of the cutterhead to increase sharply, resulting in cutterhead jamming. This is a quantitative expression of the "inherent conditions of surrounding rock instability": the smaller the index, the denser the joints, the larger the proportion of the fractured zone, and the worse the overall structure of the surrounding rock, the higher the risk of cutterhead jamming caused by collapse.
[0058] The physical significance of the rock debris deterioration index: Rock debris is a direct product of excavation at the tunnel face, and its morphological characteristics accurately reflect the actual fracture state of the surrounding rock. The more fractured the surrounding rock and the worse its self-stabilizing ability, the more uneven the particle size of the resulting rock debris, the higher the proportion of flaky and needle-like particles, and the fresher the fracture surface. This provides a quantitative expression of the "state of fractured surrounding rock products": the larger the index, the more fractured the rock debris, the higher the degree of flaky formation, and the fresher the fracture surface, reflecting substantial damage to the surrounding rock ahead and a higher risk of the cutterhead jamming due to collapse. This index is usually significantly higher in fault fracture zone excavation sections than in normal excavation sections, and there is a statistically verifiable difference between the two, making it engineering-verifiable.
[0059] The physical significance of the tunneling anomaly index: Tunneling parameters are a direct product of the interaction between the TBM and the surrounding rock, reflecting the immediate load state of the equipment under the current surrounding rock conditions. When entering unfavorable geological sections, the parameters exhibit a series of quantifiable precursor characteristics: a continuous increase in thrust indicates increased resistance to rock breaking by the cutterhead; intensified torque fluctuations indicate uneven surrounding rock ahead and intermittent obstruction of the cutterhead; increased shield pressure indicates that the surrounding rock is converging and squeezing the shield into the tunnel. This quantifies the "mechanical response state": the larger the index, the more abnormal the tunneling state, indicating an approaching risk of machine jamming. The important engineering value of this index lies in its predictive nature: the inflection point of an anomaly in a composite index usually appears earlier than the exceedance of a single parameter, buying valuable time for on-site preventative measures.
[0060] The three indices follow the spatiotemporal evolution of TBM jamming accidents, forming a complete chain: the face stability index reflects the "surrounding rock structure state"; the rock debris deterioration index reflects the "surrounding rock fracture product state"; and the tunneling anomaly index reflects the "mechanical response state". In terms of time, they constitute a progressive relationship of "precursor, manifestation, and response", and logically, they form a causal chain of "condition, verification, and consequence".
[0061] The hierarchical decoupling architecture of this invention directs the output of each level to a specific engineering problem: what are the surrounding rock conditions, has the surrounding rock suffered substantial damage, and what is the equipment status? Each level's index is itself a physical quantity with independent engineering judgment value. Even without final fusion prediction, the three indices can guide engineering practice: When the temperature drops below a certain threshold, proactive support should be strengthened in advance. When the value exceeds a certain threshold, verify the accuracy of the advanced geological forecast and review the support parameters. When the output exceeds a certain threshold, the tunneling parameters are adjusted or the equipment status is checked. This design concept, which emphasizes that "each level of output has independent engineering significance," is the core feature that distinguishes this invention from a purely black-box model.
[0062] Please refer to Figure 5 , Figure 5This is a flowchart of the encoder module and task prediction output layer of the multi-head self-attention mechanism proposed in an embodiment of this application.
[0063] like Figure 5 As shown, the encoder module of the multi-head self-attention mechanism specifically concatenates the three indices—the face stability index, the rock debris deterioration index, and the tunneling anomaly index—and their corresponding embedded feature vectors, and maps them to 256 dimensions through a linear projection layer. This ensures that the features of different modalities are comparable in numerical scale and semantic space. The 256-dimensional feature vectors of the three modalities are concatenated, and the resulting data is reshaped into three head features and their corresponding 256-dimensional feature vectors. These are then input into the encoder of the multi-head self-attention mechanism. After scaling, dot product, and softmax normalization, a weighted sum is performed. Finally, a 512-dimensional fusion discriminative feature vector H is adaptively generated through multi-head concatenation and linear projection.
[0064] Furthermore, the Softmax normalization is configured to derive the percentage contribution of the face stability index, the rock debris deterioration index, and the tunneling anomaly index to the jamming probability, as shown in the following formula:
[0065] in The percentage contribution of the three indices to mode m. Let m be the attention weight of mode m in the k-th attention head. The stability index of the tunnel face. The rock debris deterioration index. This is the tunneling anomaly index.
[0066] Furthermore, the predicted card machine probability loss function is Focal Loss, the remaining card machine distance loss function is Huber Loss, and the total loss is the weighted sum of the two. The Focal Loss formula is:
[0067] in The Focal Loss value. This represents the model's predicted probability for the true class. For focusing parameters; The Huber Loss formula is:
[0068] in The Huber Loss value. This represents the actual remaining distance of the card reader. The predicted remaining distance for the SIM card. This is the acceptable error threshold.
[0069] The total loss function is:
[0070]
[0071] in and This is the balance coefficient.
[0072] The multi-task prediction output layer module specifically inputs the fused discriminative feature vector H into the multi-task prediction output layer, adopting an architecture of "shared backbone + two task-specific branches". The feature vector H is refined to 128 dimensions through a shared fully connected layer, and then processed through two branches: Classification branch: The 128-dimensional feature vector, after passing through the branch's fully connected layer, is passed through a jam probability loss function and then into a Sigmoid activation function, compressing the output to the [0,1] interval, which is interpreted as the jam probability; Regression branch: The 128-dimensional feature vector, after passing through the branch's fully connected layer, is passed through a jam remaining distance loss function, without using an activation function, directly outputting a floating-point number, i.e., the estimated jam remaining distance. The jam probability represents the probability of a jam event occurring at a certain future tunneling distance. The estimated jam remaining distance represents how many meters are expected to occur before a jam, based on the current tunneling trend, starting from the current moment.
[0073] The classification loss reduces the weight of easily classified samples, making the model pay more attention to those samples with ambiguous features that are easily misclassified as early signs of a crash; the regression loss prevents extreme error values caused by geological changes from dominating gradient updates, thus enhancing the robustness of model training.
[0074] As an feasible approach, the probability of machine jamming, P, is divided into four levels: P ≥ 0.8 is a red alert, 0.6 ≤ P < 0.8 is an orange alert, 0.4 ≤ P < 0.6 is a yellow alert, and P < 0.4 is a blue alert. A red alert means immediately stopping construction to eliminate the potential jamming hazard; an orange alert means slowing down construction to reduce the risk of jamming; a yellow alert means continuing normal construction to prevent jamming from occurring; and a blue alert means maintaining normal construction.
[0075] Furthermore, once a machine jamming event is confirmed on-site, the system extracts data on the working face, rock debris, and tunneling sequence, along with event tags, from a period prior to the event and adds them to the incremental training set. Normal tunneling data and accident data are sampled and added at a 4:1 ratio to maintain data balance. To avoid the significant time overhead of retraining the entire model, a layered fine-tuning strategy is adopted: the parameters of the surrounding rock structure intelligent evaluation network, the rock debris feature intelligent evaluation network, and the tunneling status intelligent evaluation network are frozen; only the parameters in the multi-head self-attention mechanism module and the multi-task output fully connected layer module are updated. Using mini-batch stochastic gradient descent, fine-tuning can be completed within minutes with only a few hundred new samples. This can be performed during TBM maintenance or cutterhead replacement periods without affecting normal tunneling.
[0076] During incremental fine-tuning, in addition to using newly generated catalytic converter data, a batch of historical data (including normal and old catalytic converter samples) is randomly selected from the original training set and mixed with the new data for training. This mechanism ensures that while the model learns to identify the "hazard of new strata," it does not lose its ability to judge the "safety of old strata," making the entire model more adaptable to different geological conditions and achieving long-term stability.
[0077] As an feasible approach, while outputting the jamming probability P and tiered early warnings, the system independently determines the thresholds for three indices: when an index exceeds its warning threshold, even if the jamming probability P does not trigger the corresponding warning level, the system still outputs an independent warning message for that index as a supplementary prompt. For example, when the surrounding rock is extremely unstable, i.e., the working face stability coefficient is less than the threshold of 0.3, but the jamming probability is 0.35, which falls under the blue warning category, the system outputs "Below the warning threshold, the surrounding rock has poor self-stabilizing ability, it is recommended to strengthen advanced support and closely monitor the situation." This ensures that any serious anomaly in a single dimension is not masked by the fusion mechanism, while providing an independent risk reference for the site beyond the jamming probability.
[0078] Ultimately, the system generates a complete prediction that includes risk level, probability, remaining distance, cause breakdown, and independent early warning information. For example: Red warning, probability of machine jamming 0.87, machine jamming is highly likely 5.3 meters ahead, abnormal tunneling parameters contribute 62%, rock debris deterioration contributes 28%, and working face stability contributes 10%.
[0079] In summary, this disclosure proposes a TBM jam prediction method based on intelligent hierarchical evaluation of multi-source data. Specifically, it acquires face images, rock debris images, and tunneling sequence parameters as datasets and obtains corresponding parameters and feature vectors through three different models. After feature concatenation and input into a multi-head self-attention module, multi-task prediction is used to derive the final jam probability and jam prediction distance. In this disclosure, the three types of multi-source heterogeneous data—face images, rock debris images, and tunneling parameters—are transformed into intermediate indices with clear engineering physical meaning. Each of the three indices corresponds to an independent link in the evolution chain of TBM jam accidents. They can be independently output, independently verified, and independently guide engineering practice before fusion prediction, fundamentally overcoming the defects of existing end-to-end black box models where intermediate processes are untraceable and unexplainable. This makes the early warning results engineering verifiable and operationally instructive; it can quantify the contribution of the three indices to the final jam probability in percentage form, generating complete early warning information including "main cause + secondary cause". Meanwhile, even when a single index is abnormal but the probability of system malfunction is low, the system still outputs an independent early warning, ensuring complete interpretability and precise guidance for handling the situation. This significantly improves on-site decision-making efficiency and resource utilization. The system's early warning capabilities are strong; combining the forward-looking nature of images with the trend prediction ability of parameters, it achieves proactive identification of malfunction risks, rather than only issuing alarms when malfunctions occur, thus allowing sufficient time for on-site handling. It ensures high prediction accuracy and has strong generalization ability. Employing multimodal fusion based on multi-head self-attention, it can automatically learn the optimal weights of various data sources under different geological conditions. Compared to single-data-source methods, its fusion prediction accuracy is significantly improved, and it is more adaptable to different geological conditions.
[0080] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0081] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.
[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0083] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.
[0084] The above provides a detailed description of the TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A TBM (Traffic Machine) Prediction Method Based on Intelligent Hierarchical Evaluation of Multi-Source Data, characterized in that, include: Step S1: Obtain the tunnel face image, rock debris image, and tunneling parameter sequence data, and perform data preprocessing to obtain the preprocessed tunnel face image, preprocessed rock debris image, and preprocessed tunneling parameter sequence data; Step S2: The preprocessed tunnel face image is sent to the intelligent evaluation network of surrounding rock structure to obtain the tunnel face stability index and the tunnel face information fusion feature vector. The preprocessed rock debris image is fed into the rock debris feature intelligent evaluation network to obtain the rock debris deterioration index and rock debris information fusion feature vector; The preprocessed tunneling parameter sequence data is sent into the intelligent evaluation network for tunneling status to obtain the tunneling anomaly index and the fusion feature vector of tunneling status information. Step S3: Concatenate the face stability index with the face information fusion feature vector and project it to a specified dimension to obtain a first projection vector; concatenate the rock debris deterioration index with the rock debris information fusion feature vector and project it to a specified dimension to obtain a second projection vector; concatenate the tunneling anomaly index with the tunneling state information fusion feature vector and project it to a specified dimension to obtain a third projection vector; and send the first projection vector, the second projection vector, and the third projection vector into an encoder module based on a multi-head self-attention mechanism to generate a fusion discrimination feature vector. Step S4: Input the fused discriminative feature vector into the multi-task prediction output layer to obtain the predicted lag probability and the remaining lag distance; Step S5: Based on the jamming probability and combined with the face stability index, the rock debris deterioration index and the tunneling anomaly index, perform a graded early warning of the jamming probability, obtain early warning information and display the contribution of the three indices to the jamming probability, and analyze the actual causes of the jamming through the above data.
2. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 1, characterized in that, The intelligent assessment network for surrounding rock structures includes: Semantic segmentation sub-network: performs pixel-level segmentation of the joints and fragmented areas of the face image to obtain face feature vectors and face image segmentation mask; Physical index calculation module: Calculates physical indexes of the working face based on the segmentation mask of the working face image; Multilayer perceptron: It fuses the face feature vector with the face physical index and outputs the face stability index.
3. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 2, characterized in that, The physical parameters of the working face are joint density, bunker fragmentation index, and integrity coefficient.
4. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 1, characterized in that, The intelligent assessment network for rock debris characteristics includes: Multi-task convolutional neural network: extracts features from the rock debris image to obtain a rock debris image feature image; Particle size regression task head: Obtain the particle size distribution features of the rock debris image feature image to obtain rock debris particle size information; Shape classification task head: Classifies the feature images of the rock debris to obtain rock debris shape information; Texture feature task head: Obtain the texture features of the rock debris image feature image to obtain rock debris texture feature information; A multilayer perceptron is used to fuse the feature information obtained from the three task heads and output the rock debris deterioration index.
5. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 4, characterized in that, The rock fragment particle size information is the effective particle size d. 10 Median particle size d 30 , limited particle size d 60 Inhomogeneity coefficient and curvature coefficient The shape information is block-shaped, sheet-shaped, or needle-shaped.
6. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 1, characterized in that, The intelligent assessment network for tunneling status includes: Temporal prediction model of long short-term memory network and attention mechanism: using the aforementioned tunneling parameter sequence for prediction to obtain the corresponding predicted tunneling parameters; The feature extraction module extracts features from the tunneling parameter sequence to obtain corresponding composite index features; A multilayer perceptron is used to fuse the predicted tunneling parameters and the composite index features to output a tunneling anomaly index.
7. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 6, characterized in that, The composite index features torque fluctuation index, penetration coefficient of variation, and specific energy change rate.
8. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 1, characterized in that, The predicted card machine probability loss function is Focal Loss, the card machine remaining distance loss function is Huber Loss, and the total loss is the weighted sum of the two. The Focal Loss formula is: ; in The Focal Loss value. This represents the model's predicted probability for the true class. For focusing parameters; The Huber Loss formula is: in The Huber Loss value. This represents the actual remaining distance of the card reader. The predicted remaining distance for the SIM card. This is the acceptable error threshold. The total loss function is: in and This is the balance coefficient.
9. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 1, characterized in that, In the step of feeding the first projection vector, the second projection vector, and the third projection vector into the encoder module based on a multi-head self-attention mechanism to generate a fused discriminative feature vector, The encoder module of the multi-head self-attention mechanism is used to concatenate the first projection vector, the second projection vector and the third projection vector into three head features and corresponding feature vectors. Each head feature and its corresponding feature vector are scaled and dot-producted, then normalized by Softmax, and used as weights for weighted summation to obtain the weighted vector corresponding to each feature head. The weighted vector corresponding to each feature head is generated into a fused discriminative feature vector through multi-head concatenation and linear projection. The Softmax normalization is also configured to derive the percentage contribution of the face stability index, the rock debris deterioration index, and the tunneling anomaly index to the jamming probability, using the following formula: in The percentage contribution of the three indices to mode m. Let m be the attention weight of mode m in the k-th attention head. The stability index of the tunnel face. The rock debris deterioration index. This is the tunneling anomaly index.
10. The TBM card machine prediction method based on intelligent hierarchical evaluation of multi-source data according to claim 1, characterized in that, The method further includes: When a jamming event occurs, an online incremental learning process is provided. The face image, rock debris image, and tunneling parameter sequence data at the time of the jamming event are extracted. A hierarchical fine-tuning strategy is adopted to fine-tune the encoder module and the parameters within the multi-task prediction output layer of the multi-head self-attention mechanism. The hierarchical fine-tuning strategy involves freezing the parameters of the surrounding rock structure intelligent evaluation network, the rock debris feature intelligent evaluation network, and the tunneling status intelligent evaluation network; performing small-batch gradient updates on the encoder module of the multi-head self-attention mechanism and the multi-task prediction output layer; and mixing new event data and historical sampled data in a specified ratio for training to prevent catastrophic forgetting.
