A TBM shield system anomaly detection method based on a space-time graph neural network

Abstract

The application relates to the technical field of TBM shield system anomaly detection, and discloses a TBM shield system anomaly detection method based on a space-time graph neural network. The method aims to solve the technical problem that the existing graph neural network method lacks special design for the shield circumference geometric constraint in the TBM application, and the anomaly detection is often disconnected with the surrounding rock stability grade prediction, resulting in incomplete decision support. The application comprises the following steps: S1, data acquisition and preprocessing; S2, graph structure construction; S3, space-time feature extraction; S4, anomaly detection and classification; S5, feedback and optimization; S1 to S5 are cyclically executed to realize real-time monitoring; each step processes the data of the current time window, outputs the abnormal state, and feeds back the result to the TBM operator to form a closed-loop control. The application directly models the sensor network through the graph neural network, improves the calculation efficiency, fuses the space-time information and the geometric constraint, and improves the surrounding rock grade prediction accuracy by more than 15%.

Description

Technical Field

[0001] This invention relates to the field of TBM shield system anomaly detection technology, and in particular to a TBM shield system anomaly detection method based on spatiotemporal graph neural network. Background Technology

[0002] With the widespread application of tunnel boring machines (TBMs) in tunnel engineering, safety and efficiency during the tunneling process have become critical issues. When TBMs are tunneling under complex geological conditions, the shield system bears enormous pressure from the surrounding rock. If abnormalities occur in the circumferential distribution of the shield, such as asymmetrical displacement or pressure imbalance, it may lead to equipment jamming, rock instability, or even collapse. Traditional anomaly detection methods are mainly based on threshold alarms or statistical process control. These methods often rely on manual experience to set fixed thresholds, making them difficult to adapt to the dynamic changes and complex working conditions during TBM tunneling. For example, threshold alarm methods cannot capture the spatiotemporal correlation between shield sensors, while statistical methods have poor adaptability to non-stationary time series, resulting in a high false alarm rate.

[0003] In recent years, machine learning techniques have been introduced into the field of TBM monitoring. Models such as Support Vector Machines (SVMs) or Random Forests can improve detection accuracy, but most are limited to single-point sensor data analysis, ignoring the overall geometric constraints of the shield's circumferential distribution. The shield system consists of multiple sensor nodes, including the top shield, left shield, right shield, and wedge-shaped shields. These nodes are distributed circumferentially, and their displacement and pressure data exhibit strong spatial correlations. Traditional machine learning methods struggle to effectively model this graph-structured data and cannot capture the topological relationships between nodes, such as the mechanical interactions between adjacent shields.

[0004] Furthermore, while recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) in deep learning can handle time series data, they primarily target single-sensor signals and fail to integrate spatial dependencies between multiple sensors. Graph neural networks (GNNs), as an emerging technology, can handle non-Euclidean data and are suitable for sensor network modeling. However, existing GNN methods still have shortcomings in TBM applications: First, they typically assume that the graph structure is statically invariant, but the shield relationship in TBM tunneling dynamically adjusts with geological changes; second, there is a lack of specific designs for shield circumferential geometric constraints, such as symmetry checks and pressure balance verification; finally, anomaly detection is often disconnected from the prediction of surrounding rock stability levels, resulting in incomplete decision support.

[0005] Therefore, this invention proposes a TBM tunneling data anomaly detection system based on spatiotemporal graph neural networks. By fusing GNN and LSTM, the spatiotemporal characteristics of the shield sensor are modeled, and a geometric constraint mechanism is introduced to achieve efficient and adaptive anomaly detection. This method aims to overcome the limitations of traditional technologies and improve the safety and intelligence level of TBM tunneling. Summary of the Invention

[0006] In view of the above technical problems, this disclosure provides an anomaly detection method for TBM shield systems based on spatiotemporal graph neural networks. This method solves the following problems: First, existing GNN methods in TBM applications usually assume that the graph structure is static and unchanged, but the shield relationship is dynamically adjusted with geological changes during TBM tunneling. Second, there is a lack of specialized design for the shield's circumferential geometric constraints, such as symmetry checks and pressure balance verification. Finally, anomaly detection is often disconnected from the prediction of surrounding rock stability levels, resulting in incomplete decision support.

[0007] According to one aspect of this disclosure, an anomaly detection method for a TBM shield system based on a spatiotemporal graph neural network is provided. The shield sensor network is modeled as a graph structure, where nodes represent sensors, including displacement and pressure, and edges represent the spatial relationships between sensors, including geometric adjacency. The method includes the following steps: S1: Data Acquisition and Preprocessing; Real-time acquisition of multi-source sensor data from the TBM monitoring system, including shield displacement, pressure, and operating parameters, including thrust and torque; Data cleaning, normalization, and serialization to form time window samples; S2: Graph structure construction; Define the connection relationship between sensor nodes based on the geometric position of the shield on the circumference; Edge weights are calculated based on angle differences, the smaller the angle difference, the stronger the connection, to reflect spatial proximity; S3: Spatiotemporal feature extraction; Graph neural networks are used to process the graph data at each time step, aggregate neighbor node information, and extract spatial features; then, the output of the GNN on the time series is input into the Long Short-Term Memory (LSTM) network to capture temporal dependencies; S4: Anomaly detection and classification; weighted key time steps through an attention mechanism to output surrounding rock stability level and anomaly score; anomaly detection integrates model prediction and geometric constraint checks, symmetry verification, and generates decision suggestions; S5: Feedback and Optimization; The system automatically adjusts the threshold and updates the model parameters based on the detection results to achieve adaptive learning; The process executes S1 to S5 in a loop to achieve real-time monitoring. Each step processes the data of the current time window, outputs abnormal statuses, and feeds the results back to the TBM operator, forming a closed-loop control.

[0008] In some embodiments of this disclosure, the geometric position of the shield in S2 on the circumference corresponds as follows: the top shield corresponds to 0 degrees, and the left shield corresponds to 90 degrees.

[0009] In some embodiments of this disclosure, the graph neural network described in S3 includes a graph convolution module, a time series module, an attention module, and an output module; The graph convolutional module uses a graph attention network (GAT) as the basic graph neural network layer. The graph structure has N nodes, and each node's initial feature is a scalar. The GAT layer calculates the weights between nodes through an attention mechanism. For nodes i and j, the attention coefficient α... ij The calculation formula is: ; Among them, h i and h j Here, W is the node feature, a is the weight matrix, a is the attention vector, ∥ represents the concatenation operation, and N(i) is the set of neighbors of node i. The formula allows the model to adaptively learn edges. The network contains multiple layers of GAT, and the output of each layer is activated by the LeakyReLU activation function and Dropout regularization to prevent overfitting. The time series module uses a bidirectional LSTM to process the time series output by graph convolution; assuming a time step size of T, the node features at each time step are global pooled and then used as the input to the LSTM; the hidden state update formula of the LSTM is: ; Where, x t It is the input for time step t, h t It is a hidden state; bidirectional LSTM captures context from both the front and back directions, enhancing temporal modeling; The attention module is used to apply a multi-head attention mechanism to the LSTM output and calculate the weights between time steps; for sequence representation The attention output is: ; Where Q is the query, K is the key, V is the value matrix, and d k It is the dimension scaling factor; The output module includes a classification head and an anomaly detection head; the classification head predicts the surrounding rock grade through a fully connected layer, and the anomaly detection head outputs anomaly scores using a Sigmoid function; the optimizer uses Adam W, combined with weight decay, with the learning rate set to 0.001, and is dynamically adjusted through the Reduce LR On Plateau scheduler. The entire network structure is trained end-to-end to ensure the fusion of spatial and temporal features.

[0010] In some embodiments of this disclosure, the loss function of the graph neural network described in S3 is designed in a multi-task form to balance classification accuracy and anomaly detection reliability; the overall loss function It is classification loss and abnormal losses Weighted sum: ; Here, α and β are weighting coefficients, with α set to 1 and β set to 0.5 to balance the importance of tasks; Classification loss Weighted cross-entropy is used to address the imbalance in surrounding rock grades; let the number of categories be C, the true label of the sample be y, the predicted probability be p, and the weights be... Calculation based on category frequency: ; In the formula, N is the total number of samples, and Nc is the number of samples in class c; this is used to improve the detection sensitivity of minority classes. Abnormal loss L anom Based on mean squared error, outlier scores are encouraged to be close to their binary labels, where 0 is normal and 1 is abnormal. ; in, It predicts outlier scores. These are true anomaly labels (derived from geometric constraint violations); the loss function introduces a regularization term. This is used to ensure confidence levels, penalize excessively high entropy in predicted probabilities, and ensure that low-confidence labels only have adjacent levels with similar confidence levels. ; In the formula, It is a probability distribution The entropy is τ, the threshold (e.g., 0.7), I is the indicator function, and λ is the regularization coefficient; the formula enables the model to distribute the probability quality across adjacent levels when there is uncertainty, avoiding overconfidence. The overall loss function is optimized through gradient descent to improve the balance between accuracy and confidence in the model.

[0011] In some embodiments of this disclosure, the input data of the multi-source sensors in S1 includes real-time TBM sensor data and historical logs; the sensor data is divided into shield data and operational data; the shield data includes displacement and pressure, and the operational data includes thrust and velocity. The shield data has spatiotemporal characteristics: spatially, the sensors are distributed in a circle, with displacement values ​​ranging from 0-20 mm and pressure values ​​ranging from 0-40 MPa, and a sampling period of 2 seconds; temporally, the data is a non-stationary time series, affected by geological changes; the operational data serves as an auxiliary feature for contextual modeling. It also includes data processing of the input data from the multi-source sensors, including the following steps: a. Data cleaning: Outlier detection is performed using a sliding window, and outliers are removed using the 3σ principle; for missing values, time series interpolation is used for imputation, with the following formula: ; Where, x t These are the missing values ​​at time t; b. Normalization: Z-score standardization is used to scale the features to zero mean and unit variance; for feature x, the processed value x′ is: ; Where μ is the sum of the means and σ is the standard deviation; c. Serialization: The continuous data is divided into time windows with a window length of L=60, corresponding to 2 minutes, and a step size of S=30, corresponding to 1 minute; each window contains readings from multiple sensors, forming a graph sequence.

[0012] In some embodiments of this disclosure, the output surrounding rock stability level in S4 is an integer from 1 to 5, representing the surrounding rock state, with 1 being the most stable. The anomaly score is a continuous value between 0 and 1, with a higher score indicating a greater anomaly. The results of the geometric constraint check include displacement asymmetry, in millimeters; The output data is encapsulated in JSON format, including timestamps and confidence levels, making it easy to integrate into monitoring systems.

[0013] In S5: Feedback and Optimization, the training process of the model is divided into four stages: data preparation, model initialization, iterative training, and validation, as described in some embodiments of this disclosure; the entire process processes data in batches. First, samples are loaded from historical data to construct training and validation sets; each sample contains a graph sequence and label for a time window; model parameters are initialized using Xavier to ensure gradient stability. During the training iteration, for each batch with a batch size of 32, the following steps are performed: a. Forward propagation: Input graph sequence, extract spatial features through graph neural network layers; for each time step, graph neural network aggregates neighbor information and outputs node embeddings; then, long short-term neural network processes the time series to generate hidden states; b. Loss Calculation: Compare the predicted output with the true label to calculate the total loss; backpropagate the loss value and perform gradient clipping with a norm threshold of 1.0 to prevent explosion. c. Parameter update: The optimizer adjusts the weights, and the learning rate is dynamically reduced based on the validation loss; Training lasts for 100 epochs, with an early stopping mechanism terminating training when the validation loss no longer improves; the validation phase uses an independent dataset to evaluate the model's generalization ability; after training, the model is deployed to a real-time system to process streaming data. The entire training process ensures that the model learns spatiotemporal patterns from the data and enhances physical plausibility through geometric constraints.

[0014] The beneficial effects of this invention are as follows: (1) High efficiency: The sensor network is directly modeled by graph neural network, avoiding feature engineering and improving computational efficiency.

[0015] (2) Accuracy: By integrating spatiotemporal information and geometric constraints, the false alarm rate is reduced and the accuracy of surrounding rock grade prediction is increased by more than 15%.

[0016] (3) Adaptability: Dynamically adjust thresholds and model parameters to adapt to different geological conditions.

[0017] (4) Real-time performance: Supports streaming data processing with a latency of less than 2 seconds, meeting the real-time requirements of engineering projects.

[0018] (5) Explainability: The attention mechanism provides a basis for decision-making and is easy for operators to understand.

[0019] Deep integration of multi-dimensional data enhances the comprehensiveness of detection. It not only considers sensor readings but also explicitly models the spatial geometric relationships of shield displacement and pressure sensors through graph structures, and integrates operational parameters such as thrust and torque. This surpasses traditional single-point threshold alarms, enabling the capture of the spatiotemporal evolution patterns of anomalies at the system level.

[0020] Accurately capturing spatiotemporal dependencies improves detection accuracy. Graph attention networks adaptively learn the importance weights between sensor nodes, flexibly capturing spatial features such as asymmetric forces. Bidirectional LSTM extracts bidirectional contextual information from time series, more accurately modeling the dynamic process of shield state changes. Multi-head attention mechanisms automatically identify key time steps, enhancing sensitivity to changes in dangerous states.

[0021] Driven by both data and knowledge, the model's reliability is enhanced by an innovative approach that incorporates physical geometric constraints for verification, such as using geometric symmetry to test asymmetry. This ensures that model predictions not only rely on data but also conform to physical laws, effectively filtering out false alarms that clearly violate common sense, and making the detection results more interpretable and credible for engineering applications.

[0022] To address the challenge of data imbalance and improve the sensitivity of rare risk detection, the classification module employs frequency-based weighted cross-entropy loss, while the anomaly detection layer uses MSE loss. The multi-task learning framework explicitly enhances the model's ability to detect minority and high-risk samples, such as those exhibiting rock instability.

[0023] Quantifying prediction uncertainty and providing calibrable confidence levels, the model is encouraged to assign probabilities to adjacent levels under high uncertainty conditions, rather than making high-confidence predictions with low confidence, through entropy regularization loss. This provides a more realistic and calibrable confidence reference for the output, helping operators assess the risks of decision-making prompts.

[0024] Closed-loop adaptive learning enables long-term online optimization, supports online operation, and automatically adjusts thresholds and updates parameters based on detection results, forming a complete closed loop of data acquisition, detection, feedback, and optimization. This allows the model to adapt to the slow changes in TBM equipment wear and geological conditions, maintaining performance throughout long-term operation. Attached Figure Description

[0025] Figure 1 A schematic diagram of the structure of an anomaly detection method for a TBM shield system based on a spatiotemporal graph neural network; Detailed Implementation

[0026] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0027] Example 1

[0028] This example discloses an anomaly detection method for TBM shield systems based on spatiotemporal graph neural networks. See [link to relevant documentation]. Figure 1 , The shield sensor network is modeled as a graph structure, where nodes represent sensors (including displacement and pressure) and edges represent spatial relationships between sensors (including geometric adjacency). The process includes the following steps: S1: Data Acquisition and Preprocessing; Real-time acquisition of multi-source sensor data from the TBM monitoring system, including shield displacement, pressure, and operating parameters, including thrust and torque; Data cleaning, normalization, and serialization to form time window samples; S2: Graph structure construction; Define the connection relationship between sensor nodes based on the geometric position of the shield on the circumference; Edge weights are calculated based on angle differences, the smaller the angle difference, the stronger the connection, to reflect spatial proximity; S3: Spatiotemporal feature extraction; Graph neural networks are used to process the graph data at each time step, aggregate neighbor node information, and extract spatial features; then, the output of the GNN on the time series is input into the Long Short-Term Memory (LSTM) network to capture temporal dependencies; S4: Anomaly detection and classification; weighted key time steps through an attention mechanism to output surrounding rock stability level and anomaly score; anomaly detection integrates model prediction and geometric constraint checks, symmetry verification, and generates decision suggestions; S5: Feedback and Optimization; The system automatically adjusts the threshold and updates the model parameters based on the detection results to achieve adaptive learning; The process executes S1 to S5 in a loop to achieve real-time monitoring. Each step processes the data of the current time window, outputs abnormal statuses, and feeds the results back to the TBM operator, forming a closed-loop control.

[0029] The geometric position of the shields on the circumference described in S2 is as follows: the top shield corresponds to 0 degrees, and the left shield corresponds to 90 degrees.

[0030] The graph neural network described in S3 includes a graph convolution module, a time series module, an attention module, and an output module; The graph convolutional module uses a graph attention network (GAT) as the basic graph neural network layer. The graph structure has N nodes, and each node's initial feature is a scalar. The GAT layer calculates the weights between nodes through an attention mechanism. For nodes i and j, the attention coefficient α... ij The calculation formula is: ; Among them, h i and h j Here, W is the node feature, a is the weight matrix, a is the attention vector, ∥ represents the concatenation operation, and N(i) is the set of neighbors of node i. The formula allows the model to adaptively learn edges. The network contains multiple layers of GAT, and the output of each layer is activated by the LeakyReLU activation function and Dropout regularization to prevent overfitting. The time series module uses a bidirectional LSTM to process the time series output by graph convolution; assuming a time step size of T, the node features at each time step are global pooled and then used as the input to the LSTM; the hidden state update formula of the LSTM is: ; Where, x t It is the input for time step t, h t It is a hidden state; bidirectional LSTM captures context from both the front and back directions, enhancing temporal modeling; The attention module is used to apply a multi-head attention mechanism to the LSTM output and calculate the weights between time steps; for sequence representation The attention output is: ; Where Q is the query, K is the key, V is the value matrix, and d k It is the dimension scaling factor; The output module includes a classification head and an anomaly detection head; the classification head predicts the surrounding rock grade through a fully connected layer, and the anomaly detection head outputs anomaly scores using a Sigmoid function; the optimizer uses Adam W, combined with weight decay, with the learning rate set to 0.001, and is dynamically adjusted through the Reduce LR On Plateau scheduler. The entire network structure is trained end-to-end to ensure the fusion of spatial and temporal features.

[0031] The loss function of the graph neural network described in S3 is designed in a multi-task form to balance classification accuracy and anomaly detection reliability; the overall loss function... It is classification loss and abnormal losses Weighted sum: ; Here, α and β are weighting coefficients, with α set to 1 and β set to 0.5 to balance the importance of tasks; Classification loss Weighted cross-entropy is used to address the imbalance in surrounding rock grades; let the number of categories be C, the true label of the sample be y, the predicted probability be p, and the weights be... Calculation based on category frequency: ; In the formula, N is the total number of samples, and Nc is the number of samples in class c; this is used to improve the detection sensitivity of minority classes. Abnormal loss L anom Based on mean squared error, outlier scores are encouraged to be close to their binary labels, where 0 is normal and 1 is abnormal. ; in, It predicts outlier scores. These are true anomaly labels (derived from geometric constraint violations); the loss function introduces a regularization term. This is used to ensure confidence levels, penalize excessively high entropy in predicted probabilities, and ensure that low-confidence labels only have adjacent levels with similar confidence levels. ; In the formula, It is a probability distribution The entropy is τ, the threshold (e.g., 0.7), I is the indicator function, and λ is the regularization coefficient; the formula enables the model to distribute the probability quality across adjacent levels when there is uncertainty, avoiding overconfidence. The overall loss function is optimized through gradient descent to improve the balance between accuracy and confidence in the model.

[0032] The input data from the multi-source sensors described in S1 includes real-time TBM sensor data and historical logs. The sensor data is divided into shield data and operational data. Shield data includes displacement and pressure, while operational data includes thrust and velocity. Shield data has spatiotemporal characteristics: spatially, the sensors are distributed in a circle, with displacement values ​​ranging from 0-20 mm and pressure values ​​ranging from 0-40 MPa, and a sampling period of 2 seconds; temporally, the data is a non-stationary time series, affected by geological changes. Operational data serves as an auxiliary feature for contextual modeling. It also includes data processing of the input data from the multi-source sensors, including the following steps: a. Data cleaning: Outlier detection is performed using a sliding window, and outliers are removed using the 3σ principle; for missing values, time series interpolation is used for imputation, with the following formula: ; Where, x t These are the missing values ​​at time t; b. Normalization: Z-score standardization is used to scale the features to zero mean and unit variance; for feature x, the processed value x′ is: ; Where μ is the sum of the means and σ is the standard deviation; c. Serialization: The continuous data is divided into time windows with a window length of L=60, corresponding to 2 minutes, and a step size of S=30, corresponding to 1 minute; each window contains readings from multiple sensors, forming a graph sequence.

[0033] The output surrounding rock stability level described in S4 is an integer from 1 to 5, representing the surrounding rock condition, with 1 being the most stable. The anomaly score is a continuous value between 0 and 1, with a higher score indicating a greater anomaly. The results of the geometric constraint check include displacement asymmetry, in millimeters; The output data is encapsulated in JSON format, including timestamps and confidence levels, making it easy to integrate into monitoring systems.

[0034] In S5: Feedback and Optimization, the training process of the model is divided into four stages: data preparation, model initialization, iterative training, and validation; the entire process processes data in batches. First, samples are loaded from historical data to construct training and validation sets; each sample contains a graph sequence and label for a time window; model parameters are initialized using Xavier to ensure gradient stability. During the training iteration, for each batch with a batch size of 32, the following steps are performed: a. Forward propagation: Input graph sequence, extract spatial features through graph neural network layers; for each time step, graph neural network aggregates neighbor information and outputs node embeddings; then, long short-term neural network processes the time series to generate hidden states; b. Loss Calculation: Compare the predicted output with the true label to calculate the total loss; backpropagate the loss value and perform gradient clipping with a norm threshold of 1.0 to prevent explosion. c. Parameter update: The optimizer adjusts the weights, and the learning rate is dynamically reduced based on the validation loss; Training lasts for 100 epochs, with an early stopping mechanism terminating training when the validation loss no longer improves; the validation phase uses an independent dataset to evaluate the model's generalization ability; after training, the model is deployed to a real-time system to process streaming data. The entire training process ensures that the model learns spatiotemporal patterns from the data and enhances physical plausibility through geometric constraints.

[0035] Although some preferred embodiments of the invention 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 invention.

[0036] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An anomaly detection method for a TBM shield system based on a spatiotemporal graph neural network, wherein the shield sensor network is modeled as a graph structure, where nodes represent sensors, including displacement and pressure, and edges represent the spatial relationships between sensors, including geometric adjacency; characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing; Real-time acquisition of multi-source sensor data from the TBM monitoring system, including shield displacement, pressure, and operating parameters, including thrust and torque; Data cleaning, normalization, and serialization to form time window samples; S2: Graph structure construction; Define the connection relationship between sensor nodes based on the geometric position of the shield on the circumference; Edge weights are calculated based on angle differences, the smaller the angle difference, the stronger the connection, to reflect spatial proximity; S3: Spatiotemporal feature extraction; Graph neural networks are used to process the graph data at each time step, aggregate neighbor node information, and extract spatial features. Then, the output of the GNN on the time series is input into the Long Short-Term Memory (LSTM) network to capture the time dependencies; S4: Anomaly detection and classification; weighted key time steps through an attention mechanism to output surrounding rock stability level and anomaly score; anomaly detection integrates model prediction and geometric constraint checks, symmetry verification, and generates decision suggestions; S5: Feedback and Optimization; The system automatically adjusts the threshold and updates the model parameters based on the detection results to achieve adaptive learning; The process executes S1 to S5 in a loop to achieve real-time monitoring. Each step processes the data in the current time window, outputs abnormal statuses, and feeds the results back to the TBM operator, forming a closed-loop control.

2. The anomaly detection method for TBM shield system based on spatiotemporal graph neural network as described in claim 1, characterized in that: The geometric position of the shields on the circumference described in S2 is as follows: the top shield corresponds to 0 degrees, and the left shield corresponds to 90 degrees.

3. The anomaly detection method for TBM shield system based on spatiotemporal graph neural network as described in claim 1, characterized in that: The graph neural network described in S3 includes a graph convolution module, a time series module, an attention module, and an output module; The graph convolution module uses a graph attention network as the basic graph neural network layer; the graph structure has N nodes, and the initial feature of each node is a scalar. The GAT layer calculates the weights between nodes through an attention mechanism; for nodes i and j, the attention coefficient α ij The calculation formula is: ； Among them, h i and h j Here, W is the node feature, a is the weight matrix, a is the attention vector, ∥ represents the concatenation operation, and N(i) is the set of neighbors of node i. The formula allows the model to adaptively learn edges. The network contains multiple layers of GAT, and the output of each layer is activated by the Leaky ReLU activation function and Dropout regularization to prevent overfitting. The time series module uses a bidirectional LSTM to process the time series output by graph convolution; assuming a time step size of T, the node features at each time step are global pooled and then used as the input to the LSTM; the hidden state update formula of the LSTM is: ； Where, x t It is the input for time step t, h t It is a hidden state; bidirectional LSTM captures context from both the front and back directions, enhancing temporal modeling; The attention module is used to apply a multi-head attention mechanism to the LSTM output and calculate the weights between time steps; for sequence representation The attention output is: ； Where Q is the query, K is the key, V is the value matrix, and d k It is the dimension scaling factor; The output module includes a classification head and an anomaly detection head; the classification head predicts the surrounding rock grade through a fully connected layer, and the anomaly detection head outputs anomaly scores using a Sigmoid function; the optimizer uses Adam W, combined with weight decay, with the learning rate set to 0.001, and is dynamically adjusted through the Reduce LR On Plateau scheduler. The entire network structure is trained end-to-end to ensure the fusion of spatial and temporal features.

4. The anomaly detection method for TBM shield system based on spatiotemporal graph neural network as described in claim 1, characterized in that: The loss function of the graph neural network described in S3 is designed in a multi-task form to balance classification accuracy and anomaly detection reliability; the overall loss function... It is classification loss and abnormal losses Weighted sum: ； Here, α and β are weighting coefficients, with α set to 1 and β set to 0.5 to balance the importance of tasks; Classification loss Weighted cross-entropy is used to address the imbalance problem of surrounding rock grades; let the number of categories be C, the true label of the sample be y, the predicted probability be p, and the weights be... Calculation based on category frequency: ； In the formula, N is the total number of samples, and Nc is the number of samples in class c; this is used to improve the detection sensitivity of minority classes. Abnormal loss L anom Based on mean squared error, outlier scores are encouraged to be close to their binary labels, where 0 is normal and 1 is abnormal. ； in, It predicts outlier scores. These are true anomaly labels (derived from geometric constraint violations); the loss function introduces a regularization term. This is used to ensure confidence levels, penalize excessively high entropy in predicted probabilities, and ensure that low-confidence labels only have adjacent levels with similar confidence levels. ； In the formula, It is a probability distribution The entropy is τ, the threshold (e.g., 0.7), I is the indicator function, and λ is the regularization coefficient; the formula enables the model to distribute the probability quality across adjacent levels when there is uncertainty, avoiding overconfidence. The overall loss function is optimized through gradient descent to improve the balance between accuracy and confidence in the model.

5. The anomaly detection method for TBM shield system based on spatiotemporal graph neural network as described in claim 1, characterized in that: The input data from the multi-source sensors described in S1 includes real-time TBM sensor data and historical logs. The sensor data is divided into shield data and operational data. Shield data includes displacement and pressure, while operational data includes thrust and velocity. Shield data has spatiotemporal characteristics: spatially, the sensors are distributed in a circle, with displacement values ​​ranging from 0 to 20 mm and pressure values ​​ranging from 0 to 40 MPa, and a sampling period of 2 seconds; temporally, the data is a non-stationary time series, affected by geological changes. Operational data serves as an auxiliary feature for contextual modeling. It also includes data processing of the input data from the multi-source sensors, including the following steps: a. Data cleaning: Outlier detection is performed using a sliding window, and outliers are removed using the 3σ principle; for missing values, time series interpolation is used for imputation, with the following formula: ； Where, x t These are the missing values ​​at time t; b. Normalization: Z-score standardization is used to scale the features to zero mean and unit variance; for feature x, the processed value x′ is: ； Where μ is the sum of the means and σ is the standard deviation; c. Serialization: The continuous data is divided into time windows with a window length of L=60, corresponding to 2 minutes, and a step size of S=30, corresponding to 1 minute; each window contains readings from multiple sensors, forming a graph sequence.

6. The anomaly detection method for TBM shield system based on spatiotemporal graph neural network as described in claim 1, characterized in that: The output surrounding rock stability level described in S4 is an integer from 1 to 5, representing the surrounding rock condition, with 1 being the most stable. The anomaly score is a continuous value between 0 and 1, with a higher score indicating a greater anomaly. The results of the geometric constraint check include displacement asymmetry, in millimeters; The output data is encapsulated in JSON format, including timestamps and confidence levels, making it easy to integrate into monitoring systems.

7. The anomaly detection method for TBM shield system based on spatiotemporal graph neural network as described in claim 1, characterized in that: In S5: Feedback and Optimization, the training process of the model is divided into four stages: data preparation, model initialization, iterative training, and validation; the entire process processes data in batches. First, samples are loaded from historical data to construct training and validation sets; each sample contains a graph sequence and label for a time window; model parameters are initialized using Xavier to ensure gradient stability. During the training iteration, for each batch with a batch size of 32, the following steps are performed: a. Forward propagation: Input graph sequence, extract spatial features through graph neural network layers; for each time step, graph neural network aggregates neighbor information and outputs node embeddings; then, long short-term neural network processes the time series to generate hidden states; b. Loss Calculation: Compare the predicted output with the true label and calculate the total loss; The loss value is backpropagated and clipped using gradient pruning, with a norm threshold of 1.0 to prevent explosion. c. Parameter update: The optimizer adjusts the weights, and the learning rate is dynamically reduced based on the validation loss; Training lasts for 100 epochs, with an early stopping mechanism terminating training when the validation loss no longer improves; the validation phase uses an independent dataset to evaluate the model's generalization ability. After training, the model is deployed to a real-time system to process streaming data; The entire training process ensures that the model learns spatiotemporal patterns from the data and enhances physical plausibility through geometric constraints.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7.

Images