Shield tunneling machine attitude prediction method and system

Abstract

The invention discloses a shield tunneling machine attitude prediction method and system, and relates to the field of shield tunneling machine construction.The shield tunneling machine attitude prediction method comprises the steps that firstly, shield tunneling machine construction parameters, geological environment parameters and tunnel planning parameters are obtained and preprocessed, and related characteristics are extracted; secondly, mapping the extracted features into nodes by using a heterogeneous graph model, and updating node features in real time through weighted processing of a graph attention mechanism; then, the node features and the connection relation are input into the trained hybrid drive prediction model, and time sequence, physical and space prediction values of the attitude angle of the shield tunneling machine are obtained; and finally, constructing a Bayesian causal graph to quantify the influence of each factor on the attitude angle, carrying out dynamic weighted fusion based on a quantization result, and outputting a final predicted value and a confidence interval. The problem that the prediction effect is poor under the complex geological condition due to the fact that prediction only depends on single shield tunneling machine parameters in the prior art is solved, and accurate prediction of the attitude angle of the shield tunneling machine can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of shield machine construction, and in particular to a shield machine posture prediction method and system. Background Art

[0002] Shield tunneling is widely used in urban subway construction due to its advantages, including high excavation speeds, minimal impact on ground traffic, and minimal harm to the natural environment. The shield machine works by tunneling forward along its designed axis, excavating the soil along the way while relying on a cylindrical shield on the outside of the shield body to withstand tremendous pressure and protect the stability of the stratum. However, the shield construction environment is complex and changeable, especially in complex geological conditions such as soft soil, strata with high water content, and rock formations. Controlling the shield machine's excavation posture becomes particularly difficult, making shield machine posture prediction a key step in the shield construction process.

[0003] During the construction of a shield machine, changes in geological conditions, construction parameters, and tunnel design will affect attitude prediction. However, existing shield machine attitude prediction methods mostly rely on a single shield machine parameter or perform simple time series prediction. Simple parameter prediction cannot fully reflect this complex relationship. For example, the shield machine attitude prediction method, device, terminal equipment, and storage medium disclosed in CN112100841A only use GRU to predict the pitch angle, yaw angle, and roll angle respectively, ignoring the dynamic coupling relationship between geological conditions, construction parameters, and tunnel design. The prediction effect is significantly reduced in complex scenarios such as soft and hard uneven strata and caves.

[0004] Therefore, a shield machine posture prediction method and system are needed to solve the above problems. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention discloses a shield machine posture prediction method and system. By integrating shield machine construction parameters, geological environment parameters and tunnel planning parameters, and utilizing heterogeneous graphs, graph attention mechanism, incremental learning, hybrid drive prediction model and Bayesian causal graph technology, accurate prediction of the shield machine posture angle is achieved, and a confidence interval of the prediction result is given.

[0006] The present invention adopts the following technical solutions:

[0007] A shield machine posture prediction method comprises the following steps:

[0008] Step 1: Obtain shield machine construction parameters, geological environment parameters, and tunnel planning parameters, and preprocess the obtained parameters to extract construction time sequence characteristics, geophysical characteristics, and tunnel spatial characteristics from the preprocessed shield machine construction parameters, geological environment parameters, and tunnel planning parameters;

[0009] Step 2: The extracted construction time sequence features, geophysical features, and tunnel spatial features are input into the heterogeneous graph to map shield machine nodes, stratum nodes, and tunnel axis nodes. The shield machine nodes, stratum nodes, and tunnel axis nodes are weighted through the graph attention mechanism to form interactive connections, and the node features are updated in real time using incremental learning.

[0010] Step 3: Input the node features and node connection relationships of the heterogeneous graph into the trained hybrid drive prediction model to obtain the time series prediction value, physical prediction value, and spatial prediction value of the shield machine attitude angle;

[0011] Step 4: Quantify the impact of construction timing, geophysical and tunnel spatial factors on the current shield machine attitude angle by constructing a Bayesian causal diagram, and dynamically weighted fuse the different prediction values obtained based on the quantified results to output the final prediction value and confidence interval of the shield machine attitude angle.

[0012] Furthermore, in step one, the shield machine construction parameters include at least the propulsion speed, cutterhead speed, soil bin pressure and excavation thrust, the geological environment parameters include at least the soil type, soil moisture content, groundwater level and stratum information, and the tunnel planning parameters include at least the tunnel design axis, slope and turning radius. The sliding window method is used to filter out invalid data in the acquired parameters and perform normalization processing.

[0013] Furthermore, in step 1, a three-layer deep learning model is used to extract construction time sequence features, geophysical features, and tunnel spatial features. The working method of the three-layer deep learning model includes the following steps:

[0014] S101. Separating temporal features and spatial features in shield machine construction parameters through a gating mechanism. The gating mechanism dynamically adjusts the extraction ratio of temporal features and spatial features based on a long short-term memory network, and separates the temporal features and spatial features in the shield machine construction parameters. The construction temporal features are used to predict the change trend of the shield machine posture over time, and the construction spatial features are used to analyze the posture differences of the shield machine at different positions.

[0015] S102, using a dynamic graph convolutional network to mine the spatial correlation between the geological area and the tunnel area, mapping the geological area and the tunnel planning area as nodes of a dynamic graph, and mapping the spatial correlation between the geological environment parameters and the tunnel planning parameters as edges of the dynamic graph, wherein the dynamic graph convolutional network extracts the spatial correlation features of the geological environment parameters and the tunnel planning parameters as geophysical features and tunnel spatial features respectively through multi-layer convolution operations;

[0016] S103. Use the Voronoi subdivision principle to divide the geological area into geological units of different scales. Use the points on the tunnel axis as seed points and subdivide the geological area according to different radii to obtain geological units of different sizes. Construct a multi-scale geological unit coupling model to couple the geological unit information of different scales. During the coupling process, explore the construction space characteristics, the relative position relationship between the tunnel planning area and the geological units.

[0017] Furthermore, in step two, the shield machine node and the stratum node are connected according to the current stratum position of the shield machine, the shield machine node and the tunnel axis node are connected according to the spatial relationship between the current position of the shield machine and the tunnel axis node, the stratum unit node and the tunnel axis node are connected according to the constraint relationship of the stratum on the tunnel axis, and the stratum unit nodes are connected according to the spatial adjacent relationship of the strata.

[0018] Furthermore, the graph attention mechanism uses type-level attention to define the interaction weights between nodes of different types. The function expression is:

[0019] In formula (1), β M,S represents the interaction weight between the shield machine node and the stratum node, q M represents the query vector of the shield machine node, represents the transpose of the query vector of the shield machine node, M represents the shield machine node, k S Represents the key vector of the stratum node, S represents the stratum node, d represents the feature dimension of the key vector, which is used to scale the dot product result, and softmax is the normalization function used to map the dot product result to a probability distribution

[0020] Based on type attention, instance-level attention is used to calculate the weight distribution within nodes of the same type. The function expression is:

[0021]

[0022] In formula (2), M i,j represents the instance-level attention weight of shield machine node i and node j, a represents the attention vector, a T represents the transpose of the attention vector, W represents the feature transformation matrix, Represents the concatenation operation, which is used to fuse the features of nodes i and j. i and M j They represent the i-th and j-th TBM nodes respectively, and LeakyReLU represents the activation function.

[0023] Furthermore, the hybrid drive prediction model includes a timing branch, a physical branch and a spatial branch. The timing branch, the physical branch and the spatial branch use parallel computing to output three types of prediction values. The timing branch gradually processes the input construction timing features through a multi-layer long short-term memory network stack, updates the hidden state at each time step to output the timing prediction value of the shield machine attitude angle. The physical branch calculates the attitude angle change rate based on the mechanism model, and uses a multi-layer perceptron to learn the nonlinear mapping relationship between geological physical characteristics and attitude angles to output the physical prediction value of the shield machine attitude angle. The spatial branch uses a convolutional neural network to perform convolution and pooling operations on the input tunnel space features, and captures the spatial structure information in the relative position relationship between the construction space characteristics, the tunnel planning area and the geological unit to output the spatial prediction value of the shield machine attitude angle.

[0024] Furthermore, in step four, the cause nodes of the Bayesian causal diagram are construction timing, geophysical and tunnel space factors, and the result node of the Bayesian causal diagram is the shield machine attitude angle. The probability state of each node is updated by calculating the posterior probability distribution to obtain the quantitative results of the influence of each factor on the shield machine attitude angle, and the adaptive gating mechanism is used to dynamically weighted fuse the predicted values to output the final attitude angle and confidence interval.

[0025] Furthermore, a shield machine posture prediction system is provided, the shield machine posture prediction system comprising:

[0026] The data acquisition module uses a sensor interface to connect to the shield machine controller, gyroscope, and geological radar, and connects to the BIM model library and geological exploration report parsing engine to obtain shield machine construction parameters, geological environment parameters, and tunnel planning parameters. The data rule engine is used to clean and normalize the acquired data.

[0027] The feature extraction module uses a three-layer deep learning model to extract construction timing features, geophysical features, and tunnel spatial features from the preprocessed data. The extracted construction timing features, geophysical features, and tunnel spatial features are mapped to form a heterogeneous graph of shield machine nodes, stratum nodes, and tunnel axis nodes. Dynamic weighted interactive connections between nodes are performed through a graph attention mechanism.

[0028] The distributed prediction module obtains the time series prediction value, physical prediction value and spatial prediction value of the shield machine attitude angle in a distributed manner through the trained hybrid drive prediction model;

[0029] The prediction module is integrated to construct a Bayesian causal diagram to quantify the impact of construction timing, geophysical and tunnel spatial factors on the shield machine attitude angle. Based on the quantified results, different prediction values are dynamically weighted and fused to output the final prediction value and confidence interval of the shield machine attitude angle.

[0030] The beneficial effects of the present invention are:

[0031] 1. The present invention combines shield machine construction parameters, geological environment parameters and tunnel planning parameters, extracts construction timing characteristics, geophysical characteristics and tunnel space characteristics from different dimensions, comprehensively considers various factors affecting the shield machine posture, avoids the information loss caused by a single data source, fully reflects the complexity of shield machine posture changes, and makes the prediction results more accurate and comprehensive.

[0032] 2. The present invention inputs the extracted features into a heterogeneous graph to form shield machine nodes, stratum nodes and tunnel axis nodes, and performs weighted processing through a graph attention mechanism to form interactive connections. The heterogeneous graph structure can intuitively represent the relationship between different factors. The graph attention mechanism can dynamically allocate weights according to the degree of correlation between nodes, thereby enhancing the model's ability to capture complex relationships. It also uses incremental learning to update node features in real time, enabling the model to adapt to the ever-changing situation during shield machine construction.

[0033] 3. The present invention inputs the node features and node connection relationships of the heterogeneous graph into the trained hybrid drive prediction model to obtain the time series prediction value, physical prediction value and spatial prediction value of the shield machine attitude angle. This multi-model prediction method predicts the shield machine attitude from different angles, makes full use of various feature information, and improves the accuracy and reliability of the prediction. By constructing a Bayesian causal graph to quantify the influence of construction timing, geophysical and tunnel space factors on the current shield machine attitude angle, and dynamically weighted fusion of the different prediction values obtained based on the quantification results. The Bayesian causal graph can clearly show the causal relationship between the various factors, and the dynamic weighted fusion reasonably allocates weights according to the actual influence of each factor, so that the final prediction value is more in line with the actual situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0035] Figure 2 Schematic diagram of the process of the three-layer deep learning model in the present invention;

[0036] Figure 3 This is the overall architecture diagram of the present invention. DETAILED DESCRIPTION

[0037] The following is a combination of the embodiments of the present invention Figure 1 To the attached Figure 3The technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0038] The embodiment of the present invention discloses a shield machine posture prediction method, as shown in the attached Figure 1 , including the following steps:

[0039] Step 1: Parameter acquisition and feature extraction

[0040] Parameter acquisition

[0041] Shield machine construction parameters: Data is collected from the shield machine's control system and sensors. For example, a speed sensor installed on the shield machine measures propulsion speed; a rotation speed sensor measures cutterhead speed; a pressure sensor monitors soil bin pressure; and a thrust sensor records tunneling thrust. This data is recorded in a time series format, reflecting the real-time status of the shield machine during construction.

[0042] Geological environmental parameters: These are obtained through a combination of geological survey reports and on-site sampling and analysis. Before construction, the geological exploration team will conduct a detailed geological survey along the tunnel route to obtain information such as soil type (such as clay, sand, rock, etc.), soil moisture content, and groundwater level. Stratum information, including thickness, hardness, and bedding structure, is determined through geological drilling and geophysical exploration.

[0043] Tunnel planning parameters: These are obtained from the tunnel design documentation. The design documentation specifies the tunnel's design axis, which defines the tunnel's orientation and spatial location. Slope information describes the tunnel's longitudinal inclination. Turning radius determines the tunnel's curvature in plan.

[0044] Data preprocessing

[0045] Sliding window method to filter invalid data: For each parameter sequence, a sliding window size is set (for example, 5 data points as a window). Within the window, the data is checked to see if it exceeds the acceptable range (upper and lower limits are set based on engineering experience and parameter characteristics). If the majority of data points within the window (e.g., more than 80%) are outside the acceptable range, these data points are considered invalid and removed from the sequence. For example, propulsion speed should normally be within the range of [0, 5] m / min. If the majority of data points within the window are greater than 5 m / min, these data points may be invalid data caused by sensor failure.

[0046] The filtered data is normalized so that its value is mapped to the interval [0,1]. The normalization formula is: X is the normalized parameter, x is the original parameter, and x max and x min are the minimum and maximum values of the parameter sequence respectively.

[0047] Feature extraction

[0048] Construction time series characteristics: Analyze how construction parameters change over time. For example, calculate the mean and variance of the advancement speed over a period of time to reflect the stability and fluctuation of the advancement speed; calculate the trend of the cutterhead speed (such as the slope of the linear regression) to determine whether the cutterhead speed is increasing, decreasing, or remaining stable.

[0049] Geophysical characteristics: Extract characteristics reflecting the physical properties of the stratum based on geological environmental parameters. For example, soil type and moisture content can be used to calculate the softness and hardness of the stratum; groundwater level and stratum information can be used to assess the permeability and stability of the stratum.

[0050] Tunnel spatial characteristics: Extract the spatial geometric characteristics of the tunnel from the tunnel planning parameters. For example, the curvature of the tunnel design axis is calculated to reflect the degree of curvature of the tunnel; the rate of change of the slope is calculated to understand the longitudinal undulation of the tunnel.

[0051] Step 2: Heterogeneous graph construction and node feature update

[0052] Node Mapping: The extracted construction time series characteristics, geophysical characteristics, and tunnel spatial characteristics are mapped to the shield machine node, stratum node, and tunnel axis node, respectively. For example, construction time series characteristics such as the mean and variance of the advance speed are mapped to the shield machine node; geophysical characteristics such as the hardness and permeability of the stratum are mapped to the stratum node; and tunnel spatial characteristics such as the curvature of the tunnel design axis and the rate of change of the slope are mapped to the tunnel axis node.

[0053] Node connections: Establish connections between nodes based on their actual relationships. There's an interactive relationship between the shield machine node and the stratum node, as the shield machine comes into contact with the stratum during advancement, and the physical properties of the stratum affect the machine's advancement. The tunnel axis node is also associated with the shield machine node and the stratum node, defining the shield machine's route and the geological environment it navigates.

[0054] For each node, we calculate its attention coefficient with other nodes to measure the degree of connection between nodes. The calculation of the attention coefficient can take into account factors such as node feature similarity and spatial distance. For example, if two nodes have high feature similarity and are spatially close, the attention coefficient between them will be large.

[0055] The connections between nodes are weighted according to the attention coefficient, giving more weight to important connections. This can highlight factors that have a greater impact on the TBM posture. For example, when the TBM passes through the boundary between soft and hard strata, the weight of the connection between the stratum node and the TBM node will increase.

[0056] During shield machine construction, new construction parameters and geological environment parameters are monitored in real time. For example, sensors are used to obtain real-time information on changes in construction parameters such as propulsion speed and cutterhead speed; and equipment such as geological radar is used to monitor changes in the ground in real time.

[0057] New data is fed into the model, and an incremental learning algorithm is used to update node features. This algorithm continuously adjusts node feature representations based on new data, allowing the model to adapt to changes during construction. For example, when a shield machine encounters a new stratum, the features of the stratum node are updated based on the new geological data.

[0058] Step 3: Hybrid drive prediction of shield machine attitude angle

[0059] Select an appropriate hybrid-driven prediction model that combines the advantages of neural networks and physical models. Neural networks can learn complex nonlinear relationships in data, while physical models can introduce domain knowledge to improve the interpretability and accuracy of the model.

[0060] The hybrid drive prediction model is trained using historical data. This data includes shield machine construction parameters, geological environment parameters, tunnel planning parameters, and the corresponding actual values of the shield machine's attitude angles. By continuously adjusting the model parameters, the error between the model's predicted values and the actual values is minimized.

[0061] The updated heterogeneous graph node features and node connection relationships are input into the trained hybrid-driven prediction model.

[0062] The model outputs time-series predictions, physical predictions, and spatial predictions of the shield machine's attitude angle. The time-series predictions primarily reflect the changing trends of the shield machine's attitude over time; the physical predictions account for the impact of geophysical factors on the machine's attitude; and the spatial predictions predict the machine's attitude based on the tunnel's spatial characteristics.

[0063] Step 4: Dynamic weighted fusion of predicted values and result output

[0064] Analyze the causal relationship between construction sequence, geophysical factors, and tunnel spatial factors and the shield machine's attitude angle. For example, changes in construction sequence (such as a sudden change in propulsion speed) may lead to adjustments in the shield machine's attitude; geophysical conditions (such as the softness or hardness of the stratum) affect the shield machine's propulsion resistance, which in turn affects its attitude; and tunnel spatial characteristics (such as the curvature of the tunnel) can restrict the shield machine's travel direction, affecting its attitude.

[0065] Using methods such as Bayesian networks, a Bayesian causal diagram was constructed to quantify the causal strength between various factors. By collecting a large amount of historical data, the probability of each factor affecting the shield machine attitude angle under different circumstances was statistically analyzed to determine the causal strength.

[0066] The Bayesian causal diagram quantifies the influence of construction sequence, geophysical factors, and tunnel spatial factors on the current shield machine attitude angle. For example, if the shield machine is currently passing through the boundary between soft and hard strata, the influence of geophysical factors will increase; if the tunnel curvature is large, the influence of tunnel spatial factors will increase.

[0067] Based on the quantization results, the time series prediction value, physical prediction value and spatial prediction value are dynamically weighted and fused. The fusion formula is: P = w M p M +w S p S +w H p H , P is the final predicted value, p M 、p S and p H are respectively the time series prediction value, physical prediction value and spatial prediction value, w M 、w S and w H The corresponding weight, and w M +w S +w H = 1, the final predicted value of the shield machine attitude angle is output. Simultaneously, a confidence interval for the final predicted value is calculated based on the model's prediction error and uncertainty. The confidence interval reflects the reliability of the prediction results, providing more comprehensive information for engineering decision-making. For example, a narrow confidence interval indicates a relatively reliable prediction; a wide confidence interval indicates that further data and model analysis is needed to improve prediction accuracy.

[0068] In step 1, a three-layer deep learning model is used to extract construction time sequence features, geophysical features, and tunnel spatial features. The working method of the three-layer deep learning model is shown in the attached figure. Figure 2 As shown, the following steps are included:

[0069] S101: Separating construction timing and spatial characteristics based on a gating mechanism

[0070] Collect shield machine construction parameters, such as propulsion speed, cutterhead speed, soil bin pressure, and tunneling thrust. These parameters are recorded in a time series format, with each data point corresponding to a specific spatial position of the shield machine in the tunnel. The data is normalized so that each parameter value has the same scale, for example, using Z-score normalization: Where Y is the standardized data, y is the original data, μ is the mean, and σ is the standard deviation.

[0071] A gating mechanism is built based on the Long Short-Term Memory (LSTM) network. The input gate, forget gate, and output gate in the LSTM can control the flow of information and dynamically adjust the extraction ratio of temporal features and spatial features.

[0072] The shield machine construction parameters are input into the LSTM, and the input information is filtered through a gating mechanism. For example, the input gate determines which new information can enter the cell state, the forget gate determines which information in the cell state needs to be forgotten, and the output gate determines which information in the cell state can be output.

[0073] After LSTM processing, the output is divided into two parts: one part, as construction time series features, is used to predict the change trend of the shield machine's posture over time; the other part, as construction space features, is used to analyze the differences in the shield machine's posture at different locations. For example, the construction time series features can reflect the changes in advance speed at different time points, while the construction space features can reflect the posture characteristics of the shield machine at different cross-sections of the tunnel.

[0074] S102: Exploring the spatial correlation between geological environment parameters and tunnel planning parameters

[0075] The geological area and tunnel planning area are mapped as nodes of the dynamic graph. For example, each geological unit obtained from geological exploration is a node, and each key location in the tunnel planning (such as turning points and slope change points) is also a node.

[0076] The spatial correlation between geological environment parameters and tunnel planning parameters is mapped as edges in the dynamic graph. For example, if two geological units are spatially adjacent and have similar geological properties, there will be an edge between them; if two key locations in the tunnel planning are spatially close and have similar effects on the shield machine posture, there will also be an edge between them.

[0077] The dynamic graph convolutional network (DGCN) is used to process dynamic graphs. DGCN operates through multiple layers of convolution. In each layer, a node aggregates information about its neighboring nodes. For example, for a geological unit node, it aggregates geological environment parameter information from the surrounding adjacent geological unit nodes.

[0078] After multiple layers of convolution, spatial correlation features of geological environment parameters and tunnel planning parameters are extracted. These features are respectively used as geophysical features and tunnel spatial features. For example, geophysical features can reflect the distribution and interaction of different strata in a geological region, while tunnel spatial features can reflect the impact of the spatial layout in tunnel planning on the shield machine's posture.

[0079] S103: Capturing the relative positional relationship between construction space characteristics, tunnel planning areas, and geological units

[0080] The Voronoi mesh principle is used to divide the geological area into geological units of varying scales. Using the points on the tunnel axis as seed points, the geological area is meshed according to different radii. For example, a smaller radius yields finer-grained geological units, reflecting local geological variations; a larger radius yields coarser-grained geological units, reflecting overall geological trends.

[0081] A multi-scale geological unit coupling model is constructed to couple geological unit information at different scales. During the coupling process, the relative position relationship between the construction space characteristics, the tunnel planning area, and the geological units is taken into consideration.

[0082] For example, for a finer-grained geological unit, its relative distance and direction to a key location in tunnel planning, as well as the distribution of construction space characteristics within the geological unit, are analyzed; at the same time, the information of the fine-grained geological unit is integrated with the information of the coarser-grained geological unit to which it belongs, comprehensively considering the impact of geological factors of different scales on the shield machine posture.

[0083] Through this coupling method, the information of the geological area can be more comprehensively mined, the accuracy of geophysical feature extraction can be improved, and a richer geological basis can be provided for subsequent shield machine posture prediction.

[0084] In step 2, the connection between the TBM node and the stratum node uses the TBM's built-in positioning system (such as the Global Positioning System (GPS) and the Inertial Navigation System (INS)) and geological exploration data to accurately determine the TBM's current stratum position. For example, by scanning the stratum in front of the TBM with geological radar and combining it with a pre-drawn geological profile, it is clear which stratum (such as sand, clay, or rock) the TBM is currently located in.

[0085] After determining the stratum in which the TBM is currently located, the TBM node is connected to the corresponding stratum node. This connection can be weighted based on the stratum's impact on the TBM. For example, if the TBM is currently located in a soft stratum (such as a mud layer), which offers significant resistance to the TBM's propulsion and significantly affects its posture, the connection weight between the TBM node and the stratum node can be set higher. If the TBM is located in a hard stratum (such as a granite layer), the connection weight can be relatively low.

[0086] As the shield machine advances, its position in the stratum will constantly change. Real-time monitoring of the shield machine's position information is performed. When the shield machine enters a new stratum, the connection with the old stratum node is promptly disconnected and a connection with the new stratum node is established. The connection weight is adjusted according to the characteristics of the new stratum.

[0087] The connection between the TBM node and the tunnel axis node uses the TBM positioning data and tunnel design files to calculate the spatial distance and direction relationship between the TBM's current position and the tunnel axis node. For example, using a 3D coordinate system, the horizontal and vertical distances from the TBM's current position to the tunnel axis node, as well as the angle of deviation from the tunnel axis, are determined.

[0088] The connection is established based on the spatial relationship between the TBM and the tunnel axis. If the TBM deviates from the tunnel axis by a small distance, indicating good consistency between the TBM posture and the tunnel axis, the connection weight can be set to medium. If the deviation is large, the connection weight can be adjusted based on the degree of deviation. The greater the deviation, the smaller the weight (indicating that the deviation between the TBM current posture and the tunnel axis requires special attention), or a special mark can be used in the connection method to indicate that the TBM posture needs to be adjusted to return to the tunnel axis.

[0089] As the shield machine advances, its spatial relationship with the tunnel axis is continuously monitored. When the shield machine's position changes, the connection method and weight between the shield machine node and the tunnel axis node are updated in real time to ensure the accuracy and effectiveness of the connection.

[0090] The connection between stratigraphic unit nodes and tunnel axis nodes is based on geological exploration data and tunnel design parameters to investigate the constraints imposed by the stratigraphic elements on the tunnel axis. For example, certain hard stratigraphic elements may limit the curvature of the tunnel, keeping the tunnel axis as straight as possible within these stratigraphic sections; however, soft stratigraphic elements may allow for greater curvature of the tunnel axis. Furthermore, considering stratigraphic stability, unstable stratigraphic elements may adversely affect the stability of the tunnel axis.

[0091] Based on the constraints imposed by the strata on the tunnel axis, connections are established between stratum unit nodes and tunnel axis nodes. For strata with strong constraints, the connection weight can be set higher, indicating a significant restriction on the tunnel axis. For strata with weaker constraints, the connection weight can be set lower. For example, in rocky strata, where the tunnel axis is difficult to change direction, the connection weight between the stratum unit node and the tunnel axis node can be set to 0.8. In loose sandy strata, where the tunnel axis has some room for adjustment, the connection weight can be set to 0.5.

[0092] When geological conditions change, such as encountering special geological structures such as faults and caves, the constraint relationship of the stratum on the tunnel axis is re-evaluated, and the connection method and weight between the stratum unit node and the tunnel axis node are adjusted accordingly.

[0093] The connections between stratigraphic unit nodes are based on geological exploration data and stratigraphic distribution maps to clarify the spatial adjacency between stratigraphic units. For example, a geological cross-section can clearly show the vertical superposition or horizontal adjacency of different stratigraphic units.

[0094] Connect spatially adjacent stratigraphic unit nodes. The connection method can be determined based on the contact relationship between the stratigraphic units (e.g., conformable contact, unconformable contact) and geological properties (e.g., permeability, strength, etc.). If the stratigraphic units have close contact and similar geological properties, the connection weight can be set higher; if the contact relationship is complex or the geological properties are significantly different, the connection weight can be relatively low.

[0095] Under the influence of geological evolution and tectonic movement, the spatial relationship between stratigraphic units may change. The connections between stratigraphic unit nodes are regularly checked and adjusted based on the latest geological research results to ensure the accuracy of the connections.

[0096] The established node connections are weighted using a graph attention mechanism. This mechanism automatically learns the weight of each connection based on the node's characteristics and connection relationships. For example, for the connection between a TBM node and a stratum node, the graph attention mechanism dynamically adjusts the connection weight to better reflect the mutual influence between them, taking into account factors such as the TBM's construction parameters and the stratum's physical properties.

[0097] During shield machine construction, an incremental learning algorithm is used to update node connections in real time as new data and information are acquired. For example, when new geological structures are discovered or the shield machine's posture undergoes significant changes, the connection methods and weights between related nodes are promptly adjusted so that the heterogeneous graph can accurately reflect the current construction status and geological environment.

[0098] The graph attention mechanism uses type-level attention to define the interaction weights between nodes of different types. The function expression is:

[0099] In formula (1), β M,S represents the interaction weight between the shield machine node and the stratum node, q M represents the query vector of the shield machine node, represents the transpose of the query vector of the shield machine node, M represents the shield machine node, k S Represents the key vector of the stratum node, S represents the stratum node, d represents the feature dimension of the key vector, which is used to scale the dot product result, and softmax is the normalization function used to map the dot product result to a probability distribution

[0100] Based on type attention, instance-level attention is used to calculate the weight distribution within nodes of the same type. The function expression is:

[0101] M i,j =softmax(LeakyReLU(a T [Wh i ⊕Wh j ]))(2)

[0102] In formula (2), M i,j represents the instance-level attention weight of shield machine node i and node j, a represents the attention vector, a T represents the transpose of the attention vector, W represents the feature transformation matrix, ⊕ represents the splicing operation, which is used to fuse the features of nodes i and j, and M i and M j They represent the i-th and j-th TBM nodes respectively, and LeakyReLU represents the activation function.

[0103] The hybrid-driven prediction model includes temporal branch, physical branch and spatial branch. The temporal branch, physical branch and spatial branch use parallel computing to output three types of prediction values.

[0104] The time series branch collects time series data from the shield machine construction process, such as time-varying parameters such as propulsion speed, cutterhead speed, and soil bin pressure. The normalized data is organized into a sequence in chronological order and used as the input for the time series branch. A multi-layer LSTM network is constructed, for example, a three-layer LSTM. Each LSTM layer contains multiple memory cells, each with an input gate, a forget gate, and an output gate. At the first time step, the input construction time series feature vector is fed into each memory cell of the first LSTM layer. For example, the input gate's activation value is calculated based on the current input and the hidden state from the previous time step. The forget gate determines which information from the previous time step to retain, and the output gate determines the hidden state output for the current time step. By continuously updating the hidden state, after the first LSTM layer processes all time steps, the output serves as the input for the second LSTM layer. This process repeats until the third LSTM layer completes processing. The hidden state of the final LSTM layer passes through a fully connected layer, mapping the hidden state to the prediction space of the shield machine's attitude angle, outputting the time series prediction of the shield machine's attitude angle.

[0105] Physics branch: Build a mechanism model based on the physical principles of shield machine construction. For example, consider the force balance of the shield machine and analyze the impact of factors such as propulsion force, soil pressure, and friction on the shield machine's posture.

[0106] According to the mechanism model, calculate the rate of change of the shield machine's attitude angle. For example, through Newton's second law and the shield machine's motion equation, derive the relationship between the attitude angle change rate and various physical quantities. Collect geophysical characteristic data, such as the density, strength, and permeability of the stratum. Use these geophysical characteristics and the attitude angle change rate calculated by the mechanism model as input. Construct a multi-layer perceptron and set the appropriate number of hidden layers and neurons. For example, set 2 hidden layers with 64 neurons in each layer. Use an activation function (such as the ReLU function) to perform a nonlinear transformation on the output of each layer. Train the multi-layer perceptron through forward propagation and backpropagation algorithms to learn the nonlinear mapping relationship between geophysical characteristics and the shield machine's attitude angle. Input the new geophysical characteristics into the trained multi-layer perceptron and output the physical prediction value of the shield machine's attitude angle.

[0107] Spatial branch: Collect tunnel spatial feature data, such as the coordinates of the tunnel axis, the distribution of geological units, etc. Organize these data into a format suitable for convolutional neural network (CNN) processing, such as dividing the tunnel space into grids, each grid corresponding to a feature vector. Normalize the input tunnel spatial features so that the data have the same scale. Construct a convolutional neural network and set multiple convolution layers and pooling layers. For example, set 2 convolution layers and 2 pooling layers. In the convolution layer, use the convolution kernel to perform a convolution operation on the input tunnel spatial features. The convolution kernel slides on the feature map to calculate the features of the local area. For example, for a 3×3 convolution kernel, calculate the dot product between the convolution kernel and the corresponding feature area at each position to obtain a new feature map.

[0108] The pooling layer downsamples the output of the convolutional layer. For example, using max pooling, it selects the maximum value in each pooled area as the output, reducing the size of the feature map while retaining important features. Through convolution and pooling operations, the convolutional neural network can capture the spatial structure of the construction space characteristics and the relative position of the tunnel planning area and geological units. For example, it can identify the changes in the tunnel's orientation in different geological units and the impact of the distribution of geological units on tunnel construction. The final output of the convolutional neural network is passed through a fully connected layer and mapped to the predicted space of the shield machine's attitude angle, outputting the spatial predicted value of the shield machine's attitude angle.

[0109] The three types of prediction values output by the time series branch, the physical branch, and the spatial branch are weighted and fused. In this way, the accuracy of the shield machine attitude angle prediction is improved by comprehensively utilizing the time series information, physical mechanism information, and spatial structure information.

[0110] In step 4, a Bayesian causal diagram is constructed. The cause nodes of the Bayesian causal diagram are: construction timing factors include parameters that change with time, such as the shield machine's propulsion speed, cutterhead speed, and propulsion time; geophysical factors include indicators reflecting geological characteristics such as stratum density, strength, permeability, and porosity; tunnel spatial factors involve spatial information such as the coordinates of the tunnel axis, tunnel burial depth, and the relative position of the tunnel and surrounding geological structures.

[0111] Result node: Shield machine attitude angle, which can be subdivided into pitch angle, roll angle and yaw angle, etc., is used to describe the attitude of the shield machine in three-dimensional space.

[0112] Based on the physical principles of shield machine construction and engineering experience, the impact of each causal node on the effect node is analyzed. For example, excessively fast advancement can cause the shield machine's posture to become unstable, thus affecting the attitude angle; varying ground strengths can create varying resistance to the shield machine's advancement, further affecting the attitude angle; and tunnel depth and surrounding geological structures can affect the forces acting on the shield machine, also influencing the attitude angle. Using directed edges to connect causal and effect nodes, a Bayesian causal diagram is constructed to clarify the causal dependencies between various factors.

[0113] Based on historical data and expert experience, we initialize the prior probability distribution for each cause and effect node. For example, for propulsion speed, we can determine its probability distribution in different intervals based on historical construction data; for shield machine attitude angle, we can determine its initial probability distribution based on statistical results from similar projects.

[0114] Based on the causal relationships in the Bayesian causal graph, a likelihood function is constructed. This describes the probability of a certain state occurring at the effect node given the state of the cause node. For example, given the propulsion speed and stratum strength, the probability of a shield machine attitude angle being a specific value can be calculated. The specific form of the likelihood function can be derived through historical data fitting or physical model derivation.

[0115] The posterior probability distribution is calculated using Bayes' theorem, using the formula P(θ|D) = P(D)P(D|θ)P(θ), where θ represents the node parameters (such as the specific value of the attitude angle), D represents the observed data (i.e., the state of the node), P(θ|D) is the posterior probability distribution, P(D|θ) is the likelihood function, P(θ) is the prior probability distribution, and P(D) is the normalization constant. By calculating the posterior probability distribution, the probability state of each node is updated. For example, for the shield machine attitude angle node, the probability of its attitude angle taking different values is updated based on the new observation data (the state of the construction sequence, geophysical factors, and tunnel space factors).

[0116] By varying the state of each causal node, we observe how the posterior probability distribution of the result node (the shield machine's attitude angle) changes. For example, while keeping the states of the other causal nodes fixed and varying the propulsion speed, we calculate the magnitude of the change in the posterior probability distribution of the shield machine's attitude angle. Through sensitivity analysis, we quantify the degree of influence of each factor on the shield machine's attitude angle. The greater the degree of influence, the more significant the factor's impact on the shield machine's attitude angle.

[0117] In addition to the posterior probability-based prediction values obtained from the Bayesian causal graph, the prediction values obtained from other models in the previous steps (such as the temporal branch, physical branch, and spatial branch in the hybrid drive prediction model) can also be combined.

[0118] The adaptive gating mechanism, similar to the gating unit in a neural network, dynamically adjusts the weights of each predicted value based on the characteristics of the input data. The gating mechanism's inputs include the predicted values and related feature information (such as the current state of construction timing, geophysical factors, and tunnel spatial factors). The gating mechanism learns a mapping function to map the input features into a weight space. For example, a multilayer perceptron (MLP) is used to implement this mapping. The MLP takes the predicted values and feature information as input and outputs the weights of each predicted value.

[0119] According to the weights calculated by the adaptive gating mechanism, each prediction value is dynamically weighted and fused. The final output of the shield machine attitude angle prediction value is obtained through dynamic weighted fusion.

[0120] Calculate the confidence interval for the TBM attitude angle based on the Bayesian posterior probability distribution. For example, you can choose an interval with a high probability density in the posterior probability distribution as the confidence interval. For example, a 95% confidence interval indicates that the probability of the TBM attitude angle occurring within that interval is 95%. Alternatively, you can use Monte Carlo simulation or other methods to extract a large number of samples from the posterior probability distribution, calculate the sample mean and standard deviation, and then determine the confidence interval.

[0121] Through the above steps, the influence of construction sequence, geophysical and tunnel space factors on the shield machine attitude angle can be comprehensively considered, multiple prediction values can be dynamically integrated, and accurate shield machine attitude angle prediction results and their confidence intervals can be output, providing a reliable decision-making basis for shield machine construction.

[0122] A shield machine posture prediction system, as shown in the attached Figure 3 Shown, including:

[0123] 1. Data acquisition module

[0124] The data acquisition module is the foundation of the entire system. It connects to hardware devices such as the shield machine controller, gyroscope, and geological radar to acquire shield machine operating data. Shield machine operation parameters include: operating status, propulsion speed, steering angle, thrust, etc. Geological environment parameters include: soil type, ground pressure, soil moisture, etc., all of which affect the shield machine's operating posture. Tunnel planning parameters include: the tunnel's designed path, dimensions, inclination angle, etc.

[0125] The data acquisition module also integrates a BIM model library and a geological exploration report parsing engine, providing more contextual information for the shield machine and ensuring data accuracy and comprehensiveness. The data is cleaned and normalized using a data rules engine to remove noise and ensure data quality for subsequent analysis.

[0126] 2. Feature Extraction Module

[0127] The feature extraction module uses a three-layer deep learning model to extract different types of features from the preprocessed data: Construction time series features: These are primarily extracted from time series data, taking into account dynamic changes during the construction process. Geophysical features: These extract features related to physical properties such as soil and strata based on geological environment parameters. Tunnel spatial features: These extract spatial features related to tunnel planning and geometry.

[0128] These features are extracted through a deep learning model and converted into a heterogeneous graph, which then constructs different nodes: The TBM node represents the relevant features of the TBM itself. The stratum node represents the characteristics of different strata within the geological environment. The tunnel axis node represents the path and direction designed in tunnel planning. Using the graph attention mechanism, the system dynamically weights and interacts information between different nodes to better capture the relationships and influences between them, enhancing the accuracy of the prediction model.

[0129] 3. Distributed prediction module

[0130] This module uses a trained hybrid drive prediction model to perform time-series predictions and generate the following prediction values: Time-series predictions of the shield machine's attitude angle: These reflect the changes in the shield machine's attitude at different time points. Physical predictions: These predict the shield machine's performance under different geological environments based on geological and physical parameters. Spatial predictions: These predict the tunnel's spatial layout, topography, and other factors to ensure the shield machine can advance along its planned trajectory. This module uses distributed computing methods to accelerate the prediction process and ensure real-time feedback on prediction results during actual operations.

[0131] 4. Fusion prediction module

[0132] The fusion prediction module is the core of the entire system. It quantifies the relationship between construction timing, geophysical characteristics, and tunnel spatial characteristics through a Bayesian causal diagram, forming a causal relationship model. The application of Bayesian causal diagrams can effectively reveal the degree to which different factors affect the shield machine's attitude angle.

[0133] The system then dynamically weights and fuses the different predictions based on the quantified results. This weighted fusion approach dynamically adjusts the weights of the individual predictions based on real-time data changes, resulting in a more accurate final prediction of the shield machine's attitude angle. Furthermore, the fused prediction module outputs a confidence interval for the prediction, which is used to assess the uncertainty and reliability of the prediction results and provide decision support.

[0134] Application Scenario

[0135] This system is suitable for dynamic posture prediction during shield machine construction, effectively improving construction accuracy and reducing construction problems caused by inaccurate posture control. It optimizes construction plans and proactively identifies potential construction risks based on prediction results. Real-time monitoring and adjustment, through real-time data updates, adjusts the shield machine's posture to avoid construction deviations.

[0136] In general, this system integrates multiple advanced technologies such as sensor technology, deep learning, graph neural networks and causal inference, and has high practical value and potential.

[0137] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is clearly not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A shield machine posture prediction method, characterized in that: The following steps are involved: Step 1: Obtain shield machine construction parameters, geological environment parameters, and tunnel planning parameters, and preprocess the obtained parameters to extract construction time sequence characteristics, geophysical characteristics, and tunnel spatial characteristics from the preprocessed shield machine construction parameters, geological environment parameters, and tunnel planning parameters; Step 2: The extracted construction time sequence features, geophysical features, and tunnel spatial features are input into the heterogeneous graph to map shield machine nodes, stratum nodes, and tunnel axis nodes. The shield machine nodes, stratum nodes, and tunnel axis nodes are weighted through the graph attention mechanism to form interactive connections, and the node features are updated in real time using incremental learning. Step 3: Input the node features and node connection relationships of the heterogeneous graph into the trained hybrid drive prediction model to obtain the time series prediction value, physical prediction value, and spatial prediction value of the shield machine attitude angle; Step 4: Quantify the impact of construction timing, geophysical and tunnel spatial factors on the current shield machine attitude angle by constructing a Bayesian causal diagram, and dynamically weighted fuse the different prediction values obtained based on the quantified results to output the final prediction value and confidence interval of the shield machine attitude angle.

2. A shield machine posture prediction method according to claim 1, characterized in that: In step one, the shield machine construction parameters include at least the propulsion speed, cutterhead speed, soil bin pressure and tunneling thrust; the geological environment parameters include at least the soil type, soil moisture content, groundwater level and stratum information; the tunnel planning parameters include at least the tunnel design axis, slope and turning radius; the sliding window method is used to filter out invalid data in the acquired parameters and perform normalization processing.

3. A shield machine posture prediction method according to claim 1, characterized in that: In step 1, a three-layer deep learning model is used to extract construction timing characteristics, geophysical characteristics, and tunnel spatial characteristics. The working method of the three-layer deep learning model includes the following steps: S101. Separating temporal features and spatial features in shield machine construction parameters through a gating mechanism. The gating mechanism dynamically adjusts the extraction ratio of temporal features and spatial features based on a long short-term memory network, and separates the temporal features and spatial features in the shield machine construction parameters. The construction temporal features are used to predict the change trend of the shield machine posture over time, and the construction spatial features are used to analyze the posture differences of the shield machine at different positions. S102, using a dynamic graph convolutional network to mine the spatial correlation between the geological area and the tunnel area, mapping the geological area and the tunnel planning area as nodes of a dynamic graph, and mapping the spatial correlation between the geological environment parameters and the tunnel planning parameters as edges of the dynamic graph, wherein the dynamic graph convolutional network extracts the spatial correlation features of the geological environment parameters and the tunnel planning parameters as geophysical features and tunnel spatial features respectively through multi-layer convolution operations; S103. Use the Voronoi subdivision principle to divide the geological area into geological units of different scales. Use the points on the tunnel axis as seed points and subdivide the geological area according to different radii to obtain geological units of different sizes. Construct a multi-scale geological unit coupling model to couple the geological unit information of different scales. During the coupling process, explore the construction space characteristics, the relative position relationship between the tunnel planning area and the geological units.

4. A shield machine posture prediction method according to claim 1, characterized in that: In step 2, the shield machine node and the stratum node are connected according to the current stratum position of the shield machine, the shield machine node and the tunnel axis node are connected according to the spatial relationship between the current position of the shield machine and the tunnel axis node, the stratum unit node and the tunnel axis node are connected according to the constraint relationship of the stratum on the tunnel axis, and the stratum unit nodes are connected according to the spatial adjacent relationship of the strata.

5. A shield machine posture prediction method according to claim 4, characterized in that: The graph attention mechanism uses type-level attention to define the interaction weights between nodes of different types. The function expression is: In formula (1), β M,S represents the interaction weight between the shield machine node and the stratum node, q M represents the query vector of the shield machine node, represents the transpose of the query vector of the shield machine node, M represents the shield machine node, k S Represents the key vector of the stratum node, S represents the stratum node, d represents the feature dimension of the key vector, which is used to scale the dot product result, and softmax is the normalization function used to map the dot product result to a probability distribution Based on type attention, instance-level attention is used to calculate the weight distribution within nodes of the same type. The function expression is: M i,j =softmax(LeakyReLU(a T [Wh i ⊕Wh j ])) (2) In formula (2), M i,j represents the instance-level attention weight of shield machine node i and node j, a represents the attention vector, a T represents the transpose of the attention vector, W represents the feature transformation matrix, ⊕ represents the splicing operation, which is used to fuse the features of nodes i and j, and M i and M j They represent the i-th and j-th TBM nodes respectively, and LeakyReLU represents the activation function.

6. A shield machine posture prediction method according to claim 1, characterized in that: The hybrid drive prediction model includes a timing branch, a physical branch and a spatial branch. The timing branch, the physical branch and the spatial branch use parallel computing to output three types of prediction values. The timing branch gradually processes the input construction timing features through a multi-layer long short-term memory network stack, updates the hidden state at each time step to output the timing prediction value of the shield machine attitude angle. The physical branch calculates the attitude angle change rate based on the mechanism model, and uses a multi-layer perceptron to learn the nonlinear mapping relationship between geological physical features and attitude angles to output the physical prediction value of the shield machine attitude angle. The spatial branch uses a convolutional neural network to perform convolution and pooling operations on the input tunnel spatial features, and captures the spatial structure information in the relative position relationship between the construction space characteristics, the tunnel planning area and the geological unit to output the spatial prediction value of the shield machine attitude angle.

7. A shield machine posture prediction method according to claim 1, characterized in that: In step 4, the cause nodes of the Bayesian causal diagram are construction timing, geophysical and tunnel space factors, and the result node of the Bayesian causal diagram is the shield machine attitude angle. The probability state of each node is updated by calculating the posterior probability distribution to obtain the quantitative results of the influence of each factor on the shield machine attitude angle. The adaptive gating mechanism is used to dynamically weighted fuse the predicted values to output the final attitude angle and confidence interval.

8. A shield machine posture prediction system, characterized by: A shield machine posture prediction method according to any one of claims 1 to 7 is applied, wherein the shield machine posture prediction system comprises: The data acquisition module uses a sensor interface to connect to the shield machine controller, gyroscope, and geological radar, and connects to the BIM model library and geological exploration report parsing engine to obtain shield machine construction parameters, geological environment parameters, and tunnel planning parameters. The data rule engine is used to clean and normalize the acquired data. The feature extraction module uses a three-layer deep learning model to extract construction timing features, geophysical features, and tunnel spatial features from the preprocessed data. The extracted construction timing features, geophysical features, and tunnel spatial features are mapped to form a heterogeneous graph of shield machine nodes, stratum nodes, and tunnel axis nodes. Dynamic weighted interactive connections between nodes are performed through a graph attention mechanism. The distributed prediction module obtains the time series prediction value, physical prediction value and spatial prediction value of the shield machine attitude angle in a distributed manner through the trained hybrid drive prediction model; The prediction module is integrated to construct a Bayesian causal diagram to quantify the impact of construction timing, geophysical and tunnel spatial factors on the shield machine attitude angle. Based on the quantified results, different prediction values are dynamically weighted and fused to output the final prediction value and confidence interval of the shield machine attitude angle.

Images