Shield axis deviation prediction model, construction method thereof, prediction method and system

Abstract

The invention belongs to the field of intelligent construction, and particularly discloses a shield axis deviation prediction model, a construction method thereof, a prediction method and a system, and the method comprises the steps: training a mixed model through a training set, and enabling the trained mixed model to be the shield axis deviation prediction model; the training set comprises splicing parameters and corresponding shield axis deviation data, segmenting a shield parameter time sequence signal, and processing each obtained signal segment to obtain the splicing parameters; the hybrid model comprises an adaptive neural fuzzy inference system ANFIS and a full-connection space-time diagram neural network model FC-STGNN; the ANFIS is used for quantizing the splicing parameters, and a parameter # imgabs0 # is output; performing graph construction and graph convolution by the FC-STGNN based on the splicing parameters, and outputting a parameter # imgabs1 #; and carrying out weighted fusion on the parameters # imgabs2 # and # imgabs3 #, and outputting a shield axis deviation result. According to the invention, the precision and adaptability of shield axis deviation prediction can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent construction, and more specifically, relates to a shield axis deviation prediction model, a construction method thereof, a prediction method and a system thereof. Background Art

[0002] Shield construction is widely used in underground engineering, and the prediction of its axis deviation is crucial. With the increase in the scale and complexity of underground space development, shield construction faces many challenges: shield construction belongs to an engineering scenario with high noise and long time series, and many uncertain factors existing in the construction process may affect the prediction results of axis deviation; in addition, the coupling effect of multiple factors in shield construction further increases the difficulty of axis deviation prediction.

[0003] The prediction methods for shield axis deviation mainly include machine learning methods, deep learning methods, etc. However, these methods all have obvious shortcomings: Existing axis deviation prediction methods are difficult to comprehensively consider the uncertainty and dynamic correlation of multiple shield parameters, and there are problems such as insufficient accuracy and poor adaptability. On the one hand, traditional methods mainly capture time features or model time features and spatial features separately, ignoring the dynamic coupling correlation between parameter features, and there is a sense of fragmentation in spatio-temporal modeling; on the other hand, the black box effect of deep models leads to the lack of interpretability of their results, poor transparency of prediction logic, and difficult to measure the contribution degree of multiple factors. Summary of the Invention

[0004] In view of the above defects or improvement requirements of the prior art, the present invention provides a shield axis deviation prediction model, a construction method thereof, a prediction method and a system thereof, aiming to improve the accuracy and adaptability of shield axis deviation prediction.

[0005] To achieve the above object, according to the first aspect of the present invention, a construction method of a shield axis deviation prediction model is proposed, including the following steps: Training a hybrid model with a training set, and the trained hybrid model is the shield axis deviation prediction model; The training set includes splicing parameters and corresponding shield axis deviation data, and the acquisition method of the splicing parameters is: segmenting the time series signal of shield parameters, and processing each obtained signal segment to obtain splicing parameters including content and position information; The hybrid model includes an Adaptive Neuro-Fuzzy Inference System (ANFIS) and a Fully Connected Spatio-Temporal Graph Neural Network Model (FC-STGNN); this hybrid model takes the splicing parameters as input, and the Adaptive Neuro-Fuzzy Inference System (ANFIS) quantifies the splicing parameters and outputs parameters ; the Fully Connected Spatio-Temporal Graph Neural Network Model (FC-STGNN) constructs a graph and performs graph convolution based on the splicing parameters, and outputs parameters ; Furthermore, the parameters and are weighted and fused to output the shield axis deviation result.

[0006] As a further optimization, each signal segment is processed, including: For each signal segment , it is processed by GRU to obtain the feature encoding ; The sine position encoding is concatenated with the feature encoding to obtain the concatenated parameter .

[0007] As a further optimization, the sine position encoding is concatenated with the feature encoding to obtain the concatenated parameter , specifically: Among them, represents encoding concatenation, d represents the feature dimension, represents the dimension index, represents the auxiliary integer variable for judging parity.

[0008] As a further optimization, the adaptive neuro-fuzzy inference system ANFIS quantifies the concatenated parameter and outputs the parameter , specifically including: The Gaussian membership function is used to fuzzify the input concatenated parameter to obtain the membership degree of the parameter to the fuzzy set; furthermore, rule reasoning is performed based on the membership degree of the parameter to the fuzzy set, and then defuzzification is performed to obtain the parameter .

[0009] As a further optimization, the fully connected spatio-temporal graph neural network model FC-STGNN constructs a graph and performs graph convolution based on the concatenated parameter, and outputs the parameter , specifically including: Multiple independent moving pooling GNN layers are used to process the input in parallel. For each moving pooling GNN layer: Based on the input concatenated parameter, a fully connected graph is constructed; The dot product of the similarity matrix E of the fully connected graph and the attenuation matrix C is used to obtain the adjacency matrix ; Based on the adjacency matrix , the sliding window message passing neural network MPNN is used to perform convolution on the fully connected graph to obtain the aggregated feature ; The activation function and learnable weights are used to process the aggregated feature Update it, perform temporal dimension average pooling on the updated features, and obtain the output features ; Concatenate the output features of each moving pooling GNN layer and map and output the concatenated features through an MLP .

[0010] As a further preference, the shield parameters are one or more of the original cutting face deviation, original tail deviation, jack speed adjustment, total thrust, propulsion jack pressure, propulsion zone stroke, cutter head rotation speed, face soil pressure, slurry supply main pipeline pressure and flow rate, and slurry return main pipeline flow rate.

[0011] According to the second aspect of the present invention, a shield axis deviation prediction model is provided, which is constructed by using the above-mentioned shield axis deviation prediction model construction method.

[0012] According to the third aspect of the present invention, a shield axis deviation prediction method is provided. The shield axis deviation prediction method includes the following steps: Obtain the shield parameter time series signal and segment it, process each obtained signal segment to obtain the concatenated parameters including content and position information; input the concatenated parameters into the shield axis deviation prediction model to obtain the predicted shield axis deviation.

[0013] As a further preference, after predicting the shield axis deviation, GNN Explainer post-explain the output of the fully connected spatio-temporal graph neural network model FC-STGNN to obtain the contribution degree of each shield parameter to the prediction result and the dynamic coupling relationship between each shield parameter.

[0014] According to the fourth aspect of the present invention, a shield axis deviation prediction system is provided, including a processor, and the processor is used to execute the above-mentioned shield axis deviation prediction method.

[0015] Generally speaking, compared with the prior art through the above technical solutions conceived by the present invention, the following technical advantages are mainly possessed: 1. The hybrid model designed by the present invention adds FC-STGNN a fuzzy processing link on the basis of the ANFIS model. When facing different shield construction scenarios and data characteristics, ANFIS it can adapt to these changes through learning and adjustment, and at the same time combine FC-STGNN the ability to capture spatio-temporal dependence relationships, so that the hybrid model used for shield axis deviation prediction has better prediction accuracy, flexibility and adaptability.

[0016] 2. The present invention forms a hybrid model of FC-STGNN and ANFIS through a weighted fusion method. ANFIS optimizes the non-linear relationship from the implicit fuzzy logic level, and FC-STGNN depicts the data law from the explicit spatio-temporal dimension. The comprehensive coverage of data characteristics is achieved through the parallel operation of the two, avoiding the impact on prediction timeliness and robustness caused by sequential connection.

[0017] 3. By constructing a fully connected spatio-temporal graph FC-STGNN, the construction parameters affecting the shield axis deviation are correlated with each other from the spatio-temporal perspective, solving the singularity of the traditional scheme focusing on the sequential relationship or the sense of fragmentation of separate modeling of time and space; at the same time, GRU is further optimized to replace 1D CNN for feature extraction, conforming to the characteristics of the long time series of shield parameters.

[0018] 4. Aiming at the problem that the single FC-STGNN cannot solve the limited adaptability caused by the uncertainty of shield parameters, the ANFIS fuzzy system is introduced to quantify the uncertainty in the prediction of axis deviation. The uncertainty parameters are mapped into interpretable fuzzy logic through fuzzification, fuzzy inference and defuzzification, and the dynamic update of the fuzzy system is realized through an adaptive membership function.

[0019] 5. Aiming at the problems of the black box effect in the hybrid model and the difficulty in interpreting the graph structure of the FC-STGNN model, the post hoc explanation of GNN Explainer is used to quantify the contribution degree of shield parameters to the prediction result, solving the problem of poor transparency of the black box model, providing a traceable basis for the result of axis deviation prediction, and at the same time being able to perform a finer-grained explanation with the nodes and edges of the fully connected graph as the basic units. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic flow chart of the shield axis deviation prediction method according to the embodiment of the present invention.

[0021] Figure 2 It is an architecture diagram of the ANFIS network layer according to the embodiment of the present invention.

[0022] Figure 3 It is a schematic diagram of the FC-STGNN graph construction and graph convolution according to the embodiment of the present invention.

[0023] Figure 4 It is a post hoc explanation flow chart of GNN Explainer according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0025] A method for constructing a shield axis deviation prediction model provided by an embodiment of the present invention, as Figure 1 shown, includes the following steps: S1: Data cleaning and segmentation: Clean the raw data collected by the shield machine sensors and segment them into segments of equal length.

[0026] Specifically, the raw data is the time series signal of the shield parameters collected by the sensors; the shield parameters include several of the original cut deviation, original tail shield deviation, jack speed adjustment, total thrust, propulsion jack pressure, propulsion zone stroke, cutter head speed, face soil pressure, slurry supply main pipeline pressure and flow rate, and return slurry main pipeline flow rate.

[0027] Data cleaning and segmentation can improve the accuracy of feature extraction. On the one hand, complex precision geological conditions and equipment anomalies may both cause data fluctuations. By cleaning the data, noise reduction processing can be performed on the data to remove redundancy and unify the format, ensuring the data quality input into the model. On the other hand, the sample length continuous time series needs to be segmented into segments of equal length at a specific period for feature extraction. The initial data collected by the sensor at the time length T can be segmented into the following number of segments: where M represents the number of segments that the sequence signal collected by the sensor can be segmented into, f represents the segmentation size. When segmenting the data, the size f should be greater than the sliding window size R of the fully connected spatio-temporal graph convolution, but f being too long may cause the spatial correlation degree to be diluted by the time series when constructing the fully connected spatio-temporal graph; being too short will result in a mismatch of the rule granularity and an increase in the calculation error of the fuzzy membership degree.

[0028] These segments form a segment set , for the a-th segment , it contains the segmented signals collected by N sensors, that is, where, represents the signal of the -th sensor under the a-th segment t , and each sensor correspondingly obtains a time series signal of a shield parameter, .

[0029] S2: Feature extraction and position encoding: For each signal segment, extract the segment features, obtain the feature encoding, and perform position encoding and splicing to obtain the splicing parameters .

[0030] Furthermore, through GRU the processing of , output the feature encoding: Among them, represents the processing of the segment by GRU, and the processed feature encoding is . GRU The gating mechanism of GRU can effectively capture long-term dependencies. The data during shield tunneling construction is typical time series data, and there are long-term dependencies between the data at each time step. Using

[0031] for feature extraction and encoding can better capture the long-time sequence parameters in shield tunneling construction than 1D CNN, and improve the performance of the hybrid model. Furthermore, use sine position encoding to generate position vectors, and splice them with

[0032] Among them, represents the combination of position encoding and feature encoding, represents sine position encoding, d represents the feature dimension, represents the dimension index, represents the auxiliary integer variable for judging parity.

[0033] S3: Train the hybrid model with the training set, and the trained hybrid model is the shield axis deviation prediction model. The hybrid model includes the Adaptive Neuro-Fuzzy Inference System (ANFIS) and the Fully Connected Spatio-Temporal Graph Neural Network Model (FC-STGNN); this hybrid model takes the splicing parameter as the input. The Adaptive Neuro-Fuzzy Inference System (ANFIS) quantifies the splicing parameter and outputs the parameter ; the Fully Connected Spatio-Temporal Graph Neural Network Model (FC-STGNN) constructs a graph and performs graph convolution based on the splicing parameter and outputs the parameter ; then, weight and fuse the parameters and to output the shield axis deviation result. Specifically as follows: (1)ANFIS Quantification of Uncertainty: Based on the ANN adaptive network architecture of ANFIS and the FIS fuzzy inference, the input parameters are fuzzified, inference is performed based on fuzzy rules, and defuzzification is carried out to achieve the quantification of uncertain parameters and output parameters .

[0034] The Adaptive Neuro-Fuzzy Inference System (ANFIS) realizes the collaborative optimization of parameter self-learning and interpretable fuzzy rules by deeply integrating the fuzzy technology of fuzzy logic with the learning ability of ANN. By embedding domain knowledge in ANFIS, uncertainty is quantified and the generalization ability of the model is enhanced. When using only FC-STGNN , only the spatio-temporal dependence relationships in the data can be modeled, but the effective integration of domain knowledge is lacking; while by adding the ANFIS model in parallel in a weighted manner on the basis of FC-STGNN, the problem that the model has limited modeling ability for the highly nonlinear and strongly coupled parameter relationships in shield tunneling construction (such as the nonlinear influence of the dynamic coupling between soil pressure and shield attitude adjustment on the axis deviation, etc.) when only using the FC-STGNN model can be solved.

[0035] ANFIS modularly maps fuzzification (Layer 1), rule inference (Layers 2 - 4), and defuzzification (Layer 5) into neural network nodes. The neural network architecture includes five layers: the fuzzy layer, the rule layer, the normalization layer, the inference layer, and the defuzzification layer. As Figure 2 shown, the specific steps are as follows: (1.1) Fuzzy Layer: For various types of input parameters, different membership functions and appropriate partitioning methods need to be designed to achieve parameter fuzzification.

[0036] Since the Gaussian membership function has stronger adaptability and robustness and is suitable for the dynamically changing shield tunneling construction parameters, the Gaussian membership function is adopted in this embodiment: where represents the membership function, represents the parameter input, j represents the index of the parameter input; and represent the mean and standard deviation of the Gaussian membership function respectively.

[0037] The partitioning method adopts the dichotomy method, and uses the "high - low" semantic labels to map the data into different fuzzy sets.

[0038] (1.2) Rule layer: After being processed by the fuzzy layer, each input parameter will have a membership degree corresponding to each fuzzy set. Based on the statistical metrics (mean, variance, entropy) of the fuzzy membership degrees within the time window and domain knowledge, IF-THEN logical rules can be generated. The IF part is the antecedent, which is a combination of membership degrees, and the THEN part is the consequent, responsible for performing combined calculations. These rules are aggregated and integrated to form a systematic rule base.

[0039] The weight of each rule is determined by its firing strength: where i represents the rule index, x and y refer to different input parameters, represents the membership degree to the fuzzy set , represents the firing strength of the i -th rule.

[0040] (1.3) Normalization layer: Convert the rule firing strength into relative weights, quantify the contribution rate of each rule, and dynamically balance the competition relationship between rules.

[0041] where, represents the normalized weight of the i -th rule.

[0042] (1.4) Inference layer: Perform combined calculations on the consequent (THEN part) of each fuzzy rule to generate a quantitative output related to the input parameters. The formula for the combined calculation function is: where, represents the linear function corresponding to the i -th rule, , , represent learnable parameters.

[0043] The mathematical essence of the combined calculation lies in: dividing and fusing the conclusions of linear rules through weights, and finally constructing a global non-linear mapping that can approximate any continuous function.

[0044] (1.5) Defuzzification layer: Aggregate all rule outputs to a node, and this node takes the sum of all input signals as the total output, that is: where, represents the output result.

[0045] The adaptability of ANFIS fuzzy inference lies in the fact that the parameters of the fuzzy layer (membership degree parameters , ) and the inference layer (rule linear combination coefficients ) are dynamically adjusted to adapt to the data distribution. Therefore, it is an adaptive node, represented as a square node. The calculation logic of other network layers is fixed and performs predefined operations, so it is a fixed node, represented as a circular node.

[0046] (2) FC-STGNN captures spatio-temporal dependencies: construct a fully connected graph, use a sliding window message passing neural network MPNN for graph convolution to obtain a fully connected spatio-temporal graph, and output parameters .

[0047] Specifically, as Figure 3 shown, constructing FC-STGNN includes graph construction and graph convolution: (2.1) For the input parameter , form a set , which is the node set of the fully connected graph. Enhance the expression ability of features through the feature enhancement function and measure the similarity between different features through dot product: Among them, represents the similarity between the a -th node feature of the t -th segment and the b -th node feature of the u -th segment.

[0048] Based on this, summarize the similarity matrix of the fully connected graph: , , that is, it covers the features and their correlations of all segments and can be used as the set of edges of the fully connected graph. Thus, the fully connected graph can be obtained.

[0049] The FC-STGNN measures the temporal correlation between features by constructing an attenuation matrix. The attenuation rate of the temporal correlation between each segment is : and so on, the spatio-temporal correlation between any features within different segments is: Among them, represents the temporal correlation degree between the -th node feature of the -th segment and the -th node feature of the -th segment. Thus, construct the attenuation matrix of the fully connected graph: .

[0050] Furthermore, the dot product of the similarity matrix E and the attenuation matrix C is used to obtain the adjacency matrix, and its elements can reflect the correlation of different features in different time segments affected by the time distance: where represents the edge weight of different features in different time segments affected by the time distance, which is the normalized correlation degree adjusted by the attenuation matrix. is the matrix formed.

[0051] (2.2) The convolution of the fully connected graph means using the sliding window message passing neural network (MPNN) to capture the dynamic dependencies in the fully connected spatio-temporal graph. The fully connected spatio-temporal graph is divided into equal-sized windows along the time dimension with a step size S and a window size R. For the w-th sliding window, the l central node of the -th layer performs convolution on all nodes within the time range where represents the aggregated feature of the l -th layer, timestamp a , and sensor t , which is obtained by aggregating the information of adjacent nodes within the window. w represents the central index of the current sliding window, which is used to locate the window position in the time dimension. R represents the sliding window size. The l -th layer, timestamp a , and sensor t node features contain spatio-temporal information and position encoding.

[0052] The activation function and learnable weights are used for feature update to enhance the non-linear expression ability of features and adapt to complex spatio-temporal patterns: where represents the rectified linear activation function. represents the updated feature of the l +1-th layer, timestamp a, and sensor t, which is obtained by processing the current layer features. represents the learnable weights.

[0053] Average pooling is performed on the updated features within each window in the time dimension to extract high-level features and optimize the processing efficiency of the model for long sequence data: Among them, represents the timestamp after pooling, w , sensor t at the l +1 layer output feature, obtained by averaging the features within the window.

[0054] The input is processed in parallel using multiple independent moving pooling GNN layers. The features after temporal pooling for each layer need to be concatenated to integrate multi-scale spatio-temporal information, and finally, the concatenated features are mapped and output through an MLP. This process can be expressed as: Among them, represents the predicted output of FC-STGNN, k represents the number of parallel GNN layers, represents the process of concatenating the features of each layer. MLP represents a multi-layer perceptron, which maps the concatenated features to the final prediction result through a non-linear transformation, completing the conversion from features to output.

[0055] (3) Output weighted fusion: At the end of model parallelism, the output results of ANFIS and FC-STGNN are weighted and summed for fusion to obtain the prediction result.

[0056] Specifically, the ANFIS model and FC-STGNN in the hybrid model operate independently. At the end of the model, in the form of weighted fusion, the weight W is set to combine the output results of the two: Among them, represents the final output result of the hybrid model, W represents the weight of the ANFIS model in the hybrid model, W The higher it is, the greater the influence of the ANFIS model on the prediction result.

[0057] An axis deviation prediction method for a shield further provided by an embodiment of the present invention includes the following steps: S4: Obtain the shield parameter time series signal and segment it, process each obtained signal segment to obtain the splicing parameters including content and position information. For specific methods, refer to S1 and S2, which will not be elaborated here; input the splicing parameters into the shield axis deviation prediction model to obtain the shield axis deviation prediction value.

[0058] S5: GNN Explainer post hoc explanation: The input for post hoc explanation is the trained FC-STGNN model and the fully connected spatio-temporal graph G, FC-STGNN and the initial prediction value of 。The GNN Explainer uses edges and nodes as the granularity of explanation, reveals the decision-making basis of the model by generating interpretable subgraphs and related node features, provides post hoc explanations for the axis deviation prediction results, and improves the reliability and transparency of model explanations. Its advantage lies in explaining the output results of GNN class graph structure neural networks, and can show FC-STGNN the importance of nodes and edges in the model for the output prediction results, and solve the black-box problem of hybrid models.

[0059] Such as Figure 4 shown, it specifically includes the following steps: (1) Initialize the interpreter and mask by creating Explainer objects. Based on the given graph structure and node features, an interpretable mask can be generated: Among them, represents the mask, represents the node importance weight vector represents the edge importance weight matrix, . The initial mask is usually a matrix of all 1s to indicate that all nodes and edges are not masked.

[0060] (2) The mask covers some nodes or edges of the fully connected spatio-temporal graph G by changing its own matrix elements, thereby adjusting the fully connected graph structure to generate a candidate subgraph . If then remove the th node; if then remove the th edge. This process is expressed as: Among them, represents the node set of the subgraph, represents the edge set of the subgraph; and are one-dimensional indexes, which map the a th node feature of the t th segment and the two-dimensional information of the b th node of the u th segment into a one-dimensional vector index for easy operation of the mask matrix.

[0061] (3) Input the obtained candidate subgraph into the FC-STGNN model, and based on the difference between its prediction result and the initial prediction output, screen out the nodes and edges that have a greater impact on the model prediction result.

[0062] This process uses the mean squared error MSE which is expressed as: where, represents the prediction difference loss, represents the fully connected spatio-temporal graph model. The loss function aims to analyze the prediction logic of the existing FC-STGNN model. Its optimization object is the temporary mask weights, and the data range is the current input data. Its essence belongs to decision analysis rather than model training.

[0063] Optimize the mask through gradient ascent , and obtain the optimized mask matrix by maximizing the loss. This process is expressed as: where, represents the optimized mask matrix, argmax represents finding the mask that maximizes the loss function , that is, determining the part of the graph structure that most affects the model prediction through optimization.

[0064] (4)GNN Explainer normalizes the optimized mask to obtain the node weights and edge weights , and filters out important nodes and edges through the importance threshold to generate the final explanatory subgraph , which only retains the graph structure that contributes the most to the model prediction, that is, contains the information most critical to the model prediction result. This process is: The post hoc interpretable output includes the final explanatory subgraph and the importance weights of the optimized mask. Through these outputs, it is possible to achieve FC-STGNN the transparency of the model's black box nature, show the degree of spatio-temporal correlation between shield parameters, and analyze the factors most relevant to the shield axis offset prediction result.

[0065] ​On the one hand, GNN Explainer can view the contribution of each shield parameter to the prediction result (the importance weight of the node); on the other hand, it can also analyze the dynamic coupling relationship between parameters through edges. For example, if the edge weight between x6 (total thrust) and x5 (jacking speed) is high, it indicates that the synergistic effect between the two has a significant impact on the prediction. In addition, from the perspective of time: First, for the structure within a time step, its prediction result (i.e., the prediction subgraph) can also show the relationship between parameters within a single time step. For example, the axis deviation prediction may have x6 as the hub node, forming dense connections with x14 and x20, indicating that the total thrust indirectly controls the vertical deviation by affecting the cutter head torque and slurry feeding flow rate. Second, for the relationship between time steps, for example, if the node importance of speed is high at the t-th time step, it can illustrate that the speed adjustment at the previous moment plays a key role in the prediction of the current axis deviation.

[0066] In summary, the present invention cleans and segments the captured shield time series parameters, performs feature encoding and position encoding, and then inputs the parameters into a hybrid model composed of a fully connected spatio-temporal graph neural network (FC-STGNN) and an adaptive neuro-fuzzy inference system (ANFIS): FC-STGNN constructs a fully connected spatio-temporal graph based on the correlation matrix and decay matrix, and after using a sliding window to segment it, performs graph convolution to capture spatio-temporal dependencies; the ANFIS model then performs fuzzy processing on the input parameters, combines fuzzy rule reasoning and defuzzification; after the two models are independent and parallel, the outputs are weighted and fused; the GNN Explainer can also be used to post-explain the model decision-making process.

[0067] It is easy for those skilled in the art to understand that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for constructing a shield axis deviation prediction model, characterized in that It includes the following steps: Train the hybrid model with the training set, and the trained hybrid model is the shield axis deviation prediction model; The training set includes splicing parameters and corresponding shield axis deviation data. The acquisition method of the splicing parameters is as follows: segment the shield parameter time series signal, and process each obtained signal segment to obtain splicing parameters including content and position information; The hybrid model includes an Adaptive Neuro-Fuzzy Inference System (ANFIS) and a Fully Connected Spatio-Temporal Graph Neural Network model (FC-STGNN); the hybrid model takes the splicing parameters as input, and the Adaptive Neuro-Fuzzy Inference System (ANFIS) quantifies the splicing parameters and outputs parameters ; the Fully Connected Spatio-Temporal Graph Neural Network model (FC-STGNN) constructs a graph and performs graph convolution based on the splicing parameters and outputs parameters ; then, parameters and are weighted and fused to output the shield axis deviation result.

2. The method for constructing the shield axis deviation prediction model according to claim 1, characterized in that Processing each signal segment includes: For each signal segment , it is processed by GRU to obtain a feature encoding ; the sine position encoding is concatenated with the feature encoding to obtain a concatenated parameter .

3. The method for constructing the shield axis deviation prediction model according to claim 2, characterized in that, Perform sine positional encoding and feature encoding to obtain the concatenated parameter , specifically as follows: Among them, represents encoding concatenation, d represents the feature dimension, represents the dimension index, represents an auxiliary integer variable for judging parity.

4. The method for constructing a shield axis deviation prediction model according to claim 1, wherein The adaptive neuro-fuzzy inference system ANFIS quantifies the splicing parameters and outputs the parameters , specifically including: The Gaussian membership function is used to fuzzify the input splicing parameters to obtain the membership degrees of the parameters to the fuzzy sets; then, based on the membership degrees of the parameters to the fuzzy sets, rule reasoning is performed, and then defuzzification is carried out to obtain the parameters .

5. The method for constructing the shield axis deviation prediction model according to claim 1, wherein The fully connected spatio-temporal graph neural network model FC-STGNN constructs graphs and performs graph convolutions based on splicing parameters, and outputs parameters , specifically including: Using multiple independent moving pooling GNN layers to process the input in parallel. For each moving pooling GNN layer: Based on the input splicing parameters, construct a fully connected graph; Dot product the similarity matrix E of the fully connected graph with the attenuation matrix C to obtain the adjacency matrix ; Based on the adjacency matrix , a sliding window message passing neural network MPNN is used to perform convolution on the fully connected graph to obtain aggregated features ; Update the aggregated features using an activation function and learnable weights, and perform temporal dimension average pooling on the updated features to obtain output features ; ; Concatenate the output features of each mobile pooling GNN layer and map and output the concatenated features through an MLP .

6. The method for constructing a shield axis deviation prediction model according to any one of claims 1-5, characterized in that, The shield parameters are one or more of the original cutting face deviation, original tail shield deviation, jack speed adjustment, total thrust, propulsion jack pressure, propulsion zone stroke, cutter head rotation speed, face soil pressure, slurry supply main pipeline pressure and flow rate, and slurry return main pipeline flow rate.

7. A shield axis deviation prediction model, characterized in that, It is constructed by using the construction method of the shield axis deviation prediction model according to any one of claims 1-6.

8. A shield axis deviation prediction method, characterized in that, It includes the following steps: Obtain the shield parameter time series signal and segment it. Process each obtained signal segment to obtain splicing parameters including content and position information; input the splicing parameters into the shield axis deviation prediction model according to claim 7 to obtain the predicted shield axis deviation.

9. The shield axis deviation prediction method according to claim 8, wherein After predicting the deviation of the shield axis, through GNN Explainer post hoc interpretation of the output of the fully connected spatio-temporal graph neural network model FC-STGNN, the contribution of each shield parameter to the prediction result and the dynamic coupling relationship between each shield parameter are obtained.

10. A shield axis deviation prediction system, characterized in that It includes a processor, and the processor is used to execute the shield axis deviation prediction method according to claim 8 or 9.

Images