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

Abstract

The present invention belongs to the field of intelligent construction and specifically discloses a shield axis deviation prediction model and its construction method, prediction method, and system. The method comprises: training a hybrid model using a training set; the trained hybrid model is the shield axis deviation prediction model; the training set includes splicing parameters and corresponding shield axis deviation data; the shield parameter time series signal is segmented; each obtained signal segment is processed to obtain the splicing parameters; the hybrid model includes an adaptive neural fuzzy inference system (ANFIS) and a fully connected spatiotemporal graph neural network model (FC-STGNN); the ANFIS quantifies the splicing parameters and outputs the parameters; the FC-STGNN performs graph construction and graph convolution based on the splicing parameters and outputs the parameters; and the parameters and the sum are weightedly fused to output the shield axis deviation result. The present invention can improve the accuracy and adaptability of shield axis deviation prediction.

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 and its construction method, prediction method and system. Background Art

[0002] Shield tunneling is widely used in underground engineering, and the prediction of axis deviation is crucial. As underground space development scales and becomes more complex, shield tunneling faces numerous challenges. Shield tunneling is a high-noise, long-duration engineering scenario, and the numerous uncertainties inherent in the construction process can affect axis deviation prediction. Furthermore, the interplay of multiple factors during shield tunneling further complicates 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:

[0004] Existing axis deviation prediction methods struggle to fully account for the uncertainty and dynamic interrelationships of multiple shield parameters, resulting in insufficient accuracy and poor adaptability. On the one hand, traditional methods primarily capture temporal features or model temporal and spatial features separately, ignoring the dynamic coupling between parameter features and creating a sense of disconnected spatiotemporal modeling. On the other hand, the black-box effect of deep models results in a lack of interpretability, poor transparency in prediction logic, and difficulty measuring the contributions of multiple factors. Summary of the Invention

[0005] In response to the above defects or improvement needs of the prior art, the present invention provides a shield axis deviation prediction model and its construction method, prediction method and system, the purpose of which is to improve the accuracy and adaptability of shield axis deviation prediction.

[0006] To achieve the above object, according to a first aspect of the present invention, a method for constructing a shield axis deviation prediction model is proposed, comprising the following steps:

[0007] The hybrid model is trained using the training set, and the trained hybrid model is the shield axis deviation prediction model;

[0008] The training set includes splicing parameters and corresponding shield axis deviation data. The splicing parameters are obtained by segmenting the shield parameter time series signal and processing each obtained signal segment to obtain the splicing parameters containing content and position information;

[0009] The hybrid model includes an adaptive neural fuzzy inference system ANFIS and a fully connected spatiotemporal graph neural network model FC-STGNN; the hybrid model takes splicing parameters as input, the adaptive neural fuzzy inference system ANFIS quantifies the splicing parameters, and outputs parameters ; The fully connected spatiotemporal graph neural network model FC-STGNN performs graph construction and graph convolution based on splicing parameters, and outputs parameters ; Then the parameters and Perform weighted fusion and output the shield axis deviation results.

[0010] As a further preferred embodiment, each signal segment is processed, comprising:

[0011] For each signal segment , processed by GRU to obtain feature encoding ; Encode the sinusoidal position and feature encoding Splicing, get splicing parameters .

[0012] As a further preferred method, the sinusoidal position encoding and feature encoding Splicing, get splicing parameters , specifically:

[0013]

[0014]

[0015] in, Indicates code concatenation, d represents the feature dimension, Represents the dimension index, Auxiliary integer variable for determining parity.

[0016] As a further preferred method, the adaptive neural fuzzy inference system ANFIS quantifies the splicing parameters and outputs the parameters , specifically including:

[0017] The Gaussian membership function is used to fuzzify the input splicing parameters to obtain the membership of the parameters to the fuzzy set; then the rule reasoning is performed based on the membership of the parameters to the fuzzy set, and then the fuzzification is performed to obtain the parameters. .

[0018] As a further optimization, the fully connected spatiotemporal graph neural network model FC-STGNN performs graph construction and graph convolution based on splicing parameters, and the output parameters , specifically including:

[0019] Use multiple independent mobile pooling GNN layers to process the input in parallel. For each mobile pooling GNN layer:

[0020] Based on the input splicing parameters, a fully connected graph is constructed;

[0021] The adjacency matrix is ​​obtained by taking the dot product of the similarity matrix E of the fully connected graph and the attenuation matrix C ;

[0022] Based on the adjacency matrix , a sliding window message passing neural network MPNN is used to convolve the fully connected graph to obtain the aggregated features ;

[0023] Aggregate features using activation functions and learnable weights Update and perform time dimension average pooling on the updated features to obtain the output features ;

[0024] The output features of each mobile pooling GNN layer Splicing, and output the spliced ​​feature map through MLP .

[0025] As a further preference, the shield parameters are one or more of the original cutting deviation, original shield tail deviation, jack speed adjustment, total thrust, propulsion jack pressure, propulsion zone stroke, cutterhead speed, front soil pressure, slurry delivery main line pressure and flow, and slurry return main line flow.

[0026] According to a second aspect of the present invention, a shield axis deviation prediction model is provided, which is constructed using the above-mentioned shield axis deviation prediction model construction method.

[0027] According to a third aspect of the present invention, a shield axis deviation prediction method is provided, the shield axis deviation prediction method comprising the following steps:

[0028] The shield parameter time series signal is obtained and segmented, and each obtained signal segment is processed to obtain a splicing parameter containing content and position information; the splicing parameter is input into the shield axis deviation prediction model to obtain a predicted shield axis deviation.

[0029] As a further preferred method, after the shield axis deviation is predicted, GNN Explainer The output of the fully connected spatiotemporal graph neural network model FC-STGNN is interpreted post hoc to obtain the contribution of each shield parameter to the prediction results and the dynamic coupling relationship between each shield parameter.

[0030] According to a fourth aspect of the present invention, a shield axis deviation prediction system is provided, comprising a processor, wherein the processor is configured to execute the above-mentioned shield axis deviation prediction method.

[0031] In general, the above technical solutions conceived by the present invention have the following technical advantages compared with the existing technology:

[0032] 1. The hybrid model designed by the present invention is FC-STGNN The model is based on the ANFIS In the fuzzy processing link, when facing different shield construction scenarios and data characteristics, ANFIS You can adapt to these changes by learning and adjusting, while combining FC-STGNN The ability to capture spatiotemporal dependencies enables the hybrid model used to predict shield axis deviation to have better prediction accuracy, flexibility and adaptability.

[0033] 2. This paper forms a hybrid model of FC-STGNN and ANFIS through weighted fusion. ANFIS optimizes nonlinear relationships from the implicit fuzzy logic level, while FC-STGNN characterizes data patterns from the explicit spatiotemporal dimension. By running both in parallel, comprehensive coverage of data characteristics is achieved, avoiding the impact of successive approaches on prediction timeliness and robustness.

[0034] 3. By constructing a fully connected spatiotemporal graph (FC-STGNN), the construction parameters affecting shield axis offset are correlated from a spatiotemporal perspective. This addresses the problem of traditional approaches focusing on the singleness of temporal relationships or the disconnected nature of modeling time and space separately. Furthermore, the GRU is optimized to replace the 1D CNN for feature extraction, adapting to the long temporal characteristics of shield parameters.

[0035] 4. To address the limited adaptability of the FC-STGNN alone due to the uncertainty of shield parameters, an ANFIS fuzzy system is introduced to quantify the uncertainty in axis deviation prediction. Through fuzzification, fuzzy reasoning, and defuzzification, the uncertain parameters are mapped into interpretable fuzzy logic, and the fuzzy system is dynamically updated through adaptive membership functions.

[0036] 5. To address the black box effect in hybrid models and the difficulty in interpreting the graph structure of the FC-STGNN model, the GNN Explainer is used to quantify the contribution of shield parameters to the prediction results. This solves the problem of poor transparency of the black box model and provides a traceable basis for the results of axis offset prediction. At the same time, it can provide a more fine-grained interpretation using the nodes and edges of the fully connected graph as the basic units. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The figure is a flow chart of a shield axis deviation prediction method according to an embodiment of the present invention.

[0038] Figure 2 This is a diagram of the ANFIS network layer architecture according to an embodiment of the present invention.

[0039] Figure 3 Schematic diagram of FC-STGNN graph construction and graph convolution according to an embodiment of the present invention.

[0040] Figure 4 This is a flowchart of the post-explanation of the GNN Explainer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0042] The embodiment of the present invention provides a method for constructing a shield axis deviation prediction model, such as Figure 1 As shown, the following steps are included:

[0043] S1: Data cleaning and segmentation: The raw data collected by the shield machine sensors are cleaned and segmented into segments of equal length.

[0044] Specifically, the original data is the shield parameter time series signal collected by the sensor; the shield parameters include the original cutting deviation, the original shield tail deviation, the jack speed adjustment, the total thrust, the propulsion jack pressure, the propulsion zone stroke, the cutterhead speed, the front soil pressure, the slurry delivery main line pressure and flow, and the slurry return main line flow.

[0045] Data cleaning and segmentation can improve the accuracy of feature extraction. On the one hand, complex geological conditions and equipment anomalies can cause data fluctuations. Data cleaning can reduce noise, remove redundancy, and unify the format, ensuring the quality of data input to the model. On the other hand, continuous time series with sample lengths need to be segmented into segments of equal length at specific periods for feature extraction. The initial data collected by the sensor at a time length T can be segmented into the following number of segments:

[0046]

[0047] Among them, M represents the number of segments that the sequence signal collected by the sensor can be divided into. f Indicates the segment size. When data is segmented, the size f Should be larger than the sliding window size R of the fully connected spatiotemporal graph convolution, but f If the size is too long, the spatial correlation may be diluted by the time series when constructing a fully connected spatiotemporal graph; if the size is too short, it will lead to mismatch of rule granularity and increase the error in fuzzy membership calculation.

[0048] These fragments constitute the fragment collection , for the ath segment , contains the segmented signals collected by N sensors, that is,

[0049]

[0050] in, Indicates the ath segment The next t Each sensor obtains a shield parameter timing signal. .

[0051] S2: Feature extraction and position encoding: Extract the segment features of each signal segment, obtain the feature encoding, and perform position encoding and splicing to obtain the splicing parameters .

[0052] Further, through GRU right Processing, output feature encoding:

[0053]

[0054] in, Represents GRU's processing of the fragment, and the feature encoding after processing is . GRU The gating mechanism can effectively capture long-term dependencies. The data during shield construction is a typical time series data, and there is a long-term dependency between the data at each time step. GRU Feature extraction and encoding can better capture the long time series parameters in shield construction and improve the performance of the hybrid model compared to 1D CNN.

[0055] Furthermore, sinusoidal position coding is used to generate position vectors and Splice it so that it contains both content and location information.

[0056]

[0057]

[0058] in, Indicates combining position coding and feature coding. represents the sinusoidal position encoding, d represents the feature dimension, Represents the dimension index, Auxiliary integer variable for determining parity.

[0059] S3: The hybrid model is trained using the training set. The trained hybrid model is the shield axis deviation prediction model. The hybrid model includes the adaptive neural fuzzy inference system ANFIS and the fully connected spatiotemporal graph neural network model FC-STGNN. The hybrid model is based on the splicing parameters As input, the adaptive neuro-fuzzy inference system ANFIS calculates the splicing parameters Quantize and output parameters ; Fully connected spatiotemporal graph neural network model FC-STGNN based on splicing parameters Perform graph construction and graph convolution, output parameters ; Then the parameters and Perform weighted fusion and output the shield axis deviation results. The details are as follows:

[0060] (1) ANFIS quantification of uncertainty: Based on the ANN adaptive network architecture of ANFIS and FIS fuzzy reasoning, the input parameters are fuzzified, reasoning is performed based on fuzzy rules, and defuzzification is performed to achieve quantification of uncertain parameters and output parameters. .

[0061] Adaptive neuro-fuzzy inference system (ANFIS) achieves the coordinated 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. FC-STGNN When using the FC-STGNN model, only the spatiotemporal dependencies in the data can be modeled, but the effective integration of domain knowledge is lacking. However, by adding the ANFIS model in parallel in a weighted manner on the basis of the FC-STGNN model, the problem of the limited modeling ability of the highly nonlinear and strongly coupled parameter relationships in shield construction (such as the nonlinear influence of the dynamic coupling of earth pressure and shield posture adjustment on the axis deviation) when the FC-STGNN model is used alone can be solved.

[0062] ANFIS modularly maps fuzzification (Layer 1), rule reasoning (Layer 2-4) and defuzzification (Layer 5) into neural network nodes. The neural network architecture consists of five layers: fuzzification layer, rule layer, normalization layer, reasoning layer, and defuzzification layer. Figure 2 The specific steps are as follows:

[0063] (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.

[0064] Since the Gaussian membership function has stronger adaptability and robustness and is suitable for dynamic and variable shield construction parameters, this embodiment adopts the Gaussian membership function:

[0065]

[0066] in, represents the membership function, Indicates parameter input,j Indicates the index of the parameter input; and are the mean and standard deviation of the Gaussian membership function, respectively.

[0067] The partitioning method adopts a dichotomous approach and uses “high-low” semantic labels to map the data into different fuzzy sets.

[0068] (1.2) Rule Layer: After processing at the fuzzy layer, each input parameter will have a membership degree corresponding to a fuzzy set. Based on the fuzzified membership degrees, statistical measures of the fuzzy membership degrees within the time window (mean, variance, entropy), and domain knowledge, if-then logic rules can be generated. The if part is the antecedent, which is a combination of membership degrees, while the then part is the consequent, which is responsible for the combination calculation. These rules are aggregated and integrated to form a systematic rule base.

[0069] The weight of each rule is determined by its firing strength:

[0070]

[0071] in i Indicates the rule index, x and y Refers to different input parameters, express Fuzzy Set The membership degree of Indicates the i The excitation intensity of the rule.

[0072] (1.3) Normalization layer: Convert the rule excitation intensity into relative weights, quantify the contribution rate of each rule, and dynamically balance the competitive relationship between rules.

[0073]

[0074] in, Indicates the i The normalized weights of the rules.

[0075] (1.4) Reasoning layer: The consequent (THEN part) of each fuzzy rule is combined and calculated to generate a quantitative output related to the input parameters. The formula for the combined calculation function is:

[0076]

[0077] in, Indicates the i The linear function corresponding to the rule, , , represents a learnable parameter.

[0078] The mathematical essence of combinatorial computing lies in dividing and fusing the conclusions of linear rules through weights, and ultimately constructing a global nonlinear mapping that can approximate any continuous function.

[0079] (1.5) Defuzzification layer: Aggregate all rule outputs to a node, which takes the sum of all input signals as the total output, that is:

[0080]

[0081] in, Indicates the output result.

[0082] The adaptability of ANFIS fuzzy reasoning is that the fuzzy layer (membership parameter , ) and the inference layer (regular linear combination coefficients ) parameters are dynamically adjusted to adapt to the data distribution. Therefore, they are adaptive nodes and are represented as square nodes. The computational logic of other network layers is fixed and they perform predefined operations. Therefore, they are fixed nodes and are represented as circular nodes.

[0083] (2) FC-STGNN captures spatiotemporal dependencies: constructs a fully connected graph, uses a sliding window message passing neural network MPNN for graph convolution, and obtains a fully connected spatiotemporal graph, outputting parameters .

[0084] Specifically, such as Figure 3 As shown, building FC-STGNN includes graph construction and graph convolution:

[0085] (2.1) For input parameters , forming a set , look at the node set of the fully connected graph. Through the feature enhancement function Improve the expressiveness of features and measure the similarity between different features through dot product:

[0086]

[0087]

[0088] in, Indicates the a The first fragment t node features and b The first fragment u The similarity between node features.

[0089] Based on this, the similarity matrix of the fully connected graph is summarized: , , that is, the features of all fragments and their correlations, can be used as the set of edges of the fully connected graph. Thus, the fully connected graph can be obtained .

[0090] The FC-STGNN measures the temporal correlation between features by constructing a decay matrix. The decay rate of the temporal correlation between each segment is : By analogy, the spatiotemporal correlation between any features in different segments is:

[0091]

[0092] in, represents the temporal correlation between the t-th node feature of the a-th segment and the u-th node feature of the b-th segment. This constructs the decay matrix of the fully connected graph: .

[0093] Then, the adjacency matrix is ​​obtained by taking the dot product of the similarity matrix E and the attenuation matrix C. Its elements can reflect the association of different features in different time segments under the influence of time distance:

[0094]

[0095]

[0096] in, It represents the edge weights of different features in different time segments under the influence of time distance, and the normalized correlation after adjustment by the attenuation matrix. for The matrix formed.

[0097] (2.2) Convolution of a fully connected graph refers to the use of a sliding window message passing neural network (MPNN) to capture the dynamic dependencies in a fully connected spatiotemporal graph. The fully connected spatiotemporal graph is divided into windows of equal size along the time dimension with a step size S and a window size R. For the wth sliding window, the l Layer center node , for the time range All nodes within are convolved:

[0098]

[0099] in, Indicates the l Layer, timestamp a ,sensor t The aggregation features are obtained by aggregating the information of adjacent nodes in the window. w Represents the center index of the current sliding window, which is used to locate the window position in the time dimension. R represents the sliding window size. No.l Layer, timestamp a ,sensor t The node features include spatiotemporal information and position encoding.

[0100] Using activation functions and learnable weights Update features, enhance the nonlinear expression capability of features, and adapt to complex spatiotemporal patterns:

[0101]

[0102] in, represents the linear rectification activation function, Indicates the l The updated features of layer +1, timestamp a, and sensor t are obtained by processing the features of the current layer. represents the learnable weights.

[0103] Perform time-dimensional average pooling on the updated features within each window to extract high-level features and optimize the model's processing efficiency for long sequence data:

[0104]

[0105] in, Indicates that after pooling, the timestamp w ,sensor t In the l The output features of the +1 layer are obtained by averaging the features within the window.

[0106] Multiple independent mobile pooling GNN layers are used to process the input in parallel. The features after each layer of time pooling need to be spliced ​​to integrate multi-scale spatiotemporal information, and finally the spliced ​​feature map is output through MLP. This process can be expressed as:

[0107]

[0108] in, represents the predicted output of FC-STGNN, k represents the number of GNN parallel layers, It represents the process of splicing the features of each layer. MLP represents the multi-layer perceptron. The spliced ​​features are mapped to the final prediction results through nonlinear transformation, completing the conversion from features to output.

[0109] (3) Output weighted fusion: At the end of the model parallelization, the output results of ANFIS and FC-STGNN are weighted and fused to obtain the prediction result.

[0110] Specifically, the ANFIS model and FC-STGNN in the hybrid model run independently of each other. At the end of the model, a weighted fusion is used to set the weight W to merge the output results of the two:

[0111]

[0112] in, Represents the final output of the hybrid model, W Indicates the weight of the ANFIS model in the hybrid model, W The higher it is, the higher the influence of the ANFIS model on the prediction results.

[0113] The embodiment of the present invention further provides a shield axis deviation prediction method, comprising the following steps:

[0114] S4: Obtain the shield parameter time series signal and segment it, process each obtained signal segment, and obtain the splicing parameters containing content and position information. For details, please refer to the methods in S1 and S2, which will not be repeated here; input the splicing parameters into the shield axis deviation prediction model to obtain the shield axis deviation prediction value.

[0115] S5: GNN Explainer post-explanation: The input of the post-explanation is trained FC-STGNN Model and fully connected spatiotemporal graph G, FC-STGNN The initial prediction value of GNN Explainer uses edges and nodes as the explanation granularity, reveals the decision basis of the model by generating explainable subgraphs and related node features, and provides post-explanation for the axis deviation prediction results, thereby improving the reliability and transparency of the model explanation. GNN The output of the graph-like neural network can be explained and displayed. FC-STGNN The importance of nodes and edges in the model to the output prediction results solves the black box problem of the hybrid model.

[0116] like Figure 4 As shown, the specific steps include:

[0117] (1) By creating Explainer The object initializes the interpreter and mask. Based on the given graph structure and node features, an interpretable mask can be generated:

[0118]

[0119] in, Represents a 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 unmasked.

[0120] (2) Mask By changing its own matrix elements to cover some nodes or edges of the fully connected spatiotemporal graph G, the fully connected graph structure is adjusted to generate candidate subgraphs. .like Then remove the nodes; if Then remove the This process is expressed as:

[0121]

[0122]

[0123] in, represents the set of nodes of the subgraph, represents the edge set of the subgraph; and is a one-dimensional index, a The first fragment t node features and b The first fragment u The two-dimensional information of each node is mapped into a one-dimensional vector index to facilitate the operation of the mask matrix.

[0124] (3) The obtained candidate subgraph enter FC-STGNN The model is constructed and the difference between its prediction results and the initial prediction output is used to filter out the nodes and edges that have a greater impact on the model prediction results.

[0125] This process uses the mean square error MSE Expressed as:

[0126]

[0127] in, represents the prediction difference loss, Represents a fully connected spatiotemporal graph model. The loss function is designed to analyze the existing FC-STGNN The prediction logic of the model optimizes temporary mask weights, and the data range is the current input data. Its essence belongs to decision analysis rather than model training.

[0128] Optimizing the mask via gradient ascent , the optimized mask matrix is ​​obtained by maximizing the loss. This process is expressed as:

[0129]

[0130] in, represents the optimized mask matrix, argmax Represents finding the loss function Maximized mask , that is, optimizing to determine the part of the graph structure that most affects the model prediction.

[0131] (4) GNN Explainer normalizes the optimized mask to obtain the node weight and edge weights , passing the importance threshold Filter out important nodes and edges to generate the final explanation subgraph , which only retains the graph structure that contributes most to the model prediction, that is, contains the most critical information for the model prediction results. This process is:

[0132]

[0133]

[0134] The outputs that can be interpreted afterwards include the final explanation subgraph and the importance weights of the optimized mask. FC-STGNN The black box nature of the model is made transparent, showing the degree of temporal and spatial correlation between shield parameters and analyzing the factors most relevant to the shield axis deviation prediction results.

[0135] On the one hand, the 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 (jack speed) is high, it means that the synergistic effect of 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 results (i.e., the prediction subgraph) can also show the connection between parameters within a single time step. For example, the axis deviation prediction may be based on x6 as the hub node, forming a dense connection with x14 and x20, indicating that the total thrust indirectly controls the vertical deviation by affecting the cutterhead torque and slurry flow rate. Second, for the relationship between time steps, for example, the speed at the tth time step node is of high importance, which can indicate that the speed adjustment at the previous moment plays a key role in the prediction of the current axis deviation.

[0136] In summary, the present invention cleans and segments the captured shield timing parameters and performs feature encoding and position encoding, and then inputs the parameters into a hybrid model composed of a fully connected spatiotemporal graph neural network (FC-STGNN) and an adaptive neural fuzzy inference system (ANFIS): the FC-STGNN constructs a fully connected spatiotemporal graph based on the correlation matrix and the attenuation matrix, and after segmentation using a sliding window, performs graph convolution to capture spatiotemporal dependencies; the ANFIS model fuzzifies the input parameters, combines fuzzy rule reasoning and defuzzification; the two models run independently and in parallel, and then weightedly fuse the outputs; the model decision process can also be explained post hoc through the GNN Explainer.

[0137] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a shield axis deviation prediction model, characterized in that: The steps include: The hybrid model is trained using 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 splicing parameters are obtained by segmenting the shield parameter time series signal and processing each obtained signal segment to obtain the splicing parameters containing content and position information; The hybrid model includes an adaptive neural fuzzy inference system ANFIS and a fully connected spatiotemporal graph neural network model FC-STGNN; the hybrid model takes splicing parameters as input, the adaptive neural fuzzy inference system ANFIS quantifies the splicing parameters, and outputs parameters , specifically including: The Gaussian membership function is used to fuzzify the input splicing parameters to obtain the membership of the parameters to the fuzzy set; then the rule reasoning is performed based on the membership of the parameters to the fuzzy set, and then the fuzzification is performed to obtain the parameters. ; The fully connected spatiotemporal graph neural network model FC-STGNN performs graph construction and graph convolution based on splicing parameters, and outputs parameters , specifically including: Use multiple independent mobile pooling GNN layers to process the input in parallel. For each mobile pooling GNN layer: Based on the input splicing parameters, a fully connected graph is constructed; The adjacency matrix is ​​obtained by taking the dot product of the similarity matrix E of the fully connected graph and the attenuation matrix C ; Based on the adjacency matrix , a sliding window message passing neural network MPNN is used to convolve the fully connected graph to obtain the aggregated features ; Aggregate features using activation functions and learnable weights Update and perform time dimension average pooling on the updated features to obtain the output features ; The output features of each mobile pooling GNN layer Splicing, and output the spliced ​​feature map through MLP ; Then the parameters and Perform weighted fusion and output the shield axis deviation results.

2. The method for constructing a shield axis deviation prediction model according to claim 1, wherein: Process each signal segment, including: For each signal segment , processed by GRU to obtain feature encoding ; Encode the sinusoidal position and feature encoding Splicing, get splicing parameters .

3. The method for constructing a shield axis deviation prediction model according to claim 2, wherein: Encode the sinusoidal position and feature encoding Splicing, get splicing parameters , specifically: in, Indicates code concatenation, d represents the feature dimension, Represents the dimension index, Auxiliary integer variable for determining parity.

4. The method for constructing a shield axis deviation prediction model according to any one of claims 1 to 3, characterized in that: The shield parameters include one or more of the original cutting deviation, original shield tail deviation, jack speed adjustment, total thrust, propulsion jack pressure, propulsion zone stroke, cutterhead speed, front soil pressure, slurry delivery main line pressure and flow, and slurry return main line flow.

5. A shield axis deviation prediction method, characterized in that: The steps include: Acquire the shield parameter time series signal and segment it, process each obtained signal segment, and obtain the splicing parameter containing content and position information; input the splicing parameter into the shield axis deviation prediction model constructed by the method for constructing the shield axis deviation prediction model according to any one of claims 1 to 4, and obtain the predicted shield axis deviation.

6. The shield axis deviation prediction method according to claim 5, characterized in that: After predicting the shield axis deviation, GNN Explainer The output of the fully connected spatiotemporal graph neural network model FC-STGNN is interpreted post hoc to obtain the contribution of each shield parameter to the prediction results and the dynamic coupling relationship between each shield parameter.

7. A shield axis deviation prediction system, characterized in that: It includes a processor, which is used to execute the shield axis deviation prediction method as described in claim 5 or 6.

Images