Shield tunnel parameter prediction method and system based on deep learning
By combining self-attention convolution bidirectional long and short-term model and improved flower pollination algorithm, the problem of complex parameter changes and difficult model optimization in shield tunnel parameter prediction is solved, and higher prediction accuracy and model optimization effect are achieved .
Patent Information
- Application Number
- CN202411469356.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-21
- Publication Date
- 2025-02-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the existing shield tunnel parameter prediction methods, parameter changes are complex, time changes and other potential characteristics need to be considered, and model parameters are complex, and the optimization algorithm has problems of redundant features and inefficient calculations.
The convolutional bidirectional long-term model combined with self-attention is used to construct predictive models to improve the understanding of complex patterns and feature capture capabilities; the improved flower pollination algorithm is used for model optimization, and the elite replacement mechanism optimization algorithm is introduced to retain important features.
It improves the accuracy and adaptability of shield tunnel parameter prediction, improves the model optimization effect, and adapts to complex prediction tasks.
Smart Images
Figure CN119442863A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of intelligent construction technology, and specifically refers to a shield tunnel parameter prediction method and system based on deep learning. Background Art
[0002] Shield tunnel parameter prediction uses deep learning technology to model and predict key parameters generated during shield tunnel construction, thereby predicting potential safety issues in the construction process in advance, reducing the probability of accidents, and ensuring safety during construction.
[0003] However, in the existing shield tunnel parameter prediction methods, there is a technical problem that the changes of various parameters during the shield tunnel construction process are relatively complex, and it is necessary to consider not only the parameter changes caused by complex time changes, but also the influence of other complex potential characteristics; the parameters of the shield tunnel parameter prediction model are complex and have many levels, which increases the difficulty of calculating the optimal model parameter combination, and the traditional parameter optimization algorithm inevitably has the technical problem of many redundant features and low computational efficiency. Summary of the invention
[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a shield tunnel parameter prediction method and system based on deep learning. In view of the technical problem that various parameter changes in the shield tunnel construction process are relatively complex in the existing shield tunnel parameter prediction method, it is necessary to consider both the parameter changes caused by complex time changes and the influence of other complex potential features. This solution creatively adopts a convolutional bidirectional long-term and short-term model combined with self-attention to construct a prediction model, which improves the ability to understand complex patterns and capture features, and improves the adaptability to tunnel parameter changes in the shield tunnel parameter prediction process, thereby improving the accuracy of the overall task; In view of the technical problem that the parameters of the shield tunnel parameter prediction model are complex and have many levels in the existing shield tunnel parameter prediction method, the difficulty of calculating the optimal model parameter combination is also increased, and the traditional parameter optimization algorithm inevitably has many redundant features and low computational efficiency. This solution creatively adopts an improved flower pollination algorithm to construct a model optimization layer, and introduces an elite replacement mechanism to optimize the algorithm's ability to retain important features in the parameter combination, thereby improving the optimization effect of the complex shield tunnel parameter prediction model, and at the same time, it can also be more adaptable to complex prediction tasks.
[0005] The technical solution adopted by the present invention is as follows: The present invention provides a shield tunnel parameter prediction method based on deep learning, the method comprising the following steps:
[0006] Step S1: Shield tunnel parameter prediction modeling;
[0007] Step S2: data preprocessing layer construction;
[0008] Step S3: parameter prediction layer construction;
[0009] Step S4: Model optimization layer construction;
[0010] Step S5: Shield tunnel parameter prediction.
[0011] Furthermore, in step S1, the shield tunnel parameter prediction modeling is used to design and construct a basic framework for shield tunnel parameter prediction, specifically to construct a prediction model for shield tunnel parameters, and the prediction model specifically includes a data preprocessing layer, a parameter prediction layer and a model optimization layer.
[0012] Furthermore, in step S2, the data preprocessing layer is constructed to collect shield tunnel related parameters and perform data preprocessing, including the following steps:
[0013] Step S21: data collection, specifically collecting shield operation parameters, tunnel geological parameters and tunnel space parameters from the shield tunnel construction process to obtain shield tunnel parameters;
[0014] Step S22: Data preprocessing, specifically preprocessing the shield tunnel parameters through data cleaning, noise filtering and normalization to obtain shield tunnel enhancement data.
[0015] Further, in step S3, the parameter prediction layer is constructed to construct a prediction layer of the prediction model, specifically, a convolutional bidirectional long-term and short-term model combined with self-attention is used to construct the parameter prediction layer to obtain a shield tunnel parameter prediction model;
[0016] The convolutional bidirectional long-term and short-term model combined with self-attention includes an input layer, a convolutional network layer, a deep self-attention block, a bidirectional long-term and short-term network layer, and a fully connected output layer;
[0017] The convolutional network layer is used to extract nonlinear local features and perform feature dimensionality reduction;
[0018] The deep self-attention block is used to extract key features and suppress redundant features;
[0019] The bidirectional long-term and short-term network layer is used to mine time dependencies;
[0020] The parameter prediction layer is constructed, comprising the following steps:
[0021] Step S31: constructing an input layer, specifically receiving shield tunnel enhancement data as model input through the input layer;
[0022] Step S32: constructing a convolutional network layer, specifically setting a one-dimensional convolutional layer and a pooling layer in the convolutional network layer, extracting nonlinear local features of the model input through the one-dimensional convolutional layer, and then performing feature dimensionality reduction on the nonlinear local features through the pooling layer to obtain convolutional pooling features;
[0023] Step S33: construct a deep self-attention block, specifically by replacing all fully connected layers in the self-attention block with a deep separable convolutional layer, improving the self-attention block to obtain a deep self-attention block, and calculating the key features through the deep self-attention block, including the following steps:
[0024] Step S331: Feature mapping, specifically mapping the convolutional pooling features through three independent depth-separable convolutional layers to obtain a query vector, a key vector and a value vector. The calculation formula is:
[0025] ;
[0026] Where Q is the query vector, DSC q (·) is the depth-wise separable convolution query vector calculation function, Y is the convolution pooling feature, K is the key vector, and DSC k (·) is the depthwise separable convolution key vector calculation function, V is the value vector, DSC v (·) is the depth-wise separable convolution value vector calculation function;
[0027] Step S332: Calculate the self-attention result of each attention head. The calculation formula is:
[0028] ;
[0029] Where i is the self-attention head index, self i (Y) is the attention result of the i-th self-attention head, δ(·) is the softmax function, T is the transposition operation, dim i is the hidden dimension of the i-th self-attention head;
[0030] Step S333: Apply the multi-head self-attention mechanism to concatenate the self-attention results of each attention head, and perform a depth-separable convolution operation to obtain the multi-head self-attention. The calculation formula is:
[0031] ;
[0032] Where Mself(Y) is the multi-head self-attention, Ms conv (·) is a depth-wise separable convolution operation used to aggregate multi-head self-attention information, Concat[·] is a concatenation operation, α is an adjustment parameter used to control the ratio of attention to high-frequency information and low-frequency information, and hn is the number of self-attention heads;
[0033] Step S334: performing feature aggregation on the convolution pooling features through the deep self-attention block to obtain key features;
[0034] Step S34: constructing a bidirectional long-short term network layer, specifically constructing a bidirectional long-short term network layer including a forward long-short term network sublayer and a backward long-short term network sublayer, calculating key features through the bidirectional long-short term network layer to obtain bidirectional time series features;
[0035] Step S35: construct a fully connected output layer, calculate the bidirectional time series features through the fully connected output layer, generate and output the prediction results;
[0036] Step S36: parameter prediction model training, specifically, performing model training by constructing the input layer, constructing the convolutional network layer, constructing the deep self-attention block, constructing the bidirectional long-short term network layer and constructing the fully connected output layer to obtain a shield tunnel parameter prediction model.
[0037] Further, in step S4, the model optimization layer is constructed to optimize model parameters and performance, specifically using an improved flower pollination algorithm to construct the model optimization layer to obtain an optimal shield tunnel parameter prediction model, including the following steps:
[0038] Step S41: Initialize input parameters, specifically construct input parameter combinations, and the calculation formula is:
[0039] ;
[0040] Where IFO is the input parameter combination, f is the objective function, u is the horizontal constraint of the search space, v is the vertical constraint of the search space, and es is the scale parameter of the elite replacement mechanism;
[0041] Step S42: Initializing the population, specifically initializing the flower population position, the flower population position is used to represent the parameter combination of the shield tunnel parameter prediction model output by the model optimization layer, and calculating the fitness values of all the initialized flower population positions to obtain the best initial fitness, and sorting all the initialized flower population positions in ascending order according to the fitness value;
[0042] Step S43: subgroup division, specifically adopting a dynamic adjustment method to update the fitness, and obtaining an updated optimal fitness through the fitness update, and performing subgroup division according to the updated optimal fitness to obtain population division data;
[0043] Step S44: search selection optimization, specifically constructing a random threshold parameter and performing search selection optimization of the flower population position to obtain updated optimal flower population position data, and the calculation formula is:
[0044] ;
[0045] In the formula, is to update the optimal flower population location data, It is the flower index, is the current flower population position, is the learning rate parameter, is the optimal flower of the current iteration, is the random perturbation parameter, is the first random flower, is the second random flower, rand is the random threshold discrimination parameter, which is used to perform the search selection optimization search method. When rand>0.8, a global search is performed, otherwise a local search is performed. is the local search adjustment factor, P r is the random micro-disturbance parameter;
[0046] Step S45: Optimize the elite replacement mechanism to ensure that the high-quality flower population positions can be retained in each iteration. Specifically, the total number of sub-group flower positions in each population partition data is calculated, and the sub-group flower positions in each population partition data are randomly arranged. The flower positions with the highest fitness es are selected from the arrangement as elite replacement flower positions, and the flower positions with the lowest fitness value ranking of 5% are replaced with the elite replacement flower positions to optimize the flower pollination search algorithm.
[0047] Step S46: constructing an iteration stop condition, specifically setting the maximum number of iterations for iteration stop to 1000, and taking the absolute difference between the current best fitness and the previous best fitness to be less than 0.01 as a fitting condition, and stopping the iterative calculation when the maximum number of iterations is reached or the fitting condition is met;
[0048] Step S47: iterative optimization, specifically, based on the iteration stop condition, the model optimization layer is constructed through the operations of step S41 to step S45 to obtain the optimal shield tunnel parameter prediction model.
[0049] Furthermore, in step S5, the shield tunnel parameter prediction is specifically performed by constructing the data preprocessing layer, the parameter prediction layer and the model optimization layer to predict the shield tunnel parameters, and by using the optimal shield tunnel parameter prediction model, the shield tunnel surface settlement prediction data is obtained, and the shield tunnel surface settlement risk level is divided in combination with engineering standards.
[0050] The present invention provides a shield tunnel parameter prediction system based on deep learning, comprising: a shield tunnel parameter prediction modeling module, a data preprocessing layer construction module, a parameter prediction layer construction module, a model optimization layer construction module and a shield tunnel parameter prediction module;
[0051] The shield tunnel parameter prediction modeling module is used for shield tunnel parameter prediction modeling. Through the shield tunnel parameter prediction modeling, an infrastructure for shield tunnel parameter prediction is obtained, and a data preprocessing layer, a parameter prediction layer and a model optimization layer are constructed based on the infrastructure for shield tunnel parameter prediction.
[0052] The data preprocessing layer construction module is used for data preprocessing layer construction, through which shield tunnel enhancement data is obtained, and the shield tunnel enhancement data is sent to the parameter prediction layer construction module and the model optimization layer construction module;
[0053] The parameter prediction layer construction module is used for parameter prediction layer construction, and obtains a shield tunnel parameter prediction model through parameter prediction layer construction, and sends the shield tunnel parameter prediction model to the model optimization layer construction module;
[0054] The model optimization layer construction module is used for model optimization layer construction, and obtains the optimal shield tunnel parameter prediction model through model optimization layer construction, and sends the optimal shield tunnel parameter prediction model to the shield tunnel parameter prediction module;
[0055] The shield tunnel parameter prediction module is used for shield tunnel parameter prediction. By constructing the data preprocessing layer, parameter prediction layer and model optimization layer, shield tunnel parameter prediction is performed, and by using the optimal shield tunnel parameter prediction model, shield tunnel surface settlement prediction data is obtained.
[0056] The beneficial effects achieved by the present invention using the above scheme are as follows:
[0057] (1) In view of the technical problem that the existing shield tunnel parameter prediction methods have complex parameter changes during the shield tunnel construction process, it is necessary to consider both the parameter changes caused by complex time changes and the influence of other complex potential features. This scheme creatively adopts a convolutional bidirectional long-term and short-term model combined with self-attention to construct a prediction model, which improves the ability to understand complex patterns and capture features, and improves the adaptability to tunnel parameter changes during the shield tunnel parameter prediction process, thereby improving the accuracy of the overall task;
[0058] (2) In the existing shield tunnel parameter prediction methods, the parameters of the shield tunnel parameter prediction model are complex and have many levels, which increases the difficulty of calculating the optimal model parameter combination. The traditional parameter optimization algorithm inevitably has many redundant features and low computational efficiency. This scheme creatively uses an improved flower pollination algorithm to construct the model optimization layer. By introducing the elite replacement mechanism to optimize the algorithm's ability to retain important features in the parameter combination, it improves the optimization effect of the complex shield tunnel parameter prediction model and can also better adapt to complex prediction tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 A schematic diagram of a process flow of a shield tunnel parameter prediction method based on deep learning provided by the present invention;
[0060] Figure 2 A schematic diagram of a shield tunnel parameter prediction system based on deep learning provided by the present invention;
[0061] Figure 3 is a schematic flow chart of step S3;
[0062] Figure 4 It is a schematic diagram of the process of step S4.
[0063] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION
[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0065] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.
[0066] Example 1, see Figure 1 The present invention provides a shield tunnel parameter prediction method based on deep learning, which comprises the following steps:
[0067] Step S1: Shield tunnel parameter prediction modeling;
[0068] Step S2: data preprocessing layer construction;
[0069] Step S3: parameter prediction layer construction;
[0070] Step S4: Model optimization layer construction;
[0071] Step S5: Prediction of shield tunnel parameters.
[0072] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the shield tunnel parameter prediction modeling is used to design and construct a basic framework for shield tunnel parameter prediction, specifically to build a prediction model for shield tunnel parameters. The prediction model specifically includes a data preprocessing layer, a parameter prediction layer and a model optimization layer.
[0073] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the data preprocessing layer is constructed to collect shield tunnel related parameters and perform data preprocessing, including the following steps:
[0074] Step S21: data collection, specifically collecting shield operation parameters, tunnel geological parameters and tunnel space parameters from the shield tunnel construction process to obtain shield tunnel parameters;
[0075] The shield operation parameters include grouting volume, prop temperature, chamber soil temperature, forward speed, rotation speed, shield posture and penetration depth;
[0076] The tunnel geological parameters include friction angle, cohesion, density and Poisson's ratio;
[0077] The tunnel space parameters include the overlying stratum depth and burial depth ratio of the shield tunnel;
[0078] Step S22: data preprocessing, specifically preprocessing the shield tunnel parameters through data cleaning, noise filtering and normalization to obtain shield tunnel enhancement data;
[0079] The data cleaning includes outlier removal and missing value filling operations.
[0080] Example 4, see Figure 1 and Figure 3 , this embodiment is based on the above embodiment. In step S3, the parameter prediction layer is constructed to construct a prediction layer of a prediction model. Specifically, a convolutional bidirectional long-term and short-term model combined with self-attention is used to construct a parameter prediction layer to obtain a shield tunnel parameter prediction model;
[0081] The convolutional bidirectional long-term and short-term model combined with self-attention includes an input layer, a convolutional network layer, a deep self-attention block, a bidirectional long-term and short-term network layer, and a fully connected output layer;
[0082] The convolutional network layer is used to extract nonlinear local features and perform feature dimensionality reduction;
[0083] The deep self-attention block is used to extract key features and suppress redundant features;
[0084] The bidirectional long-term and short-term network layer is used to mine time dependencies;
[0085] The parameter prediction layer is constructed, comprising the following steps:
[0086] Step S31: constructing an input layer, specifically receiving shield tunnel enhancement data as model input through the input layer;
[0087] Step S32: constructing a convolutional network layer, specifically setting a one-dimensional convolutional layer and a pooling layer in the convolutional network layer, extracting nonlinear local features of the model input through the one-dimensional convolutional layer, and then performing feature dimensionality reduction on the nonlinear local features through the pooling layer to obtain convolutional pooling features;
[0088] Step S33: construct a deep self-attention block, specifically by replacing all fully connected layers in the self-attention block with a deep separable convolutional layer, improving the self-attention block to obtain a deep self-attention block, and calculating the key features through the deep self-attention block, including the following steps:
[0089] Step S331: Feature mapping, specifically mapping the convolutional pooling features through three independent depth-separable convolutional layers to obtain a query vector, a key vector and a value vector. The calculation formula is:
[0090] ;
[0091] Where Q is the query vector, DSC q (·) is the depth-wise separable convolution query vector calculation function, Y is the convolution pooling feature, K is the key vector, and DSC k (·) is the depthwise separable convolution key vector calculation function, V is the value vector, DSC v (·) is the depth-wise separable convolution value vector calculation function;
[0092] Step S332: Calculate the self-attention result of each attention head. The calculation formula is:
[0093] ;
[0094] Where i is the self-attention head index, self i (Y) is the attention result of the i-th self-attention head, δ(·) is the softmax function, T is the transposition operation, dim i is the hidden dimension of the i-th self-attention head;
[0095] Step S333: Apply the multi-head self-attention mechanism to concatenate the self-attention results of each attention head, and perform a depth-separable convolution operation to obtain the multi-head self-attention. The calculation formula is:
[0096] ;
[0097] Where Mself(Y) is the multi-head self-attention, Ms conv (·) is a depth-wise separable convolution operation used to aggregate multi-head self-attention information, Concat[·] is a concatenation operation, α is an adjustment parameter used to control the ratio of attention to high-frequency information and low-frequency information, and hn is the number of self-attention heads;
[0098] Step S334: performing feature aggregation on the convolution pooling features through the deep self-attention block to obtain key features;
[0099] Step S34: constructing a bidirectional long-short term network layer, specifically constructing a bidirectional long-short term network layer including a forward long-short term network sublayer and a backward long-short term network sublayer, calculating key features through the bidirectional long-short term network layer to obtain bidirectional time series features;
[0100] Step S35: construct a fully connected output layer, calculate the bidirectional time series features through the fully connected output layer, generate and output the prediction results;
[0101] Step S36: parameter prediction model training, specifically, performing model training by constructing the input layer, constructing the convolutional network layer, constructing the deep self-attention block, constructing the bidirectional long-short term network layer, and constructing the fully connected output layer to obtain a shield tunnel parameter prediction model;
[0102] By performing the above operations, in order to solve the technical problem that in the existing shield tunnel parameter prediction methods, various parameter changes during the shield tunnel construction process are relatively complex, and it is necessary to consider both the parameter changes caused by complex time changes and the influence of other complex potential features. This scheme creatively adopts a convolutional bidirectional long-term and short-term model combined with self-attention to construct a prediction model, which improves the ability to understand complex patterns and capture features, and improves the adaptability to tunnel parameter changes in the shield tunnel parameter prediction process, thereby improving the accuracy of the overall task.
[0103] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S4, the model optimization layer is constructed to optimize model parameters and performance. Specifically, the model optimization layer is constructed using an improved flower pollination algorithm to obtain an optimal shield tunnel parameter prediction model, including the following steps:
[0104] Step S41: Initialize input parameters, specifically construct input parameter combinations, and the calculation formula is:
[0105] ;
[0106] Where IFO is the input parameter combination, f is the objective function, u is the horizontal constraint of the search space, v is the vertical constraint of the search space, and es is the scale parameter of the elite replacement mechanism;
[0107] Step S42: Initializing the population, specifically initializing the flower population position, the flower population position is used to represent the parameter combination of the shield tunnel parameter prediction model output by the model optimization layer, and calculating the fitness values of all the initialized flower population positions to obtain the best initial fitness, and sorting all the initialized flower population positions in ascending order according to the fitness value;
[0108] Step S43: subgroup division, specifically adopting a dynamic adjustment method to update the fitness, and obtaining an updated optimal fitness through the fitness update, and performing subgroup division according to the updated optimal fitness to obtain population division data;
[0109] Step S44: search selection optimization, specifically constructing a random threshold parameter and performing search selection optimization of the flower population position to obtain updated optimal flower population position data, and the calculation formula is:
[0110] ;
[0111] In the formula, is to update the optimal flower population location data, It is the flower index, is the current flower population position, is the learning rate parameter, is the optimal flower of the current iteration, is the random perturbation parameter, is the first random flower, is the second random flower, rand is the random threshold discrimination parameter, which is used to perform the search selection optimization search method. When rand>0.8, a global search is performed, otherwise a local search is performed. is the local search adjustment factor, P r is the random micro-disturbance parameter;
[0112] Step S45: Optimize the elite replacement mechanism to ensure that the high-quality flower population positions can be retained in each iteration. Specifically, the total number of sub-group flower positions in each population partition data is calculated, and the sub-group flower positions in each population partition data are randomly arranged. The flower positions with the highest fitness es are selected from the arrangement as elite replacement flower positions, and the flower positions with the lowest fitness value ranking of 5% are replaced with the elite replacement flower positions to optimize the flower pollination search algorithm.
[0113] Step S46: constructing an iteration stop condition, specifically setting the maximum number of iterations for iteration stop to 1000, and taking the absolute difference between the current best fitness and the previous best fitness to be less than 0.01 as a fitting condition, and stopping the iterative calculation when the maximum number of iterations is reached or the fitting condition is met;
[0114] Step S47: iterative optimization, specifically, constructing a model optimization layer through operations from step S41 to step S45 according to the iteration stop condition to obtain an optimal shield tunnel parameter prediction model;
[0115] By performing the above operations, in the existing shield tunnel parameter prediction methods, the parameters of the shield tunnel parameter prediction model are complex and have many levels, which increases the difficulty of calculating the optimal model parameter combination. The traditional parameter optimization algorithm inevitably has many redundant features and low computational efficiency. Technical problems, this scheme creatively adopts an improved flower pollination algorithm to construct the model optimization layer, and introduces the elite replacement mechanism to optimize the algorithm's ability to retain important features in the parameter combination, thereby improving the optimization effect of the complex shield tunnel parameter prediction model, and at the same time can better adapt to complex prediction tasks.
[0116] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the shield tunnel parameter prediction is specifically performed by constructing the data preprocessing layer, the parameter prediction layer and the model optimization layer to predict the shield tunnel parameters, and by using the optimal shield tunnel parameter prediction model, the shield tunnel surface settlement prediction data is obtained, and the shield tunnel surface settlement risk level is divided in combination with the engineering standards.
[0117] Embodiment 7, see Figure 2 , this embodiment is based on the above embodiment, and a shield tunnel parameter prediction system based on deep learning provided by the present invention includes: a shield tunnel parameter prediction modeling module, a data preprocessing layer construction module, a parameter prediction layer construction module, a model optimization layer construction module and a shield tunnel parameter prediction module;
[0118] The shield tunnel parameter prediction modeling module is used for shield tunnel parameter prediction modeling. Through the shield tunnel parameter prediction modeling, an infrastructure for shield tunnel parameter prediction is obtained, and a data preprocessing layer, a parameter prediction layer and a model optimization layer are constructed based on the infrastructure for shield tunnel parameter prediction.
[0119] The data preprocessing layer construction module is used for data preprocessing layer construction, through which shield tunnel enhancement data is obtained, and the shield tunnel enhancement data is sent to the parameter prediction layer construction module and the model optimization layer construction module;
[0120] The parameter prediction layer construction module is used for parameter prediction layer construction, and obtains a shield tunnel parameter prediction model through parameter prediction layer construction, and sends the shield tunnel parameter prediction model to the model optimization layer construction module;
[0121] The model optimization layer construction module is used for model optimization layer construction, and obtains the optimal shield tunnel parameter prediction model through model optimization layer construction, and sends the optimal shield tunnel parameter prediction model to the shield tunnel parameter prediction module;
[0122] The shield tunnel parameter prediction module is used for shield tunnel parameter prediction. By constructing the data preprocessing layer, parameter prediction layer and model optimization layer, shield tunnel parameter prediction is performed, and by using the optimal shield tunnel parameter prediction model, shield tunnel surface settlement prediction data is obtained.
[0123] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0124] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.
[0125] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.
Claims
1. A shield tunnel parameter prediction method based on deep learning, characterized in that: The method comprises the following steps: Step S1: Shield tunnel parameter prediction modeling; Step S2: data preprocessing layer construction; Step S3: constructing a parameter prediction layer, specifically using a convolutional bidirectional long-term and short-term model combined with self-attention to construct a parameter prediction layer and obtain a shield tunnel parameter prediction model; The convolutional bidirectional long-term and short-term model combined with self-attention includes an input layer, a convolutional network layer, a deep self-attention block, a bidirectional long-term and short-term network layer, and a fully connected output layer; The convolutional network layer is used to extract nonlinear local features and perform feature dimensionality reduction; The deep self-attention block is used to extract key features and suppress redundant features; The bidirectional long-term and short-term network layer is used to mine time dependencies; Step S4: constructing a model optimization layer, specifically using an improved flower pollination algorithm to construct a model optimization layer to obtain an optimal shield tunnel parameter prediction model; Step S5: Prediction of shield tunnel parameters.
2. According to a method for predicting shield tunnel parameters based on deep learning according to claim 1, it is characterized in that: In step S3, the parameter prediction layer is constructed to construct the prediction layer of the prediction model, including the following steps: Step S31: constructing an input layer, specifically receiving shield tunnel enhancement data as model input through the input layer; Step S32: constructing a convolutional network layer, specifically setting a one-dimensional convolutional layer and a pooling layer in the convolutional network layer, extracting nonlinear local features of the model input through the one-dimensional convolutional layer, and then performing feature dimensionality reduction on the nonlinear local features through the pooling layer to obtain convolutional pooling features; Step S33: construct a deep self-attention block, specifically by replacing all fully connected layers in the self-attention block with a deep separable convolutional layer, improving the self-attention block to obtain a deep self-attention block, and calculating the key features through the deep self-attention block; Step S34: constructing a bidirectional long-short term network layer, specifically constructing a bidirectional long-short term network layer including a forward long-short term network sublayer and a backward long-short term network sublayer, calculating key features through the bidirectional long-short term network layer to obtain bidirectional time series features; Step S35: construct a fully connected output layer, calculate the bidirectional time series features through the fully connected output layer, generate and output the prediction results; Step S36: parameter prediction model training, specifically, performing model training by constructing the input layer, constructing the convolutional network layer, constructing the deep self-attention block, constructing the bidirectional long-short term network layer and constructing the fully connected output layer to obtain a shield tunnel parameter prediction model.
3. According to a method for predicting shield tunnel parameters based on deep learning according to claim 2, it is characterized in that: In step S33, constructing a deep self-attention block comprises the following steps: Step S331: Feature mapping, specifically mapping the convolutional pooling features through three independent depth-separable convolutional layers to obtain a query vector, a key vector and a value vector. The calculation formula is: ； Where Q is the query vector, DSC q (·) is the depth-wise separable convolution query vector calculation function, Y is the convolution pooling feature, K is the key vector, and DSC k (·) is the depthwise separable convolution key vector calculation function, V is the value vector, DSC v (·) is the depth-wise separable convolution value vector calculation function; Step S332: Calculate the self-attention result of each attention head. The calculation formula is: ； Where i is the self-attention head index, self i (Y) is the attention result of the i-th self-attention head, δ(·) is the softmax function, T is the transposition operation, dim i is the hidden dimension of the i-th self-attention head; Step S333: Apply the multi-head self-attention mechanism to concatenate the self-attention results of each attention head, and perform a depth-separable convolution operation to obtain the multi-head self-attention. The calculation formula is: ； Where Mself(Y) is the multi-head self-attention, Ms conv (·) is a depth-wise separable convolution operation used to aggregate multi-head self-attention information, Concat[·] is a concatenation operation, α is an adjustment parameter used to control the ratio of attention to high-frequency information and low-frequency information, and hn is the number of self-attention heads; Step S334: Perform feature aggregation on the convolution pooling features through the deep self-attention block to obtain key features.
4. The method for predicting shield tunnel parameters based on deep learning according to claim 3 is characterized in that: In step S4, the model optimization layer is constructed to optimize model parameters and performance, specifically using an improved flower pollination algorithm to construct the model optimization layer to obtain an optimal shield tunnel parameter prediction model, including the following steps: Step S41: Initialize input parameters, specifically construct input parameter combinations, and the calculation formula is: ； Where IFO is the input parameter combination, f is the objective function, u is the horizontal constraint of the search space, v is the vertical constraint of the search space, and es is the scale parameter of the elite replacement mechanism; Step S42: Initializing the population, specifically initializing the flower population position, the flower population position is used to represent the parameter combination of the shield tunnel parameter prediction model output by the model optimization layer, and calculating the fitness values of all the initialized flower population positions to obtain the best initial fitness, and sorting all the initialized flower population positions in ascending order according to the fitness value; Step S43: subgroup division, specifically adopting a dynamic adjustment method to update the fitness, and obtaining an updated optimal fitness through the fitness update, and performing subgroup division according to the updated optimal fitness to obtain population division data; Step S44: search selection optimization, specifically constructing a random threshold parameter and performing search selection optimization of the flower population position to obtain updated optimal flower population position data, and the calculation formula is: ； In the formula, is to update the optimal flower population location data, It is the flower index, is the current flower population position, is the learning rate parameter, is the optimal flower of the current iteration, is the random perturbation parameter, is the first random flower, is the second random flower, rand is the random threshold discrimination parameter, which is used to perform the search selection optimization search method. When rand>0.8, a global search is performed, otherwise a local search is performed. is the local search adjustment factor, P r is the random micro-disturbance parameter; Step S45: Optimize the elite replacement mechanism to ensure that the high-quality flower population positions can be retained in each iteration. Specifically, the total number of sub-group flower positions in each population partition data is calculated, and the sub-group flower positions in each population partition data are randomly arranged. The flower positions with the highest fitness es are selected from the arrangement as elite replacement flower positions, and the flower positions with the lowest fitness value ranking of 5% are replaced with the elite replacement flower positions to optimize the flower pollination search algorithm. Step S46: constructing an iteration stop condition, specifically setting the maximum number of iterations for iteration stop to 1000, and taking the absolute difference between the current best fitness and the previous best fitness to be less than 0.01 as a fitting condition, and stopping the iterative calculation when the maximum number of iterations is reached or the fitting condition is met; Step S47: iterative optimization, specifically, based on the iteration stop condition, the model optimization layer is constructed through the operations of step S41 to step S45 to obtain the optimal shield tunnel parameter prediction model.
5. The method for predicting shield tunnel parameters based on deep learning according to claim 4 is characterized in that: In step S5, the shield tunnel parameter prediction is specifically performed by constructing the data preprocessing layer, the parameter prediction layer and the model optimization layer to predict the shield tunnel parameters, and by using the optimal shield tunnel parameter prediction model, the shield tunnel surface settlement prediction data is obtained, and the shield tunnel surface settlement risk level is divided in combination with the engineering standards.
6. The method for predicting shield tunnel parameters based on deep learning according to claim 5 is characterized in that: In step S1, the shield tunnel parameter prediction modeling is used to design and construct a basic framework for shield tunnel parameter prediction, specifically to build a prediction model for shield tunnel parameters. The prediction model specifically includes a data preprocessing layer, a parameter prediction layer and a model optimization layer.
7. The method for predicting shield tunnel parameters based on deep learning according to claim 6, characterized in that: In step S2, the data preprocessing layer is constructed to collect shield tunnel related parameters and perform data preprocessing, including the following steps: Step S21: data collection, specifically collecting shield operation parameters, tunnel geological parameters and tunnel space parameters from the shield tunnel construction process to obtain shield tunnel parameters; Step S22: Data preprocessing, specifically preprocessing the shield tunnel parameters through data cleaning, noise filtering and normalization to obtain shield tunnel enhancement data.
8. A shield tunnel parameter prediction system based on deep learning, used to implement a shield tunnel parameter prediction method based on deep learning as described in any one of claims 1 to 7, characterized in that: It includes a shield tunnel parameter prediction modeling module, a data preprocessing layer building module, a parameter prediction layer building module, a model optimization layer building module and a shield tunnel parameter prediction module.
9. A shield tunnel parameter prediction system based on deep learning according to claim 8, characterized in that: The shield tunnel parameter prediction modeling module is used for shield tunnel parameter prediction modeling. Through the shield tunnel parameter prediction modeling, an infrastructure for shield tunnel parameter prediction is obtained, and a data preprocessing layer, a parameter prediction layer and a model optimization layer are constructed based on the infrastructure for shield tunnel parameter prediction. The data preprocessing layer construction module is used for data preprocessing layer construction, through which shield tunnel enhancement data is obtained, and the shield tunnel enhancement data is sent to the parameter prediction layer construction module and the model optimization layer construction module; The parameter prediction layer construction module is used for parameter prediction layer construction, and obtains a shield tunnel parameter prediction model through parameter prediction layer construction, and sends the shield tunnel parameter prediction model to the model optimization layer construction module; The model optimization layer construction module is used for model optimization layer construction, and obtains the optimal shield tunnel parameter prediction model through model optimization layer construction, and sends the optimal shield tunnel parameter prediction model to the shield tunnel parameter prediction module; The shield tunnel parameter prediction module is used for shield tunnel parameter prediction. By constructing the data preprocessing layer, parameter prediction layer and model optimization layer, shield tunnel parameter prediction is performed, and by using the optimal shield tunnel parameter prediction model, shield tunnel surface settlement prediction data is obtained.