Tunnel engineering earthquake vulnerability assessment method and system based on explainability analysis

Abstract

The application relates to the technical field of tunnel engineering, and provides a tunnel engineering earthquake vulnerability evaluation method and system based on explainability analysis, which comprises the following steps: acquiring tunnel characteristics, obtaining a tunnel earthquake damage index through a tunnel earthquake vulnerability intelligent evaluation model based on small sample learning; adopting a Shapley additive feature explanation method and a local explainable model agnostic explanation method to respectively calculate the importance scores of the tunnel characteristics on the tunnel earthquake vulnerability intelligent evaluation model prediction; and adopting a weighted geometric mean to fuse the two importance scores to obtain a fusion index importance; wherein the weights of the weighted geometric mean are calculated according to the prediction reconstruction errors of the Shapley additive feature explanation method and the local explainable model agnostic explanation method on a local perturbation data set. The contribution degrees of the influence parameters can be quantitatively and reliably revealed.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering technology, and in particular relates to a method and system for assessing the seismic vulnerability of tunnel engineering based on interpretability analysis. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Tunnels, as core infrastructure in transportation, energy, and water conservancy, are crucial projects connecting geographical barriers, ensuring unimpeded passage, and improving resource transportation efficiency. Their seismic performance directly affects the operational safety of infrastructure and is vital to regional economic development and social stability. Without in-depth research on the seismic vulnerability of tunnel engineering, tunnels will be exposed to enormous risks in the face of this powerful natural disaster. During an earthquake, unprotected tunnels may suffer severe damage: the lining may crack, peel, or even collapse; the tunnel structure may deform or shift; and landslides or collapses may occur at the tunnel entrance. Once a tunnel is damaged in an earthquake, road traffic will be blocked, making it difficult for rescue forces and supplies to reach the disaster area quickly, and the lives of trapped people will be seriously threatened. Furthermore, repairing damaged tunnels not only requires a significant investment of manpower, resources, and time, but may also have long-term negative impacts on regional economic development and social stability, resulting in incalculable losses.

[0004] Current methods for predicting tunnel seismic vulnerability can be mainly divided into traditional physical theory models and data-driven methods.

[0005] Traditional physical theory models typically rely on simplified boundary condition assumptions and idealized constitutive relations, making it difficult to accurately reflect the complex interactions between the surrounding rock and the structure under actual seismic loading. This results in significant limitations in guiding specific engineering design and reinforcement.

[0006] In data-driven approaches, the prediction accuracy of large-sample data learning methods heavily relies on massive amounts of high-quality training data. However, data from real earthquake damage is extremely scarce and cannot provide enough effective samples to support large-sample data learning methods in accurately capturing the complex linear relationship between the tunnel seismic damage index and tunnel structural parameters. Summary of the Invention

[0007] To address the technical problems mentioned above, this invention provides a method and system for assessing the seismic vulnerability of tunnel engineering based on interpretability analysis. It proposes interpretability analysis based on game theory and local proxy models. By proposing a dynamic weighted fusion index, it can quantitatively and reliably reveal the contribution of each influencing parameter, effectively promoting the seismic response analysis of tunnel engineering to move away from the traditional "black box" prediction and achieve a leap towards "transparent" analysis.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] The first aspect of the present invention provides a method for assessing the seismic vulnerability of tunnel engineering based on interpretability analysis, comprising:

[0010] Tunnel characteristics are obtained, and the tunnel seismic damage index is obtained through a tunnel seismic vulnerability intelligent assessment model based on small sample learning.

[0011] The Shapley additive feature interpretation method and the locally interpretable model agnostic interpretation method are used to calculate the importance score of each tunnel feature to the prediction of the intelligent assessment model of tunnel seismic vulnerability. The two importance scores are then fused using a weighted geometric mean to obtain the importance of the fused index. The weight of the weighted geometric mean is calculated based on the prediction reconstruction error of the Shapley additive feature interpretation method and the locally interpretable model agnostic interpretation method on the local disturbance dataset.

[0012] Furthermore, the tunnel features include tunnel structural parameters, seismic motion parameters, and geological conditions.

[0013] Furthermore, the forward propagation process of the intelligent assessment model for tunnel seismic vulnerability includes:

[0014] In the l-th layer encoder, the output of the previous layer is linearly projected into a query, key, and value matrix; based on the query, key, and value matrix, the output of each attention head is calculated, and all attention head outputs are concatenated and linearly projected before residual connection is performed. The connection result is then layer normalized to obtain the intermediate output vector.

[0015] Using a feedforward network, the intermediate output vector is transformed nonlinearly, and residual connections and layer normalization are used to obtain the final output vector of the l-th layer.

[0016] Take the position vector corresponding to the query sample from the final output vector of the last layer, and obtain the final predicted value through a linear regression layer.

[0017] Furthermore, the importance of the fusion indicator is as follows: Where ω is an adjustable weight, and the normalized importance score is obtained by the Shapley additivity feature interpretation method. Normalized importance score obtained from the locally interpretable model but unknowable interpretation method N is the sample size. Features of the i-th sample in the locally interpretable model's unknowable interpretation method The local importance score, where P is the number of tunnel features. Features in the Shapley additivity feature interpretation method The global importance score.

[0018] Furthermore, the pre-trained prior parameters of the tunnel seismic vulnerability intelligent assessment model are used to fit the prior data of the transfer table to the network.

[0019] A second aspect of the present invention provides a seismic vulnerability assessment system for tunnel engineering based on interpretability analysis, comprising:

[0020] The assessment module is configured to: acquire tunnel characteristics and obtain the tunnel seismic damage index through a tunnel seismic vulnerability intelligent assessment model based on few-sample learning;

[0021] The interpretability analysis module is configured to: use the Shapley additive feature interpretation method and the locally interpretable model-agnostic interpretation method to calculate the importance score of each tunnel feature to the prediction of the intelligent assessment model of tunnel seismic vulnerability, and use a weighted geometric mean to fuse the two importance scores to obtain the importance of the fused index; wherein, the weight of the weighted geometric mean is calculated based on the prediction reconstruction error of the Shapley additive feature interpretation method and the locally interpretable model-agnostic interpretation method on the local disturbance dataset.

[0022] Furthermore, the tunnel features include tunnel structural parameters, seismic motion parameters, and geological conditions.

[0023] Furthermore, the forward propagation process of the intelligent assessment model for tunnel seismic vulnerability includes:

[0024] In the l-th layer encoder, the output of the previous layer is linearly projected into a query, key, and value matrix; based on the query, key, and value matrix, the output of each attention head is calculated, and all attention head outputs are concatenated and linearly projected before residual connection is performed. The connection result is then layer normalized to obtain the intermediate output vector.

[0025] Using a feedforward network, the intermediate output vector is transformed nonlinearly, and residual connections and layer normalization are used to obtain the final output vector of the l-th layer.

[0026] Take the position vector corresponding to the query sample from the final output vector of the last layer, and obtain the final predicted value through a linear regression layer.

[0027] Furthermore, the importance of the fusion indicator is as follows: Where ω is an adjustable weight, and the normalized importance score is obtained by the Shapley additivity feature interpretation method. Normalized importance score obtained from the locally interpretable model but unknowable interpretation method N is the sample size. Features of the i-th sample in the locally interpretable model's unknowable interpretation method The local importance score, where P is the number of tunnel features. Features in the Shapley additivity feature interpretation method The global importance score.

[0028] Furthermore, the pre-trained prior parameters of the tunnel seismic vulnerability intelligent assessment model are used to fit the prior data of the transfer table to the network.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] This invention proposes an interpretability analysis based on game theory and local proxy models. By proposing a dynamic weighted fusion index, it can quantitatively and reliably reveal the contribution of each influencing parameter, effectively promoting the transformation of seismic response analysis in tunnel engineering from traditional "black box" prediction to "transparent" analysis.

[0031] This invention proposes a few-shot learning model based on the Transformer architecture, which utilizes transfer learning through TabPFN (Tabular Priority Learning). The pre-trained prior parameters of the data-fitted network (table prior data fitting network) enable strong generalization ability to complex nonlinear relationships under limited data conditions. It can effectively address the problem of difficult data collection for tunnel seismic response and meet the needs of accuracy and efficiency in practical application scenarios. In order to further improve the model performance, a hyperparameter optimization module based on the tree structure Parzen is integrated to realize automatic optimization of key model parameters. Attached Figure Description

[0032] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0033] Figure 1 This is a flowchart of the forward propagation of the intelligent assessment model for tunnel seismic vulnerability in Embodiment 1 of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0035] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] Example 1

[0037] This embodiment provides a method for assessing the seismic vulnerability of tunnel engineering based on interpretability analysis.

[0038] This embodiment provides a method for assessing the seismic vulnerability of tunnel engineering based on interpretability analysis, with advantages mainly reflected in three aspects:

[0039] First, a few-shot learning model based on the Transformer architecture is proposed. This model achieves strong generalization ability for complex nonlinear relationships under limited data conditions by transferring the pre-trained prior parameters of TabPFN (Tabular Prior-data Fitted Network). It can effectively address the problem of difficult collection of tunnel seismic response data and meet the needs of accuracy and efficiency in practical application scenarios.

[0040] Second, to further improve model performance, a hyperparameter optimization module based on the tree structure Parzen is integrated to achieve automatic optimization of key model parameters;

[0041] Third, an interpretability analysis module based on game theory and local agent model was constructed. By proposing a dynamic weighted fusion index, the contribution of each influencing parameter can be revealed quantitatively and reliably, effectively promoting the transformation of tunnel engineering seismic response analysis from traditional "black box" prediction to "transparent" analysis.

[0042] This embodiment provides a method for assessing the seismic vulnerability of tunnel engineering based on interpretability analysis, which includes the following steps:

[0043] Step 1: Build the dataset.

[0044] Step 101: Data collection.

[0045] The seismic vulnerability data of tunnel engineering is used to construct a raw tabular dataset for modeling: tunnel structural parameters (tunnel diameter, overburden depth, lining thickness), ground motion parameters (peak ground acceleration, source-to-site distance) and geological conditions (geological strength index) are used as input features, and the tunnel seismic damage index is used as the prediction output.

[0046] Step 102: Data preprocessing.

[0047] Correlation analysis: ;in, y and r are the means of variables x and y, respectively, and n is the sample size; when r≈0, it means that there is no significant linear relationship between the two variables.

[0048] Normalization: After constructing the tabular data, normalize the original dataset to map it to the 0-1 interval, eliminate distribution skewness, and enhance model robustness. The specific steps can be performed using the following formula: In the formula, It is the raw data. It is normalized data. The maximum value in the original data. This is the minimum value in the original data.

[0049] Step 103: Dataset partitioning.

[0050] The preprocessed dataset is divided into training and test sets in an 80% and 20% ratio. The training set is used to learn the model's training parameters, and the test set is used to evaluate the model's performance during training.

[0051] Step 2: Construction of an intelligent assessment model for tunnel seismic vulnerability based on small sample learning.

[0052] Step 201: Model architecture setup.

[0053] This embodiment constructs an intelligent assessment model for tunnel seismic vulnerability based on few-sample learning. The specific construction process is as follows:

[0054] (1) Embedding of tabular data.

[0055] (101) After normalizing the numerical features (peak ground acceleration, distance from source to station, etc.) in the original tabular data, the transformed values ​​are mapped to high-dimensional vectors through a learnable linear projection layer to obtain continuous feature embeddings. The continuous feature embeddings are calculated using the following formula:

[0056] ;

[0057] in, These are the numerical features after normalization. and For learnable weight matrices and bias vectors, This is used to embed the obtained continuous features.

[0058] (102) Assign a learnable position encoding vector to each column feature in the tabular data to facilitate the tunnel seismic vulnerability intelligent assessment model to distinguish the features of different columns; add a learnable sequence position encoding vector to each row sample in the tabular data to facilitate the tunnel seismic vulnerability intelligent assessment model to capture the sequential relationship between samples.

[0059] (103) Add the feature embedding vector of each sample to the positional encoding to obtain the input for each sample. Specifically, follow the formula below:

[0060] ;

[0061] in, This is the learnable position encoding vector corresponding to the j-th feature column. The learnable position encoding vector corresponding to the i-th sample in the input sequence It is the input vector that is directly fed into the subsequent Transformer encoding layer.

[0062] (2) Establish a pre-training architecture based on Transformer.

[0063] The core feature extractor of the intelligent assessment model for tunnel seismic vulnerability is a Transformer-based encoder. The following targeted designs were made to the structure of the intelligent assessment model for tunnel seismic vulnerability:

[0064] (A) To control computational cost and overfitting risk, the encoder depth is preset to 3 layers, each layer is equipped with a 6-head self-attention mechanism (Multi-Head Self-Attention); each Transformer encoder layer contains two core sub-layers: a multi-head self-attention sub-layer and a feedforward neural network sub-layer;

[0065] (B) First, the input vector is processed by a multi-head self-attention sublayer to capture the global dependencies within the sequence and obtain an intermediate output; this output is then used as the input to a feedforward neural network sublayer, which consists of a linear transformation-GeLU activation function-linear transformation to enhance the nonlinear transformation capability, and finally generates the output vector of the encoder of this layer.

[0066] (C) In each Transformer encoder layer, residual connections and layer normalization are applied successively after the multi-head self-attention sub-layer and the feedforward neural network sub-layer to ensure the stability of the training process and promote model convergence.

[0067] (D) To determine the optimal configuration of key hyperparameters (number of encoder layers, number of attention heads, dimension of hidden layers in the feedforward network, etc.) in the above architecture, a TPE (Tree-structured Parzen Estimator) hyperparameter optimization module is introduced. The hyperparameter optimization module uses the mean squared error (MSE) on the validation set as the loss function, constructs a probabilistic model of hyperparameters through sequential modeling, and adaptively samples new parameter combinations for iterative validation. This allows for efficient locking of high-performance configurations within a limited number of evaluations, significantly enhancing the model's generalization ability and convergence stability.

[0068] In this embodiment, as Figure 1 As shown, the specific process of forward propagation of the intelligent assessment model for tunnel seismic vulnerability is as follows:

[0069] (A) Multi-head self-attention quantum layer:

[0070] In the l-th layer Transformer encoder, the output of the previous layer is... Linear projection is a query (Q), key (K), and value (V) matrix: , , ;in, , and This is the learnable projection weight matrix for this layer. , and Let L be the matrix of query, key, and value at level l;

[0071] The calculation method for each attention head i is as follows: ;in, It is the dimension of the key vector; the query matrix of attention head i in the l-th layer is The key matrix of attention head i in the l-th layer is The value matrix of attention head i in the l-th layer is ; , and Let be the learnable projection weight matrix of attention head i in layer l;

[0072] The outputs of the six independent attention heads are concatenated and linearly projected: Among them, MSA stands for multi-head self-attention. , To output the projection matrix;

[0073] In each Transformer encoder layer, residual connections are performed, and then the connection results are normalized to obtain the intermediate output vector of the current layer's MSA. : Where LayerNorm is the layer normalization.

[0074] (B) Feedforward neural network sublayer:

[0075] Using a feedforward network (FFN), a nonlinear transformation is performed on the features at each location: ;in, and The learnable parameters are the first linear transformation parameters of the l-th layer feedforward network. and represents the learnable parameters of the second linear transformation; GeLU is a nonlinear activation function based on a Gaussian distribution.

[0076] This sub-layer also uses residual connections and layer normalization to obtain the final output vector of the l-th layer. : .

[0077] (C) Output layer:

[0078] Take the output of the last encoder layer The vector corresponding to the position of the query sample is used to obtain the final predicted value through a linear regression layer: ;in, is the location vector of the query sample output by the Lth layer Transformer, w is the weight of the offline regression, and b is the bias; the query sample is the test sample to be predicted. This sample only provides input features (tunnel diameter, overburden depth, lining thickness, peak ground acceleration, source-to-site distance, geological intensity index) and does not contain the corresponding damage index label.

[0079] Finally, the intelligent assessment model for tunnel seismic vulnerability uses mean squared error (MSE) as the loss function for optimization: Where N is the total number of samples in the current batch. Let be the predicted value for the i-th sample. Let be the true value of the i-th sample.

[0080] (3) Seismic vulnerability assessment.

[0081] The pre-trained intelligent assessment model for tunnel seismic vulnerability possesses general context learning capabilities. The model embeds all samples from the training set (including tunnel structural parameters, ground motion parameters, and geological conditions as features, along with corresponding damage index labels) into the input sequence. Subsequently, the test samples to be predicted are placed at the end of this sequence. Through this construction, the intelligent assessment model can directly infer the damage index of all test samples based on the complete "context" provided by the training set in a single forward propagation, achieving efficient few-sample learning without gradient updates.

[0082] Step 202: Performance evaluation of the intelligent assessment model for tunnel seismic vulnerability.

[0083] After the intelligent assessment model for tunnel seismic vulnerability is constructed, the mean square error (RMSE) and coefficient of determination (R²) are used. 2 As a model evaluation metric, the specific formula is as follows:

[0084] ;

[0085] ;

[0086] In the formula, n is the number of samples; , These are the true value and the predicted value of the i-th sample, respectively; Let be the mean of the target observations for n samples.

[0087] For the performance metric RMSE, the value range is from 0 to +∞. A larger value indicates a larger overall error in the intelligent assessment model for tunnel seismic vulnerability. 2 The value range is from 0 to 1, and the closer the value is to 1, the better the overall performance of the model.

[0088] Step 3: Interpretability analysis of the tunnel seismic vulnerability model.

[0089] To enable the established tunnel seismic vulnerability assessment model to quantify feature importance and overcome the limitations of a single interpretation method, this embodiment employs an importance fusion index that integrates two advanced interpretable machine learning methods: SHAP (Shapley Additive explanation) and LIME (Local Interpretable Model-agnostic Explanations). This fusion aims to quantify the importance of each feature to the seismic vulnerability model's predictions.

[0090] Step 301, SHAP method.

[0091] The SHAP method uses game theory principles to assign contribution values ​​to each feature and quantifies how each feature affects the model's predictions. The core advantage of this method lies in its solid mathematical theoretical foundation, which satisfies consistency and local accuracy, and can guarantee the fairness and stability of feature contribution allocation.

[0092] The SHAP value is calculated using the following formula:

[0093] ;

[0094] in, Indicates the first Shapley values ​​for each factor. Except All possible subsets of features other than those mentioned above. It is the set of all features, that is, different combinations of input features. It is a set of factors affecting the tunnel earthquake damage index. The model prediction values ​​obtained from the features. Indicating in the feature set Add the first After considering these features, the model predicts the tunnel earthquake damage index.

[0095] By testing on all samples By taking the average, the features can be obtained. Its global importance.

[0096] Step 302, LIME method.

[0097] The LIME method focuses on local fitting. It generates a new dataset by randomly perturbing the neighborhood of the sample to be interpreted and trains a simple interpretable model to approximate the behavior of a complex "black box" model in that local area. The advantage of LIME is its intuitiveness and its ability to provide clear local decision boundaries.

[0098] LIME obtains its interpretation by optimizing the following objective function:

[0099] ;

[0100] in, Let G represent the candidate set of local explanation models, which is the final selected best local explanation model. This is a sample to be explained. It is an established tunnel seismic vulnerability assessment model. It is an explanatory model. Defined samples Weight function within the neighborhood, It is a measure How to approximate loss function, Explanation model Complexity penalty.

[0101] Explanation Model The absolute value of the coefficient is considered as the local importance of the feature.

[0102] Step 303: Importance Integration Indicators.

[0103] Both SHAP and LIME methods have their limitations: SHAP's computation is based on the assumption of all feature combinations, which may produce unrealistic samples when feature correlation is high; LIME, on the other hand, relies on perturbation strategies and hyperparameters defined by the neighborhood. To overcome the bias of a single interpretable machine learning method and obtain a more robust importance assessment, this study proposes an importance fusion index.

[0104] The importance fusion metric is a global importance ranking derived by integrating SHAP and LIME, calculated using the following formula:

[0105] (1) Importance score normalization: The feature importance obtained by the two methods is normalized to make them have the same dimension.

[0106] For SHAP, features global importance Normalization: .

[0107] For LIME, the local importance of each feature is first averaged across all samples, followed by normalization. Where N is the sample size. Features in the LIME interpretation of the i-th sample The coefficient.

[0108] For the i-th sample, the explanation model The specific form is as follows: In the formula, Here, P represents the intercept term, and P is the number of all input features. In this embodiment, P=6. To correspond to the features The regression coefficient represents the degree of influence of this feature on the prediction result.

[0109] (2) Calculation of the importance of the fusion index.

[0110] Weighted geometric mean is more robust to extreme values ​​and better reflects the consistent contribution of features to both methods. It is used to fuse the importance of the two normalized methods.

[0111] ;

[0112] Wherein, ω (0≤ω≤1) is an adjustable weight used to balance the confidence of the two methods; this embodiment proposes a dynamic weight determination method based on the principle of local consistency, which calculates the prediction reconstruction error of SHAP and LIME on a locally perturbed dataset ( and ), to quantify its reliability:

[0113] Let the query sample to be interpreted be... M perturbation samples are generated in its neighborhood. The hyperparameter M is set to 500 for each perturbation sample. This allows us to obtain the actual output of the original intelligent assessment model for tunnel seismic vulnerability. ;

[0114] The local approximation model of SHAP satisfies: In the formula, The baseline value corresponds to the expected output when there are no features. Indicates the first Shapley values ​​for each factor; A local approximation model constructed for the SHAP method;

[0115] In the LIME method, the optimal local interpretation model is ;

[0116] The formula for calculating the prediction reconstruction error is:

[0117] ;

[0118] ;

[0119] As one implementation method, the weights are determined based on the prediction reconstruction error: This formula ensures that, for a given dataset and model, the method with higher local fitting accuracy will receive greater weight in the fusion.

[0120] As another implementation method, the above weight calculation method uses a simple inverse error ratio, but when the error of a certain method is extremely small / extremely large, the weight will be too extreme (e.g., →0, →1, which is entirely dependent on LIME, adds a non-linear penalty term to mitigate the impact of extreme values: c>1 is the smoothing constant. Strictly limit the impact of extreme errors.

[0121] If and only if the local reconstruction errors of the two methods are similar ( When the weights degenerate to ω=0.5, the fusion index degenerates into a simple geometric mean. .

[0122] Ultimately, according to The values ​​are used to sort all features, resulting in a more accurate comprehensive feature importance ranking based on multi-method validation.

[0123] This fusion model not only inherits the global consistency of SHAP and the local fidelity of LIME, but also reduces the interpretive bias caused by the limitations of a single method through mutual verification, greatly enhancing the transparency and credibility of the model's decision-making process.

[0124] The calculated importance fusion index values ​​are integrated into the tunnel seismic vulnerability assessment model to ensure that the model's predictions can be traced back to the contribution of each factor affecting tunnel seismic vulnerability, thus providing interpretability of the tunnel seismic vulnerability assessment model's predictions.

[0125] Example 2

[0126] This embodiment provides a seismic vulnerability assessment system for tunnel engineering based on interpretability analysis, including:

[0127] The assessment module is configured to: acquire tunnel characteristics and obtain the tunnel seismic damage index through a tunnel seismic vulnerability intelligent assessment model based on few-sample learning;

[0128] The interpretability analysis module is configured to: use the Shapley additive feature interpretation method and the locally interpretable model-agnostic interpretation method to calculate the importance score of each tunnel feature to the prediction of the intelligent assessment model of tunnel seismic vulnerability, and use a weighted geometric mean to fuse the two importance scores to obtain the importance of the fused index; wherein, the weight of the weighted geometric mean is calculated based on the prediction reconstruction error of the Shapley additive feature interpretation method and the locally interpretable model-agnostic interpretation method on the local disturbance dataset.

[0129] Furthermore, the tunnel features include tunnel structural parameters, seismic motion parameters, and geological conditions.

[0130] Furthermore, the forward propagation process of the intelligent assessment model for tunnel seismic vulnerability includes:

[0131] In the l-th layer encoder, the output of the previous layer is linearly projected into a query, key, and value matrix; based on the query, key, and value matrix, the output of each attention head is calculated, and all attention head outputs are concatenated and linearly projected before residual connection is performed. The connection result is then layer normalized to obtain the intermediate output vector.

[0132] Using a feedforward network, the intermediate output vector is transformed nonlinearly, and residual connections and layer normalization are used to obtain the final output vector of the l-th layer.

[0133] Take the position vector corresponding to the query sample from the final output vector of the last layer, and obtain the final predicted value through a linear regression layer.

[0134] Furthermore, the importance of the fusion indicator is as follows: Where ω is an adjustable weight, and the normalized importance score is obtained by the Shapley additivity feature interpretation method. Normalized importance score obtained from the locally interpretable model but unknowable interpretation method N is the sample size. Features of the i-th sample in the locally interpretable model's unknowable interpretation method The local importance score, where P is the number of tunnel features. Features in the Shapley additivity feature interpretation method The global importance score.

[0135] Furthermore, the pre-trained prior parameters of the tunnel seismic vulnerability intelligent assessment model are used to fit the prior data of the transfer table to the network.

[0136] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

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

Claims

1. A method for tunnel engineering seismic vulnerability assessment based on explainability analysis, characterized in that, The method comprises the following steps: obtaining tunnel characteristics, and obtaining a tunnel seismic damage index through a tunnel seismic vulnerability intelligent evaluation model based on small sample learning; Shapley additive feature interpretation method and local interpretable model agnostic interpretation method are used to calculate the importance score of each tunnel characteristic on the prediction of the tunnel seismic vulnerability intelligent evaluation model, and the weighted geometric mean is used to fuse the two importance scores to obtain the fused index importance; wherein the weight of the weighted geometric mean is calculated according to the prediction reconstruction error of the Shapley additive feature interpretation method and the local interpretable model agnostic interpretation method on the local perturbation data set; The tunnel seismic vulnerability intelligent evaluation model migrates the pre-training prior parameters of the table prior data fitting network; The fusion index importance is ; wherein ω is an adjustable weight, the normalized importance score obtained by the Shapley additive feature interpretation method , the normalized importance score obtained by the local interpretable model agnostic interpretation method , N is the number of samples, is the local importance score of the feature of the i-th sample in the local interpretable model agnostic interpretation method, P is the number of tunnel features, is the global importance score of the feature in the Shapley additive feature interpretation method.

2. The method for tunnel engineering seismic vulnerability assessment based on explainability analysis according to claim 1, wherein, The tunnel characteristics include tunnel structure parameters, ground motion parameters and geological conditions.

3. The method for tunnel engineering seismic vulnerability assessment based on explainability analysis according to claim 1, characterized in that, The forward propagation process of the tunnel seismic vulnerability intelligent evaluation model comprises the following steps: In the lth layer encoder, the output of the previous layer is linearly projected into a query, key and value matrix; based on the query, key and value matrix, the output of each attention head is calculated, and after splicing and linear projection of all attention head outputs, residual connection is performed, the connection result is layer normalized to obtain an intermediate output vector; The intermediate output vector is nonlinearly transformed by using a feedforward network, and then residual connection and layer normalization are used to obtain the final output vector of the lth layer; The position vector corresponding to the query sample in the final output vector of the last layer is taken, and the final prediction value is obtained through a linear regression layer.

4. A system for tunnel engineering seismic vulnerability assessment based on explainability analysis, characterized by, The method comprises the following steps: The evaluation module is configured to: obtain tunnel characteristics, and obtain a tunnel seismic damage index through a tunnel seismic vulnerability intelligent evaluation model based on small sample learning; The explainability analysis module is configured to: use Shapley additive feature interpretation method and local interpretable model agnostic interpretation method to calculate the importance score of each tunnel characteristic on the prediction of the tunnel seismic vulnerability intelligent evaluation model, and use weighted geometric mean to fuse the two importance scores to obtain the fused index importance; wherein the weight of the weighted geometric mean is calculated according to the prediction reconstruction error of the Shapley additive feature interpretation method and the local interpretable model agnostic interpretation method on the local perturbation data set; The tunnel seismic vulnerability intelligent evaluation model migrates the pre-training prior parameters of the table prior data fitting network; The fusion index importance is ; wherein ω is an adjustable weight, the normalized importance score obtained by the Shapley additive explanation method , the normalized importance score obtained by the local interpretable model agnostic explanation method , N is the number of samples, is the local importance score of the feature of the i-th sample in the local interpretable model agnostic explanation method, P is the number of tunnel features, is the global importance score of the feature in the Shapley additive explanation method.

5. The tunneling engineering earthquake vulnerability assessment system based on explainability analysis of claim 4, wherein, The tunnel characteristics include tunnel structure parameters, ground motion parameters and geological conditions.

6. The tunneling engineering earthquake vulnerability assessment system based on explainability analysis of claim 4, wherein, The forward propagation process of the tunnel seismic vulnerability intelligent evaluation model comprises the following steps: In the lth layer encoder, the output of the previous layer is linearly projected into a query, key and value matrix; based on the query, key and value matrix, the output of each attention head is calculated, and after splicing and linear projection of all attention head outputs, residual connection is performed, the connection result is layer normalized to obtain an intermediate output vector; The intermediate output vector is nonlinearly transformed by using a feedforward network, and then residual connection and layer normalization are used to obtain the final output vector of the lth layer; The position vector corresponding to the query sample in the final output vector of the last layer is taken, and the final prediction value is obtained through a linear regression layer.

Images