A shield tunnel surrounding rock classification method based on multi-task learning and adversarial training
By combining multi-task learning and adversarial training in a transfer learning model, the problem of poor model generalization ability in shield tunnel construction was solved. It achieved the extraction and rapid adaptation of common features across engineering data, and improved the accuracy and generalization ability of shield tunnel surrounding rock classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA RAILWAY LIUYUAN GRP CO LTD
- Filing Date
- 2025-09-09
- Publication Date
- 2026-05-08
AI Technical Summary
In existing shield tunnel construction, machine learning models have poor generalization ability, which makes it difficult to transfer knowledge across projects and shield machines. Furthermore, they cannot be effectively trained when data is sparse in the early stages of a project, affecting construction guidance.
A transfer learning model combining a multi-task learning framework and adversarial training is adopted. By decomposing and structuring different tunnel engineering data, common features are extracted and a transfer learning model is constructed to improve the model's generalization ability.
When target engineering data is scarce, the model can quickly adapt and achieve satisfactory surrounding rock classification performance, improving the model's generalization performance and classification accuracy.
Smart Images
Figure CN121167476B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of shield tunnel construction technology, and in particular to a method for classifying the surrounding rock of shield tunnels based on multi-task learning and adversarial training. Background Technology
[0002] Tunnel boring machines (TBMs) are widely used in tunnel construction. The operating parameters of these TBMs need to be matched with geological conditions, requiring TBM operators to adjust these parameters in real time. Currently, many machine learning models, such as Random Forest (RF), Convolutional Neural Networks (CNN), and Deep Neural Networks (DNN), are used for real-time rock classification. However, these models suffer from the following problems: 1. Poor generalization ability: The training data for classification models comes from a specific project. Therefore, the optimal parameter combination and learned features of the model are strongly correlated with the distribution characteristics of the training data. When the model is applied to another project, differences in TBM configuration and geological conditions can lead to deviations in the data distribution characteristics of the target dataset from the training dataset, resulting in poor model performance. 2. High retraining costs: Due to the poor generalization ability of traditional models, different projects often require training separate models. Training a single model often consumes significant computational resources and requires substantial data support. When data is sparse in the early stages of a new project, the model cannot be effectively trained, preventing it from guiding construction in the initial phase. While existing technologies attempt to improve accuracy through physical models or hybrid networks, they have not solved the problem of knowledge transfer across projects and tunnel boring machines. Therefore, there is an urgent need for a tunnel surrounding rock classification method that can efficiently utilize historical data and adapt to different engineering scenarios. Summary of the Invention
[0003] The purpose of this invention is to provide a method for classifying the surrounding rock of shield tunnels based on multi-task learning and adversarial training. By combining a multi-task learning framework with adversarial training, a transfer learning model architecture is used to extract common features among different engineering data, thereby improving the model's generalization ability and solving the problem of surrounding rock classification in scenarios where target engineering data is scarce.
[0004] To achieve the above objectives, the present invention provides the following solution:
[0005] A method for classifying surrounding rock of shield tunnels based on multi-task learning and adversarial training includes:
[0006] Collect data from different tunnel projects and perform signal decomposition and structured processing to build a database. The data includes the surrounding rock type and corresponding corrected BQ value at different tunnel segments, as well as the time series data of the tunnel boring machine's operating parameters during the construction of each ring segment, recorded by the tunnel boring machine's PLC system.
[0007] A transfer learning model is constructed using a multi-task learning framework and an adversarial training mechanism. The transfer learning model is then trained and validated based on the database to obtain the final transfer learning model.
[0008] Obtain the target tunnel engineering data, input it into the final transfer learning model, and output the predicted surrounding rock category and the corresponding corrected BQ value.
[0009] Optionally, performing the signal decomposition includes:
[0010] Outlier removal and normalization were performed on the time series data of the tunnel boring machine operating parameters during the construction of each ring segment to obtain valid time series data.
[0011] The VMD-DFA method is used to decompose the effective time series data into several sub-signals and calculate the scaling exponent of each sub-signal.
[0012] Based on the range of the scaling index, signals are divided into trend signals, fluctuation signals, and noise signals, and the noise signals are removed.
[0013] Optionally, performing the structuring process to build the database includes:
[0014] All trend and fluctuation signals are measured using a unified scale, with each ring as a data sample, to construct a structured database for different projects. The data sample includes representative values of the tunnel boring machine's operating data during the construction of each ring segment, the surrounding rock type at the location of the ring, and the corresponding corrected BQ value. The representative values of the trend signals are represented by the average value, and the representative values of the fluctuation signals are represented by the coefficient of variation.
[0015] Optionally, the transfer learning model includes an encoding layer, an attention layer, a multi-classification layer, a regression layer, a gradient reversal layer, and a data reconstruction layer. The input data is first transformed into deep features by the encoding layer, and the deep features are transformed into weighted features by the attention layer. The weighted features then flow to three different tasks, wherein the regression layer performs the first task, the multi-classification layer performs the second task, and the gradient reversal layer and the data reconstruction layer perform the third task.
[0016] Optionally, the encoding layer employs a multi-layer fully connected neural network, using the ReLU function as the activation function.
[0017] Optionally, the attention layer consists of a dense layer, a softmax layer, and a weighted layer. The dense layer is used to calculate the attention score of the deep features, using the ReLU function as the activation function. The softmax layer is used to normalize the attention score. The weighted layer is used to perform weighted evaluation on the deep features based on the normalized attention score and output the weighted features.
[0018] Optionally, the first task is a regression task, in which the regression layer predicts the corrected BQ value based on the weighted features, and classifies the surrounding rock category according to the range of the corrected BQ value as the prediction result. The regression layer is a fully connected layer and the ReLU function is used as the activation function.
[0019] Optionally, the second task is a multi-classification task, in which the multi-classification layer predicts the probability of belonging to different surrounding rock categories based on the weighted features. The multi-classification layer is a fully connected layer and uses the ReLU function as the activation function.
[0020] Optionally, the third task is a data reconstruction task, wherein the weighted features first flow through the gradient inversion layer, then are input into the data reconstruction layer, and the reconstructed data of the input data is output. The data reconstruction layer is a fully connected layer and uses the ReLU function as the activation function.
[0021] Optionally, during the training process of the transfer learning model, the first and second tasks have a positive effect on the optimization of the encoding layer, while the third task has a negative effect on the optimization of the encoding layer, forming adversarial training.
[0022] The beneficial effects of this invention are as follows:
[0023] This invention presents a transfer learning model for classifying surrounding rock in shield tunnels, based on multi-task learning and adversarial training. By employing a multi-task learning framework and adversarial training mechanism, it extracts common features across engineering data, thereby improving the model's generalization performance. Furthermore, the model requires only a small amount of fine-tuning with target engineering data to quickly adapt to the data distribution characteristics of the target engineering dataset, achieving satisfactory classification performance even when faced with sparse, unknown engineering data. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a flowchart illustrating the construction process of the transfer learning model according to an embodiment of the present invention.
[0026] Figure 2 This is a flowchart illustrating the signal decomposition process according to an embodiment of the present invention.
[0027] Figure 3 A flowchart illustrating the database creation process in an embodiment of the present invention;
[0028] Figure 4This is a basic framework diagram of the transfer learning model in an embodiment of the present invention;
[0029] Figure 5 This is a schematic diagram of the multi-task adversarial network in an embodiment of the present invention;
[0030] Figure 6 This is a flowchart illustrating the verification process of the transfer learning model in an embodiment of the present invention. Detailed Implementation
[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0032] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0033] This embodiment provides a method for classifying the surrounding rock of shield tunnels based on multi-task learning and adversarial training, including:
[0034] Collect data from different tunnel projects and perform signal decomposition and structured processing to build a database. The data includes the surrounding rock type and corresponding corrected BQ value at different tunnel segments, as well as the time series data of the tunnel boring machine's operating parameters during the construction of each ring segment, recorded by the tunnel boring machine's PLC system.
[0035] A transfer learning model is constructed using a multi-task learning framework and an adversarial training mechanism. The transfer learning model is then trained and validated based on the database to obtain the final transfer learning model.
[0036] Obtain the target tunnel engineering data, input it into the final transfer learning model, and output the predicted surrounding rock category and the corresponding corrected BQ value.
[0037] Specifically, this embodiment utilizes a transfer learning model for classifying the surrounding rock of shield tunnels, built upon multi-task learning and adversarial training. The model extracts common features across engineering data using a multi-task learning framework and adversarial training mechanism, improving its generalization performance. Furthermore, the model requires only a small amount of fine-tuning with target engineering data to quickly adapt to the data distribution characteristics of the target engineering dataset, achieving satisfactory classification performance even when faced with sparse, unknown engineering data.
[0038] In this embodiment, the transfer learning model construction process is as follows: Figure 1 As shown, it includes:
[0039] Data collection: Collect data from different projects, including the surrounding rock type and corresponding corrected BQ value at different tunnel segments, as well as time series data of the main operating parameters of the tunnel boring machine (TBM) during the construction of each ring segment, such as thrust and torque, which are significantly correlated with the surrounding rock type. The BQ value is the Basic Quality Index (BQ index), and the corrected BQ value is the value after making corrections based on the BQ value, taking into account the effects of groundwater, initial stress of the surrounding rock, and weak structural surfaces.
[0040] Signal decomposition: The VMD-DFA method is used to decompose the time series data of each ring segment into trend signal, fluctuation signal and noise signal, and retain the trend signal and fluctuation signal.
[0041] Database construction: A unified measurement scale is used, with each ring as a data sample. The sample includes representative values of the tunnel boring machine operation data during the construction of the ring segments, the surrounding rock category at the location of the ring, and the corrected BQ value corresponding to the category, to construct a structured database for different projects.
[0042] Transfer learning model construction: Integrating multi-task learning frameworks and adversarial training mechanisms to construct transfer learning models;
[0043] Method Validation: The dataset from Project 1 was divided into training and validation sets in an 8:2 ratio. The model was pre-trained based on the database from Project 1, and its effectiveness in predicting surrounding rock grades was validated. Subsequently, the pre-trained model from Project 1 was transferred to Project 2 for fine-tuning and testing. The dataset from Project 2 was also divided into a fine-tuning set and a test set in an 8:2 ratio. Data was taken from the fine-tuning set at different ratios to fine-tune the model. Simultaneously, an independent model was trained with the same amount of data to verify the advantages of transfer learning models under sparse data conditions.
[0044] Furthermore, in the data collection process of this embodiment, the original time series data of the operating parameters are obtained by the tunnel boring machine PLC system at certain time intervals. Different time intervals are determined for different projects with different construction times for each ring of segments, ensuring that the number of samples for each ring is 100 to 150, so that the information presented by the signal is representative.
[0045] Furthermore, in the signal decomposition process of this embodiment, outliers are removed from the original time series data of each ring segment using the 3σ principle to obtain valid time series data, and the valid time series data is normalized to improve the efficiency of subsequent model training.
[0046] Specifically, the VMD method is used to decompose the effective time series data into k sub-signals (IMFs). DFA is performed on each IMF to obtain the scaling exponent α of the IMF. Based on the range of α values, the IMFs are divided into trend signals, fluctuation signals and noise signals. Noise signals are removed, trend signals and fluctuation signals are retained, and identical signals are merged.
[0047] Furthermore, in the database construction process of this embodiment, the trend signals and fluctuation signals of all operating parameters are measured using a unified scale. For trend signals, the average value is used as the representative value of the trend signal in that loop; for fluctuation signals, the coefficient of variation is used as the representative value of the fluctuation signal in that loop. The coefficient of variation is calculated using the following formula:
[0048]
[0049] Furthermore, the transfer learning model constructed in this embodiment consists of a shared encoding layer, an attention layer, a multi-classification layer, a regression layer, a gradient reversal layer, and a data reconstruction layer. The encoding layer is a multi-layer fully connected neural network that uses the ReLU function as the activation function. The input data is first transformed into deep features through the encoding layer, and the deep features are then transformed into weighted features through the attention layer. The weighted features are then directed to three different tasks.
[0050] Specifically, the attention layer consists of a Dense layer, a Softmax layer, and a weighted layer, placed after the encoding layer. The Dense layer calculates the attention score for each feature, using ReLU as the activation function. The Softmax layer normalizes the attention score. The weighted layer calculates a weighted value for the original features based on the normalized attention score, strengthening features that have a greater impact on the model and weakening features that have a smaller impact. The calculation formula is as follows:
[0051] Dense layer: A ji =Activation(X ji ·W T +b ji );
[0052] Softmax layer:
[0053] Weighted layer: f′ ji =f ji *a ji ;
[0054] Among them, A ji Represents the initial attention score; α ji f represents the normalized attention score; ji 'Represents weighted features.
[0055] Task 1 is the regression task, which is the main task of this model. It uses the regression layer to predict the corrected BQ values of the samples. The predicted corrected BQ values are then categorized into different surrounding rock types based on their range, serving as the final prediction result. The regression layer is a fully connected layer, using the ReLU function as the activation function. The mean squared loss function is used for the regression task.
[0056]
[0057] Where, r i r represents the true corrected BQ value of the i-th sample. i 'Represents the corrected BQ value of the prediction for the i-th sample.
[0058] Task 2 is a multi-class classification task, serving as an auxiliary task for the model. It uses a multi-classification layer to predict the probability that a sample belongs to different rock categories. The purpose of this auxiliary task is to improve the accuracy of rock classification. The multi-classification layer is a fully connected layer, using ReLU as the activation function, and cross-entropy loss as the loss function.
[0059]
[0060] Where, p iy This represents the probability that the i-th sample is predicted as the true class y.
[0061] In addition, the weighted features flow through a special gradient inversion layer before flowing to Task 3. The gradient inversion layer does not directly participate in model training. It is located between the encoding layer and the reconstruction layer. It keeps the data format unchanged during forward propagation and multiplies the gradient by -λ during backward propagation.
[0062] Forward propagation: f → f;
[0063] Backpropagation: ▽→-λ·▽.
[0064] Task 3 is the data reconstruction task. The data reconstruction layer takes the output of the gradient inversion layer as input and the reconstructed samples as output. The goal of the reconstruction task is to extract the most representative features from the input data. These features contain complete information about the samples and are strongly correlated with them. The data reconstruction layer is a fully connected layer, using the ReLU function as the activation function. The reconstruction task uses the square of the Euclidean distance between the reconstructed data vector and the input data vector as the loss function.
[0065]
[0066] Where, x ji x represents the i-th element in the j-th input data vector. ji' represents the i-th element in the j-th reconstructed data vector.
[0067] In the model, Tasks 1 and 2 positively optimize the encoding layer, promoting the learned features to possess both the continuity of regression tasks and the discreteness of classification tasks. Task 3, due to the presence of the gradient reversal layer, has a negative effect on the encoding layer, reducing the strong correlation between features and samples. The opposing effects of Tasks 1, 2, and 3 on the encoding layer create an adversarial training mechanism, causing features to no longer be strongly correlated with the engineering dataset but instead possess commonalities. These commonalities ensure accuracy in classification tasks while improving the model's generalization performance. Although the presence of these commonalities reduces the model's accuracy on the specified engineering dataset, it improves the model's performance on other engineering datasets. The optimization rules for the encoding layer are shown in the following formula:
[0068]
[0069] Where, θ 编码层 These represent the parameters of the coding layer.
[0070] In the model, the attention layer prompts the encoding layer to focus on the features that have the greatest combined effect on the three tasks, reducing the impact of irrelevant features on the model's accuracy. The hyperparameter λ, ranging from 0 to 1, is used to balance the importance of different tasks and prevent the model's performance from being biased towards one task. It is a key hyperparameter for the model to learn common features.
[0071] Furthermore, the independent model used for method verification in this embodiment is a traditional deep neural network (DNN), which is used to directly predict the corrected BQ value of the sample.
[0072] The following uses a submarine tunnel project as an example to illustrate the method proposed in this embodiment in detail, including:
[0073] (1) Obtaining datasets from different tunnels: A certain submarine tunnel project includes two main tunnels and one service tunnel. This embodiment involves the southern main tunnel and the service tunnel. The signal decomposition process of the data is as follows: Figure 2As shown in the figure, raw time-series data of the shield tunneling machine's operating parameters were acquired from the PLC system of the south tunnel shield tunnel at 30-second intervals, and from the PLC system of the service tunnel shield tunnel at 15-second intervals. Each loop contained 100-150 data entries. The strata within a single tunneling distance were considered approximately uniform and without abrupt changes; therefore, the data recorded during one loop of tunneling approximately followed a normal distribution. Outliers were removed using the 3σ principle to obtain valid time-series data. For each operating parameter, the valid time-series data was decomposed into different individual signals (IMFs) using VMD. DFA was performed on each IMF to obtain its scaling exponent α. Based on the range of α values, the signals were classified as trend signals, fluctuation signals, or noise signals. Noise signals were removed, trend and fluctuation signals were retained, and identical signals were merged. The α values for different signals are shown in Table 1.
[0074] Table 1
[0075]
[0076] (2) Database creation process as follows Figure 3 As shown. Since the tunneling times and time series lengths differ across rings, it is necessary to unify the measurement scale for the trend and fluctuation signals of all operating parameters. For trend signals, the average value is used as the representative value for that ring; for fluctuation signals, the coefficient of variation is used as the representative value for that ring. The coefficient of variation is calculated using the following formula:
[0077]
[0078] In this way, each data sample represents a ring, containing trend and fluctuation signals of the main tunneling parameters, the surrounding rock type at the ring location, and the corrected BQ value of that type, thereby establishing a structured database.
[0079] (3) Construct a transfer learning model based on multi-task learning and adversarial training. The model framework is as follows: Figure 4 As shown, the transfer learning model consists of a shared encoding layer, a multi-classification layer, a regression layer, a gradient reversal layer, and a data reconstruction layer. The encoding layer is a multi-layer fully connected neural network. Input data first passes through the encoding layer to be transformed into deep features, which then enter the attention layer, which consists of a Dense layer and a Softmax layer. After computation by the attention layer, the features are transformed into weighted features, which are then applied to three different tasks.
[0080] The attention layer is placed at the end of the encoding layer. The Dense layer calculates the attention score for each feature, using ReLU as the activation function. The Softmax layer normalizes the attention score. The weighted layer performs a weighted evaluation of the original features based on the normalized attention score. The attention layer refines the features, making the model focus more on features useful for the task and less on features with less impact on the model. The calculation formula is as follows:
[0081] Dense layer: A ji =Activation(X ji ·W T +b ji );
[0082] Softmax layer:
[0083] Weighted layer: f′ ji =f ji *a ji ;
[0084] Among them, A ji Represents the initial attention score; α ji f represents the normalized attention score; ji 'Represents weighted features;
[0085] Task 1 is the regression task, which is the main task of this model. It uses the regression layer to predict the corrected BQ values of the samples. The predicted corrected BQ values are then categorized into different surrounding rock types based on their range, serving as the final prediction result. The regression layer is a fully connected layer that takes the features learned by the encoding layer as input and outputs the corrected BQ values of the samples. The regression task uses mean squared error (MSE) as the loss function.
[0086]
[0087] Task 2 is a multi-class classification task, serving as an auxiliary task for the model. It uses a multi-classification layer to predict the probability that a sample belongs to different surrounding rock categories. The multi-classification layer is a fully connected layer that takes features extracted by the encoding layer as input and outputs the probability that a sample belongs to different surrounding rock categories. The auxiliary task aims to improve the accuracy of surrounding rock classification, and uses cross-entropy loss as the loss function.
[0088]
[0089] Task 3 is a data reconstruction task, employing a combination of a gradient inversion layer and a data reconstruction layer. The gradient inversion layer is a special layer that does not directly participate in model training. The data reconstruction layer is a fully connected layer. Features pass through both the gradient inversion layer and the data reconstruction layer sequentially. The gradient inversion layer preserves the data format during forward propagation and multiplies the gradient by -λ during backward propagation. The data reconstruction layer takes the output of the gradient inversion layer as input and the reconstructed samples as output. The loss of the reconstruction task is measured by the square of the Euclidean distance between the vectors.
[0090]
[0091] (4) The implementation principle of the transfer learning model is as follows: Figure 5 As shown. Tasks 1 and 2 have a positive effect on optimizing the encoding layer, promoting the learned features to possess both the continuity of regression tasks and the discreteness of classification tasks, thus further improving the accuracy of surrounding rock classification. The data reconstruction layer in Task 3 makes the features learned by the encoding layer more sample-specific, i.e., strongly correlated with the samples, containing complete sample information. However, the presence of the gradient reversal layer causes Task 3 to have a negative effect on optimizing the encoding layer, reducing the strong correlation between features and samples. The opposing effects of Tasks 1, 2, and 3 on the encoding layer create an adversarial training mechanism, causing features to no longer be strongly correlated with the engineering dataset, but rather to possess commonalities. These commonalities ensure accuracy in classification tasks while improving the model's generalization performance. Although the existence of these commonalities reduces the model's accuracy on a specific engineering dataset, it improves the model's performance on other engineering datasets. The overall optimization direction of the encoding layer is shown in the following formula:
[0092]
[0093] The hyperparameter λ, ranging from 0 to 1, is used to balance the importance of different tasks and prevent the model from favoring one task. It is a key hyperparameter for the model to learn common features. The optimal value of λ needs to be determined based on a large number of test results. In this model, the value of λ is set to 0.5.
[0094] (5) Model Testing and Validation. The data set partitioning and model testing and validation process are as follows: Figure 6As shown. This invention first trains the model inside the service tunnel. Traditional CNN and DNN models are chosen for comparison. The service tunnel dataset consists of data from five rock types, with 80% of the data used as the fine-tuning set and the remaining 20% as the validation set. When the F1 score of the pre-trained model reaches above 0.85, it proves the effectiveness of the pre-trained model. Subsequently, the pre-trained model is transferred to the main tunnel for fine-tuning and testing. The main tunnel is adjacent to the service tunnel and has similar geological types, also containing data from the five rock types. The difference lies in the significant difference in the tunnel boring machine configuration, which causes different data distribution characteristics between the service tunnel and the main tunnel. The main tunnel dataset is divided into a fine-tuning set and a test set in an 8:2 ratio. Due to the class imbalance problem, the proportion of different rock types in the fine-tuning set and the test set needs to be consistent with the proportion in the total data of the main tunnel. All data from the service tunnel is used to retrain a transfer learning model, which is then directly used for testing without fine-tuning to observe the model's classification performance. Subsequently, the model was fine-tuned using different proportions of data from the fine-tuning set: 20%, 40%, 60%, and 80%. Changes in classification accuracy were observed. Simultaneously, an independent model was trained using the same amount of data, accepting only the main tunnel data. The classification performance of the fine-tuned transfer learning model was compared with other common models and the independent model. If the transfer learning model outperforms the contrastive and independent models in classification across different proportions of data, and achieves an F1 score higher than 0.85 on the main tunnel test set after fine-tuning with 80% of the dataset, then the superiority of the transfer learning model can be demonstrated.
[0095] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for classifying surrounding rock of shield tunnels based on multi-task learning and adversarial training, characterized in that, include: Collect data from different tunnel projects and perform signal decomposition and structured processing to build a database. The data includes the surrounding rock type and corresponding corrected BQ value at different tunnel segments, as well as the time series data of the tunnel boring machine's operating parameters during the construction of each ring segment, recorded by the tunnel boring machine's PLC system. A transfer learning model is constructed using a multi-task learning framework and adversarial training mechanism. This model is then trained and validated on the aforementioned database to obtain the final transfer learning model. The transfer learning model includes an encoding layer, an attention layer, a multi-classification layer, a regression layer, a gradient inversion layer, and a data reconstruction layer. Input data is first transformed into deep features by the encoding layer. These deep features are then transformed into weighted features by the attention layer. The weighted features are then directed to three different tasks: the regression layer performs the first task, the multi-classification layer performs the second task, and the gradient inversion layer and data reconstruction layer perform the third task. The first task is a regression task, in which the regression layer predicts and corrects the BQ value based on the weighted features, and classifies the surrounding rock category according to the range of the corrected BQ value as the prediction result. The regression layer is a fully connected layer, and the ReLU function is used as the activation function. The second task is a multi-classification task, in which the multi-classification layer predicts the probability of belonging to different surrounding rock categories based on the weighted features. The multi-classification layer is a fully connected layer, and the ReLU function is used as the activation function. The third task is a data reconstruction task, in which the weighted features first flow through the gradient inversion layer, then are input into the data reconstruction layer, and the reconstructed data of the input data is output. The data reconstruction layer is a fully connected layer, and the ReLU function is used as the activation function. During the training process of the transfer learning model, the first and second tasks have a positive effect on the optimization of the encoding layer, while the third task has a negative effect on the optimization of the encoding layer, forming adversarial training. Obtain the target tunnel engineering data, input it into the final transfer learning model, and output the predicted surrounding rock category and the corresponding corrected BQ value.
2. The method for classifying surrounding rock of shield tunnels based on multi-task learning and adversarial training according to claim 1, characterized in that, Performing the signal decomposition includes: Outlier removal and normalization were performed on the time series data of the tunnel boring machine operating parameters during the construction of each ring segment to obtain valid time series data. The VMD-DFA method is used to decompose the effective time series data into several sub-signals and calculate the scaling exponent of each sub-signal. Based on the range of the scaling index, signals are divided into trend signals, fluctuation signals, and noise signals, and the noise signals are removed.
3. The method for classifying surrounding rock of shield tunnels based on multi-task learning and adversarial training according to claim 2, characterized in that, The structured processing and database construction include: All trend and fluctuation signals are measured using a unified scale, with each ring as a data sample, to construct a structured database for different projects. The data sample includes representative values of the tunnel boring machine's operating data during the construction of each ring segment, the surrounding rock type at the location of the ring, and the corresponding corrected BQ value. The representative values of the trend signals are represented by the average value, and the representative values of the fluctuation signals are represented by the coefficient of variation.
4. The method for classifying surrounding rock of shield tunnels based on multi-task learning and adversarial training according to claim 1, characterized in that, The encoding layer employs a multi-layer fully connected neural network, using the ReLU function as the activation function.
5. The method for classifying surrounding rock of shield tunnels based on multi-task learning and adversarial training according to claim 1, characterized in that, The attention layer consists of a dense layer, a softmax layer, and a weighted layer. The dense layer is used to calculate the attention score of deep features, using the ReLU function as the activation function. The softmax layer is used to normalize the attention score. The weighted layer is used to perform weighted evaluation on the deep features based on the normalized attention score and output the weighted features.