Accidental explosion concrete penetration depth prediction method based on two-stage migration adversarial network

Abstract

The invention discloses an accidental explosion concrete penetration depth prediction method based on a two-stage migration adversarial network, and the method comprises the steps: (1) an instance screening stage: resampling target domain test data to construct a plurality of sub-data sets, training a base learning device, and calculating a weight coefficient; and inputting the source domain simulation data into the base learner to obtain a weighted predicted value, screening out the simulation data conforming to a target domain rule through a threshold value, and optimizing the source domain data quality. (2) a prediction stage: constructing a dynamic adversarial network model (including a weight network, a task network and a domain discriminator), distributing dynamic weights through the weight network, and aligning cross-domain feature distribution in combination with domain adversarial learning; weighted task loss and adversarial loss are jointly optimized in back propagation, so that the model adaptively fuses effective information in a hidden space, and a cross-domain generalization prediction network is generated. According to the method, the knowledge migration problem of simulation data and test data is effectively solved, and the method has excellent performance in a concrete penetration depth prediction task.

Description

Technical Field

[0001] The present invention belongs to the technical field of information processing, and particularly relates to a method for predicting the penetration depth of concrete under accidental explosion based on transfer learning. Background Art

[0002] In the field of damage effects of accidental explosions, the reasonable utilization of simulation data is of great significance for making up for the scarcity of real test data. By solving the equations related to damage dynamics, simulation data with different parameter combinations can be generated in batches, thus breaking through the experimental scale and working condition limitations of real tests. However, the reliability of simulation data is restricted by the accuracy of the constitutive model, and there are still certain differences between simulation data and real test data: in terms of the parameter coverage dimension, simulation can go beyond the limitations of actual scaled tests and generate extreme working condition data such as large equivalent charges. However, due to modeling simplification and the uncertainty of material parameters, there may still be deviations between the simulation results and the actual penetration depth of concrete; in terms of data distribution uniformity, simulation samples are usually based on a uniform sampling strategy in the parameter space, while real tests are often concentrated in typical working condition areas due to cost limitations, resulting in a mismatch in data distribution between the two. In this case, traditional transfer methods have certain limitations in transferring simulation data through a single strategy: at the instance level, the method of optimizing sample weights is usually adopted to enhance the contribution of statistically similar simulation instances to the model. However, this strategy depends on the positive effect of source domain data, and extreme working condition data is often regarded as outliers, resulting in a reduction or neglect of its weight during the transfer process, weakening its contribution to the model. At the feature level, the domain difference is usually reduced by screening key features or constructing an optimal mapping. However, due to the highly complex physical mechanisms involved in the explosion damage process, the data exhibits strong non-linearity and multi-variable coupling characteristics, and simple feature mapping is difficult to accurately extract cross-domain shared information. Summary of the Invention

[0003] Object of the Invention: In order to overcome the deficiencies of the prior art, the present invention provides a method for predicting the penetration depth of concrete under accidental explosion based on a two-stage transfer adversarial network, and constructs multiple groups of base learners based on ensemble resampling. By screening simulation samples that conform to the physical laws of the target domain through weighted predicted values, the quality of source domain data is improved; at the same time, a weight network and an adversarial learning mechanism are introduced to dynamically optimize the weights of source domain samples, forcing the model to learn domain-invariant features and reduce the domain difference, and improving the prediction performance on the target domain.

[0004] Technical Solution: To achieve the above object, the present invention provides a method for predicting the penetration depth of concrete under accidental explosion based on a two-stage transfer adversarial network. The solution includes the following steps:

[0005] (1) Instance screening stage based on resampling ensemble learning. First, the experimental data T (target domain) is sampled with replacement to generate multiple sub-datasets, and base learners are sequentially trained on the newly obtained sub-datasets. Subsequently, each base learner is applied to the original experimental data T, and the evaluation metrics of each base learner are evaluated and the weights of each base learner are calculated according to this metric. Finally, the samples in the simulation data S (source domain) are input into each trained base learner for prediction, and the weighted prediction values of the source domain samples are calculated according to their weights. By comparing with the screening threshold, it is judged whether to retain the source domain sample. Through the above steps, some simulation data that is inconsistent with the physical laws of the experimental data or contains noise is eliminated, and the quality of the source domain data is initially improved.

[0006] (2) Prediction stage integrating dynamic instance weighting and domain adversarial learning. First, a prediction network is constructed based on three sub-network structures and the model parameters are initialized. The simulation candidate set samples and experimental data after instance screening are used as the input of the prediction network. First, they are input into the weight network to assign a weight to the source domain samples, indicating the importance of the sample in the target task. Then, the weighted source domain samples and target domain samples are respectively input into the task network and the domain discriminator for adversarial learning, forcing the task network to learn shared feature representations. In the backpropagation stage, the parameters of the entire network are jointly optimized based on the weighted losses of the task network and the pre-discriminator, and finally the converged task network is output for prediction. Since the weighted loss of the source domain samples is added to the training of the adversarial network, the contribution of each source domain sample can be dynamically adjusted to adapt to the target task. And the adversarial learning mechanism aligns the distribution deviation between the simulation data and the experimental data in the implicit feature space, improving the prediction performance on the target domain.

[0007] Furthermore, the preprocessing of the original data in step (1) is to eliminate low-quality samples that are inconsistent with the true physical laws contained in the target domain before using the simulation data for prediction modeling. Step (1) includes the following steps:

[0008] (1.1) Generation of sub-datasets. The original target domain dataset T{(x1, y1), (x2, y2),...(x m , y m )) is sampled by bootstrap sampling with replacement to generate N s sub-datasets T i , and the total number of samples in each sub-dataset is the same as the total number of samples in the original target domain T. This resampling mechanism can alleviate the uneven distribution of experimental data and specifically enhance sparse extreme working condition samples, significantly improving the data coverage breadth and modeling robustness of the target domain.

[0009] (1.2) Training the base learner and calculating the corresponding weight w iSelect a model structure and initialize the model parameters. Utilize multiple sub-datasets T i to train in parallel and obtain N s different base learners BL i Next, evaluate the performance of each base learner based on the original target domain data T and assign a weight w i to each base learner, strengthening the base learners with high consistency in the target domain rules:

[0010]

[0011] Since the prediction of concrete penetration depth belongs to a regression task, the coefficient of determination is used as a measure for each base learner here. The evaluation result of the i-th base learner on the original target domain T is denoted as

[0012] (1.3) Calculate the weighted prediction value of the source domain samples. Specifically, sequentially select a single source domain sample (x, y), input the source domain sample feature x into each trained base learner BL i to obtain the prediction value of this sample. Combine the prediction results of all base learners and their corresponding weights w i to calculate the weighted prediction value of this source domain sample. The formula is as follows:

[0013]

[0014] The weight w i can reflect the generalization ability of the base learner on real data, and different base learners can cover diverse parameter intervals due to differences in the sampled subsets. Through weighted fusion, simulation samples with reasonable physical laws but large parameter spans can be co-verified by multiple base learners, while noise data is suppressed due to the influence of low-weight base learners, ultimately achieving the precise retention of effective samples.

[0015] (1.4) Screen the source domain samples according to the screening threshold. If the relative gap between the weighted prediction value of the current source domain sample and its true value y is less than the threshold, it indicates that its physical law is consistent with the target domain and can play a positive role in predicting the target domain, and this sample needs to be added to the candidate source domain set S c .

[0016]

[0017] 2. Based on the prediction network integrating dynamic instance weighting and domain adversarial learning according to claim (2), this method constructs a three-network collaborative architecture: on the basis of the adversarial training between the prediction network and the domain discriminator, an additional weight network is introduced, and the weight allocation strategy of source domain samples is dynamically optimized during the training process. At the same time, the distribution shift between simulation data and test data is reduced in the implicit parameter space after feature mapping, so as to improve the generalization ability of the model in the target domain. The step (2) includes the following steps:

[0018] (2.1) DIWAN basic framework. A prediction network integrating dynamic instance weighting and domain adversarial learning (Dynamic Instance Weighting Adversarial Networks, DIWAN) is proposed according to the domain alignment theory based on γ-divergence. This method constructs a three-network collaborative architecture: on the basis of the adversarial training between the prediction network and the domain discriminator, an additional weight network is introduced, and the weight allocation strategy of source domain samples is dynamically optimized during the training process. At the same time, the distribution shift between simulation data and test data is reduced in the implicit parameter space after feature mapping, so as to improve the generalization ability of the model in the target domain.

[0019] (2.1.1) Weight Network (h w ): Assign a weight w s to the source domain samples to quantify their contribution to the target task. The weight network adaptively adjusts the weights through the adversarial learning mechanism, suppresses the source domain samples with large distribution differences from the target domain, and enhances the transfer effect of effective samples.

[0020] (2.1.2) Task Network (h t ): The network finally used to predict the target task is composed of a feature extraction module and a prediction module. The feature extraction module is the key part of the entire adversarial learning. It learns the domain-shared feature representation through the joint optimization of the task loss and the adversarial loss, and reduces the distribution shift between simulation data and test data in the new implicit feature space; the prediction module is generally used for the prediction of specific target tasks.

[0021] (2.1.3) Domain Discriminator (Domain Discriminator, h d ): A binary classification network that takes the output of the feature extraction module as input and discriminates whether the feature comes from the source domain or the target domain. Through the backpropagation mechanism of the Gradient Reversal Layer (GRL), the adversarial loss of the domain discriminator is used to drive the alignment of feature distributions.

[0022] (2.2) DIWAN model structure. Through the collaborative optimization of dynamic weights and adversarial learning, DIWAN achieves the dual goals of feature sharing and sample screening in cross-domain tasks. First, domain adversarial training forces the feature distributions of the source domain and the target domain to align, enabling the task network to learn based on domain-invariant features. Second, the weight network dynamically adjusts the sample weights through gradient feedback, suppressing source domain samples with significant distribution differences and improving the migration efficiency. In addition, the introduction of the gradient reversal layer enables the synchronous optimization of feature alignment and weight adjustment in end-to-end training, further enhancing the generalization ability of the model. The simulation candidate data processed by the source domain instance screening algorithm in claim (1) and the test data are input into the model together after mixing.

[0023] (2.2.1) During the forward propagation process, the simulation samples first pass through the weight network h w to generate weights w s , while the target domain samples are assigned uniform weights. To maintain the comparability of the weights, the weights need to be normalized, and then the normalized instance weights and features are input into the task network and the domain discriminator for adversarial learning.

[0024] (2.2.2) The task network h t is the network used to predict the target task, mainly composed of two parts: the feature extraction module G f and the prediction module G p . Its core goal is to generate domain-invariant features f through the feature extraction module and complete the accurate prediction of the concrete penetration depth through the prediction module. Specifically, the weighted source domain samples x s and the target domain samples x t are mapped to features f by the feature extraction module G f , and then the features are input into the prediction module G p to calculate the prediction value. The task loss function is defined as the weighted sum of the source domain weighted prediction error and the target domain prediction error:

[0025]

[0026] where is the loss function related to the task, is the weight dynamically assigned by the weight network for each simulation sample, N s and N [[ID=4l]] T are the numbers of samples in the simulation candidate set and the test data respectively, y s and y t are the true labels of the simulation samples and the test samples respectively.

[0027] (2.2.3) Domain discriminator h dThe goal is to distinguish source domain features from target domain features. The features f generated by the weighted samples through the feature extraction module will also be input into the domain discriminator h d and calculate the domain classification probability corresponding to the sample features. By maximizing the classification error of the domain discriminator, the model forces the feature extraction module to generate domain-invariant features that are difficult to distinguish, thus achieving cross-domain feature alignment. Its loss function is designed as:

[0028]

[0029] where f s and f t represent the features generated by the source domain samples and target domain samples through the feature extraction module respectively, and h d (f s ) and h d (f t ) represent the classification probability values of the source domain samples and target domain samples in the adversarial network respectively. The overall training process is achieved by jointly optimizing the task loss and the adversarial loss, and the total loss function is:

[0030]

[0031] where I is the adversarial loss coefficient, used to adjust the intensity of gradient reversal. To reduce the complexity of hyperparameter tuning, I is fixed at 1 in subsequent experiments.

[0032] (2.2.4) In the backpropagation stage, a Gradient Reversal Layer (GRL) is inserted between the feature extraction module and the domain discriminator. When backpropagating, the update direction of the parameters θ f of the feature extraction module G f is opposite to the optimization direction inside the domain discriminator network, forcing the generation of domain-invariant features; while the parameters θ w of the weight network are also affected by GRL, and it dynamically adjusts w s according to the gradient information, specifically by reducing the weights of samples with larger distribution offsets, thus achieving automatic screening of source domain samples. The parameter update equations of the feature extraction module and the weight network are as follows:

[0033]

[0034] Beneficial effects: The present invention has the following advantages compared with the prior art:

[0035] (1) The multi-stage collaborative optimization of the present invention greatly improves the cross-domain prediction accuracy.

[0036] (2) The present invention has strong robustness and generalization ability, and the model performs excellently in the penetration depth prediction tasks of concrete and metal target material datasets. Brief Description of the Drawings

[0037] Figure 1 is the flowchart of the method of the present invention;

[0038] Figure 2 is the diagram showing the loss curves of three strategies during the training process in the specific embodiment;

[0039] Figure 3 is the diagram showing the ablation results of TANN on the concrete dataset; Detailed Description of the Invention

[0040] The present invention will be further described below with reference to the accompanying drawings.

[0041] As Figure 1 shown, the two-stage transfer adversarial network concrete penetration depth prediction model mainly includes the following steps: In the first stage, based on ensemble learning, the experimental data is resampled to construct multiple groups of base learners, and their weighted prediction values are used to screen the simulation samples that are consistent with the target domain rules, and the noise data deviating from the real scenario is removed; in the second stage, a weight network and an adversarial learning mechanism are introduced to dynamically adjust the weights of the screened simulation samples, and the domain feature distribution difference is minimized through joint optimization, forcing the model to learn domain-invariant features. This method uses the strategy of "screening first and then aligning" to suppress the interference of distribution shift while retaining the physical prior of the simulation data, thereby improving the cross-domain prediction accuracy in the small-sample scenario.

[0042] The specific process is as follows:

[0043] (1) Instance screening stage based on resampling ensemble learning.

[0044] The instance screening stage based on resampling ensemble learning mainly removes the noise data deviating from the real scenario by resampling the experimental data and using multiple groups of base learners.

[0045] (1.1) Sub-dataset generation. The original target domain dataset T{(x1, y1), (x2, y2), …(x m , y m )} is sampled with replacement by bootstrap sampling to generate N s sub-datasets T i , and the total number of samples in each sub-dataset is the same as that of the original target domain T. This resampling mechanism can alleviate the uneven distribution of the experimental data and specifically enhance the sparse extreme working condition samples, significantly improving the data coverage breadth and modeling robustness of the target domain.

[0046] (1.2) Train the base learner and calculate the corresponding weight w i . Select a model structure and initialize the model parameters, and use multiple sub-datasets T iParallel training is performed to obtain N s different base learners BL i . Then, the performance of each base learner is evaluated based on the original target domain data T, and a weight w i is assigned to each base learner, strengthening the base learners with high consistency of the target domain rules:

[0047]

[0048] Since the prediction of concrete penetration depth belongs to a regression task, the coefficient of determination is used as the measurement index for each base learner here. The evaluation result of the i-th base learner on the original target domain T is denoted as

[0049] (1.3) Calculate the weighted prediction value of the source domain sample. Specifically, select a single source domain sample (x, y) in turn, input the source domain sample feature x into each trained base learner BLi, and obtain the prediction value of this sample. Combine the prediction results of all base learners and their corresponding weights w i to calculate the weighted prediction value of this source domain sample . The formula is as follows:

[0050]

[0051] The weight w i can reflect the generalization ability of the base learner on real data, and different base learners can cover diverse parameter intervals due to the differences in the sampled subsets. Through weighted fusion, simulation samples with reasonable physical laws but large parameter spans can be jointly verified by multiple base learners, while noise data is suppressed due to the influence of low-weight base learners, and finally accurate retention of effective samples is achieved.

[0052] (1.4) Screen the source domain samples according to the screening threshold. If the relative gap between the weighted prediction value of the current source domain sample and its true value y is less than the threshold, it indicates that its physical law is consistent with the target domain and can play a positive role in the prediction of the target domain, and this sample needs to be added to the candidate source domain set S c .

[0053]

[0054] (2) The prediction stage that fuses dynamic instance weighting and domain adversarial learning.

[0055] This method constructs a three-network collaborative architecture: on the basis of the adversarial training between the prediction network and the domain discriminator, an additional weight network is introduced to dynamically optimize the weight allocation strategy of source domain samples during the training process, and at the same time reduce the distribution shift between simulation data and experimental data in the implicit parameter space after feature mapping, so as to improve the generalization ability of the model on the target domain.

[0056] (2.1) DIWAN basic framework. A prediction network that combines dynamic instance weighting and domain adversarial learning is proposed according to the domain alignment theory based on γ-divergence. This method constructs a three-network collaborative architecture: on the basis of the adversarial training between the prediction network and the domain discriminator, an additional weight network is introduced to dynamically optimize the weight allocation strategy of source domain samples during the training process, and at the same time reduce the distribution shift between simulation data and experimental data in the implicit parameter space after feature mapping, so as to improve the generalization ability of the model on the target domain.

[0057] The DIWAN model consists of three main components:

[0058] (2.1.1) Weight Network (h w ): Assigns a weight w s to the source domain samples to quantify their contribution to the target task. The weight network adaptively adjusts the weights through an adversarial learning mechanism, suppressing the source domain samples with large distribution differences from the target domain and enhancing the transfer effect of effective samples.

[0059] (2.1.2) Task Network (h t ): The network finally used to predict the target task, which consists of a feature extraction module and a prediction module. The feature extraction module is the key part of the entire adversarial learning. It learns the domain-shared feature representation through the joint optimization of the task loss and the adversarial loss, and reduces the distribution shift between simulation data and experimental data in the new implicit feature space; the prediction module is generally used for the prediction of specific target tasks.

[0060] (2.1.3) Domain Discriminator (h d ): A binary classification network that takes the output of the feature extraction module as input and discriminates whether the feature comes from the source domain or the target domain. Through the backpropagation mechanism of the Gradient Reversal Layer (GRL), the adversarial loss of the domain discriminator is used to drive the alignment of the feature distributions.

[0061] (2.2) DIWAN model structure. DIWAN achieves the dual goals of feature sharing and sample screening in cross-domain tasks through the collaborative optimization of dynamic weights and adversarial learning. First, domain adversarial training forces the feature distributions of the source domain and the target domain to align, enabling the task network to learn based on domain-invariant features; second, the weight network dynamically adjusts the sample weights through gradient feedback, suppressing source domain samples with significant distribution differences and improving the migration efficiency. In addition, the introduction of the gradient reversal layer enables the synchronous optimization of feature alignment and weight adjustment in end-to-end training, further enhancing the generalization ability of the model. The simulation candidate data processed by the source domain instance screening algorithm in claim (1) and the test data are input into the model together after mixing.

[0062] (2.2.1) During the forward propagation process, the simulation samples first pass through the weight network h w to generate the weights w s , while the target domain samples are assigned uniform weights. To maintain the comparability of the weights, the weights need to be normalized, and then the normalized instance weights and features are input into the task network and the domain discriminator for adversarial learning.

[0063] (2.2.2) The task network h t is the network used to predict the target task, mainly composed of two parts: the feature extraction module G f and the prediction module G p . Its core goal is to generate domain-invariant features f through the feature extraction module and complete the accurate prediction of the concrete penetration depth through the prediction module. Specifically, the weighted source domain samples x s and the target domain samples x t are mapped to the features f by the feature extraction module G f , and then the features are input into the prediction module G p to calculate the predicted values. The task loss function is defined as the weighted sum of the source domain weighted prediction error and the target domain prediction error:

[0064]

[0065] where, is the loss function related to the task, is the weight dynamically assigned by the weight network for each simulation sample, N s and N T are the numbers of samples in the simulation candidate set and the test data respectively, y s and y t are the true labels of the simulation samples and the test samples respectively.

[0066] (2.2.3) The domain discriminator h dThe goal is to distinguish source domain features from target domain features. The features f generated by the weighted samples through the feature extraction module are also input into the domain discriminator h d and calculate the domain classification probability corresponding to the sample features. By maximizing the classification error of the domain discriminator, the model forces the feature extraction module to generate domain-invariant features that are difficult to distinguish, thus achieving cross-domain feature alignment. Its loss function is designed as follows:

[0067]

[0068] where f s and f t respectively represent the features generated by the source domain samples and target domain samples through the feature extraction module, and h d (f s ) and h d (f t ) respectively represent the classification probability values of the source domain samples and target domain samples in the adversarial network. The overall training process is achieved by jointly optimizing the task loss and the adversarial loss. The total loss function is:

[0069]

[0070] where I is the adversarial loss coefficient, used to adjust the intensity of gradient reversal. To reduce the complexity of hyperparameter tuning, I is fixed at 1 in subsequent experiments.

[0071] (2.2.4) In the backpropagation stage, a gradient reversal layer (GRL) is inserted between the feature extraction module and the domain discriminator. During backpropagation, the update direction of the parameters θ f of the feature extraction module G f is opposite to the optimization direction inside the domain discriminator network, forcing the generation of domain-invariant features; while the parameters θ w of the weight network are also affected by GRL, and it dynamically adjusts w s according to the gradient information, specifically by reducing the weights of samples with larger distribution offsets, thereby achieving automatic screening of source domain samples. The parameter update equations of the feature extraction module and the weight network are as follows:

[0072]

[0073] Example:

[0074] To verify the effectiveness of the present invention, experiments were conducted on the test data and simulation data of two types of target materials, concrete and metal. By combining the high authenticity of the test data with the high coverage of the simulation data, a model capable of accurately predicting the penetration effects of different target materials was constructed to verify the advantages of the prediction method. The experiments were carried out in two aspects: (1) Ablation analysis of the instance screening stage based on resampling ensemble learning; (2) Ablation analysis of the prediction stage integrating dynamic instance weighting and domain adversarial learning.

[0075] 1. Data Preparation

[0076] In this section, the data provided by the concrete dataset was used as the experimental object, and the dataset information is shown in Table 1.

[0077] Table 1 Dataset Information

[0078]

[0079] 2. Comparative Experiments

[0080] (1) Ablation Analysis of the Instance Screening Stage Based on Resampling Ensemble Learning

[0081] The experiments compared three strategies for validation on the base learner: BP-TO (training the Baseline only with target domain data), BP-DC (directly merging the source domain and target domain data and using the mixed data to train the Baseline), and BP-SC (first screening the source domain data using the instance screening algorithm in TTAN, and then merging the screened data with the target domain data to train the Baseline).

[0082] Table 2 Parameter Information Table

[0083] dataset base learner <![CDATA[N s > δ concrete decision tree regression 12 0.8

[0084] The regression metrics of the three strategies on the concrete and metal datasets are shown in Table 3.

[0085] Table 3 Dataset Information

[0086]

[0087] 1) Results and Analysis of the Concrete Dataset

[0088] As shown in Table 3, there are significant differences in the performance of the above three strategies in the concrete penetration depth prediction task. Compared with the BP-DC and BP-TO strategies, BP-SC shows a significant performance improvement, with a significant increase in the performance in the regression task, fully verifying the important role of the instance screening algorithm in optimizing the model effect.

[0089] In summary, according to the experimental results on the concrete dataset, the BP-SC strategy successfully suppresses the negative transfer phenomenon by screening source domain samples similar to the target domain, further verifying the applicability of the instance screening algorithm in TTAN.

[0090] 2) Loss change results and analysis

[0091] The loss changes of the three strategies during training are as Figure 2 shown. The BP-SC method merged after screening the source domain data shows significant superiority during training. The screened source domain data effectively reduces noise interference and improves the model's adaptability to the target domain, thus significantly improving the model's prediction performance.

[0092] (2) Ablation analysis of the prediction stage integrating dynamic instance weighting and domain adversarial learning

[0093] In this experiment, a weight network is introduced through a domain alignment mechanism based on γ theory, which dynamically assigns weights to source domain samples during the adversarial learning process to balance the contributions of different samples to the target task. On this basis, a series of methods that only retain the basic structure and include weight adjustment or adversarial learning mechanisms are constructed for comparison. Among them, all methods are trained using the data screened by the TTAN first-stage algorithm, and the final results are shown in Table 4 and Appendix Figure 3 .

[0094] 1) Concrete dataset results and analysis

[0095] Table 4 Ablation results of the second stage of the concrete dataset

[0096]

[0097] The experimental results on the concrete dataset show that the adversarial weight learning module plays a crucial role in the TTAN model and presents the best prediction accuracy among all comparison groups.

Claims

1. A method for predicting the penetration depth of concrete by accidental explosion based on a two-stage transfer adversarial network, characterized in that It includes the following steps: (1) Instance screening stage based on resampling ensemble learning. First, perform sampling with replacement on the test data T (target domain) to generate multiple sub-datasets, and train the base learners respectively. Evaluate the performance of each base learner on the original T dataset, and calculate the corresponding weights according to the evaluation metrics. Subsequently, input the source domain simulation data S (source domain) into each base learner for prediction, and obtain the comprehensive prediction value of the sample through weighted calculation. Set a dynamic screening threshold, retain the source domain samples with prediction values higher than the threshold, and eliminate the simulation data that does not conform to the physical laws of the test data or contains noise. (2) Prediction stage integrating dynamic instance weighting and domain adversarial learning. First, construct a prediction model composed of a weight network, a task network, and a domain discriminator and initialize the parameters. Input the source domain simulation candidate set after instance screening and the target domain test data into the weight network to assign dynamic weights to the source domain samples to characterize their importance for the target task. The weighted source domain data and target domain data are respectively input into the task network and the domain discriminator for adversarial training, and the cross-domain adaptability is improved by aligning the implicit feature space distributions. During backpropagation, jointly optimize the network parameters through a weighted loss function, and its loss function includes the prediction error of the task network and the domain confusion loss of adversarial learning. This method innovatively introduces dynamic sample weights into the adversarial framework, enabling high-value source domain samples to play a dominant role in feature alignment, and at the same time effectively eliminating the distribution differences between simulation and test data through the domain adversarial mechanism. The finally converged task network can not only adaptively fuse physical mechanisms and data features, but also suppress the interference of source domain noise, significantly improving the prediction accuracy of the target domain.

2. According to claim (1), the preprocessing of the original data is to eliminate low-quality samples that are inconsistent with the true physical laws contained in the target domain before using the simulation data for prediction modeling. The step (1) includes the following steps: (1.1) Sub - dataset generation. Conduct sampling with replacement on the original target - domain dataset \(T=\{(x_1,y_1),(x_2,y_2),\cdots(x m ,y m )\}\) to generate \(N s \) sub - datasets \(T i \), where the total number of samples in each sub - dataset is the same as that of the original target domain \(T\). (1.2) Train the base learner and calculate the corresponding weight w i . Select the regression tree decision model structure, initialize the model parameters, and use multiple sub-datasets T i to train in parallel to obtain N s different base learners BL i . Then, evaluate the performance of each base learner based on the original target domain data T, and assign a weight w i to the base learners with high consistency in the target domain rules. (1.3) Calculate the weighted prediction value of the source domain samples. Specifically, sequentially select a single source domain sample (x, y), and input the feature x of the source domain sample into each trained base learner BL i to obtain the prediction value of the sample Combine the prediction results of all base learners and their corresponding weights w i to calculate the weighted prediction value of the source domain sample (1.4) Screen the source domain samples according to the screening threshold. If the relative gap between the weighted prediction value of the current source domain sample and its true value y is less than the threshold, it indicates that its physical law is consistent with the target domain and can play a positive role in predicting the target domain. The sample needs to be added to the candidate source domain set S c and 3. According to claim (2), based on the prediction network integrating dynamic instance weighting and domain adversarial learning, this method constructs a three-network collaborative architecture: an additional weight network is introduced on the basis of the adversarial training of the prediction network and the domain discriminator, dynamically optimizing the weight allocation strategy of the source domain samples during the training process, and at the same time reducing the distribution shift between the simulation data and the test data in the implicit parameter space after feature mapping, thereby improving the generalization ability of the model on the target domain. The step (2) includes the following steps: (2.1) Establish a prediction network integrating dynamic instance weighting and domain adversarial learning (Dynamic Instance Weighting Adversarial Networks, DIWAN). An additional weight network is introduced on the basis of the adversarial training of the prediction network and the domain discriminator, dynamically optimizing the weight allocation strategy of the source domain samples during the training process, and at the same time reducing the distribution shift between the simulation data and the test data in the implicit parameter space after feature mapping, thereby improving the generalization ability of the model on the target domain. (2.1.1) Weight Network, h w ): Assign a weight w to the source domain samples s to quantify their contribution to the target task. (2.1.2) Task Network (h t ): The network ultimately used to predict the target task, consisting of two parts: a feature extraction module and a prediction module. (2.1.3) Domain Discriminator, h d ): A binary classification network that takes the output of the feature extraction module as input and discriminates whether the features come from the source domain or the target domain. (2.2) DIWAN model structure. Through the collaborative optimization of dynamic weights and adversarial learning, DIWAN achieves the dual goals of feature sharing and sample screening in cross-domain tasks. First, domain adversarial training forces the feature distributions of the source domain and the target domain to align, enabling the task network to learn based on domain-invariant features. Second, the weight network dynamically adjusts the sample weights through gradient feedback, suppressing source domain samples with significant distribution differences and improving the transfer efficiency. In addition, the introduction of the gradient reversal layer enables the synchronous optimization of feature alignment and weight adjustment in end-to-end training, further enhancing the generalization ability of the model. The simulation candidate data and the test data are input into the model together after being mixed. (2.2.1) During the forward propagation process, the simulation samples first pass through the weight network h w to generate the weights w s , while the target domain samples will be assigned uniform weights. To maintain the comparability of the weights, it is also necessary to normalize the weights, and then input the normalized instance weights and features into the task network and the domain discriminator for adversarial learning. (2.2.2) Task network h t is a network used to predict target tasks, mainly consisting of a feature extraction module G f and a prediction module G p The two parts are composed of. Its core goal is to generate domain-invariant features f through the feature extraction module, and at the same time complete the accurate prediction of the concrete penetration depth through the prediction module. (2.2.3) An additional weight network is introduced based on the adversarial training of the prediction network and the domain discriminator, dynamically optimizing the weight allocation strategy of source domain samples during the training process, and at the same time reducing the distribution shift between the simulation data and the test data in the implicit parameter space after feature mapping, so as to improve the generalization ability of the model on the target domain. (2.2.4) During the backpropagation stage, a Gradient Reversal Layer (GRL) is inserted between the feature extraction module and the domain discriminator. When backpropagating, the parameter θ f of the feature extraction module G f is updated in the opposite direction to the optimization direction inside the domain discriminator network, forcing the generation of domain-invariant features; while the parameter θ w of the weight network is also affected by GRL, and it dynamically adjusts w s according to the gradient information, specifically by reducing the weights of samples with large distribution offsets, thereby achieving automatic screening of source domain samples.

Images