Shield tunneling parameter prediction method based on multi-source domain transfer learning

Abstract

The invention discloses a shield tunneling parameter prediction method based on multi-source domain transfer learning, and the method comprises the steps: obtaining real-time tunneling data of a target project, and loading a pre-stored multi-source domain tunneling data set; performing geological domain division and steady-state feature extraction to obtain a target domain spatial-temporal feature sequence and a multi-source domain spatial-temporal feature sequence; inputting the domain invariant feature into a pre-trained space-time attention feature extraction model, and extracting a corresponding domain invariant feature representation; generating a mobility evaluation weight corresponding to each sample in the multi-source domain by using a pre-constructed mobility evaluation strategy; and processing the target domain spatial-temporal feature sequence by using a pre-trained regression prediction model, dynamically adjusting the spatial-temporal attention feature extraction model, and outputting a target domain tunneling parameter prediction value. The dynamic self-adaptive migration of the shield tunneling sample level is realized, the negative migration risk caused by stratum mutation is effectively inhibited, and the generalization prediction precision and the construction safety of the model under complex and changeable working conditions are improved.

Description

Technical Field

[0001] This invention belongs to the field of key parameter prediction for tunnel boring machines, and in particular, it is a method for predicting tunnel boring machine parameters based on multi-source domain transfer learning. Background Technology

[0002] As a core method for developing underground space in modern cities, tunnel boring machines (TBMs) operate in highly complex environments with unpredictable geological changes. Real-time and accurate prediction of tunneling parameters is crucial for optimizing construction efficiency, ensuring equipment safety, and controlling surface subsidence. With the increasing intelligence of construction equipment, utilizing data-driven methods to uncover the nonlinear mapping relationships between tunneling parameters has become a research hotspot in the field of intelligent control of TBMs.

[0003] Currently, domestic and international research largely focuses on using Long Short-Term Memory (LSTM) networks or convolutional neural networks to establish tunneling parameter prediction models for single engineering scenarios. When dealing with cross-engineering applications, existing technologies mainly employ single-source domain transfer learning, measuring the distributional differences between the source and target domains using metrics such as maximum mean difference, and utilizing domain adversarial neural networks to extract common features across different domains. These methods often employ domain-level weighting strategies, assigning uniform weights to all samples based on the distributional similarity of the entire dataset, aiming to narrow the differences in statistical moments between domains within a unified feature space.

[0004] However, this traditional domain-level transfer mechanism still faces multiple challenges when dealing with the complex working conditions of tunnel boring machines (TBMs), mainly manifested in three interrelated technical problems: coarse transfer granularity, static weight logic, and feature alignment contamination. Specifically, existing methods ignore the heterogeneous physical characteristics within the same domain during TBM tunneling, treating high-quality samples from the same engineering background and low-quality samples affected by isolated boulders or abrupt geological changes equally, resulting in a lack of fine-grained discrimination in transfer weight allocation. This lack of granularity means that weight calculation is often independent of the training process, lacking dynamic feedback based on the target domain task performance, easily falling into suboptimal solutions and causing severe negative transfer phenomena. Furthermore, when performing distribution alignment, forcibly aligning all source domain samples introduces noisy features from low-quality samples into the feature space, causing domain-invariant features to be diluted by irrelevant information, further reducing the model's prediction robustness in scenarios with drastic local geological changes. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting tunnel boring machine parameters based on multi-source domain transfer learning, so as to solve the above-mentioned problems existing in the prior art.

[0006] Technical solution: A method for predicting tunnel boring machine parameters based on multi-source domain transfer learning, comprising:

[0007] Acquire real-time tunneling data of the target project and load pre-stored multi-source domain tunneling dataset;

[0008] Geological domain division and steady-state feature extraction were performed on real-time tunneling data and multi-source domain tunneling datasets to obtain spatiotemporal feature sequences of the target domain and multi-source domain.

[0009] The target domain spatiotemporal feature sequence and the multi-source domain spatiotemporal feature sequence are input into a pre-trained spatiotemporal attention feature extraction model to extract the corresponding domain-invariant feature representations.

[0010] Based on domain-invariant feature representation, a pre-constructed mobility evaluation strategy is used to generate mobility evaluation weights for each sample in the multi-source domain.

[0011] The spatiotemporal feature sequence of the target domain is processed using a pre-trained regression prediction model, and the spatiotemporal attention feature extraction model is dynamically adjusted by combining transferability evaluation weights to output the predicted values ​​of the target domain tunneling parameters.

[0012] Beneficial effects: This invention achieves dynamic adaptive migration at the sample level of tunnel boring machine (TBM) excavation, effectively suppresses the negative migration risk caused by abrupt changes in the strata, and improves the generalization prediction accuracy and construction safety of the model under complex and variable working conditions. Attached Figure Description

[0013] Figure 1 A flowchart illustrating the steps of a shield tunneling parameter prediction method based on multi-source domain transfer learning, provided in this application embodiment.

[0014] Figure 2 A flowchart illustrating the steps for obtaining the target domain spatiotemporal feature sequence and the multi-source domain spatiotemporal feature sequence provided in this application embodiment.

[0015] Figure 3 A flowchart illustrating the steps for extracting the corresponding domain-invariant feature representation provided in this application embodiment.

[0016] Figure 4 A flowchart illustrating the steps for generating mobility evaluation weights for each sample in a multi-source domain, as provided in this embodiment of the application.

[0017] Figure 5 The diagram shows the predicted tunneling speed provided in the embodiments of this application.

[0018] Figure 6 A diagram illustrating the predicted effect of the shield head horizontal deviation provided in an embodiment of this application.

[0019] Figure 7 A diagram illustrating the predicted vertical deviation of the shield head, provided in an embodiment of this application.

[0020] Figure 8 A diagram illustrating the predicted effect of the shield tail horizontal deviation provided in an embodiment of this application.

[0021] Figure 9 The image shows the predicted effect of the vertical deviation of the shield tail provided in the embodiment of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms include and have, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0024] like Figure 1 As shown, a method for predicting tunnel boring machine (TBM) parameters based on multi-source domain transfer learning includes the following steps:

[0025] Acquire real-time tunneling data of the target project and load the pre-stored multi-source domain tunneling dataset.

[0026] In this embodiment, acquiring real-time tunneling data for the target project specifically refers to using sensor arrays deployed on various subsystems of the tunnel boring machine (TBM) to collect physical quantities reflecting the current tunneling status in real time. For example, the sensor array includes pressure sensors mounted on the propulsion system cylinders, torque sensors mounted on the cutterhead drive shaft, and electromagnetic flow meters mounted on the slurry circulation pipeline. The multi-source domain tunneling dataset consists of historical tunneling data pre-stored on a local server or in a cloud database. This data originates from multiple completed TBM projects. Due to differences in the engineering geological conditions of different projects, each independent completed project's data is defined as a source domain, and the data from projects currently under construction is defined as the target domain. By loading the pre-stored multi-source domain tunneling dataset, rich historical prior knowledge can be provided for parameter prediction in the target domain.

[0027] Geological domain division and steady-state feature extraction were performed on real-time tunneling data and multi-source domain tunneling datasets to obtain spatiotemporal feature sequences of the target domain and multi-source domain.

[0028] Specifically, geological domain classification of real-time tunneling data and multi-source domain tunneling datasets refers to dividing the tunneling cross-section into preset geological domain categories based on the lithology of the cross-section strata. Specifically, geological experts classify the cross-section into four major categories based on exploration data: single soft soil strata, single hard soil strata, composite upper soft lower hard soil strata, and special strata. Soil layers within the same major category but with completely consistent geological conditions are defined as different domains. For example, Project A has an upper soft lower hard soil stratum, predominantly hard soil, while Project B has an upper soft lower hard soil stratum, predominantly soft soil; therefore, Project A and Project B are classified as different source domains. Steady-state feature extraction aims to eliminate interference from abnormal operating conditions such as shield machine shutdown and startup phases on data distribution. In this embodiment, shutdown data can be eliminated by identifying whether parameters such as advance speed and total thrust are 0, and steady-state segment data reflecting the tunneling mechanism can be retained using specific statistical truncation logic. The target domain spatiotemporal feature sequence and the multi-source domain spatiotemporal feature sequence refer to multidimensional tensors that contain time step information and multiple sensor dimensions, used to characterize the evolution of tunneling parameters as the number of rings changes.

[0029] The target domain spatiotemporal feature sequence and the multi-source domain spatiotemporal feature sequence are input into a pre-trained spatiotemporal attention feature extraction model to extract the corresponding domain-invariant feature representations.

[0030] In this embodiment, the spatiotemporal feature sequences of the target domain and those of multiple source domains are input into a pre-trained spatiotemporal attention feature extraction model. This allows the use of deep learning networks to extract common deep features across different domains. The pre-trained spatiotemporal attention feature extraction model automatically identifies key variables and key moments in the input feature sequences that significantly contribute to the prediction target. Extracting the corresponding domain-invariant feature representation involves mapping the high-dimensional original temporal features to a low-dimensional feature space. In this space, the distributional differences between the source and target domain data are minimized while preserving the essential features relevant to the geological excavation logic. By extracting domain-invariant feature representations, the domain bias caused by different geological conditions can be effectively reduced, laying the foundation for subsequent cross-domain knowledge transfer.

[0031] Based on domain-invariant feature representation, a pre-constructed transferability evaluation strategy is used to generate transferability evaluation weights for each sample in the multi-source domain.

[0032] Specifically, the transferability evaluation weights are used to quantify the reference value of each source domain sample for the target domain prediction task. Here, the samples specifically correspond to the data generated by each tunneling ring during shield tunneling construction. Because geological conditions within the same project can vary locally during shield tunneling, such as encountering boulders or sand lenses, the correlation between different tunneling rings within the same source domain and the target domain can differ significantly. By generating transferability evaluation weights, fine-grained differentiation at the sample level can be achieved. For example, the transferability evaluation weights combine the overall distribution differences of the source domain with the transferability value assessment of individual samples, accurately identifying and assigning greater weight to high-value samples while suppressing the negative impact of low-value samples on model training.

[0033] The spatiotemporal feature sequence of the target domain is processed using a pre-trained regression prediction model, and the spatiotemporal attention feature extraction model is dynamically adjusted by combining transferability evaluation weights to output the predicted values ​​of the target domain tunneling parameters.

[0034] In other words, based on domain-invariant feature representation, a pre-trained regression prediction model is used for processing, and during the training phase, the loss function of the spatiotemporal attention feature extraction model is dynamically adjusted by combining transferability evaluation weights, and the predicted values ​​of the target domain tunneling parameters are output during the inference phase.

[0035] In this embodiment, processing the spatiotemporal feature sequence of the target domain using a pre-trained regression prediction model means inputting the extracted domain-invariant features into the predictor and mapping them to obtain the final tunneling parameters. Dynamically adjusting the spatiotemporal attention feature extraction model by incorporating transferability evaluation weights means adjusting the contribution of the loss function according to the weight of each sample during the training phase. Specifically, samples with higher transferability evaluation weights have a higher proportion of prediction error and domain classification error in the total loss, forcing the model to prioritize fitting data similar to the target domain features. Through this dynamic adjustment mechanism, the model can achieve adaptive learning of target domain knowledge. The final output predicted values ​​of the target domain tunneling parameters may include key parameters such as the advance speed of the next stage, the horizontal deviation of the shield head, and the vertical deviation of the shield tail. For example, prediction accuracy can be evaluated using root mean square error and accuracy.

[0036] This embodiment achieves adaptive migration prediction of tunneling parameters in the target domain through sample-level migration assessment, solving the problems of poor model generalization performance and negative migration caused by inconsistent stratum distribution during shield tunneling.

[0037] like Figure 2 As shown, in one possible implementation, the target domain spatiotemporal feature sequence and the multi-source domain spatiotemporal feature sequence are obtained, including:

[0038] The tunneling sections are classified into pre-defined geological domains based on the lithology of the strata.

[0039] Specifically, geological engineers classify the construction sections macroscopically based on borehole exploration data and mud discharge observations during tunnel boring. Preferably, the pre-defined geological domain categories include single soft soil strata, single hard soil strata, composite soft-over-hard strata, and special strata. Special strata can specifically refer to composite strata containing spheroidal weathered boulders or cofferdam strata containing a large amount of riprap. Domain division is defined as soil layers under the same broad category but not entirely consistent geological conditions being classified as different domains. For example, in two different urban subway projects, even if both belong to a single soft soil stratum, differences in their specific water content and cohesion will still classify them as different source domains.

[0040] Based on the geological domain category, shutdown data is removed from real-time tunneling data and multi-source domain tunneling datasets, and the steady-state stage data is extracted using the left-skewed 3σ truncation method.

[0041] Alternatively, the raw tunneling data collected in real time is processed by removing shutdown data and the steady-state stage data is extracted using the left-biased 3σ truncation method.

[0042] In this embodiment, the removal of shutdown data from the real-time acquired raw tunneling data is based on the physical operating mechanism of the tunnel boring machine (TBM). For example, by monitoring five key parameters—penetration depth, total thrust, cutterhead torque, cutterhead rotation speed, and propulsion speed—data is determined to be shutdown data and removed when any one of these parameters is 0. The use of the left-skewed 3σ truncation method to extract steady-state stage data addresses the non-steady-state interference during the initial startup phase of tunneling. During TBM tunneling, the thrust is typically low as the state transitions from startup to steady state, resulting in a negatively skewed data distribution. Specifically, the mean μ and standard deviation σ of the tunneling samples are calculated. It is assumed that most of the steady-state time is within the sample mean μ minus three standard deviations σ. By performing a truncation operation, data from the right-hand segment within the range greater than μ-3σ and less than μ+3σ is retained as steady-state stage data. This effectively extracts the stable segment reflecting uniform-speed tunneling performance from the raw sequence containing the entire startup, tunneling, and shutdown process.

[0043] The 3σ criterion is used to remove outliers from the steady-state data, and a sliding window is used to divide the data according to a preset time window to obtain the spatiotemporal feature sequence of the target domain and the spatiotemporal feature sequence of the multi-source domain.

[0044] Specifically, calculate the average value x of the k-th sensor data in a certain ring sample. avg And the standard deviation σ. For example, the standard deviation σ is calculated using the Bessel formula:

[0045] σ = (∑ n=1 N (x n - xavg ) 2 / (N-1)) 0.5 ;

[0046] Where N represents the total number of parameter data samples, x n Let x represent the observed value of the nth parameter sample. avg The arithmetic mean of the samples is used. If the residual error after subtracting the mean from the nth sample exceeds three times σ, it is identified as an outlier and removed. After denoising, a sliding window is used to divide the data according to a preset time window T. Preferably, the preset time window T can be set to a typical value between 5 and 10 rings, for example, T=10. For a given input feature sequence, the sliding window is operated with a step size of 1 ring, using the expert features in the sliding window as input to predict the target parameter value of the next ring. Through the sliding window operation, the original discrete point data is transformed into a tensor structure with temporal correlation, thus obtaining the target domain spatiotemporal feature sequence and the multi-source domain spatiotemporal feature sequence.

[0047] In an exemplary embodiment, in the geological field segmentation and steady-state feature extraction, a feature selection strategy can be used to filter input features, which include:

[0048] Propulsion system parameters, including total thrust, cutterhead torque, cutterhead speed, penetration depth, and propulsion pressure of each set of hydraulic cylinders;

[0049] The parameters of the slurry circulation system include the working chamber pressure, the excavation chamber pressure, the slurry pump speed, and the total grouting volume;

[0050] Geological parameters, including unit weight, internal friction angle, and cohesion.

[0051] In this embodiment, the selection of input features is based on the multi-system coupling mechanism of the tunnel boring machine (TBM) and expert experience. The propulsion system parameters directly reflect the interaction forces between the TBM and the soil. Specifically, the propulsion pressure of each group of hydraulic cylinders can be further refined into six groups (A, B, C, D, E, and F) to characterize the TBM's attitude adjustment force. The slurry circulation system parameters reflect the pressure balance and muck removal efficiency at the excavation face. The working chamber pressure and excavation chamber pressure typically need to be maintained within a certain range to ensure surface settlement control. Geological parameters are key physical quantities describing environmental constraints; unit weight, internal friction angle, and cohesion directly determine the physical and mechanical properties of the soil. By selecting these 20 core input features, it can be ensured that the model captures the most essential soil adaptability information during the feature extraction stage.

[0052] In some alternative implementations, the unit of measurement can be based on the dimension of the input features. For example, total thrust can be measured in kilonewtons (kN), and cutterhead torque in kilonewton-meters (kNm). The unit of measurement for geological parameters can be kilonewtons per cubic meter (kN / m).3 Before constructing the feature sequence, all features must undergo Z-score standardization to eliminate the impact of dimensional differences on network weights. The standardized values ​​are distributed near the zero mean, which helps accelerate the convergence speed of deep learning models.

[0053] This embodiment illustrates the cleaning logic of tunnel boring machine (TBM) data and the selection range of expert features to ensure the quality and physical consistency of the input model data.

[0054] like Figure 3 As shown, according to one aspect of this application, the spatiotemporal attention feature extraction model includes an encoder with added spatial attention mechanism and a decoder with added temporal attention mechanism, extracting corresponding domain-invariant feature representations, including:

[0055] The encoder is used to obtain the hidden state and cell state of the tunneling parameter sequence at each time step, and the spatial attention weight of each feature dimension is calculated based on the concatenation result of the hidden state and cell state.

[0056] Specifically, the encoder employs a spatial attention mechanism to identify which expert feature contributes most to the current prediction target within a specific tunneling loop. Preferably, the encoder can be implemented using a Long Short-Term Memory (LSTM) network. By inputting the input sequence from the previous loop into the LSTM network, the corresponding hidden state h can be obtained. t-1 and unit state s t-1 The hidden states are concatenated with the unit states to obtain a concatenated vector. Based on this concatenated vector, the attention score e for each feature dimension is calculated using the following formula. t，k :

[0057] e t，k = v e T* ·tanh(W e ·[h t-1 ;s t-1 ] + U e ·x t k +b e );

[0058] Where e t，k For the attention score of the k-th feature in the t-th ring, [h t-1 ;s t-1 [b] is the concatenated vector of the hidden state and the unit state at the previous time step. e As a preset bias term, x t k W represents the k-th input feature value of the t-th ring. e U e v eFor the preset learnable weight matrix, T* This is the transpose. The attention score is processed using the softmax function to obtain the spatial attention weight α for each feature dimension. t，k :

[0059] α t，k = exp(e t，k ) / ∑ i=1 n exp(e t，i );

[0060] Where n is the total number of features. Using the softmax function ensures that the sum of all spatial attention weights is 1.

[0061] The input features are weighted using spatial attention weights, and the weighted input features are then fed into the encoder's long short-term memory network to update the encoding state at the current time step.

[0062] In this embodiment, after obtaining the spatial attention weights for each feature dimension, the original input features are rescaled. Specifically, the k-th feature value of the t-th ring is multiplied by its corresponding spatial attention weight to obtain a weighted feature value. By recombinating all the weighted feature values, the input latent variables are obtained. The input latent variables are then input into the encoder's Long Short-Term Memory network to update the current hidden state h. t and unit state s t The encoding state can prioritize the retention of system parameter information that is more relevant to the target tunneling task, effectively suppressing noise interference caused by redundant features.

[0063] The decoder calculates the temporal attention weights based on the pre-stored decoded hidden state of the previous time step and the encoder's outputs at each time step, and generates a context vector based on the temporal attention weights to extract domain-invariant feature representations.

[0064] Specifically, the decoder introduces a temporal attention mechanism to address the forgetting problem in long-term sequence prediction. In this embodiment, given the sequence of hidden states generated by the encoder at past time points, the decoder decodes the hidden state d from the previous time point. t-1 and cell state s' t-1 Combined with the encoder's hidden state h i Calculate the time attention score l of the t-th ring to the information of the i-th ring. t，i :

[0065] l t，i = v d T* ·tanh(W d ·[d t-1 ;s' t-1 ] + Ud ·h i + b d );

[0066] Among them W d U d v d b is a learnable weight vector d This is the bias term in the decoder's temporal attention mechanism. Similarly, the softmax function is used to convert the score into temporal attention weights β. t，i We perform a weighted summation of all encoder hidden states to obtain the context vector c of the t-th ring. t :

[0067] c t =∑ i=1 T β t，i ·h i ;

[0068] Where T is the length of the time window, ∑ i=1 T This represents the summation of all encoder hidden states within the time window. Context vector c t This constitutes a domain-invariant feature representation. While preserving the spatiotemporal evolution characteristics of tunnel boring machines, this vector extracts deep common knowledge across disciplines.

[0069] In some optional implementations, the number of hidden units m and q in the encoder and decoder can be set according to the experimental precision. For example, the hidden layer dimension m can be set to 128, and the number of layers can be set to 2. Through the spatiotemporal attention coupling structure, the model can still extract robust tunneling state representations under environments with drastic fluctuations in formation conditions, providing a reliable feature basis for subsequent transfer weighting.

[0070] This embodiment illustrates the process by which the spatiotemporal attention feature extraction model performs in-depth characterization of shield tunneling data through a dual attention mechanism, providing support for extracting high-quality domain-invariant features.

[0071] like Figure 4 As shown, in one embodiment of this application, generating the mobility evaluation weights corresponding to each sample in the multi-source domain includes:

[0072] Calculate the global distribution distance between each source domain and the target domain under the domain-invariant feature representation, and obtain the domain-level weights corresponding to each source domain after normalization.

[0073] In this embodiment, the maximum mean difference operator can be used to measure the overall difference between the source domain data distribution and the target domain data distribution. Specifically, assuming the number of source domains is NS, and the domain-invariant feature matrix corresponding to the i-th source domain is H-1. cSi The domain-invariant eigenvalue matrix corresponding to the target domain is H. c T The features are mapped to the regenerating kernel Hilbert space using a preset kernel function, and the squared distance between the mean points of the two sets of features is calculated. After obtaining the maximum mean difference (MMD) distance for each source domain, a negative exponential function is used for numerical mapping, and the domain-level weights w are obtained through normalization. i domain The normalization formula is:

[0074] w i domain =exp(-MMD 2 (H c Si H c T )) / ∑ k=1 NS exp(-MMD 2 (H c Sk H c T ));

[0075] Where w i domain Let be the domain-level weight of the i-th source domain, exp be an exponential function with the natural logarithm as the base, and MMD be the domain-level weight. 2 Let be the squared maximum mean difference between the i-th source domain and the target domain, and NS be the total number of source domains. Domain-level weights serve as coarse-grained prior indicators, used to measure the contribution of different tunneling projects to the current task from a macro-geological consistency perspective.

[0076] The domain-invariant feature representations of each sample in the multi-source domain are input into a pre-trained transferability evaluation network, which outputs the sample-level weights corresponding to each sample.

[0077] In a preferred implementation, the sample-level weights corresponding to each sample are output, including:

[0078] The domain-invariant feature representation is input into the transferability evaluation network, and the sample-level evaluation latent variables are extracted through a two-layer fully connected network structure.

[0079] Specifically, the transferability evaluation network consists of two fully connected neural networks. Domain-invariant feature representations are input into this network and undergo a non-linear transformation through the first fully connected layer to extract sample-level evaluation latent variables. The hidden layer dimension of the first fully connected layer is consistent with the dimension of the domain-invariant feature representation, for example, set to 128. The evaluation latent variable h... eval The calculation formula is:

[0080] heval =ReLU(W1·h c Si，j +b1);

[0081] Where h c Si，j Let W1 represent the domain-invariant feature of the j-th tunneling loop sample in the i-th source domain, and W1 and b1 be the weight matrix and bias vector of the first fully connected layer, respectively. ReLU is the linear rectified activation function. Intermediate feature information that can characterize whether a sample contains transferable knowledge is extracted from the deep features.

[0082] The evaluation latent variables are normalized using the Sigmoid activation function to obtain sample-level weights with values ​​in the range [0, 1].

[0083] In this embodiment, latent variables are mapped to scalars through a second fully connected layer, and then compressed to the range of 0 to 1 using a logistic function. Sample-level weights w i，j sample The calculation formula is:

[0084] w i，j sample =σ(W2·h eval +b2);

[0085] Where W2 and b2 are the weight matrix and bias term of the second fully connected layer, and σ represents the Sigmoid activation function. Sample-level weights enable fine-grained quantification of the migration value of different samples within the same source domain. For example, in a source domain project, most tunneling rings are located in sandy silt layers. If the j-th ring encounters a boulder stratum, the migration assessment network (TE network) can identify that the sample's features have low relevance to the target domain task and output a smaller sample-level weight.

[0086] The domain-level weights and sample-level weights are multiplied to obtain the composite weights used as the transferability evaluation weights.

[0087] In an exemplary embodiment, the composite weight w is calculated using the following formula. i，j :

[0088] w i，j =w i domain ·w i，j sample ;

[0089] Where w i domain w represents the domain-level weight corresponding to the i-th source domain. i，j sample This represents the sample-level weight corresponding to the j-th sample in the i-th source domain.

[0090] Specifically, using product operations to obtain composite weights enables two-layer collaborative weighting. Domain-level weights provide a global prior, ensuring that weight allocation does not deviate from macroscopic geological logic; sample-level weights perform fine-grained modulation, identifying and suppressing inferior samples in the source domain. The resulting composite weight w i，j The total loss function that directly affects the model.

[0091] It should be noted that the composite weights are formed by multiplying the domain-level weights and sample-level weights because: domain-level weights are calculated based on the overall distribution distance, reflecting the macroscopic similarity between the source and target domains, providing stable global prior information, and preventing excessive deviation of sample-level weights due to local noise. Sample-level weights are dynamically predicted by the transferability evaluation network based on the domain-invariant characteristics of each sample, achieving fine-grained differentiation of the transfer value of different samples within the same source domain. The multiplication of the two forms a hierarchical modulation mechanism, which retains the overall constraints at the domain level to ensure the stability of weight allocation, while introducing fine-tuning at the sample level to improve the accuracy of weight allocation. In practical applications, when the overall distribution of a source domain is similar to that of the target domain but contains a few anomalous samples, the domain-level weights give the source domain a higher base weight, while the sample-level weights reduce the weight of the anomalous samples. After multiplication, the final weight of the anomalous samples is effectively suppressed.

[0092] In some alternative implementations, the transferability evaluation network can employ a lightweight single-layer multilayer perceptron (MLP) structure to reduce computational overhead, or a deep network with dropout layers to enhance the robustness of the evaluation. Furthermore, in addition to product-based methods, composite weights can also be fused using a weighted average.

[0093] This embodiment identifies the fine-grained migration value of tunneling samples through a migration assessment network and combines it with domain-level weights to form composite weights, thus solving the problem of negative migration caused by local abrupt changes in the formation.

[0094] In a specific numerical case, suppose there are 3 source domains (NS=3), and the domain-invariant eigenvalue matrix of the target domain is H. c T After MMD distance calculation, the MMD values ​​between each source domain and the target domain are obtained. 2The values ​​are as follows: 0.15 for source domain 1, 0.08 for source domain 2, and 0.32 for source domain 3. Calculated using the domain-level weight formula: Source domain 1: exp(-0.15) / [exp(-0.15)+exp(-0.08)+exp(-0.32)]=0.860 / 2.468=0.348; Source domain 2: exp(-0.08) / 2.468=0.923 / 2.468=0.374; Source domain 3: exp(-0.32) / 2.468=0.726 / 2.468=0.294. Assume the domain-invariant feature h of the 5th tunneling ring sample in source domain 2... c S2，5 After being input into the TE network, the evaluation latent variable h is obtained through the first fully connected layer. eval =[0.82, -0.35, 1.24] T* After Sigmoid activation, the sample-level weights w are obtained. 2，5 sample =0.73. Therefore, the composite weight corresponding to this sample is: w 2，5 =w2 domain ·w 2，5 sample =0.374 × 0.73 = 0.273. In contrast, if a sample from a tunneling ring in source domain three, affected by isolated boulders, receives a sample-level weight of only 0.12 after evaluation by the TE network, then its composite weight is: w 3，j =0.294×0.12=0.035, which is much lower than that of normal samples, and thus it is effectively suppressed in the loss function.

[0095] In one possible embodiment, the spatiotemporal attention feature extraction model is dynamically adjusted by incorporating transferability evaluation weights, including:

[0096] Calculate the statistical median of the sample-level weights for each sample in the current batch, and use it as the dynamic evaluation threshold.

[0097] In this embodiment, during each iteration of training, the transferability score distribution of all source domain samples within the input batch is statistically analyzed in real time. Specifically, if the current batch size is 64, the 64 corresponding sample-level weights are obtained and sorted from smallest to largest, and the value in the middle position is taken as the dynamic evaluation threshold τ. threshold Using the dynamic median as a threshold ensures that approximately 50% of high-quality samples participate in cross-source domain distribution alignment at different stages of model training, enhancing the algorithm's adaptability to changes in feature distribution at different stages.

[0098] Based on the dynamic evaluation threshold, samples in the multi-source domain are screened, high-value samples with sample-level weights higher than the dynamic evaluation threshold are extracted, and the domain-invariant feature representations corresponding to the high-value samples are obtained.

[0099] Specifically, all source domain samples in the current batch are compared with their corresponding mobility scores, and only those with weight scores higher than τ are retained. threshold A subset of samples. High-value samples refer to the tunneling loop data that the TE network determines are highly relevant to the target domain task. By obtaining the domain-invariant feature representations corresponding to the high-value samples, a filtered feature matrix H is constructed. c Si This process excludes low-quality samples that may contain isolated boulders, sampling noise, or severe conflicts with the target strata, preventing them from interfering with the stability of the feature space during subsequent alignment.

[0100] The maximum mean difference operator is used to measure the distribution of domain-invariant feature representations of high-value samples between different source domains, generating inter-source domain moment matching constraint loss, and then applying alignment constraints to the spatiotemporal attention feature extraction model based on the inter-source domain moment matching constraint loss.

[0101] In this embodiment, the MMD distance between pairwise high-value sample subsets of the source domain is calculated. The inter-source domain moment matching constraint loss L... mmd_selective The calculation formula is:

[0102] L mmd_selective =∑ i=1 NS-1 ∑ k=i+1 NS MMD 2 (H c Si H c Sk );

[0103] The maximum mean difference operator (MMD operator) employs a Gaussian kernel function. Preferably, the kernel bandwidth σ* of the Gaussian kernel function is automatically determined using a median heuristic method, and its calculation formula is as follows:

[0104] σ* = median(||x i -x j ||) / 2 0.5 ;

[0105] Where median represents the median operation, x i and x j ||x represents the high-value sample feature vector of the input. i -x j|| represents the Euclidean distance between sample feature vectors. Using the median heuristic, the kernel bandwidth can adaptively adjust with the current feature distribution scale, solving the gradient vanishing or alignment inaccuracy problems that may occur with fixed bandwidth. Adding the generated inter-source domain moment matching constraint loss to the total loss function guides the feature extraction model to bring high-value features from different source domains closer together, forming a purer and more universally applicable domain-invariant feature space.

[0106] In a further embodiment, dynamically adjusting the spatiotemporal attention feature extraction model by incorporating transferability evaluation weights further includes:

[0107] Each source-target domain pair is processed by a predetermined independent domain discriminator, and the corresponding domain classification probability is output and the classification loss is calculated.

[0108] In this embodiment, N independent neighborhood discriminators D are designed for N source domains. i Each domain discriminator D i This is used to perform binary classification tasks, specifically determining whether an input domain-invariant feature originates from the i-th source domain or the target domain. The discriminator outputs the corresponding domain classification probability, which characterizes the model's ability to distinguish the source of the feature. The classification loss can be calculated based on the domain classification probability.

[0109] The classification loss generated by the domain discriminator is weighted based on composite weights to obtain the sample-level weighted discriminant loss.

[0110] Specifically, sample importance adjustment is introduced in adversarial training. Sample-level weighted discriminant loss L... d The calculation formula is:

[0111] L d = (1 / NS)·∑ i=1 NS [(∑ j w i，j ·l d (D i (h c Si，j ),0)) / (∑ j w i，j ) + (1 / nT)·∑ k l d (D i (h c T ), 1)];

[0112] Among them l d The cross-entropy loss is used for binary classification, where nT is the number of samples in the target domain, and h is the number of samples in the target domain. c TThis is the domain-invariant feature representation obtained after processing the target domain samples by the domain-invariant feature extraction module. By introducing composite weights, the discriminator will pay more attention to the alignment effect of high-weight samples, further enhancing the accuracy of fine-grained transfer.

[0113] Adversarial optimization of the spatiotemporal attention feature extraction model is performed using sample-level weighted discriminative loss.

[0114] Alternatively, it can be described as using sample-level weighted discriminative loss and a spatiotemporal attention feature extraction model for adversarial optimization to reduce distributional differences between different domains. A gradient inversion layer enables the features generated by the feature extractor to deceive the discriminator, achieving deep fusion of feature spaces from different domains.

[0115] In this embodiment, the sample-level adaptive weighting mechanism and selective inter-source domain moment matching work together to form a coupling enhancement effect. The mechanism of this coupling effect includes the following four aspects: First, the sample-level weighting mechanism evaluates the domain-invariant features of each source domain sample through the transferability evaluation network, identifying high-transfer-value samples that are highly relevant to the target domain prediction task, providing a screening basis for the subsequent feature alignment process. Second, selective moment matching, based on the sample-level weight screening results, only performs cross-source domain distribution alignment on high-value samples that exceed the threshold. Through the selective alignment strategy, the contamination of the feature space by low-value samples is effectively avoided. Low-value samples may contain isolated rock interference, sampling noise, or information that conflicts with the target stratum tunneling logic; including them in the alignment process will lead to a decrease in the quality of domain-invariant features. Third, the purity of the feature space is improved after selective alignment, and the domain-invariant features can more accurately represent the cross-domain shared tunneling patterns. The purified feature space, in turn, improves the discrimination accuracy of the transferability evaluation network, enabling it to more accurately distinguish between high-value and low-value samples. Fourth, the above three links form a positive feedback loop. The weighting mechanism provides a selection criterion, selective alignment improves feature quality, and improved feature quality further enhances weighted accuracy. Through the coupling enhancement effect, the model can maintain robust propulsion velocity prediction performance under complex working conditions with drastic changes in formation conditions.

[0116] In some alternative implementations, if the number of high-value samples after screening is severely imbalanced, an oversampling strategy can be used to supplement the samples from the fewer categories. Furthermore, the dynamic evaluation threshold can also be set as a percentage-based Top-K weight value.

[0117] This embodiment improves the purity of the feature space by screening and aligning high-transfer-value samples, and solves the problem of kernel parameter determination in MMD calculation.

[0118] In one possible implementation, the spatiotemporal attention feature extraction model and the transferability evaluation network are pre-trained in the following manner:

[0119] The performance feedback of the pre-stored target domain validation set is used as a supervision signal to actively optimize the migration evaluation strategy of the migration evaluation network.

[0120] During training, the source domain loss is weighted according to the sample-level weights of the transferability evaluation network output to guide the spatiotemporal attention feature extraction model to learn feature distributions that are beneficial to the target domain task.

[0121] Specifically, before model training begins, the collected target domain data is pre-divided into a target domain training set D. Ttrain and target domain validation set D T_val The optimal partitioning ratio is 8:2. Active optimization refers to optimizing the model based on the current model's performance in D, without relying on pre-set static rules. T_val The weight allocation strategy is adjusted based on the real-time regression accuracy. If the current composite weight allocation can guide the model to obtain a small root mean square error on the target domain validation set, the weight allocation strategy is deemed reasonable; otherwise, the parameters ψ of the TE network are corrected through feedback signals.

[0122] It's important to note that the Transfer Evaluation Network (TE) employs implicit supervised learning. Its output sample-level weights are not derived from manual labeling of the true labels, but rather indirectly supervised by the outer optimization signal within the meta-learning framework. Specifically, the training objective of the TE network is to minimize the prediction loss on the target domain validation set through its output weight allocation. This very objective—that the model performs better in the target domain—constitutes the implicit supervision signal for the TE network. This avoids the need for manual labeling of transfer value for each source domain sample, achieving end-to-end adaptive weight learning.

[0123] In a preferred implementation, proactive optimization employs a meta-learning two-layer optimization architecture, including:

[0124] Inner layer optimization steps: Fix the parameters of the transferability evaluation network, construct the total loss function based on the composite weights, calculate the gradients of the spatiotemporal attention feature extraction model and the regression prediction model, and update the model parameters.

[0125] In other words, the parameters of the fixed transferability evaluation network are used to weight the prediction loss of the multi-source domain based on composite weights, calculate the gradients of the spatiotemporal attention feature extraction model and the regression prediction model, and update the model parameters.

[0126] In this embodiment, the training process is executed according to a two-layer iterative logic. In the inner optimization step, the parameters ψ of the transferability evaluation network are fixed, and the weighted total loss L is calculated using multi-source domain training data and composite weights. total The main network parameters θ are updated, including the spatiotemporal attention feature extraction model, regression prediction model, and domain discriminator. The inner layer update equation is:

[0127] θ'=θ-α θ ·▽ θ L total (θ;ψ);

[0128] Where α θ The preset inner learning rate, for example, set to 1e. -4 ; θ' represents the updated master network parameters, θ represents the current master network parameters, ▽ θ L is the operator for finding the gradient with respect to the parameter θ. total Let ψ be the total loss function that incorporates composite weights, and let ψ be the parameters of the transferability evaluation network.

[0129] Outer layer optimization steps: Fix the updated model parameters, calculate the validation loss using the target domain validation set, and backpropagate the validation loss to the transferability evaluation network to update the parameters of the transferability evaluation network.

[0130] Specifically, in the outer optimization step, the updated parameters θ' of the inner layer are fixed, and the validation set samples from the target domain are input into the model to obtain the validation loss L. meta The verification loss is backpropagated to the TE network to update its parameters ψ. The outer layer update equation is:

[0131] ψ'=ψ-α ψ ·▽ ψ L meta (θ'(ψ));

[0132] Where α ψ The outer learning rate, for example, is set to 1e. -5 ;ψ' represents the parameters of the updated transferability evaluation network; ▽ ψ This is the gradient operator for the parameter ψ. Through a meta-learning two-layer optimization architecture, the TE network can predict the impact of its output weights on model performance and learn the optimal sample evaluation logic.

[0133] In a further embodiment, in the meta-learning two-layer optimization architecture:

[0134] In the outer optimization step, when calculating the parameter gradient of the mobility evaluation network, a first-order approximation method of second-order gradient is used to reduce computational complexity by ignoring the second-order derivative term;

[0135] The inner optimization step and the outer optimization step are executed alternately, and the outer optimization step is executed once after every K inner optimization steps, where K is a pre-configured optimization ratio constant.

[0136] In this embodiment, to address the computational overhead issue, the outer gradient is calculated as follows: ψ L metaWhen (θ'(ψ)), expand using the chain rule: ▽ ψ L meta =(ΨL meta / Ψθ')·(Ψθ' / Ψψ), where Ψ is the partial derivative. Since the updated main network parameter θ' depends on the transferability evaluation network parameter ψ, the expanded formula involves calculating second derivatives. A first-order approximation is preferred, i.e., ignoring the second-order terms, simplifying to: ▽ ψ L meta ≈ (ΨL meta / Ψθ')·(-α θ )·(ΨL train / Ψw)· (Ψw / Ψψ), where L train L is the training loss function used in the inner optimization step. total (θ;ψ);w represents the composite weights generated by the transfer evaluation network for each source domain sample. This approximation improves training efficiency while ensuring convergence performance. Simultaneously, an optimization scaling constant K=5 is set, meaning that a TE network policy update is performed once every 5 main network parameter updates to maintain the stability of the training process.

[0137] Furthermore, to more clearly illustrate the calculation process of the first-order approximation, the gradient propagation chain is expanded as follows: According to the chain rule, the gradient of the transfer evaluation network parameters with respect to the meta-learning validation loss in the outer optimization layer can be expressed as the product of the partial derivative of the validation loss with respect to the updated model parameters and the partial derivative of the updated model parameters with respect to the transfer evaluation network parameters. The partial derivative of the updated model parameters with respect to the transfer evaluation network parameters involves the expansion of the inner update formula. Since the inner update formula is the model parameters minus the product of the learning rate and the total loss gradient, taking the partial derivative with respect to the transfer evaluation network parameters yields the product of the negative mixed second-order partial derivative of the learning rate and the total loss. This mixed second-order partial derivative is the Hessian matrix, and its calculation requires differentiating it with respect to both the model parameters and the transfer evaluation network parameters; the computational complexity is proportional to the square of the number of parameters.

[0138] In practical engineering applications, the computational cost of accurately calculating the Hessian matrix is ​​too high. A first-order approximation method is preferred, ignoring the aforementioned second-order terms and simplifying the complex gradient calculation to the product of the partial derivative of the validation loss with respect to the updated model parameters multiplied by the negative learning rate, the partial derivative of the source domain prediction loss with respect to the composite weights, and the partial derivative of the composite weights with respect to the transferability evaluation network parameters. The update direction of the transferability evaluation network parameters is determined by several factors, including the model's prediction error on the target domain validation set under the current weight allocation, the impact of changes in the weights of each source domain sample on the source domain prediction loss, and the sensitivity of the transferability evaluation network output to changes in its parameters. The first-order approximation method achieves a good balance between computational efficiency and convergence performance, making it suitable for applications with limited computational resources in tunnel boring machine (TBM) construction sites.

[0139] In some alternative implementations, the validation loss may also include a neighborhood alignment term to simultaneously constrain the prediction and alignment performance of the weights. Furthermore, the optimization scale K can be dynamically adjusted based on the rate of decrease of the validation set loss.

[0140] This embodiment uses a meta-learning framework to enable the TE network to actively evolve its parameters based on performance feedback from the target domain validation set.

[0141] According to another aspect of this application, dynamically adjusting the spatiotemporal attention feature extraction model in conjunction with transferability evaluation weights further includes:

[0142] The orthogonal constraint loss between the shared feature matrix output by the spatiotemporal attention feature extraction model and the private feature matrices output by the pre-configured domain-specific private feature extraction models is calculated.

[0143] In this embodiment, calculating the orthogonal constraint loss ensures that the domain-invariant features (shared features) extracted by the model do not contain domain-specific private noise. In addition to the spatiotemporal attention feature extraction model, structurally identical private feature extraction models are configured for both the target domain and each source domain to capture geological characteristics within a specific engineering context. The orthogonal constraint loss L... diff The calculation formula is:

[0144] L diff =||(H c T ) T* ·H P T || F 2 +(1 / NS)·∑ i=1 NS ||(H c Si ) T* ·H p Si ||F 2 ;

[0145] Where H c T and H P T H represents the shared feature matrix (i.e., the domain-invariant feature matrix) and the private feature matrix of the target domain, respectively. c Si and H p Si Let ||·|| represent the shared feature matrix and the private feature matrix of the i-th source domain, respectively. F 2 Denotes the square of the Frobenius norm. T* This is a transpose. By minimizing the cross-covariance between shared and private features, the model is forced to decouple common knowledge of the formation (such as tunneling mechanisms) from domain-specific knowledge (such as equipment wear characteristics specific to a project), thereby improving the purity of the representation of domain-invariant features. Optionally, the orthogonal constraint loss L... diff The calculation formula can also be: L diff =||(H c T ) T* ·H P T || F 2 .

[0146] By summing the prediction loss calculated based on composite weights, the sample-level weighted discrimination loss, the source domain moment matching constraint loss, and the orthogonal constraint loss using pre-configured weight coefficients, a total loss function is constructed to adjust the parameters of the spatiotemporal attention feature extraction model.

[0147] Alternatively, the total loss function can be constructed by summing the prediction loss, sample-level weighted discrimination loss, source domain moment matching constraint loss, and orthogonal constraint loss using pre-configured weight coefficients to adjust the model parameters.

[0148] In this embodiment, the total loss function is constructed to achieve a balance among multiple technical objectives. The total loss function L... total The formula for linear weighted summation is:

[0149] L total =λ1·L y S +λ2·L y T -λ3·L d +λ4·L c +λ5·L diff +λ6·L mmd_selective ;

[0150] Where L y S L represents the sample-level weighted prediction loss. y T L represents the target domain prediction loss. d L represents the sample-level weighted discriminant loss. c L represents the domain classification loss. diff L represents the orthogonal constraint loss. mmd_selective This represents the selective source domain moment matching constraint loss. The pre-configured weight coefficients λ1 to λ6 are hyperparameters. For example, λ3 is set to 0.1, and the other coefficients can be set to 0.01. It should be noted that in adversarial training, the discriminative loss L... d The sign before the gradient inversion layer is negative, indicating that an adversarial game is implemented between the feature extractor and the discriminator. For example, the sample-level weighted discriminative loss L... d = (1 / NS)∑ i=1 NS [(∑ j=1 nSi w i，j ·l d (D i (G f c (x Si j )), 0)) / (∑ j=1 nSi w i，j )]+(1 / nT)∑ k=1 nT l d (D i (G f c (x T k )), 1); NS represents the number of source domains, nSi represents the number of samples in the i-th source domain, nT represents the number of samples in the target domain, w i，j Let l represent the composite weight of the i-th source domain and the j-th sample. d D represents the binary classification cross-entropy loss function. i Let G represent the discriminator for the i-th domain. f c Indicates a shared feature extractor, x Si j Let x represent the j-th sample from the i-th source domain. T k Let represent the k-th sample in the target domain. In the sample-level weighted discriminant loss formula, the first term is the source domain weighted discriminant loss, with a true label of 0 indicating it comes from the source domain; the second term is the target domain discriminant loss, with a true label of 1 indicating it comes from the target domain. The denominator is ∑ j=1 nSi wi，j This is a normalization factor that ensures the weighted losses within each source domain are on the same order of magnitude.

[0151] In some embodiments, weight coefficients λ1 and λ2 control the dominance of the prediction task and can be set to 1 to ensure that the model meets the basic accuracy requirements for predicting tunneling parameters; λ3 controls the adversarial training intensity and is set to 0.1 to avoid training instability caused by an overly strong discriminator while ensuring feature alignment; λ4, λ5, and λ6 serve as auxiliary constraint terms and are set to 0.01 based on empirical values, making their contribution to the total loss secondary but still having a regularization effect. In practical applications, the above coefficients can be fine-tuned according to the performance of the validation set. For example, λ3 can be appropriately reduced when the negative transfer risk is high, and λ6 can be appropriately increased when the feature distribution differences between the source domains are large.

[0152] In a further embodiment, the complete calculation method of the sample-level weighted discriminant loss is described. Specifically, the weighted discriminant loss consists of two parts: source domain discriminant loss and target domain discriminant loss. For the source domain discriminant loss, the binary cross-entropy loss of each source domain sample after passing through the domain discriminator is calculated, where a true label of 0 indicates that the sample comes from the source domain. The cross-entropy loss is weighted using the composite weights corresponding to the sample, and normalized within each source domain by a normalization factor equal to the sum of the composite weights of all samples in that source domain. The normalized weighted losses of all source domains are then averaged. For the target domain discriminant loss, the binary cross-entropy loss of each target domain sample after passing through the domain discriminator is calculated, where a true label of 1 indicates that the sample comes from the target domain. The losses of all target domain samples are then averaged. The weighted discriminant loss is the sum of the source domain discriminant loss and the target domain loss. Through this weighting method, source domain samples with high transfer value gain greater influence in adversarial training, prompting the shared feature extractor to learn the domain-invariant representations of these samples. The influence of source domain samples with low transfer value is suppressed to avoid interfering with the learning process of domain-invariant features.

[0153] In a preferred implementation, the sample-level weighted prediction loss L in the total loss function y S for:

[0154] L y S =∑ i，j (w i，j ·l(f(x i，j ), y i，j )) / ∑ i，j w i，j ;

[0155] Where l is the mean squared error loss function, f(x) i，j y represents the predicted values ​​of the tunneling parameters for the source domain samples by the regression prediction model. i，jw represents the actual sensor tag value. i，j The weights are composite. By using a formula for predicting loss at the sample level, the model can preferentially fit high-weight samples, effectively avoiding negative migration caused by local differences in the formation.

[0156] In some alternative implementations, private features can also be input into a separate prediction branch, and useful information from the private features can be fed back to the shared predictor through knowledge distillation. Furthermore, the weight coefficients λ of each loss term can be dynamically scaled according to the loss reduction rate during the training phase.

[0157] This embodiment integrates loss terms from multiple dimensions, including prediction, discrimination, alignment, and orthogonal constraints, to construct a globally optimized loss function. Orthogonal constraints are then used to achieve deep separation between shared and private features, thereby mining specific stratigraphic knowledge and improving prediction accuracy.

[0158] In one embodiment of this application, key network parameters are limited to ensure the model converges to the optimal solution. The LSTM hidden layer dimensions m and q in the feature extraction model are both set to 128. The hidden layer dimension of the transferability evaluation network (TE network) is also set to 128, consistent with the domain-invariant feature dimension. In the total loss calculation, the weight coefficient λ of the sample-level weighted discriminative loss is initially set to 0.1. The inner layer learning rate α in the meta-learning two-layer optimization is set to 1e^(-1 / 2). -4 The learning rate β of the outer evaluation network is set to 1e. -5 The optimization proportionality constant K was set to 5. During training, the dynamic adjustment of the two-layer weights exhibited significant adaptive characteristics. In the early stages of training, because the transfer evaluation network had not yet fully learned, the sample-level weight distribution was relatively uniform. At this time, the composite weights were mainly dominated by domain-level weights, and the model relied on the overall distribution similarity between the source and target domains for knowledge transfer. As training progressed, the transfer evaluation network continuously received feedback signals from the target domain validation set through a meta-learning strategy, gradually learning to identify source domain samples that were truly valuable for the target domain prediction task. In the later stages of training, the sample-level weight distribution showed obvious differentiation characteristics, with high-value samples receiving weights close to 1 and low-value samples having weights close to 0, resulting in a more refined distribution of composite weights. Through the weight evolution process from coarse-grained to fine-grained, transfer learning maintained stability at different training stages while gradually improving the accuracy of weight allocation.

[0159] Furthermore, in the training process, the target domain data is strictly divided into training and validation sets. The validation set comprises 20% of the total target domain data, serving as a performance feedback signal for outer layer optimization. The training process employs the Adam optimizer, with a batch size of 64 and a total of 100 training epochs. The preferred hardware environment is a deep learning workstation with GPU acceleration, and the software environment is built upon mainstream deep learning frameworks such as PyTorch or TensorFlow.

[0160] Optionally, during model training, the following regularization strategies are also employed to prevent overfitting: a Dropout layer is inserted between two fully connected layers in the TE network, with a dropout rate set to 0.3; L2 regularization constraints are applied to the weight matrix of the feature extraction model, with a regularization coefficient set to 1e. -4 A cosine annealing learning rate scheduling strategy is adopted, with the learning rate decaying from the initial value to 1% of the initial value according to a cosine function. Regarding inference performance, on a workstation equipped with an NVIDIA RTX 3080 GPU, the prediction inference latency for a single 10-ring time window is approximately 8-12ms, meeting the real-time requirements of shield tunneling construction at a sampling frequency of 1Hz. In actual deployment, the model can be deployed on an industrial control computer inside the shield tunneling machine cabin, or the data can be uploaded to a remote server via a 5G network for inference and to return prediction results.

[0161] Preferably, when classifying geological domains, the boundaries are defined by the typical distribution range of geological parameters. For example, the typical cohesion value for a single soft soil stratum ranges from 10 to 30 kPa, and the typical internal friction angle ranges from 5 to 15 degrees. The internal friction angle for a single hard soil stratum is typically greater than 30 degrees. For composite strata, the unit weight typically fluctuates between 18 and 22 kN per cubic meter. By providing these specific quantitative indicators, engineers in the field can accurately classify data and apply the multi-source domain migration prediction method provided in this application when facing new strata.

[0162] In some alternative implementations, when target domain data is extremely scarce, data augmentation techniques (such as synthetic data based on generative adversarial networks) can be used to expand the target domain validation set and improve the stability of the meta-learning outer layer optimization.

[0163] In another embodiment of this application, the dynamic operation of the model at an actual construction site is described, demonstrating its contribution to improving tunneling safety and efficiency. Specifically, at the tunnel boring machine (TBM) site, sensors collect data in real time at a frequency of 1 Hz. The collected raw data stream undergoes real-time preprocessing by an onboard computer, including instantaneous noise reduction based on the 3σ criterion. The preprocessed data is used to construct a real-time input sequence according to a 10-ring time window T. The real-time constructed temporal feature data is input into a pre-trained transferability assessment-spatiotemporal dynamic weighting-adversarial domain adaptation (TE-SDW-DANN) network. Internally, the model identifies current key features through a spatiotemporal attention mechanism and uses a meta-learning optimized TE network to calculate the transfer weights of the current ring relative to each historical source domain. The regression predictor outputs the prediction result for the next ring, for example, predicting that the advance speed of the next ring will decrease from 30 mm / min to 25 mm / min. Based on the predicted values ​​output by the model, construction personnel anticipate possible attitude deviations or a decrease in tunneling efficiency. For example, if a prediction indicates that the horizontal deviation of the shield head will exceed a preset threshold, the system can automatically calculate and suggest increasing the propulsion pressure of the B or C group of hydraulic cylinders to achieve early correction of the tunnel boring machine's attitude. Through closed-loop feedback, the reliance on manual experience during construction is reduced, and the level of intelligence in tunnel boring is improved.

[0164] In some alternative implementations, the prediction results can be pushed to a large screen or mobile terminal in the construction monitoring center in real time, and combined with an audible and visual alarm device to provide early warning of abnormal tunneling trends.

[0165] In a detailed embodiment, a subway tunneling project in a certain city has currently reached the 1520th ring. Geological exploration shows that this section is a composite stratum with a soft upper layer and a hard lower layer, consisting of silty clay in the upper layer and moderately weathered granite in the lower layer. The system has pre-loaded multi-source domain data from three completed projects: Source Domain 1 is a single soft soil project (5000 rings of data), Source Domain 2 is a composite stratum project with a soft upper layer and a hard lower layer (3800 rings of data), and Source Domain 3 is a single hard soil project (4200 rings of data). Real-time acquisition of tunneling data for the current rings 1510-1519 is performed, including 20-dimensional parameters such as total thrust 2850kN, cutterhead torque 1320kNm, and penetration depth 8.5mm / r. After shutdown denoising and 3σ noise reduction, an input feature tensor is constructed. A spatiotemporal attention model extracts a 128-dimensional domain-invariant feature vector. Spatial attention identified cutterhead torque and C-group cylinder pressure as the most critical features, with weights of 0.18 and 0.15, respectively. Since the current strata are closest to source domain two (soft on top, hard on the bottom), the average composite weight for source domain two is 0.42, higher than 0.28 for source domain one and 0.19 for source domain three. Sample-level weights for some hard soil layers in source domain three are below 0.1, and these are classified as low migration value samples. The regression prediction model outputs the following predictions for the next stage: advance speed 28.3 mm / min, shield head horizontal deviation prediction -2.1 mm, showing a leftward trend. The system recommends increasing the advance pressure of the B-group cylinders by 50 kPa to correct the leftward trend, which was implemented after confirmation by the construction personnel.

[0166] According to one aspect of this application, the output weight distribution of a mobility assessment network is used to identify and actively suppress negative migration, enabling the model to maintain predictive stability under extreme geological conditions, including:

[0167] Obtain the composite weights corresponding to each source domain sample within the current tunneling batch, and statistically analyze the weight distribution characteristics of the composite weights.

[0168] In this embodiment, during each round of model inference, the weight sequence generated by the mobility assessment network is exported in real time. The average and variance of the weights for that batch are calculated. In this embodiment, the negative migration risk operator R is defined as the difference between the average weight of the sample-level weights in the current batch and a preset normal threshold (e.g., 0.3). When the average migration weight of a source domain in the batch is consistently lower than this threshold, it is determined that there is a serious distribution conflict between the source domain and the current target domain's tunneling status, i.e., there is a negative migration risk.

[0169] Dynamic masking is performed based on distribution characteristics to truncate high-risk source domain samples with weights.

[0170] Specifically, a strategy combining hard thresholding and soft weighting is employed. For example, a negative migration shutdown threshold τ is set. off The value is 0.1. If the composite weight w of a certain source domain sample is...i，j If the weight falls below the cutoff threshold, the weight of that sample in the current iteration step is forcibly set to zero. This weight truncation mechanism prevents historical data that is completely contrary to the target stratum logic from entering the backpropagation process, ensuring the physical rationality of the prediction results while maintaining the model feature alignment performance.

[0171] In some alternative implementations, when severe negative migration is detected, the system can automatically reduce the weight coefficient λ3 of the neighborhood adversarial loss in the total loss function. By reducing the adversarial intensity, the model switches from a forced alignment mode to a cautious migration mode, maintaining basic tunneling parameter prediction accuracy under extreme geological abrupt changes.

[0172] According to another aspect of this application, the structure of the domain discriminator is refined, including:

[0173] Configure an independent domain discriminator for each source-target domain pair and extract discriminative features through a parameter sharing layer.

[0174] In this embodiment, NS parallel classification branches are constructed at the back end of the Domain Adaptive Network (DANN) framework for adversarial learning. Each domain discriminator contains two hidden layers, whose inputs are domain-invariant feature representations generated by a spatiotemporal attention feature extraction model. Preferably, the first fully connected layer of each domain discriminator adopts a shared weight approach to extract general domain-discriminating features; while the second fully connected layer maintains parameter independence to capture the unique distributional differences between a specific source domain and the target domain. This quasi-shared structure effectively balances computational efficiency and alignment accuracy in multi-source domain scenarios.

[0175] By combining sample-level weighted discriminative loss, adversarial optimization is performed using a gradient inversion layer.

[0176] Specifically, a gradient inversion operator (GRL) is inserted between the feature extraction model and the discriminators in each domain. During forward propagation, GRL does not modify the features; however, during backward propagation, GRL multiplies the gradients returned from the discriminators by a negative scaling factor -λ. d This information is then passed to the feature extraction model. Through this mechanism, the feature extraction model is forced to learn the feature distribution that cannot be distinguished by the discriminators. Combined with sample-level weighted discriminative loss, the model can ensure that source domain samples with higher transfer value dominate gradient updates during feature alignment, achieving more targeted cross-domain feature alignment.

[0177] In some alternative implementations, the scaling factor λ d The number of training rounds increases linearly from 0 to 1. This allows the model to focus on learning the mining logic of the source domain in the early stages and gradually increase the adversarial intensity in the later stages, ensuring smooth convergence of the multi-source domain transfer learning process.

[0178] In this embodiment, when multiple source domains exist, the collaborative work between the discriminators is ensured, and the feature extraction model achieves balanced extraction of common features from different domains.

[0179] In some optional implementations, the following boundary conditions and anomaly handling mechanisms are also included: For each input feature dimension, if the rate of change of the values ​​of 10 consecutive sampling points is less than a preset threshold, such as 0.1%, or the value exceeds the physically reasonable range, such as the total thrust exceeding 120% of the equipment's rated value, then the sensor is determined to be potentially faulty. In this case, the system uses the historical moving average of the most recent 30 rings for interpolation and issues a sensor maintenance reminder to the operator. If there is data missing within the time window, such as due to communication interruption causing data not to be collected for some rings, when the missing proportion is less than 30%, a linear interpolation method is used to complete the data; when the missing proportion exceeds 30%, the system switches to a conservative prediction mode, outputting only a reference prediction value based on the historical statistical mean, and marking it as having a low confidence level. When the negative migration risk operator R exceeds the preset warning value for 5 consecutive batches, the system automatically reduces the domain adversarial loss weight λ3 temporarily to 0, relying only on the target domain's own data for prediction, and records an anomaly log for subsequent analysis.

[0180] In some possible implementations, the spatiotemporal attention feature extraction model can adopt the following alternative structures: a self-attention architecture based on a Transformer encoder, replacing the recurrent structure of LSTM with a multi-head attention mechanism; a dilated causal convolutional architecture based on a Temporal Convolutional Network (TCN); or a topology-aware architecture based on a Graph Neural Network (GNN), treating each sensor as a graph node and modeling their interdependencies. These alternative structures can be selected based on specific engineering requirements and computational resources, while maintaining the core transfer learning framework of this application.

[0181] Furthermore, the method described in this application is not limited to the slurry balance shield tunneling machine scenario, but can also be extended to parameter prediction tasks for other underground tunneling equipment such as earth pressure balance shield tunneling machines, hard rock tunnel boring machines (TBMs), and rectangular pipe jacking machines. When applied to different equipment types, the input feature set needs to be adjusted according to the equipment characteristics. For example, for TBMs, features such as rock strength and tool wear need to be added, but the overall transfer learning framework and sample-level weight evaluation strategy remain unchanged.

[0182] In one possible embodiment, the shield tunneling parameter prediction method based on multi-source domain transfer learning can also be:

[0183] S1. Based on geological conditions, the domain is divided into source domain and target domain.

[0184] Specifically, geological classification is quite complex. Based on expert experience, the strata at the tunneling section of a slurry shield tunneling machine can generally be categorized into four main types: single soft soil strata, single hard soil strata, composite strata with soft upper and hard lower layers, and special strata. When geological conditions fall within the same major category, the parameter prediction model becomes transferable, defining soil layers under geological conditions that are defined within the same major category but not entirely consistent as different domains.

[0185] S2. Obtain the operating data of the tunnel boring machine under normal working conditions, and perform missing value processing, outlier removal and standardization on the data.

[0186] In this embodiment, shutdown data is removed, and the steady-state phase is extracted. A left-skewed 3σ truncation method is used, meaning that most of the steady-state time is within the sample mean μ minus three standard deviations σ. The right-hand segment remaining after removing the left 3σ segment is considered the steady-state phase. Outliers are removed using the 3σ criterion, and the standard deviation σ is calculated as follows:

[0187] σ=sqrt(1 / (N-1)[∑ n=1 N (x n k ) 2 -(1 / N)(∑ n=1 N x n k ) 2 ]);

[0188] Where N is the number of parameter data, x n k Let x be the value of the k-th parameter of the n-th sample. n k If the remaining error after subtracting the sample mean μ exceeds 3 times σ, it can be determined that it is an outlier and should be removed.

[0189] S3. Determine the target as parameters such as the horizontal / vertical deviation of the shield head, the horizontal / vertical deviation of the shield tail, and the average propulsion speed. Based on the feature selection strategy, process the source domain and target domain data, divide the time window, and filter features to obtain the source domain and target domain datasets.

[0190] Specifically, expert variables related to the predictor variables are selected as input variables. Not all variables in the data collected by the tunnel boring machine (TBM) are related to the aforementioned tunneling parameters; redundant features can lead to a decrease in data model performance and increase the complexity of the analysis. Therefore, feature selection should be carefully considered. The control of the TBM's tunneling speed, attitude, and energy consumption is the result of the combined effects of multiple systems, primarily influenced by the propulsion system, grouting and slurry treatment system, and geological conditions. A total of three geological attributes are preferred, and a total of 17 TBM expert features are used (k=17, n=20). Based on the domain division, a time window T is set for the source and target domains. A sliding window operation is performed on the given input X, allowing the values ​​(y1, ..., y...) in the sliding window to be... T-1 X1, X2, ..., X T ) as input, y T As output; where y T-1 X is the target variable value at time step T-1; T Let y be the input variable value at the T-th time step; T Let be the target variable value at the T-th time step.

[0191] S4. Put multiple source domain data and target domain data into the FDW-DANN (Domain Invariant Feature Differential Weighted Domain Adversarial Neural Network) network, set the hyperparameters of the network, and perform offline training of the network.

[0192] In this embodiment, the maximum mean discrepancy (MMD) distance is used to measure the difference between each source domain domain-invariant feature and the target domain domain-invariant feature;

[0193] The formula for calculating the maximum mean difference is:

[0194] MMD(X,Y)=||μ X -μ Y || H =||(1 / ns)∑ i=1 ns Φ(x i )- (1 / nT)∑ j=1 nT Φ(x j )|| H ;

[0195] Where X is the set of input variables; Y is the set of output or target variables; μ X and μ Y Let be the mean values ​​of the samples in the source and target domains, respectively, and Φ be the nonlinear function in the Hilbert space of the reproducing kernel; H is the norm in the reproducing kernel Hilbert space; ns is the number of samples in the source domain; nT is the number of samples in the target domain; xi For a single sample in the source domain, x j For a single sample in the target domain, the hyperparameters of the FDW-DANN network are set, and the network is trained offline. For example, the FDW-DANN combined model consists of three main parts: a feature extractor, a label predictor, and a domain discriminator. The feature extraction network includes a shared feature extractor and a private feature extractor, employing a Long Short-Term Memory (LSTM) network with spatial and temporal attention mechanisms. Based on the differences between source domain features and target domain features in the network's domain-invariant features, differential weighting is applied to the domain discrimination loss and prediction loss in different source domains, thereby adaptively learning samples from different source domains that are similar to the target domain. The label predictor is a regression predictor, and the shared domain features and private features extracted from the source and target domains are simultaneously input into a shared domain parameter predictor, reducing both the source and target domain label prediction losses. The domain discriminator's objective consists of the source domain discrimination loss and the target domain discrimination loss. The objective function of the domain discriminator is as follows:

[0196] R(W, b) = max u，z [-(1 / ns)∑ i=1 ns L d i (W, b, u, z) - (1 / nT)∑ i=ns+1 N L d i (W, b, u, z)];

[0197] Where ns is the number of samples in the source domain, nT is the number of samples in the target domain, and N = ns + nT is the total number of samples; W and b are the weights and bias parameters of the domain discriminator; u and z are the optimization variables of the domain discriminator; L d i Domain discrimination loss for a single sample.

[0198] Alternatively, offline training of the FDW-DANN network can also include adding a spatial attention mechanism encoder to the feature extractor network in FDW-DANN, which employs a long short-term memory network with spatial and temporal attention mechanisms. The attention score formula is as follows:

[0199] e t k =v e T* tanh(W e [h t-1 ;s t-1 ]+U e x k );

[0200] Where et k The spatial attention score for the k-th feature at time step t; [h t-1 ;s t-1 ]∈R 2m The hidden state h obtained after inputting the input sequence from the previous loop into the LSTM network. t-1 and unit state s t-1 It is composed of splices, where m is the number of hidden units in the encoder LSTM network, and R 2m Represents a real vector space of dimension 2m; x k This is the k-th input feature. e ∈R T W e ∈R T×2m U e ∈R T×T All are learnable parameters. The decoder incorporates a temporal attention mechanism, given a context vector [h1, h2, ..., h...]. T Afterwards, the hidden state d of the decoder t-1 and cell state s' t-1 The attention weight of each encoder hidden state in ring t is calculated using the following formula:

[0201] l t i =v d T* tanh(W d [d t-1 ;s' t-1 ]+U d h i ), 1≤i≤T;

[0202] β t i =exp(l t i ) / ∑ j=1 T exp(l t j );

[0203] Where [d] t-1 ;s' t-1 ]∈R 2q It is the concatenation of the hidden state and the cell state of the (t-1)th ring, where q is the number of hidden units in the decoder LSTM, and R is the number of hidden units. 2q Let v be a real vector space of dimension 2q. d ∈R m W d ∈R m×2q U d ∈R m×m It is a learnable parameter, βt i The value represents the attention score of the encoder corresponding to the hidden state in the i-th ring; l t i h is the temporal attention score for the i-th encoder hidden state at time step t; i This represents the hidden state at step i. The label predictor is a regression predictor, and the common features and private features extracted from the source and target domains are simultaneously input into the common domain parameter predictor. This reduces both the label prediction loss in the source domain and the label prediction loss in the target domain. The overall label prediction loss of the network is as follows:

[0204] L y =L y c +L y p ;

[0205] Where L y c For the loss prediction of parameters in the common domain, L y p Predict the loss for parameters in the private domain. Based on the differences between source domain features and target domain features in the network's domain-invariant features, apply differential weighting to the domain discrimination loss and prediction loss for different source domains. Adaptively learn samples from different source domains that are similar to the target domain knowledge. Assuming the number of source domains is NS, the MMD distance between the domain-invariant features of the i-th source domain and the target domain is:

[0206] l i =MMD 2 (h c Si h c T );

[0207] The moment matching loss between different source domains is the MMD loss:

[0208] L mmd =∑ i=1 NS-1 ∑ j=i+1 NS MMD 2 (h c Si h c Sj );

[0209] Where h c Si h is the field-invariant feature of the i-th source domain. c T For the domain-invariant features of the target domain, L mmd The moment matching loss is used for multiple source domains.

[0210] S5. Real-time acquisition of tunnel boring machine operation data, input into the trained model for preprocessing, input the time series feature data obtained after the current loop processing into the network, and obtain the predicted value of the target parameter; the prediction accuracy is evaluated by the root mean square error (RMSE) and accuracy (ACC).

[0211] In one exemplary embodiment, the domain is divided based on the classification of tunneling geological conditions. Geological classification is complex; the strata at the tunneling section of a slurry shield tunneling machine can generally be categorized into four main types: single soft soil strata, single hard soil strata, composite strata with soft upper and hard lower layers, and special strata. When geological conditions fall within the same major category, the parameter prediction model exhibits transferability because, in this case, there is some common knowledge between different projects, reducing the possibility of negative transfer. However, even if projects belong to the same major category of geological conditions, the specific lithology, strength, joints and fissures, groundwater, and types of soil contained in the strata at the tunneling section of different projects may vary significantly. These differences will affect the optimal tunneling parameters and performance during the shield tunneling process. In summary, soil strata under the same major category but not entirely consistent geological conditions constitute different domains.

[0212] When any of the key parameters—penetration depth, total thrust, cutterhead torque, cutterhead rotation speed, and propulsion speed—is zero, that segment is considered shutdown data. This data does not reflect the operating conditions of the tunnel boring machine (TBM) and therefore needs to be discarded. A left-skewed 3σ truncation method is preferred to separate the startup and steady-state data. Data outside the range of two standard deviations to the left of the mean is defined as startup data, while data to the right is considered steady-state data. After removing shutdown data, kernel density estimation is used to plot probability density curves for the four tunneling sections. Combining the mechanism of startup and shutdown with the probability density curves, it is known that the tunneling sections all exhibit a negatively skewed distribution. During the period from startup to steady-state, the thrust is relatively small, resulting in a rightward peak and a long tail in the frequency distribution. A left-skewed 3σ truncation is used, and the remaining right-hand segment after removing the left 3σ segment is considered the steady-state stage. After extracting the steady-state data, some outliers may still exist and need to be removed. Laida's rule (3σ criterion) can be used to improve the accuracy of the data. The 3σ criterion calculates the arithmetic mean and residual error from the measured data.

[0213] Assuming the research object is a large-diameter slurry shield tunneling machine, the target value predicted using the tunneling speed, shield head horizontal / vertical deviation, and shield tail horizontal / vertical deviation as parameters reflects the tunneling performance of the shield machine. Not all variables in the data collected by the shield are related to the above tunneling parameters; redundant features can lead to a decrease in data model performance and increase the complexity of the analysis. Therefore, feature selection should be carefully considered. The control of the shield machine's tunneling speed and attitude is the result of the combined action of multiple systems, mainly influenced by the propulsion system, grouting and slurry treatment system, and geological conditions. Using expert features from ring 1 to ring T and target values ​​from ring 1 to ring T-1, the target value of ring T is predicted. In the input feature sequence used, y' T y represents the predicted value of the Tth ring. T X represents the true value of the T-th ring. T Let y' be the observed values ​​of key features between rings T, and F be the model to be constructed. T =F(y1, ...,y) T-1 X1, X2, ..., X T The data is divided into multiple source and target domains based on the domain division principle.

[0214] An encoder with spatial attention and a decoder with temporal attention are added. The spatial attention mechanism determines which expert feature contributes most to the current prediction feature. When learning these features, the network's learnable parameters focus more on the features with high contributions, forming the encoded context vector. After obtaining the attention score from the encoder, a softmax operation is used to obtain the attention weight of the k-th feature in the t-th ring, denoted as α. t k Softmax is used to ensure that the sum of the attention weights is 1, and the formula is: α t，k = exp(e t，k ) / ∑ i=1 n exp(e t，i The input hidden variables of the LSTM are obtained through the attention weights of the t-th ring. The hidden state h of the t-th ring is obtained through the LSTM network. t and unit state s t , where h t =f1(h t-1 , X' t ), where f1() is the state transition function of the encoder LSTM; X' tThese are the latent variables for the LSTM input. After the temporal input of each loop is extracted using a spatial attention mechanism, the hidden state of each loop is obtained as the encoded context vector. The temporal attention mechanism can solve the forgetting problem caused by long time sequences. By calculating the attention score based on the information from each loop, and according to the different contributions of the information from each loop, the network learns its deep temporal relationships, and based on the decoder input y=[y1, ..., y2]... T-1 The LSTM network is used for decoding to calculate the predicted value for the next loop, given the context vector [h1, h2, ..., h...]. T Afterwards, the hidden state d of the decoder t-1 and cell state s' t-1 The attention weights for each encoder hidden state in ring t are calculated. Since each hidden state corresponds to a time component, the attention scores are used to weight and sum the hidden states obtained by the encoder to obtain the context vector c for ring t. t Furthermore, attention mechanisms were added to the time series dimension, such as: c t =∑ i=1 T β t i h i After obtaining the context vector, it can be compared with the target sequence [y1, ..., y2]. T-1 Combined, the subsequent encoder hidden states are solved, such as: y" t-1 =w[y t-1 c t-1 ]+b';d t =f2(d t-1 y” t-1 ); where the learnable parameter w∈R m+1 , b'∈R is to map the concatenated vector to the appropriate encoder input; y" t-1 Let f1 be the input feature vector of the LSTM decoder; f2() is the state transition function of the LSTM decoder. Then the predicted value of the target sequence in the T-th ring can be obtained by the following formula:

[0215] y' T =v y T* (W y [d T c T ]+b w )+b v ;

[0216] Where the learnable parameter W y ∈R q×(q+m) and b w ∈R q Concatenate vector [d] Tc T ]∈R q+m The vector dimension mapped to the decoder's hidden state allows for the learning of parameters v. y ∈R q and b v ∈R represents the parameters of the linear layer, used to produce the final predicted output y'. T ;d T Let c be the hidden state of the decoder LSTM in the T-th ring. T Let be the temporal attention context vector of the T-th ring. Since the common knowledge between each source domain and the target domain is not entirely consistent across multiple source domains, simply merging all source domain data as a single-source domain and applying single-source domain transfer methods often yields poor results, and may even lead to negative transfer due to the introduction of irrelevant or conflicting information. Therefore, the maximum mean difference distance is preferred to measure the difference between the domain-invariant features of each source domain and the domain-invariant features of the target domain, using this difference as an evaluation metric for contribution. Based on the differences between the features of different source domains and the target domain in the network's domain-invariant features, differential weighting is applied to the domain discrimination loss and prediction loss of different source domains, thereby adaptively learning samples from different source domains that are similar to the target domain in terms of knowledge. Moment matching between different source domains narrows the distribution of common features between source domain pairs, and a unique domain discriminator is designed for each source domain-target domain pair. This further narrows the distribution of common features between each source domain-target domain pair, making the expression of common features between multiple source domains and the target domain as rich as possible, thus improving the network's ability to learn common knowledge between source domain-target domain pairs. Assuming the number of source domains is N, and the MMD distance between the domain-invariant features of the i-th source domain and the target domain is l i Moment matching loss L between different source domains mmd We employ MMD loss. Therefore, when assigning different weights to samples from different source domains, the weight of the i-th source domain sample is: w i =e -li / ∑ j=1 NS e -lj ; where e -li Calculate the numerator for the weights of the i-th source domain. The orthogonal constraint loss of the network is: L diff =||(H c T ) T* ·H P T || F 2 +(1 / NS)·∑ i=1 NS ||(H c Si ) T* ·H p Si ||F 2 The parameter prediction loss, domain discrimination loss (i.e., sample-level weighted discrimination loss), and domain classification loss are respectively:

[0217] L y =(1 / NS)∑ i=1 NS w i L y Si +L y T +L y p ;

[0218] L d =(1 / NS)∑ i=1 NS w i L d i ;

[0219] L c =(1 / NS)∑ i=1 NS L c Si +L c T .

[0220] Where L y Si L is the label prediction loss for the i-th source domain; c T The domain classification loss is used for the target domain. In the label prediction loss, L... d i =(1 / nsi)∑ j=1 nSi L d (G d i (G f c (x Si j )), d Si j )+(1 / nT)∑ k=1 nT L d (G d i (G f c (x T k )), d T k );G d i G is the domain discriminator for the i-th source-target domain pair;f c For domain-shared feature extractors; x Si j d represents the j-th sample from the i-th source domain; Si j d is the domain label of the j-th sample in the i-th source domain; T k Let L be the domain label of the k-th sample in the target domain. In the domain classification loss, L... c Si =(1 / nsi)∑ j=1 nSi L c (G c (G f pSi (x Si j )), d Si j );L c Si G is the domain classification loss for the i-th source domain. f pSi G represents the source domain feature extractor for the i-th private domain; c For the domain classifier. In summary, the total loss function of the FDW-DANN network is:

[0221] L total =L y -λL d +αL c +βL diff +γL mmd ;

[0222] Where λ, α, β, and γ are the loss weight hyperparameters.

[0223] The data processing method for real-time data acquisition by the tunnel boring machine is the same as that for offline training. The final model prediction performance is evaluated using two indices: Root Mean Square Error (RMSE) and Accuracy Detection (ACC). The model prediction performance is as follows: Figures 5 to 9 As shown. In the target domain project, when the tunnel boring machine first started excavating, the soil was mainly soft soil, but around ring 380, it entered a harder soil layer. Figure 5The predicted tunneling speed figures show that the actual tunneling speed of the tunnel boring machine (TBM) has been significantly reduced to adapt to hard soil layers, generally remaining at or below 15 mm / min. If only the initial tunneling data is used for training, the model cannot learn the relevant knowledge as the TBM has not yet encountered relatively hard soil layers, and its generalization ability cannot support parameter prediction under hard soil conditions. However, after adversarial training in the FDW-DANN domain, it is clearly evident that the model possesses knowledge of TBM tunneling under hard soil conditions for tunneling speed parameter prediction. The knowledge of tunneling hard soil sections adapted from the multi-source domain has been effectively transferred to the target domain, enabling it to complete the prediction task under hard soil conditions. Furthermore, from... Figure 6 and Figure 8 It can be seen that the trends of the horizontal deviation between the shield head and tail are quite similar, both showing an upward trend from around 400 rings to around 700 rings, with a low frequency of change. This indicates that the model can learn the changing trend using data from the initial tunneling operation. However, because the tunnel boring machine in the target domain project is tunneling at a certain angle on the vertical plane, the requirements for vertical attitude control are high, which brings certain difficulties. Figure 7 and Figure 9 It can be seen that the frequency of change in the vertical deviation of the shield head / tail is higher than that in the horizontal deviation. Using newly started tunneling data to train the model, the model is insufficient to capture the changing trends and frequencies under hard soil. However, by leveraging shared knowledge from multiple source domains and combining an evaluation method for the differences between the source and target domains, the model was able to learn the changing trends and frequencies of the vertical deviation of the shield head / tail. This demonstrates the effectiveness of the shield tunneling parameter prediction method based on multi-source domain transfer learning, and the model effectively completes the target task in the target domain. For RMSE, the closer the result is to 0, the better the performance; for ACC, the closer to 100%, the better. The formulas for root mean square error and accuracy are as follows:

[0224] RMSE=sqrt((1 / n)∑ i=1 n (y i -f i ) 2 );

[0225] ACC=(1-(1 / n)∑ i=1 n |(y i -f i ) / y i |)×100%;

[0226] Among them, y i f is the true value at time t. i Let t be the predicted value at time t, and n be the total number of data points.

[0227] In summary, the shield tunneling parameter prediction method based on multi-source domain transfer learning includes constructing spatiotemporal feature sequences of the target domain and multi-source domains through geological domain segmentation and left-skewed 3σ truncation steady-state extraction; mapping domain-invariant features using a spatiotemporal attention feature extraction model and introducing a transferability evaluation network to perform sample-level transfer value evaluation; generating composite weights by combining domain-level priors and sample-level weights, and actively updating the weight strategy through a meta-learning two-layer optimization architecture with target domain verification performance as feedback; and performing selective inter-source domain moment matching using high-value samples selected based on the median of the weights.

[0228] This invention abandons the traditional domain-level uniform weighting model and introduces a sample-level transferability evaluation network, which can deeply identify the individual value of each link of data in shield tunneling. Even under the same engineering background, it can distinguish between low-quality samples affected by isolated rocks or local complex geological interference and high-value high-quality samples, achieving precise filtering and fine-grained weight allocation of transfer knowledge. It adopts a meta-learning two-layer optimization framework based on target domain performance feedback, so that weight allocation no longer depends solely on pre-calculated distribution distance, but dynamically iterates based on the model's actual prediction performance on the target domain validation set, achieving a feedback loop from the end in mind. This allows the weight strategy to evolve in real time during training, suppressing the negative transfer risk that is easily generated due to inconsistent geological distribution. By implementing a selective source domain moment matching strategy and setting a dynamic threshold, only high-value samples with high transfer scores and consistent physical logic are selected to participate in cross-source domain distribution alignment, avoiding the dilution of the feature space by low-value noise samples and improving the purity of domain-invariant features. This invention enables the model to maintain extremely high prediction robustness and generalization accuracy even under complex working conditions with drastic geological fluctuations, thereby improving the intelligence level of shield machine attitude control and tunneling parameter decision-making.

[0229] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for predicting shield tunneling parameters based on multi-source domain transfer learning, characterized in that, include: Acquire real-time tunneling data of the target project and load pre-stored multi-source domain tunneling dataset; Geological domain division and steady-state feature extraction were performed on real-time tunneling data and multi-source domain tunneling datasets to obtain spatiotemporal feature sequences of the target domain and multi-source domain. The target domain spatiotemporal feature sequence and the multi-source domain spatiotemporal feature sequence are input into a pre-trained spatiotemporal attention feature extraction model to extract the corresponding domain-invariant feature representations. Based on domain-invariant feature representation, a pre-constructed mobility evaluation strategy is used to generate mobility evaluation weights for each sample in the multi-source domain. The spatiotemporal feature sequence of the target domain is processed using a pre-trained regression prediction model, and the spatiotemporal attention feature extraction model is dynamically adjusted by combining transferability evaluation weights to output the predicted values ​​of the target domain tunneling parameters.

2. The method according to claim 1, characterized in that, The target domain spatiotemporal feature sequence and the multi-source domain spatiotemporal feature sequence are obtained, including: The tunneling sections are classified into pre-defined geological domains based on the lithology of the strata. Based on the geological domain category, shutdown data is removed from real-time tunneling data and multi-source domain tunneling datasets, and the steady-state stage data is extracted using the left-skewed 3σ truncation method. The 3σ criterion is used to remove outliers from the steady-state data, and a sliding window is used to divide the data according to a preset time window to obtain the spatiotemporal feature sequence of the target domain and the spatiotemporal feature sequence of the multi-source domain.

3. The method according to claim 1, characterized in that, The spatiotemporal attention feature extraction model includes an encoder with added spatial attention and a decoder with added temporal attention, extracting corresponding domain-invariant feature representations, including: The encoder is used to obtain the hidden state and cell state of the tunneling parameter sequence at each time step, and the spatial attention weight of each feature dimension is calculated based on the concatenation result of the hidden state and cell state. The input features are weighted using spatial attention weights, and the weighted input features are then fed into the encoder's long short-term memory network to update the encoding state at the current time step. The decoder calculates the temporal attention weights based on the pre-stored decoded hidden state of the previous time step and the encoder's outputs at each time step, and generates a context vector accordingly to extract domain-invariant feature representations.

4. The method according to claim 1, characterized in that, Generate the mobility evaluation weights for each sample in the multi-source domain, including: Calculate the global distribution distance between each source domain and the target domain under the domain-invariant feature representation, and obtain the domain-level weights corresponding to each source domain after normalization. The domain-invariant feature representations of each sample in the multi-source domain are input into a pre-trained transferability evaluation network, which outputs the sample-level weights corresponding to each sample. The domain-level weights and sample-level weights are multiplied to obtain the composite weights used as the transferability evaluation weights.

5. The method according to claim 4, characterized in that, Output the sample-level weights for each sample, and obtain the composite weights, including: The domain-invariant feature representation is input into the transferability evaluation network, and the latent evaluation variables at the sample level are extracted through a two-layer fully connected network structure. The evaluation latent variables are normalized using the Sigmoid activation function to obtain sample-level weights with values ​​in the range [0, 1]. Calculate the composite weight w using the following formula. i，j : In i，j =in i domain ·In i，j sample ; Where w i domain w represents the domain-level weight corresponding to the i-th source domain. i，j sample This represents the sample-level weight corresponding to the j-th sample in the i-th source domain.

6. The method according to claim 5, characterized in that, The spatiotemporal attention feature extraction model is dynamically adjusted by incorporating transferability evaluation weights, including: Calculate the statistical median of the sample level weights for each sample in the current batch, and use it as the dynamic evaluation threshold; Based on the dynamic evaluation threshold, samples in the multi-source domain are screened, high-value samples with sample-level weights higher than the dynamic evaluation threshold are extracted, and the domain-invariant feature representations corresponding to the high-value samples are obtained. The maximum mean difference operator is used to measure the distribution of domain-invariant feature representations of high-value samples between different source domains, generating inter-source domain moment matching constraint loss, and thereby applying alignment constraints to the spatiotemporal attention feature extraction model.

7. The method according to claim 6, characterized in that, The dynamic adjustment of the spatiotemporal attention feature extraction model by incorporating transferability evaluation weights also includes: The domain-invariant feature representation of each source-target domain pair is processed by a predetermined independent domain discriminator, and the corresponding domain classification probability is output and the classification loss is calculated. The classification loss generated by the domain discriminator is weighted based on composite weights to obtain the sample-level weighted discriminant loss. Adversarial optimization of the spatiotemporal attention feature extraction model is performed using sample-level weighted discriminant loss.

8. The method according to claim 7, characterized in that, The dynamic adjustment of the spatiotemporal attention feature extraction model by incorporating transferability evaluation weights also includes: Calculate the orthogonal constraint loss between the shared feature matrix output by the spatiotemporal attention feature extraction model and the private feature matrices output by the pre-configured domain-specific private feature extraction models; By summing the prediction loss calculated based on composite weights, the sample-level weighted discrimination loss, the source domain moment matching constraint loss, and the orthogonal constraint loss using pre-configured weight coefficients, a total loss function is constructed to adjust the parameters of the spatiotemporal attention feature extraction model.

9. The method according to claim 5, characterized in that, The spatiotemporal attention feature extraction model and the transferability evaluation network were pre-trained using the following methods: The performance feedback of the pre-stored target domain validation set is used as a supervision signal to actively optimize the migration evaluation strategy of the migration evaluation network. During training, the source domain loss is weighted according to the sample-level weights of the transferability evaluation network output to guide the spatiotemporal attention feature extraction model to learn feature distributions that are beneficial to the target domain task.

10. The method according to claim 9, characterized in that, Active optimization employs a meta-learning two-layer optimization architecture, including: Inner optimization steps: Fix the parameters of the transferability evaluation network, construct the total loss function based on the composite weights, calculate the gradients of the spatiotemporal attention feature extraction model and the regression prediction model, and update the model parameters; Outer layer optimization steps: Fix the updated model parameters, calculate the validation loss using the target domain validation set, and backpropagate the validation loss to the transferability evaluation network to update the parameters of the transferability evaluation network.

Images