Shield posture multi-step prediction and correction control method based on latent variable enhanced free transformer

Abstract

The application belongs to the technical field of intelligent shield construction, and specifically discloses a shield machine posture multi-step prediction and deviation correction control method based on latent variable enhanced Free Transformer. First, shield construction process data is acquired, a control quantity set is established and preprocessed; error features, change rate features and unit mileage features are constructed based on the preprocessed data, and after normalization, feature sequences are formed; the sequences are organized into token sequences at a single time step, and after position and mileage coding, the pre-trained Free Transformer model is input; the model aggregates all sequence information through a non-causal encoder special block, calculates latent variable distribution parameters based on the hidden state sequence, samples the latent variable, and then maps it to a conditional vector broadcast to the decoding path; finally, the future shield posture vector, the over-threshold probability and the recommended control quantity are output in parallel by the multi-task output head. The application realizes multi-step prediction and deviation correction self-adaptive to working conditions, and can improve the safety and intelligent level of shield construction.

Description

Technical Field

[0001] This application relates to the field of intelligent tunnel boring machine (TBM) construction technology, specifically to a method for multi-step prediction and correction control of TBM attitude based on latent variable enhancement Free Transformer. Background Technology

[0002] During tunnel construction, the attitude of the tunnel boring machine (TBM) (including horizontal and vertical deviations) directly affects the quality of segment assembly, ground settlement control, and construction safety, and must be monitored and actively adjusted in real time. Current engineering practices primarily rely on operator experience combined with sensor data for manual correction of the TBM's attitude, which is inefficient and prone to delayed decision-making.

[0003] In recent years, the field of intelligent tunnel boring machine (TBM) construction has attempted to introduce data-driven models to predict TBM attitude and assist in decision-making. These improvements have achieved some success, but due to the complexity of TBM construction, existing technologies still have the following shortcomings in prediction and control:

[0004] First, the attitude of the tunnel boring machine is affected by multiple coupled factors. Among them, the adjustment effect of operating parameters such as grouting pressure and propulsion speed on attitude has obvious time lag and cumulative effect. Traditional time series models (such as LSTM and GRU) are difficult to accurately characterize such long-range dynamic dependence, resulting in the rapid accumulation of multi-step prediction errors.

[0005] Secondly, key environmental factors in actual construction (such as changes in stratum lithology, groundwater distribution, and soil disturbance in front of the cutterhead) cannot be directly observed by sensors, but they have a decisive impact on the shield-soil interaction mechanism. Existing models are usually based only on observable variables and do not explicitly consider such implicit working conditions. When geological conditions change abruptly or when construction is carried out across sections, the model's generalization ability drops sharply.

[0006] Third, mainstream methods mostly output deterministic point predictions, lacking the quantification of the uncertainty of the prediction results, making it impossible to assess the risk of deviation exceeding limits, and difficult to support risk-sensitive control decisions.

[0007] Finally, existing systems generally treat prediction and control separately: the prediction module only provides future deviation estimates, while the correction strategy still relies on rule bases or manual intervention; some integrated control suggestion methods often ignore the physical constraints of the equipment in their output operating parameters, resulting in a lack of constraint feasibility guarantees for the recommended instructions. Summary of the Invention

[0008] This application provides a method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable augmented Free Transformer, which solves the problems of error accumulation, weak generalization ability, lack of quantitative prediction results and lack of command constraints in the existing technology of tunnel boring machine attitude prediction and correction control.

[0009] The technical solution of this application is as follows:

[0010] The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable augmented Free Transformer includes the following steps:

[0011] S1. Obtain data on the tunnel boring machine (TBM) construction process;

[0012] S2. Based on the preprocessed construction process data, construct error characteristics, rate of change characteristics, and unit mileage characteristics, and normalize the preprocessed construction process data and construction characteristics to obtain a normalized feature sequence.

[0013] S3. Organize the normalized feature sequence into a token sequence and encode each token. Input the encoded token sequence into the pre-trained Free Transformer model.

[0014] The Free Transformer model is configured to: generate latent variables based on the token sequence and inject the latent variables into the decoder;

[0015] The specific steps of the latent variable injection decoder are as follows:

[0016] S31. Construct at least one non-causal Free Transformer block as an encoder-specific block. This encoder-specific block aggregates the output of the shared module used by the encoder and decoder for full-sequence information aggregation.

[0017] S32. Calculate the latent variable distribution parameters based on the latent state sequence output by the encoder, and obtain the latent variables by sampling according to the latent variable distribution parameters;

[0018] S33. Obtain a condition vector with the same dimension as the Free Transformer model through linear mapping, and broadcast the condition vector to the decoding path;

[0019] S4. The multi-task output head of the Free Transformer model based on injected latent variables outputs the predicted shield attitude vector, the over-threshold probability, and the recommended control quantity, respectively.

[0020] Furthermore, the tunnel boring machine (TBM) construction process is indexed by time or mileage, including fields for earth pressure, propulsion, excavated soil, grouting, and grease.

[0021] in,

[0022] The earth pressure field includes the earth pressure setpoint and the actual earth pressure value;

[0023] The propulsion fields include the total thrust of the jacks and the propulsion speed;

[0024] The waste soil field includes the amount of excavated soil;

[0025] The grouting field includes synchronous grouting pressure, synchronous grouting volume, secondary grouting pressure, secondary grouting volume, and grouting date;

[0026] The grease field includes the amount of grease used at the tail of the shield and the pressure of grease at the tail of the shield.

[0027] Furthermore, in step S2, the error characteristics include earth pressure error, which includes observation error and relative error:

[0028] ;

[0029] ;

[0030] In the formula, This indicates the earth pressure error at time step. The observation error, Indicates time step The actual value of earth pressure, Indicates time step Earth pressure setpoint, This indicates the earth pressure error at time step. The relative error;

[0031] The rate of change characteristics include the rate of change of shield attitude deviation, earth pressure, jack thrust, advance speed, excavated soil volume, and grouting volume. Each rate of change is characterized by its first-order difference.

[0032] ;

[0033] In the formula, Indicates at time step scalar features or vector features This indicates the value of the same feature at the previous time step. This indicates the amount of change of the feature between two adjacent steps;

[0034] The unit mileage characteristics include grouting volume mileage characteristics, excavated soil volume mileage characteristics, and advance volume mileage characteristics. The unit mileage characteristics are calculated by dividing the change by the mileage change value.

[0035] Furthermore, in step S3, the token encoding includes location encoding and mileage encoding;

[0036] When performing position encoding, sinusoidal position encoding is based on the time step sequence number;

[0037] When performing mileage encoding, the cut mileage is mapped to continuous location embeddings and concatenated with the location encoding.

[0038] Furthermore, in step S31, the shared module used by the encoder and decoder consists of several FreeTransformer blocks;

[0039] The multivariate observation vector at each time step is transformed into a representation vector consistent with the hidden dimension of the Free Transformer model through a linear mapping. The representation vector at each time step is then fused with its position and odometer codes and input into the shared module.

[0040] Furthermore, in step S31, the non-causal Free Transformer block adopts a non-causal self-attention mechanism, which does not use a causal mask when calculating attention weights, so that information from all positions in the token sequence can be accessed at each time step.

[0041] Furthermore, in step S32, a pooling operation is performed along the time dimension based on the hidden state sequence to obtain a global vector representing the current construction condition;

[0042] Based on the global vector, the mean and variance parameters of the latent variable Gaussian distribution are generated as distribution parameters through a learnable linear mapping.

[0043] Furthermore, in step S32, the pooling operation is average pooling or attention pooling.

[0044] Furthermore, in step S4, the decoding module of the decoder consists of several causal Free Transformer blocks. The decoder outputs the future representation sequence according to a preset number of time steps and synchronously generates the prediction target for each time step.

[0045] The shield attitude at future time steps is regressed from the future representation sequence through a multilayer perceptron or a linear layer;

[0046] Define the engineering threshold as either a plane deviation limit or an elevation deviation limit, and output the probability of exceeding the threshold in future time steps according to the following formula:

[0047] ;

[0048] In the formula, Indicates the future number The probability of exceeding the threshold risk in a step. Represents a probability operator. Indicates the future number The value corresponding to the predicted random variable / predicted distribution sample of the monitored target. Indicates the engineering threshold;

[0049] Recommended control parameters include earth pressure setpoint, advance speed, synchronous grouting pressure, and synchronous grouting volume.

[0050] Furthermore, it is recommended that the control output adopt the following supervised output:

[0051] ;

[0052] In the formula, Indicates the future Recommended control sequence for each step. Represents the set of control variables. Indicates the decoder output. and These represent the weights and biases of the control recommendation output head in the multi-task output head, respectively.

[0053] Due to the adoption of the above technical solution, the beneficial effects of this application are as follows:

[0054] 1. This application generates a global representation reflecting the current construction context by performing time-dimensional pooling on the latent state sequence output by the encoder. Based on this, the probability distribution of latent variables is inferred, and then the latent variables are injected into the decoding process to ultimately generate prediction results and correction schemes guided by the latent variables. Focusing on actual engineering, shield tunneling faces the problem of control failure due to unpredictable conditions. For example, ground stiffness, grout diffusion efficiency, and cutterhead disturbance range are all unpredictable conditions. These unpredictable conditions, as latent factors, significantly affect various aspects of shield tunneling. Latent variables, as low-dimensional vectors automatically learned by the model, do not directly correspond to any observed values, but they implicitly encode the unpredictable conditions during construction. Each dimension of the latent variable implicitly corresponds to the intensity or combination of a certain type of unobservable physical mechanism. In the decoding stage, the latent variables are broadcast to the representation of all future prediction time steps as the generation behavior of the conditional signal modulation decoder.

[0055] The reason why latent variables can characterize unpredictable working conditions is fundamentally because the phenomenon presented in shield tunneling data—that "the same control setting produces different responses under different working conditions"—means that the observed multiple variables are insufficient to determine the future evolution of the system. The actual process is governed by a set of implicit factors not directly recorded by sensors but continuously influencing the coupling relationship between earth pressure, propulsion, soil removal, grouting, and attitude. These factors manifest at the data level as "the same historical observations and setpoint inputs, yet subsequent attitude deviation trajectories still exhibit multiple reasonable directions." This ambiguity is a typical characteristic of "latently state-driven stochastic dynamic systems." The latent variable-enhanced Free Transformer learns a posterior distribution during the training phase through a variational inference mechanism, enabling latent variables to... It becomes a "compressed summary of all observation information within the current window" and is forced to interpret residual structures that cannot be uniquely explained by explicit input variables: when the model cannot simultaneously achieve both prediction accuracy and consistency based solely on explicit features, the optimization process will naturally drive latent variables to absorb and encode high-level factors that cause differences in system response, thereby statistically mapping different operating conditions to different latent variable values.

[0056] 2. In this application, because the decoder uses a non-causal Free Transformer block, it can generate multi-step prediction results in parallel, and the output of each time step directly depends on the same latent variable. The above design can avoid the problem of the progressive propagation and amplification of preceding prediction errors in autoregressive models, and significantly improve the trajectory consistency of multi-step predictions.

[0057] 3. In this application, latent variables are automatically learned by the encoder from historical observations, implicitly encoding unpredictable working conditions in the form of a probability distribution. When faced with construction scenarios not seen during training, the model can achieve adaptive response to new working conditions through sampling of the latent variable distribution, avoiding a sharp drop in performance due to sudden changes in working conditions. Attached Figure Description

[0058] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0059] Figure 1 The flowchart of the shield machine attitude multi-step prediction and correction control method based on latent variable enhancement Free Transformer provided in this application. Detailed Implementation

[0060] As described in the background technology, the attitude of a tunnel boring machine (TBM) is influenced by multiple coupled factors. Among them, the operational parameters such as grouting pressure and propulsion speed have significant time lag and cumulative effects on attitude regulation. More importantly, key environmental factors in actual construction (such as changes in stratum lithology, groundwater distribution, and soil disturbance in front of the cutterhead) cannot be directly observed by sensors, but they have a decisive impact on the TBM-soil interaction mechanism. In addition, mainstream methods mostly output deterministic point prediction values, lacking the quantification of the uncertainty of the prediction results, and cannot assess the risk of deviation exceeding limits. Moreover, the prediction module only provides future deviation estimates, while the correction strategy still relies on rule bases or manual intervention. Several integrated control suggestion methods often ignore the physical constraints of the equipment in their output operational parameters, resulting in a lack of constraint feasibility guarantees for the recommended instructions. Based on the above problems, this application discloses a multi-step prediction and correction control method for TBM attitude based on latent variable augmented Free Transformer, as shown in the appendix. Figure 1 As shown, it includes the following steps:

[0061] S1. Obtain shield tunneling process data, use the shield tunneling process data as elements to filter and establish a set of control quantities, and preprocess the process data in the set of control quantities.

[0062] Data collected from various sensors and operational records during tunnel boring machine (TBM) construction are organized and aligned according to "timestamps" and "tunneling mileage." The TBM construction process is indexed by time and mileage, including fields for earth pressure, propulsion, excavated soil, grouting, and grease.

[0063] The earth pressure field includes the earth pressure setpoint and the actual earth pressure value;

[0064] The propulsion fields include the total thrust of the jacks and the propulsion speed;

[0065] The waste soil field includes the amount of excavated soil;

[0066] The grouting field includes synchronous grouting pressure, synchronous grouting volume, secondary grouting pressure, secondary grouting volume, and grouting date;

[0067] The grease field includes the amount of grease used at the tail of the shield and the pressure of grease at the tail of the shield.

[0068] The system can output attitude results based on construction process data. The attitude result fields include segment attitude, shield attitude, cut elevation deviation, shield tail plane deviation, and shield tail elevation deviation.

[0069] Preferably, the set of control quantities must include the earth pressure setpoint, advance speed, synchronous grouting pressure, and synchronous grouting volume.

[0070] When training the model, it is necessary to define a set of prediction targets, which should include at least the shield attitude and the segment attitude.

[0071] Data preprocessing includes data alignment, cleaning, and unit standardization. Specifically, multi-source fields are aligned using advancement time or cut mileage as the primary key; when different sampling frequencies exist, interpolation / hold / resampling methods are used to obtain a unified step size sequence. For missing values, linear interpolation or forward filling is used for gaps; long gaps are removed from the corresponding segments. For outliers, truncation or removal is performed based on physical ranges (e.g., reasonable ranges for pressure, velocity, and thrust) or statistical thresholds; if the time is disordered, it is sorted by advancement time, and if there is a bounce, it is re-sorted or removed by mileage. Pressure, thrust, velocity, excavated soil, grouting volume, etc., are unified into a fixed unit system, and the date field is timestamped to complete unit standardization.

[0072] S2. Based on the preprocessed construction process data, construct error characteristics, rate of change characteristics, and unit mileage characteristics, and normalize the processed construction process data and construction characteristics to obtain a normalized feature sequence.

[0073] Error characteristics include earth pressure error, which includes earth pressure observation error and relative error. The formula for constructing error characteristics is as follows. If other characteristics of the shield tunneling system can be observed, the formula principle for earth pressure error is still referred to, that is, the actual value is subtracted from the set value:

[0074] ;

[0075] ;

[0076] In the formula, This indicates the earth pressure error at time step. The observation error, Indicates time step The actual value of earth pressure, Indicates time step Earth pressure setpoint, This indicates the earth pressure error at time step. The relative error;

[0077] The rate of change characteristics include the rate of change of shield attitude deviation, earth pressure, jack thrust, advance speed, excavated soil volume, and grouting volume. Each rate of change is characterized by its first-order difference.

[0078] ;

[0079] In the formula, Indicates at time step Scalar or vector characteristics, such as actual earth pressure, total thrust, advance speed, and shield attitude deviation. This indicates the value of the same feature at the previous time step. This indicates the amount of change in the feature between two adjacent steps, used to characterize trends or abrupt changes.

[0080] The unit mileage characteristics include grouting volume mileage characteristics, excavated soil volume mileage characteristics, and advance volume mileage characteristics. The unit mileage characteristic is calculated by dividing the change in volume by the change in mileage. The unit mileage characteristic is used to mitigate the influence of sampling frequency variations. Advance volume mileage is acquired through sensors.

[0081] In practice, the process data consists of continuous features, which are normalized using z-scores. The normalization parameters are obtained solely from the training set. During training, input samples need to be constructed using a sliding window approach, with the historical window length and prediction compensation set. The input samples consist of the process data and the derived features from step S2 within the historical window. The prediction target set is defined as described above.

[0082] S3. Organize the normalized feature sequence into a token sequence and encode each token. Input the encoded token sequence into the pre-trained Free Transformer model. The token encoding includes location encoding and mileage encoding. In practice, each token is defined as a multi-dimensional feature vector composed of construction process data and structural features.

[0083] When performing position encoding, sinusoidal position encoding is based on the time step sequence number;

[0084] During odometer encoding, the cut-off odometer is mapped to a continuous location embedding and concatenated with the location code. The cut-off odometer is read by the system.

[0085] The Free Transformer model is configured to: generate latent variables based on the token sequence and inject the latent variables into the decoder;

[0086] The specific steps of the latent variable injection decoder are as follows:

[0087] S31. Construct at least one non-causal Free Transformer block as an encoder-specific block. This encoder-specific block aggregates the output of the shared module used by the encoder and decoder for full-sequence information, that is, aggregates all tokens in the historical time window.

[0088] The token at each time step is converted into a representation vector with the same hidden dimension as the Free Transformer model through a linear mapping. The representation vector of each time step is then fused with its position code and odometry code and input into the shared module, i.e.:

[0089] ;

[0090] In the formula, Indicates the first The model input embedding representation of the step (hidden state of layer 0). Indicates the first The input feature vector for each step is composed of process data and pose history. This represents the linear mapping weight matrix of the embedding layer. This represents the embedding layer bias vector.

[0091] After being fused with location and odometer codes, the result is .

[0092] The shared module serves both the encoder and decoder. As the first half of the Free Transformer architecture, the shared module consists of several Free Transformer blocks:

[0093] ;

[0094] in, t represents the current step, and L represents the length of the history window.

[0095] In the formula, Indicates a shared module. This represents the hidden state sequence output by the shared module.

[0096] To infer latent variables, at least one non-causal Free Transformer block is constructed as an encoder-specific block:

[0097] ;

[0098] In the formula, Indicates a dedicated block for encoders. This represents the implicit representation of the sequence output by the encoder, which is used for subsequent latent variable distribution parameterization.

[0099] The noncausal Free Transformer block employs a noncausal self-attention mechanism, which does not use a causal mask when calculating attention weights, so that each time step can access information from all positions in the token sequence.

[0100] S32. Calculate the latent variable distribution parameters based on the latent state sequence output by the encoder, and sample the latent variables according to the latent variable distribution parameters:

[0101] ;

[0102] ;

[0103] In the formula, This represents the mean vector of the posterior distribution of the latent variables. Indicates output from encoder The global representation vector obtained after pooling; and Represented by global representation Mapping to the mean The linear layer weights and biases, This represents the variance vector of the posterior distribution of the latent variable. and Represented by global representation Mapped to The linear layer weights and biases, Let z represent the standard deviation vector, and z represent the latent variable vector. Element-wise multiplication This represents the standard normal noise vector. Let I represent the normal distribution, and let I represent the identity matrix.

[0104] Pooling is performed along the time dimension based on the hidden state sequence to obtain a global vector representing the current construction condition;

[0105] Based on the global vector, the mean and variance parameters of the latent variable Gaussian distribution are generated as distribution parameters through a learnable linear mapping.

[0106] In step S32, the pooling operation is either average pooling or attention pooling.

[0107] S33. Obtain a condition vector with the same dimension as the FreeTransformer model through linear mapping, and broadcast the condition vector to the decoding path. Specifically, inject latent variables into the intermediate or specified layers of the decoding path. Condition vector:

[0108] ;

[0109] In the formula, This represents the latent variable representation after mapping, i.e., the condition vector. Indicates will Mapped to The linear transformation weight matrix, This represents the bias vector of the linear transformation.

[0110] Broadcast to decoding path:

[0111] ;

[0112] In the formula, This indicates that the shared module outputs a sequence representation. This represents the shared output representation after the latent variables are injected. Indicates the vector Broadcast to The same sequence length dimension allows it to be added element-wise to the representation at each time step.

[0113] After injection, the model can first determine the high-level operating conditions and strategies when generating future predictions / recommendations, and then make detailed predictions.

[0114] S4. The multi-task output head of the Free Transformer model based on injected latent variables outputs the predicted shield tunnel segment attitude vector, the over-threshold probability, and the recommended control quantity, respectively.

[0115] In step S4, the decoder's decoding module consists of several causal Free Transformer blocks. For multi-step prediction, the decoder outputs a future representation sequence according to a preset number of time steps and simultaneously generates the prediction target for each time step. The multi-task output head contains three outputs: attitude prediction, risk probability, and control recommendation. The decoder's raw output values ​​are input into two MLPs with different weights to obtain the attitude prediction output and the control recommendation output, respectively. After obtaining N sets of attitude prediction outputs, their out-of-limit risk is calculated to obtain the out-of-threshold probability output.

[0116] The shield attitude at future time steps is regressed using a multilayer perceptron or a linear layer to represent the future sequence.

[0117] Define the engineering threshold as either a plane deviation limit or an elevation deviation limit, and output the probability of exceeding the threshold in future time steps according to the following formula:

[0118] ;

[0119] In the formula, Indicates the future number Step (corresponding time) The probability of exceeding the threshold risk. Represents a probability operator. Indicates the future number The value corresponding to the predicted random variable / predicted distribution sample of the step monitoring target (e.g., a certain attitude deviation component). Indicates the engineering threshold;

[0120] Recommended control parameters include earth pressure setpoint, advance speed, synchronous grouting pressure, and synchronous grouting volume.

[0121] The recommended control output uses the following supervised output:

[0122] ;

[0123] In the formula, Indicates the future Recommended control sequence for each step. Represents the set of control variables. Indicates the decoder output. and These represent the weights and biases of the control recommendation output head in the multi-task output head, respectively.

[0124] Minimize the following loss during the training phase:

[0125] ;

[0126] In the formula, The posterior distribution of the encoder output. As a prior distribution, Let KL divergence be the KL divergence. As weight, For mission losses, For constraint and smoothing regularization.

[0127] The task loss includes attitude regression loss, risk classification loss, and control recommendation loss. Constraints and smoothing regularization introduce penalty terms on the boundary, rate of change, and smoothness of the recommended control quantity to ensure that the output is executable and does not fluctuate frequently.

[0128] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0129] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable augmented Free Transformer, characterized in that, Includes the following steps: S1. Obtain shield tunneling construction process data; the shield tunneling construction process is indexed by time or mileage, including fields such as earth pressure, propulsion, excavated soil, grouting, and grease. in, The earth pressure field includes the earth pressure setpoint and the actual earth pressure value; The propulsion fields include the total thrust of the jacks and the propulsion speed; The waste soil field includes the amount of excavated soil; The grouting field includes synchronous grouting pressure, synchronous grouting volume, secondary grouting pressure, secondary grouting volume, and grouting date; The grease field includes the amount of grease used at the tail of the shield and the pressure of grease at the tail of the shield; S2. Based on the preprocessed construction process data, construct error characteristics, rate of change characteristics, and unit mileage characteristics, and normalize the preprocessed construction process data and construction characteristics to obtain a normalized feature sequence. S3. Organize the normalized feature sequence into a token sequence and encode each token. Input the encoded token sequence into the pre-trained Free Transformer model. The Free Transformer model is configured to: generate latent variables based on the token sequence and inject the latent variables into the decoder; The specific steps of the latent variable injection decoder are as follows: S31. Construct at least one non-causal Free Transformer block as an encoder-specific block. This encoder-specific block aggregates the output of the shared module used by the encoder and decoder for full-sequence information aggregation. S32. Calculate the latent variable distribution parameters based on the latent state sequence output by the encoder, and obtain the latent variables by sampling according to the latent variable distribution parameters; S33. Obtain a condition vector with the same dimension as the Free Transformer model through linear mapping, and broadcast the condition vector to the decoding path; S4. The multi-task output head of the Free Transformer model based on injected latent variables outputs the predicted shield attitude vector, the over-threshold probability, and the recommended control quantity, respectively.

2. The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable enhancement Free Transformer as described in claim 1, characterized in that, In step S2, the error characteristics include earth pressure error, which includes observation error and relative error: ； ； In the formula, This indicates the earth pressure error at time step. The observation error, Indicates time step The actual value of earth pressure, Indicates time step Earth pressure setpoint, This indicates the earth pressure error at time step. The relative error; The rate of change characteristics include the rate of change of shield attitude deviation, earth pressure, jack thrust, advance speed, excavated soil volume, and grouting volume. Each rate of change is characterized by its first-order difference. ； In the formula, Indicates at time step scalar features or vector features This indicates the value of the same feature at the previous time step. This indicates the amount of change of the feature between two adjacent steps; The unit mileage characteristics include grouting volume mileage characteristics, excavated soil volume mileage characteristics, and advance volume mileage characteristics. The unit mileage characteristics are calculated by dividing the change by the mileage change value.

3. The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable enhancement Free Transformer as described in claim 2, is characterized in that, In step S3, the token encoding includes location encoding and mileage encoding; When performing position encoding, sinusoidal position encoding is based on the time step sequence number; When performing mileage encoding, the cut mileage is mapped to continuous location embeddings and concatenated with the location encoding.

4. The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable enhancement Free Transformer as described in claim 3, is characterized in that, In step S31, the shared module used by the encoder and decoder consists of several Free Transformer blocks; The multivariate observation vector at each time step is transformed into a representation vector consistent with the hidden dimension of the Free Transformer model through a linear mapping. The representation vector at each time step is then fused with its position and odometer codes and input into the shared module.

5. The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable enhancement Free Transformer according to claim 4, characterized in that, In step S31, the non-causal Free Transformer block adopts a non-causal self-attention mechanism, which does not use a causal mask when calculating attention weights, so that information from all positions in the token sequence can be accessed at each time step.

6. The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable enhancement Free Transformer according to claim 5, characterized in that, In step S32, a pooling operation is performed along the time dimension based on the hidden state sequence to obtain a global vector representing the current construction condition; Based on the global vector, the mean and variance parameters of the latent variable Gaussian distribution are generated as distribution parameters through a learnable linear mapping.

7. The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable enhancement Free Transformer according to claim 6, characterized in that, In step S32, the pooling operation is either average pooling or attention pooling.

8. The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable enhancement Free Transformer according to claim 7, characterized in that, In step S4, the decoding module of the decoder consists of several causal Free Transformer blocks. The decoder outputs the future representation sequence according to a preset number of time steps and synchronously generates the prediction target for each time step. The shield attitude at future time steps is regressed using a multilayer perceptron or a linear layer to represent the future sequence. Define the engineering threshold as either a plane deviation limit or an elevation deviation limit, and output the probability of exceeding the threshold in future time steps according to the following formula: ； In the formula, Indicates the future number The probability of exceeding the threshold risk in a step. Represents a probability operator. Indicates the future number The value corresponding to the predicted random variable / predicted distribution sample of the monitored target. Indicates the engineering threshold; Recommended control parameters include earth pressure setpoint, advance speed, synchronous grouting pressure, and synchronous grouting volume.

9. The method for multi-step prediction and correction control of tunnel boring machine attitude based on latent variable enhancement Free Transformer according to claim 8, characterized in that, The recommended control output uses the following supervised output: ； In the formula, Indicates the future Recommended control sequence for each step. Represents the set of control variables. Indicates the decoder output. and These represent the weights and biases of the control recommendation output head in the multi-task output head, respectively.

Images