Shield tunneling machine tail grease pressure prediction method and system based on causal comparison model

Abstract

The invention discloses a shield tunneling machine tail grease pressure prediction method and system based on a causal comparison model, and relates to the technical field of shield tunneling machines, and the method comprises the steps: inputting a shield tail posture parameter into a trained causal comparison model, and outputting a grease pressure prediction value; the training process of the causal comparison model is as follows: splicing and segmenting shield tail attitude parameters and grease pressure, and converting each segment into a high-dimensional feature vector; inputting the high-dimensional feature vector into a directional influence device to obtain a causal relationship feature; inputting grease pressure into a trend extractor, independently extracting trend characteristics of grease pressure historical data through a multi-layer perceptron, and outputting pressure trend characteristics; constructing a total loss function based on the causal comparison loss and the mean absolute error loss so as to adjust trainable parameters in the causal comparison model; according to the shield tunneling machine tail grease pressure prediction method and system, the grease pressure can be predicted more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of shield machines, and particularly to a method and system for predicting the grease pressure at the tail of a shield machine based on a causal contrast model. Background Art

[0002] The shield tunneling method for underground excavation depends on a large shield machine for excavation. During the tunneling process of the shield machine, a certain gap will be formed between the tail shield and the tunnel segment, connecting the tunnel soil layer with a certain pressure to the safe space inside the shield body. Therefore, a sealing device is required to isolate these two areas, which is the tail shield sealing system. With the increasing complexity of the hydrogeological conditions in shield construction, the risk of tail shield seal failure has increased significantly. Among them, the imbalance of pressure inside and outside the sealing cavity is an important reason for the failure of the tail shield seal.

[0003] In order to effectively maintain the pressure balance inside the sealing cavity, early research mainly indirectly evaluated the grease pressure inside the cavity by predicting the grease consumption, so as to adjust the pressure balance by regularly injecting grease. Specifically, a model was constructed by analyzing the correlation between tunnel parameters, geological conditions, and grease consumption. Then, combined with operation experience, the grease injection volume was quantitatively adjusted to maintain the pressure balance inside and outside the sealing cavity. However, due to its indirectness, this method always has certain errors and uncertainties.

[0004] Nowadays, with the development of sensor technology, sensors can be directly installed inside the sealing cavity to accurately measure the grease pressure. Therefore, it becomes possible to directly predict the grease pressure. However, due to the directional and sparse causal relationship between the tail shield attitude parameters and the grease pressure, the general time series prediction algorithms have poor effects. Summary of the Invention

[0005] Based on the technical problems existing in the background art, the present invention proposes a method and system for predicting the grease pressure at the tail of a shield machine based on a causal contrast model, which can more accurately predict the grease pressure.

[0006] The method and system for predicting the grease pressure at the tail of a shield machine based on a causal contrast model proposed by the present invention

[0007] The advantages of the method and system for predicting the grease pressure at the tail of a shield machine based on a causal contrast model provided by the present invention are as follows: effectively modeling the directional influence relationship between the tail shield attitude parameters and the grease pressure, and having a strong memory ability for the sparsely distributed relationship, which can more accurately predict the grease pressure. Brief Description of the Drawings

[0008] Figure 1 is a schematic structural diagram of the present invention;

[0009] Figure 2Structural schematic diagram of causal contrast loss. Detailed implementation mode

[0010] Next, the technical solution of the present invention will be described in detail through specific embodiments. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific implementations disclosed below.

[0011] As Figure 1 and 2 shown, for the shield tail grease pressure prediction method based on the causal contrast model proposed by the present invention, the shield tail attitude parameters are input into the trained causal contrast model, and the grease pressure prediction value is output;

[0012] The training process of the causal contrast model is as follows:

[0013] Step 1: Concatenate and segment the shield tail attitude parameters and grease pressure, convert each segment into a high-dimensional feature vector, and retain the time sequence information;

[0014] Among them, the shield tail attitude parameter X c includes the shield tail attitude, mud pressure, and grease injection pressure. Concatenate X c and the grease pressure variable X g into a vector X e according to the attribute dimension;

[0015] Fill the vector X e into the vector X p to ensure that the input sequence can be evenly divided by the length L seg , and split each variable into segments with a length of L seg . The specific formula is as follows:

[0016]

[0017] Among them, Reshape(X p ) is the length of X p , x d,i is the d-th shield tail attitude parameter of the i-th length, D is the total number of shield tail attitude parameters, where L is the length of the input sequence after padding, and N = L / L seg represents the total number of segments.

[0018] Embed each segment into a vector through linear projection plus position embedding to obtain a high-dimensional feature vector H that retains the time sequence information. The specific formula is shown as follows:

[0019]

[0020] Among them, represents a learnable projection matrix, represents a learnable position embedding matrix, H is an array of two-dimensional vectors, and each h in it i,d represents the embedding vector of a single-variable time segment, and d model represents the dimension of the embedding matrix, and the parameter is adjustable.

[0021] Step 2: Input the high-dimensional feature vector into the directional influencer, extract the time-dimensional features of each high-dimensional feature variable through the multi-head self-attention and the feed-forward network, use the grease pressure variable in the time-dimensional features as the query, use the shield machine parameters in the time-dimensional features as the key and value, and directionally capture the potential representation of the unidirectional influence of the shield tail attitude parameters on the grease pressure through the causal attention mechanism to obtain the causal relationship features;

[0022] Use the multi-head self-attention mechanism (MSA) and the feed-forward network to further extract the features of each high-dimensional feature variable regarding the time dimension. The formula is as follows:

[0023]

[0024] Among them, LN(.) is the layer normalization widely used in the time series prediction method, is the feature after normalizing the high-dimensional feature vector H, MLP represents the multi-layer feed-forward network, and MSA(Q, K, V) represents the multi-head self-attention mechanism, where Q, K, and V are used as the query, key, and value respectively, is the feature obtained after being processed by the self-attention mechanism, and Z time is the feature is the time-dimensional feature obtained after passing through the feed-forward network.

[0025] Use the causal attention mechanism to take into account the influence of the shield tail attitude parameters on the grease pressure. Specifically, use the grease pressure variable time in the time-dimensional feature Z as the query, as the key and value to directionally capture the influence of the shield tail attitude parameters on the grease pressure. The formula is as follows:

[0026]

[0027] Among them, MSA dim is the multi-head self-attention mechanism, with as the query, as the key and query, is the feature obtained after passing through the multi-head self-attention mechanism and layer normalization, and Z dimis the latent representation obtained after passing through a multi-layer feedforward network.

[0028] Step 3: Input the grease pressure into the trend extractor, and independently extract the trend features of the historical grease pressure data through a multi-layer perceptron, and output the pressure trend features;

[0029] Only input the grease pressure variable into the trend extractor, extract the trend features through the MLP, and obtain the pressure trend feature Y M , and the specific formula is as follows:

[0030]

[0031] where h is the prediction duration.

[0032] Among them, the MLP structure contains 2 fully connected layers, the activation function is ReLU, the output layer has no activation function, and finally the pressure trend features are output.

[0033] Step 4: Construct a total loss function based on the causal contrast loss and the mean absolute error loss to adjust the trainable parameters in the causal contrast model.

[0034] Regarding the causal contrast loss, specifically:

[0035] The vector X obtained by splicing the original shield tail attitude parameters and the grease pressure e is input into the contrast loss module and unfolded after normalization. Then, update the queue composed of dictionaries, store the current batch vector X i in it, and at the same time remove the batch that entered the queue earliest, so as to obtain the queue X after the first update q . After multiplying the current batch by the queue after the first update, obtaining the scores, find the position of the causal relationship that is most similar to the current batch except itself (that is, take the queue position corresponding to the highest similarity as the position of the causal relationship that is most similar to the current batch except itself), and the formula is as follows:

[0036]

[0037] Pos_mask = FP(Max(Sim(X i , X q ) × Self_mask(X i )));

[0038] where X i is the unfolded vector, Norm(.) is the normalization method, and Flatten(.) is to flatten the vector X eA one-dimensional vector unfolded along the time axis, B is the size of the batch, P = (D + 1) × l, D is the total number of input shield tail attitude parameters, l is the time step of the predicted grease pressure, Sim(.) is to calculate the similarity score between two vectors, FP(.) is to find the position of the current maximum similarity score, Self_mask is the self-mask, X q is the queue after the first update, and Pos_mask is the mask for positive sample pairs.

[0039] After obtaining the mask for positive pairs by finding the position of the most similar causal relationship, the time dimension features (latent feature representations) learned after performing feature extraction on the shield tail attitude parameters with respect to the time dimension and the latent representation Z representing the learned influence relationship after causal attention on the shield tail attitude parameters and grease pressure dim are concatenated to obtain the feature Z a Similarly, the feature Z will also be a unfolded after normalization to obtain the feature Z i The specific formula is as follows:

[0040]

[0041] where Concat is concatenation, Flatten is normalization, B is the size of the batch, and S is M × N × d model where N is the total number of segments into which each variable is split, and d model represents the dimension of the embedding matrix, and M is the total number of independent variables in the shield tail attitude parameters.

[0042] Using the feature Z i to update the queue to obtain the queue after the second update, calculate the dot product between the current batch and the queue after the second update to get a score, which is used to quantify the similarity of causal relationships so that the causal contrast model can push away negative samples and pull in positive samples. Then, mark the positive and negative pairs according to the positive sample mask Pos_mask and the self-mask, and then use the original input vector X i and the score calculated between it and the queue X q after its first update as weights to distinguish positive pairs with different similarities. Calculate the causal contrast loss L CC The formula is as follows:

[0043]

[0044] where is the latent representation of the causal relationship most similar to the current causal relationship, Z q is the updated queue containing the latent representation τ is the temperature parameter used to control the shape of the loss function, N is the number of negative sample pairs, and k is the index of the negative sample pair.

[0045] Therefore, the total loss function is as follows:

[0046] Add the causal relationship feature Y output by the directional effector H and the pressure trend feature Y output by the trend extractor M to obtain the feature X f , and the formula is as follows:

[0047]

[0048] X f = Y H + Y T ;

[0049] where MLP is a multi-layer perceptron, Flatten is normalization, and h is the prediction duration.

[0050] Calculate the mean absolute error loss L MAE :

[0051] L MAE = |X label - X f |;

[0052] where X label is the time series label of the grease pressure to be predicted.

[0053] Combine the mean absolute error loss L MAE and the causal contrast loss L CC to construct the total loss function L:

[0054] L = L MAE + λL CC ;

[0055] where λ is the balance coefficient.

[0056] In this embodiment, through steps one to four, the method for predicting the grease pressure at the tail of the shield machine effectively models the causal relationship between the shield tail attitude parameters and the grease pressure. When the shield tail attitude changes, it can find similar states a long time ago, so as to accurately calculate the specific impact on the grease pressure (slow increase, sudden drop, etc.), and can more accurately predict the grease pressure.

[0057] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered within the protection scope of the present invention.

Claims

1. Shield machine tail grease pressure prediction method based on causal contrast model, characterized in that, Input the shield tail attitude parameters into the trained causal comparison model and output the predicted grease pressure value; The training process of the causal contrast model is as follows: The shield tail attitude parameters and grease pressure are concatenated and segmented, and each segment is converted into a high-dimensional feature vector; The high-dimensional feature vector is input into the directional influencer. The time dimension features of each high-dimensional feature variable are extracted through multi-head self-attention and feedforward network. The grease pressure variable in the time dimension features is used as the query, and the shield tail posture parameters in the time dimension features are used as the key and value. The causal attention mechanism is used to capture the potential representation of the unidirectional influence of the shield tail posture parameters on the grease pressure, and obtain the causal relationship features. The grease pressure is input into the trend extractor, and the trend features of the grease pressure historical data are independently extracted through the multi-layer perceptron, and the pressure trend features are output; A total loss function is constructed based on the causal contrastive loss and the mean absolute error loss to adjust the trainable parameters in the causal contrastive model.

2. The shield machine tail grease pressure prediction method based on the causal contrast model according to claim 1, wherein The shield tail attitude parameters and grease pressure are spliced and segmented, and each segment is converted into a high-dimensional feature vector, specifically: The shield tail attitude parameters and the grease pressure variable are concatenated into a vector X according to the attribute dimension. e The shield tail attitude parameters include the shield tail attitude, the mud pressure, and the grease injection pressure. Pad the vector X e to form the vector X p such that the input sequence can be evenly divided by the length L seg and split each variable into segments of length L seg ; Each segment is embedded into a vector through linear projection plus position embedding to obtain a high-dimensional feature vector that retains the temporal order information.

3. The shield machine tail grease pressure prediction method based on the causal contrast model according to claim 1, characterized in that, In constructing a total loss function based on the causal contrastive loss and the mean absolute error loss to adjust the trainable parameters in the causal contrastive model, the causal contrastive loss is: The queue dictionary mechanism is used to find the positive sample pairs with the most similar causal relationship with the current sample in the historical data. The feature distance of the positive sample pairs is shortened and the negative sample pairs are pushed away through the comparison loss function, thereby strengthening the causal comparison model's ability to remember sparse causal relationships.

4. The method for predicting the grease pressure at the tail of a shield machine based on a causal contrast model according to claim 3, wherein By using the queue dictionary mechanism to find the positive sample pairs with the most similar causal relationship with the current sample in the historical data, the feature distance of the positive sample pairs is shortened by the comparison loss function, and the feature distance of the negative sample pairs is pushed away. Specifically: The vector X obtained by splicing the original shield tail attitude parameters and the grease pressure e is input into the contrastive loss module and expanded after normalization to obtain vector X i ; Update the queue composed of dictionaries, and store the current batch vector X i into it, while removing the earliest batch that entered the queue; Through the current batch vector X i and the queue X after the initial update q perform a dot product to obtain a score, and calculate the similarity in combination with the self-mask of vector X i The position of the queue corresponding to the highest similarity is used as the position of the causal relationship that is most similar to the others in the current batch except itself, and the mask of the positive sample pair is obtained; Concatenate the time - dimension features of each high - dimensional feature variable and the potential representation of the influence of the shield - tail catching posture parameters on the grease pressure unidirectionally to obtain feature Z a , and normalize feature Z a and then expand it to obtain feature Z i ; Utilize feature Z i After updating the queue, calculate the dot product between the current batch and the queue after the secondary update to obtain a score; Label the positive and negative pairs according to the positive sample mask and self-mask; Constructing causal contrastive loss.

5. The shield machine tail grease pressure prediction method based on the causal contrast model according to claim 4, characterized in that After marking positive and negative pairs according to the positive sample mask and the self-mask, the original vector X i and the queue X q after the first update are used as weights to distinguish positive pairs with different similarities.

6. The shield machine tail grease pressure prediction method based on the causal contrast model according to claim 4, characterized in that Causal contrastive loss L CC It is calculated as follows: Among them, is the potential representation of the causal relationship most similar to the current causal relationship, Z q is the update queue containing the potential representation , τ is the temperature parameter used to control the shape of the loss function, N is the number of negative sample pairs, and k is the negative sample pair index.

7. The method for predicting the grease pressure at the tail of a shield machine based on a causal contrast model according to claim 1, wherein The absolute error loss is as follows: The causal relationship feature output by the directional influencer is added to the pressure trend feature output by the trend extractor to obtain the final grease pressure prediction value; The absolute error loss is constructed using the time series label of the grease pressure to be predicted and the final grease pressure prediction value.

8. The shield machine tail grease pressure prediction method based on the causal contrast model according to claim 1, characterized in that The total loss function L is as follows: L = L MAE + λL CC ; Among them, L MAE is the mean absolute error loss, L CC is the causal contrast loss, and λ is the balance coefficient.

9. A shield machine tail grease pressure prediction system based on a causal contrast model, characterized in that, Input the shield tail attitude parameters into the trained causal comparison model and output the predicted grease pressure value; The training process of the causal contrastive model includes a splicing transformation module, a directional influencer, a trend extractor, and a loss construction module: The splicing and conversion module is used to splice and segment the shield tail attitude parameters and grease pressure, and convert each segment into a high-dimensional feature vector; The directional influencer takes a high-dimensional feature vector as input and extracts the time dimension features of each high-dimensional feature variable through multi-head self-attention and feedforward networks. The grease pressure variable in the time dimension features is used as the query, and the shield tail posture parameters in the time dimension features are used as the key and value. The causal attention mechanism is used to capture the potential representation of the unidirectional influence of the shield tail posture parameters on the grease pressure, thereby obtaining the causal relationship features. The trend extractor takes the historical data of grease pressure as input, independently extracts the trend features of the historical data of grease pressure through a multi-layer perceptron, and outputs the pressure trend features; The loss building module is used to construct a total loss function based on the causal contrastive loss and the mean absolute error loss to adjust the trainable parameters in the causal contrastive model.

10. The shield machine tail grease pressure prediction system based on the causal contrast model according to claim 9, characterized in that The loss building blocks include the causal contrast building block; The causal contrast loss construction module uses the queue dictionary mechanism to find positive sample pairs with the most similar causal relationship with the current sample in historical data. Through the contrast loss function, it shortens the feature distance of the positive sample pairs and pushes away the negative sample pairs, thereby enhancing the causal contrast model's ability to remember sparse causal relationships.

Images