Shield attitude prediction method based on multi-attention mechanism optimization combination model

Abstract

The invention provides a shield attitude prediction method based on a multi-attention mechanism optimization combination model, and the method comprises the following steps: obtaining historical feature data and shield attitude data of a target shield tunneling machine, carrying out missing value completion, tunneling section data extraction and data normalization on the historical feature data and the shield attitude data, and analyzing the correlation degree of the historical feature data and the shield attitude data through gray correlation; the input feature attention is extracted by selecting a kernel network; establishing a bidirectional long-short-term memory network, and predicting the implicit output containing the long-term trend; establishing a gating circulation unit network, and predicting implicit output containing a short-term fluctuation trend; and two groups of prediction network implicit outputs are spliced, output attention is extracted by using a multi-head attention mechanism, a dependency relationship between implicit features output by the combined model and the shield attitude is captured, and mapping from the implicit outputs to the final prediction shield attitude is realized. According to the method, the prediction precision and the prediction efficiency of the shield attitude prediction model are improved, fine control over the attitude in the shield construction stage is facilitated, and the shield construction efficiency and the construction safety are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of shield construction, and in particular to a shield posture prediction method based on a multi-attention mechanism optimized combination model. Background Art

[0002] Shield tunneling is widely used in tunnel construction for subway, municipal engineering, railway, highway, water conservancy, mining, military and other fields due to its safety, high efficiency, economy and low impact on the external environment. With the rapid development of urban construction and shield tunneling technology, my country's shield tunnel construction is gradually moving towards greater depth and larger cross-section.

[0003] However, as the tunnel section increases, the complexity of geological conditions brings uncertainty and risks to tunnel excavation. In addition, factors such as shallow burial depth and high water pressure make the shield machine more demanding in terms of shield attitude control during construction. In addition, due to the increase in the inertia of the shield system, the change of the shield attitude is also more complicated. The traditional method of adjusting the shield machine excavation parameters by manually observing the real-time attitude feedback of the shield machine often results in poor attitude adjustment such as attitude correction lag and serpentine trajectory. Therefore, in-depth research on the prediction method of shield attitude and auxiliary fine control of shield construction attitude are of great engineering significance for the safe and efficient construction of shield machines. Summary of the invention

[0004] The purpose of the present invention is to provide a shield attitude prediction method based on a multi-attention mechanism optimization combination model to improve the prediction accuracy of the shield attitude and achieve scientific adjustment and control optimization of the shield tunneling attitude.

[0005] In order to solve the above technical problems, the present invention provides a shield posture prediction method based on a multi-attention mechanism optimization combination model, comprising the following steps:

[0006] S1, obtaining and preprocessing the historical feature data of the tunneling section and the shield posture data of the target shield machine, and then using the preprocessed historical feature data as a comparison feature column and the preprocessed shield posture data as a reference column to obtain the correlation between each historical feature data and the shield posture data;

[0007] S2, arranging the correlation degree of each historical feature data in descending order, setting the feature selection dimension, and taking the feature column with the highest correlation degree as the feature data after feature selection according to the set feature selection dimension;

[0008] S3, processing the feature data after feature selection to generate a feature time series data set as the input data of the deep learning model, and dividing the feature time series data into a training set and a test set according to a certain ratio;

[0009] S4, by selecting the kernel unit network, feature attention calculation is performed on the input time series data, and implicit feature mapping with attention is adaptively formed;

[0010] S5. Construct a bidirectional long short-term memory network to predict the long-term trend implicit output of the shield posture: use the implicit feature map with attention as the input data of the bidirectional long short-term memory network through the fully connected layer, and iterate the network parameters according to the loss function model of the bidirectional long short-term memory network to obtain the optimal network parameters;

[0011] S6. Construct a gated recurrent unit network to predict the implicit output of the short-term fluctuation trend of the shield posture: use the implicit feature map with attention in step S4 as the input data of the gated recurrent network through the fully connected layer mapping, and iterate the network parameters according to the network loss function model to obtain the optimal network parameters;

[0012] S7. Combine the long-term trend implicit output and the short-term fluctuation trend implicit output, and adaptively capture the dependency between the long-term trend implicit output and the short-term fluctuation trend implicit output and the model output on the shield posture through the multi-head attention mechanism, and finally obtain the shield posture prediction result.

[0013] In the preferred solution, the preprocessing in step S1 includes missing value filling, excavation section data extraction and data normalization;

[0014] The correlation between each historical feature data and shield posture data is obtained through grey correlation analysis;

[0015] The feature data after feature selection in step S3 is subjected to time window processing according to the selected time series length.

[0016] In a preferred solution, the historical feature data in step S1 is:

[0017] X i (k),i∈[1,n],k∈[1,m],

[0018] Where X i (k) is the kth point data in the historical feature data after the i-th preprocessing, and n and m are the type and time point of the collected historical feature data respectively;

[0019] The posture data after preprocessing in step S1 is:

[0020] Y o (k),o∈[1,4],k∈[1,m],

[0021] Among them, Y o(k) represents the posture data at time point k after preprocessing, where o represents different types of posture data: head vertical posture, head horizontal posture, tail vertical posture and head horizontal posture, and m is the number of time points of the collected historical feature data;

[0022] The data of the historical feature data at the kth time point is:

[0023] {X1(k),X1(k),…,X n (k)}

[0024] In step 1, the grey correlation analysis method is used to obtain the correlation between each historical feature data and the shield posture data. Specifically:

[0025] Calculate the grey correlation coefficient:

[0026]

[0027] i∈[1,n],k∈[1,m],o∈[1,4];

[0028] Among them, Y o (k) represents the shield posture data of the oth type after preprocessing at the kth time point, X i (k) represents the historical feature data of the i-th category after preprocessing at the k-th time point, n is the type of historical feature data collected, m is the number of time points for collecting data, ρ is the resolution coefficient, and its value range is (0,∞), ξ o,i (k) represents the grey correlation coefficient between the i-th type of preprocessed historical feature data and the o-th type of preprocessed shield posture data at the kth time point.

[0029] Calculate the correlation:

[0030]

[0031] where γ i is the correlation between the historical characteristic data of the i-th shield machine and the shield machine posture data, γ i The closer it is to 1, the higher the correlation between this historical feature data and the shield posture data.

[0032] In the preferred solution, in step S3, the historical feature data after time window processing is as follows:

[0033] x(k)={[X1(k-T+1),...,X c (k-T+1)];...;[X1(k),...,X c (k)]};

[0034] Where x(k) is the network model input time window data at time point k, T is the time window time point length, and c is the number of feature data types selected after grey correlation degree sorting.

[0035] In the preferred solution, the attention mechanism calculation for the time series based on the selected kernel network in step S4 is specifically performed as follows:

[0036] S41, splitting process: taking the time window data, i.e., the multi-dimensional time series features, as input, and passing through two groups of one-dimensional convolutions of different sizes, respectively, to obtain two groups of features, U1 and U2;

[0037] S42, fusion process: The two sets of features are added together to obtain the fused feature U, and then passed through the one-dimensional global average pooling layer F gp Reduce the fusion features to a global information vector s of length L:

[0038]

[0039] In the formula, s l is the lth element of the global information vector, U l (k) is the k-th time point element of the l-th feature of the fused feature, and H is the time series length of the fused feature;

[0040] After the fully connected layer Reduce the dimension of the information vector and generate an attention information vector z with a sequence length of L / r, where r is a controllable parameter;

[0041] S43, selection operation: To strengthen important input features and suppress redundant features, softmax regression is used to obtain feature attention weights:

[0042]

[0043] Among them, A L , B L is the parameter matrix of z, exp is the exponential function with the natural constant e as the base, M A 、M B They are the feature attention weights of U1 and U2 respectively. The feature map V extracted by attention can be obtained through attention calculation. L :

[0044] V L =M A U1+M B U2,M A +M B =1.

[0045] In the preferred solution, the bidirectional long short-term memory network constructed in step S5 includes a fully connected layer and three BiLSTM layers, and the features with attention are mapped to the bidirectional long short-term memory network input through the fully connected layer; each BiLSTM layer processes the input feature data from the positive and negative directions respectively through two independent LSTM layers.

[0046] In the preferred solution, the hidden state and cell state of each time step of the forward LSTM are calculated by the following formula:

[0047] I t =σ(W Ix x t +W Ih h t-1 +b I );

[0048] F t =σ(W Fx x t +W Fh h t-1 +b F );

[0049] O t =σ(W Ox x t +W Oh h t-1 +b O );

[0050]

[0051] h t =O t ⊙tanh(C t );

[0052] Among them, I t 、F t , O t and are the input gate, forget gate, output gate and candidate state representing the current information, W and b are the network weight matrix and bias vector, σ and tanh are the Sigmoid and hyperbolic tangent functions, ⊙ represents the element-by-element product;

[0053] The reverse LSTM processes the sequence in the opposite direction, with the hidden state h′ at each time step t and cell state C′ t Also calculated by the above formula, the final bidirectional long short-term memory network output is concatenated from the hidden states in two directions: y′ t =[h t ; h′ t ].

[0054] In the preferred solution, the gated recurrent unit network constructed in step S6 includes a fully connected layer and three GRU layers; the features with attention are mapped to the gated recurrent unit network input through the fully connected layer; there are corresponding update gates and reset gates on each GRU unit layer, which respectively determine the retained information and discarded information from the previous moment to the current moment, specifically:

[0055] z t =σ(W zx x t +W zh h t-1 +b z );

[0056] r t =σ(W rx x t +W rh h t-1 +b r );

[0057]

[0058] Among them, z t is the update gate, r t is the reset gate, h t is the hidden state passed to the next moment, is a candidate hidden state, x t represents the network input mapped by the fully connected layer at the current moment, W and b are the network weight matrix and bias vector respectively, σ is the Sigmoid activation function, ⊙ represents the element-by-element product, and the final output is: y″ t =[h t ].

[0059] In a preferred embodiment, the combined long-term trend implicit output and short-term fluctuation trend implicit output in step S7 are formed by concatenating the bidirectional long short-term memory network output in step S5 with the gated recurrent unit network output in step S6: t =[y′ t ;y″ t ].

[0060] In the preferred solution, in step S7, the dependency between the two sets of implicit outputs and the model output shield posture is adaptively captured through a multi-head attention mechanism, and the specific operation is as follows:

[0061] The multi-head attention mechanism uses different linear mappings to map the input matrix of dimension m×d into N groups. Each group of independent self-attention mechanisms contains 3 matrices: value V, key K and query Q, where Q and K have the same length d K , V length is d V :

[0062]

[0063] In the formula, x i is the i-th implicit feature output by the BiLSTM and GRU networks, W i V , W i K , W i Q are three independent mapping weight matrices that can be iterated during network training. i represents the output matrix of the i-th group of self-attention operations;

[0064] Finally, the output matrix of the multi-head attention mechanism MA is obtained by concatenating the output matrices of the N groups of self-attention operations and linearly mapping them:

[0065] MA=Multihead(Q,K,V)=W O Concat(A1,...,A N );

[0066] Where W O It is the linear mapping weight matrix from the concatenation layer to the multi-head attention input matrix. The feature matrix MA output by the multi-head attention mechanism is obtained by weighted connection and remapping of the implicit features of the original two models.

[0067] The beneficial effects of the present invention are as follows: the present invention innovatively uses a combination of multiple attention mechanisms and prediction models to predict the technical solution, which improves the prediction accuracy of the shield posture to a certain extent, reduces the difficulty of model training, helps to finely control the posture, and improves the shield construction efficiency and construction safety. Grey correlation analysis GRA is used to perform preliminary feature type screening, and the features that mainly affect the shield posture are extracted as the input of the model in view of the high dimension of the characteristic data of the shield process, which effectively reduces the complexity of the model and improves the prediction efficiency; the selection kernel network SKNet is used to extract the input feature attention, which can effectively suppress the expression of redundant feature information and solve the problem that it is difficult to extract effective features from the long sequence feature input of the shield; the bidirectional long short-term memory network BiLSTM is combined with the bidirectional long short-term memory network GRU model for shield posture prediction, and the different representations of the two groups of models can be combined to enhance the expression and generalization ability of the model; finally, the multi-head attention mechanism MA is used as the mapping network from the output result of the combined prediction model to the prediction result, which can help the network automatically learn the importance and relevance of the output mapping and improve the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 The figure is a flow chart of the method of the present invention.

[0069] Figure 2 A kernel network is selected for the present invention to extract input feature attention structure diagram.

[0070] Figure 3 This is a model ablation comparison diagram of the regression prediction data and the original data obtained by predicting the shield posture using the model of the present invention.

[0071] Figure 4 This is a comparison chart of the shield posture prediction of the model of the present invention and other models in the same data interval. DETAILED DESCRIPTION

[0072] In order to verify the effectiveness of the model and method proposed in this invention, it is applied to the data of a subway tunnel in Guangzhou to predict the shield posture. To prove the effectiveness and generalization ability of the model, the data of two time periods are predicted respectively, and two sections of shield data from 3:00 on January 4, 2021 to 16:00 on February 4, 2021 and from 22:00 on March 29, 2021 to 6:00 on April 20, 2021 are selected for simulation experiments. The data is divided into training set and test set in the form of 7:3.

[0073] like Figure 1 Shown is a flow chart of the method of the present invention.

[0074] Combine the following Figures 1 to 4 , introduces a specific implementation of the present invention, a shield posture prediction method based on a multi-attention mechanism optimization combination model, specifically comprising the following steps:

[0075] Step 1, obtain the historical feature data and shield posture data of the target shield machine, and then use the missing value filling, excavation section data extraction, and data normalization methods to obtain the pre-processed excavation section historical feature data and pre-processed shield posture data. Then, the pre-processed historical feature data is used as the comparison feature column, and the pre-processed shield posture data is used as the reference column, and the correlation between each historical feature data and the shield posture data is obtained through grey correlation analysis;

[0076] The historical feature data in step 1 is:

[0077] X i (k),i∈[1,n],k∈[1,m],

[0078] Where X i (k) is the kth point data in the historical feature data after the i-th preprocessing, and n and m are the type and time point of the collected historical feature data respectively;

[0079] The posture data after preprocessing in step 1 is:

[0080] Y o(k),o∈[1,4],k∈[1,m],

[0081] Among them, Y o (k) represents the posture data at time point k after preprocessing, where o represents different types of posture data: head vertical posture, head horizontal posture, tail vertical posture, head horizontal posture, and m is the number of time points of the collected historical feature data;

[0082] The data of the historical feature data at the kth time point is:

[0083] {X1(k),X1(k),...,X n (k)};

[0084] In step 1, the grey correlation analysis method is used to obtain the correlation between each historical feature data and the shield posture data. Specifically:

[0085] Calculate the grey correlation coefficient:

[0086]

[0087] i∈[1,n],k∈[1,m],o∈[1,4],

[0088] Among them, Y o (k) represents the shield posture data of type o after preprocessing at the kth time point. i (k) represents the historical feature data of the i-th category after preprocessing at the k-th time point. n is the type of historical feature data collected, and m is the number of time points for collecting data. ρ is the resolution coefficient, and its value range is (0,∞). ξ o,i (k) represents the grey correlation coefficient between the i-th type of preprocessed historical feature data and the o-th type of preprocessed shield posture data at the kth time point.

[0089] Calculate the correlation:

[0090]

[0091] where γ i is the correlation between the historical characteristic data of the i-th shield machine and the shield machine posture data, γ i The closer it is to 1, the higher the correlation between this historical feature data and the shield posture data.

[0092] Step 2: Arrange the correlation of each historical feature data in descending order, set the feature selection dimension, and take the feature column with the highest correlation according to the set feature selection dimension as the feature data after feature selection;

[0093] Step 3: Perform time window processing on the feature data after feature selection according to the selected time series length to generate a feature time series data set as the input data of the deep learning model. Split the feature time series data into a training set and a test set according to a certain ratio;

[0094] In step 3, the historical feature data after time window processing is as follows:

[0095] x(k)={[X1(k-T+1),...,X c (k-T+1)];...;[X1(k),...,X c (k)]}'

[0096] Where x(k) is the network model input time window data at time point k, T is the time window time point length, and c is the number of feature data types selected after grey correlation degree sorting;

[0097] Step 4, by selecting the kernel unit network, feature attention calculation is performed on the input time series data, and an implicit feature map with attention is adaptively formed;

[0098] Figure 2 The kernel network is selected for the present invention to extract the input feature attention structure diagram, and the overall process is reflected in combination with the following case operation process.

[0099] The attention mechanism calculation for the time series based on the selected kernel network described in step 4 is as follows:

[0100] Step 4.1, splitting process: take the time window data, i.e., the multi-dimensional time series features, as input, and pass them through two sets of one-dimensional convolutions of different sizes. Here, the convolution kernel sizes are selected as 3×1 and 5×1 to obtain two sets of features, U1 and U2.

[0101] Step 4.2, fusion process: The two sets of features are added together to obtain the fused feature U. Then, the fused feature U is added together through a one-dimensional global average pooling layer. The fused features are reduced to a global information vector s of length L.

[0102]

[0103] In the formula, s l is the lth element of the global information vector, U l (k) is the k-th time point element of the l-th feature of the fused feature, and H is the time series length of the fused feature.

[0104] After the fully connected layer Reduce the dimension of the information vector and generate an attention information vector z with a sequence length of L / r, where r is a controllable parameter.

[0105] Step 4.3, select operation: To strengthen important input features and suppress redundant features, use softmax regression to obtain feature attention weights:

[0106]

[0107]

[0108] Among them, A L , B L is the parameter matrix of z, exp is the exponential function with the natural constant e as the base, M A 、M B They are the feature attention weights of U1 and U2 respectively. The feature map V extracted by attention can be obtained through attention calculation. L .

[0109] V L =M A U1+M B U2,M A +M B =1;

[0110] Step 5: construct a bidirectional long short-term memory network. Based on the strong memory characteristics of the bidirectional long short-term memory network for long-term trends, the long-term trend implicit output of the shield posture is predicted. The implicit feature map with attention is used as the input data of the bidirectional long short-term memory network through the fully connected layer. The network parameters are optimized iteratively according to the loss function model of the bidirectional long short-term memory network to obtain the optimal network parameters.

[0111] The bidirectional long short-term memory network constructed in step 5 includes a fully connected layer and three BiLSTM layers. The features with attention are mapped to the input of the bidirectional long short-term memory network through the fully connected layer. Each BiLSTM layer processes the input feature data from the positive and negative directions through two independent LSTM layers.

[0112] Specifically, the hidden state and cell state of each time step of the forward LSTM are calculated by the following formula:

[0113] I t =σ(W Ix x t +W Ih h t-1 +b I )

[0114] F t =σ(W Fx x t +W Fh h t-1 +b F )

[0115] O t =σ(W Ox x t +W Oh h t-1 +b O )

[0116]

[0117]

[0118] h t =O t ⊙tanh(C t )

[0119] Among them, I t 、F t , O t and They are the input gate, forget gate, output gate and candidate state representing the current information, W and b are the network weight matrix and bias vector, σ and tanh are Sigmoid and hyperbolic tangent functions, ⊙ represents the element-by-element product.

[0120] The reverse LSTM processes the sequence in the opposite direction, with the hidden state h′ at each time step t and cell state C′ t It can be calculated by a similar formula. The final bidirectional long short-term memory network output is concatenated from the hidden states in two directions: y′ t =[h t ; h′ t ].

[0121] Step 6: Construct a gated recurrent unit network. Based on the fact that the gated recurrent unit network is sensitive to short-term fluctuations, the implicit output of the short-term fluctuation trend of the shield posture is predicted. The implicit feature map with attention in step 4 is used as the input data of the gated recurrent network through the fully connected layer mapping. The network parameters are optimized iteratively according to the network loss function model to obtain the optimal network parameters.

[0122] The gated recurrent unit network constructed in step 6 includes a fully connected layer and three GRU layers. The features with attention are mapped to the input of the gated recurrent unit network through the fully connected layer. There are corresponding update gates and reset gates on each GRU unit layer, which respectively determine the retained and discarded information from the previous moment to the current moment. Specifically:

[0123] z t =σ(W zx x t +W zh h t-1 +b z )

[0124] r t =σ(W rx x t +W rh h t-1 +b r )

[0125]

[0126] Among them, z t is the update gate, r t is the reset gate, h t is the hidden state passed to the next moment, is a candidate hidden state, x t represents the network input mapped by the fully connected layer at the current moment, W and b are the network weight matrix and bias vector respectively, σ is the Sigmoid activation function, and ⊙ represents the element-by-element product. The final output is: y″ t =[h t ];

[0127] Step 7: Combine the long-term trend implicit output and the short-term fluctuation trend implicit output, and adaptively capture the dependency between the two sets of implicit outputs and the model output on the shield posture through the multi-head attention mechanism, and finally obtain the shield posture prediction result.

[0128] The combined long-term trend implicit output and short-term volatility trend implicit output in step 7 are obtained by concatenating the bidirectional long short-term memory network output in step 5 and the gated recurrent unit network output in step 6: y t =[y′ t ;y″ t ].

[0129] As described in step 7, the multi-head attention mechanism is used to adaptively capture the dependency between the two sets of implicit outputs and the model output shield posture. The specific operations are as follows:

[0130] The multi-head attention mechanism uses different linear mappings to map the input matrix of dimension m×d into N groups. Each group of independent self-attention mechanisms contains 3 matrices: value V, key K, query Q, where Q and K have the same length d K , V length is d V .

[0131]

[0132]

[0133] In the formula, x i is the i-th implicit feature output by the BiLSTM and GRU networks. iV , W i K , W i Q are three independent mapping weight matrices that can be iterated during network training. i represents the output matrix of the i-th group of self-attention operations.

[0134] Finally, the output matrix of the multi-head attention mechanism MA is obtained by concatenating the output matrices of the N groups of self-attention operations and linearly mapping them:

[0135] MA=Multihead(Q,K,V)=W O Concat(A1,...,A N )

[0136] Where W O It is the linear mapping weight matrix from the concatenation layer to the multi-head attention input matrix. The feature matrix MA output by the multi-head attention mechanism is obtained by weighted connection and remapping of the implicit features of the original two models, that is, the final shield posture prediction result.

[0137] Figure 3 A comparison chart of the regression prediction data and the original data obtained by applying the shield posture prediction based on the multi-attention mechanism optimization combination model provided in the embodiment of the present invention. The thicker solid line in the figure is the true posture value, the thinner solid line is the prediction curve of this method, and the remaining curves are the prediction curves of the ablation model. The ablation comparison model includes: bidirectional long short-term memory network (BiLSTM), gated recurrent unit (GRU), bidirectional long short-term memory network-gated recurrent unit (BiLSTM-GRU), selected kernel network-bidirectional long short-term memory network-gated recurrent unit (SKNet-BiLSTM-GRU), selected kernel network-bidirectional long short-term memory network-gated recurrent unit-multi-head attention mechanism (SKNet-BiLSTM-GRU-MA).

[0138] Table 1 Ablation model evaluation indicators

[0139]

[0140] It can be seen from Table 1 that the model and method provided by the present invention have high prediction accuracy, with a root mean square error RMSE of 0.89, a mean absolute error MAE of 0.69, and R 2 It also exceeds 0.98, which shows that the present invention is reasonable and effective in terms of model structure and method design.

[0141] Figure 4A comparison chart of the regression prediction data and the original data obtained by applying the shield posture prediction based on the multi-attention mechanism optimization combination model provided in the embodiment of the present invention. The thicker solid line in the figure is the true posture value, the fine dotted line is the prediction curve of this method, and the remaining curves are the prediction curves of the comparison model. The comparison model includes: BP neural network-support vector machine (BP-SVM), multi-layer long short-term memory network (MLSTM), gated recurrent unit-random forest (GRU-RF), selected kernel network-bidirectional long short-term memory network-gated recurrent unit-multi-head attention mechanism (SKNet-BiLSTM-GRU-MA).

[0142] Table 2 Comparison model evaluation indicators

[0143]

[0144] It can be seen from Table 2 that, in this data segment, the prediction accuracy of the model and method provided by the present invention is still relatively high, with a root mean square error RMSE of 2.06, a mean absolute error MAE of 2.28, and R 2 It also exceeds 0.98, which shows that the present invention is superior in model structure and method design.

[0145] In summary, the present invention proposes a shield posture prediction method based on the SKNet-BiLSTM-GRU-MA multi-attention mechanism optimization combination model, which realizes the accurate prediction of the shield posture, contributes to the refined control of the posture during shield tunnel construction, and improves the safety and efficiency of the construction process.

[0146] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A shield machine posture prediction method based on a multi-attention mechanism optimization combination model, characterized by: The following steps are involved: S1, obtaining and preprocessing the historical feature data of the tunneling section and the shield posture data of the target shield machine, and then using the preprocessed historical feature data as a comparison feature column and the preprocessed shield posture data as a reference column to obtain the correlation between each historical feature data and the shield posture data; S2, arranging the correlation degree of each historical feature data in descending order, setting the feature selection dimension, and taking the feature column with the highest correlation degree as the feature data after feature selection according to the set feature selection dimension; S3, processing the feature data after feature selection to generate a feature time series data set as the input data of the deep learning model, and dividing the feature time series data into a training set and a test set according to a certain ratio; S4, by selecting the kernel unit network, feature attention calculation is performed on the input time series data, and implicit feature mapping with attention is adaptively formed; S5. Construct a bidirectional long short-term memory network to predict the long-term trend implicit output of the shield posture: use the implicit feature map with attention as the input data of the bidirectional long short-term memory network through the fully connected layer, and iterate the network parameters according to the loss function model of the bidirectional long short-term memory network to obtain the optimal network parameters; S6. Construct a gated recurrent unit network to predict the implicit output of the short-term fluctuation trend of the shield posture: use the implicit feature map with attention in step S4 as the input data of the gated recurrent network through the fully connected layer mapping, and iterate the network parameters according to the network loss function model to obtain the optimal network parameters; S7. Combine the long-term trend implicit output and the short-term fluctuation trend implicit output, and adaptively capture the dependency between the long-term trend implicit output and the short-term fluctuation trend implicit output and the model output on the shield posture through the multi-head attention mechanism, and finally obtain the shield posture prediction result.

2. The shield posture prediction method based on the multi-attention mechanism optimization combination model according to claim 1 is characterized in that: The preprocessing in step S1 includes missing value filling, excavation section data extraction and data normalization; The correlation between each historical feature data and shield posture data is obtained through grey correlation analysis; The feature data after feature selection in step S3 is subjected to time window processing according to the selected time series length.

3. According to claim 1, the shield posture prediction method based on the multi-attention mechanism optimization combination model is characterized by: step The historical feature data in S1 is: X i (k),i∈[1,n],k∈[1,m], Where X i (k) is the kth point data in the historical feature data after the i-th preprocessing, and n and m are the type and time point of the collected historical feature data respectively; The posture data after preprocessing in step S1 is: Y o (k),o∈[1,4],k∈[1,m], Among them, Y o (k) represents the posture data at time point k after preprocessing, where o represents different types of posture data: head vertical posture, head horizontal posture, tail vertical posture and head horizontal posture, and m is the number of time points of the collected historical feature data; The data of the historical feature data at the kth time point is: {X1(k),X1(k),...,X n (k)} In step S1, the correlation between each historical feature data and shield posture data is obtained by grey correlation analysis method, specifically: Calculate the grey correlation coefficient: i∈[1,n],k∈[1,m],o∈[1,4]; Among them, Y o (k) represents the shield posture data of the oth type after preprocessing at the kth time point, X i (k) represents the historical feature data of the i-th category after preprocessing at the k-th time point, n is the type of historical feature data collected, m is the number of time points for collecting data, ρ is the resolution coefficient, and its value range is (0,∞), ξ o,i (k) represents the grey correlation coefficient between the historical feature data after preprocessing of the i-th type and the shield posture data after preprocessing of the o-th type at the k-th time point; Calculate the correlation: where γ i is the correlation between the historical characteristic data of the i-th shield machine and the shield machine posture data, γ i The closer it is to 1, the higher the correlation between this historical feature data and the shield posture data.

4. The shield posture prediction method based on the multi-attention mechanism optimization combination model described in claim 1 is characterized by: In step S3, the historical feature data after time window processing is: x(k)={[X1(k-T+1),…,X c (k-T+1)];...;[X1(k),...,X c (k)]}; Where x(k) is the network model input time window data at time point k, T is the time window time point length, and c is the number of feature data types selected after grey correlation degree sorting.

5. The shield posture prediction method based on the multi-attention mechanism optimization combination model according to claim 1 is characterized in that: In step S4, the attention mechanism calculation of the time series is performed based on the selected kernel network. The specific process is as follows: S41, splitting process: taking the time window data, i.e., the multi-dimensional time series features, as input, and passing through two groups of one-dimensional convolutions of different sizes, respectively, to obtain two groups of features, U1 and U2; S42, fusion process: The two sets of features are added together to obtain the fused feature U, and then passed through the one-dimensional global average pooling layer F gp Reduce the fusion features to a global information vector s of length L: In the formula, s l is the lth element of the global information vector, U l (k) is the k-th time point element of the l-th feature of the fused feature, and H is the time series length of the fused feature; Then through the fully connected layer F fc Reduce the dimension of the information vector and generate an attention information vector z with a sequence length of L / r, where r is a controllable parameter; S43, selection operation: To strengthen important input features and suppress redundant features, softmax regression is used to obtain feature attention weights: Among them, A L , B L is the parameter matrix of z, exp is the exponential function with the natural constant e as the base, M A 、M B are the feature attention weights of U1 and U2 respectively, and the feature map V extracted by attention is obtained through attention calculation. L : V L =M A U1+M B U2,M A +M B =1。 6. The shield posture prediction method based on the multi-attention mechanism optimization combination model according to claim 1 is characterized in that: The bidirectional long short-term memory network constructed in step S5 includes a fully connected layer and three BiLSTM layers. The features with attention are mapped to the input of the bidirectional long short-term memory network through the fully connected layer; each BiLSTM layer processes the input feature data from the positive and negative directions respectively through two independent LSTM layers.

7. The shield posture prediction method based on the multi-attention mechanism optimization combination model according to claim 6 is characterized in that: Specifically, the hidden state and cell state of each time step of the forward LSTM are calculated by the following formula: I t =σ(W Ix x t +W Ih h t-1 +b I ); F t =σ(W Fx x t +W Fh h t-1 +b F ); O t =σ(W Ox x t +W Oh h t-1 +b O ); h t =O t ⊙tanh(C t ); Among them, I t 、F t , O t and are the input gate, forget gate, output gate and candidate state representing the current information, W and b are the network weight matrix and bias vector, σ and tanh are the Sigmoid and hyperbolic tangent functions, ⊙ represents the element-by-element product; The reverse LSTM processes the sequence in the opposite direction, with the hidden state h′ at each time step t and cell state C′ t The calculation formula is the same as the forward LSTM. The final bidirectional long short-term memory network output is composed of the concatenation of the hidden states in two directions: y′ t =[h t ; h′ t ].

8. The shield posture prediction method based on the multi-attention mechanism optimization combination model according to claim 1 is characterized in that: The gated recurrent unit network constructed in step S6 includes a fully connected layer and three GRU layers; the features with attention are mapped to the gated recurrent unit network input through the fully connected layer; there are corresponding update gates and reset gates on each GRU unit layer, which respectively determine the retained information and discarded information from the previous moment to the current moment, specifically: z t =σ(W zx x t +W zh h t-1 +b z ); r t =σ(W rx x t +W rh h t-1 +b r ); Among them, z t is the update gate, r t is the reset gate, h t is the hidden state passed to the next moment, is a candidate hidden state, x t represents the network input mapped by the fully connected layer at the current moment, W and b are the network weight matrix and bias vector respectively, σ is the Sigmoid activation function, ⊙ represents the element-by-element product, and the final output is: y″ t =[h t ].

9. The shield posture prediction method based on the multi-attention mechanism optimization combination model according to claim 1 is characterized in that: The combined long-term trend implicit output and short-term volatility trend implicit output in step S7 are formed by concatenating the bidirectional long short-term memory network output in step S5 with the gated recurrent unit network output in step S6: and t =[and′ t ;and" t ]。 10. The shield posture prediction method based on the multi-attention mechanism optimization combination model according to claim 1 is characterized in that: In step S7, the multi-head attention mechanism is used to adaptively capture the dependency between the two sets of implicit outputs and the model output shield posture. The specific operations are as follows: The multi-head attention mechanism uses different linear mappings to map the input matrix of dimension m×d into N groups. Each group of independent self-attention mechanisms contains 3 matrices: value V, key K and query Q, where Q and K have the same length d K , V length is d V : In the formula, x i is the i-th implicit feature output by the BiLSTM and GRU networks, W i V , W i K , W i Q are three independent mapping weight matrices that are iterated during network training. i represents the output matrix of the i-th group of self-attention operations; Finally, the output matrix of the multi-head attention mechanism MA is obtained by concatenating the output matrices of the N groups of self-attention operations and linearly mapping them: MA=Multihead(Q,K,V)=W O Concat(A1,...,A N ); Where W O It is the linear mapping weight matrix from the concatenation layer to the multi-head attention input matrix. The feature matrix MA output by the multi-head attention mechanism is obtained by weighted connection and remapping of the implicit features of the original two models.

Images