TBM tunneling parameter prediction model training method and device and prediction method

Abstract

The invention relates to the technical field of geological data processing, discloses a TBM tunneling parameter prediction model training method and device and a prediction method, and aims at solving the problem that an existing TBM tunneling parameter prediction scheme is poor in accuracy. Acquiring a plurality of first sample images of a surrounding rock tunnel face tunneled by the TBM, and determining surrounding rock grade evaluation parameters, rock mass mechanical parameters and TBM tunneling parameters corresponding to the first sample images; after feature discrete matching is carried out on the surrounding rock grade evaluation parameters corresponding to the first sample images, the surrounding rock grades corresponding to the first sample images are determined according to a surrounding rock grade prediction model obtained through pre-training; the method comprises the following steps: screening rock mass mechanical parameters based on a grey relational degree method, and establishing a sample data set; and training a CNN-LSTM model according to the sample data set to obtain a TBM tunneling parameter prediction model. The TBM tunneling parameter prediction accuracy is improved, and the method is particularly suitable for rock tunneling.

Description

Technical Field

[0001] The present invention relates to the technical field of geological data processing, and in particular to a prediction model training method and device, and a prediction method for TBM excavation parameters. Background Art

[0002] TBM (Tunnel Boring Machine) is a full-face tunnel boring machine. It is a large-scale engineering machine that drives the cutterhead to rotate and the hydraulic cylinder to propel the cutterhead to the rock wall under a certain thrust. The cutter installed on the cutterhead breaks the rock to form the tunnel section in one go. TBM construction has the advantages of high automation, fast construction speed, labor saving, safety and economy, one-time formation, no influence from external climate, and can control ground subsidence during excavation, reduce the impact on ground buildings, and underwater construction does not affect ground traffic in the water. It is the most promising mechanical equipment for rock tunnel excavation.

[0003] The traditional TBM construction method is for the driver to set the TBM excavation parameters based on experience. Since this method relies on the driver's experience and is highly subjective, it is prone to risks such as machine jamming and severe cutterhead wear. In order to ensure the safety and efficiency of TBM tunnel construction, it is necessary to scientifically and effectively predict the TBM tunnel excavation parameters. The existing TBM excavation parameter predictions mainly include theoretical models, empirical models, and machine learning models. Limited by the limitations of the theoretical model itself, it is difficult to adapt to the complex problems faced in actual engineering; the empirical model has the problem of difficulty in mining the deep relationship between data, and it is difficult to play its advantages when faced with complex geological conditions and many excavation factors; machine learning, with its powerful data processing and correlation mining capabilities, provides a new idea for model prediction of TBM tunnels.

[0004] At TBM construction sites, sensors can collect hundreds of types of on-site construction parameters, with data volumes reaching hundreds of millions. However, in existing machine learning-based TBM excavation parameter prediction schemes, parameters are usually selected based on human experience, which makes it difficult to fully reveal the implicit relationship between people, machines, and rocks in TBM construction. The construction of machine learning models has problems such as shallow data mining, insufficient correlation, and poor consistency, which leads to problems such as weak robustness and overfitting, making it difficult to accurately predict TBM excavation parameters. Summary of the invention

[0005] The present invention aims to solve the problem of poor accuracy of existing prediction schemes for TBM excavation parameters, and proposes a prediction model training method and device, and a prediction method for TBM excavation parameters.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: In a first aspect, the present invention provides a method for training a prediction model of TBM excavation parameters, the method comprising: Taking the mileage pile number as a reference, a plurality of first sample images of the surrounding rock face of the TBM excavation are obtained, and surrounding rock grade evaluation parameters, rock mass mechanics parameters and TBM excavation parameters corresponding to each first sample image are determined, wherein the TBM excavation parameters include TBM ascending section excavation parameters and TBM stable section excavation parameters; After performing feature discrete matching on the surrounding rock grade evaluation parameters corresponding to each first sample image, the surrounding rock grade corresponding to each first sample image is determined according to the surrounding rock grade prediction model obtained by pre-training; After screening the rock mass mechanical parameters based on the grey correlation method, a sample data set is established, wherein the sample data set includes characteristic parameters and their corresponding true label values, wherein the characteristic parameters include surrounding rock grade, screened rock mass mechanical parameters and TBM ascending section excavation parameters, and the true label values ​​are TBM stable section excavation parameters; The CNN-LSTM model is trained according to the sample data set to obtain a TBM excavation parameter prediction model.

[0007] Furthermore, the surrounding rock grade evaluation parameters include rock mass integrity parameters, rock mass structure type parameters, rock mass hardness parameters, joint weathering parameters, groundwater parameters and initial geostress parameters.

[0008] Furthermore, the method for determining the rock mass integrity parameter and the rock mass structure type parameter includes: The pre-trained fracture extraction model is used to extract the tunnel face joint fracture map corresponding to each first sample image, and the rock mass integrity parameter and rock mass structure type parameter corresponding to each first sample image are determined according to the tunnel face joint fracture map.

[0009] Furthermore, the crack extraction model is obtained by training a deep convolutional neural network including a feature extraction layer, a size recovery layer and a skip connection structure, and the training method includes: Acquire multiple second sample images of the surrounding rock face of the TBM excavation, perform joint and fissure interpretation and annotation on each second sample image after preprocessing, and establish an image data set, wherein the image data set includes the second sample images and their corresponding annotation results; A deep convolutional neural network is trained according to the image data set, and it is determined whether the loss function and evaluation index of the deep convolutional neural network meet the first preset condition. If so, a crack extraction model is obtained, otherwise, the training parameters are adjusted and re-trained, and the evaluation index of the deep convolutional neural network includes at least accuracy, precision, recall rate, F1 score and average intersection-over-union ratio.

[0010] Furthermore, the rock mass mechanical parameters are screened based on the grey correlation method, including: Generate characteristic sequences and normalize the data in each characteristic sequence, wherein the characteristic sequence includes a comparison sequence corresponding to each rock mass mechanical parameter and a reference sequence corresponding to each TBM stable section excavation parameter; The average correlation between each comparison sequence and each reference sequence is calculated respectively, and the rock mechanics parameters corresponding to the comparison sequences with an average correlation greater than a preset correlation are selected as the rock mechanics parameters for establishing the sample data set.

[0011] Furthermore, the average correlation between each comparison sequence and each reference sequence is calculated respectively, specifically including: Sorting the data in the comparison sequence and the reference sequence, calculating the absolute difference sequence between the comparison sequence and the reference sequence, and determining the minimum and maximum absolute difference values ​​in the absolute difference sequence; Calculate the correlation coefficient between the comparison sequence and the reference sequence, the calculation formula is as follows: ; in, Indicates The first Data and The correlation coefficient between the reference sequences is Indicates Comparison sequence and The absolute difference sequence of the reference sequence The absolute difference, represents the resolution coefficient, , represents the minimum absolute difference, Indicates the maximum value of the absolute difference; The correlation between the comparison sequence and the reference sequence is calculated using the following formula: ; in, Indicates Compare the sequence with The correlation between the reference sequences is Indicates the number of data in the comparison sequence; The average degree of association between the comparison sequence and each reference sequence was calculated using the following formula: ; in, Indicates The average degree of relevance between the comparison sequences and each reference sequence, Indicates the number of reference sequences.

[0012] Furthermore, training a CNN-LSTM model according to the sample data set specifically includes: Preprocessing the feature parameters in the sample data set to generate a matrix data set, and inputting the matrix data set and its corresponding true label values ​​into the CNN-LSTM model, wherein the preprocessing at least includes outlier removal and data optimization; The CNN-LSTM model uses the CNN layer to extract features and inputs the extracted features into the LSTM layer. The LSTM layer adjusts the model parameters according to the input features and uses the back propagation algorithm. Determine whether the loss function and evaluation index of the CNN-LSTM model meet the second preset condition. If so, obtain the TBM excavation parameter prediction model; otherwise, continue to adjust the model parameters. The evaluation index of the CNN-LSTM model includes at least the mean absolute error, the root mean square error, and the mean absolute percentage error.

[0013] Furthermore, the rock mass mechanical parameters include at least uniaxial compressive strength, uniaxial tensile strength, rock mass integrity coefficient and surrounding rock wear resistance; the TBM ascending section excavation parameters and TBM stable section excavation parameters include at least thrust, penetration and torque.

[0014] In a second aspect, the present invention provides a prediction model training device for TBM excavation parameters, the device comprising a processor, a memory and a communication bus; The communication bus is used to realize the connection and communication between the processor and the memory; The processor is used to execute one or more programs in the memory to implement the steps of the TBM excavation parameter prediction model training method as described in the first aspect.

[0015] In a third aspect, the present invention provides a method for predicting TBM excavation parameters, the method comprising: Obtain surrounding rock grade evaluation parameters, rock mass mechanics parameters and TBM ascending section excavation parameters corresponding to the tunnel section to be predicted; After performing feature discrete matching on the surrounding rock grade evaluation parameters corresponding to the tunnel section to be predicted, the surrounding rock grade corresponding to the tunnel section to be predicted is determined according to the surrounding rock grade prediction model obtained by pre-training; The surrounding rock grade, rock mechanical parameters and TBM ascending section excavation parameters corresponding to the tunnel section to be predicted are input into the TBM excavation parameter prediction model to obtain the TBM stable section excavation parameters of the tunnel section to be predicted. The TBM excavation parameter prediction model is trained by the TBM excavation parameter prediction model training method as described in the first aspect.

[0016] The beneficial effects of the present invention are as follows: the present invention adopts the grey correlation method to screen the rock mechanical parameters, takes the screened rock mechanical parameters as characteristic parameters, and takes the surrounding rock grade of the surrounding rock face and the TBM ascending section excavation parameters as characteristic parameters, and constructs a TBM excavation parameter prediction model for characterizing the mapping relationship between the characteristic parameters and the TBM stable section excavation parameters. The grey correlation method is used to reveal the strength of the relationship between the characteristic parameters, screen the main characteristic parameters, and avoid the problem of overfitting. At the same time, the TBM excavation parameter prediction model of the present invention combines the advantages of convolutional neural network (CNN) and long short-term memory network (LSTM), can process spatial features and time series data at the same time, improves the prediction ability and generalization ability of the model, can adapt to the differences of multi-source field data, and improves the accuracy of TBM excavation parameter prediction. The surrounding rock grade evaluation parameters are used and the surrounding rock grade is determined based on the surrounding rock grade prediction model, which improves the accuracy and consistency of the surrounding rock grade, thereby further improving the accuracy of TBM excavation parameter prediction. The image of the surrounding rock face is used and based on the fracture extraction model, the rock integrity parameters and rock structure type parameters in the surrounding rock grade evaluation parameters are determined, which improves the accuracy and consistency of the surrounding rock grade evaluation parameters, thereby further improving the accuracy and consistency of the surrounding rock grade and TBM excavation parameter prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic flow chart of a method for training a prediction model for TBM excavation parameters provided by an embodiment of the present invention; Figure 2 A schematic diagram of a flow chart of a method for predicting TBM excavation parameters provided by an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a prediction model training device for TBM excavation parameters provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0019] In some of the processes described in the specification of the present invention and the above-mentioned drawings, multiple operations appearing in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel, and the sequence numbers of the operations are only used to distinguish different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel.

[0020] The technical solution of the embodiment of the present invention is applicable to the application scenario of TBM construction. Since the parameters in the current prediction scheme of TBM excavation parameters based on machine learning are usually selected based on human experience, this method is difficult to fully reveal the implicit relationship between people, machines and rocks in TBM construction. The construction of machine learning models has problems such as shallow data mining, insufficient correlation and poor consistency, which leads to problems such as weak robustness and overfitting, making it difficult to accurately predict TBM excavation parameters.

[0021] Based on this, the technical solution of the present invention is proposed. In an embodiment of the present invention, with the mileage pile number as a reference benchmark, multiple first sample images of the surrounding rock face of TBM excavation are obtained, and the surrounding rock grade evaluation parameters, rock mechanical parameters and TBM excavation parameters corresponding to each first sample image are determined, and the TBM excavation parameters include TBM rising section excavation parameters and TBM stable section excavation parameters; after feature discrete matching of the surrounding rock grade evaluation parameters corresponding to each first sample image, the surrounding rock grade corresponding to each first sample image is determined according to a surrounding rock grade prediction model obtained by pre-training; after screening the rock mechanical parameters based on the grey correlation method, a sample data set is established, the sample data set includes feature parameters and their corresponding true label values, the feature parameters include surrounding rock grade, screened rock mechanical parameters and TBM rising section excavation parameters, and the true label values ​​are TBM stable section excavation parameters; a CNN-LSTM model is trained according to the sample data set to obtain a TBM excavation parameter prediction model.

[0022] Specifically, the present invention screens rock mechanical parameters based on the grey correlation method, and uses the surrounding rock grade, the screened rock mechanical parameters and the TBM ascending section excavation parameters as characteristic parameters, establishes a sample data set for predicting the TBM stable section excavation parameters, and trains the CNN-LSTM model based on the sample data set, thereby constructing a TBM excavation parameter prediction model for characterizing the mapping relationship between the characteristic parameters and the TBM stable section excavation parameters. The embodiment of the present invention uses the grey correlation method to reveal the strength of the relationship between the characteristic parameters, screens the main characteristic parameters, and avoids the problem of overfitting. At the same time, the TBM excavation parameter prediction model of the present invention combines the advantages of convolutional neural network (CNN) and long short-term memory network (LSTM), can simultaneously process spatial features and time series data, improves the prediction ability and generalization ability of the model, can adapt to the differences of multi-source field data, and improves the accuracy of TBM excavation parameter prediction. Among them, the surrounding rock grade evaluation parameters are used and the surrounding rock grade is determined based on the surrounding rock grade prediction model, which improves the accuracy and consistency of the surrounding rock grade, thereby further improving the accuracy of TBM excavation parameter prediction.

[0023] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0024] Figure 1 A flow chart showing a method for training a prediction model of TBM excavation parameters provided by an embodiment of the present invention is shown. Figure 1 The prediction model training method of TBM excavation parameters provided by the embodiment of the present invention comprises the following steps: S101. Taking the mileage pile number as a reference, a plurality of first sample images of the surrounding rock face of the TBM excavation are obtained, and surrounding rock grade evaluation parameters, rock mass mechanics parameters and TBM excavation parameters corresponding to each first sample image are determined, wherein the TBM excavation parameters include TBM ascending section excavation parameters and TBM stable section excavation parameters.

[0025] In the embodiment of the present invention, the surrounding rock grade evaluation parameters include rock mass integrity parameters, rock mass structure type parameters, rock mass hardness parameters, joint weathering parameters, groundwater parameters and initial geostress parameters.

[0026] Among them, the determination methods of rock mass integrity parameters and rock mass structure type parameters include: The pre-trained fracture extraction model is used to extract the tunnel face joint fracture map corresponding to each first sample image, and the rock mass integrity parameter and rock mass structure type parameter corresponding to each first sample image are determined according to the tunnel face joint fracture map.

[0027] In an embodiment of the present invention, the crack extraction model is obtained by training a deep convolutional neural network including a feature extraction layer, a size recovery layer and a skip connection structure, and the training method includes: Acquire multiple second sample images of the surrounding rock face of the TBM excavation, perform joint and fissure interpretation and annotation on each second sample image after preprocessing, and establish an image data set, wherein the image data set includes the second sample images and their corresponding annotation results; A deep convolutional neural network is trained according to the image data set, and it is determined whether the loss function and evaluation index of the deep convolutional neural network meet the first preset condition. If so, a crack extraction model is obtained, otherwise, the training parameters are adjusted and re-trained, and the evaluation index of the deep convolutional neural network includes at least accuracy, precision, recall rate, F1 score and average intersection-over-union ratio.

[0028] In practical applications, multiple second sample images of the surrounding rock face are collected. After the on-site acquisition equipment is installed, the second sample images can be clearly collected from the front view direction to obtain images, images, videos, geological sketches and other data. When collecting, it is necessary to consider the specific conditions such as the lighting balance, acquisition conditions and equipment safety on site. The best time to collect is before the installation of the arch frame after the face is cleared and slag is discharged, and before the concrete is sprayed after the locking anchor and advance guide tube are installed. After the collected second sample images are subjected to targeted preprocessing processes such as smoke removal and histogram equalization, the collected second sample images are annotated with joints and fissures using image annotation tools with the assistance of geological engineers, and an image dataset including the second sample images and their corresponding annotation results is generated to provide data support for subsequent model training and automatic extraction.

[0029] The deep convolutional neural network in the embodiment of the present invention introduces a convolutional attention mechanism, a residual network, and an "encoder-decoder" structure of a jump connection. The deep convolutional neural network mainly includes a feature extraction layer, a size recovery layer, and a jump connection structure. Among them, the feature extraction layer uses a residual network structure, and after convolution and 4 pooling downsampling processes, extracts high-level feature information in the second sample image, and extracts important joint and crack information in the original image. The embedded lightweight convolutional attention mechanism module can suppress interference information and focus on important information in space and channels. The size recovery layer restores the size of the highly pooled feature information through deconvolution or upsampling methods, and restores it to a size consistent with the input layer, so as to realize the recognition of cracks that overlap with the original image. The jump connection structure can make the feature information obtained by each convolution pooling in the feature extraction layer and the corresponding superposition fusion of the size recovery layer to realize the capture of global information. Please refer to Table 1 for the structure of the deep convolutional neural network.

[0030] Table 1. Deep convolutional neural network structure

[0031] In the training process of the deep convolutional neural network, a learning rate warm-up strategy is adopted, that is, the initial learning rate is set to 0.0001, and the learning rate is adjusted after the model is trained for several periods. The optimization algorithm adopts the Adam algorithm, and the accuracy, precision, recall rate, F1 score and average intersection-over-union ratio are selected. When the evaluation index of the deep convolutional neural network meets the first preset condition, the training is completed and the crack extraction model is obtained.

[0032] Among them, the evaluation indicators of the deep convolutional neural network are calculated by the following formulas: ; ; ; ; ; in, Indicates the accuracy, Indicates the accuracy, represents the recall rate, represents the F1 score, represents the average intersection-over-union ratio, represents the number of pixels that the model correctly predicted as positive. represents the number of pixels for which the model correctly predicted negative examples as negative, represents the number of pixels where the model incorrectly predicts negative examples as positive, Represents the number of pixels where the model incorrectly predicted a positive example as a negative example.

[0033] After the fracture extraction model is obtained through training, each first sample image is input into the fracture extraction model to obtain the tunnel face joint fracture map corresponding to each first sample image. The rock mass integrity parameters and rock mass structure type parameters corresponding to each first sample image can be determined based on the tunnel face joint fracture map.

[0034] The rock mass hardness parameters, joint weathering parameters, groundwater parameters and initial geostress parameters in the surrounding rock grade evaluation parameters can be obtained through on-site collection.

[0035] In the embodiment of the present invention, the TBM excavation parameters include TBM ascending section excavation parameters and TBM stable section excavation parameters. The TBM ascending section excavation parameters usually show large changes, while the TBM stable section excavation parameters are relatively stable and change less.

[0036] S102, after performing feature discrete matching on surrounding rock grade evaluation parameters corresponding to each first sample image, determine the surrounding rock grade corresponding to each first sample image according to a surrounding rock grade prediction model obtained by pre-training.

[0037] In the embodiment of the present invention, the surrounding rock grade prediction model is obtained by training using a one-dimensional convolutional neural network (‌1D CNNs). The one-dimensional convolutional neural network can better process the local relationship in the sequence data, and the general expression of the neural network layer is: ; in, Indicates network The feature map of the layer output, Indicates The feature map of the layer output, Indicates The convolution kernel of the layer, represents the bias term, Represents the activation function.

[0038] In practical applications, the rock integrity parameters, rock structure type parameters, rock hardness parameters, joint weathering parameters, groundwater parameters and initial geostress parameters are discretely processed to form a one-dimensional feature vector which is input into a pre-trained surrounding rock grade prediction model to predict the surrounding rock grade corresponding to each first sample image.

[0039] S103. After screening the rock mechanical parameters based on the grey correlation method, a sample data set is established, wherein the sample data set includes characteristic parameters and their corresponding true label values, wherein the characteristic parameters include the surrounding rock grade, the screened rock mechanical parameters and the TBM ascending section excavation parameters, and the true label values ​​are the TBM stable section excavation parameters.

[0040] In an embodiment of the present invention, the rock mass mechanical parameters are screened based on the grey correlation method, specifically including the following steps: Generate characteristic sequences and normalize the data in each characteristic sequence, wherein the characteristic sequence includes a comparison sequence corresponding to each rock mass mechanical parameter and a reference sequence corresponding to each TBM stable section excavation parameter; The average correlation between each comparison sequence and each reference sequence is calculated respectively, and the rock mechanics parameters corresponding to the comparison sequences with an average correlation greater than a preset correlation are selected as the rock mechanics parameters for establishing the sample data set.

[0041] In practical applications, it is necessary to normalize the data in the feature sequence in order to eliminate the dimensional differences between the feature data. The data normalization formula in each feature sequence is as follows: ; in, Represents a certain original data in a certain feature sequence, is the minimum value in the data. is the maximum value in the data. It indicates the data obtained after normalization, ranging from 0 to 1.

[0042] In the embodiment of the present invention, the rock mass mechanical parameters include uniaxial compressive strength, uniaxial tensile strength, rock mass integrity coefficient and surrounding rock wear resistance value, then there are 4 comparison sequences, namely, the comparison sequence corresponding to the uniaxial compressive strength, the comparison sequence corresponding to the uniaxial tensile strength, the comparison sequence corresponding to the rock mass integrity coefficient and the comparison sequence corresponding to the surrounding rock wear resistance value, each of which contains the rock mass mechanical parameters corresponding to all the first sample images. The TBM stable section excavation parameters include thrust, penetration and torque. Then there are 3 reference sequences, namely, the reference sequence corresponding to thrust, the reference sequence corresponding to penetration and the reference sequence corresponding to torque, each of which contains the TBM stable section excavation parameters corresponding to all the first sample images.

[0043] In the embodiment of the present invention, respectively calculating the average correlation between each comparison sequence and each reference sequence specifically includes: Sorting the data in the comparison sequence and the reference sequence, calculating the absolute difference sequence between the comparison sequence and the reference sequence, and determining the minimum and maximum absolute difference values ​​in the absolute difference sequence; Calculate the correlation coefficient between the comparison sequence and the reference sequence, the calculation formula is as follows: ; in, Indicates The first The data and The correlation coefficient between the reference sequences is Indicates Comparison sequence and The absolute difference sequence of the reference sequence The absolute difference, represents the resolution coefficient, , represents the minimum absolute difference, Indicates the maximum value of the absolute difference; The correlation between the comparison sequence and the reference sequence is calculated using the following formula: ; in, Indicates Compare the sequence with The correlation between the reference sequences is Indicates the number of data in the comparison sequence; The average degree of association between the comparison sequence and each reference sequence was calculated using the following formula: ; in, Indicates The average degree of relevance between the comparison sequences and each reference sequence, Indicates the number of reference sequences.

[0044] For example, for the uniaxial compressive strength, the correlation coefficient of each data relative to each reference sequence in the reference sequence corresponding to the uniaxial compressive strength is first calculated, and then the correlation degree of the reference sequence corresponding to the uniaxial compressive strength relative to each reference sequence is calculated. Finally, the average correlation degree of the reference sequence corresponding to the uniaxial compressive strength relative to the reference sequence is calculated. If the average correlation degree is greater than the preset correlation degree, the uniaxial compressive strength is retained. If the average correlation degree of the uniaxial compressive strength is less than the preset correlation degree, the uniaxial compressive strength is eliminated, and so on, until the screening of all rock mechanical parameters is completed.

[0045] After obtaining the screened rock mass mechanical parameters, the screened rock mass mechanical parameters, surrounding rock grade and TBM ascending section excavation parameters are used as characteristic parameters to construct a sample data set that can represent the mapping relationship between the characteristic parameters and the TBM stable section excavation parameters.

[0046] S104. Train a CNN-LSTM model according to the sample data set to obtain a TBM excavation parameter prediction model.

[0047] In an embodiment of the present invention, training a CNN-LSTM model according to the sample data set specifically includes: Preprocessing the feature parameters in the sample data set to generate a matrix data set, and inputting the matrix data set and its corresponding true label values ​​into the CNN-LSTM model, wherein the preprocessing at least includes outlier removal and data optimization; The CNN-LSTM model uses the CNN layer to extract features and inputs the extracted features into the LSTM layer. The LSTM layer adjusts the model parameters according to the input features and uses the back propagation algorithm. Determine whether the loss function and evaluation index of the CNN-LSTM model meet the second preset condition. If so, obtain the TBM excavation parameter prediction model; otherwise, continue to adjust the model parameters. The evaluation index of the CNN-LSTM model includes at least the mean absolute error, the root mean square error, and the mean absolute percentage error.

[0048] It can be understood that the CNN-LSTM model combines the advantages of CNN's powerful spatial processing and extraction capabilities and LSTM's time series modeling capabilities. Among them, the CNN layer is mainly used for feature extraction. The pre-processed surrounding rock grade, screened rock mechanical parameters and TBM excavation parameters are sorted into a one-dimensional vector and input into the CNN layer of the CNN-LSTM model to achieve feature extraction. Please refer to Table 2 for the CNN layer network structure.

[0049] Table 2 CNN layer network structure table

[0050] The LSTM layer is composed of multiple LSTM units, each of which has an input gate, a forget gate, and an output gate, which control the flow of information. The LSTM layer takes the features extracted by the CNN layer as input and further learns the features. It contains a two-layer network structure with 60 neurons as a hidden layer, and predicts the TBM stable segment excavation parameters through a fully connected layer.

[0051] During the training process of the CNN-LSTM model, the regression task objective function uses the MSE loss function, the initial learning rate is 0.01, the category prediction uses the Softmax function, and the optimization algorithm uses the Adam algorithm. When the evaluation index of the CNN-LSTM model meets the second preset condition, the training ends and the TBM excavation parameter prediction model is obtained. The evaluation indicators of the CNN-LSTM model include mean absolute error, root mean square error, and mean absolute percentage error.

[0052] After obtaining the TBM excavation parameter prediction model, the TBM excavation parameter prediction model can be used to predict the TBM stable section excavation parameters. Figure 2 A flow chart showing a method for predicting TBM excavation parameters provided by an embodiment of the present invention is shown. Figure 2 , the method comprises the following steps: S201, obtaining surrounding rock grade evaluation parameters, rock mass mechanics parameters and TBM ascending section excavation parameters corresponding to the tunnel section to be predicted; S202, after performing feature discrete matching on the surrounding rock grade evaluation parameters corresponding to the tunnel section to be predicted, determine the surrounding rock grade corresponding to the tunnel section to be predicted according to the surrounding rock grade prediction model obtained by pre-training; S203, input the surrounding rock grade, rock mass mechanical parameters and TBM ascending section excavation parameters corresponding to the tunnel section to be predicted into the TBM excavation parameter prediction model to obtain the TBM stable section excavation parameters of the tunnel section to be predicted.

[0053] In practical applications, firstly, the surrounding rock grade evaluation parameters, rock mechanics parameters and TBM ascending section excavation parameters corresponding to the tunnel section to be predicted are obtained, and then the surrounding rock grade of the tunnel section to be predicted is determined by the same method as the first sample image. Finally, the surrounding rock grade, rock mechanics parameters and TBM ascending section excavation parameters corresponding to the tunnel section to be predicted are input into the TBM excavation parameter prediction model obtained by the TBM excavation parameter prediction model training method described in the embodiment of the present invention, so as to obtain the TBM stable section excavation parameters corresponding to the tunnel section to be predicted.

[0054] Figure 3A schematic diagram of a TBM excavation parameter prediction model training device is shown in FIG. Figure 3 , the device includes a processor, a memory and a communication bus; The communication bus is used to realize the connection and communication between the processor and the memory; The processor is used to execute one or more programs in the memory to implement the steps of the TBM excavation parameter prediction model training method described in the embodiment of the present invention.

[0055] It can be understood that since the prediction model training device for TBM excavation parameters described in the embodiment of the present invention is a device for implementing the prediction model training method for TBM excavation parameters described in the embodiment, for the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method, which will not be repeated here.

[0056] In summary, the prediction model training method and device and prediction method of TBM excavation parameters provided by the embodiment of the present invention use the grey correlation method to screen rock mechanical parameters, use the screened rock mechanical parameters as characteristic parameters, and use the surrounding rock grade of the surrounding rock face and the TBM ascending section excavation parameters as characteristic parameters to construct a TBM excavation parameter prediction model for characterizing the mapping relationship between characteristic parameters and TBM stable section excavation parameters. The grey correlation method is used to reveal the strength of the relationship between characteristic parameters, screen the main characteristic parameters, and avoid the problem of overfitting. At the same time, the TBM excavation parameter prediction model of the present invention combines the advantages of convolutional neural network (CNN) and long short-term memory network (LSTM), can process spatial features and time series data at the same time, improve the prediction ability and generalization ability of the model, can adapt to the differences of multi-source field data, and improve the accuracy of TBM excavation parameter prediction. The surrounding rock grade evaluation parameters are used and the surrounding rock grade is determined based on the surrounding rock grade prediction model, which improves the accuracy and consistency of the surrounding rock grade, thereby further improving the accuracy of TBM excavation parameter prediction. The image of the surrounding rock face is used and based on the fracture extraction model, the rock integrity parameters and rock structure type parameters in the surrounding rock grade evaluation parameters are determined, which improves the accuracy and consistency of the surrounding rock grade evaluation parameters, thereby further improving the accuracy and consistency of the surrounding rock grade and TBM excavation parameter prediction.

Claims

1. A prediction model training method for TBM excavation parameters, characterized in that: The method comprises: Taking the mileage pile number as a reference, a plurality of first sample images of the surrounding rock face of the TBM excavation are obtained, and surrounding rock grade evaluation parameters, rock mass mechanics parameters and TBM excavation parameters corresponding to each first sample image are determined, wherein the TBM excavation parameters include TBM ascending section excavation parameters and TBM stable section excavation parameters; After performing feature discrete matching on the surrounding rock grade evaluation parameters corresponding to each first sample image, the surrounding rock grade corresponding to each first sample image is determined according to the surrounding rock grade prediction model obtained by pre-training; After screening the rock mass mechanical parameters based on the grey correlation method, a sample data set is established, wherein the sample data set includes characteristic parameters and their corresponding true label values, wherein the characteristic parameters include surrounding rock grade, screened rock mass mechanical parameters and TBM ascending section excavation parameters, and the true label values ​​are TBM stable section excavation parameters; The CNN-LSTM model is trained according to the sample data set to obtain a TBM excavation parameter prediction model.

2. The prediction model training method for TBM excavation parameters according to claim 1, characterized in that: The surrounding rock grade evaluation parameters include rock mass integrity parameters, rock mass structure type parameters, rock mass hardness parameters, joint weathering parameters, groundwater parameters and initial geostress parameters.

3. The prediction model training method for TBM excavation parameters according to claim 2 is characterized in that: The method for determining the rock mass integrity parameter and the rock mass structure type parameter includes: The pre-trained fracture extraction model is used to extract the tunnel face joint fracture map corresponding to each first sample image, and the rock mass integrity parameter and rock mass structure type parameter corresponding to each first sample image are determined according to the tunnel face joint fracture map.

4. The prediction model training method for TBM excavation parameters according to claim 3 is characterized in that: The crack extraction model is obtained by training a deep convolutional neural network including a feature extraction layer, a size recovery layer and a skip connection structure, and the training method includes: Acquire multiple second sample images of the surrounding rock face of the TBM excavation, perform joint and fissure interpretation and annotation on each second sample image after preprocessing, and establish an image data set, wherein the image data set includes the second sample images and their corresponding annotation results; A deep convolutional neural network is trained according to the image data set, and it is determined whether the loss function and evaluation index of the deep convolutional neural network meet the first preset condition. If so, a crack extraction model is obtained, otherwise, the training parameters are adjusted and re-trained, and the evaluation index of the deep convolutional neural network includes at least accuracy, precision, recall rate, F1 score and average intersection-over-union ratio.

5. The prediction model training method for TBM excavation parameters according to claim 1, characterized in that: The rock mass mechanical parameters are screened based on the grey correlation method, including: Generate characteristic sequences and normalize the data in each characteristic sequence, wherein the characteristic sequence includes a comparison sequence corresponding to each rock mass mechanical parameter and a reference sequence corresponding to each TBM stable section excavation parameter; The average correlation between each comparison sequence and each reference sequence is calculated respectively, and the rock mechanics parameters corresponding to the comparison sequences with an average correlation greater than a preset correlation are selected as the rock mechanics parameters for establishing the sample data set.

6. The prediction model training method for TBM excavation parameters according to claim 5, characterized in that: Calculate the average correlation between each comparison sequence and each reference sequence, including: Sorting the data in the comparison sequence and the reference sequence, calculating the absolute difference sequence between the comparison sequence and the reference sequence, and determining the minimum and maximum absolute difference values ​​in the absolute difference sequence; Calculate the correlation coefficient between the comparison sequence and the reference sequence, the calculation formula is as follows: ； in, Indicates The first Data and The correlation coefficient between the reference sequences is Indicates Comparison sequence and The absolute difference sequence of the reference sequence The absolute difference, represents the resolution coefficient, , represents the minimum absolute difference, Indicates the maximum value of the absolute difference; The correlation between the comparison sequence and the reference sequence is calculated using the following formula: ； in, Indicates Compare the sequence with The correlation between the reference sequences is Indicates the number of data in the comparison sequence; The average degree of association between the comparison sequence and each reference sequence was calculated using the following formula: ； in, Indicates The average degree of relevance between the comparison sequences and each reference sequence, Indicates the number of reference sequences.

7. The prediction model training method for TBM excavation parameters according to claim 1, characterized in that: Training the CNN-LSTM model according to the sample data set specifically includes: Preprocessing the feature parameters in the sample data set to generate a matrix data set, and inputting the matrix data set and its corresponding true label values ​​into the CNN-LSTM model, wherein the preprocessing at least includes outlier removal and data optimization; The CNN-LSTM model uses the CNN layer to extract features and inputs the extracted features into the LSTM layer. The LSTM layer adjusts the model parameters according to the input features and uses the back propagation algorithm. Determine whether the loss function and evaluation index of the CNN-LSTM model meet the second preset condition. If so, obtain the TBM excavation parameter prediction model; otherwise, continue to adjust the model parameters. The evaluation index of the CNN-LSTM model includes at least the mean absolute error, the root mean square error, and the mean absolute percentage error.

8. The prediction model training method for TBM excavation parameters according to any one of claims 1 to 7, characterized in that: The rock mass mechanical parameters at least include uniaxial compressive strength, uniaxial tensile strength, rock mass integrity coefficient and surrounding rock wear resistance; the TBM ascending section excavation parameters and TBM stable section excavation parameters at least include thrust, penetration and torque.

9. A prediction model training device for TBM excavation parameters, characterized in that: The device includes a processor, a memory, and a communication bus; The communication bus is used to realize the connection and communication between the processor and the memory; The processor is used to execute one or more programs in the memory to implement the steps of the TBM excavation parameter prediction model training method according to any one of claims 1 to 8.

10. A method for predicting TBM excavation parameters, characterized in that: The method comprises: Obtain surrounding rock grade evaluation parameters, rock mass mechanics parameters and TBM ascending section excavation parameters corresponding to the tunnel section to be predicted; After performing feature discrete matching on the surrounding rock grade evaluation parameters corresponding to the tunnel section to be predicted, the surrounding rock grade corresponding to the tunnel section to be predicted is determined according to the surrounding rock grade prediction model obtained by pre-training; The surrounding rock grade, rock mass mechanical parameters and TBM ascending section excavation parameters corresponding to the tunnel section to be predicted are input into the TBM excavation parameter prediction model to obtain the TBM stable section excavation parameters of the tunnel section to be predicted. The TBM excavation parameter prediction model is trained by the TBM excavation parameter prediction model training method according to any one of claims 1 to 8.

Images